Relevance vector machine-based defect modelling and optimisation -- an application (Q1758850)

Summary: In presence of many correlated and autocorrelated process variables, initially the support vector machine (SVM) and later the relevance vector machine (RVM) were used for modelling the bonding defect in Hi-Cr rolls as function of explanatory variables by mapping the original input data space to high-dimensional feature space using appropriate kernels. The RVM-Bessel kernel, which turned out to be the best-fit regression model with minimum error (MSE) from among the competing kernels, was developed when the best-fit SVM-RBF kernel regression model was found associated with high absolute value of MSE and a large number of support vectors. The final sparse defect model was developed with the relevance vectors (RVs) generated while fitting the RVM-Bessel kernel model by taking recourse to hierarchical regression. Constrained optimisation treatment of the sparse defect model helped identifying the factor-setting corresponding to minimum length (0) of bonding defect. Confirmatory trial runs showed encouraging trends.