Class Incremental Fault Diagnosis under Limited Fault Data via Supervised Contrastive Knowledge Distillation

TL;DR

The paper tackles class-incremental fault diagnosis under limited fault data by introducing SCLIFD, a framework that fuses supervised contrastive knowledge distillation with cross-session feature transfer, a Marginal Exemplar Selection memory replay strategy, and a Balanced Random Forest classifier to address severe class imbalance. The method updates feature representations across sessions while preserving old knowledge, selectively replays informative boundary exemplars, and counters normal-class bias during classification. Comprehensive experiments on TE P and MFF datasets under imbalanced and long-tailed regimes show SCLIFD achieving superior accuracy and reduced forgetting, with ablations confirming the efficacy of each component. The results suggest practical impact for real-world industrial fault diagnosis where data are scarce and class distributions are uneven, and point to future work on adaptive memory sizing per class.

Abstract

Class-incremental fault diagnosis requires a model to adapt to new fault classes while retaining previous knowledge. However, limited research exists for imbalanced and long-tailed data. Extracting discriminative features from few-shot fault data is challenging, and adding new fault classes often demands costly model retraining. Moreover, incremental training of existing methods risks catastrophic forgetting, and severe class imbalance can bias the model's decisions toward normal classes. To tackle these issues, we introduce a Supervised Contrastive knowledge distiLlation for class Incremental Fault Diagnosis (SCLIFD) framework proposing supervised contrastive knowledge distillation for improved representation learning capability and less forgetting, a novel prioritized exemplar selection method for sample replay to alleviate catastrophic forgetting, and the Random Forest Classifier to address the class imbalance. Extensive experimentation on simulated and real-world industrial datasets across various imbalance ratios demonstrates the superiority of SCLIFD over existing approaches. Our code can be found at https://github.com/Zhang-Henry/SCLIFD_TII.

Figures (8)

• Figure 1: Class incremental learning for fault diagnosis under limited fault data.
• Figure 2: iCaRL's T-SNE visualization under the imbalanced case of TEP dataset. iCaRL cannot extract discriminative features from imbalanced classes.
• Figure 3: T-SNE visualization under the long-tailed case of MFF dataset using "Herding" icarlwelling2009herding. Misclassified samples tend to be scattered across the model's decision boundaries.
• Figure 4: The general framework of the proposed method SCLIFD for fault diagnosis under limited fault data, which is detailed explained in \ref{['overview']}.
• Figure 5: Experiment result comparisons of different methods on TEP and MFF Dataset.