vincenzo piuri, sicon/02, houston, tx, usa 18-21 november 2002 laser welding for automotive...
TRANSCRIPT
Vincenzo Piuri, Sicon/02, Houston, TX, USA 18-21 November 2002
LASER WELDING FOR AUTOMOTIVE COMPONENTS
This research has been carried out in collaboration with
Fiat Research Centre, Turin, Italy
Vincenzo Piuri, Sicon/02, Houston, TX, USA 18-21 November 2002
The Application
• The gear is built by joining two separated rings (a light syncronization gear and the principal gear)
• Welding is carried out with a CO2 laser
• Every product is tested using ultrasonic waves after welding for quality control
B
A
Y X
Z
Vincenzo Piuri, Sicon/02, Houston, TX, USA 18-21 November 2002
The Application• We wish on-line monitoring for
– welding quality assessment
– welding process monitoring (control)
• Welding problems are related to:
– Penetration depth
– Misalignment of coupling in mounted samples
– Porosity
– Power decrement up to 10%
– Power lack up to 10 ms
Vincenzo Piuri, Sicon/02, Houston, TX, USA 18-21 November 2002
The Application Requirements
• The error categories can be grouped in three classes: – Power Loss
– Mounting
– Porosity
• Requirements: – High monitoring performance
– Low computational load
Vincenzo Piuri, Sicon/02, Houston, TX, USA 18-21 November 2002
Signal Pre-Processing
• Simple Processing– Amplitude Demodulation
– Low Pass Filtering
• Fast Processing– 15K samples
Vincenzo Piuri, Sicon/02, Houston, TX, USA 18-21 November 2002
Feature Extraction
• Reference construction– a Cubic line has been
considered to interpolate the relevant interval of the weld watcher signal
• Processing– 1805697 Flops
Vincenzo Piuri, Sicon/02, Houston, TX, USA 18-21 November 2002
Power Errors and Features
T
F• Power Loss Errors
– Short Duration of Welding Process
– Laser Power Fluctuation
• Features– T: Duration of Effective Laser Power
– F: Maximum Power Fluctuation
T
GOOD NO GOOD
Vincenzo Piuri, Sicon/02, Houston, TX, USA 18-21 November 2002
Mount Errors and Features
• Mount Errors– Modulation in Weld
Watcher Signal
• Features– Parameters of the Cubic
line
– H-L: Cubic line Features
GOOD NO GOOD
H
L
Vincenzo Piuri, Sicon/02, Houston, TX, USA 18-21 November 2002
Porosity Errors and Features
GOOD NO GOOD
A
D
A
D
• Porosity Errors– Variations wrt to the reference signal
• Features– A: Amplitude of the discrepancy
– D: Time duration of the discrepancy
Vincenzo Piuri, Sicon/02, Houston, TX, USA 18-21 November 2002
Remarks
• Few samples are available to configure the solution
• Not all samples can be classified by the operator
• The distribution of samples for the different error typologies is unknown
Good No Good Not Classified TotalDepth Error 31 9 29 69Power Error 29 40 0 69Mount Error 47 8 14 69Porosity Error 275 10 60 345
Vincenzo Piuri, Sicon/02, Houston, TX, USA 18-21 November 2002
The Proposed Algorithm
NO GOOD
Welding SignalAcquisition
LowPass Filtering
Polynomial Fitting
NO GOOD
-
GOODNO GOOD
GOOD
Start
Laser PenetrationFeatures Extraction
Laser PenetrationClassification
GOOD
MountingFeature Extraction Power
Classification
Porosity Feature Extraction
PorosityClassification
NO GOOD
MountingClassification
Power SignalAcquisition
LowPass Filtering
Laser PowerFeatures Extraction
GOOD
GOOD
Vincenzo Piuri, Sicon/02, Houston, TX, USA 18-21 November 2002
Experimental ResultsClassifier Training Cross Validation Error of
best classifierAccuracy Interval Notes
KNN 40 samples(100%)
39 samples(~100%)
0 % ~ 0 –10 % K= 1Depth
FF-NN 28 samples(70%)
12 samples(30%)
0 % ~ 0 – 10 % Neurons= 2(Best over 100)
KNN 69 samples(100%)
68 samples(~100%)
0 % ~ 0 – 8 % K= 1Power
FF-NN 48 samples(70%)
21 samples(30%)
0 % ~ 0 – 8 % Neurons= 4(Best over 100)
KNN 55 samples(100%)
54 samples(~100%)
1.8 % ~ 0 – 10 % K= 1Mount
FF-NN 39 samples(70%)
16 samples(30%)
0 % ~ 0 – 8 % Neurons= 2(Best over 100)
KNN 215 samples(100%)
214 samples(~100%)
0.35 % ~ 0 – 4 % K= 1Porosity
FF-NN 199 samples(70%)
86 samples(30%)
0 % ~ 0 – 4 % Neurons= 4(Best over 100)
Vincenzo Piuri, Sicon/02, Houston, TX, USA 18-21 November 2002
SPARKS ANALYSIS FOR LASER CUTTING
This research has been carried out in collaboration with
TRUMPF, Ditzingen (Stuttgart), Germany
Vincenzo Piuri, Sicon/02, Houston, TX, USA 18-21 November 2002
The Application
• Laser cutting of steel/stainless steel is a complex process
• It is expected that monitoring of the sparks dynamic associated with the cutting process can provide hints about– The internal nature of the cutting process
– Indications for subsequent process monitoring and control
Vincenzo Piuri, Sicon/02, Houston, TX, USA 18-21 November 2002
The Application Requirements
• There are three cutting error categories:– Good
– No Good • Discontinuous cut
• Pearls of metal
– Ambiguous
• Requirements: – High Accuracy
– Low computational load
Vincenzo Piuri, Sicon/02, Houston, TX, USA 18-21 November 2002
Composite System Partitioning
FEATURE EXTRACTION
SC
CLASSIFIER
PEARL
CONTROL
NoGood
Good
Ambiguous
Jet /no Jet
, ,
Cut speed, gas used,...
