To improve the corrosion resistant properties of carbon steel, usually cladding process is used. It is a process of depositing a thick layer of corrosion resistant material over carbon steel plate. Most of the engineering applications require high strength and corrosion resistant materials for longterm reliability and performance. By cladding these properties can be achieved with minimum cost. The main problem faced on cladding is the selection of optimum combinations of process parameters for achieving quality clad and hence good clad bead geometry. This paper highlights an experimental study to optimize various input process parameters (welding current, welding speed, gun angle, and contact tip to work distance and pinch) to get optimum dilution in stainless steel cladding of low carbon structural steel plates using gas metal arc welding (GMAW). Experiments were conducted based on central composite rotatable design with full replication technique, and mathematical models were developed using multiple regression method. The developed models have been checked for adequacy and significance. In this study, artificial neural network (ANN) and genetic algorithm (GA) techniques were integrated and labeled as integrated ANNGA to estimate optimal process parameters in GMAW to get optimum dilution.
Prevention of corrosion is a major problem in industries. Even though it cannot be eliminated completely, it can be reduced to some extent. A corrosion resistant protective layer is made over the less corrosion resistant substrate by a process called cladding. This technique not only used to improve life of engineering components but also to reduce their cost. This process is mainly now used in industries such as chemical, textiles, nuclear, steam power plants, food processing, and petrochemical industries [
The most accepted method employed in weld cladding is GMAW. It has got the following advantages [
high reliability,
allposition capability,
ease to use,
low cost,
high Productivity,
suitable for both ferrous and nonferrous metals,
high deposition rate,
absences of fluxes,
cleanliness and ease of mechanization.
The mechanical strength of clad metal is not only highly influenced by the composition of metal but also by clad bead shape. This is an indication of bead geometry. Figure
Clad bead geometry.
This study can be divided into three parts that are experimental analysis, analytical model, and artificial intelligence model. Experimental and analytical models can be developed using conventional and statistical models. Artificial intelligent models such as ANN and GA can be used for prediction. There are many advantages using ANN for prediction such as ANN’s ability to handle nonlinear form of modeling that learn the mapping of input and output. ANN is more successful than conventional methods in terms of speed and simplicity and its capacity to learn from examples. Moreover it does not need any preliminary assumptions. Simple MATLAB Toolbox can be used for prediction.
In this study an integration system is opted for study to get an improvement in GMAW process. In this study two techniques were used: ANN is used for prediction and GA is used for optimization. These two techniques were integrated to form a new system integrated ANNGA. It is proved that the proposed hybrid system ANNGA can produce more significant results than conventional techniques.
The following machines and consumables were used for the purpose of conducting experiments:
a constant current gas metal arc welding machine (Invrtee V 350PRO advanced processor with 5–425 amp output range),
welding manipulator,
wire feeder (LF74 Model),
filler material Stainless Steel wire of 1.2 mm diameter (ER308 L),
gas cylinder containing a mixture of 98% argon and 2% oxygen,
mild steel plate (grade IS2062).
Test plates of size 300 × 200 × 20 mm were cut from mild steel plate of grade IS2062, and one of the surfaces is cleaned to remove oxide and dirt before cladding. ER308 L stainless steel wire of 1.2 mm diameter was used for depositing the clad beads through the feeder. Argon gas at a constant flow rate of 16 litres per minute was used for shielding. The properties of base metal and filler wire are shown in Table
Chemical composition of base metal and filler wire.
Materials  Elements wt%  

