Detection of laser-welding defects using neural networks BY Marc Auger A thesis submitted to the Department of Mechanical Engineering in confomiity with the requirements for the Degree of Master of Science (Engineering) Queen's University Kingston. Ontario. Canada September. 200 i Copyright O Marc Auger, 201
134
Embed
Detection of laser-welding defects using neural networks
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Detection of laser-welding defects using
neural networks
BY Marc Auger
A thesis submitted to the Department of Mechanical
Engineering in confomiity with the requirements for the
Degree of Master of Science (Engineering)
Queen's University
Kingston. Ontario. Canada
September. 200 i
Copyright O Marc Auger, 2 0 1
National Libraiy 191 .canada Bibliothèque nationaie du Canada
Acquisitions and Acquisitions et Bibliographie Services services bibliographiques
The author has granted a non- exclusive licence dowing the National Library of Canada to reproduce, loan, distribute or seil copies of this thesis in microform, paper or electronic formats.
The author retains ownership of the copyright in this thesis. Neither the thesis nor substantial extracts &om it may be p ~ t e d or othecwise reproduced without the author's permission.
L'auteur a accordé une licence non exclusive permettant à la Bibliothèque nationale du Canada de reproduire, prêter, distribuer ou vendre des copies de cette thèse sous la forme de microfiche/film, de reproduction sur papier ou sur format électronique.
L'auteur conserve la propriété du droit d'auteur qui protège cette thèse. Ni la thèse ni des extraits substantiels de celle-ci ne doivent être imprimés ou autrement reproduits sans son autorisation.
To Tiffany
Laser welding is becoming more and more important in the automotive industry
and qudity of the weld is critical for a successful application. In many cases, the increase
in welding speed provided by laser welding has caused the welding system operator to be
unable to keep up with the production rate while fully inspecting each part. Therefore,
either additional inspecton are required or some fom of real-time on-line inspection of
the weld must be provided. This is especidly necessary where the laser weld propenies
are critical to the final performance.
This thesis describes a system for the prediction of various panmeters of the
fusion zone of a weld from the emitted radiation during laser welding. A neural network
system is used to associate data from three photodiode senson to geometrical properties
of the fusion zone rneasured in cross-section.
A machine welding automotive transmission gears with a CO2 laser was used to
test the system. The neural network system was able to predict. with acceptable accuracy,
two of the most important parameten describing the geometry of the fusion zone: the
total area and the lateral position of the fusion zone relative to the weld seam. The system
shows promise in king able to predict unacceptable welds if incorporated as part of an
on-Iine quality monitoring process.
Acknowledgements
The author wishes to express his sincerest gratitude to his thesis supervisor for
their support and guidance throughout this research. Prof. P.M. Wild and Prof. A.
Ghasempoor. This work would not have been possible without the assistance of ATC
Powerlasers of Kitchener, Ontario who donated time and technical suppon on the
experimental appmus used for dl experiments. Special thanks goes out to Rob Mueller
and H o n ~ i n g Gu for their advice. rxperience and knowledge with regards to laser
welding. The assistance provided by George Pinho in performing al1 of the tests was
gatefully appreciated and duly noted. A special thanks is also given to Jack Evanecky of'
DaimleiChrysIer. Kokomo. Indiana for donating al1 the material used in the experi ments
and the use of his equipment. The shll and talent of Chris Howes and Charlie Cooney in
the Metallurgy lab were indispensable during the Iaboratory analysis. This work would
not have k e n possible without the financial support of the Centre for Automotive
Materials and Manufactunng, Kingston. Ontario. Finally, 1 would like to thank my wife
for her suppon and understanding throughout the duration of this work.
Acknow ledgem~nîs ........... .. ....... .................... ...... ...................... iv
.............. Table of Contents ... ......... ............................... .... ......................................................................... v
List of Tables .. ............. ..... .... ........o.........mw........................................................................................... ix
.. .o.. List of Figures ............................................ "...*...w....m....-..m.w........w................................................... x
.......... Appendix A ........................ ................. ................................................................ ......... 115
Appendix B ............................................................................................................................................ 116
............................................................................................................. Appendix C .......................... ..... 117
Vita ...*............... .... ................................................ ......... ..... ... ........................................ 119
List of Tables
Table 3.1 . Example of limits used for selected spectrum analysis coefficients .............. 52
Table 1.1 . Typical parameten in a rotary laser welder ................................................... 5 8
List of Figures
7 ....................................................... Figure 1.1 . Example of single-sided laser welding [3]
..................... Figure 1.2 . Tailor welded blank and siamped part of a dwr inner panel [5] 4
Figure 4.3 . Weld sequence on an un-weided plate ......................................................... 56
........................................................................... . Figure 4.4 Shield gas nozzle location 57
Figure 1.5 - Responses curve of the three photodiodes of the ATC WPM system [82] .. 59
Figure 4.6 - Weld Process Monitor from ATC Powerlasers [82] ..................................... 60
Figure 4.7 - Close-up of the WPM display screen [82] .................................................... 61
................................................... Figure 4.8 . Weld on tab only (a) weld on disk only (b) 62
Figure 4.9 - Proper lateral location for weld ................................................................. 62
Figure 4.10 . Full penetration (a) and partial penetration (b) as viewed from the underside
of the joint .......................................................................................................... 63
