prepared by s. mustafa ali zaidi, 1032109...
TRANSCRIPT
LASER CUTTING OPTIMIZATION OF NON-METALLIC
MATERIALS OVERALL QUALITY
Prepared by
S. Mustafa Ali Zaidi, 1032109
Supervised by
Dr. Imran Amin (Head of Computer Science)
A Dissertation
Submitted to the
Faculty of Computing
of the
Shaheed Zulfikar Ali Bhutto Institute of Science and Technology
In partial fulfillment of the requirements for the degree of PhD
July 2nd , 2015
ii
CERTIFICATE OF APPROVAL
This is to certify that Syed Mustafa Ali Zaidi bearing registration no. 1032109 has
completed his dissertation, entitled LASER CUTTING OPTIMIZATION OF NON-
METALLIC MATERIALS OVERALL QUALITY in partial fulfillment of the
requirements for the degree of Doctor in Philosophy (Ph.D.) in the field of Computing,
under the supervision of Dr. Imran Amin (Associate Professor & Head of Computing).
The Dissertation meets the prescribed requirements and standard as set by Computer Science
Department, Shaheed Zulfiqar Ali Bhutto Institute of Science and Technology (SZABIST),
Karachi, Pakistan.
___________________________ ___________________________
Dr. Imran Amin
Associate Professor & Head of Computing
Supervisor
Dr. Husnain Mansoor
MS/PhD-Program Manager
Computing Department
___________________________
Dr. Mohammad Altaf Mukati
Dean (Computing & Engineering Sciences)
DECLARATION
I certify that this is my own research work. The work has not, in whole or in part, been
presented elsewhere for assessment. Where material has been used from other sources, it has
been properly acknowledged. If this statement is untrue and if I am found guilty of the
plagiarism, the punitive actions against me may be taken as per the SZABIST Anti-
Plagiarism Policy.
Signatures:_____________ Date:______________
Name of Student: Syed Mustafa Ali Zaidi
Registration No: 1032109
Degree: Ph.D
Program : Computing
DEDICATED
This Dissertation is dedicated to my Parents, wife and children
vi
ABSTRACT
As the global stock of natural resources depletes the need of electricity efficient processes
emerges. Laser cutting, an advance non-contact processing technique, outweighs the old
methods such as hotwire and milling due to the requirement of retightening and replacement
of cutting tools with time. Orthogonal array and Factorial design are selected as a design of
experiment for modelling and optimization of Laser cutting process. The range adjustment of
laser machine requires knowledge of experimental design, laser cutting process and material
properties, otherwise missing values generate due to unsuccessful cutting. For this reason,
many universities are unable to utilize these machines effectively. It is essential to formulate
a technique which allows modelling the data with some missing values, consequently, it
enhance the utilization of laser machines for research and other purposes. Initially, the
qualities of output characteristic were modelled by Statistical and Neural network without
missing values and then by supervised and novel Semi-supervised learning algorithms with
missing values. The Statistical modelling results using one and two way analysis of variance
with replication were better than other data mining techniques like linear and nonlinear
regression, however, it is difficult to use these methods with missing values. Therefore,
supervised neural network modelling is carried out and the effects of its parametric change
are observed along the datasets size to model the orthogonal array. The neural network
modelling results in edge quality and kerf width signal to noise ratio, it is acceptable, the edge
quality indicates that modelling improves by pre-normalization, further improvement was
made by increasing training data size to factorial design. It is observed that for the artificial
neural network, supervised learning is not sufficient associated to orthogonal array, only due
to edge quality mean modelling, average error were higher than the acceptable limit. The
average error with factorial design was under 10%. The vast modelling experience of
supervised learning engenders the development of novel Semi-Supervised learning algorithm.
Consequently, the average error was reduced by utilizing the systematic randomize
techniques to initialize the neural network weights and increase the number of initialization
by using orthogonal array design of experiment, with up to 22% missing values. This
algorithm reduces modelling time and cost thus reduces electricity consumption. The average
error in Perspex sheet did not exceed 8.0% and 11.5% for edge quality and kerf width
respectively.
vii
The overall quality was calculated by aggregation technique of data mining and a more
generous and better aggregation is carried out by the novel combination of Fuzzy logic which
provides overall quality for the customer while saving cost, time and Electricity.
viii
ACKNOWLEDGEMENT
Thanks to the ALLAH, the Almighty, for His countless blessings and bringing me to the
position where I am today. During the progress of my Doctor of Philosophy (Ph.D.) study at
Shaheed Zulfikar Ali Bhutto Institute of Science & Technology (SZABIST), Karachi,
Pakistan, I came across many persons and institutional entities that collaborated directly and
indirectly with my research work. Without having their support, it would have been
impossible for me to finish my Ph.D. work. Here I wish to acknowledge and recognize their
massive support.
I sincerely begin with acknowledging the huge support and guidance of my advisor, Dr.
Imran Amin who gave me the opportunity to conduct research under his worthy supervision.
I received motivation; encouragement and support from him during the entire period of my
PhD study. Under his supervision and guidance, I have learned writing papers for
conferences, journals and sharing my ideas to the public. I am also thankful for the
motivation, inspiration and support received from the faculty of SZABIST (Dr Husnain
Mansoor, Dr. Bushra Saeed, Dr. Faraz Junejo, Dr Faisal Ahmed Bukhari, Mr. Asim Riaz, Mr.
Asif Qazi, Wajeeh ul Hasan), University of Malaya (Professor Nukman Bin Yousuf,
Professor Imtiaz Ahmed and Mr. Tan See Bon), UTP (Dr Fawad Hasan Junejo), NED (Dr.
Najimi Ghani Haider) and Karachi University (Dr. Ejaz Ahmed).
The Scholarship from Higher Education Commission and research funding from SZABIST
Karachi and University of Malaya for the development of this MS/PhD research are also
sincerely acknowledged. In the end, I would like to thank my family Father: Mazahir Hussain,
Mother: Kaniz Fatima (Late), wife: Aelya and children: Baqar, Shaheer and Manahil for their
unconditional support, inspiration and love.
LIST OF PUBLICATION
JOURNAL PUBLICATIONS 1. M. Zaidi, et al., “Error Assessment of Laser cutting Predictions by Semi-Supervised
Learning,” Journal of Central South University, vol. 21, pp. 3736-3745, October
2014.
2. M. Zaidi et al., “Estimation of ANN Modelling of Laser Cutting with Missing Values.”
Pensee Journal, 75(11), 159-171, 2013.
3. M. Zaidi et al., “Experimental Data Mining Techniques (Using Multiple Statistical
Methods),” IJCSI International Journal of Computer Science Issues, Elsevier, vol. 9,
2012.
4. M. Zaidi, et a.l “Evaluation of Training Algorithms for the Laser Cutting Process,”
Journal of Independent Studies and Research-Computing (JISR), vol. 9, p. 1-13, 2011.
5. M. Zaidi, "Hybrid combination of Crisp and Fuzzy aggregation for Overall quality of
Laser Cutting Process, Submitted ISI indexed Journal for review.
ORAL PRESENTATIONS
International Conferences 1. M. Zaidi et al., “Laser Cutting problem modelling Using Statistical Tools,” Paper
presented in APIEMS: Asia Pacific Industrial Engineering and Management Systems
Conference, Melaka, Malaysia, 2010.
2. M. Zaidi, et al., “Laser Cutting Quality Control of Melamine Using Artificial Neural
Networks,” Paper presented in APIEMS: Asia Pacific Industrial Engineering and
Management Systems Conference, Melaka, Malaysia, 2010.
National Conferences 1. M. Zaidi et al. “Comparative Study of Customer Quality Function, Genetic Algorithm
and Fuzzy Logic in Laser Cutting Process,” Paper presented at the SZABIST 16th
National research conference (NRC) Karachi, Pakistan, 2010.
TABLE OF CONTENTS ABSTRACT .......................................................................................................................................................... VI
ACKNOWLEDGEMENT ...................................................................................................................................VIII
LIST OF PUBLICATION ..................................................................................................................................... IX
TABLE OF CONTENTS ........................................................................................................................................ X
LIST OF TABLES .............................................................................................................................................. XIV
LIST OF FIGURES ............................................................................................................................................. XX
LIST OF SYMBOLS AND ABBREVIATION ................................................................................................ XXIII
LIST OF APPENDICES .................................................................................................................................. XXVII
1. INTRODUCTION .................................................................................................................................... 1
1.2 PROBLEM STATEMENT ............................................................................................................................. 5
1.3 PURPOSE OF STUDY ................................................................................................................................... 6
1.4 CONTRIBUTION OF STUDY ....................................................................................................................... 7
1.5 OUTLINE ......................................................................................................................................................... 7
2. LASER CUTTING .................................................................................................................................. 10
2.2 LASER CUTTING TECHNIQUES ................................................................................................................ 11
2.2.1 VAPORIZATION CUTTINNG ................................................................................................................. 11
2.2.2 FUSION CUTTING ............................................................................................................................... 11
2.2.3 CHEMICAL DEGRADATION CUTTING .................................................................................................. 12
2.3 LOW POWER CO2 LASER............................................................................................................................ 12
3. DESIGN OF EXPERIMENT ..................................................................................................................... 21
3.1 GENICHI TAGUCHI’S METHODOLOGY .................................................................................................. 22
3.2 GENICHI TAGUCHI’S METHOD APPLICATIONS ................................................................................... 23
3.3 RESPONSE SURFACE METHODOLOGY .................................................................................................. 26
3.4 FACTORIAL DESIGN ................................................................................................................................... 27
3.5 SUMMARY .................................................................................................................................................... 31
4. NEURAL NETWORK MODELLING ......................................................................................................... 33
4.1 ARTIFICIAL NEURAL NETWORK ............................................................................................................. 33
4.2 FEED-FORWARD BACK-PROPAGATION NEURAL NETWORK .......................................................... 37
4.3 COMPARISON OF TRAINING ALGORITHMS.......................................................................................... 39
4.4 BACK PROPAGATION ALGORITHM ........................................................................................................ 40
4.4.1 GRADIENT DESCENT ALGORITHM ...................................................................................................... 41
4.4.2 GRADIENT DESCENT WITH MOMENTUM .......................................................................................... 42
4.5 FASTER TRAINING ALGORITHM ............................................................................................................. 43
4.5.2 Quasi Newton Algorithm.................................................................................................................... 44
4.6 LEVENBERG-MARQUARDT ALGORITHM ............................................................................................. 44
xi
4.7 OVERTRAINING OF DATA MODEL .......................................................................................................... 46
4.8 SUMMARY .................................................................................................................................................... 47
5. OVERALL QUALITY OPTIMIZATION ..................................................................................................... 49
5.1 GENETIC ALGORITHM ............................................................................................................................... 50
5.2 MULTI-OBJECTIVE STATISTICAL METHODS ....................................................................................... 51
5.3 FUZZY AGGREGATION .............................................................................................................................. 54
5.4 SUMMARY .................................................................................................................................................... 55
6. PROPOSED MODULAR RESEARCH METHODOLOGY ............................................................................. 59
6.1 PREPROCESSING OF PROCESS MODULE ............................................................................................... 61
6.2 EXPERIMENTAL DESIGN MODULE ......................................................................................................... 62
6.2.1 SINGLE PARAMETER CHANGE AT A TIME........................................................................................... 62
6.2.2 TAGUCHI METHOD ............................................................................................................................. 63
6.2.3 FRACTIONAL FACTORIAL DESIGN ....................................................................................................... 63
6.2.4 FACTORIAL DESIGN ............................................................................................................................ 63
6.2.5 RESPONSE SURFACE METHODOLOGY (RSM) ..................................................................................... 63
6.3 MODELING AND OPTIMIZATION MODULE ........................................................................................... 64
6.3.1 ANALYSIS OF VARIANCE ..................................................................................................................... 65
6.3.2 REGRESSION ANALYSIS ....................................................................................................................... 67
6.3.3 TAGUCHI, FRACTIONAL FACTORIAL AND FACTORIAL DESIGN ........................................................... 68
6.3.4 RESPONSE SURFACE METHODOLOGY (RSM) ..................................................................................... 69
6.3.5 ARTIFICIAL NEURAL NETWORK BY SUPERVISED LEARNING ............................................................... 70
6.3.6 NEURAL NETWORK BY SEMI-SUPERVISED LEARNING ........................................................................ 72
6.4 MULTI QUALITY OPTIMIZATION MODULE .......................................................................................... 75
6.4.1 SIMPLE AGGREGATION AND CUSTOMER QUALITY FUNCTION .......................................................... 77
6.4.2 GENETIC ALGORITHM ........................................................................................................................ 77
6.4.3 FUZZY AGGREGATION ........................................................................................................................ 78
6.5 SUMMARY .................................................................................................................................................... 80
7. EXPERIMENTAL DESIGN MODULE & SETUP ......................................................................................... 83
7.1 INTRODUCTION .......................................................................................................................................... 83
7.2 SCOPE AND LIMITATION .......................................................................................................................... 87
7.3 PLAN FOR EXPERIMENT DESIGN MODULE .......................................................................................... 88
7.3.1 PROCEDURE .............................................................................................................................................. 88
7.3.2 ORTHOGONAL ARRAYS AND FACTORIAL DESIGN ......................................................................... 90
7.4 EXPERIMENTAL SETUP ............................................................................................................................. 92
7.4.1 LASER MACHINE 500 WATTS ............................................................................................................. 92
7.4.2 MEASUREMENT TOOLS OF EXPERIMENT 1 and 2 .............................................................................. 94
7.4.3 PROPERTIES OF POLYSTYRENE FOAM ................................................................................................ 96
xii
7.4.4 PROPERTIES OF PERSPEX MATERIAL .................................................................................................. 96
7.5 DATA COLLECTION ................................................................................................................................... 97
7.5.1 POLYSTYRENE FOAM and PERSPEX SHEET ......................................................................................... 97
7.6 VERIFICATION OF SIMULATED DATA ................................................................................................... 99
7.7 SUMMARY .................................................................................................................................................... 99
8. DISCUSSION OF RESULTS AND ANALYSIS .......................................................................................... 102
8.1 PREPROCESSING OF EDGE QUALITY AND KERF-WIDTH DATA .................................................... 102
8.2 ANALYSIS OF VARIANCE ....................................................................................................................... 105
8.2.1 ONE WAY ANOVA WITHOUT REPLICATION...................................................................................... 106
8.2.2 ONE WAY ANOVA WITH REPLICATION ............................................................................................. 108
8.2.3 TWO WAY ANOVA WITH REPLICATION ............................................................................................ 109
8.3 REGRESSION ANALYSIS ......................................................................................................................... 112
8.3.1 LINEAR REGRESSION ANALYSIS ........................................................................................................ 112
8.3.2 MULTIPLE LINEAR REGRESSION ....................................................................................................... 121
8.3.3 NONLINEAR REGRESSION ANALYSIS ................................................................................................ 124
8.3.4 MULTIPLE NON-LINEAR REGRESSION .............................................................................................. 128
8.4 SUPERVISED LEARNING WITH MISSING VALUE ............................................................................... 129
8.4.1 EDGE QUALITY OF PERSPEX SHEET ................................................................................................. 131
8.4.2 KERF WIDTH QUALITY OF PERSPEX SHEET ....................................................................................... 138
8.5 SEMI-SUPERVISED ALGORITHM ........................................................................................................... 141
8.5.1 EDGE QUALITY OF PERSPEX SHEET .................................................................................................. 145
8.5.2 KW QUALITY OF PERSPEX SHEET ...................................................................................................... 146
8.6 FUZZY AGGREGATION ............................................................................................................................ 149
8.7 SUMMARY .................................................................................................................................................. 160
9. CONCLUSION .................................................................................................................................... 170
10. FUTURE DIRECTION .......................................................................................................................... 175
REFERENCE ..................................................................................................................................................... 178
A. DOE & STATISTICAL MODELLIN ......................................................................................................... 185
A.1 ONE WAY ANOVA WITHOUT REPLICATION ..................................................................................... 206
A.2 ONE WAY ANOVA WITH REPLICATION .............................................................................................. 207
A.3 TWO WAY ANOVA WITH REPLICATION ............................................................................................. 210
A.4 LINEAR REGRESSION ANALYSIS ........................................................................................................ 217
A.4.1 LASER POWER AND KERF WIDTH ..................................................................................................... 217
A.4.2 CUTTING SPEED AND KERF WIDTH .................................................................................................. 219
A.4.3 ASSIST GAS PRESSURE AND KERF-WIDTH ........................................................................................ 221
A.4.4 STANDOFF DISTANCE AND KERF WIDTH......................................................................................... 223
xiii
A.5 NONLINEAR REGRESSION ANALYSIS ................................................................................................ 225
A.6 MULTIPLE NON-LINEAR REGRESSION ............................................................................................... 236
B. NEURAL NETWORK & OVERALL QUALITY .......................................................................................... 238
xiv
LIST OF TABLES
TABLE 5-1: DIFFERENCE BETWEEN DMT AND TAGUCHI METHODS .................................................................................... 54
TABLE 7-1: FOUR VARIABLES WITH THREE LEVELS ORTHOGONAL DESIGN MATRIX ................................................................. 91
TABLE 7-2: CONTROLLABLE INPUT AND FACTORS LEVELS ................................................................................................. 91
TABLE 7-3: ZL1010 SPECIFICATION ............................................................................................................................ 93
TABLE 7-4: ZLX5 SPECIFICATION ................................................................................................................................ 93
TABLE 7-5: PROPERTIES OF POLYSTYRENE FOAM ........................................................................................................... 96
TABLE 7-6: PROPERTIES OF PERSPEX SHEET (CAST ACRYLIC) ............................................................................................ 97
TABLE 8-1: MEASUREMENT OF EDGE QUALITY OF POLYSTYRENE FOAM ........................................................................... 103
TABLE 8-2: MEASUREMENT OF KERF WIDTH OF POLYSTYRENE FOAM .............................................................................. 104
TABLE 8-3: OBSERVATIONS CONSIDER LASER POWER (A) .............................................................................................. 106
TABLE 8-4: SUMMARY OF DESCRIPTIVE STATISTICS ....................................................................................................... 106
TABLE 8-5: ANOVA FOR LASER POWER .................................................................................................................... 107
TABLE 8-6: ONE WAY ANOVA WITHOUT REPLICATION ................................................................................................. 108
TABLE 8-7: ONE WAY ANOVA WITH REPLICATION ...................................................................................................... 108
TABLE 8-8: INTERACTION BETWEEN LASER POWER AND CUTTING SPEED WITH REPLICATION ................................................ 109
TABLE 8-9: INTERACTION BETWEEN LASER POWER AND CUTTING SPEED 0.2 .................................................................... 109
ABLE 8-10: INTERACTION BETWEEN LASER POWER AND CUTTING SPEED 0.7 .................................................................... 110
TABLE 8-11: INTERACTION BETWEEN LASER POWER AND CUTTING SPEED 1.2 .................................................................. 110
TABLE 8-12: TOTAL INTERACTION BETWEEN LASER POWER AND CUTTING SPEED ............................................................... 110
TABLE 8-13: ANOVA OF INTERACTION BETWEEN LASER POWER AND CUTTING SPEED ....................................................... 111
TABLE 8-14: SUMMARY OF TWO ANOVA WITH REPLICATION ....................................................................................... 111
TABLE 8-15: DESCRIPTIVE STATISTICS OF LASER POWER AND KERF WIDTH ........................................................................ 113
TABLE 8-16: CORRELATION BETWEEN LASER POWER AND KERF WIDTH ............................................................................ 113
TABLE 8-17: REGRESSION STATISTICS ........................................................................................................................ 114
TABLE 8-18: REGRESSION BETWEEN LASER POWER AND KERF WIDTH ANOVA ................................................................. 115
TABLE 8-19: LINEAR REGRESSION LINE OF LASER POWER .............................................................................................. 116
TABLE 8-20: RESIDUAL OUTPUT ............................................................................................................................... 117
TABLE 8-21: LINEAR REGRESSION ANOVA ................................................................................................................ 121
TABLE 8-22: DESCRIPTIVE STATISTICS ........................................................................................................................ 122
TABLE 8-23: CORRELATIONS ................................................................................................................................... 122
TABLE 8-24: REGRESSION STATISTICS ........................................................................................................................ 122
TABLE 8-25: MULTIPLE LINEAR REGRESSION ANOVA .................................................................................................. 123
TABLE 8-26: LINEAR REGRESSION OF MULTIVARIABLE ................................................................................................... 123
TABLE 8-27: SUMMARY OF LINEAR REGRESSION .......................................................................................................... 123
TABLE 8-28: RESIDUAL OUTPUT............................................................................................................................... 124
TABLE 8-29: REGRESSION DATA WITHOUT REPLICATION FOR LASER POWER ...................................................................... 125
xv
TABLE 8-30: REGRESSION STATISTICS ........................................................................................................................ 125
TABLE 8-31: NONLINEAR REGRESSION ANOVA FOR LASER POWER AND KW WITHOUT REPLICATION ................................... 125
TABLE 8-32: NONLINEAR REGRESSION OF LASER POWER .............................................................................................. 126
TABLE 8-33: NON-LINEAR REGRESSION ANOVA ........................................................................................................ 126
TABLE 8-34: SUMMARY OF NON- LINEAR REGRESSION ................................................................................................. 127
TABLE 8-35: REGRESSION STATISTICS ........................................................................................................................ 128
TABLE 8-36: MULTIPLE NON-LINEAR REGRESSION ANOVA .......................................................................................... 128
TABLE 8-37: NONLINEAR REGRESSION OF MULTIVARIABLE ............................................................................................ 128
TABLE 8-38: RESIDUAL OUTPUT............................................................................................................................... 129
TABLE 8-39: PRELIMINARY TRAINING BY LEVENBERG MARQUARDT ................................................................................. 131
TABLE 8-40: ENRICHMENT TRAINING BY LEVENBERG MARQUARDT ................................................................................. 133
TABLE 8-41: EQ MEAN AND S/N RATIO WITH FACTORIAL DESIGN ................................................................................. 134
TABLE 8-42: AVERAGE PERCENT ERRORS .................................................................................................................... 148
TABLE 8-43: FOUR QUALITIES FACTORIAL DESIGN DATA ................................................................................................ 149
TABLE 8-44: QUALITY QUANTIFICATION ..................................................................................................................... 150
TABLE 8-45: COMPARE RESULTS OF Q AV, CQF, FL AV AND QFL AV ............................................................................... 151
TABLE 8-46: SORTED WITH CQF, QUANTIFIED FUZZY AGGREGATION AND QUANTIFIED NORMALIZED AGGREGATION ................. 155
TABLE 8-47: ENERGY CONSUMPTION QUALITY CALCULATION FOR FACTORIAL DESIGN .......................................................... 158
TABLE 8-48: POWER LEVELS CORRESPONDING TO LASER POWER AND CUTTING SPEED ......................................................... 159
TABLE 8-49: INCLUDING POWER CONSUMED AND THEN SORTED WITH CQF, Q AV AND QFL AV ........................................... 159
TABLE 8-50: SUMMARY OF LINEAR REGRESSION ANALYSIS ............................................................................................. 162
TABLE 8-51: SUMMARY OF NONLINEAR REGRESSION ................................................................................................... 162
TABLE A-1: FOUR VARIABLES WITH THREE LEVELS FACTORIAL DESIGN MATRIX ................................................................... 185
TABLE A-2: EDGE QUALITY MEAN AND S/N RATIO OF POLYSTYRENE FOAM SHEET (13MM) ................................................. 187
TABLE A-3: KERF WIDTH OBSERVATIONS OF POLYSTYRENE FOAM SHEET (13MM) .............................................................. 188
TABLE A-4: EDGE QUALITY OBSERVATIONS OF PERSPEX SHEET (3MM) ............................................................................. 188
TABLE A-5: KERF WIDTH OBSERVATIONS FOR OF PERSPEX SHEET (3MM) .......................................................................... 191
TABLE A-6: KERF WIDTH MEAN AND SIGNAL TO NOISE RATIO OF PERSPEX SHEET (3MM) ..................................................... 194
TABLE A-7: EDGE QUALITY OBSERVATIONS OF PERSPEX SHEET (5MM) ............................................................................. 197
TABLE A-8: OBSERVATION FOR KERF WIDTH OF PERSPEX SHEET OF 5MM ......................................................................... 200
TABLE A-9: KERF WIDTH MEAN AND SIGNAL TO NOISE RATIO OF PERSPEX SHEET OF 5MM ................................................... 203
TABLE A-10: OBSERVATIONS CONSIDER LASER POWER (A) ........................................................................................... 206
TABLE A-11: SUMMARY OF DESCRIPTIVE STATISTICS .................................................................................................... 206
TABLE A-12: ANOVA FOR LASER POWER .................................................................................................................. 206
TABLE A-13: OBSERVATIONS CONSIDER CUTTING SPEED .............................................................................................. 206
TABLE A-14: SUMMARY OF DESCRIPTIVE STATISTICS .................................................................................................... 206
TABLE A-15: ANOVA FOR CUTTING SPEED ............................................................................................................... 206
TABLE A-16: OBSERVATIONS CONSIDER ASSIST GAS PRESSURE ....................................................................................... 207
xvi
TABLE A-17: SUMMARY OF DESCRIPTIVE STATISTICS .................................................................................................... 207
TABLE A-18: ANOVA FOR ASSIST GAS PRESSURE ....................................................................................................... 207
TABLE A-19: OBSERVATIONS CONSIDER STANDOFF DISTANCE ........................................................................................ 207
TABLE A-20: SUMMARY OF DESCRIPTIVE STATISTICS .................................................................................................... 207
TABLE A-21: ANOVA FOR STANDOFF DISTANCE ........................................................................................................ 207
TABLE A-22: OBSERVATIONS CONSIDER LASER POWER WITH REPLICATION ....................................................................... 207
TABLE A-23: SUMMARY OF DESCRIPTIVE STATISTICS .................................................................................................... 208
TABLE A-24: ONE WAY ANOVA WITH REPLICATION FOR LASER POWER .......................................................................... 208
TABLE A-25: OBSERVATIONS CONSIDER CUTTING SPEED WITH REPLICATION ..................................................................... 208
TABLE A-26: SUMMARY OF DESCRIPTIVE STATISTICS .................................................................................................... 208
TABLE A-27: ONE WAY ANOVA WITH REPLICATION FOR CUTTING SPEED ........................................................................ 208
TABLE A-28: OBSERVATIONS CONSIDER ASSIST GAS PRESSURE WITH REPLICATION .............................................................. 208
TABLE A-29: SUMMARY OF DESCRIPTIVE STATISTICS .................................................................................................... 209
TABLE A-30: ONE WAY ANOVA WITH REPLICATION FOR ASSIST GAS PRESSURE ............................................................... 209
TABLE A-31: OBSERVATIONS CONSIDER STANDOFF DISTANCE WITH REPLICATION .............................................................. 209
TABLE A-32: SUMMARY OF DESCRIPTIVE STATISTICS .................................................................................................... 209
TABLE A-33: ONE WAY ANOVA WITH REPLICATION FOR STANDOFF DISTANCE ................................................................. 209
TABLE A-34: ONE WAY ANOVA WITH REPLICATION .................................................................................................... 209
TABLE A-35: INTERACTION BETWEEN LASER POWER AND CUTTING SPEED WITH REPLICATION .............................................. 210
TABLE A-36: INTERACTION BETWEEN LASER POWER AND CUTTING SPEED 0.2 .................................................................. 210
TABLE A-37: INTERACTION BETWEEN LASER POWER AND CUTTING SPEED 0.7 .................................................................. 210
TABLE A-38: INTERACTION BETWEEN LASER POWER AND CUTTING SPEED 1.2 .................................................................. 210
TABLE A-39: TOTAL INTERACTION BETWEEN LASER POWER AND CUTTING SPEED .............................................................. 210
TABLE A-40: ANOVA OF INTERACTION BETWEEN LASER POWER AND CUTTING SPEED ....................................................... 211
TABLE A-41: INTERACTION BETWEEN LASER POWER AND ASSIST GAS PRESSURE WITH REPLICATION ..................................... 211
TABLE A-42: INTERACTION BETWEEN LASER POWER AND ASSIST GAS PRESSURE 0.5 ......................................................... 211
TABLE A-43: INTERACTION BETWEEN LASER POWER AND ASSIST GAS PRESSURE 2.5 ......................................................... 211
TABLE A-44: INTERACTION BETWEEN LASER POWER AND ASSIST GAS PRESSURE 4.5 ......................................................... 211
TABLE A-45: TOTAL INTERACTION BETWEEN LASER POWER AND ASSIST GAS PRESSURE ...................................................... 212
TABLE A-46: ANOVA OF INTERACTION BETWEEN LASER POWER AND ASSIST GAS PRESSURE ............................................. 212
TABLE A-47: INTERACTION BETWEEN LASER POWER AND STANDOFF DISTANCE WITH REPLICATION ....................................... 212
TABLE A-48: INTERACTION BETWEEN LASER POWER AND STANDOFF DISTANCE 1 .............................................................. 212
TABLE A-49: INTERACTION BETWEEN LASER POWER AND STANDOFF DISTANCE 5 .............................................................. 212
TABLE A-50: INTERACTION BETWEEN LASER POWER AND STANDOFF DISTANCE 10 ............................................................ 213
TABLE A-51: TOTAL INTERACTION BETWEEN LASER POWER AND STANDOFF DISTANCE ....................................................... 213
TABLE A-52: ANOVA OF INTERACTION BETWEEN LASER POWER AND STANDOFF DISTANCE ................................................ 213
TABLE A-53: INTERACTION BETWEEN CUTTING SPEED AND ASSIST GAS PRESSURE WITH REPLICATION .................................... 213
TABLE A-54: INTERACTION BETWEEN CUTTING SPEED AND ASSIST GAS PRESSURE 0.5 ....................................................... 213
xvii
TABLE A-55: INTERACTION BETWEEN CUTTING SPEED AND ASSIST GAS PRESSURE 2.5 ....................................................... 214
TABLE A-56: INTERACTION BETWEEN CUTTING SPEED AND ASSIST GAS PRESSURE 4.5 ....................................................... 214
TABLE A-57: TOTAL INTERACTION BETWEEN CUTTING SPEED AND ASSIST GAS PRESSURE.................................................... 214
TABLE A-58: ANOVA OF INTERACTION BETWEEN CUTTING SPEED AND ASSIST GAS PRESSURE ........................................... 214
TABLE A-59: INTERACTION BETWEEN CUTTING SPEED AND STANDOFF DISTANCE WITH REPLICATION ..................................... 214
TABLE A-60: INTERACTION BETWEEN CUTTING SPEED AND STANDOFF DISTANCE 1 ............................................................ 215
TABLE A-61: INTERACTION BETWEEN CUTTING SPEED AND STANDOFF DISTANCE 5 ............................................................ 215
TABLE A-62: INTERACTION BETWEEN LASER CUTTING SPEED AND STANDOFF DISTANCE 10 ................................................. 215
TABLE A-63: TOTAL INTERACTION BETWEEN CUTTING SPEED AND STANDOFF DISTANCE ..................................................... 215
TABLE A-64: ANOVA OF INTERACTION BETWEEN CUTTING SPEED AND STANDOFF DISTANCE............................................. 215
TABLE A-65: INTERACTION BETWEEN ASSIST GAS PRESSURE AND STANDOFF DISTANCE WITH REPLICATION ............................ 216
TABLE A-66: INTERACTION BETWEEN ASSIST GAS PRESSURE AND STANDOFF DISTANCE 1 ................................................... 216
TABLE A-67: INTERACTION BETWEEN ASSIST GAS PRESSURE AND ASSIST GAS PRESSURE 5 .................................................. 216
TABLE A-68: INTERACTION BETWEEN ASSIST GAS PRESSURE AND STANDOFF DISTANCE 10 ................................................. 216
TABLE A-69: TOTAL INTERACTION BETWEEN ASSIST GAS PRESSURE AND STANDOFF DISTANCE............................................. 216
TABLE A-70: ANOVA OF INTERACTION BETWEEN ASSIST GAS PRESSURE AND STANDOFF DISTANCE .................................... 217
TABLE A-71: DESCRIPTIVE STATISTICS ....................................................................................................................... 217
TABLE A-72: CORRELATION ..................................................................................................................................... 217
TABLE A-73: REGRESSION STATISTICS ....................................................................................................................... 218
TABLE A-74: REGRESSION BETWEEN LASER POWER AND KERF WIDTH ANOVA ................................................................. 218
TABLE A-75: LINEAR REGRESSION LINE OF LASER POWER .............................................................................................. 218
TABLE A-76: RESIDUAL OUTPUT ............................................................................................................................... 218
TABLE A-77: DESCRIPTIVE STATISTICS ....................................................................................................................... 219
TABLE A-78: CORRELATION ..................................................................................................................................... 219
TABLE A-79: REGRESSION STATISTICS ....................................................................................................................... 219
TABLE A-80: REGRESSION BETWEEN CUTTING SPEED AND KERF WIDTH ANOVA ............................................................... 220
TABLE A-81: LINEAR REGRESSION LINE OF CUTTING SPEED ............................................................................................. 220
TABLE A-82: RESIDUAL OUTPUT ............................................................................................................................... 220
TABLE A-83: DESCRIPTIVE STATISTICS ....................................................................................................................... 221
TABLE A-84: CORRELATION ..................................................................................................................................... 221
TABLE A-85: REGRESSION STATISTICS ....................................................................................................................... 221
TABLE A-86: REGRESSION BETWEEN ASSIST GAS PRESSURE AND KERF WIDTH ANOVA ..................................................... 222
TABLE A-87: LINEAR REGRESSION LINE OF ASSIST GAS PRESSURE.................................................................................... 222
TABLE A-88: RESIDUAL OUTPUT ............................................................................................................................... 222
TABLE A-89: DESCRIPTIVE STATISTICS ....................................................................................................................... 223
TABLE A-90: CORRELATION ..................................................................................................................................... 223
TABLE A-91: REGRESSION STATISTICS ....................................................................................................................... 223
TABLE A-92: REGRESSION BETWEEN STANDOFF DISTANCE AND KERF WIDTH ANOVA ....................................................... 224
xviii
TABLE A-93: LINEAR REGRESSION LINE OF STANDOFF DISTANCE ..................................................................................... 224
TABLE A-94: RESIDUAL OUTPUT ............................................................................................................................... 224
TABLE A-95: REGRESSION DATA WITHOUT REPLICATION FOR LASER POWER...................................................................... 225
TABLE A-96: REGRESSION STATISTICS ....................................................................................................................... 225
TABLE A-97: NON-LINEAR REGRESSION ANOVA FOR LASER POWER AND KERF WIDTH WITHOUT REPLICATION ....................... 225
TABLE A-98: NONLINEAR REGRESSION OF LASER POWER .............................................................................................. 226
TABLE A-99: RESIDUAL OUTPUT ............................................................................................................................... 226
TABLE A-100: REGRESSION DATA WITH REPLICATION FOR LASER POWER ......................................................................... 226
TABLE A-101: REGRESSION STATISTICS ..................................................................................................................... 227
TABLE A-102: NON-LINEAR REGRESSION ANOVA FOR LASER POWER AND KERF WIDTH WITH REPLICATION........................... 227
TABLE A-103: NONLINEAR REGRESSION OF LASER POWER ............................................................................................ 227
TABLE A-104: RESIDUAL OUTPUT ............................................................................................................................ 228
TABLE A-105: REGRESSION DATA WITH REPLICATION FOR CUTTING SPEED ....................................................................... 228
TABLE A-106: REGRESSION STATISTICS ..................................................................................................................... 229
TABLE A-107: NON-LINEAR REGRESSION ANOVA FOR CUTTING SPEED AND KERF WIDTH WITH REPLICATION ........................ 230
TABLE A-108: NONLINEAR REGRESSION OF CUTTING SPEED .......................................................................................... 230
TABLE A-109: RESIDUAL OUTPUT ............................................................................................................................ 230
TABLE A-110: REGRESSION DATA WITH REPLICATION FOR ASSIST GAS PRESSURE ............................................................... 231
TABLE A-111: REGRESSION STATISTICS ..................................................................................................................... 232
TABLE A-112: NON-LINEAR REGRESSION ANOVA FOR ASSIST GAS PRESSURE AND KERF WIDTH WITH REPLICATION ................ 232
TABLE A-113: NONLINEAR REGRESSION OF ASSIST GAS PRESSURE .................................................................................. 232
TABLE A-114: RESIDUAL OUTPUT ............................................................................................................................ 233
TABLE A-115: REGRESSION DATA WITH REPLICATION FOR STANDOFF DISTANCE ................................................................ 233
TABLE A-116: REGRESSION STATISTICS ..................................................................................................................... 234
TABLE A-117: NON-LINEAR REGRESSION ANOVA FOR STANDOFF DISTANCE AND KERF WIDTH WITH REPLICATION ................. 235
TABLE A-118: NONLINEAR REGRESSION OF STANDOFF DISTANCE ................................................................................... 235
TABLE A-119: RESIDUAL OUTPUT ............................................................................................................................ 235
TABLE A-120: MULTIPLE NON-LINEAR REGRESSION DATA WITH REPLICATION FOR FOUR INPUTS........................................... 236
TABLE A-121: REGRESSION STATISTICS ..................................................................................................................... 236
TABLE A-122: MULTIPLE NON-LINEAR REGRESSION DATA WITH REPLICATION ANOVA ...................................................... 236
TABLE A-123: NONLINEAR REGRESSION OF MULTIVARIABLE .......................................................................................... 237
TABLE A-124: RESIDUAL OUTPUT ............................................................................................................................ 237
TABLE B-1: TRAINING ON FACTORIAL DESIGN .............................................................................................................. 238
TABLE B-2: EDGE QUALITY MEAN TRAINING USING FACTORIAL DATASETS .......................................................................... 239
TABLE B-3: EDGE QUALITY SIGNAL TO NOISE RATIO OF FACTORIAL DATASETS ..................................................................... 240
TABLE B-4: EDGE QUALITY MEAN OF NORMALIZED DATASET .......................................................................................... 240
TABLE B-5: KERF WIDTH MEAN TRAINING OF FACTORIAL DESIGN ..................................................................................... 241
TABLE B-6: SIMULATED DATA OF FACTORIAL DESIGN .................................................................................................... 246
xix
TABLE B-7: COMPARE RESULTS OF NORMALIZED AGGREGATION, CUSTOMER QUALITY FUNCTION AND FUZZY AGGREGATION ...... 249
TABLE B-8: SORTED WITH CQF, QUANTIFIED FUZZY AGGREGATION AND QUANTIFIED NORMALIZED AGGREGATION.................. 252
TABLE B-9: ENERGY CONSUMPTION QUALITY CALCULATION FOR FACTORIAL DESIGN ........................................................... 255
xx
LIST OF FIGURES
FIGURE 1-1: BLOCK DIAGRAM OF INPUT AND OUTPUT PARAMETERS ................................................................................... 2
FIGURE 1-2: PROBLEM STATEMENT SCHEMATIC VIEW ...................................................................................................... 5
FIGURE 2-1: VAPORIZATION CUTTING PROCESS ............................................................................................................. 11
FIGURE 2-2: FUSION CUTTING SCHEMATIC DIAGRAM BY C02 LASER ................................................................................... 12
FIGURE 2-3: LASER CUTTING PROCESS ......................................................................................................................... 13
FIGURE 2-4: CO2 LASER CUTTING SHEET IN SQUARE PROFILE ............................................................................................ 17
FIGURE 3-1: COMPARISON TAGUCHI’S OA AND FACTORIAL DESIGNS NUMBER OF RUNS ........................................................ 23
FIGURE 3-2: INCREASING TREND OF NUMBER OF RUNS AS FACTORS WITH TWO LEVELS .......................................................... 28
FIGURE 3-3: INCREASING TREND OF NUMBER OF RUNS AS FACTORS WITH THREE LEVEL ......................................................... 28
FIGURE 3-4: DIFFERENCE OF NUMBER OF RUNS IN 2 AND 3 LEVELS ................................................................................... 29
FIGURE 4-1: FEED-FORWARD BACK-PROPAGATION MODEL .............................................................................................. 38
FIGURE 4-2: SEARCHING OF GLOBAL MINIMA IN ERROR SPACE .......................................................................................... 42
FIGURE 6-1: PROPOSED FRAMEWORK MODULES FOR PROCESS IMPROVEMENT ................................................................... 60
FIGURE 6-2: PREPROCESSING OF PROCESS MODULE PART OF THE PROPOSED FRAMEWORK ................................................... 62
FIGURE 6-3: EXPERIMENT DESIGN MODULE OF THE PROPOSED FRAMEWORK ...................................................................... 62
FIGURE 6-4: MODELLING MODULE OF THE PROPOSED FRAMEWORK ................................................................................. 65
FIGURE 6-5: ANALYSIS OF VARIANCE ........................................................................................................................... 66
FIGURE 6-6: REGRESSION ANALYSIS ............................................................................................................................ 67
FIGURE 6-7: MODELLING AND VERIFICATION METHODOLOGY OF SUPERVISED LEARNING ...................................................... 71
FIGURE 6-8: MODELLING AND VERIFICATION METHODOLOGY OF SEMI-SUPERVISED LEARNING .............................................. 74
FIGURE 6-9: MULTI-QUALITY MODULE OF THE PROPOSED FRAMEWORK ............................................................................ 75
FIGURE 6-10: FUZZY AGGREGATION AS A MULTI QUALITY OPTIMIZATION .......................................................................... 79
FIGURE 7-1: GENERAL MODEL OF PROCESS ................................................................................................................... 83
FIGURE 7-2: LASER CUTTING PROCESS ......................................................................................................................... 84
FIGURE 7-3: TRAINING, TESTING AND SIMULATION PROCESS ............................................................................................ 86
FIGURE 7-4: PROCEDURE OF LASER CUTTING OF PERSPEX AND POLYSTYRENE FOAM ............................................................. 89
FIGURE 7-5: PICTURE OF ZECH LASER SYSTEM .............................................................................................................. 92
FIGURE 7-6: LASER CUTTING PROCESS ......................................................................................................................... 94
FIGURE 7-7: PICTURE OF DIGITAL CALIPER .................................................................................................................... 94
FIGURE 7-8: INNER AND OUTER SIDELINE LENGTH .......................................................................................................... 94
FIGURE 7-9: SCHEMATIC DIAGRAM OF LIGHT OPTICAL MICROSCOPE................................................................................... 95
FIGURE 7-10: SCHEMATIC DIAGRAM OF ICAMSCOPE ....................................................................................................... 95
FIGURE 7-11: PERSPEX SHEET EDGE QUALITY MEASURED BY MICROSCOPE .......................................................................... 95
FIGURE 7-13: VIEWS OF DEFECTIVE CUTS ..................................................................................................................... 98
FIGURE 7-12: MAXIMUM DEVIATION BETWEEN CUT EDGES ............................................................................................. 98
FIGURE 8-1: OUTLIER ANALYSIS OF EDGE QUALITY REPLICATION OF POLYSTYRENE FOAM ..................................................... 103
xxi
FIGURE 8-2: OUTLIER ANALYSIS OF KERF WIDTH REPLICATION OF POLYSTYRENE FOAM ........................................................ 104
FIGURE 8-3: INTERACTIVE GRAPH OF LASER POWER AND KERF WIDTH ............................................................................ 113
FIGURE 8-4: LASER POWER (A) LINE FIT PLOT ............................................................................................................. 117
FIGURE 8-5: INTERACTIVE GRAPH OF CUTTING SPEED AND KERF WIDTH .......................................................................... 118
FIGURE 8-6: CUTTING SPEED (B) LINE FIT PLOT ........................................................................................................... 118
FIGURE 8-7: INTERACTIVE GRAPH OF ASSIST GAS PRESSURE AND KERF WIDTH .................................................................. 119
FIGURE 8-8: ASSIST GAS PRESSURE (C) LINE FIT PLOT .................................................................................................. 119
FIGURE 8-9: INTERACTIVE GRAPH OF STANDOFF DISTANCE AND KERF WIDTH ................................................................... 120
FIGURE 8-10: STANDOFF DISTANCE (D) LINE FIT PLOT ................................................................................................. 120
FIGURE 8-11: COMPREHENSIVE INTERACTIVE GRAPH BETWEEN ALL PARAMETERS .............................................................. 121
FIGURE 8-12: QUADRATIC GRAPH OF LASER POWER WITHOUT REPLICATION ..................................................................... 124
FIGURE 8-13: COMPARISON OF EDGE QUALITY MEAN NETWORKS ERRORS ........................................................................ 136
FIGURE 8-14: EDGE QUALITY MEAN COMPARISON IN PERCENT ERROR FOR NORMALIZED DATASETS ....................................... 137
FIGURE 8-15: AVERAGE PERCENT ERRORS COMPARISON FOR EDGE QUALITY SIGNAL TO NOISE RATIO ..................................... 138
FIGURE 8-16: COMPARISON OF KERF WIDTH MEAN FACTORIAL DATASETS ........................................................................ 139
FIGURE 8-17: VARIATIONS IN 100, 500, 1000, 3000 RE-INITIALIZATIONS ..................................................................... 143
FIGURE 8-18: EFFECT OF WEIGHT INITIALIZATIONS ON AVERAGE ERROR............................................................................ 144
FIGURE 8-19: NEURON‘S VARIATIONS IN EQ MEAN ..................................................................................................... 145
FIGURE 8-20: NEURON'S VARIATIONS IN EQ S/N ........................................................................................................ 146
FIGURE 8-21: NEURON‘S VARIATIONS IN KW MEAN .................................................................................................... 147
FIGURE 8-22: NEURON'S VARIATIONS IN KW S/N ....................................................................................................... 147
FIGURE 8-23: VIEW OF ALL AGGREGATED VALUES ........................................................................................................ 152
FIGURE 8-24: COMPARISON OF QUANTIFIED NORMALIZED AGGREGATION WITH CQF ........................................................ 152
FIGURE 8-25: FUZZY INFERENCE SYSTEM FOR POLYSTYRENE SHEET CUTTING PROCESS ......................................................... 153
FIGURE 8-26: APPLICATION OF PARALLEL RULES WITH IMPLICATION ................................................................................ 154
FIGURE 10-1: BRIEF PROCEDURE TO SOLVE THE PROBLEM ............................................................................................. 175
FIGURE A-1: OUTLIER ANALYSIS OF EDGE QUALITY OF PERSPEX GLASS SHEET (3MM)........................................................... 191
FIGURE A-2: OUTLIER ANALYSIS OF KERF WIDTH OBSERVATIONS OF PERSPEX GLASS SHEET (3MM) ........................................ 194
FIGURE A-3: OUTLIER ANALYSIS OF EDGE QUALITY OBSERVATIONS OF PERSPEX GLASS SHEET (5MM) ..................................... 200
FIGURE A-4: OUTLIER ANALYSIS OF KERF WIDTH OBSERVATIONS OF PERSPEX GLASS SHEET (5MM) ........................................ 203
FIGURE A-5: INTERACTIVE GRAPH OF LASER POWER AND KERF WIDTH ............................................................................ 217
FIGURE A-6: LASER POWER (A) LINE FIT PLOT ............................................................................................................ 218
FIGURE A-7: INTERACTIVE GRAPH OF CUTTING SPEED AND KERF WIDTH .......................................................................... 219
FIGURE A-8: CUTTING SPEED (B) LINE FIT PLOT .......................................................................................................... 220
FIGURE A-9: INTERACTIVE GRAPH OF ASSIST GAS PRESSURE AND KERF WIDTH .................................................................. 221
FIGURE A-10: ASSIST GAS PRESSURE (C) LINE FIT PLOT ................................................................................................ 222
FIGURE A-11: INTERACTIVE GRAPH OF STANDOFF DISTANCE AND KERF WIDTH ................................................................. 223
FIGURE A-12: STANDOFF DISTANCE (D) LINE FIT PLOT ................................................................................................. 224
xxii
FIGURE A-13: QUADRATIC GRAPH OF LASER POWER WITHOUT REPLICATION .................................................................... 225
FIGURE A-14: QUADRATIC GRAPH OF LASER POWER WITH REPLICATION .......................................................................... 227
FIGURE A-15: QUADRATIC GRAPH OF CUTTING SPEED WITH REPLICATION ........................................................................ 229
FIGURE A-16: QUADRATIC GRAPH OF ASSIST GAS PRESSURE WITH REPLICATION ................................................................ 232
FIGURE A-17: QUADRATIC GRAPH OF STANDOFF DISTANCE WITH REPLICATION ................................................................. 234
FIGURE B-1: NO. OF TIMES INITIALIZATION COMPARISON OF 3MM SHEET EQ MEAN .......................................................... 242
FIGURE B-2: NO. OF TIMES INITIALIZATION COMPARISON OF 5MM SHEET EQ MEAN .......................................................... 242
FIGURE B-3: NO. OF TIMES INITIALIZATION COMPARISON OF 3MM SHEET EQ S/N ............................................................. 243
FIGURE B-4: NO. OF TIMES INITIALIZATION COMPARISON OF 5MM SHEET EQ S/N ............................................................. 243
FIGURE B-5: NO. OF TIMES INITIALIZATION COMPARISON OF 3MM SHEET KW MEAN ......................................................... 244
FIGURE B-6: NO. OF TIMES INITIALIZATION COMPARISON OF 5MM SHEET KW MEAN ......................................................... 244
FIGURE B-7: NO. OF TIMES INITIALIZATION COMPARISON OF 3MM SHEET KW S/N ............................................................ 245
FIGURE B-8: NO. OF TIMES INITIALIZATION COMPARISON OF 5MM SHEET KW S/N ............................................................ 245
FIGURE B-9: COMPARISON OF QUANTIFIED AGGREGATION WITH CUSTOMER QUALITY FUNCTION .......................................... 252
xxiii
LIST OF SYMBOLS AND ABBREVIATION
% Percent
ρ Density
A Laser power
AI Artificial Intelligence
ANN Artificial Neural Network
ANFIS Adapting network fuzzy inference system
ANOVA Analysis of variance
Av Aggregated value
Ave Average
B Cutting speed
C Assisting gas pressure
CCD Central Composite Design
CE Confirmation Experiment
CGA Common Genetic Algorithm
Class Response/ output characteristic
CNC Computer numerically controlled
CO2 Carbon dioxide
CPU Central Processing Unit
CQF Customer quality function
CW Continuous wave
D Standoff distance
DE Differential Evaluation
DM Data Mining
DMT Desirability, MSE and Transformation
DOE Design of experiment
DOEs Design of experiments
etc. etcetera
EPS Expanded Polystyrene
F Overall quality of flatness
FD Factorial Design
FIS Fuzzy inference system
FL Fuzzy logic
xxiv
GA Genetic algorithm
GD Gradient Descent method
GDM Gradient Descent with Momentum
GHz Giga hertz
GMDH Group method of data handling
GUI Graphical User interface
GRA Gray relational analysis
HAZ Heat Affected Zone
i.e. That is
kW Kilo Watt
ISL Inner scrap length
L9 9 runs orthogonal array
L18 18 runs orthogonal array
L27 27 runs orthogonal array
L9(34) LRun(Number of levelsNumber of variables)
LASER Light Amplification by Stimulated Emission of Radiation
LC Laser cutting
LBC Laser beam cutting
LCP Laser cutting process
LGP Light guide plate
LM Levenberg Marquardt
LMA Levenberg-Marquardt Algorithm
LTB Larger the better
Max Maximum
MDF Medium density fiber board
Millisecond ms
Millimetre mm
mW milli Watt
Min Minimum
MOGA Multi-Objective Genetic Algorithm
Mn-Mo Manganese- Molybdenum
MRR Material removal rate
MS Membership
xxv
MSE Mean square error
ns Nanosecond
ND:YAG Neodymium-doped Yttrium Aluminium garnet
NPN Noise performance measure
NTB Normal the better
OA Orthogonal Array
OFAT One factor at a time
PC Polycarbonate
PCA Principal Component Analysis
PCR Principal Component Regression
PE Polyethylene
PSO Particle swarm optimization
PMMA Polymethyl methacrylate
PMMA Plexiglass
POC Percent Overcut
PP Polypropylene
QC Quality characteristic
QFL Quantified fuzzy logic
QFN Quad Flat No-Lead
QNA Quasi Newton algorithm
QP Quadratic Programming
R2 Coefficient of determination
RF Radio frequency
RP Rapid Prototyping RP
RSM Response surface methodology
S/N Signal to noise
Sll Side line length
STB Smaller the better
SVM Support Vector Machine
SSLA Semi-supervised learning algorithm
Ra Surface roughness
Ta Kerf taper
TEM Transverse electromagnetic
xxvi
TM Taguchi Method
TPM Target performance measure
TQM Total Quality Method
V Volume
W Overall quality of waviness
XOR Exclusive OR (Logic gate)
xxvii
LIST OF APPENDICES
APPENDIX A DOE & STATISTICAL MODELLING
APPENDIX B NEURAL NETWORK & OVERALL QUALITY
CHAPTER 1
INTRODUCTION
1
1. INTRODUCTION Laser cutting is one of the latest technologies in machining materials which effectively cuts
and engraves materials through thermal cutting process [1]. Laser beam cutting and other
modern machining techniques such as ion beam, plasma beam, electron beam, electrical
discharge, electro-chemical, chemical, abrasive water jet and water jet machining are
increasingly being used as substitutes to the conventional machining methods[1, 2].
Machining by conventional methods has limits due to the irregular shapes of the work-piece
as well as its strict tolerances. In addition, new materials are introduced every day [1]. The
use of laser technology is justified because of its unaltered cutting quality, finishing and
reliability even with its high cost (however, cost is constantly falling) [2-4]. Laser cutting is
being used in a large variety of industries such as wood, metal, glass and plastic [1, 5-13]. It
is evident that the industries are increasingly adopting laser cutting (LC) techniques for
plastic or acrylic sheets to achieve precision and fine quality cuts of varying thickness by
adjusting laser power [1, 5, 13, 14]. Ahn and Tan [15] explained that the motivation of using
laser instead of hotwire cutting in the perspective of wire stiffness and quality variation which
are nonexistent in LC due to no-contact approach. Studies show that the effect of retightening
and replacement in hotwire and milling tool restricts quality management, which gives an
edge to LC over rest of the cutting methods. [1, 15-19].
The Proposed Framework can be applied on metallic and non-metallic materials but materials
for experimentation were selected keeping in view the low material cost. Therefore, Urea
Formaldehyde (Melamine), Polystyrene foam and Perspex glass sheet has been selected for
cutting process by CO2 laser. For these experimental investigations, one factor at a time
(OFAT) has been varied to analyze the effect of input process parameters on output quality
characteristics [13, 16, 18, 20-22] also mentioned that DOE is a better technique than OFAT,
its major benefit is that it describes the relationship between input parameters and the
responses as shown in Figure 1-1 [23]. Taguchi’s orthogonal array (OA) and Factorial
design (FD) were selected as an experiment design for the purpose of modelling and
verification [2-4, 11, 12, 24, 25]. Taguchi’s OA design is a robust, Statistically efficient and
less expensive method for high quality cutting systems. Its usage can reduce the number of
observations from 81 to 9 with three replications, hence it becomes energy efficient by
2
reducing time and cost. Factorial design allows to study Treatment and Interactions effects on
the response variables as both are significant factors in many industrial processes [23].
Input Parameters
Levels
Output parameters
Attributes
Laser Power Cutting speed Assist gas pressure
Stand off distance
100 300 500 0.2 0.7 1.2 0.5 2.5 4.5 1 5 10
Edge quality Mean
Edge quality S/N ratio
Kerf width Mean
Kerf width S/N ratio
Input/output Relations
Figure 1-1: Block Diagram of Input and Output parameters
The problem of modelling can be solved by the existing techniques such as transfer function,
curve fitting and Taguchi methods, or by training of experimental data to provide sufficient
knowledge of input parameters for LC machine for efficient cutting. To build a transfer
function is not an easy task because of variations in four variables. It would be difficult to
plot four inputs with one or more output quality parameters. The one and two way ANOVA
is able to model the problem with replication by considering interaction effects. The result of
nonlinear regression is comparatively worse with replication. The results of multi-linear
regression can be used for approximation [25]. However the data sets were without missing
values. Because of missing values unbalanced analysis tables generate different prediction by
Target Performance Measurement (TPM) and Noise Performance Measurement (NPM). So
NPM is preferred as a thumb rule for more robust solution. Modelling can be improved by
ANN as it models better than regression and other Statistical techniques [2, 4, 12, 26-35].
Sivarao, et al. [9, 33, 36-38] and other researchers used ANN and Fuzzy modelling with full
FD without any missing values, which is the area where improvements can be made. Pandey
et al. [39] used large size Taguchi method of L27 with two replications and L9. In their view,
interaction plays an important role and confidence interval is better in L9, This work has also
motivated us to perform this study with OA [3]. M. Zaidi et al. [3] used back-propagation
3
training algorithm for mapping Melamine cutting experiment using OA to expand the size of
data sets to FD and then utilize in aggregation for overall quality, the results were
encouraging. Researchers also claimed that ANN prediction of surface quality in laser cutting
process lessens the cutting cost by as much as 70% of the overall manufacturing cost [33]
Levenberg Marquardt (LM) training algorithms of feed forward back-propagation was
selected for trainings based on the previous study [4, 24] but verified with FD data. The
primary purpose is to select the appropriate training algorithms of neural networks. It is clear
that LM is the best choice but Gradient Descent with Momentum (GDM) cannot be ignored
even though its values were not promising and slow. The modelling of edge quality and Kerf
width S/N were better but the modelling of mean value was achieved by enlarging the span of
data. Small datasets and small span datasets modelling are difficult such as edge quality
mean. It was observed that ANN supervised learning is not sufficient with OA. After
increasing 80% dataset of FD, the average percent errors are acceptable. This study explains
the effects of changes in neural network parameters and the size of datasets for modelling
with supervised learning [4]. For this purpose, FD were selected, it estimates the prediction
errors considering by average, maximum, minimum error and MSE for each prediction. The
huge modelling experience gained in these studies which built the basis of Semi-Supervised
learning algorithm.[24] Semi-supervised learning shows more suitable modelling results than
supervised learning [4] with at least up to 22% missing values. The results depict a better
prediction on average by utilizing the systematic randomize techniques to initialize the neural
network weights and increase the number of initialization. Even with very small dataset of
OA by GDM and LM they were able to model the problem. This algorithm makes it easy for
the researchers having less insight in soft-computing and laser cutting. This achievement
encourages new researchers/engineers to utilize their expensive machines with the adjustment
of small size experimentation even with missing values due to inappropriate range
adjustments. The motivation of this study is to find minimum dataset sizes for modelling and
estimate the errors by changing neural network attributes to achieve better modelling.
Bahar and Golnabi [40] explained that most of the researchers optimize laser cutting process
for single quality while studying single input parameter at a time. Hence, utilizing large
experimental data and increasing cost and time. With reference to [3], the overall quality is
measured on the basis of very small data sets of OA based on experimental observation. The
ANN training was performed on L9 OA aggregated data and overall quality was predicted on
the basis of simulated data. In both papers [2, 3] OA design was used to take experimental
4
data. But in [25] normalized aggregation was used instead of simple aggregation on
simulated data. The aggregation function was further checked by customer quality function
[2]. Most of the researchers used aggregation methods for more than two inputs or more than
two quality parameters because of the constraint of plotting the data within 3 dimension.
Yilbas and Rashid [41] cut 800HT alloy by CO2 laser and performed experiment using
Factorial design (FD) for overall quality of flatness (F) and waviness (W). However, authors
used simple aggregation by converting the values in score i.e. 1, 2 and 3. The overall quality
is the summation of F and W. The calculation is simple and crisp in nature.
Pandey and Dubey [39] have used kerf width family as quality parameters. Overall quality is
achieved easily with the same settings for kerf width family quality parameters. Therefore,
there is no need of compromises or adjustments and can be solved with simple aggregation.
The inclusion of edge quality raises the need of overall quality rather than only kerf width
family. Syn at el. [42] used FD with five levels and three variables. However, Mamdani
fuzzy logic results are considerably good with the current research as well as other mentioned
researcher works [9, 36, 39, 42-46]. Sivarao et al. [43] also built fuzzy modelling of surface
roughness for LC by pressure vessel of Mn-Mo 5mm. He used FD of 128 observations. But,
the accuracy is less than 90%. Sivarao et al. [45] developed Mamdani fuzzy inference system
FIS with GUI. The prediction error is quiet high, maximum error is 72% and minimum error
is 1%. In author’s view, this is due to the lack of ability of Mamdani algorithm to capture the
nonlinearity of pulsed ND:Yag LCP. Sivarao at el. [9] himself built GUI-ANFIS Sugeno FIS
model for the surface roughness and kerf width on similar data [45]. The accuracy improves
to 91% consequently, error decreases to a value of 9% as compared to experimental values
which is adequate for any modelling system.