C  SI  Mn  P  S  Al  Cr  Mo  Ni  
IS 2062  0.150  0.160  0.870  0.015  0.016  0.031  —  —  — 
ER 308L  0.03  0.57  1.76  0.021  0.008  —  19.52  0.75  10.02 
Relationship between Current and Wire Feed Rate.
The selection of the welding electrode wire is based on the matching of the mechanical properties and physical characteristics of the base metal, weld size, and existing electrode inventory [
Experimental design procedure.
The research work was planned to be carried out in the following steps [
identification of factors and responses,
finding limits of process variables,
development of design matrix,
conducting experiments as per design matrix,
recording the responses,
development of mathematical models,
checking the adequacy of developed models,
conducting conformity tests.
The following independently controllable process parameters were found to be affecting output parameters. These are wire feed rate
The basic difference between welding and cladding is the percentage of dilution. The properties of the cladding is significantly influenced by the dilution obtained. Hence control of dilution is important in cladding where a low dilution is highly desirable. When dilution is quite low, the final deposit composition will be closer to that of filler material, and hence corrosion resistant properties of cladding will be greatly improved. The chosen factors have been selected on the basis of getting minimal dilution and optimal clad bead geometry.
Few significant research works have been conducted in these areas using these process parameters, and so these parameters were used for an experimental study.
Working ranges of all selected factors are fixed by conducting trial runs. This was carried out by varying one of the factors while keeping the rest of them as constant values. Working range of each process parameter was decided upon by inspecting the bead for smooth appearance without any visible defects. The upper limit of a given factor was coded as −2. The coded value of intermediate values were calculated using (
The chosen level of the parameters with their units and notations are given in Table
Welding parameters and their levels.
Parameters  Unit  Notation  Factor levels  

−2  −1  0  1  2  
Welding current  A 

200  225  250  275  300 
Welding speed  mm/min 

150  158  166  174  182 
Contact tip to work distance  mm 

10  14  18  22  26 
Welding gun angle  Degree 

70  80  90  100  110 
Pinch  —  Ac  −10  −5  0  5  10 
Design matrix chosen to conduct the experiments was central composite rotatable design. The design matrix comprises full replication of 2^{5} (=32), factorial designs. All welding parameters in the intermediate levels (o) constitute the central points and combination of each welding parameters at either the highest value (+2) or the lowest value (−2) with other parameters of intermediate levels (0) constitutes star points. 32 experimental trails were conducted to make the estimation of linear quadratic and twoway interactive effects of process parameters on clad geometry [
The experiments were conducted at SVS College of Engineering, Coimbatore, India. In this work thirtytwo experimental runs were allowed for the estimation of linear quadratic and twoway interactive effects correspond to each treatment combination of parameters on bead geometry as shown in Table
Design matrix.
Trial no.  Design matrix  





