Nonlinear predictable feature learning with explanatory reasoning for complicated industrial system fault diagnosis

As an emerging temporal data analytical tool, graph-based predictable feature analysis (GPFA) has been successfully introduced to perform dynamic system monitoring. Nevertheless, two critical problems still need to be solved: extracting nonlinear predictable features (PFs) efficiently and explaining fault propagation path. In order to handle these two problems, this article proposes a nonlinear predictable feature learning method with explanatory reasoning for complicated industrial fault diagnosis. Unlike traditional kernel-based nonlinear machine learning methods, this approach utilizes random Fourier mapping to build the nonlinear GPFA model. This kind of nonlinear GPFA model has a more straightforward structure and achieves higher computation efficiency. Furthermore, for the sensitivity in fault detection, the probabilistic non-linear PFs are designed using Wasserstein distance (WD) to measure the difference between two probability distributions. To deal with the challenging issue of fault explanatory reasoning, a hierarchical cause-propagation-result (CPR) diagram is constructed to trace the paths of fault propagation and indicate the possible root causes. Finally, the validity of the methodology is demonstrated by simulation experiments of two systems with a continuous stirred tank reactor (CSTR) and the Tennessee Eastman (TE) process.