Figure 4.1 1 . Close-up of top surface: g d (a) pinholes (b) ........................................... 64
Figure 4.12 . Close up of top surface imperfections: concavity (a) and convexity (b) .... 64
Figure 4.13 . Close-up of bottom surface: good (a) and pinholes (b) .............................. 64
Figure 4.14 . Close-up of bottom surface imperfections: partial penetration (a) and
excessive material ejection (b) ................................................................................. 65
Figure 4.15 . Scribing the tab number (a) . wet-saw used for sectioning (b) .................... 66
Figure 4.16 - Sections mounted in epoxy puck (a) information tag (b) ........................... 66
Figure 4.17 -Stereo zoom microscope with ring light (a) image from digital camera (b) 67
Figure 4.18 - Weld area measurement with the outline tool in lmage~roQ Plus .............. 68 Q Figure 4.19 . Thickness measurement in ImageRo Plus ............................................... 69 0 ............................................ Figure 4.70 - Single point measurement in Imageho Plus 70
Figure 4.2 1 - Geometnc properties of the weld area [83] ................................................ 71 0 Figure 4.22 . Hole measurement in ImagePro Plus ........................................................ 72
PSE (total error 40 hidden nodes) .......................................................... 94
Cornparison between PSI and PSE (optimized total training error) ........... 94
Cornparison between PSI and PSE (optimized test emor) .......................... 95
PSE (individual training error) ................................................................... 96
PSE (individual test error) .......................................................................... 96
Cornparison between PSI and PSE (area training emor) ............................ 97
Cornparison between PSI and PSE (area test error) ................................... 97
Cornparison between PSI and PSE (lateral position vaining error) ........... 98
Cornparison between PSI and PSE (lateral position test error) .................. 98
Nomenclature
L-KH W-KH A-KH L-WP W-WD A-WP
length of the keyhole width of the keyhole m a of the keyhole length of the weld pool width of the weld pool width of the keyhole neural network input neural network weight neural network output neural network function maximum output value of a sigmoid function minimum output value of a sigmoid function initial value of a single input panmeter minimum value of a single initial input parameter maximum value of a single initial input panmeter normalized value of a single input parameter number of input nodes in a neural network number of output nodes in a neural network number of hidden nodes in a three-Iayer neural neiwork number of hidden nodes in a four-layer neural network target value for an output actual value for an output number of sarnples number of output varaiables
xiv
Chapter 1 Introduction
1.1 Background
The automotive industry has a significant impact on the Canadian economy, as i t
comprises a significant portion of the manufactunng GDP in Canada - (12.888 in 1998
[l]). More specifically. the province of Ontario is home to assembly plants representing
six different automotive manufacturers: Ford, Gened Moton. DaimlerChrysler, Honda,
Toyota and CAMI (a joint venture between GM and Suzuki). Parts rnanufacturen and
supplien alsc have plants spread throughout the province.
In the global marketplace. a Company must be able to produce a quality
component at a reasonable pnce to stay in operation. Cornpliance with quality standards
such as ISO 9001 and QS-9000 are now required for automotive suppliers to compete
intemationally. To remain competitive. companies must find ways of not only irnproving
part quality, but also reducing costs. Two of the most popular methods of accomplishing
these goais are automation and increased quality control. Automation is a popular choice
as it allows for the improvement in quality through the elimination of hurnan labour (and
thus human error), while simultaneousl y increasing production rates. Increased quality
control is another option as it can be implemented in both manual and automated systerns
1
to identify nonîonformances at various stages of the manufactunng process. The
availability of a wide variety of senson and monitoring equipment. dong with today's
high-powered cornputers. enables many different inspection and quality control
techniques.
1.2 Laser Welding in the Automotive lndustry
Laser welding of components is a highly automated process that is now
widespread in the automotive industry. In fusion welding processes. parts are joined by
heating such that the interface between the parts melts and mixes before cooling
(complete fusion) [2]. Cost savings and irnproved quality c m be achieved by switching
from traditional fusion welding techniques. such as resistance. MIG, and TIG, to high-
power laser welding (above 1 kW). Single-sided mess of lasers makes it possible to
create weld geometries impossible to achieve with conventional two-sided resistance
welding techniques. Ioining a piece of sheet metal to a hydroformed tube (Figure i. 1) is
an example of a single-sided weld geometry that is possible only with laser welding.
Section A-A
Top access only
Fipre 1.1 - Example of single-sided laser welding [3]
The cost savings possible with laser welding are achieved t h u g h increased
welding speed and decreased consurnables and downtime'. Some consumables found in
tnditional welding techniques are: copper tips in resistance spot welding, shield gas and
filler wire for MIG welding, tungsten elecrodes. shield gas and filler material for TIG
welding. For most steels. filler materid is not required for laser welding. Alurninum
components almost always require a tiller materid to ensure that the chernical
composition of the weld remains favounble. Research suggests that laser welding of steel
with no shield gas or a cheaper shield gas (COz vs. Ar) may be possible [4]. Despite the
advantages of laser welding, the initial capital expenditure to acquire a laser welding
system has delayed its adoption in some automotive applications.
1.3 Examples of Laser Welding in the Automotive lndustry
Laser welding has been used for the welding of transmission components for
more than 30 years. The laser welder. which replaced electron beam welders in this
application, is fixed while the cylindncal transmission components are rotated in order to
be welded. Electron bearn welden are powered by large transfomen. require thick lead
shielding and are usually opented in a vacuum environment. Laser welders, on the other
hand. are smaller and only require optical shielding.