Ac  
1  −1  −1  −1  −1  1 
2  1  −1  −1  −1  −1 
3  −1  1  −1  −1  −1 
4  1  1  −1  −1  1 
5  −1  −1  1  −1  −1 
6  1  −1  1  −1  1 
7  −1  1  1  −1  1 
8  1  1  1  −1  −1 
9  −1  −1  −1  1  −1 
10  1  −1  −1  1  1 
11  −1  1  −1  1  1 
12  1  1  −1  1  −1 
13  −1  −1  1  1  1 
14  1  −1  1  1  −1 
15  −1  1  1  1  −1 
16  1  1  1  1  1 
17  −2  0  0  0  0 
18  2  0  0  0  0 
19  0  −2  0  0  0 
20  0  2  0  0  0 
21  0  0  −2  0  0 
22  0  0  2  0  0 
23  0  0  0  −2  0 
24  0  0  0  2  0 
25  0  0  0  0  −2 
26  0  0  0  0  2 
27  0  0  0  0  0 
28  0  0  0  0  0 
29  0  0  0  0  0 
30  0  0  0  0  0 
31  0  0  0  0  0 
32  0  0  0  0  0 
In order to measure clad bead geometry of transverse section of each weld, overlays were cut using band saw from mid length. Positions of the weld and end faces were machined and grinded. The specimen and faces were polished and etched using a 5% nital solution to display bead dimensions. The clad bead profiles were traced using a reflective type optical profile projector at a magnification of ×10. Then the bead dimensions such as depth of penetration, height of reinforcement, and clad bead width were measured [
Traced Profiles (Specimen No. 2). 02A represents profile of the specimen (front side) and 02B represents profile of the specimen (rear side). The measured clad bead dimension and percentage of dilution is shown in Table
The response function representing any of the clad bead geometry can be expressed as [
The secondorder surface response model can be expressed as below:
The adequacy of the developed model was tested using the analysis of variance (ANOVA) technique. As per this technique, if the Fratio values of the developed models do not exceed the standard tabulated values for a desired level of confidence (95%) and the calculated Rratio values of the developed model exceed the standard values for a desired level of confidence (95%), then the models are said to be adequate within the confidence limit [
Artificial neural network (ANN) is biologically inspired by intelligent techniques [
Model of an Artificial Neuron.
Neural networks have many usages in the present decade. It has been successfully applied across an entire ordinary range of problems domains such as finance, medicine, energy, geology and physics where there is a problem of prediction neural network can be used successfully [
Neural network architecture.
MATLAB 7 was used for tracing the network for the prediction of clad bead geometry. Statistical mathematical model was used to compare results produced by the work. For normalizing the data, the goal is to examine the statistical distribution of values of each net input, and outputs are roughly uniform; in addition the value should be scaled to match range of input neurons [
This is basically in the range of 0 to 1 in practice it is found to between 01 and 9 [
In this study five welding process parameters were employed as input to the network. The LevenbergMarquardt approximation algorithm was found to be the best fit for application because it can reduce the MSE to a significantly small value and can provide better accuracy of prediction. So neural network model with feed forward back propagation algorithm and LevenbergMarquardt approximation algorithm was trained with data collected for the experiment. Error was calculated using (
The difficulty in using the regression equation is the possibility of overfitting the data. To avoid this, the experimental data is divided into two sets: one training set and other test data set. The ANN model is created using only training data; the other test data is used to check the of behavior the ANN model created. All variables are normalized using (
Neural network general form can be defined as a model shown above
The training process involves the derivation of weights by minimization of the regularized sum of squared error.
The complexity of model is controlled by the number of hidden level values of regularization constants and are associated with each input one for biases and one for all weights connected to output.
The effectiveness of ANN model fully depends on the trial and error process. This study considers the factors that could be influencing the effectiveness of the model developed. In the MATLAB 7 Toolbox there are five Influencing factors listed below:
network algorithm,
transfer function,
training function,
learning function,
performance function.
The ANN structure consists of three layers which are input, hidden, and output layers. It is known that ANN model is designed on trial and error basis. The trial and error is carried out by adjusting the number of layers and the number of neurons in the hidden structure. Too many neurons in hidden layer result in a waste of computer memory and computation time, while too few neurons may not provide the desired data control effect. The process is conducted using 28 randomly selected samples. Seventeen data are used for training and eleven used for testing the data. Table
Design matrix and observed values of clad bead geometry.
Trial no.  Design matrix  Bead parameters  





Ac 





1  −1  −1  −1  −1  1  6.9743  1.67345  6.0262  10.72091 
2  1  −1  −1  −1  −1  7.6549  1.9715  5.88735  12.16746 
3  −1  1  −1  −1  −1  6.3456  1.6986  5.4519  12.74552 
4  1  1  −1  −1  1  7.7635  1.739615  6.0684  10.61078 
5  −1  −1  1  −1  −1  7.2683  2.443  5.72055  16.67303 
6  1  −1  1  −1  1  9.4383  2.4905  5.9169  15.96692 
7  −1  1  1  −1  −1  6.0823  2.4672  5.49205  16.5894 
8  1  1  1  −1  −1  8.4666  2.07365  5.9467  14.98494 
9  −1  −1  −1  1  −1  6.3029  1.5809  5.9059  10.2749 
10  1  −1  −1  1  1  7.0136  1.5662  5.9833  9.707297 
11  −1  1  −1  1  1  6.2956  1.58605  5.5105  11.11693 
12  1  1  −1  1  −1  7.741  1.8466  5.8752  11.4273 
13  −1  −1  1  1  1  7.3231  2.16475  5.72095  15.29097 
14  1  −1  1  1  −1  9.6171  2.69495  6.37445  18.54077 
15  −1  1  1  1  −1  6.6335  2.3089  5.554  17.23138 
16  1  1  1  1  1  10.514  2.7298  5.4645  20.8755 
17  −2  0  0  0  0  6.5557  1.99045  5.80585  13.65762 
18  2  0  0  0  0  7.4772  2.5737  6.65505  15.74121 
19  0  −2  0  0  0  7.5886  2.50455  6.4069  15.77816 
20  0  2  0  0  0  7.5014  2.1842  5.6782  16.82349 
21  0  0  −2  0  0  6.1421  1.3752  6.0976  8.941799 
22  0  0  2  0  0  8.5647  3.18536  5.63655  22.94721 
23  0  0  0  −2  0  7.9575  2.2018  5.8281  15.74941 
24  0  0  0  2  0  7.7085  1.85885  6.07515  13.27285 
25  0  0  0  0  −2  7.8365  2.3577  5.74915  16.63287 
26  0  0  0  0  2  8.2082  2.3658  5.99005  16.38043 
27  0  0  0  0  0  7.9371  2.1362  6.0153  15.18374 
28  0  0  0  0  0  8.4371  2.17145  5.69895  14.82758 
29  0  0  0  0  0  9.323  3.1425  5.57595  22.8432 
30  0  0  0  0  0  9.2205  3.2872  5.61485  23.6334 
31  0  0  0  0  0  10.059  2.86605  5.62095  21.55264 
32.  0  0  0  0  0  8.9953  2.72068  5.7052  19.60811 
Analysis of variance for testing adequacy of the model.
Parameter  1storder terms  2ndorder terms  Lack of fit  Error terms 


