Study of weld characteristics for repair using sequential experimental design and artificial neural networks

For remanufacturing dies and moulds, it is of significance to obtain information about weld characteristics because it plays a critical role on the mechanical strength of welded area. In this paper, a novel design of experiment methodology to study the relationship between the process parameters and weld profiles utilizing sequential experiment design and artificial neural networks is proposed. For the sequential experiment design, Taguchi experimental arrays were used for the preliminary experiments to survey the main effects of process parameters on weld characteristics of laser surfacing and the overall optimal conditions were obtained for all the performances. By means of Uniform Design, a reduced number of experimental runs are employed for further regression modelling of the influence of the selected parameters on the weld characteristics. Mathematical models were established by approximating designed radial basis function neural networks (RBFNNs) which are capable of modelling any function. Validation runs were finally conducted for testing the generalisation of fitted radial basis function neural network models. The prediction errors of validation runs are less than 10 %, which indicates good generalisation of developed models.