Centroid-concentration contrastive learning for fault detection and classification in Industry 4.0

Purpose Owing to the great advances in sensor and cybernetic technologies, fault detection and classification (FDC) has become a crucial application in Industry 4.0. Recently, there has been growing attention to contrastive learning techniques to transform raw data into discriminative feature representations. The purpose of this study was to develop a novel contrastive learning model to enhance the FDC tasks for maintaining qualified performance in manufacturing processes. Design/methodology/approach In this study, a Centroid-Concentration Contrastive Learning (C2CL) framework is proposed to tackle several obstacles to representation learning in FDC applications. With the introduced centroid-orientation concept, the encoder trained by the C2CL framework can generate more discriminative feature representations with inter-class distraction and intra-class concentration. Findings The C2CL architecture employs a transformer-based encoder to effectively capture the relationships and dependencies among different parameters in long manufacturing processes. With the proposed variation-enhancement strategy, C2CL could effectively tackle the similar class-pattern issue. We integrated the centroid effect to construct the contrastive learning model. In the embedding space, we maximized the distances among centroids of classes to clearly distinguish each class, while minimizing the distances of all samples to their class centroids to concentrate the embedding vectors of all instances in the same class. Originality/value Based on the centroid concept, C2CL can derive more discriminative embeddings and enhance the classification or detection in downstream FDC tasks. Several experiments were conducted on real industrial datasets to demonstrate the superiority of C2CL over the state-of-the-art methods in terms of various evaluation metrics. We also use a case study to show the practicability of the proposed framework for fault classification in smart manufacturing.