Whether model is adequate  

SS  DF  SS  DF  SS  DF  SS  DF  

36.889  20  6.233  11  3.513  6  2.721  5  1.076  3.390  Adequate 

7.810  20  0.404  11  0.142  6  0.261  5  0.454  7.472  Adequate 

1.921  20  0.572  11  0.444  6  0.128  5  2.885  3.747  Adequate 

506.074  20  21.739  11  6.289  6  15.45  5  0.339  8.189  Adequate 
SS: sum of squares; DF: degree of freedom;
It is necessary to normalize the quantitative variable to some standard range from 0 to 1. The number of neurons’ hidden layers should be approximately equal to
In order to determine the best network structure of ANN prediction model, the randomized data set is divided into two sets: one training data set which is used for the prediction of the model, and the other test data which is used to validate the model. Seventeen data were used for training set and eleven data for test set. This to avoid the over fitting of the data. An ANN model is created and trained using MATLAB 7 ANN Toolbox. The lowest MSE obtained is for twelve hidden neurons. So a network structure is 5124: five input neurons, twelve hidden neurons, and four output neurons. Then the test data is validated against the ANN model created; the results are shown in Table
Comparison of actual and predicted values of the clad bead parameters using neural network data (test).
Trial no.  Actual bead parameters  Predicted bead parameters  Error  














1  6.9743  1.6735  6.0262  10.721  6.1945  1.85  5.9611  12.367  0.7798  −0.177  0.0651  −1.646 
2  7.6549  1.9715  5.8873  12.167  7.1815  2.1507  6.5553  10.268  0.4734  −0.179  −0.668  1.899 
3  6.3456  1.6986  5.4519  12.746  7.4954  1.5339  5.4923  9.3808  −1.15  0.1647  −0.04  3.3652 
4  7.7635  1.7396  6.0684  10.611  6.4936  1.854  6.5573  9.4799  1.2699  −0.114  −0.489  1.1311 
5  7.2683  2.443  5.7206  16.673  7.3354  2.6576  5.5657  19.104  −0.067  −0.215  0.1549  −2.431 
6  9.4383  2.4905  5.9169  15.967  7.6066  2.1045  6.4342  18.49  1.8317  0.386  −0.517  −2.523 
7  6.0823  2.4672  5.492  16.589  8.0417  2.1722  5.5126  16.874  −1.959  0.295  −0.021  −0.285 
8  8.4666  2.0737  5.9467  14.985  8.3236  2.2349  5.9031  16.972  0.143  −0.161  0.0436  −1.987 
9  6.3029  1.5809  5.9059  10.275  8.2381  1.7955  5.6022  11.219  −1.935  −0.215  0.3037  −0.944 
10  7.0136  1.5662  5.9833  9.7073  7.5899  2.4579  6.542  13.415  −0.576  −0.892  −0.559  −3.708 
11  6.2956  1.586  5.5105  11.117  7.7318  1.7647  5.8676  10.71  −1.436  −0.179  −0.357  0.407 
Comparison of actual and predicted values of the clad bead parameters using neural network data (training).
Trial no.  Actual bead parameters  Predicted bead parameters  Error  














