Rolling bearing fault diagnosis under strong noise based on feature fusion cross-multiscale convolutional network

Rolling bearing operating conditions are complex, and the vibration fault characteristics are especially weak in the early stages of fault development. This makes fault diagnosis challenging, as these subtle features are easily masked by strong background noise. To address this issue, a rolling bearing fault diagnosis method under strong noise conditions is proposed, based on a feature fusion cross-multiscale convolutional network. The proposed method initiates with signal-domain feature fusion preprocessing on the raw vibration signal and its autocorrelated signal. This is followed by the incorporation of an attention mechanism to assign adaptive weights. This mechanism effectively suppresses noise interference while enhancing fault features. Subsequently, the feature cross-extraction mechanism is proposed to improve the multi-scale convolutional structure, enhance the feature extraction capability of the multi-scale convolutional network to capture both global and local multi-scale information of the fused features. Finally, the extracted features are fed into a fully connected classification network to achieve rolling bearing fault diagnosis under strong noise. Experimental validation using publicly available rolling bearing datasets and data from dedicated test bench demonstrates that the proposed method exhibits high diagnostic accuracy and robustness under strong noise conditions.