An improved dual-channel CNN-BILSTM fusion attention model for fault diagnosis of aero-engine bearings

Accurate fault diagnosis of aero-engine bearings is vital for ensuring flight safety. Existing methods still struggle with extracted features lacking multi-dimensional representation, insufficient fault information, and ineffective feature fusion under complex conditions (e.g., varying rotational speeds) and multi-source signal inputs. As such, an improved two-channel fault diagnosis model for rolling bearings is proposed, integrating a convolutional neural network and bidirectional long short-term memory (CNN–BILSTM) architecture, enhanced by multiple improved attention mechanisms.First, the raw vibration signals were directly used as time-domain inputs and processed to obtain their frequency-domain counterparts, forming a dual-channel input to the customized and optimized CNN-BILSTM feature extraction network. Then, a one-dimensional convolutional block attention module (1DECBAM) is inserted after each of the two CNNs to retain initial features while enhancing key ones critical for fault diagnosis. Moreover, the proposed Hybrid Interaction-Fusion Attention (HIFAttn) framework incorporates a Time-Frequency Interactive Attention Mechanism (T-FIAttn) and a Local-Global Adaptive Attention Module (L-GAAM) to perform multimodal feature fusion. Specifically, the T-FIAttn is employed to capture latent feature relationships across both time and frequency domains. In addition, the L-GAAM was appended after the BILSTM layers in each channel to dynamically capture essential features. Experimental results on two aero-engine datasets demonstrate that the proposed model achieves accuracies of 99.32% and 99.94%, respectively, surpassing current state-of-the-art methods.The model also demonstrates excellent stability and robustness, even under high-noise conditions.These results indicate that the proposed model achieves high accuracy and strong generalization, making it well-suited for aero-engine bearing fault diagnosis.