ALTFN: a novel unsupervised domain adaptation bearing fault diagnosis framework using joint global and local information

Abstract

This paper proposes an Adaptive Lightweight Time-Frequency Network (ALTFN) to address insufficient fault feature representation and degraded generalization performance under varying working conditions. ALTFN integrates Adaptive Time-Frequency Transformer (ATFT) and Depthwise Separable Convolution (DSC) into a dual-branch architecture to achieve joint global-local representation learning. The ATFT branch uses short-time Fourier transform (STFT)-based spectral feature embedding and a lightweight linear self-attention mechanism to capture global time-frequency dependencies. Further, the contrast alignment constraint is introduced to learn the domain invariant representation across working conditions. The DSC branch extracts local discriminative features and incorporates feature decorrelation regularization to promote complementary representation learning. Finally, the self-paced bidirectional collaborative distillation mechanism is employed to optimize model performance, and confidence-weighted fusion is applied to improve the discrimination of hard-to-classify samples. Experimental results on JNU and HUST bearing datasets show that ALTFN has an average accuracy of more than 99% in cross-condition diagnosis tasks, indicating excellent robustness and generalization ability.