tool condition monitoring - tcm.pdf · École polytechnique montréal, politechnika warszawska...

33
Prof. Krzysztof Jemielniak Tool Condition Monitoring École Polytechnique Montréal, Politechnika Warszawska Outline TCM systems objectives & structure Process variables used in TCM Strategies of commercial TCM systems New strategy of tool wear monitoring Catastrophic tool breakage detection Concluding remarks

Upload: vuongnga

Post on 28-Feb-2019

221 views

Category:

Documents


0 download

TRANSCRIPT

1

Prof. Krzysztof Jemielniak

Tool Condition Monitoring

École Polytechnique Montréal, Politechnika Warszawska

Outline

• TCM systems objectives & structure

• Process variables used in TCM

• Strategies of commercial TCM systems

• New strategy of tool wear monitoring

• Catastrophic tool breakage detection

• Concluding remarks

2

École Polytechnique Montréal, Politechnika Warszawska

Tool life estimation

2) small batch production,

direct operator attendance

Indirect methods

3) Large scale productionTime, number of parts

freq

uen

cy

t

ool w

ear

time, No of parts

tool life

4) Automatic, unmanned production

Indirect methods

1) Laboratory

direct methods

École Polytechnique Montréal, Politechnika Warszawska

Methods used for tool wear estimation

Direct methods:(measurement of wear criteria)

Optical methods

Touch probes

Indirect methods:(changes in process variables caused by tool wear)

Acoustic Emission

Cutting forces and forcerelated measures (torque, motor current, true power)

Vibration and noise

3

École Polytechnique Montréal, Politechnika Warszawska

Definition of TCM

Tool (and Process) Condition Monitoringrefers to data gathered by sensors that is used to determine the status and performance about:

• Machine tool

• Cutting tool

• The machining process

École Polytechnique Montréal, Politechnika Warszawska

Tasks of TCM systems

Tool condition monitoring:

tool wear estimation (detection of the tool life end )

detection of catastrophic tool failure (CTF)

Detection of extensive vibration

Detection of collisions

Others (e.g. diagnostics of chip shape, detection of BUE, burrs, missing tool of workpiece, etc).

4

École Polytechnique Montréal, Politechnika Warszawska

Objectives of TCM

It allows you to:• Protect machines • Determine cutting tool status • Improve efficiency • Decrease the cycle time • Improve quality • Protect tooling and setups • Increase reliability • Optimize the process • Reduce rework • Lower scrap

École Polytechnique Montréal, Politechnika Warszawska

Tool Condition Monitoring System Configuration

5

École Polytechnique Montréal, Politechnika Warszawska

Structure of Tool & Process Condition Monitoring System

process variables

Cutting zone

sign

als

sensorsData Acquisition & Signal processing:filters, statistics

FFT, RMS,...

sign

alre

pres

enta

tion STRATEGY

Proces model,

knowledge

diagnosis,command

ACTION !

École Polytechnique Montréal, Politechnika Warszawska

Process variables used in commercial Tool & Process Condition Monitoring Systems

6

École Polytechnique Montréal, Politechnika Warszawska

Influence of Tool Wear on Cutting Forces in Turning

École Polytechnique Montréal, Politechnika Warszawska

Tool condition monitoring in drilling operations, based on signal amplitude

7

École Polytechnique Montréal, Politechnika Warszawska

Axial forces and torque for sharp and dull drills

new tools

new tools

feed

for

ce F

f(N

)

École Polytechnique Montréal, Politechnika Warszawska

Influence of CTF on power spectrum in milling

time (s)

cutt

ing

forc

e (*

0.2

kN

)

tool breakage

frequency (Hz)

frequency of edge passing

frequency of edge passing

revs

revs

nor

mal

i zed

pow

ern

o rm

ali z

ed p

ower

8

École Polytechnique Montréal, Politechnika Warszawska

Cutting Forces During CTF (Chipping) in Turning

chipping

chippings

École Polytechnique Montréal, Politechnika Warszawska

Cutting Forces During CTF (Breakage) in Turning

chipping

breakage

breakage

9

École Polytechnique Montréal, Politechnika Warszawska

Tool wear and AERMS parameters

workpiece 45,

tool: SNUN 120408 NT35;

vc= 240 m/min, ap= 2.4 mm, f = 0.3 mm/obr;

tool wear (mm)burst rate (1/s)

burst width (%)

threshold (V)

threshold (V)

École Polytechnique Montréal, Politechnika Warszawska

Development of AE until drill breaks

10

École Polytechnique Montréal, Politechnika Warszawska

Tool Life Monitoring of Small HSS Drills Using Vibration Signal

École Polytechnique Montréal, Politechnika Warszawska

Strategy of TCM Systems

processvariables

Cutting zone

sign

als

sensorsData Acquisition & Signal processing:filters, statistics

FFT, RMS,...

sign

alre

pres

enta

tion

STRATEGY

Proces model,

knowledge

diagnosis,command

ACTION !

