DYNAMIC RESISTANCE BASED INTELLIGENT RESISTANCE WELDING by MAHMOUD EL-BANNA DISSERTATION Submitted to the Graduate School of Wayne State University, Detroit, Michigan in partial fulfillment of the requirements for the degree of DOCTOR OF PHILOSOPHY 2006 MAJOR: INDUSTRIAL ENGINEERING Approved by: ______________________________ Advisor Date ______________________________ ______________________________ ______________________________ ______________________________
151
Embed
DYNAMIC RESISTANCE BASED INTELLIGENT … RESISTANCE BASED INTELLIGENT RESISTANCE WELDING by MAHMOUD EL-BANNA DISSERTATION Submitted …
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
DYNAMIC RESISTANCE BASED INTELLIGENT RESISTANCE WELDING
As stated earlier, the objective here is to develop an on-line nugget quality
classification algorithm that employs a Linear Vector Quantization (LVQ) neural network
and to investigate its efficacy on a constant current controller that employs Medium
Frequency Direct Current (MFDC) and a constant heat controller that employs
Alternating Current (AC). The results will be reported in terms of type 1 error (α ) and
type 2 error ( β ) for cold, normal, and expulsion welds. As per the definitions in Table 6,
Type 1 error (α ) (known as false alarm rate) defines the probability of rejecting the null
hypothesis, while it is true. For example, if the null hypothesis defined the weld as
expulsion weld, Type 1 error (α ) defines the probability that the weld is misclassified
as normal or cold weld, while it really is an expulsion weld. Type 2 error ( β ) (known as
failed alarm) defines the probability of not to reject the null hypothesis, while it is false.
It is important to note that that there is a trade off between type (1) error and type
(2) error. If the model is too sensitive (i.e., type (2) error is very low) it is normal to have
a larger number of false alarms (i.e., type (1) error will be high).
Constant Current Controller employing MFDC
As mentioned before, eleven batches of 300 welds each (total 3300 welds
without anchor welds counted), were performed with 10 tips dressed after each batch.
For each batch, 10 small coupons- with 5 welds each (total 55 welds each batch without
anchor weld counted)-were peeled. The total number of investigated welds is 550; 411
were found to be normal welds, 22 cold welds, and 117 welds with expulsion.
45
Table 6, Type (1) and Type (2) error
Statistical
Decision
True State of Null Hypothesis
Ho is true Ho is false
Reject Ho Type (1)
Errorα Correct
Don’t reject Ho Correct Type (2)
Error β
In all tests, the classification of nugget quality is based on resistance profile.
Figure (23), shows an illustrative dynamic resistance profile for three types of welds;
cold, normal, and expulsion, for MFDC with constant current controller. It can be seen
that these profiles are not easily distinguishable. The cold weld dynamic resistance
profile tends to be lower than the other profiles, while the expulsion weld dynamic
resistance profile tends to have a sharp drop especially towards the end.
In this test, LVQ2 network was trained on three, six, and five patterns for cold,
normal, and expulsion welds, respectively. Twelve hidden neurons were used with a
learning rate of 0.01.
Tables 7, 8, and 9 report type 1 errors (α ) and type 2 errors ( β ) for cold, normal,
and expulsion welds when using the entire dynamic resistance profile as an input vector
to the LVQ neural network. It can be seen that the percent of false alarms are lowest for
the cold weld case at 10%, 26% for normal welds, and 44% for expulsion welds. As for
type 2 errors, they are once again lowest for cold welds at 0%, 16% for expulsion welds,
and 44% for normal welds.
46
In order to reduce the dimensionality of the input resistance vector to the LVQ
neural network, different features are entertained in place of the whole vector, and
include:
• Maximum value of the input resistance vector
• Minimum value of the input resistance vector
• Mean value of the input resistance vector
• Standard deviation value of the input resistance vector
• Range value of the input resistance vector
• Root mean square (RMS) value of the input resistance vector
• First region slope (S1) value of the input resistance vector
• Second region slope (S2) value of the input resistance vector
0 50 100 150 200 250100
120
140
160
180
200
220
240
Welding Time(millisecond)
Dyn
amic
Res
ista
nce
(mic
ro o
hm)
Cold Weld
Normal Weld
Expulsion Weld
47
Figure 23, Dynamic resistance for cold, expulsion and normal welds for MFDC with constant current control
• Third region slope (S3) value of the input resistance vector
• Fourth region slope (S4) value of the input resistance vector
• Binned RMS vector of input resistance: Input resistance is divided into 5 bins and
RMS values are calculated for each bin
Table 7, Type1 and 2 errors for classification of cold welds when using the entire dynamic resistance profile with the neural network
Ho: Weld is Cold True State of Null Hypothesis
Statistical
Decision Ho is true Ho is false
Reject Ho α =0.00 1-α =1.00
Don’t reject Ho 1- β =0.90 β =0.10
Table 8, Type 1 and 2 errors for normal welds classification when using the entire dynamic resistance profile with the neural network
Ho: Weld is
Normal True State of Null Hypothesis
Statistical
Decision Ho is true Ho is false
Reject Ho α =0.45 1-α =0.55
Don’t reject Ho 1- β =0.46 β =0.54
Table 9, Type 1 and 2 errors for expulsion welds classification when using the entire dynamic resistance profile with the neural network
Ho: Weld is Expulsion True State of Null Hypothesis
48
Statistical Decision Ho is true Ho is false
Reject Ho α =0.63 1-α =0.37
Don’t reject Ho 1- β =0.69 β =0.31
Features Selection for MFDC Constant Current Control
The criteria for features selection was based on power of the test (i.e. 1- β ) for
the cold, normal, and expulsion welds as shown in Table 10 . The feature that gives the
highest classification percentages for the three types of welds will be chosen as input
for LVQ network. In order to simplify features selection, we assume that interactions
among features are neglected.
In our work, we just employed the most promising feature identified by power of
the test criteria, maximum value of the input resistance vector, as input for LVQ neural
network. Tables 11, 12, and 13 report the type 1 and 2 error results from the network
when just employing this feature. It can be seen that both types of errors are reduced by
using the maximum resistance feature instead of the entire vector of resistance for
normal and expulsion welds. On the other hand, for cold welds, the type 2 error
degrades.
Table 10, Power of the test (1- β ) for different features for MFDC controller
Feature Cold Welds Normal Welds Expulsion WeldsMaximum 99.8% 78.6% 83.0% Minimum 94.6% 13.0% 100.0% Mean 98.3% 13.7% 100.0% Standard deviation 74.9% 60.3% 72.2% Range 100.0% 38.2% 75.0% Root Mean Square (RMS) 92.1% 14.5% 100.0% Slope 1 53.6% 80.2% 79.2% Slope 2 67.7% 100.0% 30.7%
49
Slope 3 73.9% 90.1% 45.8% Slope 4 100.0% 37.4% 99.8% Bin 1 83.6% 31.3% 76.2% Bin 2 90.7% 16.0% 88.7% Bin 3 89.6% 14.5% 100.0% Bin 4 92.1% 100.0% 14.4% Bin 5 98.1% 20.6% 98.6%
Table 11, Type1 and 2 errors for cold welds classification when using the maximum of dynamic resistance profile with the neural network
Ho: Weld is Cold True State of Null Hypothesis
Statistical
Decision Ho is true Ho is false
Reject Ho α =0.00 1-α =1.00
Don’t reject Ho 1- β =0.88 β =0.12
Table 12, Type1 and 2 errors for normal welds classification when using maximum of dynamic resistance profile with the neural network
Ho: Weld is
Normal True State of Null Hypothesis
Statistical
Decision Ho is true Ho is false
Reject Ho α =0.29 1-α =0.71
Don’t reject Ho 1- β =0.81 β =0.19
Table 13, Type1 and 2 errors for expulsion welds classification when using maximum of dynamic resistance profile with the neural network
Ho: Weld is Expulsion True State of Null Hypothesis
Statistical Decision Ho is true Ho is false
Reject Ho α =0.23 1-α =0.77
Don’t reject Ho 1- β =0.87 β =0.13
50
Alternating Current (AC) with constant heat control
In this case, as mentioned before, 120 small coupons were pealed and the
quality of the welds was checked visually. The total number of investigated welds was
720; 509 were found to be normal welds, there were no cold welds, and 211 welds were
observed with expulsion.
In all tests, the classification of nugget quality is based on the dynamic resistance
profile. Figure (24), shows an illustrative dynamic resistance profile for the three types of
welds; cold, normal, and expulsion, for AC constant heat controller. It can be seen that
these profiles are not easily distinguishable. Usually, the cold weld dynamic resistance
profile tends to be lower than the other profiles, while the expulsion weld dynamic
resistance profile tends to have a sharp drop especially towards the end.
In this test, LVQ network was trained on two, six, and five patterns for cold,
normal, and expulsion welds, respectively. Twelve hidden neurons were used with a
learning rate of 0.01
0 5 10 15 20 25 3070
75
80
85
90
95
100
105
110
115
Time in half cycle
Dyn
amic
Res
ista
nce
(mic
ro o
hm)
Normal Welds
Expulsion Welds
Figure 24, Dynamic resistance for cold, expulsion and normal welds for AC with constant heat control
51
Tables 14, 15, and 16 report type 1 and type 2 errors for cold, normal, and
expulsion welds when using the entire dynamic resistance profile as an input vector to
the LVQ neural network with AC controller. Given that no cold welds were observed
during experimentation, false alarms are not applicable ‘NA’. False alarm rate for normal
welds is lowest at 5% in comparison with expulsion welds at 37%.
As for type 2 errors, the failed alarm rates were lowest for expulsion welds at 1%,
3% for cold welds, and 36% for normal welds.
