Abstract:
To address the limited representation capability of individual features and the insufficient interaction of complementary information between heterogeneous features in rolling bearing fault diagnosis under variable-speed conditions, a fault diagnosis method based on a position-aware bidirectional cross-attention network (CGBCA) was proposed. First, a compact temporal-spatial dual-branch feature extraction module was constructed to extract features from one-dimensional temporal signals and two-dimensional time-frequency representations, respectively. This design controlled the model parameters and reduced the risk of overfitting under few-shot conditions. Second, a position-aware bidirectional cross-attention module was designed. Conditional positional encoding was utilized to model the local neighborhood relationships of tokens, and coordinate biases were incorporated into the similarity computation. This explicitly constrained the attention weight allocation, thereby enhancing the bidirectional interaction and complementary fusion between the two types of features. Finally, the proposed CGBCA model was experimentally evaluated on a variable-speed rolling bearing dataset. The results showed that CGBCA achieved an average diagnostic accuracy of 98.40% ±0.31%, which is 2.26 and 3.38 percentage points higher than those of the CPE-only model and the CNN-Transformer cross-attention model (CTCA), respectively. These results fully demonstrated the effectiveness of the proposed method in the deep fusion of heterogeneous features and cross-feature interaction, providing reliable technical support for rolling bearing fault diagnosis under complex variable-speed operating conditions.