A gear fault diagnosis method based on neural network architecture search with cross-modal dynamic convolution enhancement
Multimodal fault diagnosis has attracted widespread attention in the field of rotating machinery due to the complementary information provided by multiple sensor signals. However, the design of signal fusion networks and the extraction of correlation features between multiple signals remain challenging. To address this, this paper proposes a multimodal framework based on differentiable architecture search to fuse vibration and acoustic emission signals, two signals of vastly different orders of magnitude, and to achieve automatic search for the fusion network. Simultaneously, a cross-modal dynamic convolutional module is introduced to extract and fuse correlation features between multiple signals. Experimental results on two benchmark gear fault datasets demonstrate that, compared to traditional models and existing network search methods, the proposed method exhibits stronger robustness under different operating conditions and achieves superior diagnostic performance. Ablation experiments further validate the effectiveness of the proposed interaction mechanism and joint optimization strategy.