Table 14, Type1 and 2 errors for cold welds classification when using the entire dynamic resistance profile with the neural network for AC controller
Ho: Weld is Cold True State of Null Hypothesis
Statistical
Decision Ho is true Ho is false
Reject Ho α =NA 1-α =NA
Don’t reject Ho 1- β =0.97 β =0.03
Table 15, Type1 and 2 errors for normal welds classification when using the entire dynamic resistance profile with the neural network for AC controller
Ho: Weld is
Normal True State of Null Hypothesis
Statistical
Decision Ho is true Ho is false
Reject Ho α =0.05 1-α =0.95
Don’t reject Ho 1- β =0.64 β =0.36
52
Table 16, Type1 and 2 errors for expulsion welds classification when using the entire dynamic resistance profile with the neural network for AC controller
Ho: Weld is Expulsion True State of Null Hypothesis
Statistical Decision Ho is true Ho is false
Reject Ho α =0.37 1-α =0.63
Don’t reject Ho 1- β =0.98 β =0.01
In order to once again reduce the dimensionality of the input resistance vector to
the LVQ neural network, different features are entertained in place of the whole vector
(same initial features used in MFDC constant current controller with five RMS bins).
Features screening was once again performed using the power of the test
criteria, with ignoring interactions between features. The minimum feature was used as
input for LVQ neural network as shown in Table 17.
Table 17, power of the test (1- β ) for different features for AC controller
Feature Cold Welds Normal Welds Expulsion WeldsMaximum 86.8% 97.7% 16.6% Minimum 98.3% 76.1% 95.7% Mean 96.8% 50.3% 77.5% Standard deviation 100.0% 95.4% 41.7% Range 90.4% 82.2% 88.4% Root Mean Square (RMS) 96.7% 57.9% 75.9% Slope 1 100.0% 83.4% 24.0% Slope 2 71.8% 57.5% 76.5% Slope 3 98.6% 0.0% 100.0% Slope 4 100.0% 27.9% 96.8% Bin 1 94.3% 19.6% 93.7% Bin 2 96.7% 1.5% 100.0% Bin 3 98.5% 0.3% 100.0% Bin 4 96.7% 23.7% 93.0% Bin 5 100.0% 83.0% 69.1%
53
Tables 18, 19, and 20 shows type 1 and type 2 errors for cold, normal, and
expulsion welds when using the “minimum” feature of the dynamic resistance vector as
an input to the neural network. It can be noticed that type (1) errors are reduced by
using the “minimum” feature instead of the entire dynamic resistance vector for normal
and expulsion welds.
On the other hand, type (2) errors are reduced for cold and normal welds, while it
increased for the expulsion welds, when using the “minimum” feature instead of the
entire dynamic resistance vector.
The false alarm rate for normal welds at 5% is lower than rate for expulsion
welds at 37%. The failed alarm rate for expulsion welds is the lowest at 1%, and the
rates are 3% and 36% for expulsion and normal welds.
Table 18, Type1 and 2 errors for cold welds classification when using the features of the dynamic resistance profile with the neural network for AC controller
Ho: Weld is Cold True State of Null Hypothesis
Statistical
Decision Ho is true Ho is false
Reject Ho α =NA 1-α =NA
Don’t reject Ho 1- β =1.00 β =0.00
Table 19, Type1 and 2 errors for normal welds classification when using the features of the dynamic resistance profile with the neural network for AC controller
Ho: Weld is
Normal True State of Null Hypothesis
Statistical
Decision Ho is true Ho is false
54
Reject Ho α =0.04 1-α =0.96
Don’t reject Ho 1- β =0.75 β =0.25
Table 20, Type1 and 2 errors for expulsion welds classification when using the features of the dynamic resistance profile with the neural network for AC controller
Ho: Weld is Expulsion True State of Null Hypothesis
Statistical Decision Ho is true Ho is false
Reject Ho α =0.25 1-α =0.75
Don’t reject Ho 1- β =0.96 β =0.04
2.6 Conclusions
The problem of real time estimation of the weld quality from the process data is
one of the major issues in the weld quality process improvement. This is particularly the
case for resistance spot welding. Most of the models offered in the literature to predict
nugget diameter from the process data employ measurements such as voltage and
force and are not suitable in an industrial environment for two major reasons: the input
signals for prediction model are taken from intrusive sensors (which will affect the
performance or capability of the welding cell), and, the methods often required very
large training and testing datasets.
In order to overcome these short comings, we propose a Linear Vector
Quantization (LVQ) neural network for nugget quality classification that employs the
easily accessible dynamic resistance profile as input. The goal is to make an on-line
distinction between normal welds, cold welds, and expulsion welds. Our additional goal
is to address this task when employing two types of weld controllers: Constant Current
Controller that employs Medium Frequency Direct Current and a Constant Heat
55
Controller that employs Alternating Current. The results from applying the LVQ neural
network trained using very limited data collected during the stabilization process are
very promising and are reported in detail. In addition, we report very promising results
when a reduced feature set is employed for classification rather than the complete
dynamic resistance profile. The features were selected using power of test criteria.
Overall, the results are very promising for developing practical on-line quality
monitoring systems for resistance spot-welding machines.
56
CHAPTER 3
INTELLIGENT CONSTANT CURRENT CONTROL FOR RESISTANCE SPOT
WELDING
Resistance spot welding is one of the primary means of joining sheet metal in the
automotive industry and other industries. The demand for improved corrosion resistance
has led the automotive industry to increasingly use zinc coated steel in auto body
construction. One of the major concerns associated with welding coated steel is the
mushrooming effect (the increase in the electrode diameter due to deposition of copper
into the spot surface) resulting in reduced current density and undersized welds (cold
welds). The most common approach to this problem is based on the use of simple
unconditional incremental algorithms (steppers) for preprogrammed current scheduling.
In this paper, an intelligent algorithm is proposed for adjusting the amount of current to
compensate for the electrodes degradation. The algorithm works as a fuzzy logic
controller using a set of engineering rules with fuzzy predicates that dynamically adapt
the secondary current to the state of the weld process. The state is identified by
indirectly estimating two of the main process characteristics - weld quality and expulsion
rate. A soft sensor for indirect estimation of the weld quality employing an LVQ type
classifier is designed to provide a real time approximate assessment of the weld nugget
diameter. Another soft sensing algorithm is applied to predict the impact of changes in
current on the expulsion rate of the weld process. By maintaining the expulsion rate just
below a minimal acceptable level, robust process control performance and satisfactory
weld quality are achieved. The Intelligent Constant Current Control for Resistance Spot
Welding is implemented and validated on a Medium Frequency Direct Current (MFDC)
57
Constant Current Weld Controller. Results demonstrate a substantial improvement of
weld quality and reduction of process variability due to the proposed new control
algorithm.
3.1 Introduction
The demand to improve corrosion resistance has led the auto industry to use
coated steel, which has resulted in stringent requirements on conventional weld
controllers that employ “stepper” type preprogrammed current scheduling. The main
objective of the weld current stepper is to maintain weld nugget size within acceptable
limits while at the same time minimizing electrode growth. Large current steps could
lead to an increase in electrode tip growth due to the use of high current levels. This in
turn requires even larger increases in current, thereby causing a runaway process of
electrode growth. Under these conditions, weld size would deteriorate at a rapid rate.
On the other hand, small increases in welding current result in a slow rate of electrode
tip growth, which is advantageous in terms of electrode life, provided the small
increases in current are sufficient to maintain adequate current density to produce the
required weld nugget size.
A basis for setting up a current stepper can be developed by determining the
pattern of electrode growth obtained in a particular welding cell. Typically, test
procedures suitable for this purpose include the standard electrode life test and the
dynamic/oscillating weldability lobe. Different approaches are used for setting up a weld
current stepper, including subjective methods, fixed increments, constant current
density, gradient following, and iterative approaches.
58
In a subjective or "best guess" approach, current steps are based on maintaining
a slight red glow at the electrode/sheet interface and/or regularly adjusting the current to
a value immediately below the splash or expulsion level. This approach has been found
to give significant improvements in electrode life. While acceptable results can be
achieved by this means, an extreme skill is required in determining the point at which
current is to be increased.
In a fixed (preprogrammed scheduling) increment approach, a current stepper
can be based on increasing either the heat control (i.e. phase shift control) or the actual
welding current, in fixed increments after performing a predetermined number of welds.
Generally, the increment of phase shift can be set between 1% and 5%. It was
concluded [59] that a stepper function based on a fixed increment of the heat control or
phase shift control was not a viable means of extending electrode life in many
instances.
In a constant current density approach, a stepper based on maintaining a
constant current density (current per electrode diameter) that also keeps the electrode
force constant, has been investigated by Williams [59]. It was observed that this
approach was unacceptable due to high rates of electrode growth that occurred.
In a gradient following approach, the gradient of the dynamic weldability lobe can
give a good indication of the optimum stepper. To construct the dynamic weldability
lobe, the welding current is set to achieve a weld diameter equivalent to 5 t (where “t”
is the smallest thickness between the sheet metal to be welded) and welds are
produced until the weld size falls to say 3.5 t . At that point, the current is increased
to return the weld size to 5 t and welding continued at this current level until the weld
59
size again falls to 3.5 t . The current is again increased to give 5 t weld size and
the process repeated to maintain the weld size between 5 t and 3.5 t .An
indication of current stepper requirement, in terms of the number and level of steps, can
be derived from the average slope of the dynamic weldability lobe. The problem with
this approach is the sensitivity of electrode life to the magnitude of the current steps
used to accommodate electrode growth. It is a general experience that small increases
in current at frequent intervals are more beneficial than large infrequent steps. However,
the use of smaller than ideal current steps near the start of an electrode campaign may
result in a reduction in weld size to an unacceptable level. In addition, the gradient of
the dynamic weldability lobe is influenced by coating type.
The iterative approach, developed by Williams and Holiday [60], involves
recalculating the weldability lobe limits by taking into account the higher rate of
electrode growth. The first stage involves calculation of the current I and the area A
necessary to obtain the current density I/A at the electrode contact face at the start of
the welding process. This current density would cause electrode tip growth at a certain
rate dA/dn, where ‘n’ is number of welds, which in turn would necessitate a certain rate
of current increase dI/dn. Based on this current increase and the length of the step,
defined in terms of the number of welds, a new current level is then calculated.
Similarly, a new electrode tip contact area A is calculated from the rate of increase in
the contact area dA/dn. This completes the iteration. A new current density is then
recalculated from these values, with subsequent values for the rate of tip growth dA/dn
and rate of current increases dI/dn calculated each iteration. The main disadvantage of
this approach is that it results in too rapid growth in the electrode diameter.
