Professional Documents
Culture Documents
Artificial Neural Networks For Quality Control by Ultrasonic PDF
Artificial Neural Networks For Quality Control by Ultrasonic PDF
Abstract
An artificial neural network is proposed to solve problems in the interpretation of ultrasonic oscillograms obtained by the pulse echo method.
The artificial neural network classifies resistance spot welds in several quality levels through their respective ultrasonic oscillograms. The inputs
of the artificial neural network are vectors obtained from each ultrasonic oscillogram with the help of a MATLAB® program. The training of
the artificial neural network uses supervised learning mechanism and therefore each input has the respective desired output (target). There are
four targets, one for each considered quality level. The available data set is randomly split into a training subset (to update weight values) and a
validation subset (to guard against overfitting by means of cross validation). The number of neurons in the hidden layers is selected considering the
overfitting phenomenon. This research work has the aim of contributing to the automation of quality control processes in resistance spot welding.
© 2006 Elsevier B.V. All rights reserved.
Keywords: Resistance spot welding; Non-destructive testing; Ultrasonic oscillograms; Quality control; Artificial neural networks
The chemical composition and the mechanical properties of the sheet steel
are, respectively, shown in Tables 1 and 2. The sheet thickness is 1 mm.
∗ Corresponding author. Tel.: +34 983423389; fax: +34 983185951. The considered parameters in the RSW process (single-phase AC 50 Hz
E-mail address: oml@eis.uva.es (Ó. Martı́n). equipment) are:
0924-0136/$ – see front matter © 2006 Elsevier B.V. All rights reserved.
doi:10.1016/j.jmatprotec.2006.10.011
Ó. Martı́n et al. / Journal of Materials Processing Technology 183 (2007) 226–233 227
Table 1 too few to be representative. More than six echoes are representative but they
Chemical composition of the sheet steel (wt.%) do not give significant information with regard to the information that is associ-
ated with six echoes. Therefore, the compromise solution considers the first six
C 0.05
echoes in the ultrasonic oscillogram and the input of the ANN is a 10-component
Mn 0.26
vector where (Fig. 1):
Si 0.02
P 0.012
S 0.011 • The first five components are the relative heights of the echoes: the nth com-
Al 0.033 ponent (n = 1, . . ., 5) is the height of the (n + 1)th echo (hn + 1 ) divided by the
height of the 1st echo (h1 ).
• The last five components are the distances between consecutive echoes: the
nth component (n = 6, . . ., 10) is the distance between the (n − 4)th echo and
Table 2
the (n − 5)th echo (dn − 5 ).
Mechanical properties of the sheet steel
Yield strength (MPa) 192
2.3. Classification of spot welds and characterization by
Tensile strength (MPa) 301
ultrasonic testing
Total elongation (%) 40
Hardness (HV) 104
The quality level of a spot weld is estimated from its respective ultrasonic
oscillogram. The estimation is based on the effect of the weld nugget of the spot
weld on the ultrasonic wave. Two parameters of the weld nugget are considered
• Current intensity varies between 4 and 8 kA RMS [17,18]. The recorded [5]:
values are gathered round five theoretical values (4, 5, 6, 7 and 8 kA RMS).
• Weld time varies between 4 and 20 cycles [17,18]. The recorded values are
• Weld nugget microstructure. A weld nugget is a cast microstructure with
gathered round nine theoretical values (4, 6, 8, 10, 12, 14, 16, 18 and 20
coarser grains than the parent metal, therefore a weld nugget produces
cycles).
higher attenuation than the parent metal. The heat-affected zone (HAZ) is
• Electrode sort depends on the copper alloy sort and on the thermal/mechanical
a small zone between the weld nugget and the parent metal. The HAZ is a
treatment used for the alloy. Two copper alloy sorts (Cu–Cr and Cu–Be [18])
recrystallized microstructure with fine and equiaxed grains that produce low
and three thermal/mechanical treatments [19,20] are selected, therefore there
attenuation. The influence of the HAZ on the respective ultrasonic oscillo-
are six electrode sorts.
gram can be neglected.
• Electrode force is fixed at 980.7 N.
• Weld nugget size. One-layer echoes appear between principal echoes if the
weld nugget diameter is smaller than the width of the ultrasonic beam so that
The ultrasonic spot welding testing transducer uses a captive water column
a part of reflections occurs at the interface between the two sheets. A thick
delay and a replaceable rubber membrane in order to provide good coupling
weld nugget produces higher attenuation than a thin weld nugget.
to the weld surfaces. The transducer frequency is 20 MHz and the transducer
diameter is 4.5 mm.
