Journal of Materials Processing Technology 183 (2007) 226–233

Artificial neural networks for quality control by ultrasonic

testing in resistance spot welding
Óscar Martı́n ∗ , Manuel López, Fernando Martı́n
Ingenierı́a de Materiales, Universidad de Valladolid, Escuela Técnica Superior de Ingenieros Industriales,
Paseo del Cauce s/n, Valladolid 47011, Spain
Received 8 August 2006; received in revised form 10 October 2006; accepted 11 October 2006

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

1. Introduction systems, in which the regression methods have their limitations

Resistance spot welding (RSW) is extensively used for join-
ing sheet steel in the automotive industry [1,2]. The trend to
reduce the high number of spot welds per vehicle (3000–4000
[3]) imposes the optimization and fine-tuning of reliable quality
control systems [4].
The pulse echo method is an ultrasonic non-destructive test-
ing technique suitable for the quality control in RSW [5,6]. This
method obtains from each spot weld an ultrasonic oscillogram
that allows estimating the quality level of the aforementioned
spot weld. Sometimes the ultrasonic oscillogram is hard to inter-
pret by a human expert or the task of interpreting repeatedly
oscillograms for a long time gives rise to the drop of the human
expert’s efficiency [7]. Therefore, the automation of the inter-
pretation of ultrasonic oscillograms would improve the quality
control performance in RSW. Artificial neural networks (ANN)
are mathematical models that imitate the behaviour of the human
nervous system and hence have a parallel, distributed and adap-
tive processing capable of mapping non-linear and complex
[8–10]. For this reason, the ANN are extensively used in pattern
recognition tasks [11–15]. The interpretation on each ultrasonic
oscillogram, in order to classify the respective spot weld in a cer-
tain quality level, is a pattern recognition problem, so an ANN
is proposed to carry out the automation of the interpretation of
ultrasonic oscillograms.
An ANN, just like a human being, learns by means of training.
A supervised learning mechanism is used in the training of the
ANN in which a set of input/target pairs is utilized (a target is
the desired output respective to a certain input). In the training,
the synaptic weights (each link between neurons has a synaptic
weight attached to it) are repeatedly adjusted to reduce the error
between the experimental outputs and the respective targets until
a certain value of error is achieved [16].
2. Experimental procedure
2.1. Materials and equipment