60
An alternative fuzzy control approach was developed by Messler [61] based on
electrode displacement signal to adjust power delivered to the welds in real time. The
fuzzy control scheme applied to resistance spot welding was capable of adjusting every
weld whose actual electrode displacement curve deviated from the desired or the ideal
electrode displacement curve that produces a good weld. The control actions involve:
(1) Increasing the level of applied current or % heat input anytime the actual profile falls
below the desired one, but in accordance with tuned rules to avoid under or
overshooting, and (2) Reducing or withholding current flow or heat input anytime the
actual curve rises above the desired curve in accordance with tuned rules. The main
problem with this approach is that the signal obtained from intrusive sensor (electrode
displacement) makes the applicability of this approach very difficult, if not impossible, for
industrial implementation.
Chen and Araki [62] proposed a fuzzy control algorithm to adjust the current level
during the production of the weld, by estimating different stages in the weld process
using the dynamic resistance profile. The dynamic resistance profile is divided into four
*Iin: Input Current (Start)Ip: Primary Current Is: Secondary CurrentVs: Secondary VoltageRs: Secondary Resistance
N: Number of normal welds from LVQE: Number of expulsion welds from expulsion algorithmIold: Old primary currentdα: Change of current gain
Iold
Welding Process
Intelligent Constant Current Control
*Input Current (first weld only)
Iin
Ip
Measure Secondary Current
Measure Secondary Voltage
Calculate Secondary Resistance
Fire Primary Current
LVQ based Quality Nugget
Estimation
Expulsion Detection
Fuzzy Control Algorithm
IsVs
Rs
N
E
dα
Z-1
*Iin: Input Current (Start)Ip: Primary Current Is: Secondary CurrentVs: Secondary VoltageRs: Secondary Resistance
N: Number of normal welds from LVQE: Number of expulsion welds from expulsion algorithmIold: Old primary currentdα: Change of current gain
Iold
Welding Process
Intelligent Constant Current Control
*Input Current (first weld only)
Iin
Figure 25, Fuzzy Control Scheme after the first weld
64
On the other hand, in order to get the optimum strength for the weld, the input
parameters (current, time, force) need to be targeted just below the expulsion level
[24].
Expulsion is estimated indirectly from the resistance profile. The main indicator
for expulsion, as pointed out in [24, 37, 64], is the instantaneous drop in the resistance
(Figure 26). In this chapter we use a modified version of the expulsion algorithm from
reference [55].
0 50 100 150 200 25060
80
100
120
140
160
180
Dyn
amic
Res
ista
nce
(mic
ro o
hm)
Welding Time (milli seconds)
Normal Weld
Cold Weld
Expulsion Weld
Figure 26, Secondary resistance profiles for cold, expulsion and normal welds for MFDC constant current control
Lets R(k) denote the secondary resistance value at the current millisecond cycle
(the MFDC weld process takes 233 mS), and R(k-1) and R(k-2) the previous resistance
values.
65
The soft sensing expulsion algorithm continuously checks for a resistance drop
(after the cooling period, in our experiment after 67 milliseconds) that is represented by
the following condition for the resistance:
If Max{R(k-2),R(k-1),R(k)}> Max{R(k-1),R(k)}
Then Elevel(k) = 100*R(k)}1),-Max{R(k
R(k)}1),-Max{R(k-R(k)}1),-R(k2),-Max{R(k
Else
Elevel(k) = 0
To determine if there is an expulsion in the examined weld, the following
conditions are checked against Elevel(k):
If Elevel(k) ≥ A
Or
If {Elevel(67)+…+ Elevel(k)} ≥ B,
where A and B are threshold parameters for expulsion detection (in our experiment A=3,
and B=14).
In order to enhance the indirect estimation of the weld status, another soft
sensing algorithm based on quality nugget estimation is introduced. Quality nugget
estimation employing Learning Vector Quantization (LVQ) classifier is designed to
provide a real time approximation of the weld nugget diameter.
Soft Sensing of Weld Quality
66
The nugget quality estimation algorithm is used to determine the number of
normal welds (normal welds are the welds within the specifications, i.e. they have
nugget diameter more than the minimum acceptable limit and exhibit no expulsion) for
the last window of p welds based on a LVQ neural network.
A two layer LVQ artificial neural network, Figure (27), is trained in a supervised
manner to approximate the mapping between the secondary resistance and the weld
nugget diameter. The LVQ model operates as a classifier that estimates whether the
nugget corresponding to a given secondary resistance pattern belongs to the class of
normal or cold (undersized) welds. The classes that the competitive layer finds are
dependent only on the distance between input vectors.
In this paper, the input P is a vector of dimension 167 (i.e. N=167), which is equal
to the number of millisecond samples in one weld after the pre-heat and cooling phase.
The number of hidden neurons is 12 while the number of output neurons is 3
corresponding to the three categories of welding status; cold, normal, and expulsion.
Consequently, the weight matrices W1 and W2 are of size (167X12) and (12X3),
respectively.
The LVQ model was trained on three, six, and five patterns of the secondary
resistance vector for cold, normal, and expulsion welds, respectively. Twelve hidden
neurons were used with a 0.01 learning rate.
67
Figure 27, LVQ network model [7]. P is the input vector of size N, W1, and S1, 2 are the weight matrices and the number of neurons in the first and second layer.
3.3 Fuzzy Logic Control Algorithm
The primary current for the next window of p welds is calculated by using a fuzzy
control algorithm relating the number of expulsion welds and number of normal welds.
Let "E" denote the number of expulsion welds detected from the expulsion
algorithm, "N" the number of normal welds detected from LVQ neural network, for the
last window of p welds, and dα the change of current.
We define the mechanism for adjusting the current gain based on the number of
expulsion and normal welds in the last window of p welds through the following set of
rules with fuzzy predicates:
If "E" is low AND "N" is low THEN αd = Pa
If "E" is medium AND "N" is low THEN αd = Ng/2
If "E" is high AND "N" is low THEN αd = Ng
If "E" is low AND "N" is medium THEN αd = Pa/2
If "E" is medium AND "N" is medium THEN αd = Ng/4
If "E" is high AND "N" is medium THEN αd = Ng/2
68
If "E" is low AND "N" is high THEN αd = Pa/4
If "E" is medium AND "N" is high THEN αd = Ng/8
If "E" is high AND "N" is high THEN αd = Ng/4
In the rules above low, medium, and high are fuzzy subsets defined on the [0, p]
universe for the number of expulsions "E", and the number of normal welds "N” (Figure
28). Ng and Pa are constants (fuzzy singletons) defining positive, negative, change of
the current gain.
The first three fuzzy rules deal with the case where the number of normal welds
“N” in the last window is low. Based on the number of detected expulsions three
alternative strategies for changing current level are considered:
• If the number of expulsions is low, it is reasonable to think that the state of the
welds is close to the cold welds status. Hence, it is necessary to increase
gradually the amount of current. This is done by modifying the change of
current αd .
• If the number detected expulsions is medium or high, it is reasonable to think that
the state of the welds is close to the expulsion state. Hence, it is necessary to
decrease gradually the amount of current (the amount is different in case of high
vs. medium number expulsions). This is done by modifying the change of the
current αd .
When the number of normal welds in the previous window is medium, the
strategies for adjusting the current level are as follows:
• It is reasonable to expect when we have low expulsion detection that the welds
state is approaching a cold weld. Therefore, the level of current should be
69
increased gradually. This is done by modifying the change of current αd . Note
that the amount of increase when “N” is medium ( αd = Pa/2) is less than the
case when “N” is low ( αd = Pa).
• The next case deals with medium expulsion rate, i.e. the welds state is close to
the expulsion status. This requires a gradual reduction of the current. This is
done by modifying the change of the current αd .Note that the amount of
decrease when “N” is medium ( αd = Ng/4) is less than the case when the “N” is
low ( αd = Ng/2).
• The last case appears when the expulsion rate is high. In this case the level of
current should be lowered dramatically to minimize the number of expulsions.
This is also done by modifying the change of current αd . Note that the amount
of decrease when “N” is medium ( αd = Ng/2) is less than the case when the “N”
is low ( αd = Ng).
The last three fuzzy rules consider high level of normal welds, i.e. satisfactory
weld quality. Their corresponding control strategies are:
• If we have low expulsion detection, the state of the welds will be close to the cold
weld status. Therefore, current level should be increased to prevent potential
cold welds. This is also done by modifying the change of current αd . Note that
the amount of increase when “N” is high ( αd = Pa/4) is less than both previous
cases, i.e. when “N” is low ( αd = Pa) and when “N” is medium ( αd = Pa/2).
• If we have medium expulsion detection, it is reasonable to consider that the state
of the welds is close to the expulsion welds status. Therefore the current level
70
should be decreased gradually. This is done by modifying the change of the
current αd . Note that the amount of decrease when “N” is high ( αd = Ng/8) is
less than both cases when “N” is medium ( αd = Ng/4) or when “N” is low ( αd =
Ng/2).
• In the last case, when the expulsion detection is high, the level of the current
should be significantly decreased. This is done by modifying the change of the
current αd . Note that the amount of decrease when “N” is high ( αd = Ng/4), is
less than both cases when “N” is medium ( αd = Ng/2) or when “N” is low ( αd =
Ng).
Applying the Simplified Fuzzy Reasoning algorithm [65], we obtain an analytical
expression for the change of the current αd depending on the rates of expulsion welds
“E” and normal welds “N” as follows:
∑ ∑∑ ∑
∀ ∀
∀ ∀
Δ
=
i jji
i jjiji
yx
yx
d)().(
).().( ,
νμ
νμ
α
where:
μi: linguistic value of the expulsion weld {low, medium, high}.
νj: linguistic value of the normal weld {low, medium, high}.
x: number of expulsion welds in the previous window detected from expulsion
algorithm.
y: number of normal welds in the previous window detected from LVQ.