1  7.741  1.8466  5.8752  11.4273  7.335  2.0986  6.0792  10.8222  0.406  −0.252  −0.204  0.6051 
2  7.3231  2.16475  5.72095  15.29097  6.8214  2.0617  5.6946  14.9379  0.5017  0.10305  0.02635  0.35307 
3  9.6171  2.69495  6.37445  18.54077  9.3713  2.8982  6.4084  17.4578  0.2458  0.20325  −0.0339  1.08297 
4  6.6335  2.3089  5.554  17.23138  7.4306  2.2927  5.6232  15.7908  −0.7971  0.0162  −0.0692  1.44058 
5  10.514  2.7298  5.4645  20.8755  7.8991  2.5154  5.8078  18.0664  2.6149  0.2144  −0.3433  2.8091 
6  6.5557  1.99045  5.80585  13.65762  6.5761  1.9158  5.7867  14.2039  −0.0204  0.07465  0.01915  −0.5462 
7  7.4772  2.5737  6.65505  15.74121  7.393  2.7191  6.7112  14.7525  0.0842  −0.1454  −0.0561  0.98871 
8  7.5886  2.50455  6.4069  15.77816  7.5943  2.4317  6.3834  15.9881  −0.0057  0.07285  0.0235  −0.2099 
9  7.5014  2.1842  5.6782  16.82349  7.4652  2.2814  5.7674  16.5744  0.0362  −0.0972  −0.0892  0.24909 
10  6.1421  1.3752  6.0976  8.941799  5.6583  1.44  6.2054  9.3753  0.4838  −0.0648  −0.1078  −0.4335 
11  8.5647  3.18536  5.63655  22.94721  9.9724  2.962  5.5227  18.9566  −1.4077  0.22336  0.11385  3.99061 
12  7.9575  2.2018  5.8281  15.74941  9.0693  2.6919  6.2337  17.5548  −1.1118  −0.4901  −0.4056  −1.8053 
13  7.7085  1.85885  6.07515  13.27285  6.7699  1.7807  6.109  12.8584  0.9386  0.07815  −0.0338  0.41445 
14  7.8365  2.3577  5.74915  16.63287  8.5364  2.9431  6.6735  15.9653  −0.6999  −0.5854  −0.9243  0.66757 
15  8.2082  2.3658  5.99005  16.38043  8.0083  2.371  6.0186  16.3701  0.1999  −0.0052  −0.0285  0.01033 
16  7.9371  2.1362  6.0153  15.18374  7.9441  2.1197  6.01  15.3735  −0.007  0.0165  0.0053  −0.1897 
17  8.4731  2.17145  5.69895  14.82758  8.6735  2.5165  5.4985  15.2875  −0.2001  −0.3450  0.2031  −0.4599 
Genetic algorithm is meta heuristic searching techniques, which mimics the principles of evaluation and natural genetics. These are guided by the random search which scans through entire sample space and therefore provide a reasonable solution. It was introduced by Holland (1975). It is also considered as a heuristic technique inspired by natural biological evolution process comprising selection, cross over mutation, and so forth.
In biological population genetic information is stored in the form of binary strings. The basic processes which affect the binary strings makeup in natural evolution are a selection, a crossover of genetic information between reproducing parents, a mutation of genetic information, and an elitist strategy that keeps the best individual to next generation. The main operations of GA are characterized as follows.
Selection is a method that randomly picks up chromosomes out of the population according to the evolution function. The higher the fitness function is, the more chance of an individual has to be selected. The selection pressure is defined as the degree to which the better individuals are favoured.
It takes two individuals and puts their chromosome strings at some randomly chosen position to produce two head segments and two tail segments. The tail segments are then supposed to produce two new full length chromosomes as shown in Figure
Single—Point cross over.
Mutation is applied to each child individually after cross over. It randomly alters each gene with small probability (typically .001) Table
Single mutations.
Offspring  1  0  1  0  0  1  0  0  1  0 
Mutual offspring  1  0  1  0  1  1  0  0  1  0 
The fitness function is an exponential function of one variable with a maximum
Details of individual.
Individual 

