Multi-source weighted domain adaptation guided mechanical cross-domain diagnosis method with cross-layer hybrid attention and multi-objective optimization

Significant domain-shifts between partial source-target domains heavily hinder domain adaptation and degrade transfer fault diagnosis performance. This paper proposes a multi-source weighted domain adaptation (MSWDA) framework for cross-domain diagnosis. Initially, a multi-source domain weighting strategy based on subspace similarity is designed to guide the model to prioritize learning features from high-weight source domains during domain adaptation. Subsequently, a cross-layer hybrid attention network is developed to enhance essential domain-invariant features. Furthermore, a multi-objective collaborative optimization strategy is proposed to comprehensively enhance the cross-domain diagnostic capability. Finally, target-domain transfer diagnosis is achieved using well-trained MSWDA model. Comparative experiments on two mechanical transmission datasets indicate MSWDA attains the highest diagnostic accuracy of 98.48% and 96.08% compared to advanced methods, respectively, verifying its superior capabilities for cross-domain diagnostics.