Further improvement was carried out by applying the customer quality function based on
customer specifications for better overall quality. However, the model is calculated based on
crisp logic and is unable to predict quality variable values separately. The FIS is working
with the real values of quality characteristics which provide better understanding of data
handling than CQF and normalizing the data. FIS generates better overall quality in fuzzy
manner as the rules sets is applied in parallel with aggregation. It provides unmatched results
in comparison with others. The novel combination of customer quality function, quantified
fuzzy aggregation and quantified normalized aggregation suggest overall optimized datasets.
5
In addition, electricity-efficient solution could be provided which is the requirement of
present day.
1.2 PROBLEM STATEMENT Experimental analysis has been carried out to model and optimize the laser cutting overall
quality of non-metallic sheets based on proposed framework modules with electricity-
efficient process as it is the need of time without compromising on desired quality. The
overall quality achieved by Laser cutting has an advantage over other methods like hotwire
and milling due to requirement of retightening and replacement of cutting tool with time.
Orthogonal array and Factorial design are selected as a design of experiment for modelling
and optimization. The range adjustment of laser machine input parameters requires
knowledge of DOE, laser cutting process and material properties otherwise missing values
generate due to unsuccessful cutting. It is the reason many universities have been unable to
utilize these machines to their potential. The need of a modelling technique which can model
the data with some missing values will raise the utilization of laser machines for research as
shown in Problem Statement Schematic View Figure 1-2.
Figure 1-2: Problem Statement Schematic View
Initially output qualities characteristic were modelled by Statistical and neural network
without missing values and then with missing values by supervised and novel Semi-
supervised algorithm. The Statistical modelling results of one and two way ANOVA with
Needs Overall LASER cutting quality
Issues:• Electricity efficient Process solution• Domain Experts Shortage in Universities• Small datasets
Consequence:• Need consideration of Electricity with desirable
Quality• Missing values generated
Solution:• Systematic experimentation• Better modelling techniques• Better overall quality measurement technique
6
replication were better than other data mining techniques like linear and nonlinear regression
but it is difficult to use these method with missing values. Therefore, supervised neural
network modelling is done and observed the effects of change of its parameters and size of
datasets for modelling. The neural network modelling results of edge quality and kerf width
signal to noise ratio were acceptable, however edge quality mean modelling was improved by
normalization and further improvement was made by increasing training data size to factorial
design. It is observed that ANN supervised learning is not sufficient for L9(34-2) orthogonal
array. Although only edge quality mean modelling average error were higher than required i.e.
46%. The average error with factorial design was under 10% and maximum individual error
was 25%. The vast modelling experience of supervised learning served as basis for the novel
Semi-Supervised learning algorithm and the average error was reduced, by utilizing the
systematic randomize techniques to initialize the neural network weights and increase the
number of initialization, by using orthogonal array design of experiment with up to 22%
missing values. This algorithm reduces modelling time and cost. The average error in
Perspex sheet did not exceed 8.0% to 11.5% in edge quality and kerf width.
The overall quality was calculated by data mining technique of aggregation and improved by
aggregation after normalization for equalizing the contribution of all qualities and predicts
unknown machining input setting by neural network model. Further improvement was
achieved by applying the customer quality equation based on customer specifications.
However, the model is calculated based on crisp logic and is unable to predict quality
variable values separately. A more generous and better aggregation is carried out by the novel
combination of Fuzzy logic to provide overall quality for customer and an indication for
rework at initial stage for saving cost and time.
1.3 PURPOSE OF STUDY Our goal is to propose the solution by framework modules for Process improvement using the
modular approach. The problem solves by Preprocessing of process module, Experimental
design module, Modelling and optimization module and multi-quality optimization module as
shown in Figure 6.1. Final output generates multi-objective optimized results which reduce
the financial loss of discarded items with electricity efficient solution and of desired quality.
In addition to this, built modelling algorithm to handle missing values and utilized smallest
possible datasets such as Taguchi’s orthogonal array.
7
1.4 CONTRIBUTION OF STUDY We have proposed the general “Process improvement framework modules” for process
improvement, which covers preprocessing, design of experiment, modelling and overall
quality calculation. Initially, preprocessing standard techniques were applied and avoided one
factor at a time, OFAT, preferring DOE techniques. The size of small datasets of orthogonal
array of nine observations was selected for modelling. Factorial design of eighty one
observations was selected for the estimation of modelling error and for overall quality
measurement. Performed detailed study of Statistical tools, pointed out the limitations and
solved them with Soft-computing techniques. Modelling is achieved on such small datasets of
orthogonal array including the missing values of 11% and 22% datasets by Supervised
learning and novel Semi-Supervised learning algorithm. The overall quality of any process
for two variables is a difficult task rather than Separate quality parameter. However, this
method is capable to optimize two and more than two quality parameters. This study
developed a new combination of Crisp and fuzzy aggregation for overall quality.
1.5 OUTLINE Chapter 2 (LASER CUTTING), briefly explains LASER and its application with the
justification of use of machining, it also explains Laser Cutting Techniques. In particular it
provides literature review of CO2 laser cutting to give details of its applications, machine
adjustment parameters, desired quality parameters used. Also provides concise information of
applied modeling techniques, optimization techniques and Design of experiment.
Chapter 3 (DESIGN OF EXPERIMENT), mentions the research work utilization of one
factor at a time, OFAT, technique for experimentation and shows that although the size of
experimental runs are huge in number, still are unable to study combined effects of different
variables. To overcome this problem benefits of DOE are outlined. The review of DOE
techniques used in laser cutting process is briefly mentioned. Reviews and discusses
important DOE methods used in laser cutting process. Reviews and explains vital DOE
techniques such as Taguchi Method, Factorial design and response surface methodology.
In Chapter 4 (NEURAL NETWORK MODELLING), history of Artificial Neural Network is
briefly mentioned along with a comparison of the Von Neumann architecture and brain
computing capability. It emphasizes on ANN applications while discussing and comparing
with other data mining techniques. The survey of modelling of laser cutting process by ANN
8
was carried out while considering number of inputs / outputs variables, DOE method(Size of
datasets),training algorithm and modelling methods. Explains feed-forward back-propagation
neural networks along with explanation and comparison of important training algorithms, and
justify its utility.
Chapter 5 (OVERALL QUALITY OPTIMIZATION), states that researchers generally
optimize laser cutting process for single quality and also use OFAT technique. The literature
review was performed to investigate better overall quality techniques for more than two
quality variables optimization. For more than two variables optimization, suitable data
mining techniques are discussed such as simple aggregation, normalized aggregation, genetic
Statistical techniques, Algorithm (GA) or fuzzy logic (FL). The objective is to find out the
overall quality (edge quality, kerf width, overcut and material removal rate) with Electricity-
efficiency to predict optimized input datasets. Also explains that novel Fuzzy aggregation is a
far better option for overall quality prediction.
Chapter 6 (PROCESS IMPROVEMENT FRAMEWORK MODULES), proposes Process
improvement framework modules and explains each module, with its further details for better
understanding to solve the problem of laser cutting process from pre-processing to overall
quality.
Chapter 7 (EXPERIMENTAL DESIGN MODULE & SETUP), explains the LCP with the
help of standard DOE format and points out all possible process variables. Similarly, it
describes ANN modelling process and identifies its parameters to be varied for finer training
process of ANN modelling. The Plan, Procedure and setup of experimentation with selected
DOE based machine setting, table of observation were included. The instruments
specification and work-piece material properties are recorded in this chapter with some
important definitions, formulae of quality parameters are also defined.
Chapter 8 (DISCUSSION AND ANALYSIS), briefly discussed the preprocessing of edge
quality and kerf width data. A detailed analysis of variance One/Two ANOVA with and
without replication was performed and discussed. Similarly, Linear, multiple linear, nonlinear
and multiple non-linear regression analysis were carried out and the results were discussed.
Though the First dataset was without missing values, still there is a need to improve the
modeling. Therefore, ANN supervised learning was attempted to model the problem with
9
missing values datasets of Perspex sheet. The experience of supervised learning was
discussed and provided advantage in the preparation of Semi-Supervised learning algorithm.
A detailed solution of overall quality was discussed from simple aggregation to combination
of Fuzzy aggregation.
In Chapter 9 (CONCLUSION), conclusions are outlined.
In Chapter 10 (FUTURE DIRECTION), future directions are outlined.
CHAPTER 2
LASER CUTTING
10
2. LASER CUTTING LASER is an electromagnetic radiation with the property of coherent and monochromatic
beam that propagates with insignificant divergence in the same direction and having a broad
range of frequency, power (mW to kW), laser beam density in continuous or pulsed power
and hence it finds wide applications in normal to scientific use, in medical and defense areas.
The Laser application for metallic and non-metallic processes is widely categorized into four
areas such as machining, surface, welding and laser assisted forming. It can be divided into
change of phase or require high energy to induce the phase change. This study is concerned
with laser cutting of non-metallic sheets like Perspex, melamine and Polystyrene foam which
are application of material processing [7] with the objective of achieving quality in the
cutting process[3].
Many production based industries follow total quality management (TQM) philosophy to
establish the system for continual improvement in the whole system including process
control. In process control, improvement provides more robustness, reliability and less time
to complete the process [47] . It may be the reason of selection of Laser beam cutting (LBC)
popularity. Many alternatives to conventional machining techniques are used in industries
for example electrical discharge, laser beam, water jet, abrasive water jet, ion beam, electron
beam, plasma beam, chemical and electro chemical machining. The rising applications
material processing of laser is due to requirement of high productivity, quality, material
utilization, reduction of finishing operation, processing cost, heat affected zone, kerf width,
edge quality and non-contact processing of CO2 and Nd–YAG lasers [7]. Laser cutting is a
common industrial application for organic compounds, metals, ceramics, polymers
composites, radioactive fuel and wood materials irrespective of their hardness. Diamond
cutting is ten times slower than laser in case of Textiles and hard brittle ceramics. These
materials can be cut by the combination of fusion, vaporization and chemical degradation
processes [5]. In the process non-metallic materials like carbon, plastic and wood, the laser
beam boils the surface and creates a keyhole. It creates basis of fast absorption due to
multiple times beam reflection and starting point gets deeper resulting in cut [7]. However,
Laser beam cutting has a variety of applications in fine cutting of sheet, metallic and non-
metallic and with finer precision in process [13] as pointed out in the study that laser cutting
is now commonly used in most industries. Cutting, welding and surface treatment of both
11
metallic and nonmetallic materials is carried out by various kinds of lasers at varied operating
powers. It is also mentioned by [47] that many major areas of industries have laser
applications like cutting, drilling, welding, thermal treatment and marking.
2.2 LASER CUTTING TECHNIQUES Caiazzo et al. [5] indicated that in general, a very large number of organic materials are
present which have a tendency of high surface absorption for the wavelength (10.6µm)
typically of CO2 laser. Thus, a CO2 laser with a power as low as 100 to 500 Watts may be
sufficient to cut materials such as glass, plastic, ceramics, rubber, paper, cardboard, fabrics,
wood, leather and organic compound. These materials can be cut by the combination of
fusion, vaporization and chemical degradation processes.
2.2.1 VAPORIZATION CUTTINNG Plexiglass (PMMC) is the most commonly used polymer cut by vaporization by laser cutting
consequential produce admirable finish [5]. The glossy finish is acquired by proper
adjustment of assist gas pressure parameter of laser cutting machine. The higher pressure
leaves streaks on the face of cut and it is slow leaving molten material on the face of cut. The
optimum pressure at slower side produces better finish [48, 49]. Vaporization cutting process
is shown in Figure 2.1.
Figure 2-1: Vaporization cutting process
2.2.2 FUSION CUTTING This fusion cutting phenomenon is similar to that of metal cutting with inert gases, since the
laser beam produces fusion while the covering gas removes the molten material, thus
generating the actual severing of the piece. Fusion cutting Schematic diagram by C02 laser is
shown in Figure 2.2. The cut edge and faces are macroscopically smooth with some streaks
12
which are produced by the melted material. If the cut edges of the finished product are to
acquire a glossy finish, the pressure and flow of the cutting gas (compressed air) must be
adequately low so as to make it possible for the residual molten material still present on the
cut faces and edges to solidify in a non turbulent manner [5]. Fusion cutting is mostly used
for thermoplastic polymers [50-53] such as polycarbonate (PC) and Polypropylene (PP),
polyethylene (PE) [48, 49].
Figure 2-2: Fusion cutting Schematic diagram by C02 laser
2.2.3 CHEMICAL DEGRADATION CUTTING Chemical degradation is used to cut thermosetting materials using higher power level and
temperature as compared to simple fusion cutting because of a three dimensional lattice.
Therefore, cutting speeds are generally lower for thermosetting materials as compared to
thermoplastics as reported by [15]. The schematic diagram of CO2 laser cutting shows
construction of nozzle which includes Laser beam, focusing lens, inlet of assist gas pressure
and nozzle with air and beam focused on work piece. In this process smoke is generated
which could be deposited on the cut surface [48, 50, 51] and there is need to take care of
safety issues [54].
2.3 LOW POWER CO2 LASER LASER, meaning light amplification by stimulated emission of radiation, emits infrared laser
radiations with a wavelength of 10.6 μm and posses overall efficiencies of approximately 10
to 13% mentioned. The laser-active medium in a CO2 laser is a mixture of 9.5% CO2, 13.5%
N2 and 77% He gases, where CO2 is the laser-active molecule. The stimulation of the laser-
active medium is attained by electrical discharge in the gas. There are different designs of
the CO2 laser such as: Transverse flow (cross-flow) laser, Fast-axial flow laser, Diffusion-
13
cooled slab laser and Sealed-Off laser. They can be driven in either the continues wave mode
or pulse mode as reported by [21]. LASER is applicable in ordinary to sophisticated devices,
in common use to scientific purposes, and in medical as well as defense as explained by [7].
Figure 2-3: Laser cutting process
Majumdar and Manna [7] define that the application of lasers for material processing are
broadly divided into four major categories; namely, laser-assisted forming, joining,
machining and surface engineering. The study focuses on machining process of laser cutting
of polystyrene foam and Perspex sheets which can be used as an initial work for
understanding cutting process optimization in other metallic and nonmetallic materials.
Laser can be of very low power mille watt to extremely high (1–100KW) focused power with
a high precision of spot size and delay time on to any kind of medium. Laser is better than
other electromagnetic radiations in terms of its spatial and temporal coherence, low
divergence, high continuous or pulsed power density, monochromic and propagation in a
straight line as reported by [7]. Laser material processing applications can be divided into two
classes
• Requires limited power and causes no change of phase.
• Requires high energy to induce phase change
High energy applications are cutting, welding, fusion, heat treatments on metallic and thick
materials. For limited power non-metallic materials are the area of application as well as thin
metal materials can be used. Laser power/efficiency and interaction-time are crucial as the
14
processes involves single or multiple phase changes within a very short time. In this class of
application CO2 and Nd–YAG lasers are practically used. Majumdar and Manna [7]
explained that CO2 laser is used in material machining; laser cutting is a common industrial
application for organic compounds, metals, ceramics, polymers composites and wood
materials irrespective of their hardness.
Many researchers are working in the field of machining of laser processing with different
angles. The relevant important researchers work is briefly discussed below.
Zhou and Mahdavian [13] developed the theoretical model to estimate the depth of cut with
cutting speed and laser power for different materials. They investigated theoretical and
experimental results for different materials. A semi-empirical equation is developed to
optimize or assist the cutting process and find out the laser power range selection. As CO2
laser wavelength absorption tendency for nonmetallic materials is high and often have a
characteristic of low thermal and thermal diffusion coefficients. Therefore, very low power of
60 watt of CO2 laser can cut non-metallic material like plastic board with good cutting quality
but releases toxic gases during the process which can be handled easily in glass covered table
based laser machines. Zhou and Mahdavian provide guidance for wood cutting and particle
board that keep cutting speed low for small Kerf width. The cutting of rubber material is
improved by assist gas.
However, it produces smoke in the cutting area. The theory and experiment shows that higher
energy is required for deeper cut. The theoretical model is improved by experimental result
by introducing constants in the equation by Zhou. Modeling is based on energy balance
method but ignores the loss of energy due to heat conduction in the surrounding. The
empirical equation is built on the basis of energy balance method but ignoring heat
conduction around the cutting area. Therefore, there is still need to build better modelling in
the laser cutting area. Caiazzo et al. [5] investigated plastic material cutting analytically and
systematically by CO2 laser. They used polyethylene, polypropylene and polycarbonate of
thickness ranging 2-10 mm in order to provide sufficient knowledge to the industry. He [5]
reported that the researchers conclude results after exhaustive examination of different
process parameters which is time consuming and cost ineffective. The exhaustive method is
used to optimize the input parameters. The cutting quality of polyethylene (PE) was low,
polypropylene (PP) was medium and polycarbonate (PC) was high.
15
Yusoff [10] used CO2 laser of 500 watt for the experiment of cutting Malaysian wood. He
studied wood characteristics like thickness, density and water content. The aim of analysis
was to obtain geometrical & dimensional accuracy, higher material removal rate and to avoid
wood burns, checking results of assist gas by using air or nitrogen. He also [10] concluded
that machine adjustments are dependent on the thickness, density and moisture content in
wood. The use of inert gas (nitrogen) generates better quality of kerf width. More thickness
or moisture consumed higher power. Almeida et al. [47] optimized the process parameters of
laser machine for the cutting of high cost and difficult to cut titanium material with pulsed
laser technique. They used alter six machine parameters and converted it into four level
ordinal variables and kept constant about seven parameters. They used factorial analysis. The
edge quality was improved but using large data set with high cost material. In my view the
size of experiment should be small which can apply novel semi supervised artificial neural
network or RSM.
Tan et al. [16] conducted experiments to find the optimum combination of parameters in laser
engraving process by CO2 laser machine on ceramic tile of 6 mm thickness. For optimization
Taguchi method was used for DOE. The engraving speed is the most significant factor for
depth of engrave, heat affected zone and material removal rate. However standoff distance
was most influential on kerf width. The study of M. Zaidi at al. [25] shows that interaction
effects may be possible in the laser cutting process over the selected ranges of input
parameters which is missing in the this study [16]. Nukman et al. [10] explained the effect of
machine parameters setting on the cutting quality of soft and hard wood. They demonstrated
the improvement in quality by varying assist gas pressure and used different nozzle design.
Also shows the vital importance of speed in achieving better productivity. He also discussed
noise or uncontrollable parameters such as shield gas pressure before empty cylinder. He also
explained the effect of standoff distance on beam width. The beam changes due to focal
point on the work piece which indeed affect the cutting quality [10].
Imtiaz at al. [1] performed an experimental investigation on polymeric materials and
developed a model equation relating input parameters with the output. The quality parameters
are heat affected zone, surface roughness and dimensional accuracy. Exhaustive experiments
were not conducted, only twelve sets of experiment were taken for each polymer material
using the technique of central composite design. The Response surface methodology was
16
used as a predictive model, further analysis was done by ANOVA. The analysis is based on
dimensional deviation and concludes that Polymethyl methacrylate PMMA has less heat
affected zone than PC and PP. In case of surface roughness PMMA has much better edge
quality than PC and PP. The response model can be used in the industry for production. With
reference to paper [1] laser cutting processing proved advantageous in industries and is able
to improve by increasing the ordinal variable beyond low or high. At least three numbers of
levels are more appropriate for empirical modelling and need better techniques to handle
higher polynomials or other non-linear modelling.
Berrie and Birketi [55] investigated the effects on Perspex sheet by varying lens position,
focal plane, cutting speed, applied power on the machining process of cutting and drilling
rates. The experimental results were compared with the developed theoretical models of the
thermal conductivity and the authors studied vapors removal. The results support the theories
of vapors removal. But they did not study the effects on kerf width or edge quality of Perspex
sheet.
Balkin and Lin [56] study the optimum level of the independent variables 4 to 9 and use only
2 variables with quadratic regression and ANN. The results were discussed for different
levels of input parameters. It shows that improvement is possible but number of observation
increased in L(Runs) (Levels) (Variables) . In the studied case only two variables with observation
for 4 levels will be 16 and similarly for others 25, 36, 49, 64 and 81 respectively. Therefore,
it is important to select less number of controllable parameters with optimum number of
levels.
M. Zaidi at al. [25] studied different statistical modelling techniques of Polystyrene cutting
process such as one and two way ANOVA, were able to model the problem with and without
replication. The model gives better results in case of replication showing suitable analysis
technique for given datasets. The interaction should be considered to get a better picture of
the process optimization. The 89.9% highly significant value of R2 encourages using the
multiple linear regression model because Kerf width variation can be explained by the
selected input variables. It can be used in rough modeling, simulation and optimization. The
results of nonlinear regression are worst compare to others and with replication become more
non-realistic due to increase in the number of observation over and above the fitted points
similar to linear regression. The average error reaches to 50%.
17
The best method was one way ANOVA with pooling but the current research shows that
there is one thing missing that is significantly participating in the variation of dependent
parameters i.e. interaction between two independent and one dependent parameters. The
discussion of interaction above shows that three combinations of interactions were
significantly participating in the variation of dependent variable.
Therefore, modelling can be improved with the combination of Treatments and Interaction
based design of experiment. The results of regression are highly inferior than ANN as
mentioned in [2, 25]. Many researchers handled missing value but did not apply ANN in the
Laser cutting process to solve the range adjustment problem. The modern techniques of
regression [25] had higher prediction errors. Therefore, it is better to use novel method of
Semi-supervised algorithm for process modelling [12]. The detailed experiment of Perspex
CO2cutting in square profile is shown in Figure 2.3
Figure 2-4: CO2 laser cutting Sheet in square profile
C.B. Yang, et al. [35] prepared a model of a CO2 laser cutting process of polymethyl
methacrylate PMMA using orthogonal array of L9 for four factors with three levels at the
first stage, which is similar to our modelling with novel semi-supervised learning algorithm
[12]but built empirical relationship between the input and response variables and increased
18
the data size for ANN. The modelling with L9 orthognal array in single stage is difficult
which has been done by M. Zaidi et al. in [12]. The result shows that thickness is a major
parameter in the modelling process. The increase in the size of data improves the ANN
accuracy.
Noor and Kadirgama [57] prepared a statistical linear model to predict surface roughness and
roughness height based on laser power, cutting speed and tip distance. The model is prepared
for acrylic (3mm thick) cutting by 30W pulse CO2 using the Box Behnken (RSM) design.
The first order model shows that laser power is the major factor. In the next study they [57,
58] prepared quadratic RSM model which also considered the interaction and quadratic
effects. The interaction effects are insignificant but quadratic is significant and also shows
that second order is more close to experimental values. In addition to these two studies they
[8] developed artificial intelligent (AI) model using partcle swarm optimization (PSO) to
predict the optimum surface roughness for 30W pulsed Nd-Yag cutting acrylic sheets while
using Box-Behnken Response surface method (RSM) to reduce the data size of experiments.
They also explained the effect of machine settings of cutting speed, material thickness, gap of
tip and power towards surface roughness.
Combination of slow cutting speed of cut, high laser power and optimum standoff distance
construct superior surface roughness. Some defects were found in the microstructure such as
burning, melting and wavy surface.
Meanwhile, Ciazzo et al. [5]found that all of the three materials generally follow the rule
(which the results of experiments on ferrous and nonferrous metals have already been amply
validated) according to which the value of Ra diminishes as cutting speed increases. However
Ra values are much lower compared to typical construction steel. His research showed that
the surface roughness is significantly affected by the tip distance followed by the power
requirement, cutting speed and material thickness.
2.4 SUMMARY Many researchers have modelled the laser cutting problem to build theoretical or
experimental modelling techniques [1-3, 13, 16, 25, 47, 55]. Basically the problems in laser
cutting process are
• Process is non-linear
19
• Process is multivariable
• Process adjustment parameters are large in number
• Noise can disturb the quality due to human error or environment.
• Multi-output qualities
• Day to day new materials introduce
• Day to day new laser machines introduced
Many researchers and Yang et al. [35] quote that improper controllable machine parameters
will result in poor quality of laser cutting. However, numerous researchers are doing research
on the same direction as Yang et al. i.e. adjusts machine parameters properly. Therefore, it is
important to build easy to use, low cost, less time consuming and less complicated techniques
for old and new engineers and researchers. There is a need of Statistics expert for the
modelling and analysis of non-linear and multi-variable process. A number of researchers
have used regression, ANOVA, Taguchi, RSM, factorial analysis and multi-output
optimization techniques[25] .
There are a lot of researchers using AI techniques to model the laser cutting process [2-4, 12].
But, the need of expert does not diminish such as supervised algorithms in artificial neural
networks needs AI experts and experience counts in the results. There is a need to build
some semi-supervised algorithms based on his experience of supervised learning. Laser
cutting with missing values is also an important issue to resolve which can be resolved better
by ANN ( LM back-propagation algorithm) than statistical methods [4] for experienced
researcher and Semi-supervised is much better for very small data sets [12]. Another issue of
multi-output qualities is that the techniques are very complicated sometimes [24, 30, 59-63].
Some techniques are based on crisp logic and some are fuzzy logic based. The need of hybrid
technique is a requirement of time to solve the issue in fuzzy manner with some strict
boundary of acceptable quality. But, expertise of DOE is required to select the proper design
of experiment. Normally, researchers avoid this technique and solve their problems by
changing single variable at a time or without DOE in detail which increases the number of
experiments taking more time and cost [5]. It is also possible that interaction between the
variables is present and variation in a single variable is unable to be recorded and analyze
these issues.
20
It is necessary to further investigate the DOE, modeling and overall quality optimization.
Therefore, In the next Chapter 3, we will discuss DOE in detail for better understanding and
its appropriate selection for the LCP. In Chapter 4, we will discuss ANN modeling and in
Chapter 5 overall quality optimization. Finally, propose the Process improvement Framework
modules for solution approach.
CHAPTER 3
DESIGN OF EXPERIMENT
21
3. DESIGN OF EXPERIMENT In most of the experimental investigations of Laser beam cutting (LBC) process, one factor at
a time (OFAT) has been varied to analyze the effect of input process parameters on output
quality characteristics or responses [13, 16, 18, 20-22]. Bahar and Golnabi [40] also varied
only one parameter at a time in the study of the response on output quality for laser cutting
of steel and mild steel and have not adopted any DOE technique but relied on OFAT to
observe the effect on output quality. They analyzed the changes individually on output based
on
• Speed of cutting
• Applied laser power
• Thickness of sheet
• Oxygen gas pressure
The author performed 400 runs to understand the effects of input parameters. However, this
technique requires a large number of experimental runs as just one factor is changed in each
run, with all other factors remaining constant. In this technique, the interaction effects among
different input process parameters are not considered.
To overcome these problems, it would be better to use any scientific DOE which will save
time and cost. Viles et al. [64] and Sivarao et al. [23] highlighted the significance of
Statistical DOE application and discussed problems in the preliminary stages of DOE
recommending that the scientific community should write and present more DOE
applications focusing on the preliminary stages of the methodology. It would be helpful in
bridging the gap existing between research and industry.
They mainly focused on Taguchi method, Factorial Design and Response surface
methodology in laser cutting process. Sivarao et al. [23] also mentioned that DOE is a better
technique than OFAT and its major benefit is that it shows the relationship between input
parameters and responses. It can also show the interaction effects on responses, mathematical
model developed are normally used for prediction which can predict the possible optimize
response. It also saves time due to its well planed method. Sometimes experiments will have
some error, some of which might be predictable while others are just uncontrollable. DOE
22
allows us to handle these errors while still continuing with the analysis. DOE is an excellent
choice when it comes to predicting linear behavior. However, in non- linear behavior, DOE
does not always give the best results [23]. It means there are chances of unsuccessful results.
Therefore, it is possible to select more appropriate DOE techniques for better results.
3.1 GENICHI TAGUCHI’S METHODOLOGY Taguchi’s orthogonal array (OA) design of experiments (DOE) is a robust, efficient
Statistical method for designing high quality systems at lesser expense for laser cutting
process. It designs a systematic and efficient way to optimize controllable parameters with
the assumption that there is no interaction among controllable factors and ignoring the
variation caused by uncontrollable factors. It is able to provide optimized robust input data
setting [2, 12, 16, 23].
Besseris [65] explained that Taguchi approach is drawn for a large number of the associating
parameters that changes the output quality parameters comparatively larger than other
parameters variations. Taguchi optimization of any process or product variables may be
analyzed in eight different steps.
1. Evaluate the importance of the given problem and set the required outcome which
needs to be achieved by analyzing and solving this issue.
2. Collection of implicating quality characteristics that need attention along with the
respective controlled and noise factors that may affect these quality factors.
3. Investigation and classification of a significant physical range the factors are allowed
to vary during experimentation.
4. Selection of the proper orthogonal array that will fit all the considered factors and
possible interactions along with the appropriate levels.
5. Execution of trial runs according to a randomized schedule dictated by the selected
orthogonal array.
6. Analysis of the experimental data with nonparametric tests for small samples.
7. Determination of optimum settings for all statistically significant factors and
interactions.
8. Confirmation of prediction outcomes by conduction of experimental runs at the
optimum level values obtained from the previous step.
23
Orthogonal array were discovered more than a hundred years back in 1988 by Ross, they are
still a dependable source of planning for experimenters. With reference to [66] Genichi
Taguchi’s approach to ensuring quality is based on building robust product and process
designs. He taught controlling process quality by design of product and their process instead
of inspection. He focuses on selecting factors levels of controllable parameters that reduce
variability, bringing the process on target which is the secondary objective. Taguchi’s method
is one of the better designs of experiment. As large number of variables with higher number
of levels makes Factorial design useless, however Taguchi’s method works as shown in
Figure 3-1.
Figure 3-1: Comparison Taguchi’s OA and factorial designs number of Runs
3.2 GENICHI TAGUCHI’S METHOD APPLICATIONS The aim of robust process design is to build a stable process that produces minimum
variation due to uncontrollable parameters [60]. J. Kunert et al. [67] compared Taguchi’s
combined and product array to explained robust parameters design, the experimenter looks
for settings of controllable factors close to the target, with little variation. The difference is
caused due to two sources noise and pure error. Taguchi method produces a robust design but
it is better to calculate the variation with the help of an outer array by identifying the possible
noise factors in outer array. The results of outer array are better than product array or simple
orthogonal array. However the results of Genetic algorithm are even better than outer array
Taguchi method. Chatsirirungruang and Masami Miyakawa [68] studied the usefulness of
genetic algorithm (GA) as compared to orthogonal array’s Taguchi method. They used signal
to noise ratio as a deciding factor as is a routine practice of Statisticians but sometimes this
9
3
81243 729
2187
6561
1
27
9 27 81243
729
0
1000
2000
3000
4000
5000
6000
7000
2 3 4 5 6 7 8
Num
ber
of o
bser
vatio
ns
Number of variables
Factorial Design
Orthogonal Array
Assume: Three leves of each variable
24
criteria is not true. Therefore, noise factor (outer array) is used to solve the problem more
efficiently in his all papers [68-70].
M. A. Wazed et al. [71] explained that the essential part of managing uncertainty is pointing
out maximum sources and factors of uncertainty. This article is based on detailed review on
manufacturing under uncertain conditions. Different sources and factors of uncertainty in
manufacturing/production setting have been pointed out and focus has been shifted on the
techniques and models used to deal with them. Uncertainty can also be from its sources as
follows:
• Physical deviation observed in a system because of natural uncertainty. It is also
known as physical randomness and inherent uncertainty.
• Many assumptions for simplification in prediction and analytical models are the
reason for model uncertainty.
• Measurement uncertainty due to limited capability of measurement techniques and
inferior ability of measurement instruments.
• Environment and operational uncertainty.
• Statistical uncertainty results due to incomplete Statistical data (missing value issue)
and use of improper sample size or inappropriate sample selection to approximate the
characteristics of these parameters.
• Subjective uncertainty resulted due to human errors in calculation, improper
parameter selection, judgment and fabrication.
All types of uncertainties are not present in all process systems at the same time. However,
sometimes one or more than one type of uncertainty can significantly participate in the
system variations.
W.Q. Meeker [72] explained the relationship between engineering quality, dependability and
role of Statistics, Statisticians and brief Statistical tools applied in the area of reliability.
Quality and reliability have both improved with the Statistical tools that were originally made
to control and improve quality such as process monitoring and designed experiments. Every
passing day brings advancement in scientific knowledge thus Statisticians will have an
important part to play in the area of reliability in future as well.
25
Many experimental studies used well known L9, L18 and L27 Taguchi’ orthogonal arrays [3,
12, 16, 28, 35, 39, 61] . The importance of the Taguchi method is demonstrated when many
researchers use Taguchi DOE with other methods to make hybrid combinations and produce
better results and in multi-quality cases they used Taguchi at the first stage for single and
grey relational analysis and fuzzy for multi- qualities. Following are some examples of
Taguchi DOE.
Kuo et al. [28] utilized slightly larger data sets of L18 orthogonal array DOE for the
modelling of laser processing (CO2) by back propagation with Levenberg-Marquard training
algorithm. They used 13 datasets for training and remaining 5 for validation. The average
percent error was 5% which is a better result with L18. C.B. Yang, et al. [35] prepared a
model of a CO2 laser cutting process of polymethyl methacrylate PMMA using orthogonal
array of L9 for four factors with three levels at the first stage and built factor response table
which shows that thickness is the major factor. In the second stage they used hybrid
technique with ANN. The new combined model is called progressive Taguchi neural
network model and its prediction accuracy will improve with small datasets by increasing the
size of data set with Taguchi and model with ANN. Ming-Fei et al. [61] applied the grey
relational analysis method to study the multi-quality optimization of two responses (surface
roughness and optical transmittance ratio) as a result of PMMA (thickness 6mm) CO2 laser
cutting. The orthogonal array of L9 Taguchi DOE was used for multi quality.
Pandey and Dubey [39] have used large size Taguchi method of L27 with two replications.
The use of L27 does give a better fuzzy model, so to achieve a better Taguchi model three
replications are an appropriate design. It is clear that Fuzzy model is better with L27 than L9.
However, the chance of achieving good results with L9 is the area where new researchers can
work.
Taguchi method used orthogonal arrays to solve the problems. This DOE has many qualities
such as reliability and small data size but its disadvantage is the assumption that all input
variables are independent in nature i.e. interaction effects of two or more than two variables
participates insignificantly in the output quality. M. Zaidi et al. [3, 4, 12, 24, 25] used L9
orthogonal array for Statistical and AI modelling purpose. M. Zaidi et al. [12] achieved better
modelling with L9 orthogonal array without the outer array. It is an achievement of M. Zaidi
26
et al. [12] to model with L9 array, having 12% to 22% missing values in the datasets, by
applying Semi-supervised algorithm.
3.3 RESPONSE SURFACE METHODOLOGY Response surface methodology (RSM) is an advanced mathematical and Statistical DOE
method which is useful for the modelling and optimization of manufacturing engineering
process or laser cutting process [1, 23, 57, 58]. RSM design of experiment optimizes the
output quality that is varying due to input parameters. In laser cutting process modelling the
data is collected based on selected DOE. It is used for modelling and optimizing of laser
cutting process and the size of data is two to three times that of the Taguchi method.
However, smaller than the standard Factorial design. [25]. The basic RSM used first order
linear model. However most of the laser cutting process is non-linear. Therefore, lack of fit of
first order models create the need of second order regression model [73]. Noor et al. [57]
employs the Box-Behnken design of RSM for the optimization of experimental data to
explain effect of selected significant input variables. This design of RSM is often selected
when experiment executes in non-sequential manner and the whole experiment is performed
in one session. First and second order coefficients produce better estimation. In these designs
lesser number of runs are required as compared to central composite design (CCD) with same
number of input factors. It does not contain axial points. Therefore, it is easy to set the
extreme ranges within operating range similar to factorial design as well as all factors are not
adjusted at maximum value simultaneously. Hence, the experiment is less expensive [57,
74-76].
Noor and Kadirgama [57] prepared a surface height and roughness model with input
variables, such as standoff distance, cutting speed and laser power. The work-piece is 3mm
acrylic sheet cut by pulse CO2 of 30 watts. The first order model shows that laser power is the
major factor. The analysis of variance shows P- value is not significant i.e. linear factor
model was not sufficient. In this work Surface height and roughness models show that laser
power and standoff distance are significant factors and both output qualities are directly
proportional.
It is observed that both Noor and Kadirgama et al. [57, 58] articles are similar with the
exception that the first article is linearly fit and second quadratic RSM model which checks
27
he interaction and quadratic effects. The interaction effects are insignificant but quadratic is
significant and also shows that second order is more close to experimental values.
Noor and Kadirgama eta al. [8] selected the same Box Behnken design of RSM used for
small experiment size, but changed modelling method from RSM to artificial intelligent (AI)
model. They made a combination of ANN and Particle swarm optimization (PSO). This
article shows that surface roughness is significantly affected by the priority of standoff
distance, laser power applied, speed of cut and thickness of work-piece.
Mathew et al, [14] used RSM central composite design of experiment for the modelling of
heat affected zone of Carbon Fiber reinforced plastic composites cutting by Nd:YAG
pulsed laser at the optimum process parameter ranges. The obtained results are in the
acceptable ranges.
Baş and Boyacı [77] studied the recent papers (since 2005) utilizing RSM and attempted to
find out some common mistakes and limitation of RSM. RSM model fits well but it is
possible that its prediction out of experimental data range is not predicting well. The size of
RSM observations is smaller than factorial design but it is capable to model the effects of
interactions which will produce better results than Taguchi method if interaction between the
variables is significant.
Balkin and Lin [56] studied the optimum level of the independent variables to understand the
quadratic banana function to predict Response surface methodology RSM that includes only
two variables X1 and X2. The results of the quadratic regression analysis and ANN were
discussed for the different levels of input parameters for 4, 5, 6, 7, 8 and 9 levels. The number
of observation for 4 levels will be 16 and similarly for others 25, 36, 49, 64 and 81
respectively. The results show that as the size of data increases ANN is incomparably better
than RSM. In addition to this ANN can model multiple output [32]. Dhupal, et al. [26] also
prepared ANN model to predict data on microgroove width and depth phenomena with RSM
design of experiment.
3.4 FACTORIAL DESIGN Factorial design gives the opportunity to study Treatment and interactions effects on response
variables because frequent interactions are significant factors in many industrial processes
28
[23]. It is one of the better designs of experiment. But large number of factors with higher
number of levels makes Factorial design useless.
Figure 3-2 shows that number of Runs increase as the number of factors increase when
number of levels are 2. Two levels are often used for the selection of factors which is called
screening, even though the increase in number of factors with two levels decreases the
usability of factorial design with 8 or more factors.
Figure 3-2: Increasing trend of number of Runs as factors with two levels
Figure 3-3: Increasing trend of number of Runs as factors with three level
4 8 1632
64
128
258
0
50
100
150
200
250
300
2 3 4 5 6 7 8
Num
ber o
f Run
s
Number of factors
9 27 81 243729
2187
6561
0
1000
2000
3000
4000
5000
6000
7000
2 3 4 5 6 7 8
Num
ber o
f Run
s
Number of factors
29
Figure 3-3 shows that the number of Runs also increases more as number of factors increases
when number of levels are 3. Three levels are often used for modelling and analysis. It shows
that increase in number of factors decreases the usability of factorial design with 5 or more
factors.
Figure 3-4: Difference of number of Runs in 2 and 3 levels
Figure 3-4 shows the difference between two and three levels of factorial design. It shows
that as the number of levels increase there is a remarkable increase in number of Runs which
shows that increase in levels increases number of Runs remarkably and decreases the
usability of Factorial design.
J. Ciurana, et al. [78] used factorial design of experiment for the preparation of many models
for laser milling process qualities dimensional, geometry and roughness properties. The
results of ANN with factorial designs are better than other researchers. They also compare the
results with multiple linear regressions and concluded that ANN is a better solution [4, 78].
The experiment is based on full factorial design which includes 27 observations. Two third
data is used for training and one third for testing of ANN model.
M. Zaidi et al. [4] used Full factorial design for the purpose of modelling and verification.
After getting data in the form of factorial design in the beginning, evaluate orthogonal array
which ignores the interaction effects. The results show that it is not at all as good as it sounds.
It has been proposed that the experimental design which is sequential in nature has greater
efficiency and accuracy but is avoided due high cost of experiment. The results were
4 8 16 32 64 1282589 27
81 243729
2187
6561
0
1000
2000
3000
4000
5000
6000
7000
2 3 4 5 6 7 8
Num
ber o
f Run
s
Number of factors
2 Levels
3 Levels
30
encouraging but only performed on Perspex of 3mm edge quality with missing value problem
in design of experiment. The problem with OA was that it is unable to solve the problem
within the desired percent errors in prediction, because it utilizes 60% data for training, 20%
for verification and 10% for final testing. The modeling in case of supervised learning with
missing values is possible.
Sivarao, et al. [9, 33, 36-38] has presented many papers in conferences and published many
journals to model the laser cutting process with ANN and fuzzy logic but all of them utilize
full factorial design without any missing values which is the area where improvements can
be made. The studies show factorial design is appropriate for modelling with ANN.
Sivarao et al. [43] has been working in the area of fuzzy logic modelling and has been able to
prepare the laser cut surface roughness model for Mn-Mo 5mm pressure vessel plate based on
two level factorial design consisting of 128 observations. About five observations of
experimental results deviate much more than desired value. This could be a modelling error
except in these observations the error is still greater than 10%. The author’s expectation was
0.5% which is much lower than the actual results.
M.-J. Tsai et al. [34] built both artificial and multi-regression model for QFN cutting of 6
qualities and 3 input parameters. The results of LM-back-propagation neural network were
better than multi-regression because of the selection of full factorial design. The results of
multi-regression are also acceptable because the size of data sets is sufficient and the model
prepared based with the effects of Treatments and interactions. However, sufficient data size
provides lead to ANN. Hence, if it is possible to select factorial design and it will not effect
on cost and time then factorial design is the better approach.
Yilbas and Rashid [41] also selected factorial design for cutting 800HT alloy with CO2 laser
to investigate the optimum parameters and used very simple crisp logic to assign the score to
flatness and waviness. The overall quality is the aggregation result of both parameters scores.
The striation frequency is measured from surface roughness and they built the relationship
with the dross ejected from the kerf. However, it is better to apply AI techniques to build
model and overall quality tools.
31
Yilbas [79] studied CO2 laser cutting of Stainless steel to investigate the input parameters,
effects of speed of cut, pressure of assist gas, thickness of work piece and laser pulsing
frequency in order to monitor the variations in Waviness, flatness and metallurgical changes
at the cut surface and overall cut quality utilizing factorial design of experiment. Neural
network has been utilized successfully for classification of waviness patterns. He concludes
all parameters are statistically significant on quality responses. He used scores assignment
technique for overall quality. Better results can be produced because, all the required data is
provided through factorial design but overall quality can be improved by fuzzy aggregation.
He also discussed Artificial Intelligence (AI) based techniques such as artificial neural
network (ANN) and fuzzy logic which are widely used AI techniques. In my point of view
the design of experiment used by AI techniques need some modification in design of
experiment. They need some data set out of designed matrix for checking of generalization of
model and some for testing of the model. This issue can be resolved without investing on the
experimentation, by taking the observation taken for range adjustment before the actual
experimentation.
3.5 SUMMARY It was observed during literature review of LCP for DOE point view that many researchers
used OFAT technique. However, large size experimentation runs and ignorance of combined
effects of different input variables can be overcome by using DOE. In LCP, Scientists
focused essentially on TM, FD and RSM. Taguchi’s orthogonal array DOE is a robust,
efficient method for designing high quality stable process at lesser expense for LCP due to
minimum noise. However, outer array are better than simple OA but results of AI (Genetic
algorithm) are even better than outer array.
Different sources and factors of uncertainty are mentioned but Statistical uncertainty due
incomplete data (missing value issue) is present in our problem. Many experimental studies
used orthogonal array (L9, L18 and L27). Hybrid combination of Taguchi DOE with other
methods produced better results in modelling and multi-quality optimization. However, most
of the researchers used Larger size OA than L9. If they used L9 they expand the size of
datasets by using RSM model and then used larger data sets for training by ANN. But,
accuracy depends on the selection of RSM fitting function which is difficult. The LCP is non-
linear, therefore, lack of fit of first order models creates the need of second order regression
32
model or sometimes both are unable to model. Box-Behnken and Central composite design
are similar in size but Box-Behnken range contains axial points. The axial point extreme
values of input settings are sometimes harmful for Machine. RSM model fits well but it is
possible that its prediction out of experimental data range is not predicting well.
Factorial design gives the opportunity to study Treatment and interactions effects on response
variables. It is one of the better designs of experiment. The number of Runs increase with the
increase in the number of factors or levels or both which makes FD impractical L(Runs)
(Levels) (Variables) .
Many researchers used FD for LCP for ANN modeling, fuzzy for mapping and overall
quality. FD with 2 levels is used for screening. The number of Runs with three levels and
four factors is reasonably small datasets.
Many papers with ANN and fuzzy logic utilize FD without any missing value which is the
area with margin of improvement. The articles solve overall quality with crisp logic and
Statistical methods also used FD which can be improved by fuzzy aggregation.
We selected Factorial Design and tried to model L9 orthogonal array for high cost materials.
Nine observations are used for modeling and remaining for detail verification. It has been
proposed that the experimental design which is sequential in nature has greater efficiency and
accuracy but is avoided due high cost of experiment. The problem with OA was that it is
unable to solve the problem within the desired percent errors in prediction with supervised
learning but later achieved better modelling with L9 orthogonal array, without the outer array
having 12% to 22% missing values in the datasets, by applying Semi-supervised algorithm.
The next Chapter 4 provides the history of ANN and literature review to provide guidance for
the modeling of the problem with small datasets. Chapter 5 will provide the details of overall
quality optimization using the selected DOE.
CHAPTER 4
NEURAL NETWORK MODELLING
33
4. NEURAL NETWORK MODELLING
4.1 ARTIFICIAL NEURAL NETWORK Connectionist architectures originate from Psychology, Physiology and Computer Science.
McCulloch and Pitts (1943) [80] demonstrated artificial neural network model for binary
problems. The model is binary transfer function (OR, NOT and AND) therefore, its output is
only 0 or 1. Its functionality can be represented mathematically
𝑌𝑌𝑏𝑏 = 𝑓𝑓 ��𝑊𝑊𝑖𝑖𝑏𝑏𝑋𝑋𝑖𝑖𝑏𝑏 + 𝐵𝐵𝑖𝑖𝑏𝑏𝑖𝑖
� (4.1)
Where Yb is transfer function binary output, f is function either 0 or 1 for neuron, Wib is
neighbors weighted values of input Xib and Bib is intercept or bias of straight line model.
However, no weight adjustment algorithm was presented. Animal learning theories motivated
Hebb in 1949 towards Neural Networks. The basics of Artificial Neural Network (ANN) is
artificial neuron, network topology encoding scheme and learning algorithm [81].
Frank Rosenblatt in 1958 [82] made neural network linearly separable by Perceptron. His
major contribution is to provide iterative algorithm for the weight adjustment which were
missing in McCulloch and Pitts [80] binary model.
But Minsky and Papert, 1969 [83] point out that Perceptron model is unable to classify
Exclusive OR logic gate (XOR) problem. Due to this ANN’s reputation reduced. Addition of
hidden layers with the algorithm resolved this issue and again attained attention in mid
1980’s. Connectionist architecture of artificial neural network models have been utilized in a
broad variety of problems by Rumelhart and McClelland in 1986 [84], Anderson and
McNeill in 1990 [85] and Churchland and Sejnowski in 1992 [86].
Multilayered networks are computationally complete i.e. equivalent to Turing machine [81].
ANN used common error correction (back propagation) algorithm for training [85], the
increase in number of neuron and connection improve the networks to model the problem.
The increase in neuron gives the benefit up to some extent and begins to decrease the ability
of classification upon further increase, so it needs to be adjusted to maximum accuracy by
adjusting optimum number of neurons.
34
Biological networks work in parallel and ANN follow Von Neumann architecture [85]. A
neuron response time for 1 GHz CPU is more than biological single neuron time of 1x10-3
seconds because CPU executes one step of execution in nanosecond (ns). Even then brain is
far better than a computer because of the ability of billions of neuron computational ability in
parallel [87] such as humans perform exponentially better than computers e.g. recognizing a
face from thousands of familiar faces instantly. Some salient features of Artificial Neural
Networks are non-linearity, input- output mapping, and ability to adapt to environment,
evidential response, and fault tolerance, analogous to neurobiological systems.
Imran Amin et al. [88] also mentioned inspiration of human brain in two ways, an ANN
attains knowledge through learning and ANN's knowledge is stored within synaptic weights.
ANN is based on different mathematical models that follow some properties of the biological
nervous systems and adapt the similarities of biological learning. ANN model is comprised of
a large number of interconnected processing elements that are similar to neurons and are
connected together with weighted connections that are analogous to synapses. Similar to
nervous system neural network are processing in parallel.
The ANN function is built by the connections among the elements. The ANNs can be
differentiated by feed-forward and recurrent algorithms. The selection of network is based on
the data processing technique. They can also be distinguished by the method of training used,
some ANNs use supervised and others are unsupervised training. The decision of using
supervised or unsupervised network for its training depends on data to be processed. In the
process of supervised learning of ANNs, inputs are applied to obtain output response. These
outputs are compared with the target output i.e. the target response. If the output response
differs from the desired response, the error ‘e’ in equation (4.2) for a single training pattern is
the sum of square difference i.e.
2)(21
dd yte −= ∑
(4.2)
where,
td represents Target response for dth unit,
yd represents actual output produced response for dth unit.
35
The error “e” calculates the deviation from the desired training response. To minimize the
error, optimization should be made to the network’s synaptic weights. Contrary to supervised
learning, unsupervised learning does not require a teacher; i.e. no desired output is required.
It is typically used in the background of recurrent and competitive nets. In unsupervised
learning the training set is not separated into input and output pairs during training, the neural
net receives many different excitations, or input patterns, and it randomly classifies the
patterns into categories. Although unsupervised learning does not require a teacher, it needs
guidelines to form groups. Grouping may be based on the properties of data. The basic
accepted categories are supervised and unsupervised learning but Semi-supervised learning
algorithm is prepared based on the supervised learning experiences [4]. The results of this
algorithms are more encouraging than the typical supervised learning [12].
Chavez [87] explained the Blue Brain project and future direction in ANN. He also defines
that two major problems of ANN are teaching and designing networks [87] and better
teaching (training) is the major concern in modelling module while doing training, testing and
simulation process of ANN modelling.
Dhupal, et al. [26] utilizes RSM and ANN to build model for the prediction of microgroove
width and depth phenomenon. Author built RSM model for the optimization of the response
variables and verified the optimal results with ANN predictions and experimental data. The
results were satisfactory and also proved the ability of ANN to use this tool in other laser
cutting applications.
In the extrusion process of food material problem ANN modelling executed better than
Response surface methodology (RSM) [32]. In fact RSM uses predefined linear or quadratic
regression modelling to model the surface. M. J. Tsai, et al. [34] prepared multi-regression
analysis and artificial neural network model for Quad Flat No-Lead (QFN) cutting of 3 input
and 6 qualities.
The results of LM-back-propagation neural network were better than regression model. ANN
model was trained with full factorial design. Complex and non-linear problems model by
ANN definitely gives better approximation than regression. Hence implies that it is better
than RSM. But there is a need to verify that ANN, whether works better even with the
missing values or not.
36
M. Zaidi et al. applied that multi-regression, ANOVA models which are not sufficient for
solving this problem, as a matter of fact it was shown that average error reaches 50% which is
a huge error due to small dataset of OA and missing values in observation table. This is the
reason why neural network approach was adopted in [2, 4].
Sivarao, et al. [33] used ANN in the area of laser cutting without missing values. The process
design requires subject expert participation throughout the design of experiment. But
application of genetic algorithm reduces the dependency of subject expert. They used full
factorial design L128 which is a large data size compared to M. Zaidi et al. [3, 4, 12, 25]
works using quick back-propagation training algorithm. ANN modelling is often suitable for
large number of observations and non-linear data. ANN learns from the available experience
(input-output datasets) and captures the functional relationship between the input and output
parameters. It is quoted in different papers that any continuous function can approximate to
any required accuracy by ANN network [4, 12, 29]. Sivarao, et al. [33] claimed that the
ability of ANN prediction of surface quality in laser cutting process minimizes the cutting
cost by as much as 70% of the overall manufacturing because of rework.
C.B. Yang, et al. [35] also prepared a model of a CO2 laser cutting process of polymethyl
methacrylate PMMA using orthogonal array of L9 for four factors with three levels but
triplicated the data size by statistical modelling prediction and then prepared ANN model at
the second stage with large datasets which shows OA DOE is not sufficient for ANN
modelling and also in case of supervised learning by Zaidi and Amin et al. [4, 35].
The combined progressive model Taguchi and neural network prediction improved with
small datasets. Taguchi modelling is difficult with unbalanced matrix means with missing
values and need the experimentation in the supervision of domain expert. In this paper the
deficiency of ANN on small datasets such as L9 is overcome by increasing data by Taguchi
modelling. But in our study small datasets were modeled and moved a step ahead by handling
missing values. The author admits that the Taguchi modelling is inferior so increased data
sets are also not accurate and in case of interaction this method is not useful.
Zaidi and Amin et al. [3] developed a model of laser cutting process with missing values by
LM-feed forward neural network. The results were verified with Taguchi method results. But
37
Taguchi model itself not verified by experimental data. Therefore, more study is required to
verify the modelling technique. In supervised learning the quality of model vary with the
ability of modelling of supervisor. This can be further standardized by semi-supervised
learning.
This ANN approach is sometimes not the best however, it enables to produce a less precise
model in difficult scenarios such as non-linear problems and in case of missing values [4] to
give a rough estimate. The need of a program to replace supervised learning by semi-
supervised approach create networks, vary their values and initialize synaptic weights many
times automatically to find the global minima on the error surface. The Semi-supervised
algorithm search out the better network architecture, training algorithm and most probably
global minima based model. It is learned from neural networks that it is possible to train large
systems with many inputs on the basis of relatively small data sets [3]. The resulting systems
usually have a moderately nonlinear structure.
M. Zaidi and I. Amin et al. [4] utilized supervised learning and then established semi-
supervised technique for normal and missing values data for training and simulation [12]. In
this modelling, the benefit of ANN is in learning of nonlinear, multivariable datasets of
different DOE. The model is able to predict desired input values for the optimization. By
using the trained network the machine setting time and human error on the new work-piece
can be reduced [2].
4.2 FEED-FORWARD BACK-PROPAGATION NEURAL NETWORK The simple and effective nature of feed-forward back-propagation is most widely used in
artificial neural network structures, having been used effectively in many research works
such as [2-4, 10, 24, 29-31, 35, 89]. Figure 4.1 is a supervised training neural network,
containing “n” numbers of neurons interconnected to form an input layer, with hidden layers
and then to output layer. The input and output layer nodes work as a buffer for input and
output values of the predictive model and hidden layer forwarded input relations to be
represented in the output layer.
In the beginning of training session, inputs and target values are run through the network
with randomly chosen weights for the nodes, which makes the network analogous to
newborn's brain but without knowledge [88]. With reference to paper [27] he explains the
38
concept of back-propagation mathematically. Each hidden and output neuron processes its
inputs by multiplying each input by its weight, summing the product and then passing the
sum through a non-linear transfer function to produce a result. The S-shaped sigmoid curve is
commonly used as the transfer function. The neural network learns by modifying the weights
of the neurons in response to the errors between the actual output values and the target output
values. This is carried out through the gradient descent on the sum of squares of the errors for
all the training datasets. The changes in weights are in proportion to the negative of the
derivative in the error term. One pass through the set of training patterns along with the
updating of the weights is called a cycle or epoch. Training is carried out by repeatedly
presenting the entire set of training patterns (with the weights updated at the end of each
cycle) until the average sum squared error over all the training patterns are minimized and
within the tolerance specified for the problem or reach to the limit of number of epoch.
Back
Propagation
Feed Forward
A1 A2 A3
y1 y2 y3
………………………………………………….
………………………………………………….
………………………………………………….
Output layer
Hidden layer
Input layer
Input values
t1 t2 t3 Target values
Figure 4-1: Feed-forward back-propagation model
39
At the end of the training phase, the neural network should correctly reproduce the target
output values for the training data provided the errors are minimal, i.e. convergence occurs.
The associated trained weights of the neurons are then stored in the neural network memory.
In the next phase, the trained neural network is fed a separate set of data. In this testing phase,
the neural network predictions are compared with the target output values and evaluate the
dependability of the neural network to generalize correct responses for the testing datasets
that only broadly resemble the data in the training set. No additional learning or weight
adjustments occur during the testing phase. Once the training and testing phases are found to
be successful, the neural network is ready to be used in practical applications. The neural
network will produce almost instantaneous results of the output for the practical inputs
provided. The predictions should be reliable.
After experimental observations and calculations the edge quality, kerf width, data tables are
prepared based on the requirement of DOE. The mean and S/N values are calculated and final
observation tables prepared. The training data and this trained network is then able to predict
other unknown inputs. As a result, simulated data results are nearer to the desired results and
the tolerance range gives the inputs parameter values the range applied while training [27] .
In order to obtain global minima rather than local the supervised network teacher changes the
initial network weights several times on different architectures and training algorithms. At
times modifications are made in network settings such as changes in the quantity of hidden
layers, neurons and training algorithms to obtain minimum error from the desired output with
tolerance near 10% in order to have better supervised training. This process of modifications
is called "training". A sigmoid function in the hidden layer and a linear function in the output
layer are being used by the back propagation neural network, which is applied in this study.
4.3 COMPARISON OF TRAINING ALGORITHMS In ANN modelling and simulation process first create a network consisting of input, output
with selection of training algorithm, number of hidden layers, number of neurons and
selection of transfer function with the selection of feed forward back propagation and a
training algorithm. Create a network and export to workspace. Programming is carried out in
MATLAB (The Language of Technical Computing).
In ANN wide range of learning algorithms exist, the greater part of them including the
famous back-propagation learning algorithm, are of the gradient descent type [24, 90]. The
40
network is prepared for training after initialization of network bias and weights. The network
mapped the input and output examples to get the approximate transfer function on the basis
of given datasets. The inputs and target output are iteratively minimizing the error. Some of
the training algorithms for feed forward back propagation are explained in the following
section. All the following training algorithms utilize the negative gradient of performance
function to set the weights to minimize error in desired and actual output. The negative
gradient is calculated by back propagation method.
4.4 BACK PROPAGATION ALGORITHM Even single hidden layer can be employed to approximate any nonlinear function [91] but
the selection of training algorithm is of vital importance. Some of the training algorithms for
feed forward back propagation are explained for this purpose. All these algorithms use
negative gradient and it is calculated by back propagation method. Back propagation
algorithm is simply an algorithm which changes weights and bias on the basis of previously
calculated mean squared error to reduce it. The algorithm is iterative in nature and can be
expressed in the form of a pseudo code as follows:
1. First randomly or by some systematic means initialize the weights
2. Get the output from the network with these weights
3. Calculate error by subtracting output from known input output mapping
4. Compute new weights or delta from output layer to hidden layers
5. Compute for all successive layers to input layer
6. Change the old weights with the new weights
7. Repeat from step 2 until network converges to desired error
Goh [27] model the two engineering problems to show the back-propagation neural networks
usability. These two problems are noisy data and significant participation of interaction
effects that complicates engineering problem. Both complex modelling method along with
supervisor intuition and experience are considered necessary for the forming of these
practical relationships. Fame of ANN has increased as compared to statistical methods for
nonlinear problems. The statistical methods are needed to identify significant factors first,
ANN gives low weightage to less significant variables without the interference of the
researcher. Goh [27] also mentioned that the inherent core property of ANN is that with
incomplete input information it can match correct output in pattern recognition. Hence, it can
41
work with missing values in experimental data models. This advantage increases the
utilization of ANN in both quantitative and qualitative problems.
4.4.1 GRADIENT DESCENT ALGORITHM Initialize the network with required number of layers with desired neurons and randomize the
weights. Apply input output for mapping on neural network. Now do the forward
computation of the network from book [92]. First calculate the induced level fields by the
following formula
VjL(m) = �Wji
L(m)Yi(L−1)
i
(m) (4.3)
Where
Yi(L−1)(m) is the ith neuron output signal function of preceding layer L-1at mth iteration
Wji(L)(m) is jth neuron synaptic weight in hidden layer L from ith neuron L-1(previous)
In case i=0
Yj0(L)(m) = Bj
(L)(m) is the bias applied to neuron j in layer L (hidden) and Y0(L−1)(m) =
+1 . Now apply the desired transfer function. The output signal of neuron j layer L is
Yj(L) = φj(Vj(m)).