Fitness  Chromosome 

Parent I  0.08  0.05  0001010010 
Parent II  0.73  0.000002  1011101011 
Offspring I  0.23  0.47  0011101011 
Offspring II  0.58  0.00007  1001010010 
If genetic algorithm has been correctly implemented, the population will evolve over successive generation so that fitness of the best and the average individual in each generation increases towards the global optimum. Convergence is the progression towards increasing uniformity. A gene is said to have converged when 95% of the population shares the same value (Dejony 1975). The population is said to have converged when all of the genes have converged. Traditional genetic algorithm is shown in Figure
Traditional Genetic Algorithm.
The experimental data related to welding current
The aim of the study is to find for optimum adjustments for welding current
GA search ranges.
Parameters  Range 

Welding current ( 
200−300 amp 
Welding speed ( 
150−182 mm/min 
Contact tip to work distance ( 
10−26 mm 
Welding gun angle ( 
70−110 deg 
Pinch (Ac)  −10–10 
For GA computation.
Population type  Double vector 
Population size  30 
Fitness scaling function  Rank 
Selection function  Roulette 
Reproduction elite count  2 
Cross over rate  100 
Cross over function  Intermediate 
Mutation function  Uniform 
Mutation rate  1% 
Number of generation  52 
Migration  Forward 
The objective function selected for optimizing was the percentage of dilution. The response variables bead width
minimize
subject to
Genetic algorithms are nowadays a popular tool in optimizing because GA uses only the values of objective function. The derivatives are not used in the procedure. Secondly the objective function value corresponding to a design vector plays the role of fitness in natural genetics. The aim of the study is to find the optimum adjustments for welding current, welding speed, pinch, welding angle, and contact to tip distance. Objective function selected for optimization was the percentage of dilution. The process parameters and their notations used in writing the program in MATLAB 7 software are given below:
Objective function for percentage of dilution which must be minimized was derived from (
subjected to bounds
Consider the following:
Consider
MATLAB program in GA and GA function were used for optimizing the problem. The program was written in GA and constraint bounds were applied. The minimum percentage of dilution obtained from the results was while obtained running the GA tool. The minimum percentage of dilution obtained is 10.396. the value of process parameters obtained is
Fitness value of GA function.
The methodology applied in this study involves six cases [
To estimate the minimum values of cladding parameters compared to the performance values of the experimental data, regression modeling, and ANN singlebased modeling.
To estimate the optimal process parameters values that have been within the range of minimum and maximum coded values for process parameters of experimental design that are used for experimental trial.
To estimate the optimal solution of the process parameters within the small number of iterations compared to the optimal solution of the process parameters with singlebased GA optimization.
The steps in order to implement the integrated ANNGA type A and integrated ANNGA typeB in fulfilling the three objectives are as follows.
In the experimental data module the values of dilution for different combinations of process parameters are used for modeling.
In the regression modeling schedule model was developed using cladding process parameters. A multilinear regression analysis was performed to predict dilution, and a governing equation was constructed [
A predicted model was developed using ANN. The percentage of error was calculated between predicted and actual values.
In the singlebased GA optimization, the predicted equation of the regression model would become the objective function. The minimum and maximum coded values of the process parameters of the experimental design would define the boundaries for minimum and maximum values of the optimal solution.
In the integrated ANNGA type A module which was the first integration system proposed in this study. Similar to GA singlebased optimization process, the predicted equation of the regression would become the objective function. The integration system possesses the optimal process parameters value of the single based GA optimization process combined with the process parameters of ANN system would define the boundaries for minimum and maximum values for optimal solution.
In the integrated ANNGA type B system which is the second integration system proposed in this study. Similar to type A, the predicted regression equation would become the objective function and the optimal process parameters value of singlebased GA optimization combined with process parameters value of the ANN model would define the boundaries for the minimum and maximum values of the optimal solution. This integration system proposes the process parameters values of the best ANN model to define the initial point for optimization solution. This process is shown in Figure
Integrated ANNGA optimization procedure.
The strategy of this study in implementing integrated ANNGA type A is implemented by proposing the optimal process parameters value of the GA combined with the nonoptimal process parameters value of the ANN model to define the boundaries for the minimum and maximum value for optimization solution [
Three conditions could be stated for the nonoptimal process parameters values of the ANN model (OptANN) and optimal process parameters values of the GA (OptGA) as classified in Table
Conditions to define limitation constraint bounds of integrated ANNGA.
Condition  Decision  