71
)(xμ : firing level for the expulsion membership function
)(yν : firing level for the normal membership function
:, jiΔ amount of increment/decrement when the linguistic value of expulsion welds
is “i” and the linguistic value of normal welds is “j”.( for example, if the linguistic value of
the expulsion welds is high and the linguistic value of the normal welds is low then
glowhigh N=Δ , , where Ng negative value determines the change of the current αd )
A triangular shape membership function is used in the fuzzy control scheme,
Figure (28), where this type of membership function depends on three scalar
parameters a, b, c as given by:
Figure 28, Membership functions for 'E' the number of expulsion welds and “N” the number of normal welds in the last 'p' welds
⎪⎪⎪
⎭
⎪⎪⎪
⎬
⎫
⎪⎪⎪
⎩
⎪⎪⎪
⎨
⎧
≤
≤≤−−
≤≤−−
≤
=
xc
cxbbcxc
bxaabax
ax
cbax
,0
,0
),,,;(μ
72
The parameters “a” and “c” locate the "feet" of the triangle and the parameter “b”
locates the peak.
The new target current (Inew) for the next window of p welds will be:
Inew =Iold + αd Iold
where Iold is the current in the previous window of p welds and dα is the change of the
current from fuzzy control algorithm.
3.4 Experimental Setup and Results
Proposed Intelligent Constant Current Controller algorithm was implemented in
Matlab and was experimentally tested in a supervisory control mode in conjunction with
an MFDC Constant Current Controller. Four sets of experiments were performed as
follows. The first group of tests (with/without sealer) was performed using the proposed
Intelligent Constant Current Controller. The second group (with/without sealer) was
carried out by using a conventional stepper mode. The sealer was introduced to
simulate one of the typical disturbances in a plant environment. The schematic of an
MFDC Welder [66] is shown in Figure (29). The experimental setup is illustrated in
Figure (30). Welding machine capacity is 150 KVA, with 680 lb welding force provided
from a pneumatic gun. HWPAL25 electrode type with 6.4 mm face diameter is used.
Welding time used is 233 milliseconds with 11.2 KA as initial input secondary current
from a typical welding standard schedule.
73
Figure 29, Schematics of MFDC Welder [53]
Figure 30, Schematic for set up test
74
Each group of tests consists of sixty coupons, i.e. 360 welds (for each test
without sealer), and ten coupons, i.e. 60 welds (for each test with sealer) with two metal
stacks for each coupon are used for each test. Both tests involved welding 2.00 mm
gage hot tip galvanized HSLA steel with 0.85 mm gage electrogalvanized HSLA steel.
Tables (21 and 22) show the mechanical properties and element analysis for the tested
materials. Coupon dimensions used for testing are (1"×12") with 6 welds on each
coupon with the anchor weld as the first weld.
Thirty six coupons (216 welds) without a sealer between sheet metal and ten
coupons (60 welds) with a sealer for each group of tests were examined. Cold and
expulsion welds were checked visually in each coupon.
dressing was performed.
Table 21, Mechanical properties for the tested material
Material Type 0.85 mm gage, HSLA, electrogalvanized
2.00 mm gage, HSLA, hot dip galvanized
0.2% Yield (MPa) 234 406
Tensile (MPa) 333 474
% Elongation 2 in.(51 mm) gage
38 31
The length of the moving window in the Intelligent Constant Current Controller
algorithm was p = 10, i.e. the soft sensing of expulsion and normal welds was
performed on a sequence of 10 consecutive welds. The negative and positive
consequent singleton values in the rule-base of the fuzzy control algorithm were set at
Ng= -0.09 and Pa= +0.07.
75
In the stepper mode test, an increment of one ampere per weld was used as a
stepper for this test. The initial input current was set at 11.2 kA for all tests, with no
stabilization process to simulate the actual welding setup conditions in the plant after tip
Table 22, Element analysis for the base tested materials (weight percent)
Material Coating Weight (g/m2) 0.85 mm gage, HSLA, electrogalvanized 0.70/0.64 2.00 mm gage, HSLA, hot dip galvanized 0.85/1.35
The test were performed in coupons, each coupon had 6 welds (including the
anchor weld), Figure (44). Fourteen bathes for each type of controller performed;
thirteen of them had 300 welds each, and the other left had 600 welds. After each
batch, tip dressing was performed, so total of eleven tip dressing performed in each test
Figure 41, Coupon used in CHC and CCC test
98
4.6 Results
Two sets of experiments using the Constant heat control (CHC) and the Constant
current control (CCC) were conducted to validate the fuzzy C-mean clustering algorithm
implemented in a hierarchal fashion. Using the same welding data from both controllers,
Principal Components Analysis (PCA) is used to reduce the dimension of the welding
data.
Constant heat control (CHC)
As mentioned before, fourteen batches were performed by using Constant heat
control (CHC); thirteen of them had 300 welds, while one had 600 welds. Electrode tip
dressing was performed after each batch, and the counter was being reset when the
electrode tip dressing was performed, Figure (45).
0 1000 2000 3000 4000 50000
100
200
300
400
500
600
700
Counter
Num
ber o
f Wel
ds
Figure 42, counter shows number of welds on each batch for CHC test
99
Figure (46), shows the result of the fuzzy C-mean clustering algorithm for the
training weld data when using the constant heat control CHC. Cycle secondary voltage
vector of length 24 for each weld data is used as the input for the algorithm. Two welds
performed after electrode tip dressing were used for training.
50 100 150 200 250 300 350 400 4501
2
3
4
5
6
Weld Sequence
Clu
ster
Num
ber
Clu
ster
Siz
e
2
7
8
38
190
229
Weld after tip dressing in the last cluster
Figure 43, Number of clusters obtained from training mode (6 clusters) for CHC.
Table 31, shows the number of clusters and their sizes from the training mode,
which include the welds after the electrode tip dressing in the last cluster. It can be
noticed that clusters number 4, 5, and 6 had small number of welds comparable to other
clusters, therefore a threshold can be set after cluster number 3 (i.e. in the evaluation
mode, if the weld after the tip dressing is classified to any cluster below cluster number
three, the alarm should be fired, and redressing or other actions should be considered).
100
Table 31, Clusters obtained from the training welds for CHC
Cluster No 1 2 3 4 5 6
Size 229 190 38 8 7 2
Figure (47), shows the result of the fuzzy C-mean clustering algorithm for the
evaluation of weld data when using the constant heat control CHC. In the evaluation
mode, clustering of the weld data to one of the six clusters obtained from the training
mode was performed. Eleven welds performed after electrode tip dressing were used
for validation.
500 1000 1500 2000 2500 3000 3500 4000 45001
2
3
4
5
6
7
Weld Sequence
Clu
ster
Num
ber
Weld after tip dressing in the last cluster
Clu
ster
Siz
e
76
70
342
2346
874
821
Figure 44, Clustering of weld data for CHC test
It can be seen clearly that ten out of eleven welds performed after electrode tip
dressing belongs to the last cluster (cluster number 6), while the weld performed after
the last electrode tip dressing belongs to the previous cluster (cluster number 5).
101
In conclusion, all of the welds performed after the electrode tip dressing were
classified to clusters higher than cluster number three (the threshold), therefore all of
the eleventh electrode tip dressings were performed properly.
Table 32, shows the number of clusters and their sizes from the evaluation mode
for CHC test. It can be noticed that the last two clusters (cluster 5 and 6) which contain
the welds performed after the electrode tip dressing had small sizes comparable to the
remaining clusters.
Table 32, Size of the clusters obtained from the evaluation mode for CHC
Cluster No 1 2 3 4 5 6
Size 821 847 2346 342 70 76
Constant Current Control (CCC)
Fourteen batches were performed by using Constant current control (CCC);
thirteen of them had 300 welds, while one had 600 welds. Electrode tip dressing was
performed after each batch, and the counter was being reset when the electrode tip
dressing was performed, Figure (48).
Figure (49), shows the result of the fuzzy C-mean clustering algorithm for the
training weld data when using the constant current control (CCC). Cycle secondary
resistance vector of length 28 for each weld data is used as the input for the algorithm.
Two welds performed after electrode tip dressing were used for training.
102
Table 33, shows the number of clusters and their sizes from the training mode,
with twenty four welds in the last cluster, which contains the two welds after the
electrode tip dressing.
0 1000 2000 3000 4000 50000
100
200
300
400
500
600
700
Counter
Num
ber o
f wel
ds
Figure 45, counter shows number of welds on each batch for CCC test
Table 33, Clusters obtained from the training welds for CCC
Cluster No 1 2 3 4
Size 262 146 42 24
It can be noticed that clusters number 3, and 4, had small number of welds
comparable to other clusters, therefore a threshold can be set after cluster number 2
(i.e. in the evaluation mode, if the weld after the tip dressing is classified to any cluster
103
below cluster number two alarm should be fired, and redressing or other actions should
be considered).
50 100 150 200 250 300 350 400 4501
2
3
4
5
Weld Sequence
Clu
ster
Num
ber
Clu
ster
Siz
e
262
146
42
24
Weld after tip dressing in the last cluster
Figure 46, Number of clusters obtained from training mode (4 clusters) for CCC.
Figure (50), shows the result of the fuzzy C-mean clustering algorithm for the
evaluation of weld data when using the constant current control CCC. In the evaluation
mode, clustering of the weld data to one of the four clusters obtained from the training
mode was performed. Eleven welds performed after electrode tip dressing were used
for validation.
It can be seen clearly that all the eleven welds performed after electrode tip
dressing belongs to the last cluster (cluster number 4).
104
In conclusion, all of the welds performed after the electrode tip dressing were
classified to clusters higher than cluster number two (the threshold), therefore all of the
eleventh electrode tip dressings were performed properly.
500 1000 1500 2000 2500 3000 3500 4000 45001
2
3
4
Weld Sequence
Clu
ster
Num
ber
Clu
ster
Siz
e
973
1435
1107
1016
Weld after tip dressing in the last cluster
Figure 47, Number of clusters obtained from validation mode (4 clusters) for CCC
Table 34 shows the number of the clusters and their sizes from the evaluation
mode for CCC test. It can be noticed that the size of the four clusters are very close to
each other. The last cluster which contains the welds performed after the electrode tip
dressing had the smallest size comparable to the remaining clusters.
