A hybrid strategy-based meta-learning for the gear fault diagnosis under complex working conditions

Abstract

It is difficult to ensure consistent working conditions and sufficient labelled samples in engineering practice. To address the challenge of gear fault diagnosis under such complex conditions, this paper proposes a hybrid strategy-based meta-learning model designed for few-shot and variable working condition scenarios. The proposed method comprises a multi-scale feature extractor, a fault classifier, and a meta-metric module. The feature extractor integrates dilated convolutions and a multi-head attention mechanism to effectively capture discriminative fault features across varying receptive fields and enhance feature selection. A hybrid training strategy combining supervised learning and metric-based meta-learning is employed, where in feature representations are first learned in the source domain, followed by rapid adaptation to new tasks via similarity-based matching in the meta-learning phase. This enables the model to achieve both discriminative power and cross-domain generalization with limited labelled data. Experimental validation is conducted on the public DDS gearbox dataset and a self-developed test platform. Results demonstrate that the proposed method consistently outperforms several state-of-the-art baselines in terms of accuracy and robustness, highlighting its strong potential for real-world deployment in industrial fault diagnosis under variable working conditions.