The spot welds are classified into four possible quality levels [5,21] and a
target (a two-component vector) is assigned to each one [22]:
2.2. Inputs of the artificial neural network
• Good weld: (1 1).
The binary signal associated with each ultrasonic oscillogram is transformed • Undersize weld: (0 1).
into the input of the ANN by means of a MATLAB® program [4]. The input • Stick weld: (1 0).
of the ANN must be representative of its respective ultrasonic oscillogram [15], • No weld: (0 0).
therefore the transformation program must consider the main factors that char-
acterize an ultrasonic oscillogram [5]: the attenuation of the ultrasonic wave and
the one-layer echoes (caused by ultrasonic beam reflections that occur at the 2.3.1. Good weld
interface between the two sheets) between principal echoes. On the other hand, 2.3.1.1. Weld nugget. The cast microstructure of the weld nugget has coarse
the input must not be too complex in order to avoid complicating the ANN and grains. The weld nugget thickness is enough to produce high attenuation.
using redundant data. The decision that determines the input size is how many The weld nugget diameter is larger than the width of the ultrasonic beam
echoes in the ultrasonic oscillogram are considered. Fewer than six echoes are (Figs. 2 and 3).
Fig. 1. Input of the ANN (right) obtained from its respective ultrasonic oscillogram (left).
228 Ó. Martı́n et al. / Journal of Materials Processing Technology 183 (2007) 226–233
Fig. 4. Ultrasonic oscillogram of a good weld (left) and its respective input of the ANN (right).
Ó. Martı́n et al. / Journal of Materials Processing Technology 183 (2007) 226–233 229
Fig. 7. Ultrasonic oscillogram of an undersize weld (left) and its respective input of the ANN (right).
ness of one sheet because of the discontinuity between the two sheets
(Fig. 13).
a- 1 = -i, a- M = o- (2)
Sk is the number of neurons in layer k, wk+1 (l, j) the synaptic weights, bk + 1 (l)
the biases, fk + 1 the transfer function of layer k + 1, i and o are, respectively, the
inputs and the experimental outputs of the ANN.
The Levenberg–Marquardt algorithm (called trainlm in MATLAB® ) is used
for training the ANN [16,23,24]. This algorithm proceeds as follows [25–27]:
(I) Present all inputs iq to the ANN and compute the respective outputs oq
and errors eq using the respective targets tq :
Fig. 9. Macrograph of a cross-sectioned stick weld.
e- q = -t q − o- q , q = 1, . . . , Q (3)
Fig. 10. Ultrasonic oscillogram of a stick weld (left) and its respective input of the ANN (right).
230 Ó. Martı́n et al. / Journal of Materials Processing Technology 183 (2007) 226–233
Fig. 11. Micrograph of a cross-sectioned no weld. Fig. 14. Load vs. displacement curves obtained from tensile-shear testing. (a)
Good weld, (b) undersize weld, (c) stick weld, and (d) no weld.
−1 T
x- = (J T (x- )J(x- ) + μI) J (x- )e- (x- ) (6)
(V) Recompute V(x) using x- + x- . If this new V(x) is smaller than that
computed in step I, then μ is divided by a factor β, let x- = x- + x-
and go back to step I. If V(x) is not reduced, then μ is multiplied by
β and go back to step IV (μ = 10−3 as a starting point and β = 10 are
used).
(VI) The algorithm converges when V(x) has been reduced to some error goal
Fig. 12. Macrograph of a cross-sectioned no weld. or when the norm of the gradient ∇V (x- ) is less than some predetermined
value. It can be shown that:
(II) Compute the sum of squares of errors V(x), where x are the synaptic weights
∇V (x- ) = J T (x- )e- (x- ) (7)
and the biases:
N
2.4.2. The artificial neural network architecture
V (x- ) = e2r (x- ), N = Q × SM (4)
The number of executed spot welds is 438. An ultrasonic oscillogram is
r=1
obtained from each spot weld.
(III) Compute the Jacobian matrix J(x): A human expert interprets the 438 oscillograms and assigns a target to each
one. The human expert’s efficiency, validated by tensile-shear testing [28–30]
⎛ ∂e1 (x) ∂e1 (x- ) ∂e1 (x- ) ⎞ on a sample of ten spot welds per quality level, is:
- ...
⎜ ∂x1 ∂x2 ∂xn ⎟
⎜ ∂e2 (x- ) ∂e2 (x- ) ∂e2 (x- ) ⎟ • Good weld: 100% (Fig. 14, curve (a)).
⎜ ... ⎟
⎜ ∂x1 ∂x2 ∂xn ⎟ • Undersize weld: 100% (Fig. 14, curve (b)).