Table 34, Size of the clusters obtained from the evaluation mode for CCC
Cluster No 1 2 3 4
Size 1016 1107 1435 973
105
Principal Component Analysis (PCA) with Constant heat control (CHC)
Principle Components Analysis (PCA) is a method for dimension reduction based
on finding the eigenvectors of the covariance matrix (or the correlation matrix) for the
initial random variables. Principle components themselves are particular linear
combination of the initial random variables. (more information about PCA can be
obtained in reference [73]).
Figure (51), shows the Scree plot for the same training welding data when using
Constant heat control. The Scree plot can be used to determine the number of principle
components that should be used as input for the algorithm (i.e. the first four principle
components had the highest eigen values, and they account for 94.4% of the total
variance).
Figure (52), shows the result of the fuzzy C-mean clustering algorithm for the
training weld data when using the constant heat control CHC. The first four principal
components of the cycle voltage (instead of vector of length 24) for each weld data were
used as the input for the algorithm.
On the other hand, three welds performed after electrode tip dressing were used
for training, while using the entire cycle secondary .voltage vector only two welds
performed after electrode tip dressing were used for training. It is obvious that this
increment was due to the reduction in input vector size of the algorithm.
106
5 10 15 200
100
200
300
400
Component Number
Eig
enva
lue
Figure 48, Scree plot for training CHC welding data
100 200 300 400 500 600 7001
2
3
4
5
6
Weld Sequence
Clu
ster
Num
ber
Weld after tip dressing in the last cluster
Clu
ster
Siz
e
10
21
103
208
444
Figure 49, Number and size of clusters obtained from training mode (5 clusters) for CHC when 4 principal components were used as input to the algorithm
Table 35, shows the number of clusters and their sizes from the training mode,
which include the welds after the electrode tip dressing in the last cluster. It can be
107
noticed that clusters number 4, and 5 had small number of welds comparable to other
clusters, therefore a threshold can be set after cluster number 3 (i.e. in the evaluation
mode, if the weld after the tip dressing is classified to any cluster below cluster number
three, the alarm should be fired, and redressing or other actions should be considered).
Table 35, Clusters obtained from the training welds for CHC when 4 principal components were used as input for the algorithm
Cluster No 1 2 3 4 5
Size 444 208 103 21 10
Figure (53), shows the result of the fuzzy C-mean clustering algorithm for the
evaluation of weld data when using the constant heat control CHC. The first four
principal components of the cycle voltage (vector of length 24) for each weld data were
used as the input for the algorithm. In the evaluation mode, clustering of the weld data
to one of the five clusters obtained from the training mode was performed. Ten welds
performed after electrode tip dressing were used for validation.
It can be seen clearly that all the ten welds performed after electrode tip dressing
belongs to the last cluster (cluster number 5).
In conclusion, all of the welds performed after the electrode tip dressing were
classified to clusters higher than cluster number three (the threshold), therefore all of
the tenth electrode tip dressings were performed properly.
Table 36 shows the number of clusters and their sizes from the evaluation mode
for CHC test when using 4 principal components as the input for the algorithm. It can be
108
noticed that the last cluster (cluster number 4) which contains the welds performed after
the electrode tip dressing had small size comparable to the remaining clusters.
1000 1500 2000 2500 3000 3500 4000 45001
2
3
4
5
6
Weld Sequence
Clu
ster
Num
ber
Clu
ster
Siz
e
93
107
538
2153
1638
Weld after tip dressing in the last cluster
Figure 50, Clustering of weld data for CHC test when 4 principal components were used as input for the algorithm
Table 36, Size of the clusters obtained from the evaluation mode for CHC when 4 principal components were used as the input for the algorithm
Cluster No 1 2 3 4 5
Size 1638 2153 538 107 93
109
Principal Component Analysis (PCA) with Constant current control (CCC)
Principle Components Analysis (PCA) is used to reduce the dimension of the
entire input vector (i.e. the secondary resistance vector) used by the fuzzy C-mean
clustering algorithm.
5 10 15 20 250
100
200
300
400
500
600
Component Number
Eig
enva
lue
Figure 51, Scree plot for training CCC welding data
Figure (54), shows the Scree plot for the same training welding data when using
Constant current control. The Scree plot can be used to determine the number of
principle components that should be used as input for the algorithm (i.e. the first seven
principle components had the highest eigen values, and they account for 95.0% of the
total variance).
Figure (55), shows the result of the fuzzy C-mean clustering algorithm for the
training weld data when using the constant current control CCC. The first seven
principal components of the cycle resistance (vector of length 28) for each weld data
were used as the input to the algorithm.
110
On the other hand, three welds performed after electrode tip dressing were used
for training, while using the entire cycle secondary .resistance vector only two welds
performed after the electrode tip dressing were used for training. It is obvious that this
increment was due to the reduction in input vector size of the algorithm.
100 200 300 400 500 600 7001
2
3
4
5
6
7
8
Weld Sequence
Clu
ster
Num
ber
Clu
ster
Siz
e
5
6
12
40
78
232
414
Weld after tip dressing in the last Cluster
Figure 52, Number and size of clusters obtained from training mode (7 clusters) for CCC when 7 principal components were used as input to the algorithm
Table 37, Clusters obtained from the training welds for CCC when 7 principal components were used as input to the algorithm
Cluster No 1 2 3 4 5 6 7
Size 414 232 78 40 12 6 5
Table 37, shows the number of clusters and their sizes from the training mode,
which include the three welds after the electrode tip dressing in the last cluster. It can be
noticed that clusters number 5, 6, and 7 had small number of welds comparable to the
111
other clusters, therefore a threshold can be set after cluster number 4 (i.e. in the
evaluation mode, if the weld after the electrode tip dressing is classified to any cluster
below cluster number four, the alarm should be fired, and redressing or other actions
should be considered).
1000 1500 2000 2500 3000 3500 4000 45001
2
3
4
5
6
7
8
Weld Sequence
Clu
ster
Num
ber
Clu
ster
Siz
e
Weld after tip dressing in the last Cluster
287
503
18
566
793
1300
1064
Figure 53, Clustering of weld data for CCC test when 7 principal components were used as input for the algorithm
Figure (56), shows the result of the fuzzy C-mean clustering algorithm for the
evaluation of weld data when using the constant current control CCC. The first seven
principal components of the cycle resistance (instead of entire vector of length 28) for
each weld data were used as the input for the algorithm. In the evaluation mode,
clustering of the weld data to one of the seven clusters obtained from the training mode
was performed. Ten welds performed after the electrode tip dressing were used for
validation.
112
It can be seen clearly that all the ten welds performed after electrode tip dressing
belongs to the last cluster (cluster number 7).
In conclusion, all of the welds performed after the electrode tip dressing were
classified to clusters higher than cluster number four (the threshold), therefore all of the
tenth electrode tip dressings were performed properly.
Table 38, Size of the clusters obtained from the evaluation mode for CCC when 7 principal components were used as the input for the algorithm
Cluster
No 1 2 3 4 5 6 7
Size 1064 1300 793 566 287 18 503
Table 38, shows the number of clusters and their sizes from the evaluation mode
for CCC test when using 7 principal components as the input for the algorithm. It can be
noticed that the last cluster (cluster number 7) contains all the welds performed after the
electrode tip dressing had medium size comparable to the remaining clusters.
4.7 Conclusions
The fuzzy C-mean clustering algorithm implemented in a hierarchal fashion is
used for on line detecting the electrode health condition after the tip dressing. The fuzzy
C-mean clustering algorithm consists mainly from two modes, training and validation
modes. Two welds occurred after the electrode tip dressings were used for training, and
eleven welds occurred after the electrode tip dressings were used for evaluation.
Fuzzy C-mean clustering algorithm implemented in a hierarchal fashion can be
summarized in the following steps:
113
1. Store the welding data (the entire vector of cycle resistance when using constant
current control (CCC), or the entire vector of cycle voltage when using constant
heat control (CHC)) from the weld number one until at least the weld that
occurred after the first electrode tip dressing (or use the weld data from the first
weld that occurred after the first electrode tip dressing until at least the first weld
that occurred after the second electrode tip dressing).
2. Perform Fuzzy C-mean clustering in hierarchy fashion on the training data, and
store levels of clustering with the clusters centers at each level.
3. Establish a threshold based on the size of the clusters; usually the clusters
deformed towards the end of the hierarchal will have smaller sizes comparable to
the clusters deformed towards the top of the hierarchal.
4. Classify the new weld data in a hierarchy fashion, based on the minimum
distance between the new weld data and the clusters center in each level.
5. In the evaluation mode, if the weld after the tip dressing is classified to any
cluster below the threshold, alarm should be fired and/or redressing or any other
actions should be considered.
Experiments based on Constant Heat Control (CHC) and Constant Current
Control (CCC), were performed to verify the fuzzy C-mean clustering algorithm
implemented in a hierarchal fashion. The entire vector of cycle resistance when using
constant current control (CCC), or the entire vector of cycle voltage when using
constant heat control (CHC) were used as input to the algorithm.
114
Table (39) summarizes the results obtained when using the entire cycle
resistance vector in case of CCC or the entire cycle voltage vector in case of CHC, as
inputs for the fuzzy C-mean clustering algorithm. From the training mode, a threshold
established after the third cluster in case of CCC, and after the second cluster in case of
CHC. All the first welds (used for evaluation) that occurred after the electrode tip
dressings classified above the threshold for both CHC and CCC, therefore we conclude
that all the tip dressings were performed properly.
Table 39, Number and sizes of clusters obtained when using the entire cycle resistance vector in case of CCC or the entire cycle voltage vector in case of CHC, as inputs for the fuzzy C-mean clustering algorithm implemented in a hierarchal fashion
Cluster Number Constant Heat Control
(CHC) Constant Current Control
(CCC) Training mode Evaluation mode Training mode Evaluation mode
1 229 821 262 1016 2 190 847 146 1107
3 38 2346 42 1435
4 8 342 24 973
5 7 70 NA NA
6 2 76 NA NA
It can be seen that the clustering results in Constant heat control (CHC) are
different from Constant current control in the following aspects:
• Constant heat control (CHC) had more number of clusters than Constant current
control (CCC).
