小样本下一阶元学习在滚动轴承故障诊断中的应用OA

To address the issue of model-agnostic meta-learning networks incurring overfitting due to high model complexity during training in small-sample conditions,we propose a first-order meta-learning-based rolling bearing fault diagnosis model.Initially,we form meta-tasks by randomly sampling original signals following a meta-learning strategy.Subsequently,within meta-tasks associated with known working conditions,we employ a wide-kernel convolutional network to enhance the model's capability to extract fault information from one-dimensional vibration signals,acquiring meta-knowledge.Additionally,to reduce model training complexity and mitigate overfitting,we employ a gradient-based model optimization,progressively adjusting from initial parameters to optimal training weights.Finally,leveraging the acquired meta-knowledge,our approach can realize rapid and accurate fault classification under unknown working conditions.The experimental results show that the proposed method achieves the highest diagnostic accuracy on both validations bearing datasets,proving the effectiveness and superiority of the proposed method.