Class-incremental fault diagnosis for imbalanced and long-tailed industrial data based on supervised contrastive learning

In modern industrial systems, equipment failures can lead to significant safety hazards and substantial economic losses, hence fault diagnosis technologies are of critical importance, with data-driven methods being a prominent diagnostic approach. However, practical applications are plagued by class imbalance and long-tailed distributions in fault data, while the continual emergence of novel fault types (class incrementality) challenges the generalization of existing models. To address these challenges, this paper proposes a class-incremental fault diagnosis (KCLIFD) method, Kolmogorov–Arnold network (KAN)-based supervised contrastive learning with clustering and mean-squared-error (MSE) distillation for KCLIFD. The approach employs a four-layer KAN to extract discriminative features, leverages supervised contrastive loss to enhance feature separability, and incorporates MSE distillation to mitigate catastrophic forgetting. A clustering-based memory buffer is constructed to retain representative exemplars, and a time series forest classifier is introduced to re-balance class weights. Experimental results on Tennessee Eastman Process and Multiphase Flow Facility datasets demonstrate that KCLIFD attains superior performance under both imbalanced and long-tailed conditions, effectively resolving the difficulties arising from class incrementality, imbalance, and long-tailed distributions.