• The size of the clusters towards the end of the fuzzy C-mean clustering algorithm
when using Constant heat control (CHC) are smaller than the size of the clusters
115
towards the end of the fuzzy C-mean clustering algorithm when using Constant
current control (CCC).
Principal Components Analysis (PCA) is used to reduce the dimension of the
input vector to the fuzzy C-mean clustering algorithm implemented in a hierarchal
fashion. By using Scree plot, four principal components in case of CHC, and seven
principle components in case of CCC, were used as inputs to the fuzzy C-mean
clustering algorithm. Three welds occurred after the electrode tip dressings were used
for training, and ten welds occurred after the electrode tip dressings were used for
evaluation.
Table (40) summarizes the results obtained when using seven principal
components in case of CCC and four principal components in case of CHC, as inputs to
the fuzzy C-mean clustering algorithm. From the training mode, a threshold established
after the fourth cluster in case of CCC, and after the third cluster in case of CHC. All the
first welds (used for evaluation) that occurred after the electrode tip dressings classified
above the threshold for both CHC and CCC, therefore we conclude that all the tip
dressings were performed properly.
It can be seen that the clustering results in Constant heat control (CHC), when
reducing the dimension of the inputs vector by using the principal components analysis,
is different from Constant current control in the following aspects:
• Constant heat control (CHC) had lass number of clusters than Constant current
control (CCC).
• The size of the clusters towards the end of the fuzzy C-mean clustering algorithm
when using Constant heat control (CHC) are smaller than the size of the clusters
116
towards the end of the hierarchal fuzzy C-mean clustering algorithm when using
Constant current control (CCC).
Table 40, Number and sizes of clusters obtained when using seven principal components in case of CCC and four principal components in case of CHC, as inputs
for the fuzzy C-mean clustering algorithm implemented in a hierarchal fashion
Cluster Number Constant Heat Control
(CHC) Constant Current Control
(CCC) Training mode Evaluation mode Training mode Evaluation mode
1 444 1638 414 1064 2 208 2153 232 1300
3 103 538 78 793
4 21 107 40 566
5 10 93 12 287
6 NA NA 6 18
7 NA NA 5 503
117
CHAPTER 5
CONCLUSIONS AND FUTURE WORK
For several decades, resistance spot welding has been an important process in
sheet metal fabrication. The automotive industry, for example, prefers spot welding for
its simple and cheap operation. The advantages of spot welding are many and include
the following: an economical process, adaptable to a wide variety of materials (including
low carbon steel, coated steels) and thicknesses, a process with short cycle times, and
a relatively robust process with some tolerance to fit-up variations. It is favored in the
automotive industry to join steel frame and body components, where 3000-4000 spot
welds per vehicle result in 30-40 billion welds being made in cars each year in the U.S.
alone.
However, given the uncertainty associated with individual weld quality (attributed
to factors such as tip wear, sheet metal surface debris, fluctuations in power supply
etc.). A solution used extensively in the automotive industry is to over design the
number of welds needed in a vehicle by 25% or more. Such over welding, in lieu of full
control is costly, as 7.5 to 10 billion welds may not be needed. In recent years, global
competition for improved productivity and reduced non-value added activity, is forcing
companies such as the automotive OEMs to eliminate these redundant spot welds. In
order to minimize the number of spot welds and still satisfy essential factors such as
strength, weld quality must be obtained.
118
5.1 Conclusions
The problem of real time estimation of the weld quality from the process data is
one of the major issues in the weld quality process improvement. This is particularly the
case for resistance spot welding. Most of the models offered in the literature to predict
nugget diameter from the process data employ measurements such as voltage and
force and are not suitable in an industrial environment for two major reasons: the input
signals for prediction model are taken from intrusive sensors (which will affect the
performance or capability of the welding cell), and, the methods often required very
large training and testing datasets.
In order to overcome these short comings, we propose a Linear Vector
Quantization (LVQ) neural network for nugget quality classification that employs the
easily accessible dynamic resistance profile as input. The goal is to make an on-line
distinction between normal welds, cold welds, and expulsion welds. Our additional goal
is to address this task when employing two types of weld controllers: Constant Current
Controller that employs Medium Frequency Direct Current and a Constant Heat
Controller that employs Alternating Current. The results from applying the LVQ neural
network trained using very limited data collected during the stabilization process are
very promising and are reported in detail. In addition, we report very promising results
when a reduced feature set is employed for classification rather than the complete
dynamic resistance profile. The features were selected using power of test criteria.
Overall, the results are very promising for developing practical on-line quality
monitoring systems for resistance spot-welding machines.
Based on these results from Linear Vector Quantization (LVQ), an intelligent
algorithm was proposed for adapting the current level to compensate for electrode
119
degradation in resistance spot welding. The algorithm works as a fuzzy logic controller
using a set of engineering rules with fuzzy predicates that dynamically adapt the
secondary current to the state of the weld process. A soft sensor for indirect estimation
of the weld quality employing an LVQ type classifier was designed to provide a real time
approximate assessment of the weld nugget diameter. Another soft sensing algorithm
was applied to predict the impact of the current changes on the expulsion rate of the
weld process. By keeping the expulsion rate just below a minimal acceptable level,
robust process control performance and satisfactory weld quality are achieved. The
Intelligent Constant Current Control for Resistance Spot Welding was implemented and
experimentally validated on a Medium Frequency Direct Current (MFDC) Constant
Current Weld Controller.
Results were verified by benchmarking the proposed algorithm against the
conventional stepper mode constant current control. In the case when there was no
sealer between sheet metal, it was found that the proposed intelligent control scheme
reduced the number of expulsion welds and the number of cold welds by 44% and 29%
respectively, when compared to using the stepper mode.
In the case when there was a sealer type disturbance, the proposed control
algorithm once again demonstrated robust performance reducing the number of cold
welds by 67% compared to the stepper mode, while increasing the number of expulsion
welds by only 5%.
It can be concluded that the Intelligent Constant Current Control Algorithm is
capable of successfully adapting the secondary current level according to welds state
and to maintain a robust performance. An alternative version of the algorithm that is
120
applicable to the problem of supervisory control of the weld level in the Constant Heat
Control Algorithm is under development.
Another important area explored in the thesis concerning the electrode tip
dressing. Electrode plays a major role in resistance spot welding process by
transmitting the mechanical force and the electrical current to the work piece to be
welded. Recently, Zinc sheet metal coated steel has been widely used in the automotive
industry and others to improve the corrosion resistance in auto body constructions.
However, one of the major concerns of using the coated sheet metal is that the
electrode life can be significantly shorter than the bare (uncoated) sheet metal.
In order to decrease the effect of coating on the electrode performance (i.e.
reduce the mushrooming effect), tip dressing is done frequently on the electrode;
usually from 10 to 15 times in the auto industry with the assumption that the tip dressing
is done properly. This assumption can lead to low quality in successive welds if the tip
dressing is not done properly (or not done at all).
Therefore, a fuzzy C-mean clustering algorithm implemented in a hierarchal
fashion is used for on line detecting the electrode health condition after the electrode tip
dressing is performed. The fuzzy C-mean clustering algorithm consists mainly from two
modes, training and validation.
Two different types of controller; Constant heat control (CHC) and Constant
current control (CCC), were used to verify the algorithm. In both (CHC) and (CCC) tests,
all the welds occurred after the electrode tip dressings were classified correctly in
clusters above the threshold, which means that all the tip dressing were done properly.
121
Principal components Analysis (PCA) is used to reduce the dimension of the
input vector of the fuzzy C-mean clustering algorithm. The first four principal
components when using (CHC), and the first seven principal components when using
(CCC), were used as inputs for the fuzzy C-mean clustering algorithm. Again, in both
(CHC) and (CCC) tests, all the welds occurred after the electrode tip dressings were
classified correctly in clusters above the threshold, which means that all the tip dressing
were done properly. It can be concluded from these tests that type 1 error (false alarm)
for the fuzzy C-mean clustering algorithm is zero.
5.2 Recommendations for Future Work
Based on encouraging results of this research, the following directions for the
future work are recommended:
• Implementing of Linear Vector Quantization (LVQ) algorithm with the
adaptive fuzzy control scheme on Medium Frequency Direct Current
(MFDC) with Constant heat control (CHC) (MFDC with CHC is still under
development).
• Verifying Linear Vector Quantization (LVQ) algorithm when incorporating
different types of noises such as axial and radial force misalignment, gap,
and force variation.
• Verifying Linear Vector Quantization (LVQ) algorithm with the adaptive
fuzzy control scheme on different types of materials such as aluminum,
and high strength steel.
• Developing a model that will determine when the electrode needs to be
dressed. Until now, there is no model that provides the industry about
122
when they need to do tip dressing, this work will be a milestone in the area
of spot welding.
123
APPENDIX A
EXPULSION DETECTION ALGORITHM
Expulsion refers to the ejection of molten metal from the weld fusion zone during
the spot welding process. This is undesirable due to detrimental effect on weld nugget
integrity (the loss of metal from the fusion zone can reduce weld size and result in weld
porosity, which may significantly reduce the strength and durability of the welded
joints.[64]
On the other hand, in order to get the optimum strength for the weld, the input
parameters (current, time, force) need to be targeted just below the expulsion.[24]
Expulsion can be caused by four main factors; insufficient electrode force,
excessive heating, worn electrodes, and poor sheet surface condition.
During spot welding, if an excessively high welding current or long welding time
used, the nugget radius can grow larger than the electrode contact radius, and if the
electrode pressure distribution can no longer contain the molten nugget, metal is
ejected. This type of expulsion usually happens during the mid–to-late stages of weld
growth and a considerable volume of molten metal can be lost.
Improper surface conditions or worn electrodes can cause expulsion to happen
at the faying interface or at the electrode/sheet contact surface area (splash) due to
high localized current density on both places. This type of expulsion may occur at any
point during the welding cycles.[64]
Many researchers exposed to the problem of expulsion detection from different
In our experiments, expulsion detection is based on a drop in the resistance as
shown in Figure (57).The current value of the resistance is compared to the minimum
value of the previous two resistances. If the difference is greater than a predetermined
threshold, expulsion flag is raised.
