Resistance Spot Welding (RSW) is processed by using aluminum alloy used in the automotive industry. The difficulty of RSW parameter setting leads to inconsistent quality between welds. The important RSW parameters are the welding current, electrode force, and welding time. An additional RSW parameter, that is, the electrical resistance of the aluminum alloy, which varies depending on the thickness of the material, is considered to be a necessary parameter. The parameters applied to the RSW process, with aluminum alloy, are sensitive to exact measurement. Parameter prediction by the use of an artificial neural network (ANN) as a tool in finding the parameter optimization was investigated. The ANN was designed and tested for predictive weld quality by using the input and output data in parameters and tensile shear strength of the aluminum alloy, respectively. The results of the tensile shear strength testing and the estimated parameter optimization are applied to the RSW process. The achieved results of the tensile shear strength output were mean squared error (MSE) and accuracy equal to 0.054 and 95%, respectively. This indicates that that the application of the ANN in welding machine control is highly successful in setting the welding parameters.
In automotive production, each automobile has approximately 7,000 to 12,000 spot welds. The welds are done using the Resistance Spot Welding (RSW) process which is done by a computer controlled robotic welder. The use of RSW on lightweight aluminum alloy is increasing [
The weld quality of the RSW process has been a significant problem for the automotive industry. Manual calculation of welding parameters, operator experience, and technician expertise in adjusting the parameter settings have not been consistently accurate or correct. The calculations have previously been unable to be confirmed against optimal parameters [
The problems mentioned have motivated this study, formally titled “parameter optimization for resistance spot welding of 6061T6 aluminum alloy based on artificial neural network.” The estimation of RSW parameters using ANN will be more efficiently and correctly optimized. The ANN proved to be effective in resolving both linear and nonlinear functions required for adjusting the RSW parameter settings of computer controlled robotic welders, especially in the automotive industry.
The objective of the RSW process is to generate heat rapidly in the joints of the material being welded while minimizing conduction of heat to cooler adjacent material. This heat generation can be expressed by
The resistance in RSW process [
From Figure
The resistance transformation of the material work piece affects the melting temperature in the spot weld in the RSW process and the weld ability of the material [
RSW schematic diagram.
Researchers have studied the relationship of the RSW parameters and confirm that the welding process parameters have a great influence on the weld quality [
This research studied the RSW process under actual conditions in a car body factory which used gun welds controlled by a MFDC Rexroth Bosch welder. Aluminum alloy specimens (6061T6) of both 1 and 2 mm thickness were welded at three resistance levels. Figure
The specimen dimensions (mm).
The mechanical and chemical properties based on KAISER ALUMINUM [
6061T6 aluminum alloy chemical properties.
6061  Si  Fe  Cu  Mn  Mg  Cr  Zn  Ti  Zr  Other  Max. 

Min. (wt%)  0.40  0.0  0.15  0.00  0.8  0.04  0.00  0.00  0.00  Each  0.05 
Max. (wt%)  0.8  0.7  0.40  0.15  1.2  0.35  0.25  0.15  0.05  Total  0.15 
6061T6 aluminum alloy mechanical properties.
Temper  

T6  
Tensile tests  Ultimate (MPA)  Yield (MPA)  Elongation (%) 
Min.: Max.  337 : 340  286 : 288  13.6 : 13.9 
This research used 6 mm copper alloy electrode tips. The experimental design included full factorial analysis based on low, medium, and high thickness of material, with all parameter settings varied at each level of thickness, giving
(a) is the weld joint, (b) is the tested clamping, and (c) is the graphic monitor.
In this research, 75% of the 243 experimental results of the ANN were used to train the robotic welders. 25% of the experimental results were selected randomly for testing, for comparison between the predicted shear strength and the experimental shear strength results. The ANN models were configured to establish the relationship between the welding parameters and the shear strength. The ANN model is a multilayered feedforward algorithm, as shown in Figure
The ANN structure for shear strength prediction.
The ANN was designed and tested for estimating the shear strength by using the welding parameters as input data and the shear strength as output data, using the multilayered feedforward algorithm, and the ANN was trained by the backpropagation algorithm. The final transfer function was a sigmoidal function and was a pure linear function.
This work studied the application of the ANN to calculate optimized parameters for controlling welding robots. Table
Show calculation of the mean squared error (MSE).
Number  Shear strength (kN)  ANN prediction  Squared error  Number  Shear strength (kN)  ANN prediction  Squared error  Number  Shear strength (kN)  ANN prediction  Squared error 

1 











2 











3 











4 











5 











6 











7 











8 











9 











10 











11 











12 











13 











14 











15 











16 











17 











18 











19 











20 













MSE = 0.053978 
The shear strength is obtained when applying the experimental results from the ANN. Applying the ANN calculations results in good parameter optimization as shown in Figure
ANN prediction.
The ANN model is composed of three different layers: input layer, hidden layer, and output layer. The ANN was used to calculate the shear strength at the output layer and the RSW parameters at the input layer while training and using the backpropagation algorithm. 75% of the 243 data values from the full factorial experimental results were used for training and 25% of the data were used for testing the accuracy of the ANN’s calculated values. The results show that hidden nodes, learning rate, MSE goal, and maximum learning epoch are 30, 0.001, 0.04, and 3,000,000, respectively.
When tested, 20 samples were found to have a mean squared error (MSE) of 0.047896 kN. This indicates that the ANN model is capable of predicting the shear strength for adjusting the RSW parameters under test with a small sample which are shown in Table
Calculation of the mean squared error (MSE) case of 20unit experiment.
Number  Shear strength (kN)  ANN prediction  Squared error  Number  Shear strength (kN)  ANN prediction  Squared error  Number  Shear strength (kN)  ANN prediction  Squared error 

1  3.598  3.8369  0.057073  8  3.611  3.6838  0.0053  15  3.626  3.6838  0.003341 
2  3.485  3.0774  0.166138  9  3.585  3.669  0.007056  16  3.623  3.76  0.018769 
3  3.505  3.592  0.007569  10  3.623  3.76  0.018769  17  3.623  3.76  0.018769 
4  3.458  3.669  0.044521  11  3.572  2.9412  0.397909  18  3.802  3.9422  0.019656 
5  3.848  3.8369  0.000123  12  3.566  3.592  0.000676  19  3.836  3.8369  0.000000 
6  3.878  3.9422  0.004122  13  3.84  3.9422  0.010445  20  3.485  3.0774  0.166138 
7  3.585  3.669  0.007056  14  3.525  3.592  0.004489  


MSE = 0.047896 
ANN prediction case of 20unit experiment.
From this study, it is suggested that the RSW operator can use a small sample set appropriate as input into the ANN model which would be sufficient for setting the optimization parameters. However, the accuracy of the parameters may be affected by other factors outside the scope of the ANN, such as the particular welding machine model, the surface condition of the specimen material, the cooling effect of the environment, and electrode wear.
The full factorial experimental results and ANN were developed and successfully tested in an auto body industry plant in Thailand. The results were MSE and accuracy equal to 0.054 and 95%, respectively. Calculating a reliable estimation of shear strength enables parameter settings to achieve high weld quality and reduces both the setting time process and the sample testing. Future work should test the accuracy of other algorithms used to calculate these parameters to enable comparison of the accuracy of other algorithms.
The authors declare that there is no conflict of interests regarding the publication of this paper.
The authors would like to express their profound gratitude to the UK Engineering and Supply Co. Ltd. and Thonburi Automotive Assembly Plant Co., Ltd., Thailand, for allowing use of their premises and providing research equipment to support the project.