Fault Classification of Industrial Processes based on Generalized Zero-Shot Learning

Abstract

In the process industry, the supervised learning methods cannot classify the unseen faults (i.e., those faults without training samples to participate in the establishment of the model). Although Zero-Shot Learning (ZSL) has been proposed and successfully solved the problem of unseen fault classification, it failed to classify the seen faults (i.e., those faults participate in the establishment of the model). To overcome their shortcomings, in this paper, a generalized Zero-Shot Learning (GZSL) method is proposed to classify all the faults including the seen and the unseen faults by only using the samples of the seen fault and the human-defined fault semantic attribute description information. We use a gating mechanism based on Conditional Variational Autoencoder (CVAE) and a binary classifier to distinguish the online sample into the classes of the seen and unseen faults. Thus, the GZSL problem can be transformed into a supervised fault classification problem and a ZSL fault classification problem. Firstly, we train a CVAE to generate pseudo unseen fault samples and seen fault samples. Secondly, a binary classifier is trained to classify the online samples into seen and unseen categories. Finally, the specific category of the online samples will be determined by the supervised method and ZSL method, respectively. We validate our approach on the Tennessee-Eastman benchmark process.