Prior knowledge-driven meta-learning for fault detection under varying working conditions

This paper proposes a novel prior knowledge-driven meta-learning (PKML) fault detection approach to address practical scenarios involving scarce fault samples under varying working conditions. Specifically, the PKML method is built upon a model-agnostic meta-learning framework, where meta-task samples from diverse working conditions are mapped into embedding vectors to obtain a unified representation of multi-condition features. To enhance the discriminability of the embedding representations, a learnable sample-attention mechanism is incorporated to adjust the relative contributions of support samples during condition embedding construction. Furthermore, a wide-and-deep convolutional neural network is employed as the base classifier for meta-learning, with embedded conditional layers to enhance adaptability to varying working conditions. Through a feature-wise linear modulation mechanism, the classifier parameters are dynamically adjusted according to condition-aware structural information, enabling effective adaptation and generalization to unseen working conditions. Finally, experimental results on three representative bearing fault datasets demonstrate that the proposed PKML method consistently outperforms existing approaches under varying working conditions and few-shot scenarios.