Weighted time series fault diagnosis based on a stacked sparse autoencoder

Most statistical analysis technologies use detection thresholds for fault diagnosis, which often cannot effectively characterize some specific faults in a statistical manner. However, the details and small changes in the faults can be exploited by deep learning-based feature representation. In this paper, we present a weighted time series fault diagnosis method to learn the deep correlations of faults and reduce the loss of fault information. Our model includes 2 key novel properties: (1) It can learn high-level abstract features of faults and the underlying fault patterns, which is particularly efficient for detecting incipient faults; (2) a mathematical framework of stacked sparse autoencoder-based fault diagnosis, with capabilities of multiple nonlinear mapping and complex function approximation, is presented. The monitoring performance was compared with multivariate statistical methods and conventional artificial intelligence methods on the Tennessee Eastman process data set, which is a well-known chemical industrial benchmark. The experimental results showed its performance gain over existing methods, especially for incipient faults that are difficult to detect with traditional technologies. As the details and small changes in the faults can be exploited by deep learning-based feature representation, we present a weighted time series fault diagnosis method. Our model includes 2 key novel properties: (1) It can learn high-level abstract features of faults and the underlying fault patterns, which is particularly efficient for detecting incipient faults; (2) a mathematical framework of stacked sparse auto-encoder (SSAE) -based fault diagnosis, with capabilities of multiple nonlinear mapping and complex function approximation, is presented.