If L=1 means neuron j is in the first hidden layer, set
yj(o)(m) = Xj(m) (4.4)
where
Xj(m) is the jth component of the input vector x(n). If jth neuron is in the output layer. i.e. L
= OL, apply
YjOL = Oj(m) (4.5)
The error signal is
Ej(m) = Dj(m) − Oj(m) (4.6)
Further do forward computation from book [92]. Calculate the local gradient (δs ) of the
network
Dj(m) is the jth element of the desired response vector D(m).
δjL(m) = �
EjOLφj
′ �VjOL (m)�
φj′(Vj
L(m))�δkL+1(m)Wkj
L+1(m)k
�
For jth neuron in output layer OL
For jth neuron in hidden layer L
(4.7)
Where
42
φj′(∗) is shows differentiation w.r.t the expression within brackets.
Fine-tune the synaptic weights of the neural network in layer L as per “generalized delta rule”.
The individual weight can be calculated by delta rule
WjiL(m + 1) = Wji
L(m) + [WjiL(m− 1)] + η δj
L(m)YiL−1(m) (4.8)
Where
η is the learning rate parameter.
Iteration keeps running until its stopping condition met. Iteration computation runs forward
and backward as shown in above derivation [92]. Batch gradient descent algorithm is not
very effective because learning rate is adjusted manually. If it is very small, then there is high
probability of being erroneous and getting trapped in a local minima, while a large learning
rate may cause training to become unstable due to overshooting. It is also called steepest
descent. The gradient descent algorithm often gets stuck in local minima in error surface as
shown in Figure 4.2.
Error Surface
Local Minima
Global Minima
Learning rate W1 old
W1 new
W1
Error
Figure 4-2: Searching of global minima in error space
4.4.2 GRADIENT DESCENT WITH MOMENTUM It is easy to understand that gradient decent algorithm learning can be time consuming
including the issue of local minima. The idea of momentum improves the local minima
43
concern and increases the rate of convergence considerably fast. This method resolves the
matter of batch gradient descent in case of small learning rate. The convergence speed
reduction and convergence trapped in the local minima is resolved by momentum. The
gradient descent with momentum works as a low pass filter. The momentum characteristic
ranges from 0 to 1. Zero means no momentum and 1 means high momentum. Gradient
descent with momentum (GDM) is a good algorithm as proposed in [93]; still it depends on
the user to tell it what the momentum term should be, none the less it has a probability of
finding the global minima. A mathematical representation would be:
WjiL(m + 1) = Wji
L(m) + α[WjiL(m − 1)] + η δj
L(m)YiL−1(m) (4.9)
Where α represents momentum parameter. However, probability to reach global minima
seems lesser than Quasi Newton and Levenberg Marquardt.
Ranaganth and Viswanath [31] prepared laser cutting process (2KW CO2) model of surface
roughness by Back propagation using gradient decent with momentum for mass production
and following optimization of cutting process. No specific design of experiment was used but
15 run were used for training and 5 for validation. The best selected learning rate 0.6 and
momentum 0.8 are required to achieve the desired error tolerance. The desired error goal was
met after forty one times initialize weights and 29653 iterations. They also used the technique
of weights randomization up to 41 times which shows that this is a difficult method to train
the model in supervised learning. In my point of view semi-supervised algorithm is a useful
solution for the more generalized way of modelling [12].
4.5 FASTER TRAINING ALGORITHM Gradient descent method (GD) is slower than faster training algorithms like variable learning
rate and resilient back propagation. Both of the faster algorithms are also heuristic. It is
difficult to adjust learning rate before the training. Advantage in these algorithms is that the
optimal training rate is changing during the training process as the algorithm moves in the
error space. Therefore, learning rate change during the training improves with the training
process. The dynamic and adaptive change in learning rate increases the size of learning rate
until the training process is unstable. The algorithm of variable learning with the feature of
momentum is a better choice.
44
4.5.2 Quasi Newton Algorithm The point at which the derivative of a function is zero is called the stationary point of that
particular function. Newton proposed a mathematical method to locate the stationary point on
a function, in our case error surface. The Quasi Newton Algorithm has its origins in that
function. Newton made an assumption that the area around the optimum can be approximated
to be quadratic in nature. The advantage of the Quasi Newton method is that it does not
require the computation of second derivate Hessian Matrix instead; one after another gradient
vectors are analyzed and used to update the Hessian Matrix. Quasi Newton is used as an
alternate to conjugate gradient methods and is mostly faster because there is no need to
calculate second derivatives. The modified Hessian matrix, at every iteration, is a function of
gradient. It can be explained mathematically as follows:
iiii glww 1
1−
+ −= (4.10)
Where li-1 is a Second derivative Hessian matrix of MSE index at present values of bias and
weights in equation (4.10).
But usually in the domain of neural networks this algorithm is used when training is of small
networks [93] and conjugate gradient algorithms are more suitable for large networks.
Therefore, Quasi Newton’s algorithm is better for the given case.
Secondly it must be noted that this algorithm requires a line which it uses to find the direction
in which it starts the descent, once that is set it goes into the depth of that line. The default
line function for Quasi Newton is “Charalambous’ method for 1-D minimization”.
4.6 LEVENBERG-MARQUARDT ALGORITHM The Levenberg-Marquardt training algorithm is repeatedly used as optimization algorithm. It
outperforms simple gradient descent and other conjugate gradient methods in a wide variety
of problems [94]. Utilization of Levenberg-Marquardt Algorithm (LMA) eliminates the
importance of Quasi-Newton, because it reaches to second order training pace without
calculating Hessian matrix. Therefore mean square error (MSE) is used as performance
criteria and Hessian matrix becomes the following equation; JJH T .= (4.11)
Where J is a Jacobian matrix which consists of all first order partial derivatives of network
∂ (error)/ ∂ (biases and weights) in equation (4.11) and the reason because of which partial
derivative is needed is that the error is partially dependent on many variables. As shown in
45
the equation we are taking partial change in error with respect to weights and biases. Whereas
gradient g is computed as
eJg T .= (4.12)
Where e is the error vector of the network in (4.12). The LMA can be expressed as
[ ] eJIJJww TTii
11
−
+ +−= µ
(4.13)
The intelligence lies in the µ symbol in equation (4.13), if µ is zero then it becomes equation
(4.10) i.e. Quasi Newton’s or if µ is very large then it becomes equation (4.9) i.e. Gradient
descent method with a smaller step size. As Newton’s method performs faster convergence
near error with more accuracy, therefore in LMA µ value reduces after each better iteration,
or increases in case of inferior iteration results on the basis of MSE. So that it does not get
stuck in a “local minima” and when it is away from minima it acts like a normal gradient
descent with momentum.
According to [93] the LMA is suitable for medium size datasets and fastest compared to other
training algorithms. The only drawback lies in the large memory it requires to handle the
matrices, but since our requirement is small size datasets and analysis is offline in laser
cutting process modelling. Therefore, LMA is the most suitable theoretically and the results
favor it from our modelling and simulations. Kuo et al. [28] model the Light guide plate
(LGP) manufactured by CO2 laser cutting using L18 Taguchi orthogonal array as DOE. Back
propagation neural network with Levenberg-Marquard training algorithm is used to model
the problem. They used 13 datasets for training and 5 for Test. The simulation error was
within 5%. The model is without missing values and size is double the size of L9.
J. Ciurana, et al. [78] also built up many models for laser milling process by evaluating the
roughness, geometric and dimensional characteristics. Results with ANN are better than
multiple linear regressions and it should be noted that one training algorithm was not applied
to all models. The results of resilient back-propagation algorithm predicting are optimum for
surface roughness and LM algorithm has performed better for predicting geometrical and
dimensional characteristics. The LM algorithm approach reduced the size of neural network
and produced better prediction fast conversion. The size is also three times the L9 orthogonal
array. The training is utilizing supervised learning in both the above studies. However, semi-
46
supervised learning does not need supervisor’s expertise and reduces the time consumption of
the researcher to select the best model.
4.7 OVERTRAINING OF DATA MODEL Robert P.W. Duin [95] raised the important issue that well trained networks generalize in the
sense that their outputs are (close to) correct for objects not available in the training set. If the
training set is noisy or contains errors (like overlapping classes in case of a classification
problem), a too heavily trained network may be adapted to the noise which does not
generalize. It is over-trained and produces random outputs for unseen input objects. Sivarao,
et al. [23] also mentioned that some datasets are required out of designed matrix for checking
of generalization of model and testing of the model. This disadvantage can be resolved by
modifying design of experiment or without investing on the experimentation by using the
observation obtained for range adjustment before the DOE based observation tables recording
or to rotate the some parts of the training dataset and the test dataset. Overtraining is
sometimes related to strong nonlinearity. This can be detected by the size of the weights.
Only for large weights the network is able to have a strong nonlinear functionality.
Overtraining [95] may be avoided by:
• Early stopping: Stopping the training in time, before an adaptation to the noise starts.
It is difficult to detect an optimal stopping moment without the use of a test set.
• Small networks: If the network has much less parameters than objects in the training
set, it cannot be over-trained.
• Pruning: reducing the size of a network after it is trained. In this way unimportant
neurons, causing noise, are deleted. It is necessary to retrain the network after
pruning.
• Large training sets: Networks of a fixed size cannot be over-trained for large
training sets.
• Data enrichment: artificially enlarging the training set by objects that ‘smooth out’
the noise.
• Regularization: One way to do this is to add a penalty term to the optimization
criterion for networks with large weights, as these are related to strong nonlinearity.
Over training can be avoided with other methods such as Z. Xu and Z. Mao [96] presented
support vector machine (SVM). SVM is also able to handle overtraining problem with small
47
datasets. Therefore, it can be used in laser cutting process modelling due to its ability to
model small datasets.
4.8 SUMMARY Biological networks work in parallel and ANN follow Von Neumann architecture. ANN is
capable to model non-linear problem and is analogous to neurobiological systems. ANN
attains knowledge through learning and knowledge is stored within synaptic weights. Often
ANN is distinguished into supervised and unsupervised training. The two major problems are
teaching and designing networks where better training is the major concern in modelling.
Sometimes RSM and ANN model have better results. The usage of test data for
generalization made ANN superior than RSM, regression and multi-regression,non-linear
regression and ANOVA models are not sufficient for solving this problem. However,
generally ANN is trained on FD for non-linear problems with different training algorithms
such as Quick back-propagation training algorithm was used for LCP with FD (L128) which
is large data size compared to OA. But it was claimed that ANN prediction reduces the need
of rework. With FD many researchers succeeded in modelling different materials in LCP.
However, many used Fractional factorial and L18 larger datasets without missing values.
They also used L9 to prepare a model of a CO2 LCP of PMMA sheets with L9 runs for four
factors with three levels similar to our data. But triplicate the data size by Statistical
modelling prediction and then prepared ANN model at the second stage with large datasets
which shows that they are able to model with L9 OA with ANN. Similarly, in combined
progressive model L9 was expanded with TM. Although it is difficult to model unbalanced
matrix and unable to handle interaction effects with TM.
Many researchers quoted that any continuous function can approximate to any required
accuracy by ANN network and even single hidden layer can be employed to approximate any
nonlinear function, but the selection of training algorithm is of vital importance. Regression
models are overstrained but ANN generalizes due to testing of model on unseen datasets.
Overtraining can be avoided by early stopping of training, network Pruning, reducing number
input parameters, large training data, data enrichment by duplicating datasets or adding
penalty terms. Feed-forward back-propagation ANN explained and compared the important
training algorithms and justifies its utility. Batch GD algorithm is not very effective as its
learning rate is adjusted manually. If its value is small then there is high probability of getting
48
trapped in local minima, while a large learning rate may cause training to become unstable.
But, the idea of momentum improves the local minima concern and works as a low pass filter,
it also increases the rate of convergence considerably swiftly. The LCP (CO2) model the
surface roughness by back-propagation using GDM training algorithm without using any
specific DOE using 18 runs. But, it required more time than LMA. The technique of weights
randomization used up to 41 times with GDM which indicates the importance of this
parameter. In Quasi Newton’s training algorithm, the point at which the derivative of a
function is zero is called the stationary point of that particular function. It assumed that the
area around the optimum can be approximated as quadratic in nature. Therefore, no need of
computation of second derivate (Hessian Matrix), instead one after another gradient vectors
are analyzed and used to update the Hessian Matrix. Quasi Newton algorithm is better to use
for small networks and conjugate gradient algorithms are more suitable for large networks.
Utilization of LMA eliminates the importance of Quasi-Newton, because it reaches to second
order training pace without calculating Hessian matrix. The intelligence lies in the µ symbol
in equation (4.13), if µ is zero then it becomes Quasi Newton’s and if µ is very large then it
becomes Gradient descent method with a smaller step size. As Newton’s method performs
faster convergence near error with more accuracy, therefore in LMA µ value reduces after
each better iteration or increases in case of inferior iteration results on the basis of MSE. So
that it does not get stuck in a local minima and when it is away from a minima it acts like a
normal GDM. LMA is suitable for medium size datasets and is fastest compared to others and
GDM training algorithms. It needs large memory but since our requirement is small size
datasets. GDM and LMA are more suitable theoretically and the results from our modelling
favor them, LMA is time efficient. In order to obtain global minima rather than local, the
supervised network teacher changes the initial network weights many times on different
architectures and training algorithms. At times modifications are made in network settings
such as changes in the quantity of hidden layers, neurons and training algorithms to obtain
minimum error within 10% in order to have better supervised training. Hence GDM and
LMA are the best training algorithm for the given problem.
CHAPTER 5
OVERALL QUALITY OPTIMIZATION
49
5. OVERALL QUALITY OPTIMIZATION Bahar and Golnabi [40] state that researchers usually optimize laser cutting process for single
quality and also use OFAT technique. In this dissertation Overall quality means considering
more than one quality for optimization. The solution is able to optimize two qualities at a
time. However, aggregation is better for more than two qualities such as three or four or more.
Some other significant multi-objective research has been mentioned earlier in introduction
and laser cutting literature review but some specialize advance Statistical methods are
discussed under the Multi-objective Statistical Methods heading. For more than two variables
optimization suitable data mining techniques are simple aggregation, normalized
aggregation, , Statistical methods, Genetic Algorithm (GA) or fuzzy logic (FL). The objective
is to provide a product or work-piece with better overall quality. The aim of this research is to
find out the overall quality (edge quality, kerf width, overcut and material removal rate) with
Electricity-efficient to predict optimize input datasets. The relationship between parameters
can be found by the modelling of the system by some Mathematical or Statistical methods or
Soft-computing techniques [2-4, 12, 25] . Overall quality utilizes large experimental data.
With reference to [3] the overall quality is measured on the basis of very small data sets of
OA based experimental observation. The artificial neural network (ANN) training was
performed on L9 orthogonal array aggregated data and overall quality was predicted on the
basis of simulated data. In both papers [2, 3] orthogonal array design was used to obtain
experimental data. But in [2] normalized aggregation was used instead of simple aggregation
on simulated data. The Aggregation function was further checked by customer quality
function [2]. Yilbas and Rashid [41] cuts 800HT alloy by CO2 laser and performed
experiment using Factorial design (FD) for overall quality of flatness (F) and waviness (D).
However they used simple aggregation by converting the values in score i.e. 1, 2 and 3. The
calculation is simple and crisp in nature. It was discussed in literature survey of [70] that GA
algorithm is also suitable for single response and multi-response problems. It was concluded
by Chatsirirungruang [70] that the application of GA and Taguchi method reduces average
loss and is less responding to noise i.e. improved robust system. Noor et al. [8] mentioned the
fact that the utilization of Particle swarm optimization (PSO) also has many benefits
compared to other evolutionary heuristic algorithms like GA such as faster convergence, less
number of parameters to adjust and easy application. Therefore, it is better to use partial
50
swarm or genetic algorithm in robust optimization mechanism. A combination of novel
Fuzzy aggregation is far better option for overall quality prediction.
5.1 GENETIC ALGORITHM The aim of robust process design is to build a stable process that produces minimum
variation due to uncontrollable parameters [60]. In recent times, Taguchi methodology has
become a practical technique for the improvement of productivity at low cost. Researchers
utilizing Taguchi methodology have experimentally studied the effect of input parameters on
kerf taper, kerf deviation, kerf width, heat affected zone and surface roughness in the process
of laser cutting. Taguchi methodology has been used for multi-objective optimization.
However, weight assignment is major task of multi-objective optimization. Therefore, many
researchers used hybrid techniques with Grey relational analysis [36, 61], fuzzy [39] etc. The
application of GA is also better technique and it provides better techniques of simultaneous
optimization of multiple dependent parameters [68-70] . The ability of genetic algorithm to
generate huge amount of possible population and selection on cost function basis made it an
important off line quality tool [60]. It is improbable to trap in local minima due to the
assumption of GA and does not depend on continuity, uni-modality or convexity of functions
[97]. GA using mutation, crossover and reproduction genetic operations then grow successive
generations and apply values to transfer function to predict the output quality for optimized
value [60].
Multi-objective problem can be solved by common genetic algorithm (CGA) by combining
both the quality parameters with or without weighting factor using averaging or aggregation
etc. It can be solved by multi-objective genetic algorithm (MOGA) [98]. It was concluded by
Fernandes and Vicente et. al. [98] that MOGA is probably faster and efficient. Orthogonal
array is used in this research but the selection of ordinal levels can be optimized using genetic
algorithm [99]. The study concludes that genetic algorithm can be applied at DOE level.
Chatsirirungruang and Masami Miyakawa [68] study the usefulness of GA as compared to
OA method. They used signal to noise ratio as a deciding factor as routine practice of
Statisticians but the criteria is not always true. Therefore, noise factors (outer array) are used
to solve the problem more efficiently in his all papers [68-70]. They apply GA to find
optimum controllable parameters with the support of outer array [68]. They also performed
more research by the same method including the handling of dynamic system [69].
51
The research shows that GA is useful for these problems to find optimum signal to noise ratio
and target value. The reason was consideration of interaction between the variables and more
number of levels used for continuous controllable parameters than OA to find better results
[69]. They also used the benefits of Taguchi and GA for unknown problem using the
simulations by GA [70]. Hence Genetic algorithm can be utilized for overall quality based on
the predefined cost function. But, it is preferable to use outer array at the time of DOE for
better results.
Pawar, et al. [62] applied Particle swarm optimization (PSO) to solve multi-objective
optimization problem of rough and finished grinding process parameters of three responses.
The method of constraint is handled by penalty methods as others. The PSO results are better
than GA, differential evaluation (DE) algorithm and Quadratic programming (QP). They also
studied the convergence rate, performance and accuracy of PSO algorithm. Thirty to forty
iterations are often required by PSO, which are comparatively quite lower than other
evolutionary computational methods. However, the GA has its own limitations such as risk of
replacement of a good parent string with the deteriorated child, less convergence speed, and
difficulty in selecting the controlling parameters such as population size, crossover rate, and
mutation rate. Also the results of GA presented by the authors are erroneous.
5.2 MULTI-OBJECTIVE STATISTICAL METHODS Some other significant multi-objective research has been mentioned earlier in introduction
and laser cutting literature review. But there is a need to further discuss some research. B. S.
Yilbas [79] investigated overall quality at the cut surface due to input attributes variation in
the process of CO2 laser cutting of stainless steel. The overall quality was calculated based on
cut surface flatness, material changes and Waviness. He analyzes the significance of input
attributes on mentioned qualities and found them to be significant. He utilizes scores as in
desirability method instead of weightage and sums the scores in crisp logic to calculate
overall quality. The scores are assigned in flexible range therefore crisp does not reflect the
true participation of qualities in the overall value. Fuzzy aggregation definitely improves the
results. As standard membership functions generate standard aggregation results with the
quality of applied rules and fuzzy aggregation.
52
Dubey and Vinod Yadava [59] discussed hybrid TM and RSM for the multi-response
optimization of laser cutting process. The weighted impacts of quality characteristics and cost
components are incorporated in this quality function.
Dubey [59] concluded on the basis of modelling and optimization results show significant
enhancement in kerf width and MRR qualities with hybrid TM and RSM technique. He found
that linear input parameters are significantly participated in both the output qualities. Cutting
speed and pulse frequency interaction effects are significant. Therefore, TM is not sufficient
to solve the issue. They were able to solve the problem because they used RSM with TM.
The ability to model interaction effect in RSM produces enhanced results. But, for analysis,
modelling and optimization it is a difficult method to utilize two different Statistical methods
and two different DOE. Dubey [59] shows that interaction is significant therefore it is better
to use straight RSM DOE and utilize for modelling and optimization.
Gani and I. A. Choudhury et al. [100] used TM optimization in case of end milling process
with P10 TiN coated carbide on AISI H13 Steel tool. The high speed cutting was done with
finishing and Semi-finishing conditions. They investigated interaction effect and found TM
sufficient for the problem. The main target is to find the condition to achieve good surface
roughness and with low cutting force . The optimization was achieved by using signal to
noise ratio and Pareto ANOVA methods for detailed analysis. They concluded that often high
cutting speed, low feed rate and low depth of cut produces good Surface roughness and
cutting force is kept low.
L. K. Pan et al. [30] provides optimizations of multiple qualities of laser welding of
Titanium with Nd:YAG laser. The data analysis techniques used for integrated performance
of overall quality is TM-based Grey analysis. They mentioned the presence of interactions
among pulse shape, shielding gas, welding speed, pulse frequency, laser focus and laser
energy in laser welding. The overall optimization was calculated by Grey relational grade but
the TM limitations reduce the best results. Since very complex interaction is present in this
process and modelling was done by TM. Hence the results are definitely not best. Some other
technique at the Modelling level can be used and need to change DOE technique as per
modelling and interaction requirements such as RSM.
53
Excellent work has been achieved by Phillips and Kim [63], they developed a statistical
method, the DMT method, to evaluate and optimize multiple quality characteristic problems.
The method incorporates desirability functions, a performance statistic based on the mean
squared error, and data-driven transformations to provide a systematic approach that is
adjustable to a variety of situations and easy for non experts to apply. They take weighted
geometric mean and MSE (mean square error), the DMT method allows the decision-maker
to alter the relative importance of the quality characteristics, mean, and variance to suit the
particular situation. Taguchi parameter design offers no similar flexibility. Once the data are
converted into individual desires, either Harrington’s (1965) simple geometric mean or
Derringer’s (1994) weighted geometric mean aggregation methods may be used to combine
the individual desires.
In this method if one of the variables is undesirably high and other is very low then the
aggregated suggest wrong prediction for overall quality. Table 5-1 summarizes the difference
in Taguchi and DMT method.
H. A. Eltawahni, et al. [6, 101] developed RSM (Box-Behnken design) based model of laser
cutting of MDF and AISI316L and solved the issue of overall quality in two steps by using
the desirability approach. RSM model established all factors effect on output depending on
priority and constrained based on cost and quality. Eltawahni, et al. [6, 101] is unable to
provide single optimize overall quality. The decision is based on operator to decide whether
quality is more important or cost. It is better to use fuzzy aggregation, as fuzzy can solve the
issue of quality and cost in a simpler manner than the complicated approach which the
operator used to investigate for optimization. The verification error was 17.4 to 10.04%.
However, in his second paper he was able to minimize the error below 10%. It is required
that they consider cost-effective solution while not compromising on quality.
Ming-Fei et al. [61] achieved multi-objective optimization for two qualities (optical
transmittance ratio and roughness) of CO2 laser cutting by applying grey relational analysis.
The method of grey relational analysis can be improved by considering interaction effect by
applying better techniques in the building block of measuring quality data.
Sharma and Yadava [36] developed second order RSM and Taguchi model for Aluminium
alloy sheet by ND:Yag pulsed laser . Weights of both qualities were measured by Taguchi-
54
based GRA with entropy measurement methodology. Taguchi-based GRA is an effective
approach to attain optimum values of both the responses only by orthogonal array and needs
Statistical expert. However, is unable to model with missing values because unbalanced
matrix calculation is a difficult task. M. Zaidi et al. [12] modelled L9 orthogonal array and
FD model with missing values by ANN.
Table 5-1: Difference between DMT and Taguchi methods
Method description Taguchi Method DMT Method
Design of orthogonal array Yes Yes
Noise Factors Yes Yes
Analyze Mean and variance Yes Yes
Signal to noise ratio Yes No
Desirability functions reduces multiple dimensions No Yes
Optimization of design factor levels by graph Yes Yes
5.3 FUZZY AGGREGATION Understanding of fuzzy logic in addition to classical crisp logic starts from the work of
Zadeh’s Sets theory paper which provides revolutionary ideas of fuzzy sets and
computational rule of inference [102]. The set (fuzzy) is a class of objects with scale of
number of grades of membership (MS). Fuzzy set is described by a MS function which
allocates to each object a score of MS ranging between 0 to 1. The idea of crisp sets is also
proved in fuzzy sets such as union, intersection, compliment and convexity etc. Fuzzy logic
deals with the classes of objects which don’t have accurately defined measure of MS [102]
such as class of beautiful girls and class of tall person. These sets or classes are not present in
crisp mathematics.
This concept is not only given to explain these fuzzy sets but to give a parallel framework
dealing like human for such problems [102]. Fuzzy logic is a very useful technique in the
areas of pattern recognition, information processing, classification and aggregation etc. The
concept of membership function raises the utility of fuzzy logic due to its wide variety of
functions and possibility of optimization of the shapes of the MS functions. Sivarao and
Brevern et. al [9] used fuzzy logic concept in the laser cutting process. Adapting network
fuzzy inference system ANFIS of MATLAB was selected for introducing to the researcher
55
working in the area of laser cutting modelling [9]. In the laser cutting machine higher number
of controllable variables has to be controlled as compare to other modern machine. It is more
scientific to model the problems such as kerf width and surface roughness.
With the increase in parameters, machining process increases and produces complexity which
introduces the usage of computer (neural networks, genetic algorithms and fuzzy logic) [9].
Sivarao [9] explained FIS system of MATLAB and identified the future work of comparative
study by changing the models based on FIS variables.
5.4 SUMMARY Researchers generally optimize laser cutting process for single quality and use OFAT
technique. For more than two variables optimization, suitable data mining techniques are
simple aggregation, normalized aggregation, Statistical methods, Genetic Algorithm (GA) or
fuzzy logic (FL). The aim of this research is to find out the best possible overall quality (edge
quality, kerf width, overcut and material removal rate) with Electricity-efficiency to predict
optimized input datasets.
Review shows that results of GA are better than TM and more superior with outer array.
Multi-objective problem can be solved by common genetic algorithm (CGA) by combining
both the quality parameters with or without weighting factor using averaging or aggregation.
It can be solved by multi-objective genetic algorithm (MOGA). It is probably faster and
efficient. The PSO results are better than GA, differential evaluation (DE) algorithm and
Quadratic programming (QP). They also studied the convergence rate, performance and
accuracy of PSO algorithm. Thirty to forty iterations are often required by PSO, which are
lower than other evolutionary computational methods. Hybrid TM has been used for multi-
objective optimization. However, weight assignment is a major task of multi-objective
optimization. Therefore, many researchers used hybrid techniques with Grey relational
analysis, it may be used with fuzzy aggregation. Some researchers used unnecessary hybrid
combination of TM and RSM for modelling and optimization results show significant
enhancement in kerf width and MRR qualities. They found that linear input and interaction
effects are significant in both the output qualities. It is difficult to utilize both Statistical
methods, even RSM DOE is capable to model the problem. Overall optimization of laser
welding of Titanium with Nd:YAG laser was performed with TM-based Grey analysis. They
mentioned the presence of interactions among pulse shape, shielding gas, welding speed,
56
pulse frequency, laser focus and laser energy in laser welding, but TM limitations reduces the
best results.
Desirability method, DMT, is commonly used to evaluate and optimize multiple quality
characteristic problems. The method incorporates desirability functions, a performance
statistic based on the mean squared error, and data-driven transformations to provide a
systematic approach that is adjustable to a variety of situations and easy for non experts to
apply. They take weighted geometric mean and MSE (mean square error), the DMT method
allows the decision-maker to alter the relative importance of the quality characteristics, mean,
and variance to suit the particular situation. Taguchi parameter design offers no similar
flexibility. Once the data are converted into individual desires simple geometric mean or
weighted geometric mean aggregation methods may be used to combine the individual
desires. In this method if one of the variables is undesirably high and other is very low range
then the aggregated suggest wrong prediction for overall quality.
RSM (Box-Behnken design) based model of laser cutting of MDF and AISI316L solved the
issue of overall quality in two steps by using the desirability approach. RSM model
established all factors effect on output depending on priority and constrained based on cost
and quality but is unable to provide single optimize overall quality. The decision is based on
operator to prioritize between quality and cost. It is better to use fuzzy aggregation, as fuzzy
can solve the issue of quality and cost in a simpler manner. Ming-Fei was able to achieve
multi-objective optimization for two qualities (optical transmittance ratio and roughness) of
CO2 laser cutting by applying grey relational analysis. The method of grey relational analysis
can be improved by considering interaction effect by applying better techniques in the
building block of measuring quality data. Sharma developed second order RSM and TM for
Al alloy sheet cutting by ND:Yag pulsed laser. The weights of both qualities were measured
by Taguchi-based GRA with entropy measurement methodology. It is an effective approach
to attain optimum values of both the responses only by OA and needs Statistical expert.
However, is unable to model with missing values because unbalanced matrix calculation is
not an easy task.
Understanding of fuzzy logic in addition to classical crisp logic starts from revolutionary
ideas of fuzzy sets and computational rule of inference by the work of Zadeh’s Sets theory
paper. Fuzzy set is described by an MS function which allocates to each object a score of MS
57
ranging between 0 to 1 with a wide variety of functions and possibility of optimization of the
shapes of MS functions. This concept is not only given to explain these fuzzy sets but to give
parallel framework dealing like human for such problems. Researchers used fuzzy logic for
expert systems and modeling in LCP. Adapting network fuzzy inference system ANFIS was
also workable in the area of laser cutting modelling. In the laser cutting machine higher
number of controllable variables have to be controlled as compare to other modern machine.
With the increase in parameters, machining process increases and produces complexity which
introduces the usage of computer (neural networks, genetic algorithms and fuzzy logic).
Earlier, we had not mapped input and output like others. Mapped input with aggregated
overall quality, which was aggregated after normalization on the basis L9 datasets. However
was unable to see individual values of each quality. Laser cutting of 800HT alloy by CO2
laser, experiment was performed using FD for overall quality of flatness and waviness.
However they used simple crisp aggregation by converting the values in score i.e. 1, 2 and 3.
They utilized scores as in desirability method instead of weightage and sum the scores in
crisp logic to calculate overall quality. The Aggregation function was further checked by
customer quality function used by the author. Fuzzy aggregation definitely improves the
results. As standard membership functions generate standard aggregation results with the
quality of applied rules and fuzzy aggregation. The solution with Fuzzy aggregation is further
explained in Chapter 6 and in complete detail in Chapter 8.
RSM (Box-Behnken design) based model of laser cutting of MDF and AISI316L solved the
issue of overall quality in two steps by using the desirability approach. RSM model
established all factors effect on output depending on priority and constrained based on cost
and quality but is unable to provide single optimize overall quality. The decision is based on
operator to prioritize between quality and cost. It is better to use fuzzy aggregation, as fuzzy
can solve the issue of quality and cost in a simpler manner. Ming-Fei was able to achieve
multi-objective optimization for two qualities (optical transmittance ratio and roughness) of
CO2 laser cutting by applying grey relational analysis. The method of grey relational analysis
can be improved by considering interaction effect by applying better techniques in the
building block of measuring quality data. Sharma developed second order RSM and TM for
Al alloy sheet cutting by ND:Yag pulsed laser. The weights of both qualities were measured
by Taguchi-based GRA with entropy measurement methodology. It is an effective approach
to attain optimum values of both the responses only by OA and needs Statistical expert.
58
However, is unable to model with missing values because unbalanced matrix calculation is
not an easy task.
Understanding of fuzzy logic in addition to classical crisp logic starts from revolutionary
ideas of fuzzy sets and computational rule of inference by the work of Zadeh’s Sets theory
paper. Fuzzy set is described by an MS function which allocates to each object a score of MS
ranging between 0 to 1 with a wide variety of functions and possibility of optimization of the
shapes of MS functions. This concept is not only given to explain these fuzzy sets but to give
parallel framework dealing like human for such problems. Researchers used fuzzy logic for
expert systems and modeling in LCP. Adapting network fuzzy inference system ANFIS was
also workable in the area of laser cutting modelling. In the laser cutting machine higher
number of controllable variables have to be controlled as compare to other modern machine.
With the increase in parameters, machining process increases and produces complexity which
introduces the usage of computer (neural networks, genetic algorithms and fuzzy logic).
Earlier, we had not mapped input and output like others. Mapped input with aggregated
overall quality, which was aggregated after normalization on the basis L9 datasets. However
was unable to see individual values of each quality. Laser cutting of 800HT alloy by CO2
laser, experiment was performed using FD for overall quality of flatness and waviness.
However they used simple crisp aggregation by converting the values in score i.e. 1, 2 and 3.
He utilizes scores as in desirability method instead of weightage and sums the scores in crisp
logic to calculate overall quality. The Aggregation function was further checked by customer
quality function used by the author. Fuzzy aggregation definitely improves the results. As
standard membership functions generate standard aggregation results with the quality of
applied rules and fuzzy aggregation. The solution with Fuzzy aggregation is explained in
Chapter 6 and in complete detail in Chapter 8.
CHAPTER 6
PROPOSED MODULAR RESEARCH
METHODOOLOGY
59
6. PROPOSED MODULAR RESEARCH METHODOLOGY Instead of difficult mathematical proof technique, an empiricism methodology is used to
solve the problem by Statistical and soft computing techniques. Propose the solution by
framework modules for Process improvement using the modular approach. The problem is
solved by the following modules as shown in Figure 6.1.
• Preprocessing of process module
• Experimental design module
• Modelling and optimization module
• Multi-quality optimization module
An empiricism research methodology is adopted to solve the problem by Statistical and soft
computing algorithms using proposed framework major focus on Computing rather than
Material Engineering. It is not as difficult as mathematical proof techniques and this
methodology is applied only on experimental data. Hence proposed framework does not
cover the theoretical/mathematical modelling problems.
In the process of laser cutting modelling or optimization of single or multiple qualities the
first module is used by every researcher as per their knowledge and demand. There are two
categories of researchers one who follow one factor at a time (OFAT) others who use design
of experiment (DOE). In both cases selection of material, laser cutting machine is necessary.
But proper screening of variables is not performed. The researchers normally need to inspect
the results carefully as part of procedure does not follow all the time. There are tools to
examine the observed data such as outlier analysis.
In the process of laser cutting parameter optimization the first vital stage is “Preprocessing of
Process Module”. In this stage first select the material to prove our framework workability. It
is easy to select non-metallic material because there is margin of research [39] in the area
with low power laser cutting machines and with low cost experimentation due to low material
cost. Selection of Laser cutting machine is normally based on the material properties. The
laser cutting machines are classified into low and high power. After the selection of material
and Laser cutting machine the next step is to study the possible parameters that include
controllable and uncontrollable parameters and constant parameters for the desired response
variables. This can be done by simple statistical analysis by screening or references and
domain experts’ opinion. This module is shown in Figure 6.1 and separately in Figure 6.2.
60
A. Preprocessing of Process Module
Selection of Material Selection of Laser MachineCO2 or ND:Yag
Screening controllable and uncontrollable parameter for the Desired qualities
Selection of controllable parameters for the Desired Quality
B Experiment Design ModuleTaguchi
Method L9/L18/L27
Factorial design
Fractional Factorial RSM CCD RSM BBD RSM BBD
C Modelling Module
Regression Taguchi Factorial Fractional Factorial RSM CCD RSM BBD ANN
D Multi-quality Optimization Module
RSMMulti-
objective optimization
Crisp logic aggregation
Genetic Algorithm
Fuzzy aggregation
Prediction error
Acceptable?
Sel
ect b
est
Optimize
Sel
ect b
est
Decide the fitness rules
Select best
Produce optimized results
Figure 6-1: Proposed Framework modules for Process Improvement
61
In the next stage of “Experimental design module” which is being used by the researchers
who follow DOE techniques, as shown in Figure 6.1 and Figure 6.3. The selection of DOE is
based on linear, non-linear, multi-input, single output, multi-output, number of variables and
the cost of material. The selection of DOE is a technical job based on some preliminary
analysis and experience. If the selected DOE is identified on the first stage that it will not
work then it is better to reselect another DOE. There are a lot of researchers using design of
experiment such as Taguchi’s orthogonal array, response surface methodology and factorial
design [1, 4, 12, 25, 47] .
Based on selected DOE the modelling technique is chosen from the modelling and
optimization module which is basically modelling module. But for the improvement of single
quality, it can work for optimization such as neural networks, regression analysis, Taguchi
method and factorial analysis etc. The adequacy of model is checked and it is needed to
decide whether to improve by replacing the modelling technique or reject it and go back to
DOE stage and select different DOE and perform the experiment again. If model is adequate
and requirement of one by one quality improvement is required then provide the machine
settings at this stage for production. Otherwise go to the next stage of multi-quality
optimization module which is shown in Figure 6.1 and Figure 6.9. Select the appropriate tool
for overall quality and predict best results based on fitness rules and suggest machine setting
for overall quality. It is better to used electricity-efficient solution with desired quality.
6.1 PREPROCESSING OF PROCESS MODULE For the purpose of optimization of the laser cutting parameters the essential initial phase is
“Preprocessing of Process Module”. Selection of material in order to prove framework
workability is the first step. Non-metallic materials provide a broader margin of research [39]
when it comes to low power laser cutting machines and are inexpensive as well keeping the
cost of experimentation low. Material properties provide the basis of selection of Laser
cutting machine. The laser cutting machines are classified into low and high power.
After the selection of material and Laser cutting machine the next step is to study the possible
parameters that includes controllable and uncontrollable parameters and constant parameters
for the desired response variables. The experts’ papers, experts themselves and statistical data
mining tools can identify the important parameters. This process is called screening of the
parameters. This module is shown in Figure 6.1 and separately in Figure 6.2.
62
A. Preprocessing of Process Module
Selection of Material Selection of Laser MachineCO2 or ND:Yag
Screening controllable and uncontrollable parameter for the Desired qualities
Selection of controllable parameters for the Desired Quality
Figure 6-2: Preprocessing of Process Module part of the Proposed Framework
6.2 EXPERIMENTAL DESIGN MODULE
6.2.1 SINGLE PARAMETER CHANGE AT A TIME In most of the experimental investigations of Laser beam cutting (LBC) process, one factor at
a time (OFAT) has been varied to analyze the effect of input process parameters on output
quality characteristics or responses [13, 16, 18, 20-22]. It is proved that the one parameter
change at a time process is very lengthy, difficult to perform and high cost for large number
of treatments. It is also unable to show the interaction effects of multiple parameters on
response variables. Hence, it does not always ensure the perfect modelling. Therefore, Design
of experiment (DOE) is a vital step in the study of product process quality. There are
numerous design of experiment available in research papers, some of the important DOEs are
discussed below. The analysis of DOE based data explains the relationship that the selected
DOE is appropriate otherwise revision of the DOE is required as explained in the Figure 6.1
and Figure 6.3.
B Experiment Design Module
Taguchi Method L9/
L18/L27
Factorial Design
Fractional Factorial RSM CCD RSM BBD RSM BBD
Figure 6-3: Experiment design module of the Proposed Framework
63
6.2.2 TAGUCHI METHOD The most simple and useful DOE is Taguchi Method (TM). Its small size decreases the
experimentation time and cost. It is derived from factorial design. The orthogonal array
L9 3 (4-2) has been derived from Factorial Design L(81) 3 (4). It shows that 9 runs are derived
from 81 runs of four variables with three levels. There are a number of orthogonal designs
available such as L4, L6, L8, L9, L18, L27 [103] and a lot more depending on the number of
treatments. This experiment is robust, small in size, low cost and produces excellent results
[2, 12, 16, 23] if interaction effect among the variables is negligible.
6.2.3 FRACTIONAL FACTORIAL DESIGN The fractional factorial design is also derived from factorial design, it normally reduces the
size of experiment by 1/3 or 1/2. This model results are closer than Taguchi method. It can
explain the interaction effect and produces better model than orthogonal array [73, 104].
6.2.4 FACTORIAL DESIGN The most established and useful design is Factorial Design (FD) which has the ability to
study the variation caused by independent and interaction effects on the response qualities.
The size of runs increases as the number of input parameters increases. The feasible size is 3
at the most 7 variables with two levels and 3 to 4 variables with three levels. It can be written
as L(Runs) (Levels) (Variables) therefore, in case L(81) 3 (4) means 81 runs with four variables
contains three levels. It is very important to understand that the data converted to ordinal data
of two, three or five levels definitely loses the information during the modelling based on this
ordinal data. However, Factorial design is closer to the actual model than the extracted or
derived DOEs such as Taguchi and fractional factorial design. While factorial design
experiment are unable to search optimize quality parameter directly, but are able to produce
basis for the results. However, RSM does better optimization [73, 104] .
6.2.5 RESPONSE SURFACE METHODOLOGY (RSM) Response surface methodology RSM is a statistical and mathematical method. It is helpful in
modelling and optimizing industrial machining processes which are normally two to three
times larger datasets than Taguchi method, but less than half observation as compared to
factorial design and is able to study low level interaction effects [23].
The size of RSM observations is smaller than factorial design but it is capable to model the
effects of interactions which will produce better results than Taguchi method if interaction
64
between the variables is significant [77]. Finally identify the optimize region of single quality
on three-dimensional diagram. Baş and Boyacı [77] utilize RSM and try to find out some
common mistakes and limitations of RSM. RSM model fits well but it is possible that its
prediction out of experimental data range does not predict well.
Noor and Kadirgama [57, 58] prepared a statistical model to predict surface roughness and
roughness height based on laser power, cutting speed and tip distance. The model is prepared
for acrylic (3mm thick) cutting by 30W pulse CO2 using the Box Behnken (RSM) design of
first order model. However it is better to apply quadratic model based on analysis results.
Noor and Kadirgama used Box Behnken design of experiment and applied Particle swarm
optimization in their second paper for multiple output qualities. They also defend the use of
Box Behken design of experiment. Its experiment size is smaller than Central composite
design of experiment because, axial points are absent and provide safe operating boundaries.
BB design is capable to bifurcate the order of significance and appropriate for optimization of
process by RSM and capable of linear and quadratic order modelling.
6.3 MODELING AND OPTIMIZATION MODULE Selection of modeling technique is based on the response of dependent variable trend.
Statistical modelling techniques are applied by many researchers and nowadays soft-
computing techniques are more popular. The statistical concept of orthogonal array and
factorial design is not ideal for the artificial neural network case. However, it is suited for
statistical modelling to predict better values though it is not perfect. As sampling data is never
equivalent to population data and also conversion of fraction data into three level ordinal data
loses data knowledge for perfect modelling as shown in Figure 6.4. The issue of sampling
size can be improved by selecting FD instead of OA [4]. There are a number of statistical and
soft-computing techniques available at this stage to understand these methods as well as to
compare them for better application of available tools. The possible methods are:
• One Way ANOVA
• Two Way ANOVA
• Single Variable Linear Regression Analysis
• Multivariable Regression Analysis
• Nonlinear Regression Analysis
• Multivariable nonlinear Regression Analysis
65
• Artificial neural network (Supervised learning)
• Artificial neural network (Semi-supervised learning)
The above problems can be solved by using excel. However, it is sometimes easier to handle
Statistical Software SPSS, Minitab et cetera.
Optimize the modelling results
C Modelling Module
Regres-sion Taguchi Factorial Fractional
Factorial
RSM CCD/BBD
ANN Supervised
ANN Semi-Supervised
Figure 6-4: Modelling module of the Proposed Framework
Figure 1.2 represents the four controllable input parameters with their value divided into
three stages to understand the effect of input parameters on the output parameters [93]. The
relationship between the parameters can be found by the modelling of the system by some
Mathematical, Statistical method and Soft-computing mapping. The model has been
developed for simulation and expansion of the size of datasets.
6.3.1 ANALYSIS OF VARIANCE M. Zaidi et al. [25] modeled the laser cutting data with the statistical techniques as explained
in Figure 6.5 flow chart. The data was forwarded from experimental design module based on
DOE observation tables. Perform “One way ANOVA” to understand the significance of
controllable parameters such as
• Laser Power (A)
• Cutting Speed (B)
• Assist Gas Pressure(C)
• Standoff Distance (D).
This portion is divided into two parts, with replication and without replication. The meaning
of replication is to take observations more than once to reduce random error. In the focused
66
experiment polystyrene foam was cut with replication three times. In the analysis of variance
these three assumptions are used:
• The response parameter is normally distributed.
• The variance of the response parameters is same.
• The response parameters are independent.
Start
Input Experimental data for Kerf-Width of Polystyrene foam
Conclude the Analysis
Perform Analysis of Variance
Perform One Way ANOVA to understand the significance of controllable parameters (A/B/C/D)
Perform Two Way ANOVA to understand the significance of interaction between controllable parameters (A/B/C/D)
Analyzing the significance of each parameter and also interaction effects between them
End
Figure 6-5: Analysis of variance
It is possible that the interaction effects are present. Therefore, in the second stage perform
“Two way ANOVA” to understand the significance of interaction between controllable
parameters (A/B/C/D). Analyzing the significance of each (A/B/C/D) parameter and
interaction effects between them (A and B, A and C, A and D, B and C, B and D, C and D).
67
The results indicate the real contributors of the output quality and hence the quality can be
improved by using the analysis results.
6.3.2 REGRESSION ANALYSIS
Start
Input Experimental data for Kerf-Width of Polystyrene foam
Conclude the Analysis
Draw Scatter Diagram
Calculate Correlation Coefficient to find the significance of relationship
Search the best fit to find the regression equation between two variables
Perform regression analysis for multivariable
End
If coefficient of correlation is significant
No
Yes
Figure 6-6: Regression Analysis
M. Zaidi et al. [25] modeled the laser cutting data with the Statistical techniques through
regression analysis as explained in Figure 6.6. The data was forwarded from experimental
and design module based on selected DOE observation tables. Draw Scatter plot, the scatter
68
plot shows the nature of relation between the variables. The Figure 6.6 may be Positive
linear, negative linear or curvilinear relationships or discernible. The flow chart shows that
the next step is to calculate coefficient of correlation “r” to see the significance of correlation
between the variables. If the value of “r” is not significant then predicting dependent
parameter values by regression equation is a useless practice. If “r” is significant then
determine regression equation by least square method. The purpose of regression analysis is
to see the relationship between dependent and independent parameters and predict unknown
values as and when required to reduce the time and cost of experimentation and design.
6.3.3 TAGUCHI, FRACTIONAL FACTORIAL AND FACTORIAL DESIGN Taguchi’s orthogonal array (OA) L9 was selected from experimental and design module. The
factorial design is selected for verification. However, after adequacy test the size of data set
increased from OA to fractional factorial design or factorial design observation tables.
Taguchi’ OA is one of the better methods to reduce cost of experiment, get better quality and
minimize design and development interval [3]. Dr. Taguchi had designed a method based on
OA experiments resulting in controlling variations for the experiment with optimum settings
of input controllable variables [2, 16]. OA gives sets of well balanced (minimum)
experiments. In this experiment four-three level orthogonal array L9(34) i.e. a full factorial
design (3x3x3x3) of 81 sets of experiments was reduced to 9 runs i.e. L9(34-2).
Similarly L9(34-1) is fractional factorial design which gives a much better representation as
sampling data. The size of fractional factorial design for the given matrix size is 27 data sets
which are 3 times greater than the orthogonal array of L9. At last the best possible
representation is factorial design i.e. L9(34). The size is 81 datasets. The fractional factorial
explain the interaction effect and produces better model than orthogonal array [73, 104] and
factorial design results are much better. However, Taguchi assumed independent variable
concept.
Taguchi method uses the Signal to Noise (S/N) ratio of output quality as quality parameter
for optimization with the property of robustness. It is utilized in data variation and prediction
of optimum results, these signal to noise ratio expressions are used in [2, 3, 16] and shown in
equation (7.1), (7.2) and (7.3)
.
69
The robustness reduced the noise/uncontrollable parameters effects on desired quality
parameter of product or process. Target performance (TPM), noise performance measurement
(NPM) and analysis of variance (ANOVA) explain the significance of controllable and
unknown parameter in term of percentage. The highest significant factor has higher
weighting effect on the cutting quality. The screening procedure needs to run again if the
analysis of variance results show that unknown factors participation is more than 50% in the
variation of cutting quality. It means the selected major controllable factor or factors are
missing in the list of controllable factors. M. Zaidi et al. [25] took three replications to reduce
uncontrolled error factors and human error by taking the arithmetic mean.
6.3.4 RESPONSE SURFACE METHODOLOGY (RSM) The more expressive and complicated methods is Response surface methodology (RSM). The
results of RSM technique are encouraging with small data sets which reduces the size of
experimental runs for the modelling. The surface and contour graphs are utilized to explain
the relationship between the dependent and independent parameters.
Response surface methodology RSM is a statistical and mathematical method. It is helpful in
modelling and optimizing industrial machining processes which is able to study low level
interaction effects [23]. RSM produces much better results than TM if interaction between the
variables is significant [77].
Mathew, et al. [14] response surface methodology central composite design of experiment
was used to model Carbon fiber reinforced plastic composites cutting by a pulsed Nd:YAG
laser at the optimum process parameter ranges. Predictive models have been developed for
heat affected zone (HAZ) and the taper of the cut surface. It shows RSM is capable to model
and optimize problems with Central composite design (CCD) of RSM.
Baş and Boyacı [77] mentioned that RSM model fits well but it is possible that its prediction
out of experimental data range is not predicting well. Dhupal et al. used ANN for RSM which
is capable to predict. Noor and Kadirgama [57, 58] fit first order and quadratic but quadratic
results are better for acrylic (3mm thick) cutting by 30W pulse CO2 using the Box Behnken
(RSM) design. There are many examples available to prove the adequacy of RSM. Therefore,
it can be used for optimization.
70
6.3.5 ARTIFICIAL NEURAL NETWORK BY SUPERVISED LEARNING M. Zaidi et al. [4] model laser cutting of Perspex sheet by artificial neural network with
supervised learning. The model requires preparation of training, test and Simulation datasets
for training and simulation purpose. Appropriate training algorithm is selected and utilizes
supervised learning technique.
After the selection of training algorithm, select the constant parameters for training
environment for the systematic training. Start training and observe the changes in number of
neurons and hidden layers. The results were recorded and used for analysis and discussion to
conclude the effectiveness of ANN with the given datasets with missing values replace by 0
i.e. uncut.
The results were evaluated based on novel assessment technique by minimum, maximum and
average percent error with mean square error. The model adequacy check results show that
supervised learning techniques are not sufficient for missing values with orthogonal array of
L9 observations.
Therefore, there is a need to increase the size of data for modelling or prepare some other
modelling technique and by increasing the size of datasets the model adequacy is achieved
with 60% data sets for training purpose. Essentially, training data sets size is 5 observation
only, 2 for verification and 2 for test. Therefore, there is need to increase the size of data sets.
But, Ranaganth and Viswanath [31] also model laser cutting process for surface roughness
with small datasets of fifteen runs and for testing 5 runs without missing values. They model
the process after 41 times weights randomization which shows this randomization improved
the model. This evidence motivated to prepare Semi-supervised model with small data sets of
OA by utilizing the randomization technique systematically even the limitation of missing
values.
71
Start
Define ProblemDesign of Experiment for verification
Prepare Procedure of Experimentation
Performed Experiment
Check experimental
tables
Calculate edge quality and kerf width Prepare training data
Prepare Test dataPrepare Simulation data
Compare Training AlgorithmSelect constants such as
Network type (feed forward back propagation)Training Algorithm(Levenberg Marquardt)
Performance (Mean Square Error) Hidden layers: Tangent Sigmoid
Output layers: Pure Linear
Change No. of neurons and hidden layer and prepare network
Performed Training, Testing and Simulation
Record data for discussion and analysis
End
False
Yes
Check model adequacy
False
Yes
Used for optimization or next step
Figure 6-7: Modelling and Verification Methodology of Supervised learning
72
6.3.6 NEURAL NETWORK BY SEMI-SUPERVISED LEARNING The data was forwarded from DOE observation tables for modelling. M. Zaidi et al. [12]
modelled laser cutting of Perspex sheet by artificial neural network with Semi-Supervised
learning algorithm. The model requires preparation of training, test and Simulation datasets
for training and simulation purpose. The FD has been selected for verification and orthogonal
array L9 for training in experimental and design module. After collection of experimental data
prepare training, testing and simulation data. The edge quality and kerf width data was
normalized and then applied for the training process. This technique provides better results in
the modelling of many problems for the stretching of the training data [4, 28, 105].
Mathematical expression used for Normalization is
Xnor =
(Xact − Xmin )(Xmax − Xmin )
(Nmax − Nmin ) + Dmin (6.1)
Similarly, the parameters have been put into two categories, controllable and constant for
training. Thus, an algorithm is prepared for Semi-Supervised training, simulation and
evaluation for the modelling of LCP with the exception that it handles very small size
datasets and functions even in the presence of missing values as shown in Figure 6.8.
In this algorithm some parameters have been taken as constants based on the previous study
[2-4] experience such as training algorithm, network type (feed forward back propagation)
Performance criteria, hidden layer transfer function(Tangent Sigmoid), output layer (pure
linear). However, these items can be selected for more than one setting e.g. network type can
be Gradient Descent with Momentum & learning rate, Levenberg Marquardt, and different
number of neurons can be selected.
In the process of creating networks for modelling the training data and testing data sets are
prepared. N. Qazi and H. Yeung [105] utilize stacked neural network and apply
normalization to stretch the span of applied data, using single response parameter for better
modelling. Similar experience was gained by M. Zaidi et al. in [4].
Normalization and single response model strategy is adopted from both the work experiences
[4, 105]. N. Qazi and H. Yeung [105] use Principal component regression (PCR) and P. Xiao,
et al. [106] utilize Particle swarm optimization (PSO) for the selection of initial weights for
global minimum error. However, initial weights can be adjusted by random points and may
be needed to search larger area for global minima. Even then it does not guarantee to be able
73
to reach global minima point in error space. B.J. Ranaganth and G. Viswanath [31] used
weights randomization technique and initialized about 41 times. This is difficult with
supervised learning and does not necessarily build a good model with 41 times with simple
randomization.
Similar technique of initial weight randomization up to 41 times were used by Sivarao and
Peter Brevern et al. [37]. They [38] used Dither Randomization for re-initialization of neural
network weights. However, for the selection of initial weights proven simple randomization
can be improved by Nguyen and Widrow methods [91] and initialize weights thousands of
times in semi-supervised method provide excellent models. This algorithm is also suitable for
small data sets size of OA DOE based experiments.
Kondo and Ueno [107] presented the idea to change the attributes of the feedback looped
GNDH type neural network architecture without supervisor intervention for example number
of inputs, hidden layers and neurons in each layer to improve the model accuracy based on
Akaike’s Information Criterion. Therefore, Semi-supervised algorithm automate the
supervised learning algorithm to semi-automatic and creates multiple networks by changing
number of neurons, learning rates, momentum and training algorithm with all possible
combinations, and for every combination network weights are reinitialized for 100, 500, 1000
or 3000 networks based on Nguyen-Widrow method [108]. This method recommends
applying small initial weight values in neural networks. The small values are selected by
randomization. Afterwards, weights are changed in such a way that the region of vital
importance is divided into small intervals. By assigning the initial weights in first layer, every
node is able to pick different interval from the beginning of the training session. Therefore,
every node in the hidden layer has a free will to fine-tune its size of interval and location in
error space in the process of training. Often these interval sizes are small and most of the
weight changes freeze. A famous example is that of "Truck‐Backer‐Upper" training, in it the
time of training is dramatically reduced from 2 days to a mere 4 hours [91].
The weight re-initialization value based on the error value improvement and number of
neurons in hidden layer was calculated by the following formula [35, 109]:
( )2
noninh
nnn +=
(6.2)
74
Semi Supervised Neural network modeling
Define ProblemDesign of Experiment for verification
Prepare Procedure of Experimentation
Performed Experiment
Check experimental
tables
Calculate edge quality and kerf width Prepare training, Test and Simulation data
Select constants such asNetwork type (feed forward back propagation)
Training Algorithm(Gradient Descent with Momentum & Levenberg Marquardt)
Performance (Mean Square Error) Hidden layers: Tangent Sigmoid
Output layers: Pure Linear
Record data for discussion and analysis
End
Initialize weights 100, 500, 1000, 3000 times based on Wind Nguyen-Widrow method instead of Random
Semi supervised Training, Testingand Simulation
Create network based on Change No. of neurons, learning rates, momentum
and training algorithm
Measure MSE, minimum, maximum and average error for Testing and factorial prediction
Select three best networks for new created network architecture
False
True
Figure 6-8: Modelling and Verification Methodology of Semi-Supervised learning
75
Where nnh is number of neurons in hidden layer, nni number of neuron in input layers and nno
number of neurons in output layers.
Finally, the best three networks are selected from all the networks, which were created
through the above process. The output is not just the best network but it is also the best
network architecture. If this architecture is used for further training, the results have a higher
probability of improving. The idea of selecting best networks comes from Genetic algorithm
elite values concept, Particle swarm optimization and firefly algorithms. This idea combined
with the systematic randomization method of Nguyen and Widrow [91] ensures better results.
This elite value changes on every result based on the error value comparison, thus, the
chances of losing values closer to the global minima diminishes.
6.4 MULTI QUALITY OPTIMIZATION MODULE
D Multi-quality Optimization Module
RSM Multi-objective optimization
Crisp logic aggregation
Genetic Algorithm
Fuzzy Aggregation
Sele
ct b
est
Decide the fitness rules
Produce optimized results
Figure 6-9: Multi-quality module of the Proposed Framework
Some of the other significant multi-objective research has been mentioned earlier in the
introduction and laser cutting literature review. But there is a need to discuss some research
Dubey and Vinod Yadava [59] discussed hybrid TM and RSM for the multi-response
optimization of laser cutting process. The weighted impacts of quality characteristics and cost
76
components are incorporated in this quality function. They mentioned that interaction is
significant therefore it is better to use straight RSM DOE and modelling for optimization.
L. K. Pan et al. [30] provides optimizations of multiple qualities of laser welding of
Titanium with Nd:YAG laser. The overall optimization was calculated by Grey relational
grade but complex interaction is significant in this process and modeling was done by TM.
Hence the results are definitely not best.
However Gani and I. A. Choudhury et al. [100] used Tagchi’s method (TM) and
optimization in case of end milling process with P10 TiN coated carbide on AISI H13 Steel
tool. The interaction effect is absent. Therefore, aim of optimization to provide solution was
achieved by using signal to noise ratio and Pareto ANOVA methods for detailed analysis.
H. A. Eltawahni, et al. [6, 101] developed RSM (Box-Behnken design) based model of laser
cutting of MDF and AISI316L and solved the issue of overall quality in two steps by using
the desirability approach. RSM model established all factors effect on output depending on
priority and constrained based on cost and quality. Eltawahni, et al. [6, 101] is unable to
provide single optimize overall quality. The decision is based on operator to decide whether
quality is more important or cost. It is better to use fuzzy aggregation
Ming-Fei et al. [61] achieved multi-objective optimization for two qualities (optical
transmittance ratio and roughness) of CO2 laser cutting by applying grey relational analysis.
The method of grey relational analysis can be improved by considering interaction effect by
applying better techniques in the building block of measuring quality data.
Sharma and Yadava [36] developed second order RSM and Taguchi model for Aluminium
alloy sheet by ND:Yag pulsed laser . Weights of both qualities were measured by Taguchi-
based GRA with entropy measurement methodology. Taguchi-based GRA is an effective
approach to attain optimum values of both the responses only by orthogonal array and needs
Statistical expert. However, it is unable to model with missing values because unbalanced
matrix calculation is a difficult task. M. Zaidi et al. [12] modelled L9 orthogonal array and
FD model with missing values by ANN. Therefore, it is better to model the problem with
soft-computing techniques. They are capable to handle the issue of missing values in fuzzy
aggregation by applying rules to handle it.