Sheet metai welding is the fastest growing segment of laser welding usage,
beginning in the automotive industry in the late 1980s. m e joining of two or more pieces
of flat sheet metai to mate a railored blank before stamping is a relatively new approach
t The amount of cime a piece of equipment is out of service for tepair or replacement of consumables.
to the manufacturing of body panels and has been adopted by almost al1 of the major
automotive manufacturen. Typically, the weld is a stnight line. However laser
inteptors' increasingl y offer two-dimensional welding systems.
Figure 1.2 - Tailor welded blank and stmped part of a door inner panel [SI
The Auto/SteeI Partnership [ 5 ] h a identi fied major incentives for using tailored
blanks. Weight reduction cm be achieved by using thick material only where necessary.
such as a door inner where a thicker strip of material is located on the hinge side and a
thinner. lighter piece of material is used for the rest of the door (Figure 1.2). Part
elimination is achieved because extra reinforcements are no longer required, with the
added advantage of reducing die investment and assernbly costs. An increase in structural
integrity without weight gain results from a continuous part, which also improves the
dimensional control of the final assernbly. Better material utilization results as smaller
h laser integrator is a Company that combines an off-the-shelf laser with tmling and automation equipment of their own design for a rnanufacturing faciIity at a separate site.
pieces normally discarcied, such as window or door cuiouts from liftgates and bodysides,
cm be welded together to create a tailored blmk for another part.
The most recent use of laser welding in the automotive industry has been in the
joining of stamped parts. Attaching the roof of a vehicle to the rest of the body is one of
the more difficult applications of laser welding since it involves three-dimensional
geometries (Figure 1.3). Volvo. Daimlefhrysler and Volkswagen are a few of the
manufûcturen that have adopted this technoiogy.
Figure 1 3 - Roof panel l a x r welding at Volvo 161
1.4 Quality Control in the Manufacture of Laser Welded
Components
Many quality monitoring systems have ken. and are being. developed for laser
welding. Early quality monitoring systems solely relied on destructive testing of the
completed parts. This method was time consuming. expensive, and required dedicated
test equipment and personnel with only a fraction of the total production king inspected.
Non-destructive testing has reduced the need for. and frequency of, destructive testing but
has not cornpletely eliminated it. Tensile tests of conventional dog-bone shaped
specimens with the weld parallel or normal to the mis of the tensile specimen are used
for destructive testing. but represent only a limited number of possible forming
conditions [7. 81. For tailored blanks. numerous simulative tests adapted frorn standard
sheet metal formability tests exist to cover almost dl of the possible forming conditions.
The rnost common method is the hemisphencal punch test. in which a binder clamps the
blank at its edges and a punch is forced up through the middle until failure occun (Figure
1.1). The height and force of the punch at failure is recorded and used to compare various
materials and weld configurations. Variations within this son of test include the
placement of the weld line and the shape and size of the blank. binders and punches.
1 annular binder 1
blan k
hemispherical punch
Figure 1.4 - Hemispherical simulative forming test [7]
According to ISO 139 19- 1 (91 and the Auto/Steel Partnership [LOI, the quality
level of a laser weld in steel can be detemined by inspection of a cross-section of the
weld. The determining factors for qudity are: cracks, porosity, penetration, and the
shape of the weld. Undercut, mismatc h. concavity and convexity are important factors for
the shape of the weld (Figure 1.5. 1.6, 1.7). Typically, individual manufacturen also
impose their own additional quality cnteria.
I
z CAUCES UNDER 1 .Umm CAUCES 1 .Omm AND OVER (c 1 .Omm) (= or > 1 .Omm)
(YB) < or = 15% iY/X) c or = 20% (701 < or = 15% (ZBo < or = 2Wo !Y + Z X ) c a r = 15% (Y = Z XI< or = 20% W> or = 85% W> or = 8û%
Any material mismatch m a be added to c o ~ m i t y when dciemining the toul alldwable concavitv (e. K.. IZB(I + m e n t of mismatch < or = 15 pcrccni).
Useta for r w t r u n fat p8nU w i l d u l fiom on. SM.
l Lmirs for unp~ttectionr tor qullity Irv«s
n 5 0 . t ~ t or i mm. whi~nirivu Ir 1110
8rnd.r
h r O.! t or 0.5 mm. ~hichever 11 me amiUet
n 5 O.CS t 31 0.5 mm. wn~cnevrr I S :hc srnoilcf
h dO.2mm ~ 0 . 3 1 w !Ï Inni. whtcnma 4 mm mallu
h s 0.25 t or 3 mm. W~UCIUUCI 1s tne smi i tac
h i; 0.2mt-n + O.3t w 5 mm. W N c l w w a r is
Vir m Y k r
h 5 0 . 2 m m t 0.21 01 5 mm. nnichavsr 11
tha rnuüer
Fipre 1.7 - Section of ISO 139 19-1 describing weld quality[9]
II sO.Zrnm r C.151 or 5 mm. wniChru*t ir tne srnuiof
h 5 O.2mm r O . t t or 5 mm. -vmch~var 8s Ihe srnailsr
Tns iimu ida i i to d i v i r w n r lrom t f u correct pasilion. Unims otnuwiio a p m f i d . rni cocrecr poscimn IS tnrr w h u i ;ho cintr.(inn Coinada.
