A Stacking Learning Data-Driven Method for Nonlinear KPI-Related Fault Detection

This brief studies the key performance indicator (KPI)-related fault detection problem for the nonlinear system based on the designed stacking learning data-driven (SLDD) method. The purpose of this work is to detect abnormal conditions in the running system early and determine their impact on the KPIs. In contrast to the existing kernel-based methods, the SLDD method creates several primary learners based on a set of kernel parameters, saving computation resources in the training procedure by avoiding the selection of the specific parameter. In addition, a novel KPI-related fault detection framework is constructed by integrating the Bayesian inference with multiple primary learners. In the end, the SLDD algorithm is applied to the incipient fault and sensor fault in the three-tank system to show its effectiveness.