Time–frequency modalities fusion-driven channel-adaptive contrastive learning for cross-domain fault diagnosis

While supervised learning has achieved significant success recently, its dependency on massive labeled data and poor generalization across different operational domains limit real-world applications in cross-domain fault diagnosis. Self-supervised contrastive learning offers a promising alternative by leveraging unlabeled data. However, existing methods often suffer from two critical shortcomings when applied to domain-shift scenarios: (1) Isolated feature extraction: they primarily model signals from either the time or frequency domain in isolation, failing to capture the synergistic complementarity between these two intrinsically linked perspectives. (2) Structural bias: even when both domains are used, simplistic fusion strategies and unconstrained optimization often lead the model to overly rely on a single, dominant modality, causing underutilization of the other and limiting the robustness and transferability of learned representations. To address these issues, this article proposes a novel framework named time–frequency attentive fusion with channel-adaptive contrastive learning for cross-domain fault diagnosis, which drives the fusion of time-frequency modalities. This framework employs dual encoders for time domain and frequency domain inputs, while an attention-based fusion module enables deep cross-modal interaction. In addition, channel-importance modeling and a channel-weight regularization strategy are introduced to balance branch contributions and mitigate structural bias toward a single modality, thereby enhancing domain-invariant feature learning. Experiments on two public datasets (CWRU and HUST) demonstrate that the proposed method consistently outperforms self-supervised baselines in diverse cross-domain transfer tasks and under limited label ratios, and even exceeds supervised models in some cross-domain scenarios.