11

École Polytechnique Montréal, Politechnika Warszawska

Nordman

École Polytechnique Montréal, Politechnika Warszawska

Prometec

12

École Polytechnique Montréal, Politechnika Warszawska

Artis

wear alarm (area exceeds wear limit).

area signal learned=100%,

larger area signal (e.g. through worn tools)

learnt area (bar diagram),

pre-alarm limitwear limit

pre-alarm (area exceeds wear limit)

École Polytechnique Montréal, Politechnika Warszawska

Montronix T100

13

École Polytechnique Montréal, Politechnika Warszawska

Brankamp

École Polytechnique Montréal, Politechnika Warszawska

Experiences from Brankamp Pilot Installation in Raufoss

• Installed on mechanically controlled 6-spindle automatic lathe

• 4 piezo-electric sensors:– Drilling tool on position 1– Drilling tool on position 2– Bar cut-off tool in position 6– Bar feed stopper

14

École Polytechnique Montréal, Politechnika Warszawska

Experiences from Brankamp Pilot Installation in Raufoss

R A U F O S S M A T E R I A L T E K N O L O G Iside 11

Experiences

+ No multiple tool breakage since installation+ Improved quality output, especially on

randomly occurring faults- Difficult to find correct monitoring limits: Trial

and error -approach requires skilled operators and continuity in process and operators

- Complex system, requires high level of training: problems at sick leave, vacation etc.

École Polytechnique Montréal, Politechnika Warszawska

Existing TCM systems

• TCM systems are based on phenomena correlated with tool wear (cutting force, AE, vibration)• measured signal feature (SF) depends on other process

parameters• relationship tool wear vs. measured feature is very complex • it has a statistical nature

• Reliable tool wear evaluation based on one signal feature is impossible• most commercial and experimental systems apply "one

process – one SF" strategies• basic assumption: the feature is a monotonic function of

the tool wear

• This can be the reason of still not satisfactory performance of the systems applied in industry

15

École Polytechnique Montréal, Politechnika Warszawska

Solution?

• Combination of different signal features • statistical methods, • auto-regressive modeling, • pattern recognition, expert systems • neural network - most often used

• Usually single neural network• several SFs are fed into the network inputs,• tool wear estimation is the network output.

• In some works, hierarchical tool wear monitoring strategies were proposed

• advantages of such approach will be presented here

École Polytechnique Montréal, Politechnika Warszawska

Questions

Does operator have to find and enter limit values, points of measurements etc?

Why the operator have to know anything about the signals values?

300-250∆T = ----------- =....

400-250

He can calculate used up part of tool life! 0,33

NO!

???

What can he do with this knowledge?

16

École Polytechnique Montréal, Politechnika Warszawska

Tool Condition Indicator

What is interesting for machine tool operator?

used up part of the tool life:

∆T = t / Twhere t – cutting time as performed so far

T – tool life

In laboratories: tool wear measures (VBB, KT)

crater on rake face

wear land of flank facenotch

wear

notch wear

not used in factory floor

École Polytechnique Montréal, Politechnika Warszawska

Basics of TCM Learning

SF1

SF

The first tool life

1st cycle.: No of cuts, min and max of the signals

n= 1

After each cycle: store the measure

value SFOp

SF2SF3

2 3 ...

. . .

After the end of 1st tool life transform array SFOp into SF∆T using:

n∆T= ---

N

SFN

N

Sequential tool life

SF1

SF

After each part: store SFOpand evaluate ∆T using SF∆T

SF2SF3

. . .

n= 1 2 3 ...