77
6.4.1 SIMPLE AGGREGATION AND CUSTOMER QUALITY FUNCTION Yilbas and Rashid [41] cut 800HT alloy by CO2 laser and performed experiment using
Factorial design (FD) for overall quality of flatness (F) and waviness (D). But used simple
aggregation by converting the values in score i.e. 1, 2 and 3. The overall quality = F+D. The
calculation is simple and crisp in nature. In the next work B. S. Yilbas [79] investigated
overall quality at the cut surface due to input attributes variation in the process of CO2 laser
cutting of stainless steel. The similar method applied for overall quality. He utilizes scores as
in desirability method instead of weightage and sums the scores in crisp logic to calculate
overall quality. The scores are assigned in flexible range therefore crisp does not reflect the
true participation of qualities in the overall value.
The aggregation function is used in data mining techniques to evaluate overall effects or
quality. But, there are ways to give priority to any quality in the total aggregated value by
assigning higher weigtage or converting the values into score and select the score criteria
which automatically gives higher priority to selected quality. Normalization is the method
which equalizes the priority of every participant in final aggregation value [2, 3, 11]. The
normalized aggregation is calculated by equation ( 8.16) and ( 8.17).
The customer quality function can be calculated after normalization and converted into
ordinal values using Table 8-44 by applying equation (8.18), which simply used the max
function among all qualities and assign maximum value to customer quality function. This
rule is completely in the favour of the customer. Therefore, Fuzzy aggregation is definitely
the superior method of aggregation due to many standard membership functions available and
rules applied in parallel. So, it produces balanced results from both manufacturer and
customer point of view and provides the indication of rework.
6.4.2 GENETIC ALGORITHM The aim of robust process design is to build a stable process that produces minimum
variation due to uncontrollable parameters [60]. Taguchi’s approach is unable to do multi-
objective optimization and constraints handling. Genetic algorithm (GA) provides the
techniques of simultaneous optimization of multiple dependent parameters. The ability of
genetic algorithm to generate huge amount of possible population and selection on cost
function basis made it an important off line quality tool [60]. It is improbable to trap in local
minima due to the assumption of GA and does not depend on continuity, uni-modality or
78
convexity of functions [97] . GA using mutation, crossover and reproduction genetic
operations then grow successive generations and apply values to transfer function to predict
the output quality for optimized value [60]. It is a better method to solve with outer array in
orthogonal array to handle noise factors [68-70]provided that the transfer function is available
for making cost function. Chatsirirungruang and Masami Miyakawa [68] study the
usefulness of GA as compared to OA method and concluded that GA produces good results.
Multi-objective problem can be solved by common genetic algorithm (CGA) by combining
both the quality parameters with or without weighting factor using averaging or aggregation
etc. It can be solved by multi-objective genetic algorithm (MOGA) [98]. It was concluded by
Fernandes and Vicente et. al. [98] that MOGA is probably faster and efficient. Pawar, et al.
[62] applied Particle swarm optimization (PSO) to solve multi-objective optimization
problem of rough and finished grinding process parameters of three responses. The method of
constraint is handled by penalty methods as others. The PSO results are better than GA,
differential evaluation (DE) algorithm and Quadratic programming (QP). They also studied
the convergence rate, performance and accuracy of PSO algorithm.
6.4.3 FUZZY AGGREGATION M. Zaidi et al. [110] performed this work to solve the issue of overall quality of laser cutting
work-piece. The data was collected from modelling and optimization module of best
predicted datasets. Instead of difficult mathematical proof technique an empiricism
methodology is used to solve the problem by soft computing techniques. Recently, many
researchers have applied intelligent algorithms such as Genetic Algorithm and fuzzy logic for
overall quality in [9, 60, 68-70, 98, 99, 111, 112]. The comparative study leads to find better
techniques as in iterative methodology. The methodology is explained briefly in the flow
chart as shown in Figure 6.10. The definitions and formulae are given below[2].
The input parameters with properties are explained in Figure 1.2. The primary output
qualities are edge quality and kerf width. The percent overcut and material removal rate can
be calculated by the side line length and kerf width. The definitions and formulae [2] can be
seen from equation (7.4), (7.5), and (7.6). The adequacy of overall quality is checked based
on selection of DOE, modelling results and the problem statement. After the collection of
data normalized aggregation values are calculated using equation ( 8.16), ( 8.17) and
customer quality function calculated by (8.18).
79
Start
Define Problem
Gather useful data of Polystyrene foam modelling
Suitability of GA for application
Study Normalized Aggregation functionStudy customer quality function
Study Genetic Algorithm GA for overall qualityStudy Fuzzy Logic FL for overall quality
True
False
Use Fuzzy inference system (FIS) Connect input/output with FIS
Select membership function on simulated quality datasets for all I/P and O/P
Fine tune membership functionDefine defuzzification
Apply simulated data for evaluation
Record data for discussion and analysis
End
Suitability of FIS/FL for application
False
True
Analyze results and predict better input data sets based on novel combination of CQF, Quantified fuzzy aggregation and Quantified
normalized aggregation. Calculate Energy consumption quality for factorial design with the help of Eltawahni et al. [106] Formulas of cutting
cost and flow rate. Including power consumed in factorial table and again
sorted with Similar novel combination.
Predict machine setting for overall quality with energy efficient solution
Figure 6-10: Fuzzy Aggregation as a Multi Quality Optimization
80
The selection criteria and reasoning has been discussed in the discussion part of overall
quality. If selected technique is fuzzy aggregation, then use Matlab’s Fuzzy inference system
(FIS) and connect laser cutting qualities values as a input and normalized aggregation values
with output for the mapping. Connect input with appropriate membership function and
optimize its range and available parameters which changes the membership function shapes.
The rules are applied by the theme of customer quality function (CQF) to get fuzzy
aggregation value. Fine-tune FIS based on mapping data. Record the fuzzy aggregation
values in observation tables. Utilize quality quantification information by Table 8-44 for the
conversion of normalized and fuzzy aggregation values into quantified normalized and fuzzy
aggregation for the selection of best overall quality. As we know the tags
• Excellent 0<X≤0.25
• Desired 0.25<X≤0.5
• Worst >0.5
Finally, the decision is made based on novel combination of quantified customer quality
function, fuzzy aggregation and quantified normalized aggregation. However, the concept of
electricity –efficient solution is considered along with the desired quality.
6.5 SUMMARY Framework modules proposed for Process improvement using four modular approaches as
shown in Figure 6.1.
• Preprocessing of process module
• Experimental design module
• Modelling and optimization module
• Multi-quality optimization module
An empiricism research methodology is adopted to solve the problem by Statistical and soft
computing algorithms using proposed framework, focusing mainly on Computing rather than
Material Engineering. This Framework is applied only on experimental data in the process of
LC of single or multiple qualities modeled by “Preprocessing of Process Module” as per their
knowledge and demand. There are two categories of methodology which are followed OFAT
and DOE. In both cases, it is easy to select non-metallic material because there is margin of
81
research in the area with low power laser cutting machines and with low cost experimentation
to prove the workability of the framework. Selection of Laser cutting machine is normally
based on the material properties. The laser cutting machines are classified into low and high
power. Screening of parameters was performed for the selection of controllable and
uncontrollable parameters and constant parameters for the desired response variables. This
can be done by screening or references and domain experts’ opinion.
In the next stage of “Experimental design module” selection of DOE is based on linear, non-
linear, multi-input, single output, multi-output, number of variables and the cost of material.
If the selected DOE will not work then it is better to reselect another DOE. Many researchers
use Single parameter change at a time, TM, Fractional Factorial design, FD and RSM
techniques which are briefly explained in this chapter. The orthogonal array L9 3 (4-2) has been
derived from Factorial Design L(81) 3 (4). It shows that 9 runs are derived from 81 runs of four
variables with three levels. There are a number of orthogonal designs available such as L4,
L6, L8, L9, L18, L27 and a lot more depending on the number of treatments. The fractional
factorial design L27 3 (4-1) is also derived from FD, it normally reduces the size of experiment
by 1/3 or 1/2. The feasible size is 3, at the most 7 variables with two levels and 3 to 4
variables with three levels. It can be written as L(Runs) (Levels) (Variables) therefore, in case L(81)
3 (4) means 81 runs with four variables contains three levels. As sampling data is never
equivalent to population data and data converted to ordinal data of two, three or five levels
loses the information during the modelling based on this ordinal data. However, Factorial
design is closer to the actual model than the extracted or derived DOEs such as Taguchi and
fractional factorial design. RSM is helpful in modelling and optimizing industrial machining
processes which are normally two to three times larger datasets than TM, but has less than
half observations as compared to FD and is able to study low level interaction effects. The
size of RSM is smaller than FD but it is capable to model the effects of interactions which
will produce better results than TM. Finally, identify the optimize region of single quality on
three-dimensional diagram. RSM model fits well but it is possible that its prediction out of
experimental data range does not predict well.
Based on selected DOE, the modelling technique is chosen from the “Modelling and
optimization module” for the improvement of single quality, it can work for optimization
such as neural networks, regression analysis, Taguchi method and factorial analysis etc. The
adequacy of model is checked and it is needed to decide whether to improve by replacing the
82
modelling technique or reject it and go back to DOE stage and select different DOE and
perform the experiment again. If model is adequate then go to the next stage of multi-quality
optimization module. The issue of sampling size can be improved by selecting FD instead of
OA. The applied methods/Algorithms flowcharts of “One Way ANOVA, Two Way ANOVA,
Single Variable Linear Regression Analysis, Multivariable Regression Analysis, Nonlinear
Regression Analysis, Multivariable nonlinear Regression Analysis, ANN (Supervised
learning) and ANN (Semi-supervised learning)” modelling techniques are explained for
implementation. The model has been developed for simulation and expansion of the size of
datasets.
Some of the other significant multi-objective research has been mentioned earlier in the
introduction and laser cutting literature review. But there is a need to discuss some research
using hybrid Statistical Techniques for the multi-response optimization of laser cutting
process such as RSM with TM, TM with Grey relational grade, TM with Pareto ANOVA,
grey relational analysis, Taguchi-based GRA with entropy measurement methodology, RSM
(Box-Behnken design) with desirability approach and the decision is based on operator to
prioritize between quality or cost. However, these methods do not attempt to solve with
missing values because unbalanced matrix calculation is a difficult task. The applied novel
Fuzzy aggregation combination is capable to handle the issue of missing values.
The chapter also explains the following techniques
• Simple aggregation and customer quality function
• Genetic algorithm
• Fuzzy aggregation
Finally, the decision is made based on novel combination of quantified customer quality
function, fuzzy aggregation and quantified normalized aggregation. However, the concept of
electricity –efficient solution is considered along with the desired quality.
CHAPTER 7
EXPERIMENTAL DESIGN MODULE & SETUP
83
7. EXPERIMENTAL DESIGN MODULE & SETUP
7.1 INTRODUCTION In the laser cutting experiment of X1, X2, X3…..Xn controllable independent parameters, the
Y1, Y2, Y3…..Yn dependent parameters have to be focused for the improvement of the sheet
cutting quality. The laser cutting data is designed for the estimation of the errors of training,
testing and simulation of small datasets in the presence of n% missing values in orthogonal
array. Like any other process as shown in Figure 7.1, a well planned experiment is performed
to identify the controllable variable effect on output response to investigate the reasons for
changes in the output [73].
Training and Simulation/Laser cutting ProcessInput Output
Con
trolla
ble
fact
ors
Unc
ontro
llabl
e fa
ctor
s
X1 X2 X3 ………….Xi
U1 U2 U3 ………….Xi
Y1 Y2 Y3 ………….Yi
Figure 7-1: General model of process
The constant factors have been selected based on the M. Zaidi et al., [2, 4] work in order to
reduce the complexity of the modelling process. The factorial design based inputs were
applied on selected network and calculated overall mean square error, maximum percent
error, minimum percent error and average percent error for training and simulation intended
for the selection of best model. The simulated results are compared with each other based on
average percent error and only three elite class models were selected. The best were selected
on the basis of sorting on average, mean square error and maximum percent error. This novel
error assessment method reduces human involvement and is able to evaluate different
84
combinations with thousands of times re-initialization of network weights instead of 10 to 12
times in supervised learning. The semi supervised algorithm is able to apply ANN far better
than supervised learning.
Experiments are carried out by researchers as a rule to determine something about a
particular system or process. The meaning of an experiment is a test. A planned experiment is
a test or set of tests conducted to identify the controllable variable effect on output response
to investigate the reasons for changes in the output and its main contributors [2, 4, 73] . A
systematic experimental process is called Design of Experiment (DOE). In the design of
experiment there are a number of factors which affect the output quality. The laser cutting
process modelling by data mining techniques such as ANN, under investigation can be
represented by the model proposed by Montgomery [73] as shown in Figure 7.3. In this
model X1, X2, X3,,..Xn are controllable parameters for example laser cutting machine has
many variables that can be controlled.
Process
Training, testing and Simulation
InputWork-piece
OutputKerf width, Edge
quality etc.
Controllable factorPower, cutting speed, frequency, standoff distance, pressure etc.
Uncontrollable factorTemperature, humidity, varing cylinder gas pressure, voltage
fluctuation etc.
ConstantNozzle diameter, delay time,
thickness of work-piece, assist gas pressure etc.
Figure 7-2: Laser cutting process
The following parameters are generally taken into consideration:
Controllable parameters
• Power (Watt)
• Cutting speed (mm/ min)
• Frequency (Hz)
• Standoff Distance (mm)
• Focal Distance (mm)
• Pressure (Bar)
• Duty cycle (%)
85
The output qualities generally under investigation by many researchers include:
Output quality parameters
• Measure the exhaust gas by gas chromatography
• Kerf Width
• Surface Roughness
• Heat affected zone
• Electrical energy consumption
• Edge quality
• Percent overcut, ought to be calculated by using kerf width value
• Material removal rate, ought to be calculated by using kerf width value
The variables U1, U2, U3,…….Ui are uncontrollable parameters which may be controlled. In
addition to these controllable factors, some other factors have been taken as constant to
reduce the complexity of problem. They include:
Constant parameters:
• Nozzle Diameter: 2mm or less
• Delay Time: 2 seconds
• Thickness of work piece: mm
• Assist Gas: Compressed Air/Nitrogen/Oxygen
• Corner Power: 70%
• Actual Sideline Length: 40 mm
• Operating mode is Continuous wave
Some unknown or noise factors may include:
• Temperature, Season dependent can be control by robustness
• Humidity, Season dependent can be control by robustness
• CO2 gas pressure in cylinder from start to end
• Backash in mechanical machine gears
• Fluctuation in electrical voltages causes variation in parameters
• Effect of controllable factors on each other
• Inherent errors in laser cutting system
• Variation in sheet due variation in thickness and material
86
The statistical analysis in [93] shows that output quality factors are not independent variables
and calculation of Percent over cut (POC) and Material removal rate (MRR) shows that they
are calculated by side line length and cutting speed. It is concluded that it is sufficient to train
and simulate only kerf width and edge quality of polystyrene foam. Therefore POC and MRR
will be calculated by the existing data which will reduce the work of training and simulation
of these two output quality factors. The Taguchi method (orthogonal array) was selected as
DOE and only 9 runs were performed with three time replications.
In 2nd design of experiment based study, laser cutting process and ANN training process are
the major processes. The process of laser cutting has been explained in the above discussion
as shown in Figure 7.1. But, this time Factorial design was selected as DOE and Perspex
glass sheet was chosen as material. The target of the second experiment is to verify the
simulation results with better modelling by ANN as shown in Figure 7.3.
Process
Training, testing and Simulation
InputExperimental
data
OutputCalculation, &
Novel error assessment
Controllable factorNo. neurons, learning rate,
momentum, no. of initialization, training algorithm etc
Uncontrollable factorInitial weights, Human mistakes in
observation etc
ConstantNetwork type, Training
Algorithms, Performance method, Hidden layers & output
layers transfer functions
Figure 7-3: Training, testing and simulation process
Similarly, training, testing and simulation on experimental data are assumed as a process.
There are a number of factors which affect the output quality of the model. This process, can
be represented by the model proposed in [73]. In this model X1, X2, X3,……..Xn are
controllable parameters. They include:
• Number of neurons in hidden layer
• Number of hidden layers
• Performance function
87
• Number of epochs
• Training algorithm
• Leaning rate
• Momentum
• Number of times initialization performed etc.
In addition to controllable factors, some other factors have been taken as constant to reduce
the complexity of problem such as:
• Network type (feed forward back propagation)
• Training Algorithm(Levenberg Marquardt)
• Performance (Mean Square Error)
• Transfer function (Hidden layers): Tangent Sigmoid
• Transfer function (Output layers): Pure Linear
• The process of training is supervised learning
The variables U1, U2, U3,….Ui are uncontrollable parameters which may be controlled. Some
unknown or noise factors may include:
• Initial weights
• Human mistakes in observation etc.
The variation in output variables were measured by the novel error assessment method which
reduces human involvement.
7.2 SCOPE AND LIMITATION The scope of research is to select the small size DOE, improve modelling and optimization
techniques based on data mining techniques such as Statistical or Applied Soft-Computing
focusing mainly on the Computing aspect. Therefore, the whole focus is on data mining
techniques and laser cutting process is selected for this purpose. The scope of the proposed
framework is experimental work i.e. empiricism research methodology.
88
Proposed framework scope does not cover the theoretical/mathematical modelling problems.
The model was trained on orthogonal array and accuracy of modelling was measured on
factorial design. The supervised and semi-supervised algorithm was used for modelling in the
special case of missing value.
The Laser cutting machine used in our experiment was Zech laser machine which is mainly
used for low power applications and its max power is 500 watt, which is sufficient for plastic,
polymeric and organic material. The machine is installed in University of Malaya (UM),
Malaysia and the experiment was conducted in University of Malaya in co-ordination with
Professor, Dr. Nukman Bin Yusoff. There is a huge distance between UM and SZABIST
therefore experiment was performed less number of times. Also cost is also a constraint
regarding material selection. After adding more data during the industrial work, the ANN
model can be improved progressively.
7.3 PLAN FOR EXPERIMENT DESIGN MODULE The laser cutting of Polystyrene foam and Perspex glass sheets materials was selected for the
experiment for improvement of machining process. The procedure is explained in brief and
safety of personnel was taken into consideration which includes protection from released
fumes before performing the experiment. The precautionary steps are taken based on the
selection material and type of machines used in the process [113]. The next step is to find out
the process parameters.
7.3.1 PROCEDURE The Procedure of Laser cutting of Perspex and polystyrene foam is shown in Figure 7.4. In
the first experiment Polystyrene foam sheets were selected as a material. CO2 Laser cutting
machine of 500 watts was used for this experiment. Taguchi Method (orthogonal array) was
selected as a DOE with only 9 observations. In M. Zaidi [2] ANN modelling based on only
nine observations were used for training, testing and Simulation. However, verification was
done by comparing with statistical results. The statistical modelling was performed with
possible suitable methods on such a small data sets without the missing values in order to do
better modelling and study the importance of interaction effects.
89
Laser cutting process Start
Selected: Materials are Polystyrene foam and Perspex glass sheets Cutting profiles is square Constant, input and output parameters list down Nozzle diameter
Prepared: Table for observation based on design of experiment
Measured: Average thickness of work piece by digital caliper
Machine Startup:Fit selected NozzleApply constant values to the experimentation processMachine mode ContinuousZech Carbon Dioxide Laser Machine switched onSupply of gas 1 and gas 2 to be switched on.Place work piece on the platformFocused nozzle onto work-piece in z-axis direction.To order the model from nest open C-CUT program.
Start Experimentation: Adjust input parameters rangeCut the material in three replications / repetitions.
Measure the Quality parameters by digital caliper and scope
Store Data: Store data in worksheets Store Pictures
Perform: AnalysisModellingOptimizationValidation
Report writing
Figure 7-4: Procedure of Laser cutting of Perspex and polystyrene foam
90
Similarly, in the 2nd experiment Perspex glass sheet were selected as a material. CO2 Laser
cutting machine of 500 watts was used for this experiment. Factorial design was selected as a
DOE with only 81 observations. In my previous work ANN modelling based on only nine
observations were not verified experimentally. Therefore, Factorial design is selected for
verification and also for the best possible smallest size for modelling by supervised neural
network.
A Semi-Supervised neural network algorithm was built to improve the modelling on
orthogonal array based DOE. This small size data was successfully utilized by Semi-
Supervised neural network algorithm.
7.3.2 ORTHOGONAL ARRAYS AND FACTORIAL DESIGN Dr. Genichi Taguchi’s quality engineering design of experiments (DOE) is a robust, efficient
statistical method of total quality method (TQM) for designing high quality systems at lesser
expense for laser cutting process [3]. M. Zaidi and I. Amin [3] mentioned that the concept of
design of experiment (DOE) method reduces cost of experiment, provides better quality and
minimizes design and development interval. It designs a systematic and efficient way to
optimize controllable parameter as explained in [3]. Orthogonal Arrays (OA) provide a set of
well balanced experiments. In this experiment four-three level orthogonal array L9(34) whose
equivalent full factorial design (3x3x3x3) i.e. 81, reduce in size to 9 test sets. It converts test
range into factors and levels. DOE using Orthogonal Array resulted in efficient and to the
point test set without losing information. Many researchers used the following Signal to
Noise (S/N) ratios of Taguchi's method [2, 3, 16].
)/(log10 2210 σyNTB = (7.1)
)(log10 2210 ySTB +−= σ
(7.2)
++
−= 2
2
210311log10yy
LTB σ
(7.3)
Where NTB is normal the better in equation (7.1), STB smaller the better in (7.2) and LTB is
Larger the better in (7.3).Where y� is mean of output quality replicated values and σ2 is
variance between the replicated values.
91
Table 7-1: Four variables with three levels Orthogonal design matrix
Run Input Variable Measurement(Units) Laser Power
A
Cutting speed
B
Assist gas Pressure
C
Standoff Distance
D
Replication Mean S/N Ratio R1 R2 R3
1 100 0.2 0.5 1 2 100 0.7 2.5 5 3 100 1.2 4.5 10 4 300 0.2 2.5 10 5 300 0.7 4.5 1 6 300 1.2 0.5 5 7 500 0.2 4.5 5 8 500 0.7 0.5 10 9 500 1.2 2.5 1
Table 7-2: Controllable input and factors levels
Input Factors Level 1 Level 2 Level 3 Units Laser Power (LP) 100 300 500 Watt Cutting speed(CS) 0.2 0.7 1.2 m/min Assist gas Pressure (AGP) 0.5 2.5 4.5 Bars Standoff Distance (SD) 1 5 10 Mm
Table 7-1 shows the Experiment orthogonal design matrix which includes four input
controllable parameters with three stage ordinal data type in orthogonal array pattern and two
quality parameters for the measurement point of view for both the experiments performed for
Polystyrene foam and Perspex sheet. Appendix A (Table A-1) shows the experimental four
variables with three levels Factorial design matrix and two quality parameters for the
measurement point of view for both the experiments performed for Polystyrene foam and
Perspex sheet. The repeatability is achieved by three time replication. As the experiment is
performed with the same method, same researcher, same measurement tools, repetition over a
short time, same lab and same maintained conditions in lab.
Table 7-2 explains input factors three levels to cover the whole range of input data and also
shows their units for Polystyrene foam and Perspex sheet. The available laser cutting
92
machines are 0 to100 and 0 to 500 watts. Therefore, the whole work statements and claimes
based on the limitation of maximum possible laser power of 500 Watts.
7.4 EXPERIMENTAL SETUP In this section the experimental setup of both Polystyrene foam and Perspex sheet cutting is
explained which includes:
• Laser Machine
• Measurement tools
• Work piece material
• Data collection
• Verification of simulation data
7.4.1 LASER MACHINE 500 WATTS Both the experiments were conducted in Malaya University. Zech Laser Austria system (CO2
laser cutting machine) was used for cutting Polystyrene foam thickness of 13 mm and
Perspex sheet of thickness of 3 mm and 5 mm using the laser cutting machine as shown in
Figure 7.5 and Figure 7.6. It shows the laser cutting workstation, the model is ZL1010
specifications are given in Table 7-3 and Laser generation system ZLX5 specifications in
Table 7-4 using AutoCAD, C-Cut® for Laser cutting and Zechlaser for automatic machining
mode software.
\
Figure 7-5: Picture of Zech Laser System
93
Table 7-3: ZL1010 Specification
S. No. Parameter Ranges
1 Max. Cutting Speed 7500 mm/min
2 Maximum Rapid Power 10,000 mm/min
3 Precision Of Position +/- 0.10 mm
4 Precision Of Outline +/-0.10 mm
5 Weight Of Machine 550 Kilogram
6 Height Of Working Table 840 mm
7 X-Axis 1000 mm
8 Y-Axis 1000 mm
9 Z-Axis 100 mm
10 Max. Weight Of Working Piece 120 Kilogram
11 Size Of Machine
Length, Breadth, Height
(5280 m3)
2000 mm,1660 mm,1600 mm
Table 7-4: ZLX5 Specification
S. No. Parameter Ranges
1 Weight 630 Kilogram
2 Cooling Power 5 kW
3 Water Cooling 16 -18 ºC , 12 liters/min
4 Electrical Power 400V, 50Hz, 16A
5 Gas Mixture Co2 -N2 –He 7-28-65 (%)
6 Mode Structure Gaussian modes ,TEM01
7 Beam Divergence 2 mrad
8 Beam Diameter 14 mrad
9 Power Output CO2 laser up to 500 Watt
10 Laser system Continuous wave CO2 laser
94
Figure 7-6: Laser cutting process
7.4.2 MEASUREMENT TOOLS OF EXPERIMENT 1 and 2 For the selection of measurement tool, first identify the type, possible range and accuracy. In
this experiment digital caliper and microscope were used. The digital Vernier is easier for
measurement and reading in digital format. Routine calibration is used to avoid errors in
measurement, it is less prone to human error as compared to analog vernier caliper. It can
measure internally and externally in the range of 0 to 150mm thickness and least count
0.01mm as shown in Figure 7.7 . In these experiments, it is used to measure Polystyrene foam
and Perspex sheet thickness, standoff distance side line lengths are also shown in Figure 7.8 .
Figure 7-8: Inner and outer sideline length
Figure 7-7: Picture of digital caliper
Kerf width
Lin
Sheet
Lout
Work-piece
95
Light optical microscope (icamcope) were used to zoom in and show the geometry of edge
quality as shown in Figure 7.11 . The edge surfaces of specimens were examined using these
instruments. The results were taken by magnification of 40x and 100x in “icamscope”.
Figure 7-9: Schematic diagram of light optical microscope
Figure 7-10: Schematic diagram of icamscope
Figure 7-11: Perspex sheet edge quality measured by microscope
Edge Quality
96
7.4.3 PROPERTIES OF POLYSTYRENE FOAM Polystyrene foam sheet is used as work piece material for this experiment. It is usually white
and made of expanded polystyrene beads. Polystyrene's chemical formula is (C8H8)n, it
contains the chemical elements carbon and hydrogen. Expanded polystyrene (C8H8)n is a
generic term for polystyrene and styrene copolymers. Expanded Polystyrene is supplied to
molders in the form of a polystyrene bead. Normally used in packing, models and building
insulation etc. Some of the important properties of work piece material are listed in
Table 7-5.
Table 7-5: Properties of Polystyrene foam
Properties Range
Density 1.05 g/cm3
Density of EPS 25–200 kg/m3
Dielectric constant 2.4–2.7
Electrical conductivity (s) 10−16 S/m
Thermal conductivity (k) 0.08 W/(m·K)
Melting point 240 °C
Linear expansion coefficient (a) 8 10−5 /K
Specific heat (c) 1.3 kJ/(kg·K)
The thermal conductivity and specific heat are considered to be temperature dependent. The
convection coefficient and density are considered to be temperature independent also
mentioned by [15] . The cushioning is dependent on bead shape and size as well as density.
The strength increases with density. A key benefit of Expanded Polystyrene is that it is
recyclable and its disadvantage is that it is not consumed by time even buried in earth.
7.4.4 PROPERTIES OF PERSPEX MATERIAL Perspex sheet is used as work piece material for this experiment. The material properties of
Perspex sheet is shown in Table 7-6 [114].
97
Table 7-6: Properties of Perspex sheet (Cast Acrylic)
Property setting Units Value
Physical property
Relative Density g/cm³ 1.19
Water Absorption % 0.2
Mechanical property
Tensile Strength at yield 5mm/min MPa 75
Elongation at break 5mm/min % 4
Flexural Modulus 2mm/min MPa 3210
Property setting Units Value
Flexural Strength at yield 2mm/min MPa 116
Rockwell Hardness M Scale 102
Thermal property
Vicat Softening Temperature ˚C >110
Coefficient of Thermal Expansion mm/m°C 0.077
Optical property
Light Transmission 3mm sheet % >92
Refractive Index 1.49
Electrical property
Dielectric Strength kV/mm-1 15
Surface Resistivity Ω m-2 >1014
7.5 DATA COLLECTION
7.5.1 POLYSTYRENE FOAM and PERSPEX SHEET Laser cutting of Polystyrene foam(13mm sheet), Perspex glass sheet 3mm and 5 mm data
was observed and collected in the observation tables of edge quality are shown in Appendix
A (Table A-2, Table A-4 and Table A-7) respectively.
98
EDGE QUALITY: Is a distance between two maximum effective parallel straight lines
(tangents) after cutting straight profile. It also includes perpendicular and flatness deviation
as shown in Figure 7.12.
KERF Width:
Laser cutting of Polystyrene foam 13mm Perspex glass sheets 3mm and 5mm data was
observed and collected in the given table format and complete observation tables of Kerf
width are shown in Appendix A (Table A-3, Table A-5, Table A-6 and Table A-9
respectively.
Figure 7-13: Views of defective cuts
The width of the path of laser beam which is moving through the work piece is defined as a
kerf. Measurement for kerf width was performed at the straight section cut by the laser. The
kerf width is a distance of the cut surfaces at the upper edge of cut or with existing melting of
top edge immediately below, as caused by the laser beam as shown in Figure 7.8.
𝐾𝐾𝐾𝐾𝐾𝐾𝑓𝑓 𝑊𝑊𝑖𝑖𝑊𝑊𝑊𝑊ℎ =(𝑊𝑊𝑊𝑊𝐾𝐾𝑊𝑊 𝑝𝑝𝑖𝑖𝑝𝑝𝐾𝐾 𝐿𝐿𝑊𝑊𝑜𝑜𝑊𝑊 −𝑊𝑊𝑊𝑊𝐾𝐾𝑊𝑊 𝑝𝑝𝑖𝑖𝐾𝐾𝑝𝑝𝐾𝐾 𝐿𝐿𝑖𝑖𝑖𝑖 )
2
(7.4)
Edge quality
Sheet 40 mm
Figure 7-12: Maximum deviation between cut edges
99
PERCENT OVER-CUT: Is explained in terms of percentage over-cut of the measured
length of work piece length by the following simple formula.
Percent Overcut =
(Measured 𝐿𝐿𝑊𝑊𝑜𝑜𝑊𝑊 − Actual 𝐿𝐿𝑊𝑊𝑜𝑜𝑊𝑊 ) Actual 𝐿𝐿𝑊𝑊𝑜𝑜𝑊𝑊
× 100 (7.5)
Negative sign in percent over-cut shows that the measured side line length is smaller than
actual side line length and positive is vice versa as shown in Figure 7.8.
MATERIAL REMOVAL RATE (MRR): In laser cutting experiment it is a function of
kerf width, cutting speed and work piece thickness. 𝑀𝑀𝑀𝑀𝑀𝑀 = 𝑇𝑇ℎ𝑖𝑖𝑝𝑝𝑊𝑊𝑖𝑖𝐾𝐾𝑖𝑖𝑖𝑖 × 𝐶𝐶𝑜𝑜𝑊𝑊𝑊𝑊𝑖𝑖𝑖𝑖𝐶𝐶 𝑖𝑖𝑝𝑝𝐾𝐾𝐾𝐾𝑊𝑊 × 𝐾𝐾𝐾𝐾𝐾𝐾𝑓𝑓 𝑤𝑤𝑖𝑖𝑊𝑊𝑊𝑊ℎ (7.6)
7.6 VERIFICATION OF SIMULATED DATA Factorial Design was selected for the Confirmation experiment (CE) of orthogonal array
based modelling. The statistical and artificial neural network based prediction modelling
ability was confirmed in the second experiment of Perspex glass sheet. It is necessary to
prove the simulation ability. Therefore, experiment was performed on Perspex sheet and the
results were ensured.
7.7 SUMMARY LCP is explained pictorially based on DOE literature including list of possible parameters of
Controllable, quality, Constant and uncontrollable or noise. These parameters have to be
focused for the improvement of the sheet cutting quality. The laser cutting data is designed
for the estimation of the errors of training, testing and simulation of small datasets in the
presence of n% missing values in OA or FD. Like any other process as shown in Figure 7.1
and Figure 7.2, a well planned experiment is performed to identify the controllable variable
effect on output response to investigate the reasons for changes in the output and finally
expand the size of existing datasets. The constant factors have been selected based on the
previous experience in order to reduce the complexity of the modelling process.
Similarly, training, testing and simulation on experimental data are assumed as a process and
mentioned all parameters in the list and also shown pictorially. This model again shows list of
possible parameters of Controllable variables, quality parameter, Constant parameter and
uncontrollable parameters. These parameters have to be focused for the improvement of the
modelling and change each parameter as per algorithm shown in Figure 6.8.
100
In the earlier work FD based inputs were applied on selected network and calculated overall
mean square error, maximum percent error, minimum percent error and average percent error
for training and simulation intended for the selection of best model. The simulated results are
compared with each other based on average percent error and only three elite class models
were selected. The best were selected on the basis of sorting on average, mean square error
and maximum percent error. This novel error assessment method reduces human involvement
and is able to evaluate different combinations with thousands of times re-initialization of
network weights instead of 10 to 12 times in supervised learning. The Material engineering
formulas show that output quality factors like Percent over cut (POC) and Material removal
rate (MRR) are not independent, they can be calculated by side line length and cutting speed.
It is concluded that it is sufficient to train and simulate only Kerf width and Edge quality of
polystyrene foam and other parameters can be calculated. The first experiment observation
tables are based on OA and in 2nd FD is selected for experimentation of Perspex sheet. The
target of the second experiment is to verify the simulation results with better modelling by
ANN.
The Plan, Procedure and setup of experimentation with selected DOE based machine setting,
table of observation were mentioned in this chapter. The laser cutting of Polystyrene foam
and Perspex glass sheets materials is explained in brief and safety of personnel was taken into
consideration which includes protection from released fumes before performing the
experiment. CO2 Laser cutting machine of 500 watts was used for both experiments. TM
(OA) was selected as a DOE with only 9 observations. The experimental verification was not
done with Polystyrene data so it was verified with TM results. Similarly, in the 2nd
experiment of Perspex glass sheet FD was selected as a DOE with only 81 observations. FD
is selected for verification and also for the best possible smallest size for modelling by
supervised neural network.
Many researchers used Signal to Noise (S/N) ratios of TM. The observation table is prepared
from experimental data and Controllable input and Appendix A (Table A-1) shows the
experimental four variables with three levels FD matrix and two quality parameters for the
measurement point of view for both the experiments performed for Polystyrene foam and
Perspex sheet.
101
In the end experimental setup of both materials sheet cutting is explained which includes
Laser Machine specification, Measurement tools, Work piece material properties, Data
collection and Verification of simulation data. Both the experiments were conducted in
Malaya University. Zech Laser Austria system (CO2 laser cutting machine) was used for
cutting Polystyrene foam thickness of 13 mm and Perspex sheet of thickness of 3 mm and 5
mm using the laser cutting machine. In this experiment digital caliper and microscope were
used. Light optical microscope (icamcope) was used to show the geometry of edge quality.
The edge surfaces of specimens were examined using these instruments. The results were
taken by magnification of 40x and 100x in “icamscope”.
Polystyrene foam and Perspex sheet is used as work piece material for these experiments.
The material properties of both the materials are mentioned in the chapter.
Laser cutting of Polystyrene foam (13mm sheet), Perspex glass sheet 3mm and 5 mm data
was observed and collected in the observation tables for edge quality and Kerf width. Design
was selected for the Confirmation experiment (CE) of orthogonal array based modelling. The
statistical and artificial neural network based prediction modelling ability was confirmed in
the second experiment of Perspex glass sheet. It is necessary to prove the simulation ability.
Therefore, experiment was performed on Perspex sheet and the results were ensured.
CHAPTER 8
DISCUSSION OF RESULTS AND ANALYSIS
102
8. DISCUSSION OF RESULTS AND ANALYSIS The problem is solved using the Processed Framework for the Improvement of process. The
whole effort is to improve the techniques or algorithm to improve the process optimization.
The chapter discussed the preprocessing of process and after selection of material, machine
and screening they demonstrate and discussed the outlier analysis. Modelling and
optimization module is vital part of this dissertation. A detailed analysis of variance One/Two
ANOVA with and without replication was performed and discussed. Similarly, Linear,
multiple linear, nonlinear and multiple non-linear regression analysis were carried out and the
results were discussed. Though the First dataset was without missing values, still there is a
need to improve the modeling. Therefore, ANN supervised learning was attempted to model
the problem with missing values datasets of Perspex sheet. The experience of supervised
learning was discussed and provided advantage in the preparation of Semi-Supervised
learning algorithm. A detailed solution of overall quality was discussed from simple
aggregation to combination of Fuzzy aggregation.
8.1 PREPROCESSING OF EDGE QUALITY AND KERF-WIDTH DATA Experiments of laser cutting on polystyrene foam sheet had been performed in order to
understand the relation between cutting parameters and the resulting cutting quality. This was
achieved using Taguchi’s method. Some variable parameters were given higher attention
during the laser cutting process such as edge quality, kerf width, percent overcut and material
removal rate as it affected the quality of laser cutting process. CO2 laser machine with
maximum continuous wave output power of 500W was employed for carrying out the
experiment. Polystyrene foam sheets with 13 mm thickness were used as work pieces. To
focus the laser beam, a ZnSe lens of 127mm focal length was installed in the lens holder.
Compressed air was introduced co-axially with the laser beam through a converging conical
nozzle of about 2 mm exit diameter. The cutting output quality parameters included laser
power (A), cutting speed (B), assisting gas pressure (C) and standoff distance (D). A/ B/ C/ D are the controllable parameters of the experiment. The selection of input test sets
is based on orthogonal array. To avoid uncontrollable and human factor errors, data was
replicated three times and reduced by averaging. To avoid the following human errors outlier
analysis is used. The possible errors can be detected by outlier analysis for example:
• Wrong measurement once in three times • Missing values
103
• Miss matched reading etc.
Table 8-1: Measurement of Edge Quality of Polystyrene foam
Run Laser
Power
A
Cutting
Speed
B
A. Gas
Pressure
C
Standoff
Distance
D
Edge Quality Edge
Quality Mean Replication
R1 R2 R3
1 100 0.2 0.5 1 3.5 4 3.5 3.667
2 100 0.7 2.5 5 2.2 3 3 2.733
3 100 1.2 4.5 10 1.5 1.5 2 1.667
4 300 0.2 2.5 10 1.5 1.5 2 1.667
5 300 0.7 4.5 1 1 1 1 1.000
6 300 1.2 0.5 5 2.5 2.5 3 2.667
7 500 0.2 4.5 5 1.5 2 3 2.167
8 500 0.7 0.5 10 1.5 2 2 1.833
9 500 1.2 2.5 1 1.5 2.5 2 2.000
Figure 8-1: Outlier analysis of Edge Quality replication of Polystyrene foam
Figure 8.1 shows that the replicate values are reasonably closed to each other except run
number 7 which is also acceptable which shows better repeatability. Choua at al. [115]
defined that before the calculation of signal to noise ratio in case of 0 value in algorithm used
zero value directly as in run 5.
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
0 1 2 3 4 5 6 7 8 9
Repl
icat
ion
Run
Replication 1
Replication 2
Replication 3
104
Table 8-2: Measurement of Kerf Width of Polystyrene foam
Run LP A
CS B
AGP C
SD D
Kerf Width Kerf Width Mean (y1)
Replication 1 2 3
1 100 0.2 0.5 1 1.58 1.615 1.525 1.573
2 100 0.7 2.5 5 1.73 1.255 1.48 1.488
3 100 1.2 4.5 10 1.66 1.695 1.86 1.738
4 300 0.2 2.5 10 1.94 1.885 1.915 1.913
5 300 0.7 4.5 1 1.765 1.775 1.985 1.842
6 300 1.2 0.5 5 1.655 1.715 1.855 1.742
7 500 0.2 4.5 5 2.01 2.04 1.855 1.968
8 500 0.7 0.5 10 1.975 1.94 2.29 2.068
9 500 1.2 2.5 1 1.79 1.89 2.08 1.920
Figure 8-2: Outlier analysis of Kerf width replication of Polystyrene foam
Figure 8.2 shows that the replicate values are reasonably close to each other except run
number 2 and 8 which shows better repeatability. The idea is to use these input parameters
and output quality parameter for training of data by neural network, to be able to predict
output variable against the given input parameters so as to complete a full factorial table
without experiment.
1
1.2
1.4
1.6
1.8
2
2.2
2.4
0 1 2 3 4 5 6 7 8 9
Repl
icat
ion
Run
Replication 1
Replication 2
Replication 3
105
Similarly, experiments of laser cutting on Perspex glass sheet had been performed using FD
DOE. Selected quality parameters were given higher attention during the LCP as it affected
the quality of laser cutting process. CO2 laser machine with maximum continuous wave
output power of 500W was employed for carrying out the experiment. Perspex glass sheet
with 3 mm and 5 mm thickness were used as work pieces and other machine parameters were
remain same.
In the process of Perspex glass sheet (3mm) cutting, some of the machine settings are unable
to cut at 21, 24 and 27th “run”s as shown in Appendix A (Figure A-1, Table A-4, Figure A-2
and Table A-6) due to mediocre range adjustments of input parameters in Appendix A (Table
A-3). Similarly, In the process of Perspex glass sheet (5mm) cutting, some machine settings
are unable to cut at 11, 12, 15, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 48, 49, 50,51, 53, 54, 74,
75, 77, 78, 80 and 81 “runs” as shown in Appendix A (Figure A-3, Table A-7, Figure A-4
and Table A-8) due to mediocre range adjustments of input parameters of Appendix A (Table
A-3). Therefore, researcher is unable to measure edge quality and kerf width of Perspex sheet
qualities in uncut condition and assigned 0 value in observation tables.
M. Zaidiet al. [25] discussed the possible solutions of non-linear multivariable by
experimental data mining techniques using polystyrene foam cutting data [2]. Taguchi
method is a very useful technique to reduce the time and cost of the experiment but it ignores
all kind of interaction effects. The results are not much encouraging and motivated to study
Laser cutting process of non-linear multivariable modeled by one and two way analysis of
variance, also linear and nonlinear regression analysis.
Four controllable input parameters value range converted to three stages ordinal type variable
to understand the effect of input parameters on the output parameters [2]. The relationship
between the parameters can be found by the modelling of the system by some mathematical
or Statistical methods.
8.2 ANALYSIS OF VARIANCE In the beginning one way ANOVA was performed to understand the significance of
controllable parameters (Laser power (A), cutting speed (B), assist gas pressure(C) and
standoff distance (D) [2]. This analysis is divided into two parts with and without replication.
In the focused experiment polystyrene foam was cut, replicated three times as shown in Table
106
8-3 Appendix A (Table A-10 to Table A-21). Kerf width is selected for analysis and
modeling as it is the easier than edge quality and without any unsuccessful cutting.
8.2.1 ONE WAY ANOVA WITHOUT REPLICATION In one way ANOVA the effects of input parameters are analyzed one by one on laser cutting
quality of kerf width. Three replicated observations were including in the calculation in case
of with replication but for without replication only one set of replication is used. Smaller kerf
width means better cutting quality therefore, target performance measurement (TPM) is
smaller the better and signal to noise ratio is larger the better. The continuous input variables
were transformed in three levels and built an input table based on orthogonal array (OA),
which reduced the number of observations and time of experiment [18].
Table 8-3: Observations consider Laser Power (A)
100 300 500
1.573 1.913 1.968
1.488 1.842 2.068
1.738 1.742 1.920
Table 8-4: Summary of descriptive Statistics
Groups Count Sum Average Variance
100 3 4.800 1.600 0.016
300 3 5.497 1.832 0.007
500 3 5.957 1.986 0.006
The analysis of individual input parameters on Kerf Width Mean were done based on the data
recorded in Appendix A (Table A-3). Following formulas from equation (8.1) to (8.7) were
used for one way ANOVA with or without replication analysis shown in Table 8-3 to Table
8-21 by using SPSS Software.
𝑋𝑋�𝑗𝑗 = �𝑋𝑋𝑖𝑖𝑗𝑗 /𝑖𝑖𝑗𝑗
𝑖𝑖𝑗𝑗
𝑖𝑖=1
( 8.1)
𝑖𝑖𝑗𝑗2 = �(𝑋𝑋𝑖𝑖𝑗𝑗 − 𝑋𝑋�𝑗𝑗 )2
𝑖𝑖𝑗𝑗
𝑖𝑖
𝑖𝑖=1
( 8.2)
𝑆𝑆𝑆𝑆𝑇𝑇𝑀𝑀 = �𝑖𝑖𝑗𝑗 �𝑋𝑋�𝑗𝑗+𝑋𝑋��
𝑊𝑊
𝑗𝑗=1
2 ( 8.3)
107
𝑀𝑀𝑆𝑆𝑇𝑇𝑀𝑀 = �𝑆𝑆𝑆𝑆𝑇𝑇𝑀𝑀𝑊𝑊 − 1
𝑊𝑊
𝑗𝑗=1
( 8.4)
𝑆𝑆𝑆𝑆𝑆𝑆 = ��𝑖𝑖𝑗𝑗 − 1�𝑖𝑖𝑗𝑗 2𝑊𝑊
𝑗𝑗=1
( 8.5)
𝑀𝑀𝑆𝑆𝑆𝑆 = 𝑆𝑆𝑆𝑆𝑆𝑆𝑖𝑖𝑇𝑇 − 𝑊𝑊
��𝑖𝑖𝑗𝑗 − 1�𝑖𝑖𝑗𝑗 2𝑊𝑊
𝑗𝑗=1
( 8.6)
𝐹𝐹 =𝑀𝑀𝑆𝑆𝑇𝑇𝑀𝑀𝑀𝑀𝑆𝑆𝑆𝑆
( 8.7)
Where
𝑋𝑋�𝑗𝑗 𝐾𝐾𝐾𝐾𝑝𝑝𝐾𝐾𝐾𝐾𝑖𝑖𝐾𝐾𝑖𝑖𝑊𝑊𝑖𝑖 𝑆𝑆𝑆𝑆𝑆𝑆𝑝𝑝𝑆𝑆𝐾𝐾 𝑆𝑆𝐾𝐾𝑆𝑆𝑖𝑖 𝑓𝑓𝑊𝑊𝐾𝐾 𝑊𝑊𝐾𝐾𝐾𝐾𝑆𝑆𝑊𝑊𝑆𝑆𝐾𝐾𝑖𝑖𝑊𝑊 𝑗𝑗
𝑋𝑋𝑖𝑖𝑗𝑗 𝐾𝐾𝐾𝐾𝑝𝑝𝐾𝐾𝐾𝐾𝑖𝑖𝐾𝐾𝑖𝑖𝑊𝑊𝑖𝑖 𝑉𝑉𝑆𝑆𝑆𝑆𝑜𝑜𝐾𝐾 𝑊𝑊𝑓𝑓 𝑊𝑊𝑏𝑏𝑖𝑖𝐾𝐾𝐾𝐾𝑜𝑜𝑆𝑆𝑊𝑊𝑖𝑖𝑊𝑊𝑖𝑖 𝑓𝑓𝑊𝑊𝐾𝐾 𝑊𝑊𝐾𝐾𝐾𝐾𝑆𝑆𝑊𝑊𝑆𝑆𝐾𝐾𝑖𝑖𝑊𝑊 𝑗𝑗
𝑖𝑖𝑗𝑗 𝐾𝐾𝐾𝐾𝑝𝑝𝐾𝐾𝐾𝐾𝑖𝑖𝐾𝐾𝑖𝑖𝑊𝑊𝑖𝑖 𝑖𝑖𝑜𝑜𝑆𝑆𝑏𝑏𝐾𝐾𝐾𝐾 𝑊𝑊𝑓𝑓 𝑊𝑊𝑏𝑏𝑖𝑖𝐾𝐾𝐾𝐾𝑜𝑜𝑆𝑆𝑊𝑊𝑖𝑖𝑊𝑊𝑖𝑖 𝑓𝑓𝑊𝑊𝐾𝐾 𝑊𝑊𝐾𝐾𝐾𝐾𝑆𝑆𝑊𝑊𝑆𝑆𝐾𝐾𝑖𝑖𝑊𝑊 𝑗𝑗
𝑖𝑖𝑗𝑗2 𝐾𝐾𝐾𝐾𝑝𝑝𝐾𝐾𝐾𝐾𝑖𝑖𝐾𝐾𝑖𝑖𝑊𝑊𝑖𝑖 𝑖𝑖𝑆𝑆𝑆𝑆𝑝𝑝𝑆𝑆𝐾𝐾 𝑜𝑜𝑆𝑆𝐾𝐾𝑖𝑖𝑆𝑆𝑖𝑖𝑝𝑝𝐾𝐾 𝑓𝑓𝑊𝑊𝐾𝐾 𝑊𝑊𝐾𝐾𝐾𝐾𝑆𝑆𝑊𝑊𝑆𝑆𝐾𝐾𝑖𝑖𝑊𝑊 𝑗𝑗
𝑋𝑋� 𝐾𝐾𝐾𝐾𝑝𝑝𝐾𝐾𝐾𝐾𝑖𝑖𝐾𝐾𝑖𝑖𝑊𝑊𝑖𝑖 𝑂𝑂𝑜𝑜𝐾𝐾𝐾𝐾𝑆𝑆𝑆𝑆𝑆𝑆 𝑖𝑖𝑆𝑆𝑆𝑆𝑝𝑝𝑆𝑆𝐾𝐾 𝑆𝑆𝐾𝐾𝑆𝑆𝑖𝑖/𝐴𝐴𝑜𝑜𝐾𝐾𝐾𝐾𝑆𝑆𝐶𝐶𝐾𝐾
𝑆𝑆𝑆𝑆𝑇𝑇𝑀𝑀 𝐾𝐾𝐾𝐾𝑝𝑝𝐾𝐾𝐾𝐾𝑖𝑖𝐾𝐾𝑖𝑖𝑊𝑊𝑖𝑖 𝑆𝑆𝑜𝑜𝑆𝑆 𝑊𝑊𝑓𝑓 𝑖𝑖𝑠𝑠𝑜𝑜𝑆𝑆𝐾𝐾𝐾𝐾 𝑊𝑊𝑜𝑜𝐾𝐾 𝑊𝑊𝑊𝑊 𝑊𝑊𝐾𝐾𝐾𝐾𝑆𝑆𝑊𝑊𝑆𝑆𝐾𝐾𝑖𝑖𝑊𝑊
𝑀𝑀𝑆𝑆𝑇𝑇𝑀𝑀 𝐾𝐾𝐾𝐾𝑝𝑝𝐾𝐾𝐾𝐾𝑖𝑖𝐾𝐾𝑖𝑖𝑊𝑊𝑖𝑖 𝑀𝑀𝐾𝐾𝑆𝑆𝑖𝑖 𝑖𝑖𝑠𝑠𝑜𝑜𝑆𝑆𝐾𝐾𝐾𝐾 𝑊𝑊𝑜𝑜𝐾𝐾 𝑊𝑊𝑊𝑊 𝑊𝑊𝐾𝐾𝐾𝐾𝑆𝑆𝑊𝑊𝑆𝑆𝐾𝐾𝑖𝑖𝑊𝑊
𝑆𝑆𝑆𝑆𝑆𝑆 𝐾𝐾𝐾𝐾𝑝𝑝𝐾𝐾𝐾𝐾𝑖𝑖𝐾𝐾𝑖𝑖𝑊𝑊𝑖𝑖 𝑆𝑆𝑜𝑜𝑆𝑆 𝑊𝑊𝑓𝑓 𝑖𝑖𝑠𝑠𝑜𝑜𝑆𝑆𝐾𝐾𝐾𝐾 𝑊𝑊𝑜𝑜𝐾𝐾 𝑊𝑊𝑊𝑊 𝐾𝐾𝐾𝐾𝐾𝐾𝑊𝑊𝐾𝐾
𝑀𝑀𝑆𝑆𝑆𝑆 𝐾𝐾𝐾𝐾𝑝𝑝𝐾𝐾𝐾𝐾𝑖𝑖𝐾𝐾𝑖𝑖𝑊𝑊𝑖𝑖 𝑀𝑀𝐾𝐾𝑆𝑆𝑖𝑖 𝑖𝑖𝑠𝑠𝑜𝑜𝑆𝑆𝐾𝐾𝐾𝐾 𝑊𝑊𝑜𝑜𝐾𝐾 𝑊𝑊𝑊𝑊 𝐾𝐾𝐾𝐾𝐾𝐾𝑊𝑊𝐾𝐾
𝐹𝐹 𝐾𝐾𝐾𝐾𝑝𝑝𝐾𝐾𝐾𝐾𝑖𝑖𝐾𝐾𝑖𝑖ts 𝐹𝐹 𝑜𝑜𝑆𝑆𝑆𝑆𝑜𝑜𝐾𝐾 𝑓𝑓𝑊𝑊𝐾𝐾 𝐹𝐹 𝑊𝑊𝐾𝐾𝑖𝑖𝑊𝑊
Table 8-5: ANOVA for Laser Power
Source of Variation SS df MS F P-value F critical
Between Groups 0.226 2 0.113 11.568 0.009 5.143
Within Groups 0.059 6 0.010
Total 0.285 8
The Table 8-3, Table 8-4 and Table 8-5 shows the analysis between Laser Power and Kerf
Width Mean as a sample. The F critical value shows that Laser Power is significantly
participating in the variation in kerf width quality.
With reference to Table 8-5 and Table 8-6 input controllable parameter Laser power’s P and
F values reject null hypothesis Ho. Similarly, Appendix A (Table A-10 to Table A-21) shows
the analysis between Laser power, cutting speed, assist gas pressure and standoff distance
108
with kerf width mean one by one. Therefore population means of different groups are not
equal. The Table 8-6, P and F values accept Ho in case of cutting speed, assist gas pressure
and standoff distance means they are insignificantly participating in the variation of kerf
width quality.The one way ANOVA without replication analysis shows that better kerf width
predicted inputs are Laser power (100 watt), Cutting speed (0.7 m/s), Assist gas pressure (2.5
bar) and Standoff distance (5 mm).
Table 8-6: One way ANOVA without replication
Treatments F P-value F critical
Laser Power 11.568 0.009 5.143 Cutting Speed 0.007 0.993 5.143 Assist gas pressure 0.0997 0.907 5.143 Standoff distance 0.620 0.569 5.143
8.2.2 ONE WAY ANOVA WITH REPLICATION With reference to Appendix A (Table A-22, Table A-23, Table A-24) and Table 8-7 Laser
power is significant. Table 8-7 shows that P and F values accept Ho in case of cutting speed,
assist gas pressure and standoff distance are insignificantly participating in the variation of
kerf width quality but F value is more than 1, also indicating the standoff distance
significance as compared to cutting speed and assist gas pressure. The results are improved
with replication. The analysis show that kerf width predicted the same input data set and
similar prediction on the variance and mean point of view but in case of without replication
variance results are not acceptable.
Similarly, Appendix A (Table A-25 to Table A-34) show the analysis between cutting speed, assist gas pressure and standoff distance with kerf width mean one by one.
Table 8-7: One way ANOVA with replication
Treatments F P-value F critical
Laser Power 16.066 0.000 3.403 Cutting Speed 0.021 0.979 3.403 Assist gas pressure 0.285 0.755 3.403 Standoff distance 1.691 0.206 3.403
109
The analysis of individual input parameters on Kerf Width Mean were done based on the data
recorded in Appendix A (Table A-3). Formulas (equation (8.1) to (8.7)) were used for one
way ANOVA with or without replication analysis such as Table 8-3 to Table 8-21.
In one way analysis of variance results with replication and without replication support the
benefit of replication. The results of one way ANOVA show that the use of replication
improves the F value. It means replication improves the ability of bifurcation between
controllable and uncontrollable variations.
8.2.3 TWO WAY ANOVA WITH REPLICATION In Table 8-8 to Table 8-13 the analysis shows the means and variance due to interaction
between the two parameters ignoring other variables for each Laser Power and at cutting
Speed of 0.2, 0.7 and 1.2 in separate tables.
Table 8-8: Interaction between Laser Power and Cutting Speed with replication
A/B 100 300 500
0.2 1.580 1.940 2.010
1.615 1.885 2.040
1.525 1.915 1.855
0.7 1.730 1.765 1.975
1.255 1.775 1.940
1.480 1.985 2.290
1.2 1.660 1.655 1.79
1.695 1.715 1.89
1.860 1.855 2.08
Table 8-9: Interaction between Laser Power and Cutting speed 0.2
100/0.2 300/0.2 500/0.2 Total
Count 3 3 3 9
Sum 4.720 5.740 5.905 16.365
Average 1.573 1.913 1.968 1.818
Variance 0.002 0.001 0.010 0.037
T
110
able 8-10: Interaction between Laser Power and Cutting speed 0.7
100/0.7 300/0.7 500/0.7 Total
Count 3 3 3 9
Sum 4.465 5.525 6.205 16.195
Average 1.488 1.842 2.068 1.799
Variance 0.056 0.015 0.037 0.091
Table 8-11: Interaction between Laser Power and Cutting speed 1.2
100/1.2 300/1.2 500/1.2 Total
Count 3 3 3 9
Sum 5.215 5.225 5.760 16.200
Average 1.738 1.742 1.920 1.800
Variance 0.011 0.011 0.022 0.019
Table 8-12: Total Interaction between Laser Power and Cutting speed
Count 9 9 9
Sum 14.400 16.490 17.870
Average 1.600 1.832 1.986
Variance 0.030 0.012 0.021
The reason to perform the analysis is to understand how the Kerf Width behaves when
subjected to combination of parameters. With reference to Table 8-13,”Sample” is used for
Cutting Speed. The F value is less than the F critical value and P value is not less than 0.05,
which indicates that the there is no significant difference between Cutting Speed’s Means,
concluding that it will not play an important role in the variation of the Kerf Width.
“Columns” mean Laser power. The F value is greater than F critical value and P value is less
than 0.05 indicating that there is significant difference between Laser Power Means
concluding that it will play an important role in the variation of the Kerf Width.
“Interaction” stands for effect of Laser Power and Cutting Speed on Kerf Width. The F value
is not greater than F critical value and P value is not less than 0.05 showing that there is no
significant difference between means which is shown in Table 8-13.
111
Table 8-13: ANOVA of Interaction between Laser Power and Cutting speed
Source of Variation SS df MS F P-value F critical
Sample 0.002 2 0.001 0.057 0.945 3.555
Columns 0.678 2 0.339 18.457 4.37x10-5 3.555
Interaction 0.174 4 0.043 2.365 0.092 2.928
Within 0.331 18 0.018
Total 1.185 26
The summary of interaction result of six combinations of two way ANOVA with replication
is listed in Table 8-14. Similarly, detail analysis is indicated in Appendix A (Table A-35 to
Table A-70). The Table 8-14 shows interaction significance in the variation of the Kerf
Width.
Table 8-14: Summary of Two ANOVA with replication
S. No. Treatments F P-value F critical
1 Laser Power & Cutting speed 2.365 0.092 2.928 2 Laser Power & A. gas pressure 2.020 0.135 2.928 3 Laser power and standoff distance 0.402 0.805 2.928 4 Assist gas pressure and Cutting speed 11.220 9x10-5 2.928 5 Cutting speed and Standoff distance 9.602 0.0002 2.928 6 Assist gas pressure and standoff distance 9.257 0.0003 2.928
1. Interaction between laser power and cutting speed is insignificantly participating in
the variation of kerf width.
2. Interaction between laser power and assist gas pressure is insignificantly participating
in the variation of kerf width.
3. Interaction between laser power and standoff distance is insignificantly participating
in the variation of kerf width.
4. Interaction between assist gas pressure and cutting speed is significantly participating
in the variation of kerf width.
5. Interaction between cutting speed and standoff distance is significantly participating
in the variation of kerf width.
6. Interaction between assist gas pressure and standoff distance is significantly
participating in the variation of kerf width.
