Meta-Class-Incremental Training Strategy for Enhancing Lifelong Intelligent Diagnosis Performance Under Varying Operating Conditions

Class-incremental learning methods can improve diagnostic performance in scenarios where new fault modes emerge continuously throughout the lifecycle of machinery. However, challenges such as difficulties in transfer diagnosis and limited model generalization still hinder the effective application of class-incremental learning. To address these challenges, this study presents an innovative meta-class-incremental training strategy aimed at improving lifelong intelligent diagnosis in class-incremental models under varying operating conditions. The proposed meta-class-incremental training strategy is developed through incorporating an improved meta-learning method into the class-incremental learning framework. This strategy employs both meta-training and fast adaptation techniques to significantly improve the model generalization ability in class-incremental transfer diagnosis scenarios. The proposed meta-class-incremental training strategy for lifelong class-incremental transfer diagnosis was verified upon a planetary gearbox dataset. Extensive experiment results have shown the advantageous for the proposed meta-class-incremental training strategy in different class-incremental transfer diagnosis scenarios, solidly outperforming several class-incremental learning methods.