Gauranga Lal Datta

Prediction of weld bead geometry and penetration in shielded metal-arc welding using artificial neural networks

Abstract Bead geometry (bead height and width) and penetration (depth and area) are important physical characteristics of a weldment. Several welding parameters seem to affect the bead geometry and penetration. It was observed that high arc-travel rate or low arc-power normally produced poor fusion. Higher electrode feed rate produced higher bead width making the bead flatter. Current, voltage and arc-travel rate influence the depth of penetration. The other factors that influence the penetration are heat conductivity, arc-length and arc-force. Longer arc-length produces shallower penetration. Too small arc-length may also give rise to poor penetration, if the arc-power is very low. Use of artificial neural networks to model the shielded metal-arc welding process is explored in this paper. Back-propagation neural networks are used to associate the welding process variables with the features of the bead geometry and penetration. These networks have achieved good agreement with the training data and have yielded satisfactory generalisation. A neural network could be effectively implemented for estimating the weld bead and penetration geometric parameters. The results of these experiments show a small error percentage difference between the estimated and experimental values.

Genetic algorithm for optimization of welding variables for height to width ratio and application of ANN for prediction of bead geometry for TIG welding process

This paper explains an integrated method with a new approach using experimental design matrix of experimental designs technique on the experimental data available from conventional experimentation, application of neural network for predicting the weld bead geometric descriptors and use of genetic algorithm for optimization of process parameters. The properties of the welded joints are affected by a large number of welding parameters. Modeling of weld bead shape is important for predicting the quality of welds. In an attempt to model the welding process for predicting the bead shape parameters (also known as bead geometry parameters) of welded joints, modeling and optimization of bead shape parameters in tungsten inert gas (TIG) welding process has been tried in the present work. Multiple linear regression technique has been used to develop mathematical models for weld bead shape parameters of TIG welding process, considering the effects of main variables as well as two factor interactions. Also by using the same experimental data, an attempt has been made to predict the bead shape parameters using back-propagation neural network. To optimize the process parameters for the desired front height to front width ratio and back height to back width ratio, genetic algorithmic approach has been applied.

Forward and reverse mappings in green sand mould system using neural networks

The quality of castings in a green sand mould is influenced significantly by its properties, such as green compression strength, permeability, mould hardness, and others, which depend on input parameters. The relationships of these properties with the input parameters, like sand grain size and shape, binder, water, etc. are complex in nature. In the neural network based forward mapping, mould properties are expressed as the functions of input parameters, whereas attempts can also be made to determine an appropriate set of input parameters, to ensure a set of desired properties, in reverse mapping. In the present work, the problems related to both the forward as well as reverse mappings in green sand mould system were tackled by using a back-propagation neural network (BPNN) and a genetic-neural network (GA-NN). Batch mode of training had been provided to both the networks with the help of one thousand data, generated artificially from the regression equations obtained earlier by the authors. The performances of the developed models had been compared among themselves for 20 randomly generated test cases. The results show that GA-NN outperforms the BPNN and that both the NN approaches are able to carry out the reverse mapping effectively.

Linear and non-linear statistical modelling of green sand mould system

Abstract In the present work, design of experiments (DOE) technique with response surface methodology was used to develop both linear and non-linear models, to establish the input–output relationships in green sand mould system. Grain fineness number (GFN), amount of clay, amount of water and number of strokes (degree of ramming) were considered as the input variables (parameters), which control the outputs (i.e. mould properties). Full factorial DOE was utilised for the linear model, whereas central composite design and Box-Behnken designs were used to develop the non-linear models. Experiments were conducted to measure the green sand mould properties, such as permeability, green compression strength, mould hardness and bulk density. The adequacy of all the developed models was checked through statistical analysis. Twenty random test cases were considered, to validate the models and compare their performances. A model that is statistically adequate and gives minimum percentage of deviation in prediction was adjudged as the best model for a particular response.

Neural Network-Based Approaches for Forward and Reverse Mappings of Sodium Silicate-Bonded, Carbon Dioxide Gas Hardened Moulding Sand System

Back-propagation neural network and genetic-neural network were developed to predict mould properties of sodium silicate-bonded, carbon dioxide gas hardened moulding sand system from the input process parameters. The performance of back-propagation neural network was found to be comparable with that of the best statistical regression model in predicting the mould properties, whereas genetic-neural network showed large deviations from the target values. Both the said neural networks had been developed as the reverse mapping tools and the performance of genetic-neural network was found to be marginally better than the other. The performances of above neural networks were seen to be dependent on the nature of error surface.

