Prototype Recalibration-Driven Incremental Learning Framework for Bearing Fault Diagnosis Without Exemplars

Abstract

As key components of rotating machinery, bearings continually encounter new fault classes in industrial applications. However, the existing deep learning-based diagnosis methods are limited by the requirement that all fault classes must be known during training. This constraint affects the applicability of the intelligent fault diagnosis (IFD) models in real-world scenarios. Incremental fault diagnosis methods can effectively enable models to accumulate knowledge for fault classes, but most require retaining training examples of previously learned fault classes as exemplars for new training tasks. This article proposes a prototype recalibration-driven incremental learning (PRIL) framework to address the challenge of continuous fault diagnosis for bearings without exemplars. Specifically, a feature prototypes recalibration mechanism is employed to align feature prototypes with the feature spaces, allowing features generated from feature prototypes of old fault classes adaptable to the latest training tasks. Additionally, a contrastive learning-based pretrained feature extractor is utilized to enhance the generalization ability of model in continuous fault diagnosis tasks. A distance-based incremental prototype classifier is designed to enable the model balance knowledge between different fault classes. Finally, a case study on continuous fault diagnosis is conducted to assess the effectiveness of the proposed method. The results demonstrate that PRIL can continually accumulate fault diagnosis knowledge while effectively mitigating catastrophic forgetting without retaining historical training data.