Neural representations for quality-related kernel learning and fault detection

Quality-related modeling and monitoring which aim at the key performance indicators have received wide attention in the research community. The widely used kernel-based methods mainly map process variables into kernel space without considering the relationship between the high-dimension features and quality indicators; therefore, the modeling performance of such transform cannot be guaranteed. For quality-related kernel learning, we propose a framework consisting of flexible neural transform and fixed kernel mapping. In this framework, neural network is used to learn representations for predicting quality indicators in the following kernel regression models. For monitoring the quality-related and quality-independent information, we present a solution for relevant subspaces decomposition and the diagnostic logic is summarized based on the quality-related and quality-independent statistics. The effectiveness of the proposed method is evaluated by simulations and real industrial-scale process.