Time-frequency hybrid multilevel feature alignment domain adaptation network for cross-condition fault diagnosis of rolling bearings

Abstract

Traditional fault diagnosis methods suffer from performance degradation when applied under varying operating conditions. This presents a critical challenge to the analysis of industrial systems, where distribution shifts in sensor data severely degrade diagnostic reliability. Domain adaptation approaches have shown promise in cross-condition scenarios, but current methods face two fundamental limitations: (1) insufficient feature hierarchy for capturing multiscale fault signatures and (2) incomplete domain alignment, neglecting conditional distribution discrepancies. This paper proposes a Time-Frequency Hybrid Multilevel Feature Alignment Domain Adaptation Network (TF-MFA-DA) that integrates three coordinated design components: (a) a time-frequency hybrid feature extractor, in which the time-domain branch uses multiscale convolution and Transformer encoding to capture local impulsive patterns and global temporal dependencies, while the frequency-domain branch extracts spectral-statistical information; (b) a dual-alignment mechanism enforcing both marginal and class-conditional distribution matching through global and conditional domain discriminators; and (c) a weighted maximum mean discrepancy (WMMD) based adaptive feature weighting module for enhanced cross-domain transferability. Extensive experiments on two publicly available datasets demonstrate that the proposed framework achieves competitive diagnostic performance and improved stability under diverse operating conditions.