112
The interaction 4, 5 and 6 are significant but had been ignored at the time and added in
pooled error. For better optimization considering all the above interaction in [2, 3, 16] will
give better results or apply soft-computing techniques. It is not a good practice to assume
without analysis that input parameters are independent from each other. In most of the
experimental investigations of Laser beam cutting, one factor at a time has been varied to
analyze the effect of input process parameters on output quality characteristics or responses
[13, 16, 18, 20-22] which shows inferior analysis results.
8.3 REGRESSION ANALYSIS In Appendix A (Table A-3), the relationship can be studied by collecting the experimental
data and afterwards drawing scatter plot. The scatter plot shows the nature of relation
between the variables. They may be positive linear, negative linear or curvilinear
relationships. Then calculate coefficient of correlation “r” to see the significance of
correlation between the variables. If the value of r is not significant then predicting dependent
parameter values by regression equation is a useless practice. If r is significant then determine
regression equation by least square method as shown in equation (8.8).
𝑌𝑌𝑖𝑖 = 𝑏𝑏1𝑋𝑋𝑖𝑖 + 𝑏𝑏0 (8.8)
Where
b1 is slope of the regression line or coefficient of independent parameter
b0 is y intercept
The regression analysis shows the relationship between dependent and independent
parameters and predicts unknown values as and when required to reduce the time and cost of
experimentation and design.
8.3.1 LINEAR REGRESSION ANALYSIS Figure 8.3 shows the scatter plot between Kerf Width Mean and Laser power. It shows the
positive linear regression relationship, coefficient of correlation r is positive, +0.885 is
significant between laser power and kerf width. It also shows the curve fitting and 95%
confidence interval. R2 also shows 78.3% variation i.e. significant. F and P values shows
significant role. In linear regression coefficient of line were calculated.
113
Figure 8-3: Interactive graph of Laser Power and Kerf Width
In Table 8-15 description of the dependent (Kerf width) and independent (Laser power)
variables are calculating Mean and Standard deviation by SPSS Statistics Software. The units
of Laser power is Watts and Kerf width mean is mm.
Table 8-15: Descriptive Statistics of Laser power and Kerf width
Parameters Mean Std. Deviation N
Kerf Width Mean (dependent) 1.806 0.189 9
Laser Power (Independent) 300.000 173.205 9
In Table 8-16, Pearson Correlation coefficient r is shown and its significance is calculated by
SPSS using the formula in the following equation (8.9).
Table 8-16: Correlation between Laser power and Kerf width
Kerf Width Laser Power
Pearson Correlation Kerf Width Mean 1.000 0.885
Laser Power 0.885 1.000
Sig. (1-tailed) Kerf Width Mean
Laser Power
. .001
.001 .
N Kerf Width Mean 9 9
Laser Power 9 9
114
r =
∑ XY − ∑ Xni=1 ∑ YXn
i=1N
ni=1
��∑ X2ni=1 −
(∑ X2ni=1 )
N � �∑ Y2ni=1 −
(∑ Y2ni=1 )
N �
(8.9)
Where
r is the Coefficient Correlation and is the correlation between the measured and
predicted values of dependent parameter.
X is independent/input variable
Y is dependent/output variable
n is number of observations
Table 8-17: Regression Statistics
R Square 0.783
Adjusted R Square 0.752
Standard Error 0.094
Observations 9
In Table 8-17 R2, adjusted R2 and standard error have been shown to explain the goodness of
fit. R2 is the variance of the dependent parameter that can be explicated by the input
parameters. It measures the degree of association between input and output parameters. It
can also be explained by the following equation (8.10).
𝑀𝑀2 =𝑆𝑆𝑆𝑆𝑀𝑀𝑆𝑆𝑆𝑆𝑇𝑇
(8.10)
Where
R2 is Coefficient of Determination
SSR is Sum of Square due to variation in Regression
SST is Sum of Square due to Total variation including error / undefined.
Adjusted R2 explains the variation in the dependent by independent parameter. In formula of
Adjusted R2 account for the number of independent parameters and sample size. Therefore,
adjusted R2 is generally considered to be a more accurate goodness-of-fit measure than R2.
Mathematically shown below:
𝐴𝐴𝑊𝑊𝑗𝑗𝑜𝑜𝑖𝑖𝑊𝑊𝐾𝐾𝑊𝑊 𝑀𝑀2 = 1 − (1 − 𝑀𝑀2) �
(𝑖𝑖 − 1)(𝑖𝑖 − 𝑝𝑝 − 1)�
( 8.11)
Where
115
Adjusted R2 is Adjusted R Square
n is Sample size
p is no of independent parameters in the model
The standard error is similar to root mean square error for sampling distribution. In Table
8-17 standard deviation of dependent variable shows that change in input setting by Laser
Power causes change in dependent variable. Therefore there is a need to find, through
correlation and regression analysis, the significance of change. The calculated results of
coefficient of correlation in Table 8-16 shows the positive and significant relation in single
tailed test basis. The results show that laser power effect on Kerf Width is significant even
when the value of alpha is 1%.
In Table 8-17 R2 value shows that the variation in Kerf width is 78.3% due to Laser Power
showing that it is a highly significant parameter in controlling the quality compared to other
parameters. The adjusted R2 is smaller than R2 due to small sample size. In Table 8-17
analysis of variance results also shows the significant role of Laser power by F value and
significance of F.
Table 8-18: Regression between Laser power and Kerf width ANOVA
d.f. SS MS F Significance F
Regression 1 0.223 0.223 25.280 0.0015
Residual 7 0.062 0.009
Total 8 0.285
In Table 8-18 analysis of variance is performed. Regression, Residual, Total are the breakup
of variance in output quality, The “Total” variance is divided into two parts, variance which
can be associated or related to independent parameter like in regression and the variance
which is not associated with independent parameter like in error. Sum of Squares is related to
the three Total, Model and Residual sources of variance. The Total variance is split in the
variance clarified by independent variables i.e. Regression and the variance clarified by
independent variables i.e. Residual. Degree of freedom is related with the causes of
variance. The Total variance has N-1, Regression number of coefficients (including
intercept) anticipated minus 1 and the Error total minus the degree of freedom of the model.
Mean Square is explained as the Sum of Squares divided by their degree of freedom. The F-
116
test is the Mean Square Regression divided by the Mean Square error as explained earlier in
equation ( 8.7). The p-value is compared to tolerance value alpha which is by default 0.05
used to test the null hypothesis. The coefficients or intercept mean square is 0.
Table 8-19: Linear regression line of Laser Power
Coefficients Standard Error t Stat P-value Lower 95% Upper 95%
Intercept 1.517 0.065 23.163 7.09 x10-8 1.362 1.672
A 0.001 0.000 5.028 1.52 x10-3 0.001 0.001
In Table 8-19 coefficient 1.517 means the y intercept of a least square line and 0.001 is a
coefficient of dependent variable A, which is a slope of a line. Laser Power (A) is the
independent parameter in this analysis using equation (8.8) of linear regression. The standard
errors are
𝜎𝜎𝑏𝑏0 = �1𝑁𝑁 +
𝑌𝑌�2
∑ (𝑌𝑌𝑖𝑖 − 𝑌𝑌�)ni=1
2
𝜎𝜎𝜀𝜀
( 8.12)
𝜎𝜎𝑏𝑏1 = �𝑌𝑌�2
∑ (𝑌𝑌𝑖𝑖 − 𝑌𝑌�)ni=0
2
𝜎𝜎𝜀𝜀
( 8.13)
Where 𝜎𝜎𝜀𝜀 = 𝑆𝑆𝑆𝑆𝑆𝑆𝑁𝑁−2
SSE is sum of squared residual
is the ith value of output paramaeters
Y is output parameter ith value
Y� is output parameters mean value
𝜎𝜎𝑏𝑏0 , 𝜎𝜎𝑏𝑏1 are standard error of coefficients
𝑊𝑊 =
𝑝𝑝𝑊𝑊𝐾𝐾𝑓𝑓𝑓𝑓𝑖𝑖𝑝𝑝𝑖𝑖𝐾𝐾𝑖𝑖𝑊𝑊𝑆𝑆𝑊𝑊𝑆𝑆𝑖𝑖𝑊𝑊𝑆𝑆𝐾𝐾𝑊𝑊 𝐾𝐾𝐾𝐾𝐾𝐾𝑊𝑊𝐾𝐾
( 8.14)
The value of t statistic is measured by equation 8.14. The value of t statistic gives the p- value
i.e. probability value in t-distribution. If the value of p is less than tolerance (alpha) value
then null hypothesis is rejected i.e. significant difference caused by the independent
parameter on dependent variable. In default case of 95% confidence lower and upper limit of
the y intercept and slope of line is mentioned in the last two columns. It is related to p value
and shows the tolerance space in numerical and on graph by placing two lines under and over
117
the fit line as shown in Figure 8.3. In Table 8-19 y intercept and slope of the line are
calculated. T test value and p value rejected the null hypothesis that Mean values are not
equal. The Laser Power causes significant variation in Kerf width.
Table 8-20: Residual output
Observation Predicted Kerf width Mean Residuals
1 1.613 -0.040
2 1.613 -0.125
3 1.613 0.125
4 1.806 0.107
5 1.806 0.036
6 1.806 -0.064
7 1.999 -0.030
8 1.999 0.070
9 1.999 -0.079
In Table 8-20 and Figure 8.4 predicted Kerf Width is measured by regression line fitting
equation (8.9) and residual is measured from difference between experimental value and the
predicted value. The standard residuals are measured by experimental value minus predicted.
The regression line model is based on laser power and predicted minimum, maximum, and
average percent errors are 5.17%, 21.55% and 12.95% respectively. The values show that the
average error is considerably high i.e. more than 10%.
Figure 8-4: Laser Power (A) Line fit Plot
0.000
0.500
1.000
1.500
2.000
2.500
100 100 100 300 300 300 500 500 500
Expe
rim
enta
l and
Pred
icte
d Ke
rf W
idth
Laser Power
Kerf Width
Predicted Kerf Width
118
Figure 8-5: Interactive graph of Cutting Speed and Kerf Width
Figure 8.5 shows r is -0.041 which is insignificant between cutting speed and kerf width. R2
also shows 0.2% variation i.e. insignificant. F and P values shows insignificant role. In linear
regression coefficient of line were calculated. The detailed analysis are attached in Figure 8.6
and Appendix A (Table A-77 to Table A-82).
Figure 8-6: Cutting speed (B) Line fit Plot
0.000
0.500
1.000
1.500
2.000
2.500
0.2 0.7 1.2 0.2 0.7 1.2 0.2 0.7 1.2
Expe
rim
enta
l and
Pred
icte
d Ke
rf W
idth
Cutting Speed
Kerf Width
Predicted Kerf Width
119
Figure 8-7: Interactive graph of Assist Gas Pressure and Kerf Width
Scatter plot in Figure 8.7 shows r is positive 0.026 and is insignificant between assist gas
pressure and kerf width. R2 also shows 1.6% variation i.e. insignificant. F and P values also
shows insignificant role. In linear regression coefficient of line were calculated. The detailed
analysis are attached in Figure 8.8 and Appendix A (Table A-83 to Table A-88).
Figure 8-8: Assist gas Pressure (C) Line Fit Plot
0
0.5
1
1.5
2
2.5
0.5 2.5 4.5 2.5 4.5 0.5 4.5 0.5 2.5
Expe
rim
enta
l and
Pred
icte
d Ke
rf W
idth
Assist Gas Pressure
TPM
Predicted TPM
120
Figure 8-9: Interactive graph of Standoff Distance and Kerf Width
Scatter plot in Figure 8.9 shows r is positive 0.312 and is insignificant between standoff
distance and kerf width. R2 shows that the 9.8% variation in kerf width due to standoff
distance is significant than two insignificant factors. F value shows that it is insignificant and
P value of linear regression coefficient of line shows insignificance. The detailed analysis are
attached in Figure 8.10 and Appendix A (Table A-89 to Table A-94).
Figure 8-10: Standoff Distance (D) Line Fit Plot
0
0.5
1
1.5
2
2.5
1 5 10 10 1 5 5 10 1
Expe
rim
enta
l and
Pred
icte
d Ke
rf W
idth
Standoff Distance
Kerf Width
Predicted Kerf Width
121
Table 8-21: Linear Regression ANOVA
Laser
Power
Cutting
Speed
Assist gas
pressure
Standoff
Distance
R +0.885 -0.041 +0.126 0.312
Significant r Significant Insignificant Insignificant Insignificant
F 25.280 0.012 0.113 0.758
Sig. F 0.0015 0.914 0.746 0.413
Intercept 1.517 1.819 1.772 1.725
Coefficient X 0.001 -0.018 0.014 0.015
P-value 1x10-3 0.0914 0.746 0.413
Min error 5.17% 6.2% 1.38% 6.38%
Max. error 21.55% 54.83% 54.83% 53.97%
Average error 12.95% 25.77% 25.07% 25.94%
The Table 8-21 shows that laser power causes significant variation in dependent variable
(Kerf width mean). But cutting speed, assist gas pressure, standoff distance causes
insignificant role in the variation of dependent variable. The maximum error model based on
significant factor is 21.55% which is just above acceptable limits but insignificant factor
results more than 50% which is completely unacceptable.
8.3.2 MULTIPLE LINEAR REGRESSION
Figure 8-11: Comprehensive Interactive graph between all parameters
Scatter plots are drawn in Figure 8.11 to observe relationship between the parameters. It
shows the linear regression relationship drawn 95% confidence interval lines below and
122
above if it is possible with the current scale of the plot. The data points lie in between them.
The relation between Kerf width and input parameters are laser power positive, cutting speed
slightly negative, assist gas pressure slightly positive and standoff distance is positive.
Table 8-22: Descriptive Statistics
Mean Std. Deviation N
Kerf Width Mean 1.806 0.1887 9
Laser Power 300.00 173.205 9
Cutting Speed 0.700 0.4330 9
Assist Gas pressure 2.500 1.7321 9
Stand of distance 5.33 3.905 9
Table 8-23: Correlations
Kerf Width Mean
Laser Power
Cutting Speed
Assist Gas pressure
Standoff Distance
Pearson Correlation Kerf Width Mean 1.000 0.885 -0.041 0.126 0.312
Laser Power 0.885 1.000 0.000 0.000 0.000 Cutting Speed -0.041 0.000 1.000 0.000 0.000 Assist Gas pressure 0.126 0.000 0.000 1.000 0.000 Stand of distance 0.312 0.000 0.000 0.000 1.000
Kerf Width Mean
Laser Power
Cutting Speed
Assist Gas pressure
Standoff Distance
Sig. (1-tailed) Kerf Width Mean . 0.001 .458 0.373 0.207 Laser Power 0.001 . .500 0.500 0.500 Cutting Speed 0.458 0.500 . 0.500 0.500 Assist Gas pressure 0.373 0.500 0.500 . 0.500 Stand of distance 0.207 0.500 0.500 0.500 . N Kerf Width Mean 9 9 9 9 9 Laser Power 9 9 9 9 9 Cutting Speed 9 9 9 9 9 Assist Gas pressure 9 9 9 9 9 Stand of distance 9 9 9 9 9
Table 8-24: Regression Statistics
The coefficient of correlation r is 0.948 and significant on single tailed test. R2 value shows
that the variation in kerf width is 89.9% it can be explained by independent parameter which
Multiple R 0.948
R Square 0.899
Adjusted R Square 0.797
Standard Error 0.085
Observations 9.000
123
is highly significant in controlling the quality as compared to the variation due to
uncontrollable parameters.
Table 8-25: Multiple linear Regression ANOVA
D.o.f SS MS F Significance F
Regression 4 0.256 0.064 8.858 0.0288
Residual 4 0.029 0.007
Total 8 0.285
Table 8-26: Linear regression of multivariable
Coefficients Standard Error t Stat P-value Lower 95%
Intercept 1.4147 0.097 14.566 0.0001 1.1450
( A ) 0.0010 0.000 5.557 0.0051 0.0005
( B ) -0.0183 0.069 -0.264 0.8047 -0.2110
( C ) 0.0138 0.017 0.793 0.4723 -0.0344
( D ) 0.0151 0.008 1.963 0.1212 -0.0063
The mathematical expression of multiple linear regression as shown in equation (8.15). 𝑌𝑌𝑖𝑖 = 𝑏𝑏1𝑋𝑋1𝑖𝑖 + 𝑏𝑏2𝑋𝑋2𝑖𝑖 + 𝑏𝑏1𝑋𝑋3𝑖𝑖 + 𝑏𝑏1𝑋𝑋4𝑖𝑖 + 𝑏𝑏0 (8.15)
In Table 8-27 Laser Power is a significant parameter in the variation of kerf width and at a
low level so is standoff distance. The value of R2 encourages to use the regression modelling
techniques but their maximum residual and average residual errors do not allow to
recommend the method to be used as shown in Appendix A (Table A-76, Table A-82, Table
A-88 and Table A-94).
Table 8-27: Summary of linear Regression
The overall coefficient of correlation is improved to positive 0.984 and is significant. R2
encourages using the model because kerf width is 89.9% due to independent parameters
R2 F sig. P , t test Max. error Average error Remarks
A 78.3% 0.001 0.001 21.55% 12.95% Significant
B 0.2% 0.914 0.091 54.83% 25.77% Insignificant
C 1.6% 0.74 0.746 54.83% 25.07% Insignificant
D 9.8% 0.412 0.413 53.97% 25.94% Insignificant
124
which is a highly significant parameter in controlling the quality as compared to the
variation due to uncontrollable parameters.
Analysis of variance results show the significant role of independent controllable parameters
by F-value, t-test and p-value. Kerf width residual predicted value shows that the minimum,
maximum, and average percent errors are 0.71%, 20.66%, 7.99% respectively. The values
show that the error is considerably low i.e. around 8%. Therefore, this model can be used to
see an approximate trend of idea dependent variable, hence can be used as an empirical
formula. However, the results are good without missing value data.
Table 8-28: Residual Output
Observation Predicted Kerf width Mean Residuals
1 1.5294 0.0440
2 1.6081 -0.1198
3 1.7019 0.0364
4 1.8856 0.0278
5 1.7680 0.0737
6 1.7642 -0.0226
7 2.0303 -0.0620
8 2.0417 0.0267
9 1.9241 -0.0041
8.3.3 NONLINEAR REGRESSION ANALYSIS
Figure 8-12: Quadratic graph of Laser Power without replication
y = -1E-06x2 + 0.001x + 1.454R² = 0.794
0.000
0.500
1.000
1.500
2.000
2.500
0 100 200 300 400 500 600
Kerf
Wid
th M
ean
Laser Power
Kerf Width
Poly. (Kerf Width)
125
Regression analysis between Laser power and Kerf width without replication is considered.
For the analysis in the beginning Scatter plot is drawn in Figure 8.12 showing the nonlinear
regression quadratic equation without replication. The coefficient of correlation r is 0.891
(positive), significant and R2 value shows that the variation in Kerf width is 79.4% and F-
value due to Laser Power that shows that it is a highly significant parameter in controlling the
quality compared to other parameters.
Table 8-29: Regression data without replication for Laser Power
S. No. Laser Power A A2 Kerf Width Mean
1 100 10000 1.573
2 100 10000 1.488
3 100 10000 1.738
4 300 90000 1.913
5 300 90000 1.842
6 300 90000 1.742
7 500 250000 1.968
8 500 250000 2.068
9 500 250000 1.920
Table 8-30: Regression Statistics
Multiple R 0.891
R Square 0.794
Adjusted R Square 0.725
Standard Error 0.099
Observations 9
Table 8-31: Nonlinear Regression ANOVA for Laser Power and KW without replication
d.f. SS MS F Significance F
Regression 2 0.2261 0.1130 11.5684 0.0087
Residual 6 0.0586 0.0098
Total 8 0.2847
Y intercept and equation shown on Figure 8.12 is based on nonlinear regression of laser
power. T-test and p-value accepted null hypothesis. The Laser Power causes no significant
126
variation in Kerf width in non- linear model i.e. is unable to model the problem. The
prediction error is more than 50%.
Table 8-32: Nonlinear regression of Laser Power
Coefficients Standard Error t Stat P-value
Intercept 1.454 0.130 11.154 3.10E-05
A 0.002 0.001 1.457 0.195
A2 -9x10-7 1x10-6 -0.564 0.593
Scatter plots of laser cutting, cutting speed, assist pressure and standoff distance with
replication shown in Appendix A (Table A-100 to Table A-119, Figure A-14, Figure A-15,
Figure A-16 and Figure A-17) and detail parameters such as values of r, R2, F-value, T-test
and p-value are shown in Table 8-33. This information shows that laser power is a significant
parameter based on r, R2 and F-values. But, finally T-test and p-values show that laser power
is an insignificant parameter and unable to model the problem. The prediction error is more
than 50%. The cutting speed, assist gas pressure and standoff distance are insignificantly
participating in the variation of kerf width based on r, R2, F-value, T- test and p-values. Kerf
Width residual values in predicted value tables shows that the minimum, maximum, and
average percent errors are 1.628%, 23.85%, 11.94% respectively. The values shows that the
error is high i.e. more than 10%.
Table 8-33: Non-linear Regression ANOVA
A Without
Rep.
A With
Rep.
B With
Rep.
C With
Rep.
D With
Rep.
R +0.891 +0.757 +0.042 +0.152 +0.351
Sig. r Significant Significant Insignificant Insignificant Insignificant
F 11.568 16.066 0.021 0.285 1.691
Sig. F 0.0087 3x 10-5 0.979 0.755 0.206
T test X 1.457 1.7169 -0.140 -0.403 -0.865
P-value 0.195 0.0989 0.889 0.691 0.395
T test X2 -0.564 -0.665 0.107 0.536 1.206
P-value 0.593 0.5124 0.915 0.597 0.240
127
The Null hypothesis is rejected in Linear Regression, ANOVA analysis of Laser Power effects
on Kerf Width and in multivariable case also rejects the null hypothesis. The data is
nonlinear but regression accepts the null hypothesis due to data points’ pattern and error
squared in X2 values.
The value of curve fitting parameters like coefficient of correlation and R2 are better in
nonlinear case but null hypothesis is accepted which is not desirable in this case. The residual
values are also not better than the single and multi linear regression.
Table 8-34: Summary of Non- linear Regression
Input variable % R2 F sig. P of t test H0
Without replication
Laser cutting 79.4 0.009 0.195 Accepted
With replication
Laser cutting 57.2 3x10-5 0.989 Accepted
Cutting speed 0.2 0.979 0.889 Accepted
Assist gas Pressure 2.3 0.755 0.691 Accepted
Standoff distance 12.4 0.206 0.395 Accepted
The results of nonlinear regression show that Laser power is the most important parameter.
Its R2 value decreases with the replication. The pattern of the data is above or below the least
square point. In case of without replication the model touches only one point out of 3 while in
case of replication it only touches one point out of 9 points. The error is increased due to
replication and it will remain increasing if more observations are considered as in factorial
design [47]. The null hypothesis H0 is accepted for all variables which show that nonlinear
regression with or without replication cannot explain the variations in dependent variables.
The residual error due to predicted values increased in nonlinear case rather than in linear
prediction. N. Yusoff et al. [10] explain many nonlinear relations by using only one
independent and one dependent variable and kept others constant but in our case three other
independent variables are changing along with the considered variable. Therefore, this
modelling technique cannot be recommended on the basis of the above Table 8-34 results.
The detailed analysis is shown in Appendix A (Table A-95 to Table A-119).
128
8.3.4 MULTIPLE NON-LINEAR REGRESSION Regression analysis between inputs with kerf width is explained. r is positive 0.849 which
shows it is significant and R2 shows that the variation in kerf width is 72.1% due to
independent parameters which is a sufficiently significant parameter in controlling the quality
compared to the variation due to unknown variables. Analysis of variance shows the
significant role of independent controllable parameters by F value.
Table 8-35: Regression Statistics
Multiple R 0.849 R Square 0.721 Adjusted R Square 0.597 Standard Error 0.136 Observations 27
Nonlinear regression of multivariable calculates the coefficient of quadratic equations. The
hypothesis by T test value and p value of A, B, C and D and their square values are accepted
i.e. the independent parameters cause insignificant variation in kerf width.
Table 8-36: Multiple Non-linear Regression ANOVA
d.f. SS MS F Significance F
Regression 8 0.854 0.107 5.811 0.0010
Residual 18 0.331 0.018
Total 26 1.185
Table 8-37: Nonlinear Regression of multivariable
Coefficients Standard Error t Stat P-value Lower 95% Upper 95%
Intercept 1.498 0.160 9.357 2.46 x10-8 1.162 1.834
A 0.002 0.001 1.840 0.082 0.000 0.003
A2 -9.86 x10-7 1.38 x10-6 -0.713 0.485 -3.89 x10-6 1.92 x10-6
B -0.073 0.316 -0.230 0.821 -0.738 0.592
B2 0.039 0.221 0.176 0.863 -0.426 0.504
C -0.046 0.071 -0.652 0.522 -0.195 0.103
C2 0.012 0.014 0.868 0.397 -0.017 0.041
D -0.042 0.032 -1.328 0.201 -0.109 0.025
D2 0.005 0.003 1.850 0.081 -0.001 0.011
129
Kerf Width predicted residual shows that the minimum, maximum, and average percent
errors are 1.82%, 59.48%, 55.81% respectively. The values show that the error is more than
50%. Detail analysis is attached with Appendix A (Table A-120 to Table A-124).
Table 8-38: Residual Output
Observation
Predicted KERF WIDTH MEAN
Residuals
1 1.573 0.007 2 1.573 0.042 3 1.573 -0.048 4 1.488 0.242 5 1.488 -0.233 6 1.488 -0.008 7 1.738 -0.078 8 1.738 -0.043 9 1.738 0.122 10 1.913 0.027 11 1.913 -0.028 12 1.913 0.002 13 1.842 -0.077 14 1.842 -0.067 15 1.842 0.143 16 1.742 -0.087 17 1.742 -0.027 18 1.742 0.113 19 1.968 0.042 20 1.968 0.072 21 1.968 -0.113 22 2.068 -0.093 23 2.068 -0.128 24 2.068 0.222 25 1.920 -0.130 26 1.920 -0.030 27 1.920 0.160
8.4 SUPERVISED LEARNING WITH MISSING VALUE Statistical work shows that the modelling is possible with it. However, the modelling with
missing value DOE based data analysis is challenging. Normally, there are methods to
130
complete the missing table with interpolation, experienced guess etc. In laser cutting process
during the DOE based experiment sometimes the metallic or nor metallic sheets are cut
incompletely [25]. Therefore, the value of kerf width or edge quality measurement is not
possible. The observation data was unbalanced. In case of unbalanced data matrix the
statistical analysis is very difficult and the results are awkward. The ability of recognition of
patterns with incomplete information motivated to utilize the ANN to handle missing values.
Often larger size datasets rather than simple orthogonal array are used by many researchers
without the missing values.
Orthogonal array based experimental data were trained by Artificial Neural Network (ANN)
for the modelling of laser cutting process of Perspex glass sheet. The simulation results were
compared with factorial design experimental observations. The factorial design is selected for
estimation of error on each dataset of observation table with missing values. The ANN
simulation evaluation results are expected to depict better generalization based on very small
training datasets by applying feed forward back-propagation. The network parameters
variation shows the benefits and disadvantages on neural network model. The increment in
size of training datasets based on factorial design generates better accuracy in simulations as
compared to orthogonal array modelling. ANN modelling possibly will be utilized by the new
researcher without much trouble in case of missing values in observation tables.
The problem solution methodology is shown in Figure 6.7. M. Zaidi et al. [4] modeled laser
cutting of Perspex sheet by Artificial neural network with supervised learning. The model
requires preparation of training, test and Simulation datasets for training and simulation
purpose.
Training, testing and simulation on experimental data are assumed as a process. There are a
number of factors which affect the output quality of the model. This process, can be
represented by the model proposed in [73]. Constant factors are selected in the beginning for
the sake of simplification of modelling on the basis of theoretical and literature review such
as:
• Network type (feed forward back propagation)
• Training Algorithm (Levenberg Marquardt)
• Performance (Mean Square Error)
131
• Transfer function (Hidden layers): Tangent Sigmoid
• Transfer function (Output layers): Pure Linear
• The process of training is supervised learning
• Regression coefficient > 0.99
8.4.1 EDGE QUALITY OF PERSPEX SHEET Levenberg Marquardt LM algorithm was selected for training and explained under the
heading 4.3 to 4.6 in detail and an independent study was also performed by [4, 12, 24] which
shows LM training algorithm is potentially better than other back-propagation algorithms.
This training was performed with two output parameters (mean and signal to noise ratio with
three replications) of single quality (edge quality) on orthogonal array consisting of only nine
observations with missing value. The statistical modelling utilizes all data for modelling. But,
MATLAB’s Neural Networks Toolbox by default utilizes 60% data for training, 20% for
validation and 20% for ultimate testing of generalization. The limitation of orthogonal dataset
is that only five datasets were used for training, two for validation and two for testing. As
discussed earlier, Perspex sheet 3mm was cut by CO2 laser cutting machine in factorial
design to provide sufficient data to verify simulation design based on orthogonal array. For
the purpose of error estimation many training sessions were carried out but some selected
sessions results were recorded in Table 8-39. The average percent errors are 6.8% and 7.8%
(mean and signal to noise ratio), enrichment technique is applied for improvement with
duplicating the datasets for training of models [2, 3, 95].
Table 8-39: Preliminary training by Levenberg Marquardt
S. No. No. of
Datasets
No. of
neurons
Max.
% error
Min.
% error
Average
% error Results
1 1 10 114.85 3.610-10 41.263 Training
33.505 5.64E-12 7.578
2 1 20-20 98.282 1.086 26.744 Training
64.143 0.867 15.631
3 1 9-9 14.642 0.006 6.844 Training
23.184 0.005 7.807
132
In the beginning a model for training is prepared for both the output (mean and signal to
noise ratio) of edge quality of 3 millimeter thick Perspex sheet based on orthogonal array.
With reference to Table 8-39 (serial number 1) the input layer with four input, hidden layer
with 10 neurons and output layer with two output, the training results show very high level of
error on average 41% and 7% (edge quality mean and signal to noise ratio respectively)
however the regression coefficient and performance of three training, validation and test data
sets are sufficiently better. The single hidden layer is a universal approximate [95, 116]. It
was the better result among the other network settings of different number of neurons in
hidden layer.
With reference to serial number 2 the number of layers is three with two hidden layers with
20-20 neurons in each hidden layer. The training was worse and error on average was 26%
and 15% (edge quality mean and signal to noise ratio respectively). In the process of training
the weights are initialized many times to get the global minima. Similarly, with reference to
serial number 3 only number of neurons in hidden layer was changed from 20-20 to 9-9. Now
the results are comparatively better even though the results are not good for simulation. The
training error on average is 6% and 7% (edge quality mean and signal to noise ratio
respectively) and corresponding maximum percent error in the training process is 14% and
23%, which is higher than expected.
It is concluded from Table 8-39 that the weights are initialized many times to get the global
minima. The model consists of four inputs and two outputs. With reference to serial number 3,
number of neurons in two hidden layer was 9-9. Now the results are comparatively better
even though the results are not good for simulation. The training error on average is 6% and
7% (edge quality mean and signal to noise ratio respectively) and corresponding maximum
percent error in the training process is 14% and 23%, which is higher than expected which
clearly shows the disadvantage of small dataset available for training and so there is less
generalization.
With reference to Table 8-40 the enriched datasets provide improvement in training up to
4.7% and 3%. Hence, enrichment improves the training. This technique is also used in
training models of Melamine and Polystyrene foam materials and similarly in current
experiment of Perspex sheet cutting. However, the average simulation percent error of mean
and S/N ratio is 19% and 24% which is not up to international standards. The basic reason is
133
very small datasets with the presence of missing values and the data is multi-variable
nonlinear. The whole model is based on five datasets, which is not sufficient to predict the
population data.
For the purpose of improvement enrichment technique with two times all datasets are used in
training models of Melamine and Polystyrene foam materials[2, 3] and similarly in current
experiment of Perspex sheet cutting. With reference to Table 8-40, serial number 1 the
number of layers are three i.e. two hidden layers with 9-9 neurons and training results show
improvement by the enrichment of data. The error on average is 4% and 3% (edge quality
mean and signal to noise ratio respectively) with the regression coefficient and performance
of three training, validation and test data sets are sufficiently better. The simulation based on
factorial datasets show that the average error is 19%, 24% for mean and signal to noise ratio
respectively and the corresponding maximum errors are at very high level i.e. 83%, 109% in
comparison with the experimental results. The training and simulation model of edge quality
signal to noise ratio is better than edge quality mean based on average error because of its
larger range of values trained. One of the major deficiencies in the model is the size of the
dataset.
Table 8-40: Enrichment Training by Levenberg Marquardt
S.
No.
No. of
neurons
Max.
% error
Min.
% error
Average
% error
Results
1 9-9 14.643 0.006 4.78 Training
12.685 0.002 3.169
83.741 0.0047 19.33 Simulation
109.517 0.0020 24.03
2 9-9 2.027 0.077 0.59 Training
1.105 0.045 0.37
108.744 0.000 23.79 Simulation
104.671 0.000 25.54
3 9-9 2.966 0.033 1.07 Training
1.105 0.045 0.51
77.256 0.517 18.28 Simulation
74.196 0.065 18.80
134
With reference to serial number 2 after extensive search of minimum error by initializing
weight of network many times, the training results are improved and error on average is
0.59% and 0.36% (edge quality mean and signal to noise ratio respectively). The simulation
of factorial input table results show the average errors are 23%, 25% and corresponding
maximum errors are at a higher level i.e. 108%, 104% in comparison with the experimental
values. Overtraining reduces the generalization in model.
Similarly, with reference to serial number 3 try to find better training and simulation, the
average training error are 1% and 0.5% (edge quality mean and signal to noise ratio
respectively) and corresponding maximum error is 1.9% and 0.5% which are better results.
The simulation based on factorial datasets show the average errors are 18% and 18%. The
corresponding maximum errors are at a higher level i.e. 77%, 74% in comparison with the
experimental values. The results show better generalization in model. Simulation with two
hidden layers improved.
Similarly, with reference to Table 8-40, serial number 1, 2 and 3 try to find better training
and simulation model. The average training error are 1% and 0.5% (edge quality mean and
signal to noise ratio respectively) and corresponding maximum error is 1.9% and 0.5% which
are better results. The simulation based on factorial datasets show better generalization in this
model and reaches to average errors 18%, 18%. The corresponding maximum errors are at a
higher level i.e. 77%, 74% in comparison with the experimental values. The simulation with
two hidden layers improved.
Table 8-41: EQ Mean and S/N Ratio with Factorial Design
S.
No.
No. of
Neurons
Max.%
error
Min.%
error
Average
% error Results Rem.
1
20.20.20
12.37 0.0003 0.663 Train
Factorial
design
2.2 0.0001 0.186
12.37 0.0003 14.135 Test
2.2 0.0001 9.115
81.00 11.000 46.000 Simulation
0.29 0.000 0.123
135
Simulation is improved in orthogonal design training model but not up to the desired level.
The results show that the size of training model of orthogonal is not sufficient for Perspex
sheet process cutting model. In MATLAB 2010 the training is performed on five datasets,
two datasets are used for validation and two for test datasets. It shows that the whole model is
based on the five datasets which are not sufficient to predict the population. It is decided on
the basis of above trainings that training will be performed on 74 datasets with 7 data sets
used exclusively for testing the model and then performing the factorial simulation. The
results were observed on the basis of training, testing and simulation data.
With reference to Table 8-41, serial number 1 the number of hidden layers increases to three
with 20-20-20 neurons. The factorial design training results are improved and error on
average is 0.66% and 0.18% (edge quality mean and signal to noise ratio respectively). After
training 7 Test datasets are applied for prediction on the given model and compared with the
experimental results. The errors on average are 12%, 2% and maximum errors are level i.e.
14%, 9%. The test results are encouraging. The simulation results show that the average
errors are 46%, 0.12% (edge quality mean and signal to noise ratio respectively) and
corresponding maximum errors are at a higher level i.e. 81%, 0.29%, in comparison with the
experimental results the training of edge quality mean is clearly inferior than signal to noise
ratio of edge quality. The results can be improved by changing number of neurons. With
reference to Table 8-41(EQ Mean and S/N Ratio with FD) and Appendix B (Table B-1), the
simulation results of 30-30-30 neurons in each layer and model is better by decreasing the
number of neurons to 20-20-20 which shows excellent results for the signal to noise ratio of
edge quality i.e. 0.29, upon further decreasing the number of neurons to 10-10-10 and 5-5-5
degrades the results. Hence, better results are 20-20-20 and 30-30-30 neurons as compared to
other models in this table. The major problem to resolve the issue of modelling is edge
quality mean. It is also better to reduce the number of three hidden layers to two or one. The
results also show that modelling both output parameters separately may provide better results.
The models were made for outputs edge quality mean values and signal to noise ratios with
controllable input variables. The results seem suitable but the maximum value in terms of
error percent is much higher than expected. The idea to improve the modelling and
simulation single output is mapped with input variables i.e. Edge quality mean [105].
136
Figure 8-13: Comparison of edge quality mean networks errors
In the beginning a model was prepared for edge quality of 3 mm thick Perspex sheet using
Factorial design. With reference to Figure 8.13 and Appendix B (Table B-2 serial number 1,
2, 3, 9 and 10 the number of layers are three i.e. two hidden layer with number of neurons 5-
5, 10-10, 30-30, 40-40 and 50-50. The training results show percent error on average is high
with 5-5 neurons, higher in 10-10, similarly increasing with 30-30 and improved with 40-40
neurons and further decreased with 50-50 neurons also shown in Figure 8.13 serial 10. With
reference to serial number 4, 5, 6, 7 and 8 the number of layers are four i.e. three hidden
layers. The training and simulation results show percent error on average low with 10-10-10
number of neurons in each hidden layer, similarly increasing with 20-20-20 and 30-30-30
even with 40-40-40 neurons and again decrease with 50-50-50 neurons.
The results demonstrate that minimum training error does not guaranty the best
generalization. If early stopping of training is avoided the percent errors in training reduces
but the simulation results will be inferior due to over training and reduction in the
generalization. The results of “Test errors” show the possible best generalization. Therefore,
serial number 1 is the best possible percent average errors of training, test and simulation of
mean edge quality are 7.69%, 9.87% and 17.47% respectively. The maximum percent errors
of training, test and simulation are 28.76%, 25.5 and 53.41 respectively. However, 10-10
neurons in hidden layer simulation errors are improved even Test errors are inferior than 5-5
neuron. Therefore, better model is probably the Test error based. But, the simulation results
0
10
20
30
40
50
60
Ave
rage
% e
rror
Neurons in hidden layers
Training
Testing
Simulation
137
are 14.3% i.e. little high due to missing values. Hence, this model needs to be improved. It is
better to normalize the data for spreading the range of edge quality mean data.
Figure 8-14: Edge quality mean comparison in percent error for normalized datasets
The models were made for both mean values and signal to noise ratios. The results of signal
to noise ratios are already better, but mean data training results can be improved by
normalization to stretch the range of mean data. The results are looking suitable even at
maximum error.
Normalization is applied on the training data of edge quality mean and found better results
with 10-10-10 neurons in Figure 8.14 and Appendix B (Table B-4). The dataset is same as
earlier i.e. factorial design. The average simulation errors are encouraging. The early stopping
of training provides better generalization models. Therefore, the training and testing on
average percent errors are higher than previous case but simulation results are better and
average errors is 6.9% which is definitely better, without normalizing the dataset the error
was 14.3%. Hence, this training technique improved results by normalization (stretching data
sets) of edge quality mean data. The ANN modelling shows small error with hidden layers
10-10-10 and 20-20 neurons. The training, testing and simulations error are 2%, 9% and 3%,
but maximum possible individual error is 25% due to the fact of missing values. The results
are better for modelling. The models were made for both mean values and signal to noise
ratios. The results of signal to noise ratios are already better, but mean data training results
improved by normalization. The results are looking suitable even at maximum error.
0
5
10
15
20
25
30
10-10 20-20 30-30 10-10-10 20-20-20 30-30-30 40-40-40
Ave
rage
% e
rror
Neurons in hidden layers
Training
Testing
Simulation
138
Figure 8-15: Average percent errors comparison for edge quality signal to noise ratio
Now study the signal to noise ratio model training of edge quality, of 3mm thickness Perspex
sheet on Factorial design array. With reference to Figure 8.15 and Appendix B (Table B-3)
the best training datasets of are 20 and 40-40 neurons in hidden layer in Figure 8.15. The
training dataset size is same. Average simulation errors are low. The early stopping of
training provides better generalization model therefore the training and testing on average
percent errors are higher than previous case but results are better and simulation on average
errors is 3.3% which is also better. Hence, this technique is sufficient for the training.
Modelling of signal to noise ratio of edge quality of Perspex sheet using Factorial design
array produces better results with 10, 20 and 6 neurons with single hidden layer. The
training results of hidden layer containing 20 neuron average errors are lower than 10 and 6
neurons. The average percent errors of training, testing and simulations are 6%, 3%, and 6%
but maximum possible individual error is 25%. The results are better for modelling. The 40-
40 neurons in hidden layer average percent errors of training, testing and simulations are
2%, 9%, and 3% and maximum possible individual error is 25%. The results are better with
20 and 40-40 neurons in hidden layers as shown in Figure 8.15.
8.4.2 KERF WIDTH QUALITY OF PERSPEX SHEET The models were made for both mean values and signal to noise ratios of dependent variables
kerf width. The results of signal to noise ratios are already better, but mean data training
results can be checked without normalization. The results are looking suitable even at
maximum error. With reference to Figure 8.16 and Appendix B (Table B-5) serial number 1
0
5
10
15
20
25
10 20 6 10-10 20-20 40-40 20-20-20
Ave
rage
% e
rror
Neurons in hidden layers
Training
Testing
Simulation
139
and 2 the number of layers is two i.e. one hidden layer with 10 and 20 neurons. The training
results show percent error on average are better. With reference to Figure 8.16 and Appendix
B (Table B-5), serial number 3, 4 and 5 the number of layers is three i.e. two hidden layer
with 10-10, 20-20 and 30-30 neurons. The training results show percent error on average has
not improved. Therefore, further training is performed and with reference to serial number 1
and 2 of Figure 8.16 and Appendix B (Table B-5) the number of layers is four i.e. three
hidden layer with 10-10-10 and 20-20-20 neurons. The training results show percent error on
average are improved in 10-10-10 neurons. The better average percent errors of training,
testing and simulations are 3%, 16%, and 4% and maximum possible individual error is 35%.
The results are better for modeling. Therefore, in the above table results of serial 6 modelling
results are better than others as shown in Figure 8.16 even with 20-20-20 neurons.
Figure 8-16: Comparison of Kerf width mean factorial datasets
The best neural network model contains 20, 10-10 and 10-10-10 neurons in hidden layers as
shown in Figure 8.16. The average simulation errors are 3.84, 3.13 and 4.39. The results
show that on average simulation errors are acceptable, however better models can be
prepared by single, double and triple hidden layers. Hence the idea of single hidden layer
network architecture is successful as many researchers mentioned that single layer hidden
layer architecture is capable to approximate any arbitrary function [91]. It can be improved
by normalization. Hence, this training technique is sufficient for training and improvement,
which is possible to obtain by normalization of kerf width mean. The kerf width signal-to-
noise ratio range is sufficiently large. Therefore, its training is not an issue and can be
performed easily.
0
5
10
15
20
25
30
35
40
10 20 10-10 20-20 30-30 10-10-10 20-20-20
Ave
rage
% e
rror
Neurons in hidden layers
Training
Testing
Simulation
140
The different studies [117, 118] of machining with different methods are utilizing ANN since
a long time for the relationship between the two variables which is very simple and between
multi-input/output problems. However, they do not always define each and every aspect of
distribution of datasets for training, validation and testing. The calculation of signal to noise
ratio will be able to provide robust modelling solution. In case of training based on
orthogonal design L9 or fractional factorial or major part of the factorial design dataset, if
the prediction are made on full factorial dataset then the chance of maximum error in all
prediction raises because missing values are around 11% in OA and 1.3% in factorial
design. In this study, we have learned from OA and predicted the full factorial design values.
The average error is 19 to 24% which is reasonably close due to these reasons. The actual
size of OA used for training is 5 which should be 9 for better ANN mapping. However, any
missing value will increase smaller error on average and larger near the missing input
dataset. Prediction is not tested only on the known input with missing values in the
observation tables.
Statistical modelling techniques are applied by many researchers and nowadays soft-
computing techniques are more popular. The statistical concept of orthogonal array and
factorial design is not ideal for the artificial neural network case. However, it is suited for
Statistical modelling to predict better values which is not perfect. As sampling data is never
equivalent to population data and also conversion of fraction data into three level ordinal
data loses data knowledge for perfect modelling. The issue of sampling size can be improved
by selecting FD instead of OA.
The ANN technique is better than regression model. The results of ANN with factorial
designs are better than other researchers. They also compare the results with multiple linear
regressions and concluded that ANN is a better solution [4, 78]. M.-J. Tsai et al. [34] also
built both artificial and multi-regression model for QFN cutting of 6 qualities and 3 input
parameters. The results of LM-back-propagation neural network were better than multi-
regression because of the selection of full factorial design. The regression model just fits the
curve on the experimental data by least square method and then predicts data based on the
fitted model which can be over fitted and losses generalization. Sometimes the results of ANN
modelling are not of high-quality. As the ANN is similar to regression but regression utilizes
the whole data for curve fitting and ANN uses 60% data for training and rest for validation
141
and testing. For medium size datasets ANN results are better but for small datasets like OA
the rest of the data, after reserving for validation and testing, is insufficient for training.
Therefore, on the small datasets like OA it is recommended to utilize nine datasets for
training and other four data sets for validation and testing so they may be added for ANN
modelling to generate even better results than statistical techniques in special case of small
datasets. This modelling work shows the effect of early stopping, enrichment techniques,
normalization, change in number of hidden layers, change in number of neurons, size of
datasets, training algorithm and effect of missing values in the training datasets, which is
very difficult to handle with statistics and has not been addressed in earlier studies of ANN
due to unease with Statistics. It is also important to mention that design of experiment for
ANN is different than the routine statistical modelling tools because the extra datasets need
to be noted for validation, testing as well as for prediction on unseen examples. For extra
datasets there is no need to experiment more because the data observed during range
adjustment may be noted and utilized for this purpose.
8.5 SEMI-SUPERVISED ALGORITHM There are different methods for resolving the matter of missing value. Few of the methods are
mentioned here of which some are modern and some traditional [119]. The missing values
can be generated due to some of following reasons:
• Uncontrollable factors / Random Error
• Human errors
• Forget to note down
• Range adjustment
The range adjustment of laser cutting is difficult for a new researcher. Applying DOE based
machine adjustment and work-piece is unable to cut every set of inputs, then these values are
considered as missing values. The Statistical traditional methods try to measure the missing
values however errors are higher due to covariance. Sometimes Statisticians analyze data
without using missing values if the data is more related with less variance and is a larger size
datasets. In our view for small data sets like L9 it is difficult to make models with missing
values. Some of the methods are [119]:
• Deletion List Wise
• Deletion Pair-Wise
• Weighted Arithmetic
142
• Normal Ratio Method
• Inverse Distance Method
Modern methods are built for more accuracy and some of these methods are [119]:
• Time Series Analysis and Regression
• ANN
• Stochastic Techniques of Interpolation
• Maximum Likelihood
• Multiple Imputations
The results of regression are quite inferior to ANN that are mentioned in [2-4, 25, 34, 78].
Many researchers handled missing value but did not apply ANN in the Laser cutting process
to solve the range adjustment problem. The modern techniques of regression [25] had higher
prediction errors. Therefore, it is better to use novel method of Semi-supervised algorithm for
process modelling.
For the purpose of simplification, some parameters were kept constant based on experience
of supervised learning work [4]. The best possible training algorithms were discussed under
the heading 4.1, 4.2, 4.3 and 4.4 in detail and Levenberg Marquardt algorithm and Gradient
Descent with Momentum were selected. An independent study was performed by M. Zaidiet
al. [4, 24] to select the training algorithm. This training was performed with two output
parameters (mean and signal to noise ratio with three replications) of single quality (edge
quality) on orthogonal array consisting of nine observations preprocessed data with missing
values. The training set needed validation and test datasets. Therefore, nine datasets were
used for training, three for validation and three for testing.
M. Zaidi [4] discussed some recorded results of edge quality mean and signal to noise ratio
modelling for the purpose of modelling and verification. Different training sessions were run
with supervised learning mode and measure of average percent errors in normal training were
6.8% and 7.8%, which improves by applying enrichment technique (duplicating the dataset)
on training data (Melamine and Polystyrene foam materials) [2, 95]. It improves up to 4.7%
and 3%, but simulation average percent errors are 19% and 24% [4], which indicate higher
error level because of small datasets, nonlinear multi-variables and insufficient number of re-
143
initializations of neural networks weights, as it is a human limitation because of the training
being supervised. The previous research has increased the size of dataset from orthogonal to
factorial design for the training and testing which improves the results but the option of re-
initialization had not been tested by the author.
Figure 8-17: Variations in 100, 500, 1000, 3000 re-initializations
Therefore, there is a need to increase the number of re-initialization and necessitate a semi-
supervised algorithm to decrease the human load. ANN Training and simulations were
performed by fixing number of neurons to 4 and changing the number of initializations in the
algorithm and have observed the standard deviation of average percent error on factorial
input predictions in Figure 8.17 which is probabilistic in nature. Hence shows that there is a
chance of improvement by changing the number of re-initialization from 100 to 500, 1000
and 3000 times. The small dataset OA is utilized for all models of edge quality/kerf widths of
3mm or 5mm Perspex sheets. Increase in number of re-initializations of neural network
weights will generate better results as shown in Figure 8.18. One of the major lead of Semi-
supervised on supervised learning is a huge number of re-initializations of the network
weights such as 100 or 1000 or 3000 times but in other cases it is 5 to 10 times.
Therefore, it can be concluded that Semi Supervised learning is required to produce better
models. Because in supervised learning it is difficult to do this job even for 100 times. The
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
EQ Mean 3mm
EQ S/N 3mm
KW Mean 3mm
KW S/N
3mm
EQ Mean 5mm
EQ S/N 5mm
KW Mean 5mm
KW S/N
5mm
Mea
n /
Stan
dard
dev
iati
on
Standard deviation
144
use of higher speed processors as Intel Core I3 based systems to perform the Semi-supervised
study to observe clearly produces noteworthy improvements.
Figure 8-18: Effect of weight initializations on average error
As in supervised learning the edge quality model of 3mm were studied and the average
percent error were 19% (mean) and 24% (S/N) but in case of Semi-supervised they were not
more than 9.5% and 11%. These results encourage utilizing the Semi-supervised algorithm
technique based on supervised learning experiences. The error at most is 12.0 with 22%
missing values in 5mm models as shown in Figure 8.18. The average percent error in case of
edge quality and kerf width for 3mm and 5mm was calculated by changing number of re-
initializations of weights of the network. The mean and standard deviation was calculated for
each case. The results are depicted in Figure 8.17 and Figure 8.18 respectively which shows
that this parameter can be utilized for the improvement of neural network modelling and is
also capable of modelling small datasets such as orthogonal array.
In the [4] previous study only 3mm material was used but in this study an additional 5mm
work piece is also used for experimentation, training & modelling. The number of missing
values doubled with the same input parameters ranges. But, the modelling error increased no
more than 2.5%. The modelling of missing data encourages the cutting of new materials with
small experiment sets even when missing values exist where more time, cost and domain
expert are needed to adjust the range to perform experiment without missing values. The
4
5
6
7
8
9
10
11
12
13
14
100 500 1000 3000
Ave
rage
% e
rror
Number of times intialized
EQ 3mm Mean
EQ 3mm S/N
KW 3mm Mean
KW 3mm S/N
EQ 5mm Mean
EQ 5mm S/N
KW 5mm Mean
KW 5mm S/N
145
results of single output quality were better than multiple outputs [4]. Figure 8.18 shows that
3000 time re-initialization will generate better results.
The modelling is improved by running the program for GDM and LM training algorithms.
Different learning rates (0.01 0.1 0.2) and Momentum (0.01 0.1 0.2 0.3) values are the
manual input requirement of GDM. Both the training algorithms are run repeatedly for each
number of neuron.
8.5.1 EDGE QUALITY OF PERSPEX SHEET Edge quality EQ mean with 3mm Perspex sheet average percent error (9.9%) is minimum
with 5 neurons with the selection of LM algorithm as shown in Figure 8.19. Similarly, with
5mm Perspex sheet average percent error (11.5%) is minimum with 2 neurons with the
selection of LM algorithm as shown in Figure 8.19. The increase in error is 1.6% due to the
increase in number of missing values from 11% to 22% in training data. The error reduced
by increasing the number of re-initialization of Appendix B (Figure B-1 and Figure B-2)
showing that 12000 time re-initialization will generate better results than 3000 times. After
3000 times re-initialization the process was repeated for 12000 times and improvement was
observed in edge quality mean value of 3 and 5mm. Therefore, the result shown in Figure
8.19 was based on 12000 time re-initialization.
Figure 8-19: Neuron‘s variations in EQ mean
8.0
9.0
10.0
11.0
12.0
13.0
14.0
15.0
16.0
17.0
2 3 4 5 6 7 8 9 10
Ave
rage
per
cent
err
or
Number of neurons
EQ 3mm mean
EQ 5mm mean
146
Edge quality EQ S/N with 3mm Perspex sheet average percent error (9.1%) is minimum with
5 neurons with the selection of LM algorithm as shown in Figure 8.20. Similarly, with 5mm
Perspex sheet average percent error (10.8%) is minimum with 2 neurons with the selection of
LM algorithm as shown in Figure 8.20.
The increase in error is 1.7% due to the increase in number of missing values from 11% to
22% in training data. Appendix B (Figure B-3 and Figure B-4) shows edge quality S/N values
of 3 and 5mm sheet cutting process model 12000 time re-initialization and is able to generate
better results than 3000 times. After 3000 times re-initialization the process was repeated for
12000 times and the improvement was observed in edge quality S/N value of 3 and 5mm.
Therefore, the result shown in Figure 8.20 was based on 12000 time re-initialization.
Figure 8-20: Neuron's variations in EQ S/N
8.5.2 KW QUALITY OF PERSPEX SHEET Kerf width KW mean with 3mm Perspex sheet average percent error (4.7%) is minimum with
8 neurons with the selection of GDM algorithm as shown in Figure 8.21.
The minimum error of KW Mean with LM was 5.0%, which is only 0.3 percent higher. The
difference is not remarkable and can be omitted or reduced by more re-initializations while
saving training time in hours. Similarly, with 5mm Perspex sheet average percent error
8.0
9.0
10.0
11.0
12.0
13.0
14.0
15.0
2 3 4 5 6 7 8 9 10
Ave
rage
per
cent
err
or
Number of neurons
EQ 3mm S/N
EQ 5mm S/N
147
(8.0%) is minimum with 3 neurons with the selection of LM algorithm. The increase in error
is 3.3% due to the increase in number of missing values from 11% to 22% in training data.
Figure 8-21: Neuron‘s variations in KW mean
Edge quality KW S/N with 3mm Perspex sheet average percent error (5.3%) is minimum
with 5 neurons with the selection of LM algorithm as shown in Figure 8.22. Similarly, with
5mm Perspex sheet average percent error (8.7%) is minimum with 3 neurons with the
selection of LM algorithm. When the missing values double it increases by 3.4% only.
Figure 8-22: Neuron's variations in KW S/N
4
5
6
7
8
9
10
2 3 4 5 6 7 8 9 10
Ave
rage
per
cent
err
or
Number of neurons
KW 3mm mean
KW 5mm mean
4.05.06.07.08.09.0
10.011.012.013.014.015.016.017.018.0
2 3 4 5 6 7 8 9 10 11
Ave
rage
per
cent
Number of Neuron
KW 3mm S/N
KW 5mm S/N
148
The different studies [73, 118] of machining with different methods have been utilizing ANN
since a long time for the relationship between two variables and between multi-input/output
problems. However, they do not always define each and every aspect of distribution of
datasets for training, generalization and testing. The optimized datasets are normally
predicted which are in case of Taguchi method, more close to the desired value and also
more robust in nature. Therefore, the predictions mostly produce better results.
However, in case of training based on orthogonal design L9, fractional factorial or major
part of the factorial design dataset, if the predictions are made on full factorial dataset then
the chance of maximum error in all prediction raises. In this study, we have learned from OA
and predicted the full factorial design values. The average percent error in supervised
learning and semi-supervised are listed in Table 8-42. The results show remarkable
improvement with 11% (3mm) and 22% (5mm) missing values yet error increases in small
amount. The error is calculated against the whole factorial inputs.
Table 8-42: Average percent errors
Supervised Semi Supervised
Edge quality Edge quality Kerf Width
Mean S/N Mean S/N Mean S/N
3mm 19% 24% 9.9% 9.1% 4.7% 5.3%
5mm - - 11.5% 10.8% 8.0% 8.7%
To understand the missing values, assume a collection of points, which are to be used for the
formation of a regression line. With more points the regression will have a higher probability
of being accurate. Similarly, in our case, data used for checking generalization and testing
gives smaller error on average but larger error near the missing points.
Statistical modelling techniques are applied by many researchers and nowadays soft-
computing techniques are used as well. The statistical concept of orthogonal array and
factorial design is not suitable for the artificial neural network case because by nature the
orthogonal and factorial design can only represent approximate population values. It loses
knowledge because of less number of samples for the representation of a population. In other
words it only represents the generalization aspect of a sample while for neural networks such
a sample can be used for training but as far as validation and testing is concerned, it results
149
in a regression model which has higher probability of having errors. Conversion of fractional
data into ordinal data also loses a lot of information that the artificial neural network can be
modeled with, using real non-reprocessed data. There is a similar loss of information in
factorial design due to usage of ordinal three level datasets. The factorial design is better
than orthogonal design because of its larger size than OA.
The ANN technique is better than regression model. The regression model just fits the curve
on the experimental data by least square method and then predicts data based on the fitted
model. The results of ANN are better because the training data works similar to regression,
but the results are degraded because of decrease in size of datasets due to need of datasets
for validation and testing. Using these extra datasets will improve generalization and
accuracy. Hence, provide better results than statistical modelling techniques.
8.6 FUZZY AGGREGATION Table 8-43: Four qualities factorial design data
Run A B C D Simulated results
EQ KW POC MRR
1 100 0.2 0.5 1 3.670 1.552 2.356 6.74E-08
2 100 0.2 0.5 5 2.729 1.842 3.089 6.98E-08
3 100 0.2 0.5 10 2.090 1.882 3.868 6.80E-08
4 100 0.2 2.5 1 2.833 1.440 2.413 6.31E-08
5 100 0.2 2.5 5 2.991 1.547 3.207 6.37E-08
. . . . . . . . .
. . . . . . . . .
75 500 1.2 0.5 10 2.305 1.962 4.168 3.24E-07
76 500 1.2 2.5 1 1.756 1.920 4.031 4.85E-07
77 500 1.2 2.5 5 2.383 1.935 4.118 4.85E-07
78 500 1.2 2.5 10 1.624 1.950 4.109 4.82E-07
79 500 1.2 4.5 1 0.625 1.977 3.842 4.85E-07
80 500 1.2 4.5 5 1.125 1.957 3.933 4.85E-07
81 500 1.2 4.5 10 1.688 1.892 4.049 4.82E-07
150
The data of Table 8-43 is taken from Polystyrene foam experimental data as shown in
Appendix B (Table B-6). It contains the unique key of serial number or run which can be
used as a substitute of four inputs (Laser power (A), cutting speed (B), assist gas pressure (C)
and standoff distanced (D)) and output datasets in further representation for ease.
For equal participation of all four quality factors values normalized between 0 to 1. For
simplicity, the fraction (continuous) data is converted to ordinal type data i.e. 1, 2 and 3 else
0 (Unsuccessful cutting) as shown in Table 8-44.
Table 8-44: Quality quantification
Value Quality Customer Range
1 Excellent 0<X≤0.25
2 Desired 0.25<X≤0. 5
3 Worst 0.5<X≤1
Often the industry sends the cutting quality specifications along with the acceptable
tolerances. Therefore, good or desired quality is also called as per specification of the
customer. Excellent means better then desired quality and worst means unacceptable for the
customer as shown in Table 8-44. The conversion of variable at this stage does not harm the
optimization objective. The aggregation was performed using the following formula between
the range 0 to 4 [2] .
Av = Nor(EQ) + Nor(KW) + Nor(POC) + (1 − Nor(MRR)) ( 8.16)
Where Av is Aggregated value, Nor(EQ), Nor(KW), Nor(POC) and Nor(MRR) are
normalized edge quality, kerf width, percent over cut and material removal rate respectively.