After the end of tool life transform array SFOp into SF∆T,

learn (gather experience)

SFN

N

17

École Polytechnique Montréal, Politechnika Warszawska

Transformation SFOp into SF∆T

1. Smoothing SFop to eliminate random changes of SF: SFOp SFOpFlt

École Polytechnique Montréal, Politechnika Warszawska

Transformation SFOp into SF∆T

2. Assignment ∆T to each operation number (∆T =n/N) and transforming SFOpFltinto 20 elements array: SFOpFlt SF∆Tcurrent

18

École Polytechnique Montréal, Politechnika Warszawska

Transformation SFOp into SF∆T

3. Taking into account previous experience:

SF∆T = 0.25 SF∆Tcurrent + 0.75 SF∆Tprev

École Polytechnique Montréal, Politechnika Warszawska

SF∆T

∆T

Evaluation of Used Up Part of Tool Life

∆T

SF

n

Search out in array SF∆T value closest to SF obtained during machining last part

SF∆T

∆T

SF

n

The search starts from previous result: if recent value is lower, ∆T does not changes

∆T

Search only 6 elements of the array (30%T) reduces influence of accidental high values

SF∆T

∆T

∆T

SF

n

30%

19

École Polytechnique Montréal, Politechnika Warszawska

SF∆T

∆T

Evaluation of Used Up Part of Tool Life

SF

∆T

30%

Evaluation cannot be lower then 0.7 ∆T resulting from previous experience

SF∆T

∆T

∆T=70*n/NB

SF

n

Limited search of the SF∆T array enables, at least to some extent, to utilize the signal features which are not monotonic versus used up portion of the tool life

École Polytechnique Montréal, Politechnika Warszawska

Machine Tool & Sensors

Turning Center Venus 450

AE sensor

Cutting force sensor

20

École Polytechnique Montréal, Politechnika Warszawska

Workpiece & Tool

Workpiece: steel C 45 bars, 160 mm machined down to 85 mm

ap=1,5 and ap = 2 mm

f = 0.1 mm/rev

vc = 150 m/min

École Polytechnique Montréal, Politechnika Warszawska

Results of Experiments

21

École Polytechnique Montréal, Politechnika Warszawska

Results of Experiments

• 88 operations

• 8 worn out tools

• 1st tool used for system learning

• 7 tools used for estimation of tool wear monitoring accuracy

École Polytechnique Montréal, Politechnika Warszawska

Results – Signal Feature Value vs. Operation Number

The average value of the feed force versus number of operation in three selected tool lives

22

École Polytechnique Montréal, Politechnika Warszawska

Results – Transformation SFOp into SF∆T

After the 1st tool life:

1. transformation of SFOp into SFOpFlt

2. transformation of SFOpFlt into SF∆T

After the 2nd tool life:

1. transformation of SFOp into SFOpFlt

2. transformation of SFOpFlt and SF∆Tprevinto SF∆T

École Polytechnique Montréal, Politechnika Warszawska

Results – Evaluation of Used Up Part of Tool Life

23

École Polytechnique Montréal, Politechnika Warszawska

Signal Features

maximum another

average 2smedian

2s

varianceaverage

Altogether 98 signal features obtained from 4 signals – Ff, Fp, Fc and AERMS

École Polytechnique Montréal, Politechnika Warszawska

Feature Selection

2nd degree polynomial approximation

Fc,MaxFfMed2s

FfMed2sFc,Max

Root Mean Square Error (RMSE), threshold = 0.05

0,046113 0,189998

RMSE=n

TTn

iii∑

=

∆−∆1

2)(

i – number of operationn – number of operations in tool wear

24

École Polytechnique Montréal, Politechnika Warszawska

Feature Selection

0,082725Variance(Fc (j))FcVar

0,077072Variance(AE(l))AEVar2s

0,070857RMS(Ff(j))FfRMS

0,062425Median(AE(l))AEMed2s

0,062076Mod(Ff(l))FfMod2s

0,061438Max(Ff(l))Ffmax2s

0,058544Min(Ff(j))Ffmin

0,055442Sum(Ff(j))/1000Ffsred

0,053809Max(Ff(k))-Suma(Ff(l))/1000Ffmaxpocz-sred

0,047488Median(Ff(j))FfMed

0,047267Sum(Ff(l))/1000Ffsred2s

0,046814RMS(Ff(l))FfRMS2s

0,046113Median(Ff(l))FfMed2s

0,044602Min(AE(l))AEmin2s

0,044247Median(AE(j))AEMed

0,030035Moment 3 degree(Ff(j))FfMom3

0,025272Moment 4 degree (Ff(j))FfMom4

0,021929Standard DeviationFf(j))FfStDev

0,020465Variance (Ff(j))FfVar

RMSE formulaFeature

RMSE<0,05

École Polytechnique Montréal, Politechnika Warszawska

FfRMS2s

FfSred2s FfMed

Elimination of Similar Features

RMSE between features

0,009750Criterion: RMSE < 0.1

0,003096

0,002382

FfMed2s

25

École Polytechnique Montréal, Politechnika Warszawska

min. of AERMS from the 2nd sec.variance of Ff

Feature Integration

FfMed2s

FfVar

AEMed

?