0 50 100 150 200 25060
80
100
120
140
160
180
Dyn
amic
Res
ista
nce
(mic
ro o
hm)
Welding Time (milli seconds)
Normal Weld
Cold Weld
Expulsion Weld
Figure 54, Dynamic resistance for cold, expulsion and good welds for MFDC constant current control
125
APPENDIX B
FUZZY C-MEANS CLUSTERING
Fuzzy C-means (FCM) is a method of clustering which allows one piece of data
to belong to two or more clusters. This method (developed by Dunn 1973 and improved
by Bezdek 1981) is frequently used in pattern recognition. It is based on minimization of
the following objective function:
Jm = 2
1 1ji
N
i
C
j
m
ijcxu −∑∑
= = ,1 ∞≤≤ m
where m is any real number greater than 1, uij is the degree of membership of xi in the
cluster j, xi is the ith of d-dimensional measured data, cj is the d-dimension center of the
cluster, and ||*|| is any norm expressing the similarity between any measured data and
the center.
Fuzzy partitioning is carried out through an iterative optimization of the objective
function shown above, with the update of membership uij by:
uij =
∑=
−
⎟⎟
⎠
⎞
⎜⎜
⎝
⎛
−
−C
k
m
ki
ji
cx
cx
1
12
1
and the cluster centers cj by:
126
cj = ∑
∑
=
=N
i
m
ij
N
ii
m
ij
u
u x
1
1.
This iteration will stop when maxij ( ) ( ){ }uu k
ij
k
ij −+1 < ε , where ε is a termination
criterion between 0 and 1, whereas k are the iteration steps. This procedure converges
to a local minimum or a saddle point of Jm.
The algorithm is composed of the following steps:
1. Initialize U=[uij] matrix, U(0)
2. At k-step: calculate the centers vectors C(k)=[cj] with U(k)
cj = ∑
∑
=
=N
i
m
ij
N
ii
m
ij
u
u x
1
1.
3. Update U(k) , U(k+1)
uij =
∑=
−
⎟⎟
⎠
⎞
⎜⎜
⎝
⎛
−
−C
k
m
ki
ji
cx
cx
1
12
1
4. If || U(k+1) - U(k)||<ε then STOP; otherwise return to step 2.
127
REFERENCES
[1] Scanmaster, "Spot Weld Inspection in the Automotive Industry."
[2] A. W. Society, Welding Handbook, 5 ed., 1968.
[3] ASTM, "Standard Test methods for Measuring Contact Resistance for Electrical
Connections," Philadelphia, PA 1985.
[4] VOGT, 2004, p. Spot Weld Inspection
[5] Anon, "Evaluating the Integrity of Spot Weld Joints in Metal," Insight, vol. 40, pp.
778-779, NOV 1998.
[6] A. Denisov, C. Shakarji, B. Lawford, R. Maev, and J. Paille, "Spot weld analysis
with 2D ultrasonic Arrays," Journal of Research of the National Institute of
Standards and Technology, vol. 109, pp. 233-244, MAR-APR 2004.
[7] B. Doyum, "Ultrasonic Examination of Resistance Spot Welds," Non Destructive
Testing (NDT), vol. 8, 2003.
[8] H. Lee, M. Wang, R. Maev, and E. Maeva, "A study on using scanning acoustic
microscopy and neural network techniques to evaluate the quality of resistance
spot welding," International Journal of Advanced Manufacturing Technology, vol.
22, pp. 727-732, NOV 2003.
[9] V. T. Pronyakin, N. K. Rybakov, and Y. N. Panchenko, "Ultrasonic flaw detection
in the welded joints of thin-walled articles," Russian Journal of Nondestructive
Testing, vol. 33, pp. 215-218, APR 1997.
[10] D. Roberts, J. Mason, and C. Lewis, "Ultrasonic spot weld testing with automatic
classification," Science and Technology of Welding and Joining, vol. 7, pp. 47-50,
FEB 2002.
128
[11] D. R. Roberts, J. Mason, and C. Lewis, "Ultrasonic spot weld testing: attenuation
study," Insight, vol. 42, pp. 720-724, NOV 2000.
[12] B. Rooks, "Advances in resistance welding for body-in-white," Assembly
Automation, vol. 23, pp. 159-162, 2003.
[13] J. X. Tao and J. K. Hu, "Ultrasonic method for evaluation of resistance spot
welding," Progress in Natural Science, vol. 11, pp. S190-S193, MAY 2001.
[14] J. Tusek and T. Blatnik, "Ultrasonic detection of lack of fusion in spot welds,"
Insight, vol. 44, pp. 684-688, NOV 2002.
[15] H. S. Yu and B. G. Ahn, "A study on ultrasonic test for evaluation of spot
weldability in automotive materials," KSME International Journal, vol. 13, pp. 775-
782, NOV 1999.
[16] K. Asibu and C. Chien, "Investigation of Monitoring Systems for Resistance Spot
Welding," Welding Journal, 2002.
[17] A. De, O. Gupta, and L. Dorn, "An Experimental Study of resistance spot welding
in 1 mm thick sheet of low carbon steel," in Proceedings of institute of
mechanical Engineering, Part B, 1996, pp. 341-347.
[18] Dorn, Lutz , Xu, and Ping, "Influence of the mechanical properties of resistance
welding machine on the quality of the spot weld," Welding and Cutting, vol. 1, pp.
12-16, 1993.
[19] S. Gedeon , C. Ulrich, and T. Eager, "Measurement of Dynamic Electrical and
Mechanical Properties of resistance spot welds," Welding Journal, pp. 387-385,
1987.
129
[20] J. E. Gould, l. R. Lebman, and S. Holmes, "A Design of Experiment Evaluation of
Factors Affecting The Resistance Spot Weldability of High Strength Steel," in
AWS Sheet Metal Welding Conference VII, Detroit, MI, 1996.
[21] O. Hahn, L. Budde, and D. Hanitzsch, "Investigation on the Influence of the
mechanical properties of spot welding tongs on the welding process," Welding
and Cutting, vol. 1, pp. 6-8, 1990.
[22] M. J. Karagoulis, "A nuts - and Bolts Approach to the control Resistance
welding," Welding Journal, pp. 27-31, july 1994.
[23] H. J. Krause and B. Lehmkuhl, "Measuring the dynamic mechanical
characteristics of spot and projection welding machines - Measured parameters,
measuring procedures and initial results," Technical Report 1984.
[24] P. Podrzaj, I. Polajnar, J. Diaci, and Z. Kariz, "Expulsion detection system for
resistance spot welding based on a neural network," Measurement Science &
Technology, vol. 15, pp. 592-598, MAR 2004.
[25] T. Satoh, Katayama, J., and Okumura, S.,, "Effects of mechanical properties of
spot welding machine on electrode life for mild steel," International Institute of
Welding, 1988, p. 912.
[26] A. Stiebel, "Thermal Force Feedback System (TFF)," 1990.
[27] H. Chang, Y. Cho, S. Choi, and H. Cho, "A proportional -Integral Controller for
resistance spot welding using nugget expansion," ASME Journal of Dynamic
Systems , Measurement and Control, vol. 111, pp. 332-336, 1989.
130
[28] H. S. Cho and D. W. Chun, "A Microprocessor-Based Electrode Movement
Controller for Spot Weld Quality Assurance," IEEE Transactions on Industrial
Electronics, vol. 32, pp. 234-238, 1985.
[29] K. Haefner, B. Carey, B. Bernstein, K. Overton, and M. Dandrea, "Real-Time
Adaptive Spot-Welding Control," Journal of Dynamic Systems Measurement and
Control-Transactions of the Asme, vol. 113, pp. 104-112, MAR 1991.
[30] K. I. Johnson, "Resistance welding quality control techniques," Metal
Construction and British Welding Journal, vol. 5, pp. 176-181, 1973.
[31] M. Jou, C. Li, and R. Messler, "Controlling resistance spot welding using neural
network and fuzzy logic," Science and Technology of Welding and Joining, vol. 3,
pp. 42-50, 1998.
[32] S. A. Gedeon, C. D. Ulrich, and T. W. Eager, "Measurement of Dynamic
Electrical and Mechanical Properties of resistance spot welds," Welding Journal,
pp. 387-385, 1987.
[33] A. Stiebel, Ulmer, C.,Kodrack, D., and Holmes, B. B., "Monitoring and control of
spot weld operations," in SAE , International Congress and Exposition, Detroit,
Michigan, 1986, p. # 860579.
[34] C. Tsai, W. Dai, D. Dickinson, and J. Papritan, "Analysis and Development of a
Real-Time Control Methodology in Resistance Spot-Welding," Welding Journal,
vol. 70, pp. S339-S351, DEC 1991.
[35] D. N. Waller, & Knowlson, P.M.,, "Electrode separation applied to quality control
in resistance welding," Welding Journal, vol. 12, pp. 168-174, 1965.
131
[36] R. T. Wood, L. W. Bauer, J. F. Bedard, B. M. Bernstein, J. Czechowski, M. M.
Dandrea, and R. A. Hogle, "A Closed-Loop Control-System for 3-Phase
Resistance Spot-Welding," Welding Journal, vol. 64, pp. 26-30, 1985.
[37] D. Dickinson, J. Franklin, and A. Stanya, "Characterization of Spot-Welding
Behavior by Dynamic Electrical Parameter Monitoring," Welding Journal, vol. 59,
pp. S170-176, 1980.
[38] W. L. Roberts, "Resistance variations during spot welding," Welding Journal, vol.
30, pp. 1004-19, 1951.
[39] W. Savage, E. Niplles, and F. Wassel, "Dynamic contact resistance of series spot
welds," Welding Journal, vol. 57, pp. 43-50, 1978.
[40] P. H. Thornton, A. R. Krause, and R. G. Davies, "Contact resistances in spot
welding," Welding Journal, vol. 75, pp. S402-S412, DEC 1996.
[41] J. G. Kaiser, G. J. Dunn, and T. W. Eagar, "The Effect of Electrical-Resistance
on Nugget Formation During Spot-Welding," Welding Journal, vol. 61, pp. S167-
S174, 1982.
