An Incremental Learning Framework for Mechanical Fault Diagnosis With Bi-Level Multiscale Convolutional Attention

Fault diagnosis for rotating machinery is important for optimizing productivity and enhancing safety. However, in practical engineering, data-driven methods are not only challenged by the problem of insufficient fault data, but also often cannot achieve continuous learning and online diagnosis of newly emerging fault types in constantly changing operating environments. To address these issues, this article proposed a continuous few-shot incremental learning method based on bi-level multiscale convolutional attention (BMCA) mechanism. First, infrared thermal imaging data are used as input signals, and a variational encoder (VAE) synthesis replay bank is constructed to automatically replenish and retain the most representative samples for relearning. Next, a cross-channel dynamic spatial (CCDS) convolutional attention mechanism is proposed to achieve a dynamic allocation of attention weights in both channel and spatial feature dimensions. Finally, the update scale of the model is constrained by the designed focus-knowledge distillation (FKD) loss function, and the weights of the small samples as well as the loss contribution of the hard-to-categorize samples are dynamically adjusted. The experimental results of bearing data based on infrared thermal imaging show that the diagnostic accuracy of this method can still reach 96.68% under the condition of small samples, and the incremental learning strategy effectively alleviates the negative effects of catastrophic forgetting and insufficient samples.