Adaptive Working Condition Recognition With Clustering-Based Contrastive Learning for Unsupervised Anomaly Detection

In real industrial processes, machines usually run under variable working conditions, which impose challenges for anomaly detection. To complete anomaly detection for machines under variable working conditions, we develop a reconstruction-based autoencoder called clustering-based contrastive learning autoencoder (CBCL-AE). It integrates clustering-based contrastive learning (CBCL) to perform clustering in the feature space and enhance the differentiation of features from different working conditions, thereby achieving adaptive working condition recognition. Considering the crucial role of the clustering of CBCL, we theoretically and experimentally demonstrate its convergence property during the training process, which directly determines the effectiveness of CBCL-AE. CBCL-AE's superiority has been validated on three public datasets and two private datasets collected from an actual industrial process. These validations highlight its superiority over five state-of-the-art models in unsupervised anomaly detection.