Envelope cross-attention based adaptive aggregation Transformer for fault diagnosis

Abstract In the field of intelligent fault diagnosis, traditional deep learning methods still suffer from several inherent limitations: the inadequate exploitation of shallow-layer impact features and the progressive dilution of shallow fault-sensitive features during hierarchical network stacking, both of which ultimately degrade the overall diagnostic performance. To this end, a novel intelligent diagnostic model named the envelope cross-attention based adaptive aggregation Transformer (ECAAT) is proposed for bearing fault diagnosis. Specifically, ECAAT utilizes a time-domain Transformer encoder as the backbone and introduces a historical-layer adaptive aggregation mechanism to selectively fuse features from multiple historical layers, improving the capability of inter-layer feature propagation. Meanwhile, an envelope-spectrum cross-attention residual injection mechanism is adopted. This scheme incorporates frequency-domain discriminative features of fault impulse modulation into the time-domain backbone via residuals, enhancing the model’s detection of fault impact characteristics and corresponding modulation patterns. Experiments on three publicly available bearing datasets demonstrate the effectiveness of ECAAT for bearing fault diagnosis under the evaluated protocols. Ablation studies further support the contributions of historical-layer aggregation and envelope-spectrum cross-attention to diagnostic performance.