Neural networks

median of Ff from the 2nd sec. median of AERMS

AEMin2s

École Polytechnique Montréal, Politechnika Warszawska

Feature Integration

Artificial Neural Network Feed Forward Back Propagation (FFBP)

• 10 neurons in hidden layer (tested 4 to 20 neurons)

• Learning rate 0,14 (tested for 0,02 to 0,3 with step 0,02)

• Momentum rate 0,65 (tested for 0,05 to 0,9 with step 0,05)

26

École Polytechnique Montréal, Politechnika Warszawska

Feature Integration by Neural Networks -Results

FfMed2s FfVar

Untypical course of SF

dominates the network’s

output and can not be

compensated by typical ones

∆TNN

École Polytechnique Montréal, Politechnika Warszawska

Feature Integration by Neural Networks –reasons of failure

Number of learning cases is not big enough

The system should be trained (learnt) using the data from the first tool life only Number of inputs is

dependent on number of learning cases

Number of features must be limited

27

École Polytechnique Montréal, Politechnika Warszawska

Hierarchical Algorithm

1. Estimation of ∆T from all signal features separately

2. Integration of estimations– neural network

– averaging without results outside 3σ

École Polytechnique Montréal, Politechnika Warszawska

Hierarchical Algorithm

1. Estimation of ∆T based on single feature

FfMed2s

FfVar

AEMed

AEMin2s

28

École Polytechnique Montréal, Politechnika Warszawska

Hierarchical Algorithm

NN

2. Integration of estimations

Average

École Polytechnique Montréal, Politechnika Warszawska

Tool Wear Estimation – Conclusions

1. Assumption about monotonic, increasing character of SF(∆T) function, excludes many useful features. Taking full advantage of all SFs correlated with the tool condition can significantly improve TCM systems reliability.

2. The difficulties related to reversing of non-motonic SF(∆T) functions can be overcome by transforming them into array and limited range search of SF value closest to the measured one.

3. The signal feature integration minimizes the diagnosis uncertainty, reducing the randomness in one SF and providing more reliable tool condition estimation.

4. Decomposition of the multi SF tool wear estimation into hierarchical algorithms has several advantages over the single step approach:

• many more SFs can be used,

• non-monotonic SF(∆T) functions can be used

29

École Polytechnique Montréal, Politechnika Warszawska

Cutting Force Change Pattern Accompanying Two Types of CTF

École Polytechnique Montréal, Politechnika Warszawska

CTF Detection Based on Limit Value

assumed

actual

30

École Polytechnique Montréal, Politechnika Warszawska

CTF Detection Based on Pattern Recognition

École Polytechnique Montréal, Politechnika Warszawska

Strategy of CTF Detection Developed in TH Aachen

31

École Polytechnique Montréal, Politechnika Warszawska

Strategy of CTF Detection Developed in WUT

École Polytechnique Montréal, Politechnika Warszawska

Evaluation of the WUT CTF Detector Performance – Example 1

Tool breakage followed by the chipping of the cutting edge

C45, SNUN S30S

32

École Polytechnique Montréal, Politechnika Warszawska

Evaluation of the WUT CTF Detector Performance – Example 2

Cutting force and CTF detector reactions during stepwise change of the depth of

cut and the cutting edge chipping

ZL25M, SPUN H10S

École Polytechnique Montréal, Politechnika Warszawska

Dynamic Limit Strategy for CTF Detection in Milling

33

École Polytechnique Montréal, Politechnika Warszawska

Conclusions – problems in TCM

• TCM systems are based on phenomena correlated with tool wear (cutting force, AE, vibration)• basic assumption: the feature is a monotonic function of

the tool wear• relationship tool wear vs. measured feature is very complex • it has a statistical nature

• Reliable tool wear evaluation based on one signal feature is impossible• most commercial and experimental systems apply "one

process – one SF" strategies• this can be the reason of still not satisfactory performance

of the systems applied in industry

• Detection of catastrophic tool failure based on constant limit strategies (most often used) can detect only a major tool breakage, usually far too late

École Polytechnique Montréal, Politechnika Warszawska

Conclusions – possible solutions

• Reliable tool wear monitoring can be achieved by combination of different signal features • neural network - most often used

• number of features must be limited, thus NN seem to be not effective enough

• Decomposition of the multi SF tool wear estimation into hierarchical algorithms has several advantages over the single step approach: • many more SFs can be used, • non-monotonic SF(DT) functions can be used

• Detection of catastrophic tool failure is possible only using dynamic limit strategies