In addition to destructive test methods, there is a large number of non-destructive
quaiity assurance methods available. in generd. these non-destructive methods fa11 into
three categories: pre-pmcess, pst-process and in-process. Re-process techniques use
Figure 4.24 - Data file for a single cab: 06-26-16-10- 16-S 1-Tl
Investigation of the time intervals between datapoints during welding revealed
occasionai gaps due to the Windows-based operating system. The placement of the gaps
was random in nature and the frequency of the gaps increased when the ce11 was
continuously running in production mode with both spindles operational. Linear
73
interpolation between the data points on either side of the gap was used to mate a file
with constant time intervals (Figure 4.25). The value of the interpolation was rounded to
the nearest multiple of four to be consistent with the original sampled data.
Figure 4-25 - Gnph of original and interpolated data for a single sensor
A time iag was found to exist between the start of signal acquisition and the start
of welding. A second program was therefore written that eliminated starting data points
until a preset percentage of the maximum value of one of the three senson was reached
(Fi gure 1.26).
Figure 4.26 - Graph of original and interpolated data at 5% of maximum signal value
4.5.1 Spectrum Analysis
The data was converted from the time domain to the frequency domain using
Maximum Enuopy Method (MEM). MEM (also known as an dl-poles model) is an
alternative to Fast Fourier Transform m) analysis. The advantages of this system are
that i t can be quicker to mn than FFT and it has the ability to fit sharp spectral peaks [84].
The number of coefficients selected determines the order or number of poles in the
approximation. The number of poles used in an approximation determines the mount of
features ihat can be identified (Figure 4.17). .4 smaller number of poles requires less
analysis time and creates a smoother output specmm. If tw rnany poles are selected than
this method may show phantom peaks when compared to anal ysis.
. l S .- 7
frequency f
Figure 13.7.1. S;implc ourput of maximum cnaopy specrml esumtion. Thc input signai consists of 5 17 sarnpies of the sum of two sinusoids of vcry ncarly the u m c fkquency, plus white noise with about c q u d powcr. Shown IS an rxpandcd parnon of thc full Nyquist frcqucncy interval (which would cxtcnd from zero IO 0.5). The dashed spccml esumatc uses 20 polcs; the doncd. M: the solid. 150 With the I q c r numbcr of ples. the mcthod can rrsolve the dininct sinusoids; but the h t noise background is b e g i ~ i n g to show spunous peaks. (Note logmrhmic scale.)
Fipre 4.27 -Approximations of a sinusoidal function using the MEM with different number of
A Computer program incorporating code from Numerical Recipes in C [84] wiis
wntten to perform the analysis on al1 the sarnple data. In the current research 5 p l e s
were selected as any more led to the creation of phantom peaks.
Chapter 5 Experimental Results
5.1 Introduction
The results of training and testing numerous neural networks to associate t h e
weld pool parmeten to the shape and relative position of the Fusion in gear welding is
presented in this chapter. The analysis of the geometncal properties used to determine
the quality of the weld is also discussrd in this chapter.
5.2 Geometrical Properties of the Fusion Zone
The image analysis software, Imagepro, is an excellent 1001 as it allows for
automatic calculation of the geometrical properties of the fusion zone (Figure 4.31). Four
measurements were performed on the weld cross-sections in order to generate the set of
propenies chosen to describe the fusion zone: presence of holes. area of holes. thickness
of the disk. md a single point at the outer corner of the disk that is used to detemine the
lateral position of the fusion zone. Precision of these measurements was determined by
performing repeated runs on a *mup of nndody selected sarnples. Studeni's t-
distribution was used as it is applicable for small sarnple populations.
5.2.1 Area of the Fusion Zone
The weld m a including al1 holes and excluding the heat affected zone ( H M ) was
measured. The analysis software had the ability to automatically trace an area based on
brightness, contmt and colour differences in the image. Unfortunately this feature was
only useful dong the top and bottom outer edges of the weld and required manual
intervention in order to accurately follow the edge of the actual fusion area. Repeated
rneasurements were performed on selected samples that indicated an error estimate of
2%.
5.2.2 Thickness of the Disk
Thickness measurements were performed by dnwing lines on the top and bottom
surfaces of the disk. The tolennce on the thickness of the disk resulted in the top and
bottom surfaces not always being parallel. The average thickness of the disk over the
length of the two lines was used for analysis and estimated to Vary by +/- 0.07mm. Al1
samples were found to be within the manufacturer's specified thickness of 4.45 +/-
O. 15mm.
5.2.3 Area of the Holes
Unlike the total weld area, it was possible to use the automatic tmcing feature in
the image analysis software to measure the area of the holes. This was possible as the
holes appear darker in the digital images. The automatic hole measurement reduced
possible human enors but. due to the small overall area of most holes. a precision error of
5% is estirnated. A C++ program was written to count the number of holes and to
calculate a ratio of the total area of holes to the totai weld area including holes to reduce
the absolute error.
5.3 Neural Network Architecture
As described in Chapter 3, a multilayer feedforward neural network is considered
for the task of modeling the relation between sensor signds and geometrical properties of
the fusion zone. The number of inputs. outputs and hidden nodes describes the
architecture of such a network. The input and output node counts are detemined by the
mode1 requirements while the number of hidden layen and nodes in each layer are
determined by model performance. The mode1 performance is also affected by the design
of the training sets, data preparation and presentation to the network.
Unfortunately. only genenl guidelines exist for how to design a neural network or
the training patterns. In order to achieve the best possible results. different ways of
presenting the experirnental data to the network and architecture of the neural network
itself had to be exarnined.
