Dual-mechanism knowledge distillation for imbalanced class-incremental industrial fault diagnosis

Industrial fault diagnosis models used in changing process conditions need to learn new fault classes without losing their ability to recognize previously learned faults. In class-incremental learning, this loss of old-class performance is caused by parameter updates with new-class data and is known as catastrophic forgetting. In industrial diagnosis, this problem is often coupled with class imbalance, because normal samples are abundant whereas fault samples are limited. As a result, incremental updating may not only weaken old-class representations but also bias the model toward majority classes. This article proposes a Dual-Mechanism Knowledge Distillation method for Imbalanced Class-Incremental Industrial Fault Diagnosis (DMKDFD). The framework includes three main parts. First, the Balanced Contrastive Knowledge Distillation (BCKD) module combines supervised contrastive learning and weighted balanced distillation to improve class separability while reducing the drift of old-class features. Second, a Balanced Random Forest (BRF) classifier is trained on the learned features to reduce the decision bias caused by class imbalance. Third, an Adaptive Exemplar Selection (AES) strategy updates the memory buffer by keeping both boundary samples and stable samples for later incremental sessions. Experiments on the Tennessee Eastman Process (TEP) and Multiphase Flow Facility (MFF) datasets show that DMKDFD provides better incremental diagnostic performance under imbalanced class distributions.