Decoupling incremental classifier and representation learning based continual learning machinery fault diagnosis framework under long-tailed distribution

Continual learning fault diagnosis (CLFD) has gained growing interest in mechanical systems for its ability to accumulate and transfer knowledge in dynamic fault diagnosis scenarios. However, existing CLFD methods typically assume balanced task distributions, neglecting the long-tailed nature of real-world fault occurrences, where certain faults dominate while others are rare. Due to the long-tailed distribution among different mechanical conditions, excessive attention has been focused on the dominant type, leading to performance degradation in rarer types. In this paper, decoupling incremental classifier and representation learning (DICRL) is proposed to address the dual challenges of catastrophic forgetting introduced by incremental tasks and the bias in long-tailed CLFD (LT-CLFD). The core innovation lies in the structural decoupling of incremental classifier learning and representation learning. An instance-balanced sampling strategy is employed to learn more discriminative deep representations from the exemplars selected by the herding algorithm and new data. Then, the previous classifiers are frozen to prevent damage to representation learning during backward propagation. Cosine normalization classifier with learnable weight scaling is trained using a class-balanced sampling strategy to enhance classification accuracy. Experimental results demonstrate that DICRL outperforms existing continual learning methods across multiple benchmarks, demonstrating superior performance and robustness in both LT-CLFD and conventional CLFD. DICRL effectively tackles both catastrophic forgetting and long-tailed distribution in CLFD, enabling more reliable fault diagnosis in industrial applications.