A Multi-modal Fusion Fault Diagnosis Method Based on RepLKNET and BiGRU-Global Attention

In response to the issues of insufficient feature extraction ability and poor noise resistance and limited adaptability in traditional gearbox fault diagnosis methods under strong noise environments, this paper proposes a new multi-modal fusion method based on the Reparameterized Local Kernel Networks (RepLKNET) and Bidirectional Gated Recurrent Unit (BiGRU) with Global Attention (GATT). First, to solve the problem of insufficient feature expression ability of traditional models in noisy environments, BiGRU is used to extract features from time-series signals, and Global Attention is introduced to enhance BiGRU's focus on key time-step features. Meanwhile, RepLKNet is employed to enhance feature representation ability in image processing. Finally, to improve the model's noise resistance and adaptability, a multi-modal feature fusion method based on both models is proposed, which fuses image samples with original time-series signals, extracting information from both the time sequence and spatial structure perspectives. The effectiveness of the model is verified on a gearbox test bench dataset. The results show that the proposed model maintains an accuracy of over 95% even under noisy conditions and varying rotational speeds, demonstrating its strong feature extraction ability, noise resistance and adaptability.