A classification-driven neuron-grouped SAE for feature representation and its application to fault classification in chemical processes

Abstract

Recently, deep learning has become a popular tool for fault detection and diagnosis in chemical processes to learn complex nonlinear features. However, the features extracted from most traditional deep networks are only good representation for the raw input data, which may contain irrelevant information with the faults and are not beneficial for the fault classification task. Thus, a novel classification-driven neuron-grouped stacked autoencoder (CG-SAE) is proposed for hierarchical fault-related feature representation in this paper. During the pre-training phase of CG-SAE, the hidden neurons in each CG-AE are divided into several groups, the number of which is equal to that of the data categories. Each group is mainly associated to one type of fault. Moreover, the supervised information is introduced without additional weights by comparing the error between the actual label and the average activation of the corresponding group. In this way, the extracted deep features can contain category-related information and be clustered to different categories. In terms of the classification accuracy and training efficiency, CG-SAE performs better than some advanced supervised methods as verified by the Tennessee Eastman (TE) benchmark and an industrial hydrocracking process.