Dual-noise autoencoder combining pseudo-labels and consistency regularization for process fault classification

Abstract

In the classification of industrial process faults, the collected process fault data has the problem of having more irrelevant fault information, limited labels, and a significant impact of noise, which affects the prediction accuracy of the classification model. To address these problems, this paper proposes a semi-supervised dual-noise autoencoder method that integrates pseudo-labels and consistency regularization (PR-SNAE). Based on normal samples, the differences between faulty samples and normal samples are enhanced through dissimilarity analysis. Two types of noise are introduced into the enhanced samples to improve the robustness of the model. A stacked supervised autoencoder (SSAE) network is trained using a small amount of labelled data. The deep feature information is extracted to establish a preliminary fault classification model. Pseudo-labels are generated for unlabelled samples to overcome the problem of insufficient labels for fault data. In the adjustment stage of the classification model, a loss function that integrates pseudo-labels and consistency regularization is proposed to prevent overfitting and poor robustness of the model. Simulation experiments were conducted on the Tennessee Eastman (TE) benchmark process and three-phase flow process, and the results verified the effectiveness of the proposed method.