Identification of the key manufacturing parameters impacting the prediction accuracy of support vector machine (SVM) model for quality assessment

In the context of manufacturing, support vector machines (SVM) are commonly used to predict quality, i.e., predict the characteristics of a product according to the manufacturing parameters. The prediction accuracy of a SVM model is affected by a number of factors: training set size, data set quality, etc. Manufacturing datasets are usually prone to measurement uncertainties. Such uncertainties affect the observed values of the manufacturing parameters, thereby affecting the predictive performance of the SVM. To address this issue, several works in the literature have been proposed to improve the robustness of SVM to measurement uncertainties. These works, however, do not evaluate the contribution of the uncertainties of each parameter to the overall impact. For this reason, this paper focuses on quantifying the impact of the uncertainties of each parameter on the accuracy of the SVM prediction. Three approaches are proposed to do so. The first two approaches are based on Monte-Carlo simulation and allow providing quantitative measures that represent the impact of the uncertainties of each manufacturing parameter on the accuracy of the SVM. On the other hand, the third approach relies on simple statistical tools in order to estimate the impact of the uncertainties of each parameter. The proposed approaches would eventually make it possible to identify the uncertainties of the parameters that mostly affect the SVM. Such parameters are referred to as key measurement uncertainties. Identifying the key measurement uncertainties would provide a better understanding of how the SVM is affected by uncertainties, as it would provide a strong basis for improving the robustness of SVM in future works. The proposed approaches are applied to four datasets, and their performances are discussed and compared.