A new class of fault detection and diagnosis methods by fusion of spatially distributed and time-dependent features

Abstract

Nonlinear, non-Gaussian, and dynamic features pose a great challenge for complex fault detection and fault diagnosis (FDD). Focusing on fault detection, independent component analysis (ICA) and adversarial autoencoder (AAE) are fused to form a new method for nonlinear non-Gaussian latent variable extraction: ICA–AAE. In addition, a strategy for establishing more accurate fault detection thresholds using tail distribution features is presented. Furthermore, a new class of fault diagnosis frameworks to fully exploit the information obtained from normal samples is developed. Fault data are first re-represented using the established ICA–AAE model. Then, the low-dimensional spatial distribution features with their inherited high-dimensional temporal dependencies are synthesized into image information using an image-based approach, and a spatio-temporal fusion fault diagnosis method is implemented using a convolutional neural network (CNN). Tennessee Eastman (TE) process results show that the proposed methods can achieve more accurate fault detection and diagnosis.