Nor(Av) = (V − Vmin col ) Vmax col ⁄ ( 8.17)
Where Nor(Av) is Normalized Aggregated value, V is current value, Vmin col is minimum
column value and Vmax col is maximum column value.
It will help to perform customer quality function in a more simplified way in equation (8.18)
than M. Zaidi et al. used in [2].
𝐶𝐶𝐶𝐶𝐹𝐹 = Max�VEQ or VKW or VPOC or VMRR � (8.18)
151
Where CQF is customer quality function, VEQ, VEQ, VEQ and VEQ are edge quality, kerf width,
percent over cut and material removal rate quantified ordinal values.
The customer quality function performs by crisp logic and aggregation function performed
with fuzzy logic. The existing problem is solved by Matlab’s FIS toolbox. The results are
shown in Table 8-45 and complete table is in Appendix B (Table B-7) which consists of
quantified normalized aggregated values (Q.Av), customer quality function (CQF),
defuzzified fuzzy aggregation value (FL Av) and quantified fuzzy aggregation value
(QFL Av).
Table 8-45: Compare results of Q Av, CQF, FL Av and QFL Av
Run Ordinal outputs Overall quality
EQ KW POC MRR Q AV CQF FL Av QFL Av
1 2 2 1 1 1 2 0.45 2
2 1 2 1 1 1 2 0.48 2
3 2 1 1 2 1 2 0.48 2
4 3 2 1 1 2 3 0.46 2
. . . . . . . . .
. . . . . . . . .
81 2 3 3 3 3 3 0.48 2
The overall quality was measured by the aggregation function of fuzzy logic. This addition
will prepare novel combination with crisp and simple data mining aggregation technique. The
normalized aggregation was performed with equation ( 8.16) after normalization of each
output parameter by equation (8.17). The simple aggregation was performed by equation
( 8.17). The results show that normalized aggregation value shows a better picture of overall
quality as compared to other two as shown in Appendix B (Table B-7). But CQF shows the
real picture of the overall quality as shown in Figure 8.23. However, quantified fuzzy
aggregation shows a more appropriate picture of the overall quality.
Normalized aggregation is unable to reflect worst quality any one of the aggregated quality
all the time. However, customer quality can do this by quantifying every quality before
aggregation and accept maximum value as CQF value, shown in Table 8-45 of Run 4 that
152
normalized aggregation value is 2 which shows that the overall quality is desirable but CQF
shows 3 i.e. worst quality, which protects the customer to accept overall desired quality
because one of the qualities is worst. It also secures the product quality inspection as shown
in Figure 8.24. This function is also based on crisp logic and it is completely in favor of the
customer rights and is damaging to the manufacturer.
Figure 8-23: View of all aggregated values
Figure 8-24: Comparison of quantified Normalized aggregation with CQF
0
0.5
1
1.5
2
2.5
3
3.5
1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67 70 73 76 79
Aggregation Customer quality function
Fuzzy aggregation Quantified Fuzzy Aggregation
1
2
3
4
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33
Aggregation
Customer quality function
153
The defuzzified fuzzy aggregation FL Av has a benefit of member ship function and applies
parallel rule which can protect the rights of customer and lower aggregation value shows that
minor repair can save a lot of revenue of the product cost which eventually saves the
customer’s cost and time, as shown in run 4 that edge quality is worst but by repairing it
product can be acceptable.
Instead of using the process of normalization and aggregation equation, fuzzy logic is directly
applied to output qualities. In Fuzzy logic there are many possible membership functions
other than normalized function. Therefore, the possibility to transform, stretch or compress
the simulated data is much higher. Fuzzy logic is applied using Matlab (2010a) software
utility called “Fuzzy Inference System”.
Figure 8-25: Fuzzy inference system for polystyrene sheet cutting process
154
The following steps have been taken
• Define input variables function and range. Four input simulated quality parameters
were created.
• Define output variables function and range. Fuzzy aggregation function was created.
• Link inputs and output with model name of Polystyrene sheet laser cutting process as
“Poly2” using Mamdani algorithm which consists of rules to be followed for
aggregation.
• The rules apply in parallel as shown in Figure 8.26.
• Implication method min is used which just truncated output on the basis of logic
operators on right most column and first four columns are inputs.
Figure 8-26: Application of parallel rules with implication
In this problem following membership functions are used
• “zmf” Opened from left side polynomial curve asymmetrical.
• “gaussmf” Gaussian membership function
155
• “sigmf” Sigmoid membership function
• In crisp logic consequent may be true or false based on Antecedent, but in fuzzy it
might be partially true or false. If-then rule in fuzzy solves first antecedent which
means fuzzifying the four inputs and applying fuzzy operators and assign results to
consequent.
• The results are aggregated and then defuzzified i.e. giving single number.
The fuzzy logic results are closer to the CQF but not its replacement. The fuzzy rules are
applied simultaneously and cater for the effect of participation of simulated results in fuzzy
manner, it shows the degree of Excellent, Good or Worst based on the inputs value, its
member ship function and applied rules. The results of FIS cannot be utilized in place of CQF
but its property of using any membership function and considering the degree of participation
in aggregation gives the idea to utilize this function for optimization and rework function
with CQF and normalized aggregation.
Table 8-46: Sorted with CQF, Quantified fuzzy aggregation and quantified normalized
aggregation
Run EQ KW POC MRR Agg. CQF FL AV QFL AV
1 2 2 1 1 1 2 0.45 2
2 1 2 1 1 1 2 0.48 2
3 2 1 1 2 1 2 0.48 2
10 2 2 1 2 2 2 0.44 2
8 2 2 1 1 2 2 0.45 2
6 2 2 2 2 2 2 0.48 2
7 1 3 1 2 2 3 0.36 2
14 2 3 1 2 2 3 0.38 2
18 1 3 2 3 2 3 0.55 3
64 3 2 3 3 3 3 0.82 3
65 1 3 3 3 3 3 0.83 3
73 3 3 3 2 3 3 0.86 3
156
For this purpose the result is sorted in novel order of Customer quality function, quantified
fuzzy aggregation and quantified normalized aggregation values. CQF has the highest
priority because if any one of the output quality is worst i.e. 3 it is not acceptable for any
customer. The quantified fuzzy aggregation has second priority because it aggregates the
value based on simple aggregation with selected membership functions and also with the
rules applied in CQF but in fuzzy manner. The normalized aggregation function helped to
achieve a more comparative list for better optimization. A part of this list is shown in Table
8-46 and complete table in Appendix B (Table B-8).
The sorted Table 8-46 shows overall input data setting of Run 1, 2 and 3. In addition to this
Run 7 and 14 can be utilized by rework because CQF value is 3 but QFL and Normalized
aggregated (AV) values are 2. These settings results show that the quantified value of Kerf
width is worst but other qualities are excellent and desired showing that overall quality is a
low degree of worst. However, run 18 and up to end are shown in Table 8-46 and the full
table in Appendix B (Table B-8) has QFL value at 3 which shows that the product quality is
worst to the high degree so it is better to reject rather than rework. This rework function can
guide us as to whether the product is suitable for rework and save the cost of the loss of
whole product. In many cases the result from FIS is desirable/good/2 but in CQF the result is
worst/useless/3. The overall quality fuzzy model can be improved by fine-tuning of
membership function’s boundaries, intersection points, width and shape. These parameters of
FIS system can be fine tuned by genetic algorithm.
In overall quality problem genetic algorithm is learned and studied for the implementation
point of view. A detailed study was performed for the implementation of genetic algorithm
[60, 68-70, 98, 99, 111] and it was concluded that for the following reasons GA is not
suitable.
• It is an experimental study based on OA without outer array.
• Factorial design dataset is filled with neural network modelling with quality
parameters consisting of three levels (1-3).
• GA is applied instead of experimentation and the initial population is generated on
random basis and selection is based on cost function. Therefore, transfer function is
necessary.
157
• Therefore, GA is not suitable for multi-objective overall quality calculation at this
stage.
The genetic algorithm can be applied in our problem but it is out of the scope at the moment.
It is possible to apply genetic algorithm by performing the following tasks.
• Select appropriate DOE for GA.
• Model the laser cutting process for all qualities.
• Use ANN trained model for simulation.
• Prepare cost function for minimum the better.
• Generate initial population of input datasets and predict quality parameter of the
applied input datasets.
• The quality functions values were applied in cost function and selected the population
based on cost function.
• In the end, there are a number of optimized datasets that will be gathered by the
utilization of genetic algorithm.
The overall quality was solved by combination of customer quality function, quantified fuzzy
aggregation and quantified normalized aggregation. Neural network helps in modelling of
laser cutting process. The process industry usually needs to improve multiple quality
parameters which are resolved by novel combination of Fuzzy aggregation. Consumption of
Electricity is also an important factor there is a pressing need to improve energy efficient
processes which lead to reduction in the manufacturing cost by increasing cutting speed and
decrease the utilization of laser power per meter cutting. Therefore, the study adds to the
electricity-efficient solution and shows the new quality parameter calculation which will
include in the overall quality as shown in Table 8-47 and possible energy levels in three
levels factorial design are 9 as shown in Table 8-48. With reference to Table 8-49 the results
are appropriate even for energy point of view based on the assumption that laser power is the
major contributor and it is more energy efficient if assumed that change in cutting speed
causes change in energy consumption. Therefore, the units are watts . m/hr.
Eltawahni et al. [101] looked into the CO2 laser cutting variables of MDF and explained the
cutting cost.
𝐶𝐶𝑜𝑜𝑊𝑊𝑊𝑊𝑖𝑖𝑖𝑖𝐶𝐶 𝑝𝑝𝑊𝑊𝑖𝑖𝑊𝑊 =
2.654 + 1.376 × 𝐴𝐴 + 1.3718 × 10−5 × 𝐹𝐹0.051 × 𝐵𝐵
( 8.19)
𝐹𝐹𝑆𝑆𝑊𝑊𝑤𝑤 𝐾𝐾𝑆𝑆𝑊𝑊𝐾𝐾 = 𝐹𝐹(1 ℎ⁄ ) = 492 × 𝑊𝑊2(C + 1) ( 8.20)
158
Where,
A is the power (KW)
F is flow rate (1/h)
B is cutting speed (mm/min)
d is nozzle diameter (mm)
C is assist gas pressure (bar).
Table 8-47: Energy consumption quality calculation for factorial design
Run Laser Power
Cutting Speed
A.Gas pressure
Standoff distance
Energy Consumed
1 100 0.2 0.5 1 8.33 . . . . . . 9 100 0.2 4.5 10 8.33 10 100 0.7 0.5 1 2.38 . . . . . .
18 100 0.7 4.5 10 2.38 19 100 1.2 0.5 1 1.39 . . . . . .
27 100 1.2 4.5 10 1.39 28 300 0.2 0.5 1 25.00 . . . . . .
36 300 0.2 4.5 10 25.00 37 300 0.7 0.5 1 7.14 . . . . . .
45 300 0.7 4.5 10 7.14 46 300 1.2 0.5 1 4.17 . . . . . .
54 300 1.2 4.5 10 4.17 55 500 0.2 0.5 1 41.67 . . . . . .
63 500 0.2 4.5 10 41.67 64 500 0.7 0.5 1 11.90 . . . . . .
72 500 0.7 4.5 10 11.90 73 500 1.2 0.5 1 6.94 . . . . . .
81 500 1.2 4.5 10 6.94
159
Table 8-48: Power levels corresponding to Laser power and cutting speed
Levels Power levels Laser Power Cutting speed
1 1.39 100 1.2
2 2.38 100 0.7
3 8.33 100 0.2
4 6.94 300 1.2
5 7.14 300 0.7
6 8.33 300 0.2
7 11.90 500 1.2
8 25.00 500 0.7
9 41.67 500 0.2
Table 8-49: Including power consumed and then sorted with CQF, Q AV and QFL AV Run A B C D EQ KW POC MRR Q AV CQF FL AV QFL AV Power
1 100 0.2 0.5 1 2 2 1 1 1 2 0.45 2 8.33
2 100 0.2 0.5 5 1 2 1 1 1 2 0.48 2 8.33
3 100 0.2 0.5 10 2 1 1 2 1 2 0.48 2 8.33
10 100 0.7 0.5 1 2 2 1 2 2 2 0.44 2 2.38
8 100 0.2 4.5 5 2 2 1 1 2 2 0.45 2 8.33
6 100 0.2 2.5 10 2 2 2 2 2 2 0.48 2 8.33
7 100 0.2 4.5 1 1 3 1 2 2 3 0.36 2 8.33
14 100 0.7 2.5 5 2 3 1 2 2 3 0.38 2 2.34
77 500 1.2 2.5 5 2 3 3 3 3 3 0.79 3 6.94
64 500 0.7 0.5 1 3 2 3 3 3 3 0.82 3 11.90
65 500 0.7 0.5 5 1 3 3 3 3 3 0.83 3 11.90
73 500 1.2 0.5 1 3 3 3 2 3 3 0.86 3 11.90
The above equations indicate that laser power (A) is the major contributor in cost. They also
identified that the power consumed by motion controller does not differ with different speed
[101]. It means difference in power consumption with 0.2 m/min or 0.7 m/min or 1.2 m/min
is negligible. So, assumed constant power consumed in motion controller. It means Flow rate
of assist gas pressure is also another parameter which affects the cost per unit length.
Reducing power and increasing speed results in the reduction of cost. Therefore, for the sake
160
of simplicity we calculate cost based on laser power only. As the cost of chiller electrical
power, compressor power, exhaust system power and manpower are constant. With reference
to Table 8-49 the results are appropriate even for cost point of view.
H. A. Eltawahni, et al [6, 101] applied limitation on input, output, overall quality and cost
using desirability approach. In overall quality they are unable to give single optimized results,
entrusting the operator to decide whether money is important or quality. They mentioned that
they were able to reduce the operating cost of cutting up to 71%. But need to compromise on
quality due to cost-effective solution. In our case electricity-effective solution was provided
without compromising on quality. The overall quality can be compromised in some other
cases as many researchers face in cost or other quality parameters [6, 101].
8.7 SUMMARY Experiments of laser cutting non-metallic sheets, based on proposed framework modules,
have been performed in order to model and optimize the laser cutting quality. This was
initially achieved by using orthogonal array as DOE. Some variable parameters were focused
during the laser cutting process such as edge quality, kerf width, percent overcut and material
removal rate as it affected the quality of laser cutting process. However, percent over cut and
material removal rate can be calculated by the kerf width. The combination of input sets is
based on Taguchi’s orthogonal array design. To avoid uncontrollable and human errors, data
was replicated three times and normalized by averaging.
In the work of M. Zaidi et al. [3] data response tables were made without using
preprocessing techniques. Therefore, one value has been taken zero which causes difficulties
in the calculations of target (TPM) and noise performance measurement (NPM) consequently,
unbalancing the analysis table. Therefore, the TPM and NPM recommendations differ and
NPM was preferred as thumb rule of Statisticians[66].
Further, M. Zaidi et al. [25] discussed the possible solutions of non-linear multivariable by
experimental data mining techniques using polystyrene foam cutting data [2]. The results are
not encouraging to study laser cutting process of non-linear multivariable modeled by one
and two way analysis of variance, also linear and nonlinear regression analysis. This detailed
study was started by M. Zaidi et al. [3] by predicting best input data sets using TPM and
NPM. Applying sorting, curve fitting and aggregation technique to solve the overall quality
161
of multi-input output problem. This work was presented in international conference and it
was realized that there is a need of systematic approach to handle the whole problem in
modular form as shown in Figure 6.1.
After data collection “Modulling and optimization Module” was focused for the improvement
of laser cutting process using Statistical methods. The one way ANOVA shows in Table 8-5
and Table 8-6 that feeds controllable parameter Laser power’s P and F values reject null
hypothesis Ho. The Table 8-6 shows that cutting speed, assist gas pressure and standoff
distance insignificantly participating in the variation of kerf width quality. The analysis also
shows that better kerf width predicted as laser power (100 watt), cutting speed (0.7 m/s),
assist gas pressure (2.5 bar) and standoff distance (5 mm).
In comparison, the results of “one way analysis of variance” with both type either replication
or without replication favor the replication. It improves the F value. It means, replication
improves the ability of bifurcation between controllable and uncontrollable variations.
The results of “two way ANOVA” show the interaction results between the two variables and
ignoring among three variables as they were insignificant. The results are concluded in Table
8-14.
The interaction 4, 5 and 6 were significant and was ignored as they will be considered as
uncontrollable variables or pooled error. For better optimization considering all the above
interaction in [2, 3, 16] will give better results or apply Semi-supervised learning algorithm
for modelling. It is not a good practice to assume without analysis that input parameters are
independent from each other in Statistical analysis. One factor at a time has been varied to
analyze the effect of input process parameters on output quality characteristics or responses
[13, 16, 18, 20-22] which shows inferior analysis results due to wrong assumption.
The Table 8-50 shows that the laser power causes significant variation on Kerf width mean.
However, the parameters such as cutting speed, assist gas pressure, standoff distance causes
insignificant role in the variation of dependent variable. The average error model based on
significant factor (approximately 12.95%) which is barely acceptable but insignificant factor
having the value of more than 25% is absolutely unacceptable. In addition to this Kerf width
modelling is easier than edge quality mean and also the data is without any missing value.
162
Table 8-50: Summary of linear regression analysis
Laser
Power
Cutting
Speed
Assist gas
pressure
Standoff
distance
R +0.885 -0.041 +0.126 0.312
Significant r Significant Insignificant Insignificant Insignificant
Min error 5.17% 6.2% 1.38% 6.38%
Max. error 21.55% 54.83% 54.83% 53.97%
Average error 12.95% 25.77% 25.07% 25.94%
The overall coefficient of correlation is improved to positive 0.984 and it is significant. R2
favors the model usage as the kerf width is 89.9% due to planned input variations Kerf width
residual predicted value shows that the minimum, maximum, and average percent errors are
0.71%, 20.66%, 7.99% respectively. Hence, it can be used as an empirical formula without
missing value data.
Table 8-51: Summary of nonlinear Regression
Input Factors % R2 F sig. P of t test Max. error Average error H0
Without replication
Laser power 79.4 0.009 0.195 23.85% 11.94% Insignificant
With replication
Laser power 57.2 3x10-5 0.989 59.48% 55.81% Insignificant
Cutting speed 0.2 0.979 0.889 93.84% 84.75% Insignificant
Assist gas pressure 2.3 0.755 0.691 89.46% 8.89% Insignificant
Standoff distance 12.4 0.206 0.395 82.37% 76.14% Insignificant
The results of nonlinear regression Table 8-51 show that Laser power is the key parameter.
Its R2 value decreases with the replication. The error is increased due to replication and it will
remain increasing if more observations are considered as in factorial design [47]. The null
hypothesis H0 is accepted for all variables which show that nonlinear regression with or
without replication cannot explain the variations in output quality. The residual error due to
predicted values increases in nonlinear case rather than in linear prediction.
163
Kerf Width predicted residual shows that the minimum, maximum, and average percent
errors are 1.82%, 59.48%, 55.81% respectively. The values show that the error is more than
50%. This modelling technique cannot be recommended on the basis of the above Table 8-4
results.
It is concluded that one way ANOVA is able to model the problem with and without
replication as shown in Table 8-6 and Table 8-7. The model gives better results in the case of
replication which shows suitable analysis technique for given datasets. It is also observed that
interaction should be considered to get the better picture of the process optimization as shown
in Table 8-50 [25]. R2 encourages using multiple linear regression model as the Kerf width
variation is 89.9% which is greatly significant. It can be used in rough modeling, simulation
and optimization. The results of nonlinear regression are worst compared to the others
moreover, with replication it becomes more non-realistic because of large number of
observations over and above the fitted points similar to linear regression. The average error
reaches to 50%. The best method was one way ANOVA with pooling but the current research
shows that there is one thing missing that is significantly participating in the variation of
dependent parameters i.e. interaction between two independent and one dependent
parameters. The literature survey shows that ANN model better than regression and other
Statistical techniques [2, 4, 12, 26-35], but need to verify for missing value.
To solve the issue of missing value, the Proposed Framework modules provide guideline to
apply Preprocessing of process module as shown in Figure 6.1 and Figure 6.2. The outlier
analysis and visual inspection of observation tables clearly show that the missing values are
present and the video and picture gallery also show that the sheet was uncut a few times.
The studies show factorial design is appropriate for modelling with ANN. But in this study a
full factorial design is used for L9 orthogonal array and also for larger size of datasets for the
improvement in modelling and verification with supervised learning algorithm of ANN.
This research encourages new researchers to use supervised learning of ANN in case of
missing values. Improvements in the edge quality mean value modelling could be achieved
by normalization. It enlarges the range in case of small range of datasets. The huge modelling
experience gained in this study built the basis of Semi-supervised learning algorithm. This
research verified the results of OA by datasets of FD. The modelling of edge quality signal to
164
noise ratio and kerf width S/N were better but the modelling of kerf width and edge quality
mean can be achieved by normalization of quality parameters. Kerf width mean modelling is
easier. However, edge quality mean is difficult and requires larger datasets for modelling.
Other researchers used 9 datasets for modelling and verified with one or two runs. But, in this
study the prediction was verified by the experimental values of full factorial of 81 runs. So,
the modelling capability of ANN with some missing values was never studied before by
predicting on the whole FD input datasets. The OA data utilizes 5 runs for training, 4 runs for
validation, testing and single missing value rest for the verification. Even though the best
possible results of edge quality mean and S/N ratios in terms of average percent error are
46% and 0.123%. Therefore, this modelling does not meet the desire requirement but S/N
ratio can be utilized. After increasing dataset size as factorial design then edge quality mean
modelling average percent error is 6.9% and signal to noise ratio is 3% which is really
encouraging. The kerf width mean average error is also acceptable i.e. 6.4%. The average
percent error in simulation was under 10% and maximum possible individual percent error is
25%. The process of training explains the effects of change in number of neurons, number of
hidden layers, selection of suitable training algorithm, benefit of enrichment, early stopping,
re-initialization of weights, performance measurement and size of datasets It is better to use
Semi-supervised ANN learning algorithm to model with OA (smaller dataset size).
Previous research suggests larger data set size from orthogonal to factorial design for the
training and testing consequently improves the results, however, the option of re-initialization
had not been tested by the authors. Increase in number of re-initializations of neural network
weights will generate better results as shown in Figure 8.18. One of the major advantages of
Semi-supervised on supervised learning is a huge number of re-initializations of the network
weights such as 100 or 1000 or 3000 times however, in other cases it is 5 to 10 times.
Therefore, it can be concluded that the Semi-Supervised learning is indispensible to produce
better models. As in supervised learning it is difficult to do this job even for 100 times.
In the supervised learning, the edge quality model of 3mm were studied and the average
percent error were 19% (mean) and 24% (S/N) but in case of Semi-supervised the same
parameters were not more than 9.5% and 11%. These results encourage utilizing the Semi-
supervised algorithm technique based on supervised learning experiences. The error at most
is 12.0 with 22% missing values in 5mm models as shown in Figure 8.18. The average
165
percent error in case of edge quality and kerf width for 3mm and 5mm was calculated by
changing number of re-initializations of weights of the network. The mean and standard
deviation were calculated for each case. The results are depicted in Figure 8.17 and Figure
8.18 respectively which shows that this is also an important parameter that can be utilized for
the improvement of neural network modelling. It is also capable of modelling small datasets
such as orthogonal array.
Earlier [4] only 3mm material was used but later an additional 5mm work piece is also used
for experimentation, training & modelling. The number of missing values doubled with the
same input parameters ranges. But, the modelling error increased no more than 2.5%. The
modelling of missing data encourages the cutting of new materials with small experimental
data sets even when missing values exist. However, it is time consuming, expensive and
needs domain expert to adjust the range for performing experiments without missing values.
The results of single output quality were better than multiple outputs [4]. Figure 8.18 shows
that 3000 time re-initialization will generate better results.
The modelling is improved by running GDM and LM training algorithms. Both the training
algorithms are run repeatedly for each number of neuron.To understand the missing values,
assume a collection of points, which are to be used for the formation of a regression line.
With more points the regression will have a higher probability of being accurate. Similarly, in
our case, data used for checking generalization and testing gives small errors on average
results but larger error near the missing points.
Conversion of fractional data into ordinal data also loses a lot of information that the artificial
neural network can be modeled with, using real non-reprocessed data. There is a similar loss
of information in factorial design due to usage of ordinal three level datasets. The factorial
design is better than orthogonal design because of its larger size than OA.
M. Zaidi et al. [12] in the process of laser cutting modelling, orthogonal array based
experimental data was trained by Semi-supervised learning algorithm of ANN, using GDM
and LM training algorithm. In only one model out of eight GDM predicted better than LM.
The GDM based model average percent error was 4.7% and second lowest was 5%, which
was trained by LM. These results also conclude that utilization of LM training algorithm
reduce training time to hours instead of days. It is true that the re-initialization of neural
166
network initial weights is an important factor. The new idea of Semi-supervised learning
definitely improves the modelling accuracy containing up to 22% missing values. This
algorithm will make it easy for the researchers having less insight in soft-computing and laser
cutting. This work reduces modelling time, cost and encourages utilizing existing machines
for the cutting of new materials. The average percent error in 3mm and 5mm Perspex sheet
did not exceed 11.5% and 10.8% in edge quality. Similarly, it also did not exceed 8.0% and
8.7% in case of kerf width. The major achievement is to model orthogonal array with missing
values in a very short time. Therefore, this modelling technique can be used for difficult and
high cost materials for the purpose of modelling and optimization.
M. Zaidi et al. [110] solve the issue of overall quality of laser cutting work-piece using
multi-quality optimization module. The overall quality was calculated by aggregation and
improved it by normalization for equalizing the contribution of all qualities and predicts
unknown machining input setting by neural network model [3]. The normalization was
estimated by equation (8.21) and (8.22). Further improvement was achieved by applying the
customer quality equation based on customer specifications. However, the model is
calculated based on crisp logic and is unable to predict quality variable values separately [2].
The predicted values’ trend was verified by the OA datasets trend. The solution was so
simple and unable to bifurcate, that the total aggregated value rose due to individual factor
domination or it is the accumulated effect of all qualities. Based on customer requirements,
weightage is given to the output quality and the range of the quality tag can be set. Many
researchers have applied intelligent algorithms for overall quality in [9, 11, 60, 68-70, 98, 99,
111, 112] and some others applied data mining techniques [39, 41] and several other
references mentioned in literature review and discussion.
Sum of all qualities = Nor. (Edge quality) + Nor. (Kerf width)
+ Nor. (Overcut) + (1- Nor. (Material removal rate))
(8.21)
Normalized Aggregation = Normalized (Sum of all qualities) (8.22)
The overall quality is based on ANN therefore, the issue of interaction does not effect on the
quality of modelling. However, Ming-Fei et al. [61] and Sharma et al. [36] used statistical
techniques in the modelling, they ignored the interaction effects which generates inferior
results. Semi-supervised learning algorithm overcomes both the issues of missing values and
167
interaction effects. M. Zaidi et al. [12] prepared a more generous and better aggregation,
carried out by the novel combination of Fuzzy logic to provide overall quality for customer
and an indication for rework at initial stage for saving cost and time, by the novel
combination of CQF, Q FL AV and Q AV.
The fuzzy logic results are closer to the CQF. The optimizations of membership function
generate closest result to CQF but not as Crisp aggregation results. The fuzzy rules are
applied simultaneously and cater for the effect of participation of simulated results in fuzzy
manner, it shows the degree of Excellent, Good or worst based on the inputs value, its
member ship function and applied rules. The results of FIS cannot be utilize in place of CQF
but its property of using any membership function and considering the degree of participation
in aggregation gives the idea to utilize this function for optimization and rework function
with CQF and quantified normalized aggregation. For this purpose the result is sorted in
novel combination of Customer quality function, quantified fuzzy aggregation and quantified
normalized aggregation values. CQF has the highest priority because if any one of the output
quality is worst i.e. 3 it is not acceptable for any customer. The quantified fuzzy aggregation
has second priority because it aggregates the value based on simple aggregation with selected
membership functions and also with the rules applied in CQF but in fuzzy manner. The
normalized aggregation function helped to achieve a more comparative list for better
optimization. A part of this list is shown in Table 8-46 and complete table in Appendix B
(Table B-8). The sorted Table 8-46 shows overall input data setting of Run 1, 2 and 3. In
addition, Run 7 and 14 can be utilized by rework because CQF value is 3 but QFL AV and
Normalized aggregated (AV) values are 2. These settings results show that the quantified
value of Kerf width is worst but other qualities are excellent and desired, it shows that overall
quality is a low degree of worst. However, run 77 and up to end are shown in Table 8-46 and
the full table in Appendix B (Table B-8) has QFL AV value at 3 which shows that the product
quality is worst to a high degree so it is better to reject rather than rework. Sometimes, one of
the qualities is worst such as Kerf width while other qualities are excellent and desired,
showing that overall quality is a low degree of worst. In this case, rework is suitable which
saves the cost and time of manufacturing organization. It is also possible to sell the product to
the customers with higher tolerances.
The overall quality fuzzy model can be improved by fine-tuning of membership function’s
boundaries, intersection points, width and shape. These parameters of FIS system can be fine
168
tuned by genetic algorithm.In overall quality problem, genetic algorithm is learned and
studied for the implementation point of view. A detailed study was performed for the
implementation of genetic algorithm [60, 68-70, 98, 99, 111] and it was concluded that for
the following reasons GA is not suitable for our study.
• It is an experimental study based on OA without outer array.
• GA is applied instead of experimentation and the initial population is generated on
random basis and selection is based on cost function. Therefore, cost function is
required.
The genetic algorithm can be applied in our problem however ANN trained model for
simulation is used. Prepare cost function for minimum the better for KW, EQ and POC and
maximum the better for MRR.
The process industry usually needs to improve multiple quality parameters which are
resolved by novel combination of Fuzzy aggregation. It is also an important factor to improve
energy efficient processes which lead to reduction in the manufacturing cost by increasing
cutting speed and decreasing the utilization of laser power per meter cutting. Therefore, the
study adds an energy efficient solution and shows the new quality parameter calculation
which will include in the overall quality as shown in Table 8-47 and possible levels in three
levels factorial design are 9. With reference to Table 8-49 the results are appropriate even
from energy point of view. Assuming that changes in cutting speed causes change in energy
consumption. Therefore, the units are watts. m/hr. Eltawahni et al. [101] looked into the CO2
laser cutting variables of MDF and explained the cutting cost. The equation ( 8.19) and ( 8.20)
indicate that laser power (A) is the major contributor in cost. They also identified that the
power consumed by motion controller does not differ with different speed [101]. It means
that the difference in power consumption with 0.2 m/min or 0.7 m/min or 1.2 m/min is
negligible. So, it is assumed constant power is consumed in motion controller. It means Flow
rate of assist gas pressure is also another parameter which affects the cost per unit length.
Reducing power and increasing speed results in the reduction of cost. Therefore, for the sake
of simplicity, we calculate cost based on the laser power only. As the cost of chiller electrical
power, compressor power, exhaust system power and manpower are constant. With reference
to Table 8-49 the results are appropriate even from cost point of view.
169
H. A. Eltawahni, et al [6, 101] applied limitation on input, output, overall quality and cost
using desirability approach. In overall quality they are unable to give single optimized results,
entrusting the operator to decide whether money is important or quality. In our case
electricity-effective solution was provided without compromising on quality.
CHAPTER 9
CONCLUSION AND FUTURE DIRECTION
170
9. CONCLUSION Laser cutting experiment on non-metallic sheets, based on proposed framework modules,
have been performed in order to model and optimize the laser cutting quality. Edge quality,
kerf width, percent overcut and material removal rate were focused as it affected the quality
of laser cutting process. In the initial work, calculations of TPM and NPM were unbalanced
due to missing values [3]. Therefore, TPM and NPM recommendations were different and
NPM was preferred as a thumb rule of Statisticians. It was realized that there is a need of
systematic approach to handle the whole problem in modular form as shown in Figure 6.1.
Further, detailed Statistical techniques were applied to model the problem. The one way
ANOVA in Table 8-5 and Table 8-6 shows that Laser power’s is a significant factor while,
cutting speed, assist gas pressure and standoff distance insignificantly participate in the
variation of kerf width quality. The results are favorable with or without replication in one
way and two way ANOVA. However, the results of replications are best. The analysis shows
that better kerf width predicted at laser power (100 watt), cutting speed (0.7 m/s), assist gas
pressure (2.5 bar) and standoff distance (5 mm). The results of “two way ANOVA” show the
interaction 4, 5 and 6 were significant between the two variables as shown in Table 8-50. It
was ignored as their error will be added with noise or pooled error.
The Table 8-51 linear regression summary shows that the laser power is significant and
cutting speed, assist gas pressure, standoff play an insignificant role on Kerf width mean. The
average error model based on significant factor (approximately 12.95%) is acceptable but
other factors having the value of more than 25% is absolutely unacceptable. Kerf width
residual predicted value of multiple linear regression shows that the minimum, maximum,
and average percent errors are 0.71%, 20.66%, 7.99% respectively. Hence, it can be used as
an empirical formula without missing value data. The results in Table 8-52 show that
nonlinear regression with or without replication cannot explain the variations in output
quality and results of replication were more inferior. Kerf Width predicted residual shows
that the minimum, maximum, and average percent errors are 1.82%, 59.48%, 55.81%
respectively.
It is concluded that one way ANOVA results are best among the Statistical models with
replication as shown in Table 8-7 provided that missing value do not exist. However, results
171
show that ANN models better than regression and other Statistical techniques. [2, 4, 12, 26-
35]. However, it needs to be verified for missing value problem.
To solve the issue of missing value apply Preprocessing module (outlier analysis) as shown
in Figure 6.1 and Figure 6.2. The studies show factorial design is appropriate for modelling
with ANN. But, unable to model supervised learning model with OA. The supervised
learning modelling of edge quality signal to noise ratio and kerf width S/N were better than
kerf width mean and edge quality mean. Kerf width mean modelling is easier than edge
quality mean which is difficult and requires larger datasets for modelling. The modelling
capability of ANN with some missing values was never studied before by predicting on the
whole FD input datasets. The OA data is utilized 5 runs for training, 4 runs for validation,
testing and single missing value, remaining is for verification. Even though the best possible
results of edge quality mean and S/N ratios in terms of average percent error are 46% and
0.123%. Therefore, this modelling does not meet the desired requirement but S/N ratio can be
utilized.
After increasing dataset size as factorial design then edge quality mean modelling average
percent error is 6.9% and signal to noise ratio is 3% which is really encouraging. The kerf
width mean average error is also acceptable i.e. 6.4%. The average percent error in simulation
was under 10% and maximum possible individual percent error is 25%. The process of
training explains the effects of change in number of neurons, number of hidden layers,
selection of suitable training algorithm, benefit of enrichment, early stopping, re-initialization
of weights, performance measurement and size of datasets. Increase in number of re-
initializations of neural network weights will generate better results as shown in Figure 8.18.
In the supervised learning, the edge quality model of 3mm were studied and the average
percent error were 19% (mean) and 24% (S/N) but in case of Semi-supervised the same
parameters were not more than 9.5% and 11%. These results encourage utilizing the Semi-
supervised algorithm technique based on supervised learning experiences. The error at most
is 12.0 with 22% missing values in 5mm models as shown in Figure 8.18.
M. Zaidi et al. [12] used GDM and LM training algorithm in ANN. In only one model out of
eight GDM predicted better than LM. The GDM based model average percent error was 4.7%
and second lowest was 5%, which was trained by LM. These results also conclude that
172
utilization of LM training algorithm reduce training time to hours instead of days. The
number of missing values doubled with the same input parameters ranges. But, the modelling
error increased no more than 2.5%. The modelling of missing data encourages the cutting of
new materials with small experimental data sets even when missing values exist. However, it
is time consuming, expensive and needs domain expert to adjust the range for performing
experiments without missing values. The results of single output quality were better than
multiple outputs [4]. Figure 8.18 shows that 3000 time re-initialization will generate better
results.
This algorithm will make it easy for the researchers having less insight in soft-computing and
laser cutting. This work reduces modelling time, cost and encourages utilizing existing
machines for the cutting of new materials. It is concluded that Semi-supervised learning
algorithm is capable of modelling small datasets such as OA.
M. Zaidi et al. [110] solved the issue of overall quality of laser cutting work-piece using
multi-quality optimization module. It was calculated by aggregation and improved by
normalization for equalizing the contribution of all qualities and predicts unknown machining
input setting by neural network model [3]. The normalization was estimated by equation
(8.21) and (8.22). Further improvement was achieved by applying the customer quality
equation based on customer specifications. However, the model is calculated based on crisp
logic and is unable to predict quality variable values separately [2]. The predicted values’
trend was verified by the OA datasets trend. The solution was so simple and unable to
bifurcate, that the total aggregated value rose due to individual factor domination or it is the
accumulated effect of all qualities. Based on customer requirements, weightage is given to
the output quality and the range of the quality tag can be set. The overall quality is based on
ANN (Semi-supervised learning algorithm) therefore, the issue of interaction does not effect
the quality of modelling. M. Zaidi et al. [12] prepared a more generous and better aggregation,
carried out by the novel combination of Fuzzy logic to provide overall quality for customer
and an indication for rework at initial stage for saving cost and time, by the novel
combination of CQF, Q FL AV and Q AV.
The fuzzy logic results are closer to the CQF. The fuzzy rules are applied simultaneously and
cater for the effect of participation of simulated results in fuzzy manner, it shows the degree
of Excellent, Good or worst based on the inputs value, its member ship function and applied
173
rules. The property of using any membership function and considering the degree of
participation in aggregation gives the idea to utilize this function for optimization and rework
function with CQF and quantified normalized aggregation. For this purpose the result is
sorted in novel combination. CQF has the highest priority because if any one of the output
quality is worst i.e. 3 it is not acceptable for any customer. The quantified fuzzy aggregation
has second priority because it aggregates the value based on simple aggregation with
selected membership functions and also with the rules applied in CQF but in fuzzy manner.
The normalized aggregation function helped to achieve a more comparative list for better
optimization. A part of this list is shown in Table 8-46 and complete table in Appendix B
(Table 12-8). The sorted Table 8-46 shows overall input data setting of Run 1, 2 and 3.
In addition, Run 7 and 14 can be utilized by rework because CQF value is 3 but QFL AV and
Normalized aggregated (AV) values are 2. These settings results show that the quantified
value of Kerf width is worst but other qualities are excellent and desired, it shows that overall
quality is a low degree of worst. However, run 77 and up to end are shown in Table 8-46 (full
table in Appendix B (Table B-8) having QFL AV value at 3 which shows that the product
quality is worst to the high degree so it is better to reject rather than rework. Sometimes, one
of the qualities is worst such as Kerf width is worst but other qualities are excellent and
desired, showing that overall quality is a low degree of worst. In this case, rework is suitable
which saves the cost and time of manufacturing organization. It is also possible to sell the
product to the customers with higher tolerances.
Electricity is also an important factor to improve energy efficient processes which lead to
reduction in the manufacturing cost by increasing cutting speed and decreasing the utilization
of laser power per meter cutting as shown in Table 8-49 the results are appropriate even from
energy point of view. Assuming that changes in cutting speed causes change in energy
consumption. Therefore, the units are watts. m/hr. The equation ( 8.19) and ( 8.20) indicate
that laser power (A) is the major contributor in cost and power consumed by motion
controller which does not differ with different speed [101]. It means Flow rate of assist gas
pressure is also another parameter which affects the cost per unit length. Reducing power and
increasing speed results in the reduction of cost. Therefore, for the sake of simplicity, we
calculate cost based on the laser power only. As the cost of chiller electrical power,
compressor power, exhaust system power and manpower are constant. With reference to
Table 8-49 the results are appropriate even from cost point of view. Eltawahni, et al [6, 101]
174
used desirability approach for overall quality but were unable to give single optimized results,
entrusting the operator to decide whether money is important or quality. In our case
electricity-effective solution was provided without compromising on quality.
CHAPTER 10
FUTURE DIRECTION
175
10. FUTURE DIRECTION
In my view the design of experiment used by AI techniques requires some modifications.
Some data set out of designed matrix are required for checking of generalization of model
and some for testing of the model. This issue can be resolved without investing on
experimentation, by taking the observation taken for range adjustment before the actual
experimentation. Range adjustment will be carried out systematically and this observation
table will be used in ANN modelling.
Future research can be focused on running process industry to improve the quality and
reduction in cost by optimizing the use of Electricity. The modelling and optimization can be
applied in the light of “Proposed Framework Module”.
Figure 10-1: Brief Procedure to solve the problem
Titanium Tungsten Inert Gas (TIG) welding is more complex as compared to other welding
processes applied in process industry. Welding guidelines are available however “experience
and practice is the trainer”. The cost of Titanium pipe is very high and in case of rejection
more pipe is required which can take up to 6 months to procure. Therefore, bad weld costs
money and time. A 10% Radiography (RT) rejection in welding is considered acceptable.
Smooth movement of welder’s hand, maintaining uniform speed and constant distance from
•Understand the problem and aim
•Screening of variables for this purpose
•Select the data for modelling
Preprocessing & Experimental data
•Select the modelling method •Model the problem by Semi-
supervised method or more appropriate
Modelling•Apply appropriate
techhnique• If apply Fuzzy agregation
then optimize memebership functions
Optimization
176
work-piece joint will produce better results. But due to human error it is difficult to maintain
consistency. The cost of labor, material, workshop capital and workshop running are
expensive. There is a need to develop a setting to reduce the rejections.
Normally vendor documents suggest the welding parameters, however recommended values
are only guidelines, quality also relies on skill and experience. Hence our proposed
Framework can be applied in this problem for consistent and improved results. Evaluation of
the problem shows that following are the constant parameters
1. Temperature < 26 °C
2. Relative Humidity < 30%
3. Air tight Environment
4. Stable power supply
5. Metal Thickness 0.1875”
Following are the controllable parameters
1. Shielding Gases for TIG
Argon
Helium
Mixtures
2. Tungsten Electrode diameter
0.125” to 0.156”
3. Current in Ampere
180 to 225
4. Distance between Electrode and Work-piece
5mm to 50 mm
Speed of Electrode on joint in first pass
Quality will be assessed by the
1. RT
2. Heat affected zone
3. Joint roughness.
For high accuracy in measuring, instead of using a Caliper other measuring tools could be
used for small least count and better repeatability.
177
A mechanical structure will be prepared around the pipe where TIG electrode can move with
different speed settings and the distance between the electrode and joint can be adjusted.
There are four controllable variables and each have 3 ordinal levels. The output quality will
be assessed by the three variables. The DOE is L9 orthogonal array and range adjustment
observation will be used to model the problem by Semi-supervised model. The Fuzzy
aggregation is more appropriate to resolve the issue of overall optimization because quality
parameters are more than 3. The overall quality fuzzy model can be improved by fine-tuning
of membership function’s boundaries, intersection points, width and shape. These parameters
of FIS system can be fine tuned by genetic algorithm.
REFERENCE
178
REFERENCE
[1] I. A. Choudhury and S. Shirley, "Laser cutting of polymeric materials: An experimental investigation," Optics and laser technology, vol. 42, pp. 503-508, 2010.
[2] M. Zaidi and I. Amin, "Laser cutting parameters optimization of polystyrene foam using taguchi method and artificial intelligence tools," Master of Computer Science, Computer Science, SZABIST Karachi, Pakistan, 2010.
[3] M. Zaidi, et al., "Laser Cutting Quality Control of Melamine Using Artificial Neural Networks " in APIEMS: Asia Pacific Industrial Engineering and Management Systems Conference, Melaka, Malaysia, 2010.
[4] M. Zaidi, et al., "Estimation of ANN Modelling of Laser Cutting with Missing Values," Pensee Journal, vol. 75, pp. 159-171, 2013.
[5] F. Caiazzo, et al., "Laser cutting of different polymeric plastics (PE, PP and PC) by a CO2 laser beam," Journal of Materials Processing Technology, vol. 159, pp. 279-285, 2005.
[6] H. A. Eltawahni, et al., "Effect of CO2 laser cutting process parameters on edge quality and operating cost of AISI316L," Optics and laser technology, vol. 44, pp. 1068-1082, 2012.
[7] J. D. Majumdar and I. Manna, "Laser processing of materials," SADHANA, Academy Proceedings in Engineering Sciences, Indian Academy of Sciences vol. 28, pp. 495–562, 2003.
[8] M. M. Noor, et al., "Particle swarm optimisation prediction model for surface roughness," International Journal of Physical Sciences, vol. 6, pp. 3082-3090, 2011.
[9] Sivarao, et al., "A New Approach of Adaptive Network-Based Fuzzy Inference System Modeling in Laser Processing-A Graphical User Interface (GUI) Based," Journal of Computer Science, vol. 5 pp. 704-710, 2009.
[10] N. Yusoff, et al., "Selected Malaysian Wood CO2-Laser Cutting Parameters And Cut Quality," American Journal of Applied Sciences, vol. 5, pp. 990-996, 2008.
[11] M. Zaidi and I. Amin, "Hybrid combination of Crisp and Fuzzy aggregation for Overall quality of Laser Cutting Process," Submited, Journal of Central South University, vol. 21, December 2014.
[12] M. Zaidi, et al., "Error Assessment of Laser cutting Predictions by Semi-Supervised Learning," Journal of Central South University, vol. 21, pp. 3736-3745, October 2014.
[13] B. H. Zhou and S. M. Mahdavian, "Experimental and theoretical analyses of cutting nonmetallic materials by low power CO2-laser," Journal of Materials Processing Technology, vol. 146, pp. 188-192, 2004.
[14] J. Mathew, et al., "Parametric studies on pulsed Nd:YAG laser cutting of carbon fibre reinforced plastic composites," Journal of Materials Processing Technology, vol. 89-90, pp. 198-203, 1999.
[15] D. G. Ahn, et al., "Investigation into thermal characteristics of linear hotwire cutting system for variable lamination manufacturing (VLM) process by using expandable polystyrene foam," International Journal of Machine Tools & Manufacture, vol. 42, pp. 427–439, 2002.
[16] S. B. Tan and Y.Nukman, "A study of laser engraving on Ceramic material," Graduation of Engineering, Department Of Engineering Design And Manufacture, Faculty of Engineering, University Of Malaya, Kuala Lumpur, 2008.
[17] H. Golnabi and M. Bahar, "Investigation of optimum condition in oxygen gas-assisted laser cutting," Optics and laser technology, vol. 41, pp. 454-460, 2009.
179
[18] I. Uslan, "CO2 laser cutting : kerf width variation during cutting," Journal of engineering manufacture, vol. 219, pp. 571-577, 2005.
[19] J. ZHU, et al., " An 8-Axis Robot Based Rough Cutting System for Surface," Journal of Advance Mechanical Design, Systems, and Manufacturing, vol. 2, pp. 812-823, 2008.
[20] Yang, et al., "Development of a rough set-based fuzzy neural network for online monitoring of microdrilling," The International Journal of Advanced Manufacturing Technology, vol. 41, pp. 219-225, 2009.
[21] C. Wandera and D. A. Salminen, "Laser cutting of austenitic stainless steel with a high quality laser beam," Master of Science in Technology, Department of Mechanical Engineering, Lappeenranta University Of Technology, Lappeenranta, Finland, 2006.
[22] B. S. Yilbas, et al., "Laser Cutting of Rectangular Blanks in Thick Sheet Steel:Effect of Cutting Speed on Thermal Stresses," ASM International, ASM International, vol. 19, pp. 177-184, 2010.
[23] Sivarao, et al., "DOE Based Statistical Approaches in Modeling of Laser Processing – Review & Suggestion," International Journal of Engineering & Technology IJET-IJENS, vol. 10, 2010.
[24] A. Zaman and M. Zaidi, "Evaluation of Training Algorithms for the Laser Cutting Process," Journal of Independent Studies and Research-Computing, vol. 9, pp. 1-13, 2011
[25] M. Zaidi, et al., "Experimental Data Mining Techniques (Using Multiple Statistical Methods)," IJCSI International Journal of Computer Science Issues, , vol. 9, pp. 132-138, January 2012.
[26] D. Dhupal, et al., "Optimization of process parameters of Nd:YAG laser microgrooving of Al 2TiO5 ceramic material by response surface methodology and artificial neural network algorithm," Proceedings of the Institution of Mechanical Engineers, Part B: Journal of Engineering Manufacture, vol. 221, pp. 1341-1351, 2007.
[27] A. T. C. Goh, "Back-propagation neural networks for modeling complex systems," Artificial Intelligence in Engineering vol. 9, pp. 143-151, 1994.
[28] C.-F. J. Kuo, et al., "Application of an LM-neural network for establishing a prediction system of quality characteristics for the LGP manufactured by CO2 laser," Optics & Laser Technology vol. 43 (2011) pp. 529–536, 2011.
[29] R. Kuo, "Intelligent tool wear system through artificial neural networks and fuzzy modeling," Journal of Artificial Intelligence in Engineering, vol. 5, pp. 229-242, 1998.
[30] L. K. Pan, et al., "Optimizing multiple quality characteristics via Taguchi method-based Grey analysis," Journal of Materials Processing Technology, vol. 182, pp. 107–116, 2007.
[31] B. J. Ranaganth and G. Viswanath, "Application of artificial neural network for optimising cutting variables in laser cutting of 304 grade stainless steel," International Journal of Applied Engineering and Technology vol. 1, pp. 116-112, 2011.
[32] N. Shihani, et al., "Modeling of extrusion process using response surface methodology and artificial neural networks," Journal of Engineering Science and Technology, vol. 1, pp. 31-40, 2006.
[33] Sivarao, et al., "Neural Network Multi Layer Perceptron Modeling of Surface Quality in Laser Machining," presented at the International Conference on Intelligent and Advance Systems, 2007.
180
[34] M.-J. Tsai, et al., "Optimal laser-cutting parameters for QFN packages by utilizing artificial neural networks and genetic algorith," Journal of materials processing technology, vol. 208 pp. 270–283, 2008.
[35] C. B. Yang, et al., "Combining the Taguchi method with artificial neural network to construct a prediction model of a CO2 laser cutting experiment," International Journal of Advanced Manufacturing Technology, pp. 1-9, 2011.
[36] A. Sharma and V. Yadava, "Modelling and optimization of cut quality during pulsed Nd:YAG laser cutting of thin Al-alloy sheet for straight profile," Optics & Laser Technology, vol. 44, pp. 159-168, 2012.
[37] Sivarao, et al., "Neural Network Modeling and Validation for Micro Quality Evaluation in Laser Processing," International Journal of Intelligent Information Technology Application, vol. 2, pp. 84-90, 2009.
[38] Sivarao, et al., "Modeling, Testing and Experimental Validation of Laser Machining Micro Quality Response by Artificial Neural Network," International Journal of Engineering & Technology IJET, vol. 9, pp. 161-166, 2009.
[39] A. K. Pandey and A. K. Dubey, "Taguchi based fuzzy logic optimization of multiple quality characteristics in laser cutting of Duralumin sheet," Optics and Lasers in Engineering, vol. 50, pp. 328-335, 2012.
[40] M. Bahar and H. Golnabi, "Optimum experimental condition in oxygen gas-assisted low power Nd:Yag laser cutting," Modern Physics Letters B, vol. 23, pp. 877-890, 2009.
[41] B. S. Yilbas and M. Rashid, "CO2 laser cutting of Incoloy 800 HT alloy and its quality assessment," Lasers in Engineering, vol. 12, pp. 135-145, 2002.
[42] C. Z. Syn, et al., "Approach to prediction of laser cutting quality by employing fuzzy expert system," Expert Systems with Applications, vol. 38, pp. 7558-7568, 2011.
[43] Sivarao, "Computational modeling of surface roughness in CO2 laser cutting quality evaluation," in International Conference on Intelligent and Advanced Systems 2007, Kuala Lumpur, 2007, pp. 528-532.
[44] S. Subramonian, et al., "Modelling of laser processing cut quality by an adaptive network-based fuzzy inference system," Proceedings of the Institution of Mechanical Engineers, Part C: Journal of Mechanical Engineering Science, vol. 223, pp. 2369-2381, 2009.
[45] Sivarao, et al., "Mamdani Fuzzy Inference System Modeling to Predict Surface Roughness in Laser Machining," International Journal of Intelligent Information Technology Application, vol. 2, pp. 12-18, 2009.
[46] C. Saleh, et al., "optimization of fuzzy membership function using genetic algorithm to minimize the mean square error of credit status prediction," in APIEMS, Melaka Malaysia, 2010.
[47] I. A. Almeida, et al., "Optimization of titanium cutting by factorial analysis of the pulsed Nd:YAG laser parameters," Journal of Materials Processing Technology, vol. 179, pp. 105-110, 2006.
[48] P. A. Atanassov, "CW CO2 laser cutting of plastics," in SPIE, Edinburgh, 1997, pp. 772-775.
[49] S. I. Powel, "CO2 Laser Cutting, LASER 5: Laser Materials Processing for Industry," in IITT International Gournay Sur Marne, 1989, pp. 54-71.
[50] X. Schmidt and C. A. Marlies, Eds., Principles of High-polymer Theory Anti Practice. New York: McGraw-Hill, 1948, p.^pp. Pages.
[51] J. Bandrup and E. Immergut, Eds., Polymer Handbook. New York: Wiley/Inter science, 1989, p.^pp. Pages.
181
[52] L. E. Nielson and E. L. Nielsen, Eds., Mechanical Properties of Polymers. New York: Marcel Dekker, 1974, p.^pp. Pages.
[53] D. M. V. Krevelen, Ed., Properties of Polymers. Amsterdam: Elsevier, 1990, p.^pp. Pages.
[54] J. Powell, et al., "Filtration of the fumes generated by industrial CO2 laser cutting," Journal of Laser Applications, vol. 8, pp. 241-246, 1996.
[55] P. G. BERRIE and F. N. BIRKETI, "The Drilling and cutting of Polymethyl methacrylate (Perspex) By CO@ Laser," 1980.
[56] S. D. Balkin and D. K. J. Lin, "A neural network approach to response surface methodology," Communications in Statistics - Theory and Methods, vol. 29, pp. 2215-2227, 2000.
[57] M. M. Noor, et al., "Development of statistical model to predict Ra and Rz in laser cutting," Tallinn, 2010, pp. 401-408.
[58] M. M. Noor, et al., "Prediction modelling of surface roughness for laser beam cutting on acrylic sheets," vol. 83-86, ed. Manama, 2010, pp. 793-800.
[59] A. K. Dubey and V. Yadava, "Multi-objective optimisation of laser beam cutting process," Optics & Laser Technology, vol. 40, pp. 562-570, 2008.
[60] B. Forouraghi, "A Genetic Algorithm for Multiobjective Robust Design," Applied Intelligence vol. 12, pp. 151-161, 2000.
[61] Y.-S. H. Ming-Fei Chena, Wen-Tse Hsiaob, Tse-Hung Wua, Shih-Feng Tsengb and Kuo-Cheng Huangb, "Optimized laser cutting on light guide plates using grey relational analysis," Optics and Lasers in Engineering, vol. 49, pp. 222-228, 2011.
[62] P. J. Pawar, et al., "Multiobjective Optimization of Grinding Process Parameters Using Particle Swarm Optimization Algorithm," Materials and Manufacturing Processes, vol. 25, pp. 424 - 431, 2010.
[63] P. W. Phillips and K.-J. Kim, "Taguchi Parameter Design with Multiple Quality Characteristics," Quality Management journal vol. 4, pp. 26-40, 1999.
[64] E. Viles, et al., "Planning Experiments, the First Real Task in Reaching a Goal," Quality Engineering, Taylor & Francis Group vol. 21, pp. 44–51, 2009.
[65] G. J. Besseris, "Analysis of an Unreplicated Fractional-Factorial Design Using Nonparametric Tests," Quality Engineering, vol. 20, pp. 96-112, 2008.
[66] S. Sarin. (1997) Teaching Taguchi’s Approach to Parameter Design. Quality Progress.
[67] J. Kunert, et al., "An Experiment to compare Taguchi’s Product Array and the Combined Array," Journal of Quality Technology ASQ, vol. 39, 2007.
[68] P. Chatsirirungruang, "Application of Genetic Algorithm to Numerical Experiment in Static Robust Parameter Design," J Jpn Ind Manage Assoco, vol. 59, pp. 312-319, 2008.
[69] P. Chatsirirungruang and M. Miyakawa, "Application of Genetic Algorithm to Numerical Experiment in Robust Parameter Design for Signal Multi-Response Problem," International Journal of Manageent Science and Engineering Management vol. 4, pp. 49-59, 2009.
[70] P. Chatsirirungruang, "Application of Genetic Algorithm and Taguchi method in Dynamic Robust Parameter design for Unknown problems," Int J Adv Manuf Technol, vol. 47, pp. 993-1002, 2010.
[71] M. A. Wazed, et al., "Uncertainty Factors in Real Manufacturing Environment" Australian Journal of Basic and Applied Sciences, vol. 3, pp. 342-351, 2009.
[72] W. Q. Meeker and L. A. Escobar, "Reliability: The Other Dimension of Quality," Quality Technology &Quantitative Management, vol. 1, pp. 1-25, 2003.
182
[73] D. C. Montgomery, Design and Analysis of Experiments, Third Edition ed. New York: John Wiley & Sons, 1991.
[74] N. R. Draper and H. Smith, Applied Regression Analysis. NewYork: John Wiley, 1981.
[75] G. E. P. Box and N. R. Draper, Empirical model-building and response surfaces. New York: John Wiley & Sons, 1987.
[76] G. E. P. Box and D. W. Behnken, "Some new three level designs for the study of quantitative variables," Technometrics, pp. 455-475, 1963.
[77] D. Baş and İ. H. Boyacı, "Modeling and optimization I: Usability of response surface methodology," Journal of Food Engineering, vol. 78, pp. 836-845, 2007.
[78] J. Ciurana, et al., "Neural Network Modeling and Particle Swarm Optimization (PSO) of Process Parameters in Pulsed Laser Micromachining of Hardened AISI H13 Steel," Materials and Manufacturing Processes, vol. 24(3), pp. 358-368, 2010.
[79] B. S. Yilbas, "Study of parameters for CO2 laser cutting process," Materials and Manufacturing Processes, vol. 13, pp. 517-536, 1998.
[80] W. S. McCulloch and W. Pitts, "A logical calculus of the ideas immanent in neural nets," in Bulletin of Mathematical Biophysics vol. 5, ed, 1943, pp. 115–137.
[81] G. F. Luger and W. A. Stubblefield, Artificial Intelligence (Structures and Strategies for Complex Problem Solving), Third ed. MA 01867-3999 USA: Addion Wesley Longman, 1998.
[82] F. Rosenblatt, "The perceptron: A probabilistic model for information storage and organization in the brain," Psychological Review, vol. 65, pp. 386-408, 1958.
[83] M. Minsky and S. Papert, Perceptrons. . Cambridge, MA: MIT Press,, 1969. [84] J. L. McClelland and D. E. Rumelhart, Parallel distributed processing: MIT Press
Cambridge, MA, USA, 1986. [85] D. Anderson and G. McNeill, "Artificial Neural Networks Technology," Kaman
Sciences Corporation, 258 Genesse Street Utica, New York 13502-4627 A011, August 20, 1992 1992.
[86] N. Qian, "On the momentum term in gradient descent learning algorithms," Neural Networks, vol. 12, pp. 145-151, 1999.
[87] A. Chavez, "Artificial Neural Networks," Maths, California State University, Northridge, California, 2010.
[88] I. Amin, et al., "Automated people counting by using low-resolution infrared and visual cameras," Elsevier Editorial System(tm) for Measurement., vol. 41, pp. 589-599 2008.
[89] S. M. Bohte, et al., "Error-back-propagation in temporally encoded networks of spiking neurons," Neurocomputing, vol. 48, pp. 17-37, 2002.
[90] H. Demuth, et al., "Neural Network Toolbox™ 6 User’s Guide," in Matlab Documentation vol. 6.0.4, Release 2010a ed, 2010.
[91] D. Nguyen and B. Widrow, "Improving the Learning Speed of 2-Layer Neural Networks by Choosing Initial Values of the Adaptive Weights," in IJCNN International Joint Conference on Neural Networks, San Diego, CA , USA, 1990, pp. 21-26.
[92] S. Haykin, Neural Networks and Learning Machines, 3rd ed. New Jersey: Pearson Education Inc., 2009.
[93] Bulman, Elementary Statistics, Sixth Edition ed.: McGraw-Hill International Edition, 2007.
