Counterfactual-augmented few-shot contrastive learning for machinery intelligent fault diagnosis with limited samples

Capturing sufficient and balanced data for intelligent fault diagnosis is significantly consumptive in practice. It is tricky and demand-oriented to identify faults accurately and reliably with limited samples. For this issue, contrastive learning is a promising attempt by learning discriminative representations (DRs). It is a research gap for contrast learning to apply in limited and imbalanced samples due to negative sample pairs, contrastive loss, and mechanistic interpretations. In this study, counterfactual-augmented few-shot contrastive learning (CAFCL) is proposed for intelligent fault diagnosis with limited samples. Firstly, a feature weight network is designed to exploit optimal features (OFs) with sparsity. Moreover, OFs are reliably unique with the assistance of weight updaters and feature sparsifiers. Next, counterfactual augmentation with OFs is defined for negative samples, whose plausibility is discussed in terms of causation. Thirdly, few-shot contrastive learning (FCL) is customized for intelligent fault diagnosis with limited samples, which reconciles global associations and local diffusions. Furthermore, model decisions are indicated by FCL via DRs, which are deemed to coincide with the fault mechanism of mechanical components. The feasibility and effectiveness of CAFCL are verified by various experiments on fault diagnosis. Results show that CAFCL is superior in few-shot intelligent fault diagnosis with promising engineering applications.