J(x- ) = ⎜ ⎟ (5)
⎜ .. .. .. .. ⎟ • Stick weld: 80%. Sometimes it is very difficult to distinguish a good weld
⎜ . . . . ⎟
⎝ ⎠ from a stick weld. In fact, the two errors are two good welds (Fig. 14, curve
∂eN (x- ) ∂eN (x- ) ∂eN (x- ) (c)).
...
∂x1 ∂x2 ∂xn • No weld: 100% (Fig. 14, curve (d)).
Fig. 13. Ultrasonic oscillogram of a no weld (left) and its respective input of the ANN (right).
Ó. Martı́n et al. / Journal of Materials Processing Technology 183 (2007) 226–233 231
Fig. 16. Validity intervals for the components of experimental output vectors.
3. Results
Table 3
Determination of the number of neurons in the hidden layers
ANN Training MSE Validation MSE
Ó. Martı́n et al. / Journal of Materials Processing Technology 183 (2007) 226–233
o 1.0007 1 1.0001 1.0001 3.07E − 05 0.99904 1.0005 0.99994 1.0002 −6.89E − 05 1.0004 0.051161 1.0004 1.0003 0.00012059 6.41E − 05 3.72E − 05 −7.71E − 05 −0.0018457 1.0003
0.00028102 0.00042181 1.0002 −0.00065469 1.0005 0.00026482 0.00027534 1.0002 1.0013 0.99969 1.0001 0.99862 1.0001 1.0001 1.0016 1.0007 1.0005 0.99977 0.078448 1.0001
t 1 1 1 1 0 1 1 1 1 0 1 0 1 1 0 0 0 0 0 1
0 0 1 0 1 0 0 1 1 1 1 1 1 1 1 1 1 1 0 1
√ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √
/× (δ = 0.25)
0.52381 0.33333 0.33333 0.69231 0.57143 0.42857 0.2381 0.33333 0.38095 0.33333 0.66667 0.2381 3.3333 1 0.88889 0.33333 0.083333 0.2381 0.28571 0.7619 0.42857
0.28571 0.28571 0.28571 0.30769 0.52381 0.33333 0.095238 0.19048 0.38095 0.2381 0.047619 0.14286 0.33333 0.80952 0.66667 0.33333 0.58333 0.2381 0.2381 0.095238 0.33333
0 0.095238 0.14286 0 0.33333 0.19048 0 0.2381 0.28571 0.095238 0.42857 0.095238 1.5 0.61905 0.55556 0.19048 0.083333 0.095238 0.19048 0.42857 0.095238
0 0 0.095238 0 0.14286 0.095238 0 0.14286 0.19048 0.095238 0.047619 0 0.33333 0.38095 0.33333 0 0.33333 0.047619 0.095238 0.047619 0
0 0 0 0 0.047619 0 0 0.047619 0.047619 0.047619 0.28571 0 1.1667 0.33333 0.22222 0 0.083333 0 0 0.38095 0
18 15 18 16 18 19 20 20 19 22 18 20 10 11 20 22 8 20 20 22 16
19 15 17 19 18 19 18 18 18 19 9 17 9 10 20 20 12 21 19 10 17
0 14 18 0 18 19 0 19 17 20 11 20 11 9 20 20 9 20 21 10 18
0 0 18 0 18 19 0 19 19 20 10 0 10 10 19 0 12 20 21 10 0
0 0 0 0 18 0 0 19 17 21 10 0 10 9 20 0 11 0 0 11 0
1.0001 0.99931 0.99986 1 0.99991 0.99989 1.0001 0.99996 1.0003 1 6.22E − 05 1.0002 6.41E − 05 −0.00045997 1.0002 1.0001 2.51E − 05 1.0001 0.99998 5.79E − 05 1.0003
1 1.0001 0.99975 1 1.0006 0.99986 1 1.0006 1.0003 1.0007 1 1.0001 1.0007 0.020984 0.00066878 1.0001 1.0001 1.0006 1 1.0011 1.0001
1 1 1 1 1 1 1 1 1 1 0 1 0 0 1 1 0 1 1 0 1
1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 1 1 1 1 1 1
√ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √
0.52381 0.38095 0.6875 0.58333 0.7 0.2381 0.6 0.69231 0.095238 0.57143 0.47619 0.047619 0.71429 0.58824 0.66667 0.47619 0.73333 0.66667 0.8125 0.10526
0.42857 0.28571 0.4375 0.33333 0.5 0.19048 0.4 0.38462 0.61905 0.38095 0.14286 0.47619 0.42857 0.29412 0.5 0.38095 0.5 0.53333 0.38889 0.5 0.68421
0.38095 0.047619 0.25 0.16667 0.1 0.095238 0.2 0.15385 0.095238 0.095238 0.38095 0.095238 0.21429 0.11765 0.41667 0.28571 0.27778 0.4 0.16667 0.25 0.10526
0.2381 0 0.125 0 0.1 0.095238 0.13333 0 0.33333 0.095238 0.047619 0.33333 0.071429 0 0.25 0.19048 0.16667 0.2 0.11111 0.0625 0.47368