5.3.1 Evaluating Performance of the Neural Moâel
The performance of the neud model is evaluated using root mean square (RMS)
emor. Root mean square e m r has a few desirable properties. which have made it the
method of choice for evaluating the performance of neural network models. These
include the ease of calculation. emphasizing the large emn and the ease with which the
denvative of the emor can be computed for optimization purposes [79].
The RMS e m r for an individual output is defined as.
where. t is the target value for the output a is the actual value for the output r is the number of samples
Each neural model is evaiuated twice: once against the training set consisting of
patterns with which the neural network was trained and once with a test set. The test set
consists of data set aside for evaluating the generalization capability of the neural model.
The neural network model has not seen these pattems during training.
In each case, the total error for the network is calculated by averaging the
individual RMS errors over the entire set. i.e.,
where. s is the number of output variables in the network.
Initial tnining was performed by repeatedly presenting the training set to the
network until a minimum training error was reached. Once a minimum training error was
reached the network was presented with a test set. This method creates the risk of
ovemaining the network. however. and indication of the upper bound on the test error
and a lower bound on the training error cm be found. Once a reasonable upper bound on
test emr has been found, the optimal training will be attempted by comparing the test
and training enors as outline in Section 3.2.2.
Based on discussion with ATC and DairnlerChrysler a target of less than 10%
error on the individual output parameters was deemed to be acceptable.
5.3.2 Training using Sequential Data
As previously mentioned, five tabs are welâed on each gear. The initial training
set, therefore. consisted of the data from the fint four of the five tabs from each gear. The
data from the remaining tabs was used for the test set. The training set was presented in
chronologicai order to the neural network. Figure 5.1 shows the total error for the
network when trained with sequential data. The training error reduces as the number of
hidden nodes increases. The test error follows a general trend of increasing with the
addition of hidden nodes. This trend is not entirely consistent as the test error for 12
hidden nodes is srniiller than the networks trained for 10 or 8 nodes. The decreasing total
training error and increasing total test error with the addition of hidden nodes indicate
that the network was overtrained. Even though the network was overtrained the values
can be interpreted as an upper bound on the test error and a lower bound on the training
error. Similar behaviour is seen when the individuai results are presented (Figure 5.2.
Figure 5.3). The area of the fusion zone and the lateral position of the fusion zone
typically have the lowest individual errors whereas the thickness of the disk and the
number of holes present in the fusion zone have the highest individual errors.
Figure 5.1- Sequential training set (total enor)
Number of hldôan no&$
Figure 5.2 - Sequential training set (individual training enor)
=5 O
t8 16 14 12 IO 8 6 4 2 1
Numbef of hidden nodes
4- x-pos
+axis major '
+ a s minor + thickness
I - lholes - hole ratio
Figure 5.3 - Sequential training set (individual test error)
5.3.3 Training by Randomizing the Data
A second attempt at training w;is undertaken by randomizing the order of the data
present in the training set. A test set was fonned by removing fifty patterns frorn the
data. Minor decreases in the total training error and small increases in the test error were
found in cornparison to the sequential training set (Figure 5.4). The difference between
the total test and training error and the increasing total test error with an increase in the
number of hidden nodes indicate that overtraining occurred. The individual training error
of the fusion zone was found to have decreased (Figure 5.5). It is interesting to note that
the error fluctuation as a function of the number of hidden nodes, is reduced as a result of
this randomization (Figure 5.6).
Fipre 5.4- Results for randornized training set (total error)
1 -a- area i aspect
' x-pos f - 1
I -a+ a#s major '
/ +a#s minor i
1 +hid<ness ' l
1 -#holes ~
Figure 5.5 - Results for randomized training set (individual training error)
2 - hole ratio : 0 -
18 16 14 12 10 8 6 4 2 1
Himbsr d h W n -8
a r e a 1 ! t
aspecï i /
x-pos ! ' 1 .+sas major 1 ,
+a#$ minor 1 1
+ ttiickness / I i !
15; 1 -#holes
- hole ratio :
O 18 16 14 12 10 8 6 4 2 1
Numôar al hidckn nackr
Figure 5.6 - Results for randomized uaining set (individual test error)
5.3.4 Input Normalization
Since outputs of sigmoid functions used in the neural network are limited to
values between [O, 11, the outputs have k e n nomalized to the range of [0.1.0.9] as
previousl y discussed in C hapter 3. Theoreticdl y. nonnaiizing the inputs is not required.
but previous experience has shown that such nomalization may improve the performance
of the network [76]. Training was repeated with ali the input and output values
normalized between O. 1 and 0.9.
Improvements were observed in leaming the uaining set while no noticeable
change in the performance of the test set was evident (Figure 5.7). Increasing total test
erron with the addition of hidden nodes and the difference between total test and training
errors indicate that ovenraining has occurred. The error in several of the individual
training panmeten. including the area of the fusion zone and lateral position of the
fusion zone, appear to converge toward a minimum of 4% and 5% respectively when 8 or
more hidden nodes are used (Figure 5.8). Irnpmvements of the test error for the fusion
zone area and its lateral position are achieved, both under 10%. with the exception of 16
hidden nodes (Figure 5.9).
pq . test
Figure 5.7- Fully nortnalized training set (total error)
i t a r a , I
! apect ! - x - r n Il l-rcaps major i / os minarll
/ ! c t t i i c b a i l
;-#i~ler jl
Figure 5.8 - Fully nomdized training set (individual training error)
Figure 5.9 - Fully n o m l i z e d training set (individual test error)
5.3.5 Two Hidden Layer Neural Network
Although one hidden layer is sufficient for a neural network to l e m any input-
output relation. a second hidden layer cm sometimes improve training results.