[94] A. Ranganathan, "The Levenberg-Marquardt Algorithm," Journal of Materials Processing Technology, vol. 179 pp. 105–110, 2004.
183
[95] R. P. W. Duin, "Learned from Neural Networks," Pattern Recognition Letters Elsevier, vol. 17, pp. 529-536, 1996.
[96] Z. Xu and Z. Mao, "Comparisons of Element Yield Rate Prediction Using Feed-Forward Neural Networks and Support Vector Machine " in Control and Decision Conference (CCDC), 2010 Chinese, 2010
[97] D. B. Fogel and M. Mitchel, "An introduction to genetic algorithms," Bulletin of Mathematical Biophysics, vol. vol. 59, pp. 199-205, 1997.
[98] V. R. Fernandes, et al., "Multi-objective genetic algorithm applied to spectroscopic ellipsometry of organic-inorganic hybrid planar waveguides," 2010 Optical Society of America, Portogal and Tokyo2010.
[99] L. Yongxian, et al., "Intelligent Algorithm of Robust Design for mechanism (orthogonal Optimization Based on the Analysis of Variance Ratio)," presented at the 2010 International Conference on easuring Technology and Mecatronics Automation, 2010.
[100] J. A. Ghani, et al., "Application of Taguchi method in the optimization of end milling parameters," Journal of Materials Processing Technology, vol. 145, pp. 84–92, 2004.
[101] H. A. Eltawahni, et al., "Investigating the CO2 laser cutting parameters of MDF wood composite material," Optics & Laser Technology, vol. 43, pp. 648–659, 2011.
[102] L. A. Zadeh, "Fuzzy Sets," vol. Information and Control, pp. 338-335, 1965. [103] G. Taguchi and S. Konishi, Orthogonal Arrays and Linear Graphs: MI: ASI Press,
1987. [104] K. Chen and Y. L. Yao, "Process Optimisation in Pulsed Laser Micromachining with
Applications in Medical Device Manufacturing," The International Journal of Advanced Manufacturing Technology, vol. 16, pp. 243-249, 2000/03/01 2000.
[105] N. Qazi and Y. Hoi, "Estimation of weights to combine trained neural networks using linear estimation techniques," in Multitopic Conference (INMIC), 2011 IEEE 14th International, 2011, pp. 1-7.
[106] P. Xiao, et al., "Combined Training of Recurrent Neural Networks with Particle Swarm Optimization and Backpropagation Algorithms for Impedance Identification," in IEEE Swarm Intelligence Symposium (SIS 2007), 2007.
[107] T. Kondo and J. Ueno, "Revised GMDH-type neural network algorithm with a feedback loop identifying sigmoid function neural network," International Journal of Innovative Computing, Information and Control vol. 2, p. 985—996, 2006.
[108] "Transactions of the North American Manufacturing Research Institution of SME 2007 - Papers Presented at NAMRC 35," 2007.
[109] R. Fisher, Statistical methods for research worker. London: Oliver & Boyd, London, 1925.
[110] M. Zaidi and N. Haider, "Comparative Study of Customer Quality Function, Genetic Algorithm and Fuzzy Logic in Laser Cutting Process," presented at the SZABIST 16th National research conference (NRC) Karachi, Pakistan, 2010.
[111] A. Alipoor and M. Fesharaki, "Multi objective optimization for object recognition," presented at the International Conference on Education Technology and Computer (ICETC), 2010.
[112] K. Khaldoun, et al., "A Neuro-Fuzzy Reasoning System for Mobile Robot Navigation," Jordan Journal of Mechanical and Industrial Engineering, vol. 3, pp. 77 - 88, 2009.
[113] L. H. Vassie, et al., "Fumes generated during laser processing of polyvinyl chloride (PVC)," Optics & Laser Technology, vol. 27, pp. 31-37, 1995.
[114] (2014). The Plastic Shop.co.uk. Available: http://www.theplasticshop.co.uk/
184
[115] J.-H. Choua, et al., "Application of the Taguchi-genetic method to design an optimal grey-fuzzy controller of a constant turning force system," Journal of Materials Processing Technology, vol. 105, pp. 333-343, 2000.
[116] K. Funahashi, "On the Approximate Realization of Continuous Mappings by Neural Networks," Neural Networks, vol. 2, pp. 183-192, 1989.
[117] F. J. Pontes, et al., "Artificial neural networks for machining processes surface roughness modeling," The International Journal of Advance Manufacturing Technology, vol. 49, pp. 879–902, 2010.
[118] A. M. Zain, et al., "Review of ANN Technique for Modeling Surface Roughness Performance Measure in Machining Process," presented at the 2009 Third Asia International Conference on Modelling & Simulation, 2009.
[119] H. K. Ghuge and D. G. Regulwar, "Artificial Neural Network Method for Estimation of Missing Data," International Journal of Advanced Technology in Civil Engineering, vol. 2, 2013.
APPENDIX A
185
A. DOE & STATISTICAL MODELLIN Table A-1: Four variables with three levels Factorial design matrix
Run Input variable Measurement
Laser Power
A
Cutting speed
B
Assist gas Pressure
C
Standoff Distance
D
Mean S/N Ratio
1 100 0.2 0.5 1 2 100 0.2 0.5 5 3 100 0.2 0.5 10 4 100 0.2 2.5 1 5 100 0.2 2.5 5 6 100 0.2 2.5 10 7 100 0.2 4.5 1 8 100 0.2 4.5 5 9 100 0.2 4.5 10
10 100 0.7 0.5 1 11 100 0.7 0.5 5 12 100 0.7 0.5 10 13 100 0.7 2.5 1 14 100 0.7 2.5 5 15 100 0.7 2.5 10 16 100 0.7 4.5 1 17 100 0.7 4.5 5 18 100 0.7 4.5 10 19 100 1.2 0.5 1 20 100 1.2 0.5 5 21 100 1.2 0.5 10 22 100 1.2 2.5 1 23 100 1.2 2.5 5 24 100 1.2 2.5 10 25 100 1.2 4.5 1 26 100 1.2 4.5 5 27 100 1.2 4.5 10 28 300 0.2 0.5 1 29 300 0.2 0.5 5 30 300 0.2 0.5 10
186
Run Input variable Measurement
Laser Power
A
Cutting speed
B
Assist gas Pressure
C
Standoff Distance
D
Mean S/N Ratio
31 300 0.2 2.5 1 32 300 0.2 2.5 5 33 300 0.2 2.5 10 34 300 0.2 4.5 1 35 300 0.2 4.5 5 36 300 0.2 4.5 10 37 300 0.7 0.5 1 38 300 0.7 0.5 5 39 300 0.7 0.5 10 40 300 0.7 2.5 1 41 300 0.7 2.5 5 42 300 0.7 2.5 10 43 300 0.7 4.5 1 44 300 0.7 4.5 5 45 300 0.7 4.5 10 46 300 1.2 0.5 1 47 300 1.2 0.5 5 48 300 1.2 0.5 10 49 300 1.2 2.5 1 50 300 1.2 2.5 5 51 300 1.2 2.5 10 52 300 1.2 4.5 1 53 300 1.2 4.5 5 54 300 1.2 4.5 10 55 500 0.2 0.5 1 56 500 0.2 0.5 5 57 500 0.2 0.5 10 58 500 0.2 2.5 1 59 500 0.2 2.5 5 60 500 0.2 2.5 10 61 500 0.2 4.5 1 62 500 0.2 4.5 5 63 500 0.2 4.5 10 64 500 0.7 0.5 1 65 500 0.7 0.5 5 66 500 0.7 0.5 10
187
Run Input variable Measurement Laser Power
A
Cutting speed
B
Assist gas Pressure
C
Standoff Distance
D
Mean S/N Ratio
67 500 0.7 2.5 1 68 500 0.7 2.5 5 69 500 0.7 2.5 10 70 500 0.7 4.5 1 71 500 0.7 4.5 5 72 500 0.7 4.5 10 73 500 1.2 0.5 1 74 500 1.2 0.5 5 75 500 1.2 0.5 10 76 500 1.2 2.5 1 77 500 1.2 2.5 5 78 500 1.2 2.5 10 79 500 1.2 4.5 1 80 500 1.2 4.5 5 81 500 1.2 4.5 10
Table A-2: Edge quality mean and S/N ratio of Polystyrene foam Sheet (13mm)
Run Input Variable Measurement Laser Power
A
Cutting speed
B
Assist gas Pressure
C
Standoff Distance
D
Replication Edge quality Mean
S/N Ratio R1 R2 R3
1 100 0.2 0.5 1 3.50 4.00 3.50 3.67 -11.312 2 100 0.2 0.5 5 2.20 3.00 3.00 2.73 -8.856 3 100 0.2 0.5 10 1.50 1.50 2.00 1.67 -4.565 4 100 0.2 2.5 1 1.50 1.50 2.00 1.67 -4.565 5 100 0.2 2.5 5 1.00 1.00 1.00 1.00 0.000 6 100 0.2 2.5 10 2.50 2.50 3.00 2.67 -8.570 7 100 0.2 4.5 1 1.50 2.00 3.00 2.17 -7.225 8 100 0.2 4.5 5 1.50 2.00 2.00 1.83 -5.371 9 100 0.2 4.5 10 1.50 2.50 2.00 2.00 -6.284
188
Table A-3: Kerf width observations of Polystyrene foam Sheet (13mm)
Run Input Variable Measurement Laser Power
A
Cutting speed
B
Assist gas Pressure
C
Standoff Distance
D
Replication Edge quality Mean
S/N Ratio R1 R2 R3
1 100 0.2 0.5 1 1.58 1.62 1.53 1.573 -3.940 2 100 0.7 2.5 5 1.73 1.26 1.48 1.488 -3.563 3 100 1.2 4.5 10 1.66 1.70 1.86 1.738 -4.819 4 300 0.2 2.5 10 1.94 1.89 1.92 1.913 -5.637 5 300 0.7 4.5 1 1.77 1.78 1.99 1.842 -5.324 6 300 1.2 0.5 5 1.66 1.72 1.86 1.742 -4.834 7 500 0.2 4.5 5 2.01 2.04 1.86 1.968 -5.893 8 500 0.7 0.5 10 1.98 1.94 2.29 2.068 -6.350 9 500 1.2 2.5 1 1.79 1.89 2.08 1.920 -5.692
Table A-4: Edge quality observations of Perspex Sheet (3mm)
Run Input Variable Measurement Laser Power
A
Cutting speed
B
Assist gas Pressure
C
Standoff Distance
D
Replication Edge qualityMean
S/N Ratio
R1 R2 R3
1 100 0.2 0.5 1 0.14 0.15 0.15 0.147 16.67
2 100 0.2 0.5 5 0.1 0.1 0.09 0.097 20.28
3 100 0.2 0.5 10 0.06 0.06 0.06 0.060 24.44
4 100 0.2 2.5 1 0.14 0.14 0.14 0.140 17.08
5 100 0.2 2.5 5 0.07 0.07 0.07 0.070 23.10
6 100 0.2 2.5 10 0.06 0.05 0.05 0.053 25.41
7 100 0.2 4.5 1 0.18 0.18 0.18 0.180 14.89
8 100 0.2 4.5 5 0.06 0.06 0.06 0.060 24.44
9 100 0.2 4.5 10 0.12 0.11 0.11 0.113 18.90
10 100 0.7 0.5 1 0.29 0.29 0.29 0.290 10.75
11 100 0.7 0.5 5 0.12 0.12 0.12 0.120 18.42
12 100 0.7 0.5 10 0.09 0.09 0.09 0.090 20.92
13 100 0.7 2.5 1 0.21 0.23 0.23 0.223 13.01
14 100 0.7 2.5 5 0.07 0.06 0.07 0.067 23.49
15 100 0.7 2.5 10 0.08 0.08 0.08 0.080 21.94
16 100 0.7 4.5 1 0.29 0.29 0.29 0.290 10.75
17 100 0.7 4.5 5 0.11 0.1 0.1 0.103 19.70
18 100 0.7 4.5 10 0.18 0.14 0.14 0.153 16.19
189
Run Input Variable Measurement Laser Power
A
Cutting speed
B
Assist gas Pressure
C
Standoff Distance
D
Replication Edge qualityMean
S/N Ratio
R1 R2 R3
19 100 1.2 0.5 1 0.24 0.28 0.28 0.267 11.45
20 100 1.2 0.5 5 0.13 0.13 0.13 0.130 17.72
21 100 1.2 0.5 10 Unsuccessful cutting
22 100 1.2 2.5 1 0.17 0.17 0.17 0.170 15.39
23 100 1.2 2.5 5 0.11 0.11 0.11 0.110 19.17
24 100 1.2 2.5 10 Unsuccessful cutting
25 100 1.2 4.5 1 0.25 0.25 0.25 0.250 12.04
26 100 1.2 4.5 5 0.1 0.12 0.1 0.107 19.39
27 100 1.2 4.5 10 Unsuccessful
28 300 0.2 0.5 1 0.15 0.15 0.15 0.150 16.48
29 300 0.2 0.5 5 0.09 0.11 0.09 0.097 20.23
30 300 0.2 0.5 10 0.09 0.13 0.11 0.110 19.03
31 300 0.2 2.5 1 0.07 0.07 0.07 0.070 23.10
32 300 0.2 2.5 5 0.03 0.04 0.04 0.037 28.61
33 300 0.2 2.5 10 0.05 0.05 0.05 0.050 26.02
34 300 0.2 4.5 1 0.14 0.15 0.14 0.143 16.87
35 300 0.2 4.5 5 0.09 0.08 0.08 0.083 21.56
36 300 0.2 4.5 10 0.08 0.07 0.07 0.073 22.67
37 300 0.7 0.5 1 0.11 0.11 0.11 0.110 19.17
38 300 0.7 0.5 5 0.11 0.11 0.11 0.110 19.17
39 300 0.7 0.5 10 0.07 0.07 0.07 0.070 23.10
40 300 0.7 2.5 1 0.17 0.17 0.17 0.170 15.39
41 300 0.7 2.5 5 0.1 0.09 0.09 0.093 20.58
42 300 0.7 2.5 10 0.1 0.1 0.1 0.100 20.00
43 300 0.7 4.5 1 0.22 0.22 0.22 0.220 13.15
44 300 0.7 4.5 5 0.08 0.09 0.08 0.083 21.56
45 300 0.7 4.5 10 0.11 0.11 0.11 0.110 19.17
46 300 1.2 0.5 1 0.23 0.23 0.23 0.230 12.77
47 300 1.2 0.5 5 0.08 0.08 0.08 0.080 21.94
48 300 1.2 0.5 10 0.1 0.1 0.1 0.100 20.00
49 300 1.2 2.5 1 0.21 0.21 0.21 0.210 13.56
50 300 1.2 2.5 5 0.07 0.07 0.07 0.070 23.10
190
Run Input Variable Measurement Laser Power
A
Cutting speed
B
Assist gas Pressure
C
Standoff Distance
D
Replication Edge qualityMean
S/N Ratio
R1 R2 R3
51 300 1.2 2.5 10 0.11 0.11 0.11 0.110 19.17
52 300 1.2 4.5 1 0.29 0.29 0.29 0.290 10.75
53 300 1.2 4.5 5 0.11 0.11 0.11 0.110 19.17
54 300 1.2 4.5 10 0.1 0.09 0.09 0.093 20.58
55 500 0.2 0.5 1 0.12 0.12 0.15 0.130 17.64
56 500 0.2 0.5 5 0.08 0.08 0.08 0.080 21.94
57 500 0.2 0.5 10 0.09 0.09 0.09 0.090 20.92
58 500 0.2 2.5 1 0.11 0.11 0.11 0.110 19.17
59 500 0.2 2.5 5 0.07 0.08 0.07 0.073 22.67
60 500 0.2 2.5 10 0.06 0.06 0.06 0.060 24.44
61 500 0.2 4.5 1 0.14 0.14 0.14 0.140 17.08
62 500 0.2 4.5 5 0.12 0.12 0.12 0.120 18.42
63 500 0.2 4.5 10 0.11 0.11 0.11 0.110 19.17
64 500 0.7 0.5 1 0.17 0.17 0.17 0.170 15.39
65 500 0.7 0.5 5 0.1 0.12 0.1 0.107 19.39
66 500 0.7 0.5 10 0.16 0.13 0.13 0.140 17.01
67 500 0.7 2.5 1 0.11 0.11 0.11 0.110 19.17
68 500 0.7 2.5 5 0.06 0.06 0.06 0.060 24.44
69 500 0.7 2.5 10 0.13 0.13 0.13 0.130 17.72
70 500 0.7 4.5 1 0.21 0.21 0.21 0.210 13.56
71 500 0.7 4.5 5 0.13 0.1 0.1 0.110 19.07
72 500 0.7 4.5 10 0.12 0.12 0.12 0.120 18.42
73 500 1.2 0.5 1 0.15 0.13 0.13 0.137 17.26
74 500 1.2 0.5 5 0.1 0.1 0.1 0.100 20.00
75 500 1.2 0.5 10 0.08 0.08 0.08 0.080 21.94
76 500 1.2 2.5 1 0.11 0.11 0.11 0.110 19.17
77 500 1.2 2.5 5 0.08 0.08 0.08 0.080 21.94
78 500 1.2 2.5 10 0.1 0.1 0.1 0.100 20.00
79 500 1.2 4.5 1 0.19 0.19 0.19 0.190 14.42
80 500 1.2 4.5 5 0.09 0.08 0.08 0.083 21.56
81 500 1.2 4.5 10 0.1 0.1 0.1 0.100 20.00
191
Figure A-1: Outlier analysis of edge quality of Perspex glass sheet (3mm)
Table A-5: Kerf width observations for of Perspex Sheet (3mm)
Input Variable Measurements Run Laser
Power (watts)
Cutting Speed
(m/min)
Assist Gas
pressure (bar)
Stand off
distance (mm)
Inner side line length (mm)
Replications
Outer scrap length (mm)
Replications Lin1 Lin2 Lin3 Lout1 Lout2 Lout3
1 100 0.2 0.5 1 20.76 20.77 20.79 19.67 19.66 19.66 2 100 0.2 0.5 5 20.82 20.82 20.82 19.5 19.52 19.5 3 100 0.2 0.5 10 21.3 21.35 21.35 19.23 19.23 19.23 4 100 0.2 2.5 1 20.75 20.76 20.76 19.65 19.64 19.65 5 100 0.2 2.5 5 20.88 20.82 20.87 19.49 19.49 19.49 6 100 0.2 2.5 10 21.24 21.24 21.26 19.04 19.03 19.04 7 100 0.2 4.5 1 20.7 20.71 20.76 19.78 19.78 19.78 8 100 0.2 4.5 5 20.82 20.89 20.89 19.59 19.61 19.6 9 100 0.2 4.5 10 21.28 21.22 21.24 19.19 19.2 19.21 10 100 0.7 0.5 1 20.57 20.6 20.63 19.65 19.68 19.69 11 100 0.7 0.5 5 20.68 20.72 20.72 19.6 19.61 19.6 12 100 0.7 0.5 10 20.89 20.94 20.93 19.41 19.42 19.42 13 100 0.7 2.5 1 20.49 20.52 20.54 19.78 19.79 19.78 14 100 0.7 2.5 5 20.72 20.79 20.73 19.63 19.63 19.63
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67 70 73 76 79
Edge
qua
lity
obse
rvat
ions
Runs
Replicate 3
Replicate 2
Replicate 1
192
Input Variable Measurements Run Laser
Power (watts)
Cutting Speed
(m/min)
Assist Gas
pressure (bar)
Stand off
distance (mm)
Inner side line length (mm)
Replications
Outer scrap length (mm)
Replications Lin1 Lin2 Lin3 Lout1 Lout2 Lout3
15 100 0.7 2.5 10 21.11 21.11 21.15 19.37 19.39 19.38 16 100 0.7 4.5 1 20.64 20.64 20.64 19.81 19.79 19.79 17 100 0.7 4.5 5 20.69 20.71 20.73 19.76 19.75 19.75 18 100 0.7 4.5 10 21.12 21.07 21.05 19.31 19.29 19.29 19 100 1.2 0.5 1 20.52 20.5 20.5 19.87 19.88 19.87 20 100 1.2 0.5 5 20.63 20.6 20.7 19.71 19.71 19.7 21 100 1.2 0.5 10 Unsuccessful cutting 22 100 1.2 2.5 1 20.56 20.52 20.53 19.9 19.93 19.89 23 100 1.2 2.5 5 20.65 20.68 20.69 19.75 19.74 19.73 24 100 1.2 2.5 10 Unsuccessful cutting 25 100 1.2 4.5 1 20.49 20.49 20.49 19.8 19.81 19.81 26 100 1.2 4.5 5 20.63 20.62 20.59 19.79 19.74 19.73 27 100 1.2 4.5 10 Unsuccessful cutting 28 300 0.2 0.5 1 21 21 21 19.37 19.37 19.37 29 300 0.2 0.5 5 21.18 21.18 21.18 19.13 19.12 19.12 30 300 0.2 0.5 10 21.47 21.45 21.42 18.89 18.89 18.9 31 300 0.2 2.5 1 20.93 20.91 20.99 19.56 19.55 19.57 32 300 0.2 2.5 5 21.17 21.16 21.18 19.26 19.22 19.22 33 300 0.2 2.5 10 21.67 21.65 21.63 18.74 18.74 18.73 34 300 0.2 4.5 1 20.84 20.85 20.86 19.48 19.51 19.48 35 300 0.2 4.5 5 21.21 21.21 21.21 19.38 19.38 19.38 36 300 0.2 4.5 10 21.53 21.54 21.54 18.85 18.85 18.85 37 300 0.7 0.5 1 20.8 20.83 20.79 19.58 19.58 19.58 38 300 0.7 0.5 5 20.97 20.99 21 19.38 19.38 19.38 39 300 0.7 0.5 10 21.27 21.25 21.26 19.02 19.05 19.02 40 300 0.7 2.5 1 20.82 20.79 20.81 19.62 19.62 19.62 41 300 0.7 2.5 5 21 20.98 20.98 19.41 19.4 19.41 42 300 0.7 2.5 10 21.43 21.44 21.44 18.97 18.97 18.97 43 300 0.7 4.5 1 20.82 20.82 20.8 19.63 19.63 19.63 44 300 0.7 4.5 5 20.94 20.95 20.99 19.44 19.44 19.44 45 300 0.7 4.5 10 21.35 21.33 21.32 19.02 19.03 19.03 46 300 1.2 0.5 1 20.81 20.82 20.82 19.67 19.67 19.67 47 300 1.2 0.5 5 20.85 20.85 20.85 19.46 19.46 19.46 48 300 1.2 0.5 10 21.09 21.11 21.09 19.32 19.32 19.32 49 300 1.2 2.5 1 20.71 20.69 20.67 19.72 19.72 19.72 50 300 1.2 2.5 5 20.91 20.9 20.89 19.4 19.42 19.4 51 300 1.2 2.5 10 21.34 21.34 21.34 19.08 19.08 19.08
193
Input Variable Measurements Run Laser
Power (watts)
Cutting Speed
(m/min)
Assist Gas
pressure (bar)
Stand off
distance (mm)
Inner side line length (mm)
Replications
Outer scrap length (mm)
Replications Lin1 Lin2 Lin3 Lout1 Lout2 Lout3
52 300 1.2 4.5 1 20.68 20.68 20.68 19.74 19.74 19.74 53 300 1.2 4.5 5 20.91 20.93 20.91 19.47 19.47 19.47 54 300 1.2 4.5 10 21.07 21.07 21.07 19.11 19.11 19.11 55 500 0.2 0.5 1 21.14 21.14 21.14 19.31 19.31 19.31 56 500 0.2 0.5 5 21.31 21.34 21.34 19.05 19.05 19.05 57 500 0.2 0.5 10 21.54 21.54 21.54 17.75 17.78 17.76 58 500 0.2 2.5 1 20.98 20.98 20.97 19.41 19.41 19.41 59 500 0.2 2.5 5 21.31 21.31 21.31 19.11 19.11 19.11 60 500 0.2 2.5 10 21.7 20.69 21.7 18.92 18.91 18.92 61 500 0.2 4.5 1 21.06 21.06 21.06 19.49 19.51 19.51 62 500 0.2 4.5 5 21.21 21.21 21.21 19.24 19.24 19.24 63 500 0.2 4.5 10 21.63 21.65 21.65 18.72 18.71 18.71 64 500 0.7 0.5 1 20.87 20.87 20.87 19.49 19.51 19.49 65 500 0.7 0.5 5 21 21.03 21.03 19.28 19.28 19.28 66 500 0.7 0.5 10 21.4 21.4 21.4 18.94 18.94 18.94 67 500 0.7 2.5 1 20.89 20.93 20.91 19.57 19.59 19.57 68 500 0.7 2.5 5 21.09 21.09 21.1 19.31 19.33 19.33 69 500 0.7 2.5 10 21.52 21.52 21.52 18.85 18.85 18.85 70 500 0.7 4.5 1 20.84 20.84 20.84 19.49 19.51 19.49 71 500 0.7 4.5 5 21.04 21.04 21.04 19.43 19.43 19.43 72 500 0.7 4.5 10 21.38 21.41 21.4 18.94 18.99 18.98 73 500 1.2 0.5 1 20.81 20.81 20.76 19.59 19.61 19.6 74 500 1.2 0.5 5 20.98 20.98 20.98 19.44 19.44 19.45 75 500 1.2 0.5 10 21.26 21.22 21.26 19.19 19.2 19.18 76 500 1.2 2.5 1 20.79 20.81 20.81 19.62 19.62 19.59 77 500 1.2 2.5 5 20.95 20.92 20.93 19.44 19.44 19.44 78 500 1.2 2.5 10 21.41 21.46 21.45 18.98 18.97 18.98 79 500 1.2 4.5 1 20.73 20.72 20.71 19.72 19.72 19.72 80 500 1.2 4.5 5 21.09 21 21.05 19.38 19.38 19.38 81 500 1.2 4.5 10 21.29 21.31 21.3 19.1 19.11 19.09
194
Figure A-2: Outlier analysis of Kerf width observations of Perspex glass sheet (3mm)
Table A-6: Kerf width mean and signal to noise ratio of Perspex Sheet (3mm)
Input Variable Measurements
Run Laser Power (watts)
Cutting Speed (m/min)
Assist Gas pressure (bar)
Standoff distance (mm)
Mean S/N Ratio
1 100 0.2 0.5 1 0.56 5.11 2 100 0.2 0.5 5 0.66 3.65 3 100 0.2 0.5 10 1.05 -0.44 4 100 0.2 2.5 1 0.56 5.11 5 100 0.2 2.5 5 0.68 3.30 6 100 0.2 2.5 10 1.11 -0.87 7 100 0.2 4.5 1 0.47 6.52 8 100 0.2 4.5 5 0.63 3.96 9 100 0.2 4.5 10 1.02 -0.20 10 100 0.7 0.5 1 0.46 6.68 11 100 0.7 0.5 5 0.55 5.16 12 100 0.7 0.5 10 0.75 2.48 13 100 0.7 2.5 1 0.37 8.71 14 100 0.7 2.5 5 0.56 5.06
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67 70 73 76 79
Kerf
wid
th o
bser
vati
ons
Runs
Replication 1
Replication 2
Replication 3
195
Input Variable Measurements
Run Laser Power (watts)
Cutting Speed (m/min)
Assist Gas pressure (bar)
Standoff distance (mm)
Mean S/N Ratio
15 100 0.7 2.5 10 0.87 1.19 16 100 0.7 4.5 1 0.42 7.50 17 100 0.7 4.5 5 0.48 6.40 18 100 0.7 4.5 10 0.89 1.00 19 100 1.2 0.5 1 0.32 9.99 20 100 1.2 0.5 5 0.47 6.57 21 100 1.2 0.5 10 0.00 0.00 22 100 1.2 2.5 1 0.32 10.02 23 100 1.2 2.5 5 0.47 6.62 24 100 1.2 2.5 10 0.00 0.00 25 100 1.2 4.5 1 0.34 9.33 26 100 1.2 4.5 5 0.43 7.33 27 100 1.2 4.5 10 0.00 0.00 28 300 0.2 0.5 1 0.82 1.78 29 300 0.2 0.5 5 1.03 -0.24 30 300 0.2 0.5 10 1.28 -2.12 31 300 0.2 2.5 1 0.69 3.20 32 300 0.2 2.5 5 0.97 0.28 33 300 0.2 2.5 10 1.46 -3.27 34 300 0.2 4.5 1 0.68 3.35 35 300 0.2 4.5 5 0.92 0.77 36 300 0.2 4.5 10 1.34 -2.56 37 300 0.7 0.5 1 0.61 4.24 38 300 0.7 0.5 5 0.80 1.90 39 300 0.7 0.5 10 1.12 -0.95 40 300 0.7 2.5 1 0.59 4.53 41 300 0.7 2.5 5 0.79 2.05 42 300 0.7 2.5 10 1.23 -1.82 43 300 0.7 4.5 1 0.59 4.56 44 300 0.7 4.5 5 0.76 2.38 45 300 0.7 4.5 10 1.15 -1.24 46 300 1.2 0.5 1 0.57 4.83 47 300 1.2 0.5 5 0.70 3.16 48 300 1.2 0.5 10 0.89 1.03 49 300 1.2 2.5 1 0.49 6.28 50 300 1.2 2.5 5 0.75 2.54 51 300 1.2 2.5 10 1.13 -1.06 52 300 1.2 4.5 1 0.47 6.56
196
Input Variable Measurements
Run Laser Power (watts)
Cutting Speed (m/min)
Assist Gas pressure (bar)
Standoff distance (mm)
Mean S/N Ratio
53 300 1.2 4.5 5 0.72 2.81 54 300 1.2 4.5 10 0.98 0.18 55 500 0.2 0.5 1 0.92 0.77 56 500 0.2 0.5 5 1.14 -1.14 57 500 0.2 0.5 10 1.89 -5.52 58 500 0.2 2.5 1 0.78 2.12 59 500 0.2 2.5 5 1.10 -0.83 60 500 0.2 2.5 10 1.22 -1.99 61 500 0.2 4.5 1 0.78 2.18 62 500 0.2 4.5 5 0.99 0.13 63 500 0.2 4.5 10 1.47 -3.32 64 500 0.7 0.5 1 0.69 3.26 65 500 0.7 0.5 5 0.87 1.21 66 500 0.7 0.5 10 1.23 -1.80 67 500 0.7 2.5 1 0.67 3.52 68 500 0.7 2.5 5 0.89 1.06 69 500 0.7 2.5 10 1.34 -2.51 70 500 0.7 4.5 1 0.67 3.46 71 500 0.7 4.5 5 0.81 1.88 72 500 0.7 4.5 10 1.21 -1.68 73 500 1.2 0.5 1 0.60 4.48 74 500 1.2 0.5 5 0.77 2.29 75 500 1.2 0.5 10 1.03 -0.24 76 500 1.2 2.5 1 0.60 4.48 77 500 1.2 2.5 5 0.75 2.54 78 500 1.2 2.5 10 1.23 -1.81 79 500 1.2 4.5 1 0.50 6.02 80 500 1.2 4.5 5 0.83 1.58 81 500 1.2 4.5 10 1.10 -0.83
197
Table A-7: Edge quality observations of Perspex Sheet (5mm)
Run Input Variable Measurement Laser Power
A
Cutting speed
B
Assist gas Pressure
C
Standoff Distance
D
Replication Mean S/N Ratio R1 R2 R3
1 100 0.2 0.5 1 0.2 0.19 0.19 0.19 14.27
2 100 0.2 0.5 5 0.08 0.08 0.09 0.08 21.56
3 100 0.2 0.5 10 0.13 0.1 0.1 0.11 19.07
4 100 0.2 2.5 1 0.11 0.11 0.11 0.11 19.17
5 100 0.2 2.5 5 0.09 0.09 0.09 0.09 20.92
6 100 0.2 2.5 10 0.11 0.11 0.11 0.11 19.17
7 100 0.2 4.5 1 0.14 0.14 0.14 0.14 17.08
8 100 0.2 4.5 5 0.07 0.07 0.07 0.07 23.10
9 100 0.2 4.5 10 0.09 0.09 0.09 0.09 20.92
10 100 0.7 0.5 1 0.2 0.2 0.18 0.19 14.26
11 100 0.7 0.5 5 Unsuccessful cutting
12 100 0.7 0.5 10 Unsuccessful cutting
13 100 0.7 2.5 1 0.23 0.21 0.23 0.22 13.01
14 100 0.7 2.5 5 0.06 0.06 0.06 0.06 24.44
15 100 0.7 2.5 10 Unsuccessful cutting
16 100 0.7 4.5 1 0.21 0.19 0.19 0.20
17 100 0.7 4.5 5 0.09 0.09 0.09 0.09
18 100 0.7 4.5 10 Unsuccessful cutting
19 100 1.2 0.5 1 Unsuccessful cutting
20 100 1.2 0.5 5 Unsuccessful cutting
21 100 1.2 0.5 10 Unsuccessful cutting
22 100 1.2 2.5 1 Unsuccessful cutting
23 100 1.2 2.5 5 Unsuccessful cutting
24 100 1.2 2.5 10 Unsuccessful cutting
25 100 1.2 4.5 1 Unsuccessful cutting
26 100 1.2 4.5 5 Unsuccessful cutting
27 100 1.2 4.5 10 Unsuccessful cutting
28 300 0.2 0.5 1 0.14 0.16 0.14 0.15 16.65
29 300 0.2 0.5 5 0.11 0.1 0.1 0.10 19.70
30 300 0.2 0.5 10 0.14 0.14 0.16 0.15 16.65
31 300 0.2 2.5 1 0.09 0.09 0.09 0.09 20.92
198
Run Input Variable Measurement Laser Power
A
Cutting speed
B
Assist gas Pressure
C
Standoff Distance
D
Replication Mean S/N Ratio R1 R2 R3
32 300 0.2 2.5 5 0.05 0.05 0.05 0.05 26.02
33 300 0.2 2.5 10 0.06 0.06 0.06 0.06 24.44
34 300 0.2 4.5 1 0.07 0.07 0.07 0.07 23.10
35 300 0.2 4.5 5 0.07 0.07 0.07 0.07 23.10
36 300 0.2 4.5 10 0.06 0.06 0.06 0.06 24.44
37 300 0.7 0.5 1 0.14 0.14 0.14 0.14 17.08
38 300 0.7 0.5 5 0.09 0.09 0.09 0.09 20.92
39 300 0.7 0.5 10 0.11 0.11 0.11 0.11 19.17
40 300 0.7 2.5 1 0.19 0.19 0.19 0.19 14.42
41 300 0.7 2.5 5 0.07 0.07 0.07 0.07 23.10
42 300 0.7 2.5 10 0.1 0.1 0.1 0.10 20.00
43 300 0.7 4.5 1 0.07 0.07 0.07 0.07 23.10
44 300 0.7 4.5 5 0.05 0.05 0.05 0.05 26.02
45 300 0.7 4.5 10 0.08 0.08 0.08 0.08 21.94
46 300 1.2 0.5 1 0.24 0.24 0.24 0.24 12.40
47 300 1.2 0.5 5 Unsuccessful cutting
48 300 1.2 0.5 10 Unsuccessful cutting
49 300 1.2 2.5 1 0.27 0.27 0.27 0.27 11.37
50 300 1.2 2.5 5 0.09 0.09 0.09 0.09 20.92
51 300 1.2 2.5 10 Unsuccessful cutting
52 300 1.2 4.5 1 0.15 0.15 0.15 0.15 16.48
53 300 1.2 4.5 5 Unsuccessful cutting
54 300 1.2 4.5 10 Unsuccessful cutting
55 500 0.2 0.5 1 0.12 0.13 0.12 0.12 18.17
56 500 0.2 0.5 5 0.1 0.1 0.1 0.10 20.00
57 500 0.2 0.5 10 0.15 0.13 0.14 0.14 17.06
58 500 0.2 2.5 1 0.12 0.12 0.12 0.12 18.42
59 500 0.2 2.5 5 0.08 0.08 0.08 0.08 21.94
60 500 0.2 2.5 10 0.07 0.07 0.07 0.07 23.10
61 500 0.2 4.5 1 0.11 0.11 0.11 0.11 19.17
62 500 0.2 4.5 5 0.08 0.08 0.08 0.08 21.94
63 500 0.2 4.5 10 0.1 0.1 0.1 0.10 20.00
199
Run Input Variable Measurement Laser Power
A
Cutting speed
B
Assist gas Pressure
C
Standoff Distance
D
Replication Mean S/N Ratio R1 R2 R3
64 500 0.7 0.5 1 0.17 0.16 0.16 0.16 15.73
65 500 0.7 0.5 5 0.08 0.08 0.07 0.08 22.28
66 500 0.7 0.5 10 0.11 0.11 0.11 0.11 19.17
67 500 0.7 2.5 1 0.18 0.18 0.18 0.18 14.89
68 500 0.7 2.5 5 0.07 0.07 0.07 0.07 23.10
69 500 0.7 2.5 10 0.11 0.11 0.11 0.11 19.17
70 500 0.7 4.5 1 0.1 0.09 0.09 0.09 20.58
71 500 0.7 4.5 5 0.08 0.08 0.08 0.08 21.94
72 500 0.7 4.5 10 0.12 0.12 0.12 0.12 18.42
73 500 1.2 0.5 1 0.2 0.21 0.19 0.20 13.97
74 500 1.2 0.5 5 Unsuccessful cutting
75 500 1.2 0.5 10 Unsuccessful cutting
76 500 1.2 2.5 1 0.17 0.17 0.17 0.17 15.39
77 500 1.2 2.5 5 Unsuccessful cutting
78 500 1.2 2.5 10 Unsuccessful cutting
79 500 1.2 4.5 1 0.13 0.12 0.12 0.12 18.17
80 500 1.2 4.5 5 Unsuccessful cutting
81 500 1.2 4.5 10 Unsuccessful cutting
200
Figure A-3: Outlier analysis of Edge quality observations of Perspex glass sheet (5mm)
Table A-8: Observation for Kerf width of Perspex Sheet of 5mm
Input Variable Measurements
Run Laser Power (watts)
Cutting Speed
(m/min)
Assist Gas
pressure (bar)
Stand off
distance (mm)
Inner side line length (mm)
Replications
Outer scrap length (mm) Replications
Lin1 Lin2 Lin3 Lout1 Lout2 Lout3 1 100 0.2 0.5 1 20.73 20.75 20.72 19.63 19.62 19.62 2 100 0.2 0.5 5 20.82 20.81 20.84 19.41 19.39 19.38 3 100 0.2 0.5 10 21.1 21.11 21.12 19.15 19.17 19.18 4 100 0.2 2.5 1 20.68 20.72 20.71 19.7 19.71 19.73 5 100 0.2 2.5 5 20.8 20.83 20.82 19.49 19.47 19.46 6 100 0.2 2.5 10 21.21 21.23 21.21 19.12 19.1 19.14 7 100 0.2 4.5 1 20.71 20.73 20.7 19.75 19.73 19.73 8 100 0.2 4.5 5 20.89 20.9 20.87 19.57 19.56 19.59 9 100 0.2 4.5 10 21.21 21.19 21.2 19.13 19.1 19.11 10 100 0.7 0.5 1 20.46 20.48 20.45 19.89 19.86 19.88 11 100 0.7 0.5 5 Unsuccessful cutting 12 100 0.7 0.5 10 Unsuccessful cutting 13 100 0.7 2.5 1 20.57 20.55 20.59 19.9 19.89 19.87
0
0.05
0.1
0.15
0.2
0.25
0.3
1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67 70 73 76 79
Edge
qua
lity
obse
rvat
ions
Runs
Replication 1
Replication 2
Replication 3
201
Input Variable Measurements
Run Laser Power (watts)
Cutting Speed
(m/min)
Assist Gas
pressure (bar)
Stand off
distance (mm)
Inner side line length (mm)
Replications
Outer scrap length (mm) Replications
Lin1 Lin2 Lin3 Lout1 Lout2 Lout3 14 100 0.7 2.5 5 20.65 20.67 20.68 19.76 19.76 19.76 15 100 0.7 2.5 10 Unsuccessful cutting 16 100 0.7 4.5 1 20.55 20.59 20.57 19.8 19.82 19.81 17 100 0.7 4.5 5 20.7 20.68 20.71 19.83 19.86 19.84 18 100 0.7 4.5 10 Unsuccessful cutting 19 100 1.2 0.5 1 Unsuccessful cutting 20 100 1.2 0.5 5 Unsuccessful cutting 21 100 1.2 0.5 10 Unsuccessful cutting 22 100 1.2 2.5 1 Unsuccessful cutting 23 100 1.2 2.5 5 Unsuccessful cutting 24 100 1.2 2.5 10 Unsuccessful cutting 25 100 1.2 4.5 1 Unsuccessful cutting 26 100 1.2 4.5 5 Unsuccessful cutting 27 100 1.2 4.5 10 Unsuccessful cutting 28 300 0.2 0.5 1 20.93 20.95 20.96 19.36 19.34 19.33 29 300 0.2 0.5 5 21 21.03 21 19.14 19.11 19.15 30 300 0.2 0.5 10 21.46 21.45 21.48 18.71 18.69 18.67 31 300 0.2 2.5 1 21 20.98 21.03 19.57 19.56 19.54 32 300 0.2 2.5 5 21.11 21.13 21.11 19.25 19.23 19.21 33 300 0.2 2.5 10 21.56 21.55 21.54 18.69 18.69 18.71 34 300 0.2 4.5 1 20.97 20.99 20.96 19.42 19.4 19.43 35 300 0.2 4.5 5 21.26 21.23 21.24 19.15 19.11 19.13 36 300 0.2 4.5 10 21.56 21.57 21.59 18.6 18.65 18.63 37 300 0.7 0.5 1 20.77 20.78 20.79 19.57 19.59 19.6 38 300 0.7 0.5 5 20.86 20.88 20.88 19.43 19.46 19.47 39 300 0.7 0.5 10 21.14 21.11 21.1 19.21 19.23 19.2 40 300 0.7 2.5 1 20.76 20.77 20.79 19.74 19.71 19.72 41 300 0.7 2.5 5 20.93 20.91 20.9 19.57 19.53 19.55 42 300 0.7 2.5 10 21.38 21.36 21.35 19.17 19.15 19.16 43 300 0.7 4.5 1 20.74 20.77 20.76 19.56 19.54 19.54 44 300 0.7 4.5 5 20.95 20.94 20.92 19.4 19.42 19.43 45 300 0.7 4.5 10 21.3 21.35 21.38 19.09 19.07 19.06 46 300 1.2 0.5 1 20.6 20.61 20.63 19.78 19.79 19.81 47 300 1.2 0.5 5 Unsuccessful cutting 48 300 1.2 0.5 10 Unsuccessful cutting 49 300 1.2 2.5 1 20.69 20.7 20.7 19.85 19.86 19.88
202
Input Variable Measurements
Run Laser Power (watts)
Cutting Speed
(m/min)
Assist Gas
pressure (bar)
Stand off
distance (mm)
Inner side line length (mm)
Replications
Outer scrap length (mm) Replications
Lin1 Lin2 Lin3 Lout1 Lout2 Lout3 50 300 1.2 2.5 5 20.85 20.87 20.87 19.64 19.67 19.66 51 300 1.2 2.5 10 Unsuccessful cutting 52 300 1.2 4.5 1 20.72 20.71 20.74 19.9 19.87 19.88 53 300 1.2 4.5 5 Unsuccessful cutting 54 300 1.2 4.5 10 Unsuccessful cutting 55 500 0.2 0.5 1 21.13 21.11 21.1 19.22 19.24 19.25 56 500 0.2 0.5 5 21.28 21.29 21.31 18.97 18.99 18.99 57 500 0.2 0.5 10 21.58 21.56 21.59 18.6 18.62 18.64 58 500 0.2 2.5 1 21.12 21.14 21.15 19.37 19.39 19.4 59 500 0.2 2.5 5 20.84 20.83 20.86 19.63 19.6 19.61 60 500 0.2 2.5 10 21.53 21.51 21.5 18.67 18.69 18.68 61 500 0.2 4.5 1 21.05 21.02 21.03 19.32 19.3 19.3 62 500 0.2 4.5 5 21.26 21.28 21.29 19.09 19.08 19.06 63 500 0.2 4.5 10 21.64 21.65 21.67 18.59 18.6 18.6 64 500 0.7 0.5 1 20.87 20.85 20.89 19.45 19.43 19.48 65 500 0.7 0.5 5 21.02 21.04 21.01 19.36 19.38 19.39 66 500 0.7 0.5 10 21.25 21.24 21.22 19.04 19.06 19.08 67 500 0.7 2.5 1 20.85 20.83 20.86 19.72 19.74 19.76 68 500 0.7 2.5 5 20.9 20.92 20.93 19.48 19.44 19.46 69 500 0.7 2.5 10 21.33 21.31 21.34 18.94 18.95 18.96 70 500 0.7 4.5 1 20.8 20.82 20.84 19.5 19.52 19.53 71 500 0.7 4.5 5 21.09 21.11 21.09 19.27 19.24 19.25 72 500 0.7 4.5 10 21.43 21.41 21.4 19.04 19.02 19.01 73 500 1.2 0.5 1 20.76 20.78 20.79 19.75 19.74 19.74 74 500 1.2 0.5 5 Unsuccessful cutting 75 500 1.2 0.5 10 Unsuccessful cutting 76 500 1.2 2.5 1 20.72 20.74 20.75 19.73 19.76 19.75 77 500 1.2 2.5 5 Unsuccessful cutting 78 500 1.2 2.5 10 Unsuccessful cutting 79 500 1.2 4.5 1 20.78 20.79 20.76 19.67 19.68 19.7 80 500 1.2 4.5 5 Unsuccessful cutting 81 500 1.2 4.5 10 Unsuccessful cutting
203
Figure A-4: Outlier analysis of Kerf width observations of Perspex glass sheet (5mm)
Table A-9: Kerf width mean and signal to noise ratio of Perspex Sheet of 5mm
Input Variable Measurements
Run Laser Power (watts)
Cutting Speed
(m/min)
Assist Gas pressure
(bar)
Standoff distance
(mm)
Mean S/N Ratio
1 100 0.2 0.5 1 0.56 5.11 2 100 0.2 0.5 5 0.72 2.91 3 100 0.2 0.5 10 0.97 0.25 4 100 0.2 2.5 1 0.50 6.11 5 100 0.2 2.5 5 0.67 3.45 6 100 0.2 2.5 10 1.05 -0.41 7 100 0.2 4.5 1 0.49 6.22 8 100 0.2 4.5 5 0.66 3.65 9 100 0.2 4.5 10 1.04 -0.37 10 100 0.7 0.5 1 0.29 10.64 11 100 0.7 0.5 5 Unsuccessful cutting 12 100 0.7 0.5 10 Unsuccessful cutting 13 100 0.7 2.5 1 0.34 9.32 14 100 0.7 2.5 5 0.45 6.87
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
1 4 7 101316192225283134374043464952555861646770737679
Ker
wid
th o
bser
vati
ons
Runs
Replication 1
Replication 2
Replication 3
204
Input Variable Measurements
Run Laser Power (watts)
Cutting Speed
(m/min)
Assist Gas pressure
(bar)
Standoff distance
(mm)
Mean S/N Ratio
15 100 0.7 2.5 10 Unsuccessful cutting 16 100 0.7 4.5 1 0.38 8.40 17 100 0.7 4.5 5 0.43 7.39 18 100 0.7 4.5 10 Unsuccessful cutting 19 100 1.2 0.5 1 Unsuccessful cutting 20 100 1.2 0.5 5 Unsuccessful cutting 21 100 1.2 0.5 10 Unsuccessful cutting 22 100 1.2 2.5 1 Unsuccessful cutting 23 100 1.2 2.5 5 Unsuccessful cutting 24 100 1.2 2.5 10 Unsuccessful cutting 25 100 1.2 4.5 1 Unsuccessful cutting 26 100 1.2 4.5 5 Unsuccessful cutting 27 100 1.2 4.5 10 Unsuccessful cutting 28 300 0.2 0.5 1 0.80 1.92 29 300 0.2 0.5 5 0.94 0.55 30 300 0.2 0.5 10 1.39 -2.84 31 300 0.2 2.5 1 0.72 2.81 32 300 0.2 2.5 5 0.94 0.51 33 300 0.2 2.5 10 1.43 -3.09 34 300 0.2 4.5 1 0.78 2.18 35 300 0.2 4.5 5 1.06 -0.48 36 300 0.2 4.5 10 1.47 -3.37 37 300 0.7 0.5 1 0.60 4.49 38 300 0.7 0.5 5 0.71 2.97 39 300 0.7 0.5 10 0.95 0.43 40 300 0.7 2.5 1 0.53 5.59 41 300 0.7 2.5 5 0.68 3.33 42 300 0.7 2.5 10 1.10 -0.84 43 300 0.7 4.5 1 0.61 4.36 44 300 0.7 4.5 5 0.76 2.38 45 300 0.7 4.5 10 1.14 -1.10 46 300 1.2 0.5 1 0.41 7.74 47 300 1.2 0.5 5 Unsuccessful cutting 48 300 1.2 0.5 10 Unsuccessful cutting 49 300 1.2 2.5 1 0.42 7.60 50 300 1.2 2.5 5 0.60 4.39 51 300 1.2 2.5 10 Unsuccessful cutting
205
Input Variable Measurements
Run Laser Power (watts)
Cutting Speed
(m/min)
Assist Gas pressure
(bar)
Standoff distance
(mm)
Mean S/N Ratio
52 300 1.2 4.5 1 0.42 7.53 53 300 1.2 4.5 5 Unsuccessful cutting 54 300 1.2 4.5 10 Unsuccessful cutting 55 500 0.2 0.5 1 0.94 0.55 56 500 0.2 0.5 5 1.16 -1.25 57 500 0.2 0.5 10 1.48 -3.40 58 500 0.2 2.5 1 0.88 1.16 59 500 0.2 2.5 5 0.62 4.22 60 500 0.2 2.5 10 1.42 -3.03 61 500 0.2 4.5 1 0.86 1.28 62 500 0.2 4.5 5 1.10 -0.83 63 500 0.2 4.5 10 1.53 -3.68 64 500 0.7 0.5 1 0.71 3.00 65 500 0.7 0.5 5 0.82 1.69 66 500 0.7 0.5 10 1.09 -0.74 67 500 0.7 2.5 1 0.55 5.14 68 500 0.7 2.5 5 0.73 2.75 69 500 0.7 2.5 10 1.19 -1.50 70 500 0.7 4.5 1 0.65 3.72 71 500 0.7 4.5 5 0.92 0.71 72 500 0.7 4.5 10 1.20 -1.55 73 500 1.2 0.5 1 0.52 5.73 74 500 1.2 0.5 5 Unsuccessful cutting 75 500 1.2 0.5 10 Unsuccessful cutting 76 500 1.2 2.5 1 0.49 6.11 77 500 1.2 2.5 5 0.00 0.00 78 500 1.2 2.5 10 0.00 0.00 79 500 1.2 4.5 1 0.55 5.24 80 500 1.2 4.5 5 Unsuccessful cutting 81 500 1.2 4.5 10 Unsuccessful cutting
206
A.1 ONE WAY ANOVA WITHOUT REPLICATION Kerf width observations of Polystyrene foam Sheet (13mm) used for analysis Table A-3. Table A-10: Observations consider Laser Power (A)
100 300 500 1.573 1.913 1.968 1.488 1.842 2.068 1.738 1.742 1.920
Table A-11: Summary of descriptive Statistics
Groups Count Sum Average Variance 100 3 4.800 1.600 0.016 300 3 5.497 1.832 0.007 500 3 5.957 1.986 0.006
Table A-12: ANOVA for Laser Power Source of Variation SS df MS F P-value F crit Between Groups 0.226 2 0.113 11.568 0.009 5.143 Within Groups 0.059 6 0.010 Total 0.285 8
Table A-13: Observations consider Cutting speed
0.2 0.7 1.2 1.573 1.488 1.738 1.913 1.842 1.742 1.968 2.068 1.920
Table A-14: Summary of descriptive Statistics
Groups Count Sum Average Variance 0.2 3 5.455 1.818 0.046 0.7 3 5.398 1.799 0.085 1.2 3 5.400 1.800 0.011
Table A-15: ANOVA for Cutting Speed Source of Variation SS df MS F P-value F critical Between Groups 0.001 2 0.000 0.007 0.993 5.143 Within Groups 0.284 6 0.047 Total 0.285 8
207
Table A-16: Observations consider Assist gas pressure 0.5 2.5 4.5
1.573 1.488 1.738 1.742 1.913 1.842 2.068 1.920 1.968
Table A-17: Summary of Descriptive Statistics
Groups Count Sum Average Variance 0.5 3 5.383 1.794 0.063 2.5 3 5.322 1.774 0.061 4.5 3 5.548 1.849 0.013
Table A-18: ANOVA for Assist Gas Pressure Source of Variation SS df MS F P-value F critical Between Groups 0.009 2 0.005 0.0997 0.907 5.143 Within Groups 0.276 6 0.046 Total 0.285 8
Table A-19: Observations consider Standoff Distance
1 5 10 1.573 1.488 1.738 1.842 1.742 1.913 1.920 1.968 2.068
Table A-20: Summary of Descriptive Statistics
Groups Count Sum Average Variance 1 3 5.335 1.778 0.033 5 3 5.198 1.733 0.058 10 3 5.720 1.907 0.027
Table A-21: ANOVA for Standoff Distance Source of Variation SS df MS F P-value F critical Between Groups 0.049 2 0.024 0.620 0.569 5.143 Within Groups 0.236 6 0.039 Total 0.285 8
A.2 ONE WAY ANOVA WITH REPLICATION Table A-22: Observations consider Laser Power with replication
S. No. 100 300 500 S. No. 100 300 500 1 1.580 1.940 2.010 6 1.480 1.985 2.290 2 1.615 1.885 2.040 7 1.660 1.655 1.790 3 1.525 1.915 1.855 8 1.695 1.715 1.890 4 1.730 1.765 1.975 9 1.860 1.855 2.080 5 1.255 1.775 1.940
208
Table A-23: Summary of Descriptive Statistics Groups Count Sum Average Variance
100 9 14.400 1.600 0.030 300 9 16.490 1.832 0.012 500 9 17.870 1.986 0.021
Table A-24: One way ANOVA with replication for Laser Power Source of Variation SS Df MS F P-value F critical Between Groups 0.678 2 0.339 16.066 0.000 3.403 Within Groups 0.507 24 0.021 Total 1.185 26
Table A-25: Observations consider Cutting Speed with replication
S. No. 0.2 0.7 1.2 S. No. 0.2 0.7 1.2 1 1.580 1.730 1.660 6 1.915 1.985 1.855 2 1.615 1.255 1.695 7 2.010 1.975 1.790 3 1.525 1.480 1.860 8 2.040 1.940 1.890 4 1.940 1.765 1.655 9 1.855 2.290 2.080 5 1.885 1.775 1.715
Table A-26: Summary of Descriptive Statistics
Groups Count Sum Average Variance 0.2 9 16.365 1.818 0.037 0.7 9 16.195 1.799 0.091 1.2 9 16.200 1.800 0.019
Table A-27: One way ANOVA with replication for Cutting Speed Source of Variation SS Df MS F P-value F critical Between Groups 0.002 2 0.001 0.021 0.979 3.403 Within Groups 1.183 24 0.049
Table A-28: Observations consider Assist gas pressure with replication
S. No. 0.5 2.5 4.5 S. No. 0.5 2.5 4.5 1 1.580 1.730 1.660 6 1.855 1.915 1.985 2 1.615 1.255 1.695 7 1.975 1.790 2.010 3 1.525 1.480 1.860 8 1.940 1.890 2.040 4 1.655 1.940 1.765 9 2.290 2.080 1.855 5 1.715 1.885 1.775
209
Table A-29: Summary of Descriptive Statistics Groups Count Sum Average Variance
0.5 9 16.150 1.794 0.060 2.5 9 15.965 1.774 0.066 4.5 9 16.645 1.849 0.019
Table A-30: One way ANOVA with replication for Assist Gas Pressure Source of Variation SS Df MS F P-value F critical Between Groups 0.027 2 0.014 0.285 0.755 3.403 Within Groups 1.157 24 0.048 Total 1.185 26
Table A-31: Observations consider Standoff Distance with replication
S. No. 1 5 10 S. No. 0.5 2.5 4.5 1 1.580 1.730 1.660 6 1.985 1.855 1.915 2 1.615 1.255 1.695 7 1.790 2.010 1.975 3 1.525 1.480 1.860 8 1.890 2.040 1.940 4 1.765 1.655 1.940 9 2.080 1.855 2.290 5 1.775 1.715 1.885
Table A-32: Summary of Descriptive Statistics
Groups Count Sum Average Variance 1 9 16.005 1.778 0.035 5 9 15.595 1.733 0.062 10 9 17.160 1.907 0.033
Table A-33: One way ANOVA with replication for Standoff Distance Source of Variation SS df MS F P-value F crit Between Groups 0.146 2 0.073 1.691 0.206 3.403 Within Groups 1.039 24 0.043 Total 1.185 26
Table A-34: One way ANOVA with replication Treatments F P-value F critical Laser Power 16.066 0.000 3.403 Cutting Speed 0.021 0.979 3.403 Assist gas pressure 0.285 0.755 3.403 Standoff distance 1.691 0.206 3.403
210
A.3 TWO WAY ANOVA WITH REPLICATION Table A-35: Interaction between Laser Power and Cutting Speed with replication
A/B 100 300 500 0.2 1.580 1.940 2.010
1.615 1.885 2.040 1.525 1.915 1.855
0.7 1.730 1.765 1.975 1.255 1.775 1.940 1.480 1.985 2.290
1.2 1.660 1.655 1.79 1.695 1.715 1.89 1.860 1.855 2.08
Table A-36: Interaction between Laser Power and Cutting speed 0.2 100/0.2 300/0.2 500/0.2 Total Count 3 3 3 9 Sum 4.720 5.740 5.905 16.365 Average 1.573 1.913 1.968 1.818 Variance 0.002 0.001 0.010 0.037
Table A-37: Interaction between Laser Power and Cutting speed 0.7 100/0.7 300/0.7 500/0.7 Total Count 3 3 3 9 Sum 4.465 5.525 6.205 16.195 Average 1.488 1.842 2.068 1.799 Variance 0.056 0.015 0.037 0.091
Table A-38: Interaction between Laser Power and Cutting speed 1.2 100/1.2 300/1.2 500/1.2 Total Count 3 3 3 9 Sum 5.215 5.225 5.760 16.200 Average 1.738 1.742 1.920 1.800 Variance 0.011 0.011 0.022 0.019
Table A-39: Total Interaction between Laser Power and Cutting speed Count 9 9 9 Sum 14.400 16.490 17.870 Average 1.600 1.832 1.986 Variance 0.030 0.012 0.021
211
Table A-40: ANOVA of Interaction between Laser Power and Cutting speed Source of Variation SS df MS F P-value F crit
Sample 0.002 2 0.001 0.057 0.945 3.555 Columns 0.678 2 0.339 18.457 4.37x10-5 3.555 Interaction 0.174 4 0.043 2.365 0.092 2.928 Within 0.331 18 0.018 Total 1.185 26
Table A-41: Interaction between Laser Power and Assist Gas Pressure with replication
A/C 100 300 500 0.5 1.580 1.655 1.975
1.615 1.715 1.940 1.525 1.855 2.290
2.5 1.730 1.940 1.790 1.255 1.885 1.890 1.480 1.915 2.080
4.5 1.660 1.765 2.010 1.695 1.775 2.040 1.860 1.985 1.855
Table A-42: Interaction between Laser Power and Assist Gas Pressure 0.5
100/0.5 300/0.5 500/0.5 Total Count 3 3 3 9 Sum 4.720 5.225 6.205 16.150
Average 1.573 1.742 2.068 1.794 Variance 0.002 0.011 0.037 0.060
Table A-43: Interaction between Laser Power and Assist Gas Pressure 2.5
100/2.5 300/2.5 500/2.5 Total Count 3 3 3 9 Sum 4.465 5.740 5.760 15.965
Average 1.488 1.913 1.920 1.774 Variance 0.056 0.001 0.022 0.066
Table A-44: Interaction between Laser Power and Assist Gas Pressure 4.5
100/4.5 300/4.5 500/4.5 Total Count 3 3 3 9 Sum 5.215 5.525 5.905 16.645
Average 1.738 1.842 1.968 1.849 Variance 0.011 0.015 0.010 0.019
212
Table A-45: Total Interaction between Laser Power and Assist Gas Pressure Count 9 9 9 Sum 14.400 16.490 17.870 Average 1.600 1.832 1.986 Variance 0.030 0.012 0.021
Table A-46: ANOVA of Interaction between Laser Power and Assist Gas Pressure Source of Variation SS df MS F P-value F crit Sample 0.027 2 0.014 0.747 0.488 3.555 Columns 0.678 2 0.339 18.457 4.37 x10-5 3.555 Interaction 0.148 4 0.037 2.020 0.135 2.928 Within 0.331 18 0.018 Total 1.185 26
Table A-47: Interaction between Laser Power and Standoff Distance with replication
A/D 100 300 500 1 1.580 1.765 1.790 1.615 1.775 1.890 1.525 1.985 2.080 5 1.730 1.655 2.010 1.255 1.715 2.040 1.480 1.855 1.855
10 1.660 1.940 1.975 1.695 1.885 1.940 1.860 1.915 2.290
Table A-48: Interaction between Laser Power and Standoff Distance 1 100/1 300/1 500/1 Total Count 3 3 3 9 Sum 4.720 5.525 5.760 16.005 Average 1.573 1.842 1.920 1.778 Variance 0.002 0.015 0.022 0.035
Table A-49: Interaction between Laser Power and Standoff Distance 5 100/5 300/5 500/5 Total Count 3 3 3 9 Sum 4.465 5.225 5.905 15.595 Average 1.488 1.742 1.968 1.733 Variance 0.056 0.011 0.010 0.062
213
Table A-50: Interaction between Laser Power and Standoff Distance 10 100/10 300/10 500/10 Total Count 3 3 3 9 Sum 5.215 5.740 6.205 17.160 Average 1.738 1.913 2.068 1.907 Variance 0.011 0.001 0.037 0.033
Table A-51: Total Interaction between Laser Power and Standoff Distance Count 9 9 9 Sum 14.400 16.490 17.870 Average 1.600 1.832 1.986 Variance 0.030 0.012 0.021
Table A-52: ANOVA of Interaction between Laser Power and Standoff Distance Source of Variation SS df MS F P-value F crit Sample 0.146 2 0.073 3.982 0.037 3.555 Columns 0.678 2 0.339 18.457 0.000 3.555 Interaction 0.030 4 0.007 0.402 0.805 2.928 Within 0.331 18 0.018 Total 1.185 26
Table A-53: Interaction between Cutting speed and Assist Gas Pressure with replication C/B 0.2 0.7 1.2 0.5 1.580 1.975 1.655 1.615 1.940 1.715 1.525 2.290 1.855 2.5 1.940 1.730 1.790 1.885 1.255 1.890 1.915 1.480 2.080 4.5 2.010 1.765 1.660 2.040 1.775 1.695 1.855 1.985 1.860
Table A-54: Interaction between Cutting Speed and Assist Gas Pressure 0.5 0.2/0.5 0.7/0.5 1.2/0.5 Total Count 3 3 3 9 Sum 4.720 6.205 5.225 16.150 Average 1.573 2.068 1.742 1.794 Variance 0.002 0.037 0.011 0.060
214
Table A-55: Interaction between Cutting Speed and Assist Gas Pressure 2.5 0.2/2.5 0.7/2.5 1.2/2.5 Total Count 3 3 3 9 Sum 5.740 4.465 5.760 15.965 Average 1.913 1.488 1.920 1.774 Variance 0.001 0.056 0.022 0.066
Table A-56: Interaction between Cutting Speed and Assist Gas Pressure 4.5 0.2/4.5 0.7/4.5 1.2/4.5 Total Count 3 3 3 9 Sum 5.905 5.525 5.215 16.645 Average 1.968 1.842 1.738 1.849 Variance 0.010 0.015 0.011 0.019
Table A-57: Total Interaction between Cutting Speed and Assist Gas Pressure Count 9 9 9 Sum 16.365 16.195 16.200 Average 1.818 1.799 1.800 Variance 0.037 0.091 0.019
Table A-58: ANOVA of Interaction between Cutting Speed and Assist Gas Pressure
Source of Variation SS df MS F P-value F crit Sample 0.027 2 0.014 0.747 0.488 3.555 Columns 0.002 2 0.001 0.057 0.945 3.555 Interaction 0.825 4 0.206 11.220 9.58 x10-5 2.928 Within 0.331 18 0.018 Total 1.185 26
Table A-59: Interaction between Cutting speed and Standoff Distance with replication B/D 0.2 0.7 1.2 1 1.580 1.765 1.790 1.615 1.775 1.890 1.525 1.985 2.080 5 2.010 1.730 1.655 2.040 1.255 1.715 1.855 1.480 1.855 10 1.940 1.975 1.660 1.885 1.940 1.695 1.915 2.290 1.860
215
Table A-60: Interaction between Cutting Speed and Standoff Distance 1 0.2/1 0.7/1 1.2/1 Total Count 3 3 3 9 Sum 4.720 5.525 5.760 16.005 Average 1.573 1.842 1.920 1.778 Variance
0.002 0.015 0.022 0.035
Table A-61: Interaction between Cutting Speed and Standoff Distance 5 0.2/5 0.7/5 1.2/5 Total Count 3 3 3 3 Sum 5.905 4.465 5.225 15.595 Average 1.968 1.488 1.742 1.733 Variance 0.010 0.056 0.011 0.062
Table A-62: Interaction between Laser Cutting Speed and Standoff Distance 10 0.2/10 0.7/10 1.2/10 Total Count 3 3 3 9 Sum 5.740 6.205 5.215 17.160 Average 1.913 2.068 1.738 1.907 Variance 0.001 0.037 0.011 0.033
Table A-63: Total Interaction between Cutting Speed and Standoff Distance Count 9 9 9 Sum 16.365 16.195 16.200 Average 1.818 1.799 1.800 Variance 0.038 0.091 0.019
Table A-64: ANOVA of Interaction between Cutting Speed and Standoff Distance Source of Variation SS df MS F P-value F crit Sample 0.146 2 0.073 3.982 0.0370 3.555 Columns 0.002 2 0.001 0.057 0.9451 3.555 Interaction 0.706 4 0.176 9.602 0.0002 2.928 Within 0.331 18 0.018 Total 1.185 26
216
Table A-65: Interaction between Assist Gas Pressure and Standoff Distance with replication C/D 0.5 2.5 4.5
1 1.580 1.790 1.765 1.615 1.890 1.775 1.525 2.080 1.985 5 1.655 1.730 2.010 1.715 1.255 2.040 1.855 1.480 1.855
10 1.975 1.940 1.660 1.940 1.885 1.695 2.290 1.915 1.860
Table A-66: Interaction between Assist Gas Pressure and Standoff Distance 1 0.5/1 2.5/1 4.5/1 Total Count 3 3 3 9 Sum 4.720 5.760 5.525 16.005 Average 1.573 1.920 1.842 1.778 Variance 0.002 0.022 0.015 0.035
Table A-67: Interaction between Assist Gas Pressure and Assist Gas Pressure 5 0.5/5 2.5/5 4.5/5 Total Count 3 3 3 9 Sum 5.225 4.465 5.905 15.595 Average 1.742 1.488 1.968 1.733 Variance 0.011 0.056 0.010 0.062
Table A-68: Interaction between Assist Gas Pressure and Standoff Distance 10 0.5/10 2.5/10 1.2/4.5 Total Count 3 3 3 9 Sum 6.205 5.740 5.215 17.160 Average 2.068 1.913 1.738 1.907 Variance 0.037 0.001 0.011 0.033
Table A-69: Total Interaction between Assist Gas Pressure and Standoff Distance Count 9 9 9 Sum 16.150 15.965 16.645 Average 1.794 1.774 1.849 Variance 0.060 0.066 0.019
217
Table A-70: ANOVA of Interaction between Assist Gas Pressure and Standoff Distance Source of Variation SS df MS F P-value F crit
Sample 0.146 2 0.073 3.982 0.0370 3.555 Columns 0.027 2 0.014 0.747 0.4877 3.555 Interaction 0.680 4 0.170 9.257 0.0003 2.928 Within 0.331 18 0.018 Total 1.185 26
A.4 LINEAR REGRESSION ANALYSIS
A.4.1 LASER POWER AND KERF WIDTH
Figure A-5: Interactive graph of Laser Power and Kerf Width
Table A-71: Descriptive Statistics
Parameters Mean Std. Deviation N Kerf Width Mean(dependent) 1.806 0.189 9 Laser Power (Independent) 300.000 173.205 9
Table A-72: Correlation
Kerf Width Laser Power Pearson Correlation Kerf Width Mean 1.000 0.885 Laser Power 0.885 1.000 Sig. (1-tailed) Kerf Width Mean
Laser Power . .001
.001 . N Kerf Width Mean 9 9 Laser Power 9 9
218
Table A-73: Regression Statistics R Square 0.783 Adjusted R Square 0.752 Standard Error 0.094 Observations 9
Table A-74: Regression between Laser Power and Kerf width ANOVA d.f. SS MS F Significance F Regression 1 0.223 0.223 25.280 0.0015 Residual 7 0.062 0.009 Total 8 0.285
Table A-75: Linear regression line of Laser Power Coefficients Standard Error t Stat P-value Lower
95% Upper 95%
Intercept 1.517 0.065 23.163 7.09 x10-8 1.362 1.672 A 0.001 0.000 5.028 1.52 x10-3 0.001 0.001
Table A-76: Residual output
Observation Predicted KERF WIDTH MEAN Residuals 1 1.613 -0.040 2 1.613 -0.125 3 1.613 0.125 4 1.806 0.107 5 1.806 0.036 6 1.806 -0.064 7 1.999 -0.030 8 1.999 0.070 9 1.999 -0.079
Figure A-6: Laser Power (A) Line fit Plot
0.000
0.500
1.000
1.500
2.000
2.500
100 100 100 300 300 300 500 500 500
Expe
rim
enta
l and
Pred
icte
d Ke
rf W
idth
Laser Power
Kerf Width
Predicted Kerf Width
219
A.4.2 CUTTING SPEED AND KERF WIDTH
Figure A-7: Interactive graph of Cutting Speed and Kerf Width
Table A-77: Descriptive Statistics Mean Std. Deviation N Kerf Width Mean 1.80578 .188667 9 Cutting Speed .700 .4330 9
Table A-78: Correlation
Table A-79:
Regression Statistics Multiple R 0.042 R Square 0.002 Adjusted R Square -0.141 Standard Error 0.202 Observations 9.000
Kerf Width Cutting Speed Pearson Correlation Kerf Width Mean 1.000 -0.041 Cutting Speed -0.041 1.000 Sig. (1-tailed)
Kerf Width Mean Cutting Speed
. .458 .458 .
