A Novel Bearing Fault Diagnosis Method Based on Multifeature Fusion Attention-Guided Mechanism With Noise Robustness

Fault diagnosis (FD) of rolling bearings is crucial for ensuring the reliability and safety of rotating machinery. However, substantial noise interference has caused significant difficulties in improving the robustness of FD. To address this issue, a multifeature fusion attention-guided mechanism with wide first-layer kernels convolutional neural network (MFA-WCNN) is proposed. The position information is introduced into the attention module by feature fusion to catch the relative position or other information of fault fluctuations segment in the whole cycle, which extracts more discriminative fault-related features from noise-contaminated signals. Concretely, a feature extraction module (CFE-Module) is proposed to utilize the different levels of features of rolling bearings, by constructing a convolutional neural network (CNN) with wide first-layer kernels to extract the position information from the low-level features. Furthermore, a feature learning adaptive adjustment module (FLA-Module) is constructed to extract advanced features containing semantic information, simultaneously ignoring the irrelevant information and noise in the fusion features. These two modules allow MFA-WCNN to extract and generate multilevel fusion features with position details and semantic information, which promotes the improvement of FD performance with strong noise. The evaluation experiments conducted on two testing platforms show that the network has excellent FD ability under strong noise and unknown noise.