Unfonunately, when an additional hidden layer is added, there is an increase in the
amount of time required for training results. More than two hidden layers are seldom
used as this does not. in general, improve the performance of a neural network.
Analysis was perfonned using the fully nomalized training set described in
Section 5.3.4 on a small selection of possible two hidden layer architectures. The total
training error of al1 of the two hidden layer networks were comparable to the erron for
the networks containing ten or more nodes on one hidden layer. The best total test error
achieved using two hidden layers, 12%, was comparable to the second worst total error
using the single hidden layer fully normalized network (Figure 5.10). Overtraining was
dso found to be present with two hidden layea as there was a significant difference
between the total test and training emr. Cornparisons of the individual erroa were
87
similar in nature but varied in magnitude (Figure 5.1 1, Figure 5.12). Therefore. adding a
second hidden layer does not appear to improve the performance of the network.
- - I 14 l
A
, 8 12 1
V
s 10
5 8 1. test 1 6
4 ! 2
O i
Figure 5.10- Two hidden layers (total error)
12 t -a- area I
A 10 a s w , 8 - XQOS 1 Y
8 t
j ++ a#s major 1 + a i s minor
S ~ t h i c k n e s s ! a 4 1
I
Numkr of hiddan nodm
Figure 5.1 1 - Two hidden layers (individual training error)
Figure 5.12- Two hidden layen (individual test error)
5.3.6 Elirnination of Samples Containing Porosity
As previously rnentioned, the main objective of this work is to estimate the shape
and relative location of the fusion zone. Inspection of the individual errors revealed that
samples containing porosity were a major contributor to the total error. As rwt mean
square error emphasizes the large erron, this effect would dominate the network output.
In an attempt to further improve the performance of the neural network in estimating the
area of the fusion zone and its lateral position. it was decided to remove samples
containing porosity. In addition, the two parameters associated with porosi ty. number of
holes and porosity ratio. were elirninated from the training and test sets. From this point
fonvard the training sets created by removing porous sarnples will be refened as the PSE
(porous sarnples excluded), and the training set including porous sarnples will be refemd
as the PSI (porous sarnples included). Note that boih of these data sets are normalized as
described in Section 5.3.4.
5.3.7 Improving Generalization by Adding Noise
Due to the large number of parameters, neural networks are very sensitive to
overfitting. i.e.. they will leam specific pattern characteristics in the training set at the
expense of general input-output relations. One way to reduce the ovenitting problem is to
use a larger training set. It is important to address this problem since. by eliminating the
sample containing porosity. the training set has become smaller.
Genention of training pattems is often expensive andor time consurning. as is the
case in this work. Other methods of expanding the training set must, therefore, be
explored. One such method is to generate additional training cases by superimposing
random noise on a measured set. This may improve the ability of the trained neural
network to handle noisy data that will be presented to it later and will also reduce the
likelihood of overfitting [85]. When additional data is generated. careful attention must
be paid to ensure that data created from the same original pattems is not present in both
the training and test sets. Keeping data generated from the same original pattern separate
ensured that a pmper test for generaiization occurred. The precision error in measuring a
parameter was used to determine the amount of noise added to the data.
A ciramatic decrease in the total training error was found with the implementation
of the aforernentioned changes to the PSE. achieving a minimum near 3 1 (Figure 5.13).
The total test error, however, is sirnilar to the PSI. No additional data was added to the
PSI set. Similar to the previous network configurations the difference between the total
training and test enors indicate that overtraining occurred.
Figure 5.13- PSE (total error)
The PSE total tnining error showed significant decreases over the PSI error
(Figure 5.14). The decrease in the total training error with an increase in the number of
hidden ndes was common between the two different networks. However, similar
decreases were not found with the total error in the test sets (Figure 5.15). The total test
error is similar in magnitude regardless of how many hidden nodes were used.
Figure 5.14 - Comparison between PSI and PSE (total training enor)
Fipre 5.15- Comparison beiween PSI and PSE (total test error)
5.3.8 Optimizing ktwork Training
Due to the significant decrease in training error achieved with the PSE data and
to ensure that ovenraining does not occur training was repeated to include evaluation of
the test error at several steps. Thirteen incremenis in training error mging from 15% to
O. 1% were selected to stop training and evaluate the test error for both the PSI and PSE
network. For a given number of hidden nodes the total test and training error cm be seen
to decrease with increased training epochs. For the PSI network the total test and training
error begin to diverge below the training of 0.75% (Figure 5.16). The PSE network
exhibited a similar difference between the total training and test error but at a lower
training error of 0.5% (Figure 5.17).
a train t e s t
Figure 5.16 - PS l (total error 4-û hidden nodes)
35
8g 20
. H test 1 8 15 = 10
5 O
Figure 5.17 - PSE (totai error 40 hidden nodes)
The uaining of the network was considered optimized and capable of
gnenlization at the point before the test and training error began to diverge. When these
training values are compareci it cm still be seen that the PSE training set resulted in lower
total training and test error (Figure 5.18, Figure 5.19). The lowest total test and training
errors for the PSE sarnples were achieved when 70 or 30 hidden nodes were used.