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
Chemical composition of the sheet steel (wt.%)
C 0.05
Mn 0.26
Si 0.02
P 0.012
S 0.011
Al 0.033
Table 2
Mechanical properties of the sheet steel
Yield strength (MPa) 192
Tensile strength (MPa) 301
Total elongation (%) 40
Hardness (HV) 104
• Current intensity varies between 4 and 8 kA RMS [17,18]. The recorded
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
gathered round nine theoretical values (4, 6, 8, 10, 12, 14, 16, 18 and 20
cycles).
• Electrode sort depends on the copper alloy sort and on the thermal/mechanical
treatment used for the alloy. Two copper alloy sorts (Cu–Cr and Cu–Be [18])
and three thermal/mechanical treatments [19,20] are selected, therefore there
are six electrode sorts.
• Electrode force is fixed at 980.7 N.
The ultrasonic spot welding testing transducer uses a captive water column
delay and a replaceable rubber membrane in order to provide good coupling
to the weld surfaces. The transducer frequency is 20 MHz and the transducer
diameter is 4.5 mm.
2.2. Inputs of the artificial neural network
The binary signal associated with each ultrasonic oscillogram is transformed
into the input of the ANN by means of a MATLAB® program [4]. The input
of the ANN must be representative of its respective ultrasonic oscillogram [15],
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
interface between the two sheets) between principal echoes. On the other hand,
the input must not be too complex in order to avoid complicating the ANN and
using redundant data. The decision that determines the input size is how many
echoes in the ultrasonic oscillogram are considered. Fewer than six echoes are
too few to be representative. More than six echoes are representative but they
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
echoes in the ultrasonic oscillogram and the input of the ANN is a 10-component
vector where (Fig. 1):
• The first five components are the relative heights of the echoes: the nth com-
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
the (n − 5)th echo (dn − 5 ).
2.3. Classification of spot welds and characterization by
ultrasonic testing
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
[5]:
• Weld nugget microstructure. A weld nugget is a cast microstructure with
coarser grains than the parent metal, therefore a weld nugget produces
higher attenuation than the parent metal. The heat-affected zone (HAZ) is
a small zone between the weld nugget and the parent metal. The HAZ is a
recrystallized microstructure with fine and equiaxed grains that produce low
attenuation. The influence of the HAZ on the respective ultrasonic oscillo-
gram can be neglected.
• 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
a part of reflections occurs at the interface between the two sheets. A thick
weld nugget produces higher attenuation than a thin weld nugget.
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]:
• Good weld: (1 1).
• Undersize weld: (0 1).
• Stick weld: (1 0).
• No weld: (0 0).
2.3.1. Good weld
2.3.1.1. Weld nugget. The cast microstructure of the weld nugget has coarse
grains. The weld nugget thickness is enough to produce high attenuation.
The weld nugget diameter is larger than the width of the ultrasonic beam
(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. 2. Micrograph of a cross-sectioned good weld.
Fig. 3. Macrograph of a cross-sectioned good weld.
2.3.1.2. Ultrasonic oscillogram. The sequence of echoes is short because of
the high sound attenuation by the cast microstructure of a thick weld nugget.
The distance between consecutive echoes is the combined thickness of the two
sheets because the ultrasonic beam reflections occur at the outer surface of the
two sheets (Fig. 4).
2.3.2. Under size weld
2.3.2.1. Weld nugget. The weld nugget is a cast microstructure. The weld
nugget diameter is smaller than the width of the ultrasonic beam (Figs. 5 and 6).
2.3.2.2. Ultrasonic oscillogram. The ultrasonic beam reflections that occur at
the interface between the two sheets give rise to one-layer echoes between prin-
cipal echoes (Fig. 7).
Fig. 4. Ultrasonic oscillogram of a good weld (left) and its respective input of the ANN (right).
Fig. 5. Micrograph of a cross-sectioned undersize weld.
Fig. 6. Macrograph of a cross-sectioned undersize weld.
2.3.3. Stick weld
2.3.3.1. Weld nugget. The weld nugget is a cast microstructure but is thinner
than the weld nugget of a good weld. The weld nugget diameter is larger than
the width of the ultrasonic beam (Figs. 8 and 9).
2.3.3.2. Ultrasonic oscillogram. The sequence of echoes is longer than the
sequence in a good weld because the sound attenuation is lower. The distance
between echoes is the combined thickness of the two sheets (Fig. 10).
2.3.4. No weld
2.3.4.1. Weld nugget. There is no melting therefore the grain is not as coarse
as the grain of a cast microstructure (Figs. 11 and 12).
2.3.4.2. Ultrasonic oscillogram. The sequence of echoes is the longest because
the sound attenuation is very low. The distance between echoes is the thick-
 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).

2.4. The artificial neural network model

2.4.1. Training algorithm

The Neural Network Toolbox in the MATLAB® 6.1 package is used for the
simulation of the ANN [23,24].
In the present work, a back-propagation multilayer feedforward ANN is
used.
For a M layer ANN (input layer included), the output of neuron l in layer
k + 1 is [25]:
 Sk 

k+1 k+1 k+1 k k+1
a (l) = f w (l, j)a (j) + b (l) , k = 1, . . . , M − 1 (1)
j=1
Fig. 8. Micrograph of a cross-sectioned stick weld.
where

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.

