A multi-level bidirectional cross-attention framework for multi-sensor fault diagnosis of rotating machinery

Abstract Compared to single-sensor inputs, multi-sensor signals can provide more comprehensive operational information for fault diagnosis. However, how to effectively model the interactions among different sensors to achieve efficient utilization of multi-source information remains a key challenge in multi-sensor fault diagnosis. To address this issue, this paper proposes a multi-sensor fault diagnosis framework based on multi-level bidirectional cross-attention. First, to enhance the information correlation among sensors, the framework introduces a bidirectional cross-attention mechanism on top of feature-level fusion, which establishes symmetric dependency relationships between sensors. Second, the proposed framework further models inter-sensor feature interactions from both the feature level and the channel level, enabling the model to perform information exchange across different representation levels, thereby enhancing the expressiveness of the fused features. Finally, experimental results on three fault diagnosis datasets demonstrate that, compared with single-sensor models and various multi-sensor fusion methods, the proposed framework achieves superior diagnostic performance and exhibits stronger robustness and stability. Related ablation experiments further validate the effectiveness of the bidirectional modeling strategy and the multi-level modeling mechanism in enhancing the representation capability of multi-sensor features.