基于元学习的小样本跨域自适应方法在液压泵故障诊断中的应用OA

To address the problems of scarce fault samples and insufficient cross-condition generalization in hydraulic pump fault diagnosis,we propose a meta-learning-based few-shot cross-domain adaptation method.First,different pressure-flow combinations are defined as operating-condition domains.Then,a multi-scale one-dimensional convolutional neural network is employed as the feature extractor to learn highly discriminative multi-scale features directly from raw vibration signals.During meta-training,the bi-level optimization mechanism of model-agnostic meta-learning enables the model to rapidly adapt to unseen operating conditions using only a few support samples.To further improve robustness,a dual-objective loss function is constructed by combining a label-smoothed classification loss with an L2 regularization term,thereby alleviating overfitting and enhancing cross-domain stability.Experiments are conducted on two unseen target operating-condition domains with multi-episodic evaluation.After five adaptation steps,the proposed method achieves average accuracies of 98.90%and 96.80%,respectively.Compared with existing methods,the proposed method reduces dependence on training samples while maintaining high diagnostic accuracy,demonstrating its effectiveness and practical potential in few-shot cross-condition fault classification.