(IV) Obtain x:

−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

The aim is to get a well-trained ANN capable of performing

Fig. 15. Hyperbolic tangent function (tansig) and identity function (purelin).
The overfitting phenomenon may appear in the training when the ANN
memorizes the training data instead of constructing an input–output mapping
for the problem in question. Therefore the overfitting phenomenon gives rise
to the drop of the ability to generalize of the ANN. The data are organized in
input/target pairs. The total data set has 438 pairs and is randomly split into two
subsets [26,31,32]:
• Training subset. With 376 input/target pairs for training the ANN (Q = 376).
In the training, the synaptic weights are repeatedly updated to reduce an error
function.
• Validation subset. With 62 input/target pairs for avoiding overfitting and
achieving good generalization by means of cross validation. Training is
stopped if the error with respect to validation subset begins to increase (early
stopping). This procedure is run together with the training procedure but the
synaptic weights are not updated.
The number of neurons in the input and output layers of the ANN are,
respectively, 10 and 2 (S1 = 10; S4 = 2). There are two hidden layers (M = 4)
with the same numbers of neurons in each one. The transfer function (f2 , f3 ) for
the hidden layers (Fig. 15) is the hyperbolic tangent function (called tansig in
MATLAB® ), the transfer function (f4 ) for the output layer (Fig. 15) is the identity
function (called purelin in MATLAB® ) [23,24]. In order to determine the number
of neurons in the hidden layers it is necessary to consider that if too few neurons
are used the ANN is not capable of fitting the input–output mapping but if too
many neurons are used overfitting can happen [26]. The training mean square
error (MSE) decreases monotonically with increasing number of neurons in the
hidden layers. The validation MSE decreases initially with increasing number
of neurons in the hidden layers but after a certain value is reached, the validation
MSE increases due to the overfitting phenomenon [24]. Table 3 shows that this
certain value of the number of neurons in the hidden layers is seven (S2 = 7;
S3 = 7).
the human expert’s function and classifying spot welds in quality
levels from their respective ultrasonic oscillograms, so the ANN
must have an appropriate ability to generalize.
The 62 input vectors used for cross validation are used for
assessing the ability to generalize of the previously trained
ANN.
An experimental output is a two-component vector. A com-
ponent value in the 1 ± δ interval is considered as 1 and
a component value in the 0 ± δ interval is considered as 0.
Other values of components are considered non-valid [22]
(Fig. 16).
The 62 input vectors are presented to the ANN that com-
pares each experimental output vector with its respective √ tar-
get (Table 4 shows each comparison and its success, , or
failure, ×). In these comparisons, the success rate depends
on the accuracy level, and therefore, depends on the inter-
val radius (δ) (Fig. 17). If δ > 0.25, the success rate is 100%
(Fig. 17).
An unequivocal ANN classification is essential and more
important than a good success rate. An equivocal ANN classi-
fication may appear if the component value of the experimental
output vector is between 0 and 1 and is ≥0.5 distant from its
respective target. The unequivocal ANN is guaranteed by a suc-
cess rate = 100% for a δ = 0.5 (Fig. 17).
Sometimes it is difficult for a human expert to distinguish a
good weld from a stick weld. This problem may cause serious
setbacks. The ANN is capable of distinguishing between two
similar spot welds classified in different quality levels (a good
weld and a stick weld), showing a good fault tolerance (capacity
for robust computation).

Table 3
Determination of the number of neurons in the hidden layers
ANN Training MSE Validation MSE