N Kerf Width Mean 9 9 Cutting Speed 9 9
220
Table A-80: Regression between Cutting speed and Kerf width ANOVA Df SS MS F Significance F Regression 1 0.001 0.001 0.012 0.9144 Residual 7 0.284 0.041 Total 8 0.285
Table A-81: Linear regression line of cutting speed Coefficients Standard Error t Stat P-value Lower 95% Upper 95% Intercept 1.819 0.133 13.642 0.0000 1.504 2.134 ( B ) -0.018 0.165 -0.111 0.0914 -0.407 0.371
Table A-82: Residual output Observation Predicted KERF
WIDTH MEAN Residual
s Observation Predicted KERF
WIDTH MEAN Residuals
1 1.815 -0.242 6 1.797 -0.055 2 1.806 -0.318 7 1.815 0.153 3 1.797 -0.058 8 1.806 0.262 4 1.815 0.098 9 1.797 0.123 5 1.806 0.036
Figure A-8: Cutting Speed (B) Line Fit Plot
0.000
0.500
1.000
1.500
2.000
2.500
0.2 0.7 1.2 0.2 0.7 1.2 0.2 0.7 1.2
Expe
rim
enta
l and
Pred
icte
d Ke
rf W
idth
Cutting Speed
Kerf Width
Predicted Kerf Width
221
A.4.3 ASSIST GAS PRESSURE AND KERF-WIDTH
Figure A-9: Interactive graph of Assist Gas Pressure and Kerf Width
Table A-83: Descriptive Statistics Mean Std. Deviation N Kerf Width Mean 1.806 0.189 9 Assist Gas pressure 2.500 1.732 9
Table A-84: Correlation
Table A-85: Regression Statistics Multiple R 0.126 R Square 0.016 Adjusted R Square -0.125 Standard Error 0.2 Observations 9
Kerf Width Assist Gas Pressure Pearson Correlation Kerf Width Mean 1.000 0.126 Assist Gas pressure 0.126 1.000 Sig. (1-tailed)
Kerf Width Mean Assist Gas pressure
. 0.373 0.373 .
N Kerf Width Mean 9 9 Assist Gas pressure 9 9
222
Table A-86: Regression between Assist Gas Pressure and Kerf Width ANOVA df SS MS F Significance F Regression 1 0.005 0.005 0.113 0.7462 Residual 7 0.280 0.040 Total 8 0.285
Table A-87: Linear regression line of Assist Gas Pressure Coefficients Standard Error t Stat P-value Lower
95% Upper 95%
Intercept 1.772 0.122 14.527 0.0000 1.483 2.060 ( C ) 0.014 0.041 0.337 0.746 -0.083 0.110
Table A-88: Residual output Observation Predicted KERF WIDTH MEAN Residuals
1 1.778 -0.205 2 1.806 -0.318 3 1.833 -0.095 4 1.806 0.107 5 1.833 0.008 6 1.778 -0.037 7 1.833 0.135 8 1.778 0.290 9 1.806 0.114
Figure A-10: Assist gas Pressure (C) Line Fit Plot
0
0.5
1
1.5
2
2.5
0.5 2.5 4.5 2.5 4.5 0.5 4.5 0.5 2.5
Expe
rim
enta
l and
Pred
icte
d Ke
rf W
idth
Assist Gas Pressure
TPM
Predicted TPM
223
A.4.4 STANDOFF DISTANCE AND KERF WIDTH
Table A-89: Descriptive Statistics Mean Std. Deviation N Kerf Width Mean 1.806 0.189 9 Standoff distance 5.33 3.905 9
Figure A-11: Interactive graph of Standoff Distance and Kerf Width
Table A-90: Correlation
Table A-91: Regression Statistics Multiple R 0.313 R Square 0.098 Adjusted R Square -0.031 Standard Error 0.192 Observations 9.000
Kerf Width Standoff Distance Pearson Correlation Kerf Width Mean 1.000 0.312 Standoff Distance 0.312 1.000 Sig. (1-tailed)
Kerf Width Mean Standoff Distance
. 0.207 0.207 .
N Kerf Width Mean 9 9 Standoff Distance 9 9
224
Table A-92: Regression between Standoff Distance and Kerf Width ANOVA
df SS MS F Significance F Regression 1 0.028 0.028 0.758 0.4128 Residual 7 0.257 0.037 Total 8 0.285
Table A-93: Linear regression line of Standoff Distance Coefficients Standard
Error t Stat P-value Lower
95% Upper 95%
Intercept 1.725 0.112 15.350 0.000 1.460 1.991 ( D ) 0.015 0.017 0.871 0.413 -0.026 0.056
Table A-94: Residual output
Observation Predicted TPM Residuals 1 1.740 -0.167 2 1.801 -0.313 3 1.876 -0.138 4 1.876 0.037 5 1.740 0.101 6 1.801 -0.059 7 1.801 0.167 8 1.876 0.192 9 1.740 0.180
Figure A-12: Standoff Distance (D) Line Fit Plot
0
0.5
1
1.5
2
2.5
1 5 10 10 1 5 5 10 1
Expe
rim
enta
l and
Pred
icte
d Ke
rf W
idth
Standoff Distance
Kerf Width
Predicted Kerf Width
225
A.5 NONLINEAR REGRESSION ANALYSIS Table A-95: Regression data without replication for Laser Power
S. No. Laser Power A A2 Kerf Width Mean 1 100 10000 1.573 2 100 10000 1.488 3 100 10000 1.738 4 300 90000 1.913 5 300 90000 1.842 6 300 90000 1.742 7 500 250000 1.968 8 500 250000 2.068 9 500 250000 1.920
Figure A-13: Quadratic graph of Laser Power without replication
Table A-96: Regression Statistics Multiple R 0.891 R Square 0.794 Adjusted R Square 0.725 Standard Error 0.099 Observations 9 Table A-97: Non-linear Regression ANOVA for Laser Power and Kerf width without replication d.f. SS MS F Significance F Regression 2 0.2261 0.1130 11.5684 0.0087 Residual 6 0.0586 0.0098 Total 8 0.2847
y = -1E-06x2 + 0.001x + 1.454R² = 0.794
0.000
0.500
1.000
1.500
2.000
2.500
0 200 400 600
Kerf
Wid
th M
ean
Laser Power
Kerf Width
Poly. (Kerf Width)
226
Table A-98: Nonlinear regression of Laser Power Coefficients Standard Error t Stat P-value Lower 95% Upper 95% Intercept 1.454 0.130 11.154 3.10 x10-5 1.135 1.773 A 0.002 0.001 1.457 0.195 -0.001 0.004 A2 -9.86 x10-7 1.75 x10-6 -0.564 0.593 -5.26 x10-6 3.29 x10-6
Table A-99: Residual output Observation Predicted Kerf width Mean Residuals
1 1.600 -0.027 2 1.600 -0.112 3 1.600 0.138 4 1.832 0.081 5 1.832 0.009 6 1.832 -0.091 7 1.986 -0.017 8 1.986 0.083 9 1.986 -0.066
Table A-100: Regression data with replication for Laser Power
S. No. Laser Power A A2 Kerf Width Replications 1 100 10000 1.58 2 100 10000 1.615 3 100 10000 1.525 4 100 10000 1.73 5 100 10000 1.255 6 100 10000 1.48 7 100 10000 1.66 8 100 10000 1.695 9 100 10000 1.86 10 300 90000 1.94 11 300 90000 1.885 12 300 90000 1.915 13 300 90000 1.765 14 300 90000 1.775 15 300 90000 1.985 16 300 90000 1.655 17 300 90000 1.715 18 300 90000 1.855 19 500 250000 2.01 20 500 250000 2.04 21 500 250000 1.855
227
S. No. Laser Power A A2 Kerf Width Replications 22 500 250000 1.975 23 500 250000 1.94 24 500 250000 2.29 25 500 250000 1.79 26 500 250000 1.89 27 500 250000 2.08
Figure A-14: Quadratic graph of Laser Power with replication
Table A-101: Regression Statistics Multiple R 0.757 R Square 0.572 Adjusted R Square 0.537 Standard Error 0.145 Observations 27 Table A-102: Non-linear Regression ANOVA for Laser Power and Kerf width with replication d.f. SS MS F Significance F Regression 2 0.678 0.339 16.066 3.73E-05 Residual 24 0.507 0.021 Total 26 1.185
Table A-103: Nonlinear regression of Laser Power Coefficients Standard Error t Stat P-value Lower 95% Upper 95% Intercept 1.4543 0.1106 13.145 1.85 x10-12 1.2260 1.6826 A 0.0016 0.0009 1.7169 0.0989 -0.0003 0.0034 A2 -9.86 x10-7 1.48 x10-6 -0.665 0.5124 -4.05 x10-6 2.07 x10-6
y = -1E-06x2 + 0.001x + 1.454R² = 0.572
0
0.5
1
1.5
2
2.5
0 200 400 600
Kerf
Wid
th
Laser Power
Kerf Width
Poly. (Kerf Width )
228
Table A-104: Residual Output
Observation Predicted Kerf width Mean
Residuals
1 1.600 -0.020 2 1.600 0.015 3 1.600 -0.075 4 1.600 0.130 5 1.600 -0.345 6 1.600 -0.120 7 1.600 0.060 8 1.600 0.095 9 1.600 0.260 10 1.832 0.108 11 1.832 0.053 12 1.832 0.083 13 1.832 -0.067 14 1.832 -0.057 15 1.832 0.153 16 1.832 -0.177 17 1.832 -0.117 18 1.832 0.023 19 1.986 0.024 20 1.986 0.054 21 1.986 -0.131 22 1.986 -0.011 23 1.986 -0.046 24 1.986 0.304 25 1.986 -0.196 26 1.986 -0.096 27 1.986 0.094
Table A-105: Regression data with replication for Cutting Speed
Cutting Speed B B2 Kerf width Mean 1 0.2 0.04 1.58 2 0.2 0.04 1.615 3 0.2 0.04 1.525 4 0.2 0.04 1.94 5 0.2 0.04 1.885 6 0.2 0.04 1.915 7 0.2 0.04 2.01 8 0.2 0.04 2.04 9 0.2 0.04 1.855
229
Cutting Speed B B2 Kerf width Mean 10 0.7 0.49 1.73 11 0.7 0.49 1.255 12 0.7 0.49 1.48 13 0.7 0.49 1.765 14 0.7 0.49 1.775 15 0.7 0.49 1.985 16 0.7 0.49 1.975 17 0.7 0.49 1.94 18 0.7 0.49 2.29 19 1.2 1.44 1.66 20 1.2 1.44 1.695 21 1.2 1.44 1.86 22 1.2 1.44 1.655 23 1.2 1.44 1.715 24 1.2 1.44 1.855 25 1.2 1.44 1.79 26 1.2 1.44 1.89 27 1.2 1.44 2.08
Figure A-15: Quadratic graph of Cutting Speed with replication
Table A-106: Regression Statistics Multiple R 0.042 R Square 0.002 Adjusted R Square -0.081 Standard Error 0.222 Observations 27
y = 0.038x2 - 0.072x + 1.831R² = 0.001
0
0.5
1
1.5
2
2.5
-0.3 0.2 0.7 1.2
Kerf
WId
th
Cutting Speed
Ker Width
Poly. (Ker Width)
230
Table A-107: Non-linear Regression ANOVA for Cutting Speed and Kerf width with replication
d.f. SS MS F Significance F Regression 2 0.002 0.001 0.021 0.979 Residual 24 1.183 0.049 Total 26 1.185
Table A-108: Nonlinear regression of Cutting Speed Coefficients Standard Error t Stat P-value Lower 95% Upper 95% Intercept 1.831 0.145 12.658 4.11 x10-12 1.533 2.130 B -0.073 0.518 -0.140 0.889 -1.142 0.997 B2 0.039 0.363 0.107 0.915 -0.709 0.787
Table A-109: Residual Output
Observation Predicted Kerf width Mean
Residuals
1 1.818 -0.238 2 1.818 -0.203 3 1.818 -0.293 4 1.818 0.122 5 1.818 0.067 6 1.818 0.097 7 1.818 0.192 8 1.818 0.222 9 1.818 0.037 10 1.799 -0.069 11 1.799 -0.544 12 1.799 -0.319 13 1.799 -0.034 14 1.799 -0.024 15 1.799 0.186 16 1.799 0.176 17 1.799 0.141 18 1.799 0.491 19 1.800 -0.140 20 1.800 -0.105 21 1.800 0.060 22 1.800 -0.145 23 1.800 -0.085 24 1.800 0.055 25 1.800 -0.010 26 1.800 0.090 27 1.800 0.280
231
Table A-110: Regression data with replication for Assist Gas Pressure
S. No. Assist Gas Pressure(C) C2 Kerf Width Mean 1 0.5 0.25 1.58 2 0.5 0.25 1.615 3 0.5 0.25 1.525 4 0.5 0.25 1.655 5 0.5 0.25 1.715 6 0.5 0.25 1.855 7 0.5 0.25 1.975 8 0.5 0.25 1.94 9 0.5 0.25 2.29 10 2.5 6.25 1.73 11 2.5 6.25 1.255 12 2.5 6.25 1.48 13 2.5 6.25 1.94 14 2.5 6.25 1.885 15 2.5 6.25 1.915 16 2.5 6.25 1.79 17 2.5 6.25 1.89 18 2.5 6.25 2.08 19 4.5 20.25 1.66 20 4.5 20.25 1.695 21 4.5 20.25 1.86 22 4.5 20.25 1.765 23 4.5 20.25 1.775 24 4.5 20.25 1.985 25 4.5 20.25 2.01 26 4.5 20.25 2.04 27 4.5 20.25 1.855
232
Figure A-16: Quadratic graph of Assist Gas Pressure with replication
Table A-111: Regression Statistics Multiple R 0.152 R Square 0.023 Adjusted R Square -0.058 Standard Error 0.220 Observations 27
Table A-112: Non-linear Regression ANOVA for Assist Gas Pressure and Kerf width with replication d.f. SS MS F Significance F Regression 2 0.027 0.014 0.285 0.755 Residual 24 1.157 0.048 Total 26 1.185
Table A-113: Nonlinear regression of Assist Gas Pressure Coefficients Standard Error t Stat P-value Lower 95% Upper 95% Intercept 1.815 0.111 16.281 1.8 x10-14 1.585 2.045 C -0.046 0.115 -0.403 0.691 -0.284 0.191 C2 0.012 0.022 0.536 0.597 -0.034 0.058
y = 0.012x2 - 0.046x + 1.814R² = 0.023
0
0.5
1
1.5
2
2.5
-0.5 0.5 1.5 2.5 3.5 4.5
Kerf
WId
th
Asssit Gas Pressure
Ker Width
Poly. (Ker Width)
233
Table A-114: Residual Output Observation Predicted Kerf Width Mean Residuals
1 1.794 -0.214 2 1.794 -0.179 3 1.794 -0.269 4 1.794 -0.139 5 1.794 -0.079 6 1.794 0.061 7 1.794 0.181 8 1.794 0.146 9 1.794 0.496 10 1.774 -0.044 11 1.774 -0.519 12 1.774 -0.294 13 1.774 0.166 14 1.774 0.111 15 1.774 0.141 16 1.774 0.016 17 1.774 0.116 18 1.774 0.306 19 1.849 -0.189 20 1.849 -0.154 21 1.849 0.011 22 1.849 -0.084 23 1.849 -0.074 24 1.849 0.136 25 1.849 0.161 26 1.849 0.191 27 1.849 0.006
Table A-115: Regression data with replication for Standoff Distance
Standoff Distance (D) D2 Kerf Width Mean 1 1 1 1.58 2 1 1 1.615 3 1 1 1.525 4 1 1 1.765 5 1 1 1.775 6 1 1 1.985 7 1 1 1.79 8 1 1 1.89 9 1 1 2.08
234
Standoff Distance (D) D2 Kerf Width Mean 10 5 25 1.73 11 5 25 1.255 12 5 25 1.48 13 5 25 1.655 14 5 25 1.715 15 5 25 1.855 16 5 25 2.01 17 5 25 2.04 18 5 25 1.855 19 10 100 1.66 20 10 100 1.695 21 10 100 1.86 22 10 100 1.94 23 10 100 1.885 24 10 100 1.915 25 10 100 1.975 26 10 100 1.94 27 10 100 2.29
Figure A-17: Quadratic graph of Standoff Distance with replication
Table A-116: Regression Statistics Multiple R 0.351 R Square 0.124 Adjusted R Square 0.050 Standard Error 0.208 Observations 27
y = 0.005x2 - 0.042x + 1.815R² = 0.123
0
0.5
1
1.5
2
2.5
0 5 10
Kerf
WId
th
Standoff Distance
Ker Width
Poly. (Ker Width)
235
Table A-117: Non-linear Regression ANOVA for Standoff Distance and Kerf width with replication
d.f. SS MS F Significance F Regression 2 0.146 0.073 1.691 0.206 Residual 24 1.039 0.043 Total 26 1.185
Table A-118: Nonlinear regression of Standoff Distance Coefficients Standard Error t Stat P-value Lower 95% Upper 95% Intercept 1.815 0.103 17.686 2.86 x10-15 1.604 2.027 D -0.042 0.049 -0.865 0.395 -0.143 0.058 D2 0.005 0.004 1.206 0.240 -0.004 0.014
Table A-119: Residual Output Observation Predicted Kerf Width Mean Residuals
1 1.778 -0.198 2 1.778 -0.163 3 1.778 -0.253 4 1.778 -0.013 5 1.778 -0.003 6 1.778 0.207 7 1.778 0.012 8 1.778 0.112 9 1.778 0.302 10 1.733 -0.003 11 1.733 -0.478 12 1.733 -0.253 13 1.733 -0.078 14 1.733 -0.018 15 1.733 0.122 16 1.733 0.277 17 1.733 0.307 18 1.733 0.122 19 1.907 -0.247 20 1.907 -0.212 21 1.907 -0.047 22 1.907 0.033 23 1.907 -0.022 24 1.907 0.008 25 1.907 0.068 26 1.907 0.033 27 1.907 0.383
236
A.6 MULTIPLE NON-LINEAR REGRESSION Table A-120: Multiple Non-linear Regression data with replication for four inputs
A A2 B B2 C C2 D D2 Kerf Width 1 100 10000 0.2 0.04 0.5 0.25 1 1 1.58 2 100 10000 0.2 0.04 0.5 0.25 1 1 1.615 3 100 10000 0.2 0.04 0.5 0.25 1 1 1.525 4 100 10000 0.7 0.49 2.5 6.25 5 25 1.73 5 100 10000 0.7 0.49 2.5 6.25 5 25 1.255 6 100 10000 0.7 0.49 2.5 6.25 5 25 1.48 7 100 10000 1.2 1.44 4.5 20.25 10 100 1.66 8 100 10000 1.2 1.44 4.5 20.25 10 100 1.695 9 100 10000 1.2 1.44 4.5 20.25 10 100 1.86 10 300 90000 0.2 0.04 2.5 6.25 10 100 1.94 11 300 90000 0.2 0.04 2.5 6.25 10 100 1.885 12 300 90000 0.2 0.04 2.5 6.25 10 100 1.915 13 300 90000 0.7 0.49 4.5 20.25 1 1 1.765 14 300 90000 0.7 0.49 4.5 20.25 1 1 1.775 15 300 90000 0.7 0.49 4.5 20.25 1 1 1.985 16 300 90000 1.2 1.44 0.5 0.25 5 25 1.655 17 300 90000 1.2 1.44 0.5 0.25 5 25 1.715 18 300 90000 1.2 1.44 0.5 0.25 5 25 1.855 19 500 250000 0.2 0.04 4.5 20.25 5 25 2.01 20 500 250000 0.2 0.04 4.5 20.25 5 25 2.04 21 500 250000 0.2 0.04 4.5 20.25 5 25 1.855 22 500 250000 0.7 0.49 0.5 0.25 10 100 1.975 23 500 250000 0.7 0.49 0.5 0.25 10 100 1.94 24 500 250000 0.7 0.49 0.5 0.25 10 100 2.29 25 500 250000 1.2 1.44 2.5 6.25 1 1 1.79 26 500 250000 1.2 1.44 2.5 6.25 1 1 1.89 27 500 250000 1.2 1.44 2.5 6.25 1 1 2.08 Table A-121: Regression Statistics Multiple R 0.849 R Square 0.721 Adjusted R Square 0.597 Standard Error 0.136 Observations 27 Table A-122: Multiple Non-linear Regression data with replication ANOVA
d.f. SS MS F Significance F Regression 8 0.854 0.107 5.811 0.0010 Residual 18 0.331 0.018 Total 26 1.185
237
Table A-123: Nonlinear Regression of multivariable Coefficients Standard Error t Stat P-value Lower 95% Upper 95%
Intercept 1.498 0.160 9.357 2.46 x10-8 1.162 1.834 A 0.002 0.001 1.840 0.082 0.000 0.003 A2 -9.86 x10-7 1.38 x10-6 -0.713 0.485 -3.89 x10-6 1.92 x10-6 B -0.073 0.316 -0.230 0.821 -0.738 0.592 B2 0.039 0.221 0.176 0.863 -0.426 0.504 C -0.046 0.071 -0.652 0.522 -0.195 0.103 C2 0.012 0.014 0.868 0.397 -0.017 0.041 D -0.042 0.032 -1.328 0.201 -0.109 0.025 D2 0.005 0.003 1.850 0.081 -0.001 0.011
Table A-124: Residual Output
Observation Predicted Kerf width Mean Residuals 1 1.573 0.007 2 1.573 0.042 3 1.573 -0.048 4 1.488 0.242 5 1.488 -0.233 6 1.488 -0.008 7 1.738 -0.078 8 1.738 -0.043 9 1.738 0.122 10 1.913 0.027 11 1.913 -0.028 12 1.913 0.002 13 1.842 -0.077 14 1.842 -0.067 15 1.842 0.143 16 1.742 -0.087 17 1.742 -0.027 18 1.742 0.113 19 1.968 0.042 20 1.968 0.072 21 1.968 -0.113 22 2.068 -0.093 23 2.068 -0.128 24 2.068 0.222 25 1.920 -0.130 26 1.920 -0.030 27 1.920 0.160
APPENDIX B
238
B. NEURAL NETWORK & OVERALL QUALITY Table B-1: Training on Factorial design
S. No. No. of neurons Max. % error Min. % error Average % error Results
1 20-20-20 12.374 0.0003 0.663 Train
2.2 0.0001 0.186
14.135 0.0003 12.37 Test
9.115 0.0001 2.2
81.000 11.000 46.000 Simulation
0.290 0.000 0.123
2 10-10-10 66.376 0.247 16.440 Train
9.533 0.002 1.272
66.376 0.002 25.173 Test
16.437 0.247 9.533
28.608 0.000 18.323 Simulation
83.740 0.005 21.338
3 5-5-5 51.685 0.527 19.819 Train
67.427 0.224 21.509
51.685 0.224 14.201 Test
67.427 0.527 16.673
109.517 0.002 26.268 Simulation
108.743 0.000 26.133
4 30-30-30 12.143 0.002 0.931 Train
2.764 0.0002 0.393
16.275 0.0002 12.143 Test
13.192 0.002 2.764
104.671 0.000 28.513 Simulation
77.256 0.517 19.343
5 30-30-30 10.220 0.0002 0.318 Train
3.088 2.2e-05 0.145
15.140 2.20e-5 10.220 Test
239
Table B-2: Edge quality mean training using factorial datasets
S. No. No. of
Neurons Max.% error
Min.% error
Average % error
Overall Regression
Results
1 5-5 28.757 0.017 7.691 0.910 Train
25.499 2.601 9.869 Test
57.414 0.049 17.472 Simulation
2 10-10 54.752 0.066 11.899 0.902 Train
53.614 0.447 20.029 Test
61.376 0.337 14.305 Simulation
3 30-30 45.797 0.001 4.215 0.911 Train
29.776 8.080 15.987 Test
67.506 0.213 17.732 Simulation
4 10-10-10 34.691 0.007 3.059 0.984 Train
42.789 1.695 14.067 Test
80.650 0.656 17.894 Simulation
5 20-20-20 12.309 0.002 0.667 0.994 Train
31.960 2.262 17.052 Test
83.781 0.432 18.819 Simulation
6 30-30-30 21.789 0.004 1.489 0.994 Train
18.301 1.138 11.822 Test
72.636 0.099 18.265 Simulation
7 40-40-40 20.361 0.060 1.965 0.986 Train
23.958 0.702 11.657 Test
90.140 0.246 18.495 Simulation
8 65-65-65 71.222 0.005 6.162 0.922 Train
54.716 1.047 19.281 Test
103.241 0.592 22.471 Simulation
9 40-40 1.587 0.000 0.115 0.999 Train
14.564 1.852 8.960 Test
83.677 0.372 17.660 Simulation
10 50-50 14.995 0.000 0.678 0.900 Train
87.975 21.624 57.457 Test
92.221 0.104 21.184 Simulation
240
Table B-3: Edge quality signal to noise ratio of factorial datasets
S. No. No. of
Neurons Max.% error
Min.% error
Average %error
Overall Regression
Results
1 10 26.066 0.016 4.431 0.963 Train 19.470 1.290 9.868 Test 26.066 0.016 4.901 Simulation 2 20 25.010 0.002 6.424 0.922 Train 12.830 0.147 3.784 Test 25.010 0.002 6.196 Simulation 3 6 23.623 0.063 6.950 0.913 Train 16.044 1.256 9.640 Test 26.066 0.016 4.901 Simulation 4 10-10 25.567 0.051 1.840 0.969 Train 23.367 0.215 15.292 Test 25.567 0.051 3.003 Simulation 5 20-20 25.679 0.013 1.678 0.949 Train 55.579 2.029 22.590 Test 55.579 0.013 3.485 Simulation 6 10-10-10 24.781 0.047 2.672 0.962 Train 21.998 1.136 9.546 Test 24.781 0.047 3.266 Simulation 7 20-20-20 29.493 0.257 6.985 0.943 Train 15.159 0.509 7.706 Test 29.493 0.257 7.047 Simulation
Table B-4: Edge quality mean of normalized dataset
S. No. No. of
Neurons Max.% error
Min.% error
Average % error
Overall Regression
Results
1 10-10 29.778 0.027 6.482 0.953 Train 39.168 4.180 17.639 Test 35.254 0.573 9.892 Simulation 2 20-20 29.272 0.000 2.242 0.954 Train 61.590 3.893 16.318 Test 29.759 0.006 6.723 Simulation 3 30-30 66.666 0.134 4.886 0.884 Train 69.056 9.207 23.385 Test 44.857 0.157 7.701 Simulation 4 10-10-10 14.891 0.085 3.299 0.957 Train 39.119 2.200 12.234 Test 31.493 0.019 6.953 Simulation 5 20-20-20 17.088 0.003 1.963 0.977 Train
241
S. No. No. of
Neurons Max.% error
Min.% error
Average % error
Overall Regression
Results
59.009 1.929 22.556 Test 31.892 0.079 7.398 Simulation 6 30-30-30 26.804 0.000 0.664 0.983 Train 76.089 1.852 21.426 Test 27.018 0.011 6.232 Simulation 7 40-40-40 24.207 0.004 1.434 0.966 Train 48.022 8.332 27.453 Test 47.686 0.034 7.707 Simulation
Table B-5: Kerf width mean training of factorial design S. No. No. of
Neurons Max.% error
Min.% error
Average % error
Overall Regression
Results
1 10 43.005 0.040 4.201 0.9861 Train 57.876 3.687 15.697 Test 57.876 0.040 5.194 Simulation 2 20 48.976 0.000 2.518 0.9767 Train 42.996 0.993 17.762 Test 48.976 0.000 3.835 Simulation 3 10-10 20.876 0.000 0.887 0.9631 Train 59.445 3.544 26.821 Test 59.445 0.000 3.128 Simulation 4 20-20 110.949 0.000 5.848 0.9986 Train 72.252 7.785 35.447 Test 110.949 0.000 8.406 Simulation 5 30-30 48.116 0.633 11.506 0.8802 Train 31.662 0.762 16.352 Test 48.116 0.633 11.925 Simulation 6 10-10-10 35.301 0.006 3.257 0.9614 Train 32.369 1.423 16.321 Test 35.301 0.006 4.386 Simulation 7 20-20-20 61.063 0.000 5.778 0.9722 Train 25.190 1.146 13.453 Test
61.063 0.000 6.441 Simulation
242
Figure B-1: No. of times initialization comparison of 3mm sheet EQ Mean
Figure B-2: No. of times initialization comparison of 5mm sheet EQ Mean
9.5
10.0
10.5
11.0
11.5
12.0
2 3 4 5 6 7 8 9 10
Ave
rage
per
cent
err
or
Number of neurons
3000 times initialize
12000 times initialize
11.0
11.5
12.0
12.5
13.0
13.5
14.0
14.5
15.0
15.5
16.0
2 3 4 5 6 7 8 9 10 11
Ave
rage
per
cent
err
or
Number of Neuron
3000 times initialize
12000 times initialize
243
Figure B-3: No. of times initialization comparison of 3mm sheet EQ S/N
Figure B-4: No. of times initialization comparison of 5mm sheet EQ S/N
8
8.5
9
9.5
10
10.5
11
11.5
2 3 4 5 6 7 8 9 10
Ave
rage
per
cent
err
or
Number of neuron
3000 times initialize
12000 times initialize
10.0
10.5
11.0
11.5
12.0
12.5
13.0
13.5
14.0
14.5
2 3 4 5 6 7 8 9 10 11
Ave
rage
per
cent
err
or
Number of neuron
3000 times initialize
12000 times initialize
244
Figure B-5: No. of times initialization comparison of 3mm sheet KW Mean
Figure B-6: No. of times initialization comparison of 5mm sheet KW Mean
4.0
4.5
5.0
5.5
6.0
6.5
7.0
2 3 4 5 6 7 8 9 10
Ave
rage
per
cent
err
or
Number of neuron
3000 times initialize
12000 times initialize
7.5
8.0
8.5
9.0
9.5
10.0
2 3 4 5 6 7 8 9 10 11
Ave
rage
per
cent
err
or
Number of neuron
3000 times initialize
12000 times initialize
245
Figure B-7: No. of times initialization comparison of 3mm sheet KW S/N
Figure B-8: No. of times initialization comparison of 5mm sheet KW S/N
5.06.07.08.09.0
10.011.012.013.014.015.016.017.018.0
2 3 4 5 6 7 8 9 10 11
Ave
rage
per
cent
err
or
Number of neuron
3000 times initialize
12000 times initialize
7.0
8.0
9.0
10.0
11.0
2 3 4 5 6 7 8 9 10 11
Ave
rage
per
cent
err
or
Number of neuron
3000 times initialize
12000 times initialize
246
Table B-6: Simulated data of factorial design
Run LP
A
CS
B
AGP
C
SD
D
Simulated results
EQ KW POC MRR
1 100 0.2 0.5 1 3.670 1.552 2.356 6.74E-08
2 100 0.2 0.5 5 2.729 1.842 3.089 6.98E-08
3 100 0.2 0.5 10 2.090 1.882 3.868 6.80E-08
4 100 0.2 2.5 1 2.833 1.440 2.413 6.31E-08
5 100 0.2 2.5 5 2.991 1.547 3.207 6.37E-08
6 100 0.2 2.5 10 2.086 1.878 3.799 6.62E-08
7 100 0.2 4.5 1 1.846 1.485 2.434 6.56E-08
8 100 0.2 4.5 5 1.545 1.500 3.259 7.98E-08
9 100 0.2 4.5 10 1.264 1.837 3.736 6.61E-08
10 100 0.7 0.5 1 3.968 1.489 2.228 7.21E-08
11 100 0.7 0.5 5 2.584 1.563 2.696 2.53E-07
12 100 0.7 0.5 10 1.484 1.877 3.871 7.76E-08
13 100 0.7 2.5 1 2.926 1.485 2.236 8.64E-08
14 100 0.7 2.5 5 2.730 1.488 2.763 2.25E-07
15 100 0.7 2.5 10 1.211 1.832 3.888 6.60E-08
16 100 0.7 4.5 1 1.406 1.636 2.220 4.05E-07
17 100 0.7 4.5 5 1.934 1.641 2.801 4.57E-07
18 100 0.7 4.5 10 1.539 1.694 3.832 4.40E-07
19 100 1.2 0.5 1 2.869 1.507 2.304 4.79E-07
20 100 1.2 0.5 5 2.146 1.484 2.589 4.35E-07
21 100 1.2 0.5 10 1.400 1.827 3.513 3.42E-07
22 100 1.2 2.5 1 2.348 1.623 2.249 4.85E-07
23 100 1.2 2.5 5 2.215 1.624 2.612 4.84E-07
24 100 1.2 2.5 10 1.078 1.680 3.627 4.45E-07
25 100 1.2 4.5 1 1.669 1.735 2.196 4.78E-07
26 100 1.2 4.5 5 2.211 1.735 2.586 4.58E-07
27 100 1.2 4.5 10 1.670 1.738 3.664 4.56E-07
28 300 0.2 0.5 1 2.988 1.632 3.519 8.01E-08
247
Run LP
A
CS
B
AGP
C
SD
D
Simulated results
EQ KW POC MRR
29 300 0.2 0.5 5 2.175 1.862 3.789 2.77E-07
30 300 0.2 0.5 10 1.542 1.917 3.924 1.30E-07
31 300 0.2 2.5 1 2.743 1.644 3.524 8.66E-08
32 300 0.2 2.5 5 2.302 1.643 3.764 8.10E-08
33 300 0.2 2.5 10 1.670 1.913 3.821 6.56E-08
34 300 0.2 4.5 1 2.405 1.748 3.504 4.20E-07
35 300 0.2 4.5 5 2.265 1.683 3.710 1.06E-07
36 300 0.2 4.5 10 1.276 1.862 3.766 6.57E-08
37 300 0.7 0.5 1 3.836 1.649 3.051 2.70E-07
38 300 0.7 0.5 5 2.795 1.650 3.732 3.05E-07
39 300 0.7 0.5 10 1.669 1.913 4.100 3.19E-07
40 300 0.7 2.5 1 2.212 1.805 3.091 4.85E-07
41 300 0.7 2.5 5 2.896 1.718 3.719 4.20E-07
42 300 0.7 2.5 10 1.948 1.870 4.034 6.95E-08
43 300 0.7 4.5 1 1.000 1.842 3.129 4.82E-07
44 300 0.7 4.5 5 1.135 1.778 3.662 4.57E-07
45 300 0.7 4.5 10 1.375 1.759 3.942 4.37E-07
46 300 1.2 0.5 1 3.045 1.829 2.731 4.85E-07
47 300 1.2 0.5 5 2.670 1.772 3.590 4.53E-07
48 300 1.2 0.5 10 2.014 1.877 4.098 3.22E-07
49 300 1.2 2.5 1 2.109 1.902 2.681 4.85E-07
50 300 1.2 2.5 5 2.273 1.827 3.648 4.85E-07
51 300 1.2 2.5 10 1.278 1.774 4.087 4.57E-07
52 300 1.2 4.5 1 0.888 1.837 2.680 4.85E-07
53 300 1.2 4.5 5 1.714 1.782 3.623 4.81E-07
54 300 1.2 4.5 10 1.718 1.757 4.028 4.57E-07
55 500 0.2 0.5 1 2.107 1.672 4.274 9.25E-08
56 500 0.2 0.5 5 1.599 1.715 4.036 3.18E-07
57 500 0.2 0.5 10 1.200 1.905 3.925 3.17E-07
58 500 0.2 2.5 1 2.051 1.870 4.1372 1.12E-07
248
Run LP
A
CS
B
AGP
C
SD
D
Simulated results
EQ KW POC MRR
59 500 0.2 2.5 5 1.658 1.895 3.869 3.24E-07
60 500 0.2 2.5 10 1.081 2.057 3.804 7.26E-08
61 500 0.2 4.5 1 2.353 1.972 3.916 4.81E-07
62 500 0.2 4.5 5 2.170 1.968 3.768 8.55E-08
63 500 0.2 4.5 10 1.095 2.084 3.763 6.51E-08
64 500 0.7 0.5 1 3.204 1.856 4.221 3.62E-07
65 500 0.7 0.5 5 2.793 1.862 4.108 3.22E-07
66 500 0.7 0.5 10 1.830 2.068 4.125 3.14E-07
67 500 0.7 2.5 1 1.970 1.911 4.115 4.85E-07
68 500 0.7 2.5 5 2.747 1.928 3.927 4.45E-07
69 500 0.7 2.5 10 2.330 2.095 4.015 3.43E-07
70 500 0.7 4.5 1 0.865 1.977 3.903 4.85E-07
71 500 0.7 4.5 5 1.047 1.964 3.787 4.84E-07
72 500 0.7 4.5 10 1.335 1.925 3.929 4.36E-07
73 500 1.2 0.5 1 3.356 1.884 4.117 4.85E-07
74 500 1.2 0.5 5 3.094 1.886 4.294 4.49E-07
75 500 1.2 0.5 10 2.305 1.962 4.168 3.24E-07
76 500 1.2 2.5 1 1.756 1.920 4.031 4.85E-07
77 500 1.2 2.5 5 2.383 1.935 4.118 4.85E-07
78 500 1.2 2.5 10 1.624 1.950 4.109 4.82E-07
79 500 1.2 4.5 1 0.625 1.977 3.842 4.85E-07
80 500 1.2 4.5 5 1.125 1.957 3.933 4.85E-07
81 500 1.2 4.5 10 1.688 1.892 4.049 4.82E-07
249
Table B-7: Compare results of Normalized aggregation, Customer quality function and fuzzy
aggregation
Run Ordinal outputs EQ KW POC MRR Aggregation Customer
quality function
Fuzzy aggregation
Quantified fuzzy
aggregation 1 2 2 1 1 1 2 0.45 2
2 1 2 1 1 1 2 0.48 2
3 2 1 1 2 1 2 0.48 2
4 3 2 1 1 2 3 0.46 2
5 3 1 1 2 2 3 0.42 2
6 2 2 2 2 2 2 0.48 2
7 1 3 1 2 2 3 0.36 2
8 2 2 1 1 2 2 0.45 2
9 2 1 1 3 2 3 0.49 2
10 2 2 1 2 2 2 0.44 2
11 2 1 3 3 2 3 0.39 2
12 3 1 1 3 2 3 0.48 2
13 3 1 1 3 2 3 0.47 2
14 2 3 1 2 2 3 0.38 2
15 1 2 3 3 2 3 0.48 2
16 3 1 2 3 2 3 0.49 2
17 2 3 3 1 2 3 0.61 3
18 1 3 2 3 2 3 0.55 3
19 2 3 2 2 2 3 0.48 2
20 1 3 3 1 2 3 0.49 2
21 2 2 3 1 2 3 0.48 2
22 2 2 3 2 2 3 0.49 2
23 1 3 3 1 2 3 0.49 2
24 1 3 3 3 2 3 0.49 2
25 1 3 3 2 2 3 0.49 2
26 1 3 3 1 2 3 0.50 2
27 1 2 3 2 2 3 0.49 2
250
Run Ordinal outputs EQ KW POC MRR Aggregation Customer
quality function
Fuzzy aggregation
Quantified fuzzy
aggregation 28 2 2 3 3 2 3 0.49 2
29 1 3 3 2 2 3 0.68 3
30 3 3 2 2 2 3 0.48 2
31 2 3 3 1 2 3 0.49 2
32 3 1 1 3 3 3 0.49 2
33 2 3 3 2 3 3 0.49 2
34 3 2 3 2 3 3 0.63 3
35 3 1 1 3 3 3 0.49 2
36 2 2 3 2 3 3 0.49 2
37 3 2 3 2 3 3 0.59 3
38 1 3 3 2 3 3 0.63 3
39 2 3 3 1 3 3 0.48 2
40 1 3 3 2 3 3 0.63 3
41 3 1 1 3 3 3 0.63 3
42 3 3 3 2 3 3 0.48 2
43 2 3 3 2 3 3 0.63 3
44 3 2 3 3 3 3 0.63 3
45 3 2 3 2 3 3 0.63 3
46 1 3 3 3 3 3 0.57 3
47 2 3 3 2 3 3 0.49 2
48 3 1 2 3 3 3 0.48 2
49 2 3 3 2 3 3 0.53 3
50 1 3 3 3 3 3 0.51 3
51 3 3 3 1 3 3 0.49 2
52 2 2 3 3 3 3 0.52 3
53 2 2 3 3 3 3 0.63 3
54 3 2 2 3 3 3 0.49 2
55 1 3 3 3 3 3 0.49 2
56 2 3 3 3 3 3 0.62 3
251
Run Ordinal outputs EQ KW POC MRR Aggregation Customer
quality function
Fuzzy aggregation
Quantified fuzzy
aggregation 57 2 3 3 3 3 3 0.63 3
58 3 3 3 1 3 3 0.49 2
59 2 3 3 3 3 3 0.62 3
60 2 3 3 3 3 3 0.49 2
61 2 3 3 3 3 3 0.63 3
62 3 2 3 3 3 3 0.49 2
63 3 3 3 1 3 3 0.49 2
64 3 2 3 3 3 3 0.82 3
65 1 3 3 3 3 3 0.83 3
66 3 3 2 3 3 3 0.48 2
67 1 3 3 3 3 3 0.63 3
68 3 3 3 2 3 3 0.63 3
69 3 3 3 3 3 3 0.54 3
70 1 3 3 3 3 3 0.63 3
71 3 3 3 2 3 3 0.63 3
72 3 3 3 2 3 3 0.63 3
73 3 3 3 2 3 3 0.86 3
74 2 3 3 3 3 3 0.73 3
75 2 3 3 3 3 3 0.48 2
76 2 3 3 3 3 3 0.68 3
77 2 3 3 3 3 3 0.79 3
78 2 3 3 3 3 3 0.48 2
79 2 3 3 3 3 3 0.63 3
80 3 3 3 3 3 3 0.67 3
81 2 3 3 3 3 3 0.48 2
252
Figure B-9: Comparison of quantified aggregation with customer quality function
Table B-8: Sorted with CQF, Quantified fuzzy aggregation and Quantified Normalized
aggregation
RUN EQ KW POC MRR Q AV CQF FL AV QFL AV
1 2 2 1 1 1 2 0.45 2
2 1 2 1 1 1 2 0.48 2
3 2 1 1 2 1 2 0.48 2
10 2 2 1 2 2 2 0.44 2
8 2 2 1 1 2 2 0.45 2
6 2 2 2 2 2 2 0.48 2
7 1 3 1 2 2 3 0.36 2
14 2 3 1 2 2 3 0.38 2
11 2 1 3 3 2 3 0.39 2
5 3 1 1 2 2 3 0.42 2
4 3 2 1 1 2 3 0.46 2
13 3 1 1 3 2 3 0.47 2
1
2
3
4
1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67 70 73 76 79
Aggregation
Customer quality function
253
RUN EQ KW POC MRR Q AV CQF FL AV QFL AV
12 3 1 1 3 2 3 0.48 2
15 1 2 3 3 2 3 0.48 2
19 2 3 2 2 2 3 0.48 2
21 2 2 3 1 2 3 0.48 2
30 3 3 2 2 2 3 0.48 2
9 2 1 1 3 2 3 0.49 2
16 3 1 2 3 2 3 0.49 2
20 1 3 3 1 2 3 0.49 2
22 2 2 3 2 2 3 0.49 2
23 1 3 3 1 2 3 0.49 2
24 1 3 3 3 2 3 0.49 2
25 1 3 3 2 2 3 0.49 2
27 1 2 3 2 2 3 0.49 2
28 2 2 3 3 2 3 0.49 2
31 2 3 3 1 2 3 0.49 2
26 1 3 3 1 2 3 0.5 2
39 2 3 3 1 3 3 0.48 2
42 3 3 3 2 3 3 0.48 2
48 3 1 2 3 3 3 0.48 2
66 3 3 2 3 3 3 0.48 2
75 2 3 3 3 3 3 0.48 2
78 2 3 3 3 3 3 0.48 2
81 2 3 3 3 3 3 0.48 2
32 3 1 1 3 3 3 0.49 2
33 2 3 3 2 3 3 0.49 2
35 3 1 1 3 3 3 0.49 2
36 2 2 3 2 3 3 0.49 2
47 2 3 3 2 3 3 0.49 2
51 3 3 3 1 3 3 0.49 2
54 3 2 2 3 3 3 0.49 2
55 1 3 3 3 3 3 0.49 2
254
RUN EQ KW POC MRR Q AV CQF FL AV QFL AV
58 3 3 3 1 3 3 0.49 2
60 2 3 3 3 3 3 0.49 2
62 3 2 3 3 3 3 0.49 2
63 3 3 3 1 3 3 0.49 2
18 1 3 2 3 2 3 0.55 3
17 2 3 3 1 2 3 0.61 3
29 1 3 3 2 2 3 0.68 3
50 1 3 3 3 3 3 0.51 3
52 2 2 3 3 3 3 0.52 3
49 2 3 3 2 3 3 0.53 3
69 3 3 3 3 3 3 0.54 3
46 1 3 3 3 3 3 0.57 3
37 3 2 3 2 3 3 0.59 3
56 2 3 3 3 3 3 0.62 3
59 2 3 3 3 3 3 0.62 3
34 3 2 3 2 3 3 0.63 3
38 1 3 3 2 3 3 0.63 3
40 1 3 3 2 3 3 0.63 3
41 3 1 1 3 3 3 0.63 3
43 2 3 3 2 3 3 0.63 3
44 3 2 3 3 3 3 0.63 3
45 3 2 3 2 3 3 0.63 3
53 2 2 3 3 3 3 0.63 3
57 2 3 3 3 3 3 0.63 3
61 2 3 3 3 3 3 0.63 3
67 1 3 3 3 3 3 0.63 3
68 3 3 3 2 3 3 0.63 3
70 1 3 3 3 3 3 0.63 3
71 3 3 3 2 3 3 0.63 3
72 3 3 3 2 3 3 0.63 3
79 2 3 3 3 3 3 0.63 3
255
RUN EQ KW POC MRR Q AV CQF FL AV QFL AV
80 3 3 3 3 3 3 0.67 3
76 2 3 3 3 3 3 0.68 3
74 2 3 3 3 3 3 0.73 3
77 2 3 3 3 3 3 0.79 3
64 3 2 3 3 3 3 0.82 3
65 1 3 3 3 3 3 0.83 3
73 3 3 3 2 3 3 0.86 3
Table B-9: Energy consumption quality calculation for factorial design
Run Laser Power
Cutting Speed
A.Gas pressure
Standoff distance
Energy Consumed
1 100 0.2 0.5 1 8.33 2 100 0.2 0.5 5 8.33 3 100 0.2 0.5 10 8.33 4 100 0.2 2.5 1 8.33 5 100 0.2 2.5 5 8.33 6 100 0.2 2.5 10 8.33 7 100 0.2 4.5 1 8.33 8 100 0.2 4.5 5 8.33 9 100 0.2 4.5 10 8.33 10 100 0.7 0.5 1 2.38 11 100 0.7 0.5 5 2.38 12 100 0.7 0.5 10 2.38 13 100 0.7 2.5 1 2.38 14 100 0.7 2.5 5 2.38 15 100 0.7 2.5 10 2.38 16 100 0.7 4.5 1 2.38 17 100 0.7 4.5 5 2.38 18 100 0.7 4.5 10 2.38 19 100 1.2 0.5 1 1.39 20 100 1.2 0.5 5 1.39 21 100 1.2 0.5 10 1.39 22 100 1.2 2.5 1 1.39 23 100 1.2 2.5 5 1.39 24 100 1.2 2.5 10 1.39 25 100 1.2 4.5 1 1.39 26 100 1.2 4.5 5 1.39 27 100 1.2 4.5 10 1.39
256
Run Laser Power
Cutting Speed
A.Gas pressure
Standoff distance
Energy Consumed
28 300 0.2 0.5 1 25.00 29 300 0.2 0.5 5 25.00 30 300 0.2 0.5 10 25.00 31 300 0.2 2.5 1 25.00 32 300 0.2 2.5 5 25.00 33 300 0.2 2.5 10 25.00 34 300 0.2 4.5 1 25.00 35 300 0.2 4.5 5 25.00 36 300 0.2 4.5 10 25.00 37 300 0.7 0.5 1 7.14 38 300 0.7 0.5 5 7.14 39 300 0.7 0.5 10 7.14 40 300 0.7 2.5 1 7.14 41 300 0.7 2.5 5 7.14 42 300 0.7 2.5 10 7.14 43 300 0.7 4.5 1 7.14 44 300 0.7 4.5 5 7.14 45 300 0.7 4.5 10 7.14 46 300 1.2 0.5 1 4.17 47 300 1.2 0.5 5 4.17 48 300 1.2 0.5 10 4.17 49 300 1.2 2.5 1 4.17 50 300 1.2 2.5 5 4.17 51 300 1.2 2.5 10 4.17 52 300 1.2 4.5 1 4.17 53 300 1.2 4.5 5 4.17 54 300 1.2 4.5 10 4.17 55 500 0.2 0.5 1 41.67 56 500 0.2 0.5 5 41.67 57 500 0.2 0.5 10 41.67 58 500 0.2 2.5 1 41.67 59 500 0.2 2.5 5 41.67 60 500 0.2 2.5 10 41.67 61 500 0.2 4.5 1 41.67 62 500 0.2 4.5 5 41.67 63 500 0.2 4.5 10 41.67 64 500 0.7 0.5 1 11.90 65 500 0.7 0.5 5 11.90 66 500 0.7 0.5 10 11.90 67 500 0.7 2.5 1 11.90
257
Run Laser Power
Cutting Speed
A.Gas pressure
Standoff distance
Energy Consumed
68 500 0.7 2.5 5 11.90 69 500 0.7 2.5 10 11.90 70 500 0.7 4.5 1 11.90 71 500 0.7 4.5 5 11.90 72 500 0.7 4.5 10 11.90 73 500 1.2 0.5 1 6.94 74 500 1.2 0.5 5 6.94 75 500 1.2 0.5 10 6.94 76 500 1.2 2.5 1 6.94 77 500 1.2 2.5 5 6.94 78 500 1.2 2.5 10 6.94 79 500 1.2 4.5 1 6.94 80 500 1.2 4.5 5 6.94 81 500 1.2 4.5 10 6.94