Genetic algorithm-tuned entropy-based fuzzy C-means algorithm for obtaining distinct and compact clusters

A modified approach had been developed in this study by combining two well-known algorithms of clustering, namely fuzzy c-means algorithm and entropy-based algorithm. Fuzzy c-means algorithm is one of the most popular algorithms for fuzzy clustering. It could yield compact clusters but might not be able to generate distinct clusters. On the other hand, entropy-based algorithm could obtain distinct clusters, which might not be compact. However, the clusters need to be both distinct as well as compact. The present paper proposes a modified approach of clustering by combining the above two algorithms. A genetic algorithm was utilized for tuning of all three clustering algorithms separately. The proposed approach was found to yield both distinct as well as compact clusters on two data sets.

Modelling of input–output relationships in cement bonded moulding sand system using neural networks

Abstract Cement bonded sand moulds can be used to cast ferrous metals with a good dimensional control. To determine input–output relationships in the cement bonded moulding sand system, both forward and reverse mappings were carried out using feed forward neural networks trained with the help of a back propagation algorithm and a genetic algorithm, separately. In the forward mapping, mould properties, namely compression strength and hardness, were predicted for different combinations of process parameters, such as percentages of cement, of accelerator and of water and testing time. In the reverse mapping, the process parameters were determined as the functions of mould properties. A batch mode of training had been provided to the neural networks with the help of one thousand training data generated artificially using the conventional statistical regression equations derived earlier by the authors. The performances of the developed models were compared among themselves and with those of the statistical regression model, for twenty randomly generated test cases. Neural network based approaches had proved their ability to carry out both the mappings. In forward mapping, the results of the neural network based approaches were found to be comparable with those of conventional regression analysis. Moreover, the genetic algorithm trained neural network was seen to perform better than the back propagation trained neural network for both the forward and reverse mappings.

Forward and reverse modeling of electron beam welding process using radial basis function neural networks

An attempt has been made in the present study to model input-output relationships of an electron beam welding process in both forward as well as reverse directions using radial basis function neural networks. The performance of this network is dependent on its architecture significantly, which, in turn, depends on the number of hidden neurons, as the number of input nodes and that of output neurons can be decided beforehand for modeling a particular process. Input-output data can be clustered based on their similarity among them. The number of hidden neurons of this network is generally kept equal to that of clusters made by the data-set. Two popular fuzzy clustering algorithms, namely fuzzy C-means and entropy-based fuzzy clustering have been used for grouping the data into some clusters. As both these algorithms have inherent limitations, a modified clustering algorithm has been proposed by eliminating their demerits and combining their advantages. Radial basis function neural network developed using the proposed clustering algorithm is found to perform better than that designed based on the above two well-known clustering algorithms.

Prediction of Weld Bead Profile Using Neural Networks

In fusion welding process, a monolithic structure is obtained at the weldment between the two pieces to be joined. This monolithic structure called the welded zone has got different shapes, which depend upon the type of welding process and its parameters. The strength of the welded structure depends upon the extent to which the metal penetrates between the joint and the volume of the parent metal that gets melted. The shape of the weldment governs mechanical properties of the structure. The weldment shape is generally represented by bead width, bead height and bead penetration. In this study, two neural network-based approaches have been developed to predict the locus of weld fusion zone.

Hybrid optimization scheme for radial basis function neural network

Radial Basis Function Neural Network (RBFNN) is a curve fitting tool in a higher dimensional space. The nature of this surface depends mainly on the number of neurons in the hidden layer. The number of hidden neurons is decided by the number of clusters into which the data-set gets divided. It has been shown that the accuracy in prediction depends upon the quality of the clusters. To obtain good quality clusters, in this study, a hybrid optimization scheme of running a genetic algorithm in the outer loop, while simultaneously running a back-propagation algorithm in the inner loop, has been adopted. The number of hidden neurons is kept the same with that of clusters formed by an algorithm proposed here, apart from the popular fuzzy-c-means and entropy-based clustering algorithms. RBFNN developed using the proposed clustering algorithm is found to perform better than that obtained utilizing the other two clustering algorithms. The method has been successfully implemented in both forward and reverse mappings of electron beam welding process.