10-1-1-2 9.501E − 02 8.681E − 02

10-2-2-2 9.359E − 02 8.244E − 02
10-3-3-2 3.240E − 02 4.970E − 02
10-4-4-2 1.041E − 02 1.269E − 02
10-5-5-2 2.507E − 03 1.040E − 02
10-6-6-2 1.424E − 03 1.014E − 02
10-7-7-2 1.336E − 06 9.871E − 04
10-8-8-2 4.074E − 07 1.000E − 02
Fig. 17. Success rate vs. interval radius (δ).
 232
Table 4
Sixty-two input vectors non-used in the training classified by the ANN
i 0.78947 0.85714 0.5625 0.72222 0.095238 0.8 0.75 0.83333 0.52381 0.28571 0.42857 0.71429 0.47619 0.57143 4.5 6 0.15385 0.2381 0.95238 0.41667
0.68421 0.61905 0.5 0.44444 0.85714 0.6 0.5 0.58333 0.38095 0.85714 0.19048 0.52381 0.28571 0.33333 1.5 1 0.76923 0.42857 0.90476 0.25
0.57895 0.47619 0.3125 0.33333 0.047619 0.4 0.4 0.41667 0.28571 0.28571 0.095238 0.047619 0.095238 0.095238 4.5 5.5 0.15385 0.047619 0.7619 0.16667
0.36842 0.28571 0.125 0.22222 0.61905 0.3 0.3 0.16667 0.14286 0.57143 0 0.095238 0 0 1 0.5 0.46154 0.2381 0.61905 0
0.26316 0.19048 0.0625 0.055556 0.095238 0.2 0.2 0 0.047619 0.21429 0 0 0 0 2 3 0.15385 0.047619 0.57143 0
16 16 17 20 9 13 16 16 15 10 18 18 18 18 7 9 10 11 6 17
15 17 16 20 9 13 15 16 18 10 19 18 20 19 11 10 10 9 10 18
15 17 17 21 10 13 17 17 18 9 20 10 20 20 9 10 10 10 10 18
15 16 16 20 11 15 16 17 17 11 0 9 0 0 10 10 10 12 10 0
15 18 18 19 10 12 15 0 17 10 0 0 0 0 10 10 20 10 10 0