[42] S. R. Patange, T. Anjaneyulu, and G. P. Reddy, "Microprocessor-Based
Resistance-Welding Monitor," Welding Journal, vol. 64, pp. 33-38, 1985.
[43] Y. Cho and S. Rhee, "New Technology for measuring dynamic resistance and
estimating strength in resistance spot welding," Measurement Science &
Technology, vol. 11, pp. 1173-1178, 2000.
[44] S. R. Lee, Y. J. Choo, T. Y. Lee, M. H. Kim, and S. K. Choi, "A quality assurance
technique for resistance spot welding using a neuro-fuzzy algorithm," Journal of
Manufacturing Systems, vol. 20, pp. 320-328, 2001.
132
[45] S. C. Wang and P. S. Wei, "Modeling dynamic electrical resistance during
resistance spot welding," Journal of Heat Transfer-Transactions of the ASME,
vol. 123, pp. 576-585, JUN 2001.
[46] M. Jou, "Real time monitoring weld quality of resistance spot welding for the
fabrication of sheet metal assemblies," Journal of Materials Processing
Technology, vol. 132, pp. 102-113, JAN 10 2003.
[47] A. A. Denisov, C. M. Shakarji, B. B. Lawford, R. G. Maev, and J. M. Paille, "Spot
weld analysis with 2D ultrasonic Arrays," Journal of Research of the National
Institute of Standards and Technology, vol. 109, pp. 233-244, MAR-APR 2004.
[48] H. T. Lee, M. Wang, R. Maev, and E. Maeva, "A study on using scanning
acoustic microscopy and neural network techniques to evaluate the quality of
resistance spot welding," International Journal of Advanced Manufacturing
Technology, vol. 22, pp. 727-732, NOV 2003.
[49] J. X. Tao, J. K. Hu, K. Y. Zhou, Y. H. Hu, Y. H. Huang, X. Chen, and Z. Y. Pan,
"Ultrasonic method for evaluation of resistance spot welding," Progress in Natural
Science, vol. 11, pp. S190-S193, MAY 2001.
[50] W. T. C.-M. A. Stiebel, "Thermal Force Feedback System (TFF)," 1990.
[51] A. E. Kannatey and C. S. Chien, "Investigation of Monitoring Systems for
Resistance Spot Welding," Welding Journal, 2002.
[52] C. L. Tsai, W. L. Dai, D. W. Dickinson, and J. C. Papritan, "Analysis and
Development of a Real-Time Control Methodology in Resistance Spot-Welding,"
Welding Journal, vol. 70, pp. S339-S351, DEC 1991.
133
[53] M. Jou, C. J. Li, and R. W. Messler, "Controlling resistance spot welding using
neural network and fuzzy logic," Science and Technology of Welding and Joining,
vol. 3, pp. 42-50, 1998.
[54] Y. Park and H. Cho, "Quality evaluation by classification of electrode force
patterns in the resistance spot welding process using neural networks,"
Proceedings of the Institution of Mechanical Engineers, Part B (Journal of
Engineering Manufacture), vol. 218, p. 1513, 2004.
[55] H. Hasegawa and M. Furukawa, "Electric Resistance Welding System," U.S
Patent 6130369, 2000.
[56] J. Wales, "Wikipedia Free Encyclopedia ", 2003.
[57] Y. Nishiwaki and Y. Endo, "Resistance welding Controller," W. T. Corporation,
Ed. USA, April 10, 2001.
[58] W. Li and D. Cerjanec, "A comparative study of AC and MFDC resistance spot
welding," Anaheim, CA, United States, 2004, p. 99.
[59] N. T. Williams and J. D. Parker, "Review of resistance spot welding of steel
sheets Part 2 - Factors influencing electrode life," International Materials
Reviews, vol. 49, pp. 77-108, APR 2004.
[60] N. T. Williams, R. J. Holiday, and J. D. Parker, "Current stepping programmes for
maximizing electrode campaign life when spot welding coated steels," Science
and Technology of Welding and Joining, vol. 3, pp. 286-294, 1998.
[61] R. W. Messler, J. Min, and C. J. Li, "An intelligent control system for resistance
spot welding using a neural network and fuzzy logic," in IEEE Industry
Applications Conference Orlando, FL, USA, Oct. 8-12, 1995, pp. 1757-1763.
134
[62] X. Chen and K. Araki, "Fuzzy adaptive process control of resistance spot welding
with a current reference model," in Proc. of the IEEE International Conference on
Intelligent Processing Systems, ICIPS
Beijing, China, 1998, p. 190.
[63] S. Lee, Y. Choo, T. Lee, M. Kim, and S. Choi, "A quality assurance technique for
resistance spot welding using a neuro-fuzzy algorithm," Journal of Manufacturing
Systems, vol. 20, pp. 320-328, 2001.
[64] M. Hao, K. A. Osman, D. R. Boomer, C. J. Newton, and P. G. Sheasby, "On-line
nugget expulsion detection for aluminum spot welding and weldbonding," SAE
Trans. Journal of Materials and Manufacturing, vol. 105, pp. 209-218, 1996.
[65] R. R. Yager and D. P. Filev, Essentials of Fuzzy Modeling & Control New York:
John Wiley & Sons
1994.
[66] W. Li, F. Eugene, and D. Cerjanec, "Energy Consumption in AC and MFDC
Resistance Spot Welding," in Sheet Metal Welding Conference XI Sterling
Heights, MI, May 11-14, 2004.
[67] N. T. Williams and W. Waddell, "The Importance of Electrode Tip Growth When
Welding Zinc Coated Steels," in International Institute of Welding , Doc.III-1032-
94, 1994.
[68] S. W. Howes and J. S. H. Lake, in Sheet metal Welding Conference Detroit, MI,
AWS Detroit section, 1990.
135
[69] R. Holliday, J. Parker, and N. Williams, "Prediction of Electrode Campaign Life
When Spot Welding Zinc Coated Steels Incorporating Electrode Tip Dressing
Operations," Ironmaking & Steelmaking, vol. 23, pp. 157-163, 1996.
[70] D. P. Lu F., "Model for estimating electrode face diameter during resistance spot
welding," Science and Technology of Welding and Joining, vol. 4, pp. 285-289,
1999.
[71] J. Matejec and M. Zelenak, "Resistance Spot Welding of Zinc Coated Steel
Sheets with Electrode Dressing," International Institute of Welding, Doc. no.:III-
1023-93, 1993.
[72] F. J. Ganowski and N. T. Williams, "Advances in Resistance Spot and Seam
Welding of Zinc-Coated Steel Strip," vol. 49, 1972.
[73] R. A. Johnson and D. W. Wichern, Applied Multivariate Statistical Analysis,
Fourth ed.: Prentice -Hall, 1998.
[74] W. Li, J. Ni, and S. J. Hu, "On-line Expulsion Detection and Estimation for
Resistance Spot Welding," Technical Paper - Society of Manufacturing
Engineers, 1999.
136
ABSTRACT
DYNAMIC RESISTANCE BASED INTELLIGENT RESISTANCE WELDING
by
MAHMOUD EL-BANNA
MAY 2006
Advisor: Dr. Ratna Babu Chinnam and Dr. Dimitar Filev
Major: Industrial Engineering (Manufacturing)
Degree: Doctor of Philosophy
Resistance spot welding (RSW) is one of the most popular processes employed
for sheet metal assembly. Although used in mass production for several decades, RSW
poses several major problems, most notably, huge variation in weld quality. The
strategy employed by the automobile OEMs to reduce the risk of part failure is to often
require more welds to be performed than would be needed to maintain structural
integrity if each weld was made reliably. Advances over the last decade in the area of
non-intrusive electronic sensors, signal processing algorithms, and computational
intelligence, coupled with drastic reductions in computing and networking hardware
costs, have now made it possible to develop non-intrusive intelligent resistance welding
systems that overcome the above shortcomings.
The research develops an Intelligent Resistance Welding System that improves
the weld quality and reduces the number of welds needed. In particular, there are three
specific research achievements: 1) Development of a resistance welding monitoring
137
system based on Linear Vector Quantization (LVQ) algorithm for accurate in-process
non-destructive classification of nugget quality by using the dynamic resistance (or
voltage) profile, 2) Development of a fuzzy control scheme for adapting the controller
set point for weld quality enhancement, and 3) Development of an algorithm for on-line
evaluation of the electrode condition right after tip dressing.
The fuzzy control scheme developed for adapting the welding controller set point
relies on two soft sensors for expulsion detection as well as weld quality evaluation. The
objective is to operate the welding process just beneath the expulsion level conditions to
achieve optimum weld strength. The adaptive fuzzy control scheme was successful in
reducing the number of bad welds, cold or expulsion welds, when used on Medium
Frequency Direct Current (MFDC) constant current control against the traditional
stepper/no stepper techniques.
Fuzzy C-means clustering algorithm implemented in a hierarchal fashion is used
to evaluate the electrode condition right after tip dressing. The algorithm was
successfully verified on constant current and constant heat alternating current
controllers.
138
AUTOBIOGRAPHICAL STATEMENT
MAHMOUD EL-BANNA
Education
• Currently attending PhD program at Industrial and Manufacturing Engineering, Wayne State University, Detroit, Michigan. (3.8 GPA). (Area of concentration: Intelligent Manufacturing Systems).
• Master of Science in Industrial Engineering (MSIE), University of Jordan, Amman, Jordan. Feb. 2001. Top 5%. (Area of concentration: Manufacturing Processes)
• Bachelor of Science in Mechanical Engineering (BSME), University of Jordan, Amman, Jordan. July 1998 Top 5%. (Area of concentration: Manufacturing and Quality)
Experience Oct/2004 – Present: Project Engineer at Ford Motor Company Wayne State University, Detroit, MI, USA. Project: Intelligent Resistance Welding. Feb/2002 – Oct/2004: Graduate Research Assistant
Wayne State University, Detroit, MI, USA. Projects: Hot Metal Gas Forming, Hydro Forming (HMGF). Magnesium Manufacturability. Carbon Nanotube. June/98 – Dec/2001: Manufacturing Engineer