0.14286 0 0 0 0 0 0 0 0.047619 0 0.19048 0.047619 0 0 0.083333 0.095238 0.055556 0.13333 0.055556 0 0.15789
17 17 16 16 17 17 16 17 11 21 17 11 18 18 9 16 19 10 19 20 10
17 17 17 16 17 15 17 17 10 18 10 10 16 17 9 17 18 9 19 19 10
16 16 16 15 16 16 17 17 10 17 9 10 18 18 10 17 20 9 20 20 10
17 0 17 0 18 16 18 0 9 19 11 10 18 0 9 16 18 10 20 19 10
17 0 0 0 0 0 0 0 12 0 9 9 0 0 9 17 18 9 20 0 11
1.0039 0.99936 0.99988 1.0002 0.99984 1.0001 0.99987 1.0003 −0.00010755 0.99984 4.22E − 05 −0.00010774 0.99986 1.0003 −0.0012505 1.0006 0.99986 −0.0013334 1 0.99984 −7.53E − 05
−0.0013388 1.0001 1.0007 1.0001 1.0005 0.84737 0.99988 1.0001 1.0002 1.0005 0.98555 1.0002 0.99988 1.0001 −0.0013126 0.77748 0.79971 −0.0015875 0.99992 1.0003 1.0001
1 1 1 1 1 1 1 1 0 1 0 0 1 1 0 1 1 0 1 1 0
0 1 1 1 1 1 1 1 1 1 1 1 1 1 0 1 1 0 1 1 1
√ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √
Ó. Martı́n et al. / Journal of Materials Processing Technology 183 (2007) 226–233 233
4. Conclusions [12] A. Roy, P. Barat, S.K. De, Material classification through neural networks,
Ultrasonics 33 (1995) 175–180.
[13] R. Wendel, J. Dual, Application of neural networks to quantitative non-
The object of this work is achieved because the proposed
destructive evaluation, Ultrasonics 34 (1996) 461–465.
neural model has shown its effectiveness as a tool to carry out [14] R. Jordan, F. Feeney, N. Nesbitt, J.A. Evertsen, Classification of wood
the automation of the classification of resistance spot welds in species by neural network analysis of ultrasonic signals, Ultrasonics 36
quality levels through their respective oscillograms obtained by (1998) 219–222.
ultrasonic testing: [15] G. Barrera, M.A. Fabián, C.A. Ugalde, Detección, identificación y clasi-
ficación de defectos usando RNA y un manipulador robótico de 2 G.L.
(algoritmos Kohonen y MLP), Rev. Metal. Madrid 38 (2002) 163–172.
• The proposed ANN produces good results in the classification [16] J. McBride, S. Malinov, W. Sha, Modelling tensile properties of gamma-
of input vectors non-used in the training. The ANN proves its based titanium aluminides using artificial neural network, Mater. Sci. Eng.
adaptability, its robustness and therefore its ability to gener- A 384 (2004) 129–137.
alize. [17] Y. Cho, S. Rhee, Experimental study of nugget formation in resistance spot
welding, Weld. J. 82 (8) (2003) 195s–201s.
• The selection of the significant characteristics of the phe-
[18] J.M. Gerken, R.S. Brown, D.A. DeAntonio, D.W. Dickinson, R.H. Foxall,
nomenon in question in order to make the inputs of the ANN is I.A. Oehler, L.M. Schetky, M.J. Shields, T.A. Siewert, L.W. Weller, Resis-
very important. The way of representing the ultrasonic oscil- tance spot welding, in: K. Mills, J.R. Davis, B.R. Sanders (Eds.), Welding,
lograms by means of 10-component vectors, that contain the Brazing and Soldering, Metals Handbook, vol. 6, 9th ed., American Society
relative heights of the echoes and the distance between con- for Metals, Metals Park, OH, 1983, pp. 469–493.
[19] G. Joseph, in: K.J.A. Kundig (Ed.), Copper: Its Trade, Manufacture, Use
secutive echoes, is appropriate.
and Environmental Status, ASM International, Materials Park, OH, 1999,
pp. 277–280.
