Contrastive Learning-Based Feature-Consistency Distillation Network for Weak Fault Diagnosis of Harmonic Drive

The deep learning (DL) technology has contributed excellent intelligent diagnosis algorithms and network structures for fault diagnosis of mechanical equipment. However, it requires high-quality fault samples to train models, which significantly challenges the weak fault diagnosis because the early fault characteristics of signals are extremely weak and easily overwhelmed by background noise. To resolve this problem, a weak fault diagnosis method called a contrastive learning-based feature-consistency distillation (CL-FCD) network is proposed, including feature extraction, feature augmentation, and classification modules. First, a multidilation rate dilated convolutional block is developed to extract weak fault features at different scales receptive fields, which can improve the feature representation of weak faults. Second, a feature-consistency distillation loss is designed in the feature augmentation module to align the feature distribution between the weak fault and severe fault by the contrastive learning technology, where the severe fault features are obtained through a convolutional neural network (CNN) trained on severe fault samples. Finally, the classifier pretrained on severe faults is employed to diagnose the weak fault. In this way, weak features can be augmented to achieve the same high discriminatory power as severe faults. The effectiveness and generalization of the proposed method are verified on a rolling bearing dataset and harmonic drive dataset, which outperform existing methods in weak fault diagnosis.