A sensitive efficient multiple predictable feature extraction and fusion method for complicated industrial fault detection

Abstract

In recent years, graph-based predictable feature analysis (GPFA) has emerged as a potential tool in industrial fault diagnosis. However, the basic GPFA can only extract the linear predictable feature information, which is incompetent in the scenarios of the complicated linear-nonlinear hybrid data characteristic. To handle this issue, an improved fault detection method, called Sensitive Efficient Multiple Graph-based Predictable Feature Analysis (SEM-GPFA), is presented for mining the linear and nonlinear predictable features simultaneously. In this method, a multiple predictable feature fusion framework is constructed to utilize the complementary advantages of linear and nonlinear features fully. The linear predictable features are extracted by the basic GPFA, while the nonlinear predictable features are captured by the nonlinear GPFA model. To address the high computational complexity of traditional kernel methods, a random Fourier mapping method is used to improve the nonlinear feature extraction approach, enhancing operational efficiency. Considering that the specific fault information may be concealed in massive features of the model, a fault-sensitive feature highlighting strategy is designed by assigning relatively large weights to emphasize the influence of significant fault features. Finally, case studies on the Continuously Stirred Tank Reactor (CSTR) process and the Tennessee Eastman (TE) chemical system are conducted to demonstrate the superiority of the proposed method.