Lower bound  Upper bound  
(1) (Opt_{ANN}) < (Opt_{GA})  Opt_{ANN}  Opt_{GA} 
(2) (Opt_{ANN}) > (Opt_{GA})  Opt_{GA}  Opt_{ANN} 
(3) (Opt_{ANN}) = (Opt_{GA})  Nearest lower bound of the coded value of experimental design  Nearest upper bound of 
By using the conditions stated in Table
As per Figure
By using the objective function formulated in (
Fitness value of ANNGA TypeA.
It can be seen that the set values of optimal process parameters that lead to the minimum dilution
Similar to integrated ANNGA type A approach, the objective function formulated in (
The results of the integrated ANNGA type B by using MATLAB Optimization Toolbox fitness function are shown in Figure
Fitness value of ANNGA TypeB.
From Figure
Consider
Theoretically, to validate the results of the proposed approach of this study, the optimal process parameters values of the integrated ANNGA will be transferred into the regression model equation. Equation (
In this study, discussion is carried out to highlight all the objectives of the study and is separated into three parts which are evaluation of the minimum dilution value, evaluation of the optimal process parameters, and evaluation of the number of iteration of the integrated ANNGA results.
Figures
Optimum dilution should be between 8 and 15% from a previous study. This objective is fulfilled in this case.
The minimum predicted dilution value of the regression model is 9.875. Therefore, with dilution is 9.7467, It can be concluded that both integration systems have given the more minimum result of the dilution compared to regression model. Consequently, integrated ANNGA type A and integrated ANNGA type B have reduced the value of dilution and optimized it between 8 to 15.
The minimum predicted Dilution value of the ANN model is 9.353; it can be concluded that both integration systems have given the more optimum result of the dilution compared to ANN model. Consequently, integrated ANNSA type A and integrated ANNGA type A are well within the limits of standard dilution.
The minimum predicted dilution value of the GA was 10.375. Therefore, with dilution 9.745, it can be concluded that both integrated systems have given the more minimum of the Dilution compared to GA technique.
For the second objective of this study, the optimal values of the integrated ANNGA type A and integrated ANNSA type B for each process parameter are within the range of minimum and maximum values of experimental design; thus, this study concludes that the second objective of this study is fulfilled.
Integrated ANNGA type A and integrated ANNGA type B are approximately the same as or lower than the number of iterations by GA; thus, this study concludes that the third objective of this study is fulfilled.
A fivelevelfivefactor full factorial design matrix based on central composite routable design technique was used for building the mathematical model.
ANN tool box available in MATLAB was used for the prediction of bead geometry.
In this study two models, ANN and GA are used for prediction and optimization of bead geometry.
In GMAW process bead geometry plays an important role in economising the material. This study effectively used the integration method for optimization.
This study proposed two integration systems, integrated ANNGA type A and integrated ANNGA type B, in order to estimate the optimal solutions of process parameters that lead to minimum dilution found to satisfy three conditions. Overall, integrated ANNGA type A and integrated ANNGA type B have been very effective in estimating optimum values of dilution compared to experimental data and regression method. In the second process the optimal value of process parameters recommended by integrated type A and integrated type B are well within the range of minimum and maximum value of process parameters of experimental design. In the third issue it was found that the proposed integration system satisfies the number of iterations than the single system of optimization.