Robust Supervised Contrastive Learning for Fault Diagnosis Under Different Noises and Conditions

Abstract

Fault diagnosis is of vital importance to maintain safety and reliability of mechanical equipment. Intelligent diagnostic methods have achieved high performance in recent researches. However, in industrial application, machines will suffer complex noises and the operating condition is varying as well, which leads to domain shift and performance degradation. As a promising alternative to supervised learning, self-supervised contrastive learning follows a contrastive paradigm to extract robust feature representation. Nevertheless, the training stage of self-supervised learning suffers from the lack of label information, thus its classification accuracy is inferior to supervised approaches. To address this problem, a supervised contrastive learning method, which incorporates the merits of supervised learning and self-supervised learning, is proposed for fault diagnosis. First, dataset is splitted and two views are generated from each original sample via four data augmentation strategies. Then label information is integrated with contrastive loss function by treating views of the same class as positive pairs. Eventually, generalized Gaussian distribution is adopted to develop the noise model. Experiments of multi-noise as well as multi-working condition are implemented. Experimental results demonstrate that the proposed method outperforms other supervised and self-supervised approaches in fault diagnosis of aero-engine bevel gear.