Composite System
No Jet
Vincenzo Piuri, Sicon/02, Houston, TX, USA 18-21 November 2002
Features Extraction
beta sxbeta sx
gammagamma
beta dxbeta dx
AlphaAlpha
gammagamma
: inclination angle : opening angle of the main jet : opening angle of the whole jet
Vincenzo Piuri, Sicon/02, Houston, TX, USA 18-21 November 2002
Finding the Jet Starting Point
100 200 300 400 500
100
200
300
400
0 100 2000
2
4
6x 10
650 100 150
100
200
300
400
500
600
100 200 300 400 500
100
200
300
400
100 200 300 400 500
100
200
300
400
0 100 2001
2
3
4
5x 10
750 100 150
100
200
300
400
500
600
100 200 300 400 500
100
200
300
400
Radon transform
Profile extraction
Direction of the main jet
Vincenzo Piuri, Sicon/02, Houston, TX, USA 18-21 November 2002
Estimating the Angles
THE AND ANGLESTHE AND ANGLES
• Median filtering• Threshold binarization• Cumulate intensity in rows• Find left/right edges of the spark• Separated left/right linear regression passing trough the vertex
Vincenzo Piuri, Sicon/02, Houston, TX, USA 18-21 November 2002
Pearl Identification
FeedForward Neural Network
(2 hidden units,1 output good/no good unit)
NEURAL
NETWORK
Vincenzo Piuri, Sicon/02, Houston, TX, USA 18-21 November 2002
The Final SystemSLAPS-EU
Activity Status Report - 10 March 2000
Camera
Sensors
Sampler
Movie
Information fromthe field
N images + Information from the sensors
Classification Supervisor
Classification Data file
Index evaluationImage
Evaluation JET-P.
JET?
JET presence
Presence ofPearls of burr
NO
Image Classifier
YES Vertex evaluationalpha
beta
gamma
(x,y)
alpha
alpha
beta
gamma
Classificationof a singleimage
Metal typeMetal thicknessCut speedFocus positionGas typeGas pressureLaser powerNozzleCut distance
Final ClassifierN Classifications & JET-presence
JET presence
Final CutClassification
Classification system
A
B
B1
C
B2
B3
Vincenzo Piuri, Sicon/02, Houston, TX, USA 18-21 November 2002
Performance
NotePerformance
Using human estimates over 121 imagesthe behavior of the angles-processing module fits suitably the sparks
Error < 3°processing
, ,
Using validation images30/30Pearls
84/84Classification good/no good/ambiguous
Vincenzo Piuri, Sicon/02, Houston, TX, USA 18-21 November 2002
LASER SPOT WELDING FOR ELECTRONIC COMPONENTS
This research has been carried out in collaboration with
Philips CFT - Centre for Industrial Technology
Philips Centre for Industrial Technology
Vincenzo Piuri, Sicon/02, Houston, TX, USA 18-21 November 2002
Electron Gun for Cathode Ray Tube
1 - Generation of free electrons by cathode2 - Beam shaping using ‘electric field lenses’3 - Acceleration of electrons
Hdeflection
Vincenzo Piuri, Sicon/02, Houston, TX, USA 18-21 November 2002
Manual Classification
Top viewspot weld
Bottom viewspot weld OK
Bottom viewspot weld bad
Acceptable gap
Too large gap
Vincenzo Piuri, Sicon/02, Houston, TX, USA 18-21 November 2002
CCD Camera
On-AxisLaser Reflection
Laser OutputMonitor
TemperatureSensor (2x)
Off-AxisReflection
Plume(not visible)
FiberInserts hereLaser power input
Fibre
On-axisreflection
Camera
Laser powermonitoring
X-Y scannermirrors
Off-axisreflection
SoundThermalemission
Plumeemission
Work piece
Vincenzo Piuri, Sicon/02, Houston, TX, USA 18-21 November 2002
Feature Extraction
Spot welding of Grid 1
-0.5
0
0.5
1
1.5
2
2.5
3
3.5
-1.648 -0.848 -0.048 0.752 1.552 2.352 3.152 3.952 4.752 5.552 6.352 7.152
Time [ms]
Volts
Laser input power
On-axis reflection
Off-axis reflection
Plume emission
Temperature
cooling slope
turning point
time of firston-axis minimum
Process starting time
Meanpower level
Vincenzo Piuri, Sicon/02, Houston, TX, USA 18-21 November 2002
Automatic Classifiers
Pros Cons
NeuralNetworks
High accuracy Requires propertraining (and re-training)
Difficult to refuseclassification
Estimation ofclassificationaccuracy only
K-NearestNeighborClassifier
No trainingrequired
Does not necessarilylead to classifications
Vincenzo Piuri, Sicon/02, Houston, TX, USA 18-21 November 2002
Results from theNearest Neighbour Classifier
1NN 2NN
StripStrip (3)
Bracket
Bracket(15)
CuramikCuramik(18)
MAI (32)
MAI (18)
98.15 1.8597.67 2.33
75.92 24.08
75.44 24.56
86.04 13.96
78.65 21.35
77.37 22.63
84.86 15.14
95.74 3.24 1.0295.63 3.20 1.17
61.53 25.43 13.04
62.25 25.04 12.71
73.06 19.64 7.3075.01 17.14 7.85
64.22 23.73 12.05
63.70 24.75 11.55
O.K. O.K. N.O.K.N.O.K. ??