Figure 5.18 - Comprison ktween PSI and PSE (optimized total training error)
1
, a PSI , PSE
Figure 5.19 - Cornparison between PSI and PSE (optimized test e m )
Of the parameten under investigation, the fusion zone area and its lateral position
are the most important when it cornes to quality control. The fusion zone area and its
latenl position are among the parmeten with the lowest individual training and test
erron (Figure 5.20. Figure 5.2 1). The training erron for the fusion zone area for PSE
samples are typically sevenl percent lower than the PSI samples (Figure 5.22). Training
errors of less than 4% were achieved in the former case. The test emors for the fusion
zone area were half the magnitude of the PSI samples for over half the cases (Figure
5.23). Test errors below 4% were achieved for the PSE samples for several different
hidden nodes combinations used dunng training.
O [ r i 200 180 120 70 60 50 40 30 20 10
Number of hidden nodes
1 aspect i l i I j : - x-pos
! / i * axis major, i 1 1 -+ axis minor i I I
/ - thickness i
Figure 520 - PSE (individual training emr)
I
200 180 120 70 60 50 40 30 20 10
Number of hidden nodes
+ area aspect x- pos I
* axîs major l
-e- axis minor + thickness : i
Figure 521 - PSE (individual test error)
Figure 5.22 - Comparison between PSI and PSE (area training error)
n 12
8 10 - 8 I PSI
L 6 U)
.PSE ' a
2
O 200 180 120 70 60 50 40 30 20 10
Number of hMâen nodes
Figure 5.23 - Comparison between PSI and PSE (area test error)
The training error in the Iateral position of the fusion zone for PSE showed a
small decrease in magnitude over PSI for most variations in the number of hidden nodes
(Figure 5.24). The test e m r on the lateral position of the fusion zone was lower in al1 but
two of the cases. On several occasions the magnitude of the PSE data was almost half the
PSI data (Figure 5.25). Significant improvements to the test error and minor
improvements in the training error were observed when training with PSE data cornpared
to PSI data. Individual training and test errors below 5% for the area of the fusion zone
and its lateral position were achieved for PSE data when 70 hidden nodes were used.
Figure 5.24 - Comparison between PSI and PSE (laterd position training error)
.7
op 12 v L 10 I
8 P S I i
O I
u, 6 P S E ;
B 4 I
2 I
O 1
200 180 120 70 60 50 40 30 20 10
Numbet of hiâden nodes
Figure S.= - Comparison between PSI and PSE (lateral position test error)
The neural network model has been shown to be capable of estimating
geometncai properties of the fusion zone frorn laser-welding sensor data. To ensure the
success of the neural model. careful attention was paid to the manner in which the
training data is presented to the network. Ali the input and output data used to generate
the training set were normalized to an identical range to facilitate training. Adding a
second hidden layer to a network did not improve the performance and, therefore, was
not further pursued. It was dso found that reducing the dernand on the neurd network by
eliminating hard to predict outputs irnproved training. The most significant reduction in
error was achieved by generating supplementary training cases through the addition of
random noise to the measured data set.
Chapter 6 Conclusions
6.1 Contributions
A weld monitoring system using neural networks has k e n developed that can
estimste the cross sectionai area and lateral location of the fusion zone in a laser welded
transmission gear. The estimates are based on measurements from ultra-violet. visible
and infrared photodiodes. It has been shown that through proper signal processing,
judicious choice of architecture and application of appropriate techniques, a neural
network c m be trained to correlate photodiode signais to fusion zone properties with
reasonable accuracy.
6.2 Concluding Remarks
The work described in this thesis was aimed at developing a system for estimating
the shape and location of the fusion zone in a laser gear welding application. In general, a
good weld cm be characterited as having a fusion zone that is as deep as the material is
thick. The fusion zone should have a consistent width from the top to bottom surface of
the joint. Maintaining a consistent depth and width of the fusion zone in a weld results in
a better joint that is able to evenly distribute any loading throughout the entire weld
region.
Existing quality standards recornrnend the practice of destructively testing a
random selection of a representative sample of parts. The best method to determine the
shape and size of a laser weld is to cut a section through the joint and perfonn a
metallographic inspection on the fusion zone.
Due to the Iength of time required to cut. polish and mesure the fusion zone in a
welded transmission gear component. many bad parts may be produced before a problem
is noticed. From an economical standpoint, it is impractical to destruciively inspect al1
parts produced. Therefore, non-destructive inspection techniques must be used, in
addition to periodical destructive tests, in order to determine the quality of the paris
produced.
Neural networks were investigated as a tool for predicting charactenstics of the
fusion zone in a laser weld as part of a quality control system. Using information
gathered from previously welded and inspected components, a training set was created
and used to train a neural network. The main advantage of using neural networks is that
the network iiself determines relationships between the input and output signals. In laser
welding, the relationship between photodiode signals of the rnolten weld pool and the
size a d shape of the fusion zone are not well known due to complex physical
interactions present during the process.
Different methods of generating the training sets and different neural networks
intemal architecture were investigated. 1t was found that improvements in the error of
the training and test sets could be achieved by randornizing the order in which data was
presented to the system and by using fully nomalized data. The improvements were
most noticeable in the test error for individual parameters. Presentation of the data in a
random order removes the possibility of learning any trends that may exists between
groups of welds. This is especially imponant as the data for the training set is generated
by incrementally varying the welding parameters that the network may interpret as king
a requirement for a good weld. Normalizing the input data between the sarne upper and
lower values as the output data ensures that al1 the data is treated equally by the neural
network.