References [20] J.C. Harkness, A. Guha, Beryllium–copper and beryllium–nickel alloys,
in: K. Mills, J.R. Davis, J.D. Destefani, D.A. Dieterich, G.M. Crankovic,
[1] M. Jou, Real time monitoring weld quality of resistance spot welding for H.J. Frissell (Eds.), Metallography and Microstructures, Metals Handbook,
the fabrication of sheet metal assemblies, J. Mater. Process. Technol. 132 vol. 9, 9th ed., American Society for Metals, Metals Park, OH, 1985, pp.
(2003) 102–113. 392–398.
[2] S. Agashe, H. Zhang, Selection of schedules based on heat balance in [21] J. Krautkrämer, H. Krautkrämer, Ultrasonic Testing of Materials, 4th ed.,
resistance spot welding, Weld. J. 82 (7) (2003) 179s–183s. Springer-Verlag, Berlin, 1990.
[3] E. Bayraktar, D. Kaplan, M. Grumbach, Application of impact tensile test- [22] I.E. Dror, M. Zagaeski, C.F. Moss, Three-dimensional target recognition
ing to spot welded sheets, J. Mater. Process. Technol. 153–154 (2004) via sonar: a neural network model, Neural Netw. 8 (1995) 149–160.
80–86. [23] N. Selvakumar, P. Radha, R. Narayanasamy, M.J. Davidson, Prediction of
[4] O. Martı́n, PhD Thesis, Ciencia de los Materiales e Ingenierı́a Metalúrgica, deformation characteristics of sintered aluminium preforms using neural
Universidad de Valladolid, 2004. networks, Model. Simul. Mater. Sci. Eng. 12 (2004) 611–620.
[5] T. Mansour, Ultrasonic testing of spot welds in thin gage steel, in: P. [24] K.K. Tho, S. Swaddiwudhipong, Z.S. Liu, J. Hua, Artificial neural network
McIntire (Ed.), Non-destructive Testing Handbook, v. 7 Ultrasonic Test- model for material characterization by indentation, Model. Simul. Mater.
ing, 2nd ed., American Society for Non-destructive Testing, USA, 1991, Sci. Eng. 12 (2004) 1055–1062.
pp. 557–568. [25] M.T. Hagan, M.B. Menhaj, Training feedforward networks with the Mar-
[6] D.J. Spinella, J.R. Brockenbrough, J.M. Fridy, Trends in aluminum resis- quardt algorithm, IEEE Trans. Neural Netw. 5 (1994) 989–993.
tance spot welding for the auto industry, Weld. J. 84 (1) (2005) 34–40. [26] C.M. Bishop, Neural Networks for Pattern Recognition, Oxford University
[7] G. Barrera, M.A. Fabián, M. Vélez, L. Villaseñor, Inteligencias artificiales Press, New York, 1995.
y ensayos ultrasónicos para la detección de defectos, Rev. Metal. Madrid [27] D.P. Bertsekas, Non-linear Programming, 2nd ed., Athena Scientific, Bel-
37 (2001) 403–411. mont, MA, 1999.
[8] K. Hornik, M. Stinchcombe, H. White, Multilayer feedforward networks [28] Specimen dimensions and procedure for shear testing resistance spot, seam
are universal approximators, Neural Netw. 2 (1989) 359–366. and embossed projection welds, ISO 14273, 2000.
[9] N.S. Reddy, A.K.P. Rao, M. Chakraborty, B.S. Murty, Prediction of grain [29] M. Zhou, S.J. Hu, H. Zhang, Critical specimen sizes for tensile-shear testing
size of Al-7Si Alloy by neural networks, Mater. Sci. Eng. A 391 (2005) of steel sheets, Weld. J. 78 (9) (1999) 305s–313s.
131–140. [30] M. Vural, A. Akkus, On the resistance spot weldability of galvanized
[10] A. Jiahe, X. Jiang, G. Huiju, H. Yaohe, X. Xishan, Artificial neural network interstitial free steel sheets with austenitic stainless steel sheets, J. Mater.
prediction of the microstructure of 60Si2MnA rod based on its controlled Process. Technol. 153–154 (2004) 1–6.
rolling and cooling process parameters, Mater. Sci. Eng. A 344 (2003) [31] S. Haykin, Neural Networks: A Comprehensive Foundation, 2nd ed.,
318–322. Prentice-Hall, Upper Saddle River, NJ, 1999.
[11] R.H. Silverman, A.S. Noetzel, Image processing and pattern recognition [32] S. Guessasma, C. Coddet, Microstructure of APS alumina–titania coatings
in ultrasonograms by backpropagation, Neural Netw. 3 (1990) 593–603. analysed using artificial neural network, Acta Mater. 52 (2004) 5157–5164.