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
4. Conclusions
The object of this work is achieved because the proposed
neural model has shown its effectiveness as a tool to carry out
the automation of the classification of resistance spot welds in
quality levels through their respective oscillograms obtained by
ultrasonic testing:
• The proposed ANN produces good results in the classification
of input vectors non-used in the training. The ANN proves its
adaptability, its robustness and therefore its ability to gener-
alize.
• The selection of the significant characteristics of the phe-
nomenon in question in order to make the inputs of the ANN is
very important. The way of representing the ultrasonic oscil-
lograms by means of 10-component vectors, that contain the
relative heights of the echoes and the distance between con-
secutive echoes, is appropriate.
References
[1] M. Jou, Real time monitoring weld quality of resistance spot welding for
the fabrication of sheet metal assemblies, J. Mater. Process. Technol. 132
(2003) 102–113.
[2] S. Agashe, H. Zhang, Selection of schedules based on heat balance in
resistance spot welding, Weld. J. 82 (7) (2003) 179s–183s.
[3] E. Bayraktar, D. Kaplan, M. Grumbach, Application of impact tensile test-
ing to spot welded sheets, J. Mater. Process. Technol. 153–154 (2004)
80–86.
[4] O. Martı́n, PhD Thesis, Ciencia de los Materiales e Ingenierı́a Metalúrgica,
Universidad de Valladolid, 2004.
[5] T. Mansour, Ultrasonic testing of spot welds in thin gage steel, in: P.
McIntire (Ed.), Non-destructive Testing Handbook, v. 7 Ultrasonic Test-
ing, 2nd ed., American Society for Non-destructive Testing, USA, 1991,
pp. 557–568.
[6] D.J. Spinella, J.R. Brockenbrough, J.M. Fridy, Trends in aluminum resis-
tance spot welding for the auto industry, Weld. J. 84 (1) (2005) 34–40.
[7] G. Barrera, M.A. Fabián, M. Vélez, L. Villaseñor, Inteligencias artificiales
y ensayos ultrasónicos para la detección de defectos, Rev. Metal. Madrid
37 (2001) 403–411.
[8] K. Hornik, M. Stinchcombe, H. White, Multilayer feedforward networks
are universal approximators, Neural Netw. 2 (1989) 359–366.
[9] N.S. Reddy, A.K.P. Rao, M. Chakraborty, B.S. Murty, Prediction of grain
size of Al-7Si Alloy by neural networks, Mater. Sci. Eng. A 391 (2005)
131–140.
[10] A. Jiahe, X. Jiang, G. Huiju, H. Yaohe, X. Xishan, Artificial neural network
prediction of the microstructure of 60Si2MnA rod based on its controlled
rolling and cooling process parameters, Mater. Sci. Eng. A 344 (2003)
318–322.
[11] R.H. Silverman, A.S. Noetzel, Image processing and pattern recognition
in ultrasonograms by backpropagation, Neural Netw. 3 (1990) 593–603.
233
[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-
destructive evaluation, Ultrasonics 34 (1996) 461–465.
[14] R. Jordan, F. Feeney, N. Nesbitt, J.A. Evertsen, Classification of wood
species by neural network analysis of ultrasonic signals, Ultrasonics 36
(1998) 219–222.
[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.
[16] J. McBride, S. Malinov, W. Sha, Modelling tensile properties of gamma-
based titanium aluminides using artificial neural network, Mater. Sci. Eng.
A 384 (2004) 129–137.
[17] Y. Cho, S. Rhee, Experimental study of nugget formation in resistance spot
welding, Weld. J. 82 (8) (2003) 195s–201s.
[18] J.M. Gerken, R.S. Brown, D.A. DeAntonio, D.W. Dickinson, R.H. Foxall,
I.A. Oehler, L.M. Schetky, M.J. Shields, T.A. Siewert, L.W. Weller, Resis-
tance spot welding, in: K. Mills, J.R. Davis, B.R. Sanders (Eds.), Welding,
Brazing and Soldering, Metals Handbook, vol. 6, 9th ed., American Society
for Metals, Metals Park, OH, 1983, pp. 469–493.
[19] G. Joseph, in: K.J.A. Kundig (Ed.), Copper: Its Trade, Manufacture, Use
and Environmental Status, ASM International, Materials Park, OH, 1999,
pp. 277–280.
[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,
H.J. Frissell (Eds.), Metallography and Microstructures, Metals Handbook,
vol. 9, 9th ed., American Society for Metals, Metals Park, OH, 1985, pp.
392–398.
[21] J. Krautkrämer, H. Krautkrämer, Ultrasonic Testing of Materials, 4th ed.,
Springer-Verlag, Berlin, 1990.
[22] I.E. Dror, M. Zagaeski, C.F. Moss, Three-dimensional target recognition
via sonar: a neural network model, Neural Netw. 8 (1995) 149–160.
[23] N. Selvakumar, P. Radha, R. Narayanasamy, M.J. Davidson, Prediction of
deformation characteristics of sintered aluminium preforms using neural
networks, Model. Simul. Mater. Sci. Eng. 12 (2004) 611–620.
[24] K.K. Tho, S. Swaddiwudhipong, Z.S. Liu, J. Hua, Artificial neural network
model for material characterization by indentation, Model. Simul. Mater.
Sci. Eng. 12 (2004) 1055–1062.
[25] M.T. Hagan, M.B. Menhaj, Training feedforward networks with the Mar-
quardt algorithm, IEEE Trans. Neural Netw. 5 (1994) 989–993.
[26] C.M. Bishop, Neural Networks for Pattern Recognition, Oxford University
Press, New York, 1995.
[27] D.P. Bertsekas, Non-linear Programming, 2nd ed., Athena Scientific, Bel-
mont, MA, 1999.
[28] Specimen dimensions and procedure for shear testing resistance spot, seam
and embossed projection welds, ISO 14273, 2000.
[29] M. Zhou, S.J. Hu, H. Zhang, Critical specimen sizes for tensile-shear testing
of steel sheets, Weld. J. 78 (9) (1999) 305s–313s.
[30] M. Vural, A. Akkus, On the resistance spot weldability of galvanized
interstitial free steel sheets with austenitic stainless steel sheets, J. Mater.
Process. Technol. 153–154 (2004) 1–6.
[31] S. Haykin, Neural Networks: A Comprehensive Foundation, 2nd ed.,
Prentice-Hall, Upper Saddle River, NJ, 1999.
[32] S. Guessasma, C. Coddet, Microstructure of APS alumina–titania coatings
analysed using artificial neural network, Acta Mater. 52 (2004) 5157–5164.