Minor improvements in the training set error were found when a second hidden
layer was added to the neural network. However, the test set error increased and the time
required for training dramatically increased with the addition of a second hidden layer.
This confirmed that a second hidden layer is very rarely required in pattern recognition
networks.
The training set error was typically more than halved when the network was
trained with data from welds that contained no porosity. The mount of data in the
training set was dso increased by adding a perceniage of random noise, based on the
precision error of the measured parameters, to the data used in the previous training sets.
The individual parmeters that were best predicted by the neural network were the
fusion area and the laterai position (x-pos) of the fusion zone. The depth of the fusion
zone (major mis) was also well leamed by the network. When only full penetration
welds are under investigation, the area and the lateral position of the fusion zone are
quality measures that c m be used to determine the quality of a weld.
The work presented has shown that the cross sectional area and the lateral
position of the fusion zone in a laser weld can be predicted with reasonable certainty.
However, other parameten that were investigated such as the number and area of the
holes could not be predicted.
Future work should concentrate on improving the accuracy of the prediction of
the weld ares and its geomeuy, as these mesures can be used as a quality mesure. This
could be accomplished by increasing the number of variations within the training set and
the number of samples used to generate the training set. Methods of increasing the size of
the training set through artificial means. such as adding random noise, should be funher
investigated as it is much easier to genente more samples with a cornputer ihan getting
materiül and tirne to be used on a production machine.
The addition of a dynamic neunl network after the static neural network may also
be able to improve the accuracy of the network. while reducing the arnount of training
required. It may be possible for such a network to predict accurateiy outside of the
training set used for training the static network. The largest benefit of such a system
would be that a network would not have to be retrained after adjustments are made to the
processing parmeten.
Unfonunaiely, the system presented did not work very well when attempting to
identify the area and number of holes present within a sampie. This is most likely due to
the relatively small size of the holes in cornparison to the totd weld area. Instead of
trying to teach a neural network to identify the characteristics of the holes in a given
simple, it should merely be asked to try and determine the presence of pomsity.
Finally. the system must be able to function properly in a production environment
with minimum intervention. In this case, the addition of a second dynamic neural
network to improve the operating range of the system would prove invaluable as it would
reduce the need for retraining after any adjustrnents in the system.
References
[ l ] Industry Canada. Automotive and transponation branch. Statistical review of the
Canadian automotive industrv: 1999 Edition. Ottawa, ON. 1999.
[2] The new lexicon Webster's encvclowdic dictionam of the enalish laneuage -
Canadian eciition. New York:Lexicon Publications, 1988.
(31 Porsche Engineering Services, Inc. Cntrali - eht steel auto closures eneineenne rewn.
CD-ROM. April2000.
[4j R.E. Mueller. "Laser weldmg of hem flange joints." Proceedinas of ICALEO 2000.
91 (2000):Fll-F21.
[5] Auto/Steel Partnenhip. Tailor blank desim and manufactunng manual volume Ii,
business and economic considerations. Southfield, MI, 1996.
[6] Lanson. J.K.. L. Hanicke. "Multi-materiai approach with integrated joining
technologies in the new Volvo S80." SAE. Paper No. 1999-0 1-3 147. 1999.
[7] Auger, M., Abdullah. K.. Jeswiet, J., Wild P. and L. Clapham. "Determination of
weld line characteristics in tailored blanks." SAE. Paper No. 2000-0 1-266 1. 2000.
[SI Sanden. F.I. and R.H. Wagoner. "Forming of tailor-welded blanks." Metallurgical
and Materials Transactions A. 27A( l9%):2605-26 16.
[9] International Standard. ISO 13919-1. Welding - electron and laser-beam welded
joints - miiâance - on aualitv levels for irn~rfections - art 1:steel. Switzerland.
Reciws in C - The Art of Scientific Com~uting, Me1boume:Cambridge University
Press, 1992.
[85] Reed. R., Oh, S. and R.J. Marks. "Regularization using jittered training data."
International Joint Conference on Neurd Networks. Baltimore, MD 1992.
Appendix A
Portion of sample output file 06-26-16-10-16.Sl.txt
UV 116 116
Time (s) UV 76 16 20 - O
IR Time (s) ' 0.05 0.051
IR 40 48
VIS VIS 4a 56
Spindle 1 standard
Appendix B
setting and tabk of test variations
Spindle 1 standard setting I
Laser Power 1 94% of 8kW I Pulse freauencv 1 20 OOOHz I
06-26- 1 3-27-29-SI standard settings 06-26-1 3-28-32-SI standard settings 06-26- 1 3-33-32-SI x-axis towards tab 1.264 06-26- 1 3-36-08-SI x-axis 1.284 ,
û6-26-13-38-21 -SI x-axis away from tab 1 224 û6-26- 13-41 -1 O-SI x-axis 1.204 06-26-1 3-45-1 6-SI z-ais up to 9.27 and x-axis 1.204 06-26- 13-464-SI z-axis 9.27 1 mm up and x-axis to 1.244 , 06-26- 1 3-48-47-SI z-axis 9.23 2mm up
106-26-1 4-50-34-SI [z-axis 9.27 1 mm up and speeâ 65 1
Variations .bad noule location bad nonle location standard settings standard settings standard settings standard settings standard settings standard settings standard settings standard settings standard settings