A decoupling method for compound faults of rolling bearings based on multi-modal bidirectional cross-attention fusion and contrastive learning

Abstract In compound fault diagnosis of rolling bearings, traditional single-modal analysis and existing multimodal fusion methods suffer from issues including insufficient cross-modal interaction, limited adaptive weighting capability, and poor discriminability in the feature space, which leads to ambiguous decoupling of compound fault features. Which leads to ambiguous decoupling of compound fault features. For this purpose, this paper proposes a decoupling method for compound faults of rolling bearings based on multimodal bidirectional cross-attention fusion and contrastive learning. First, an autoencoder extracts temporal features from vibration signals, while adaptive rectangular convolution dynamically captures multi-scale spatial features of time-frequency images. Secondly, a bidirectional cross-attention mechanism is constructed to achieve cross-modal semantic alignment through signal-to-image and image-to-signal dual guidance. Furthermore, a dynamic gating mechanism is introduced, which generates adaptive fusion weights based on the characteristics of the samples. It dynamically adjusts contribution ratio of each modality to avoid information masking. Finally, orthogonal subspace projection maps multimodal features into independent fault subspaces. When combined with contrastive learning, this approach enhances both intra-class compactness and inter-class separability, forming clear decoupling boundaries for precise fault decoupling. Experiments on PU and lab datasets show that the proposed method achieves average decoupling accuracies of 99.36% and 98.76% under various conditions. Ablation studies further confirm the synergistic effectiveness of bidirectional cross-attention, dynamic gating, and contrastive learning, offering an industrially viable approach for decoupling compound faults in rolling bearings.