Bearing Fault Diagnosis Using CNN with Multi-scale Residual Adaptive Feature Fusion Module

Rolling bearings are key components in the operation of rotating machinery and are vital to modern industry. However, bearings operate in complex environments, resulting in the vibration signals collected are diverse, noisy and complex. Traditional Convolutional Neural Networks (CNNs) struggle to adapt to varying scales of features, hindering its ability to simultaneously capture fine-grained details and broader contextual information essential for accurate fault detection. To address this limitation, this paper proposes a convolutional neural network with multi-scale residual adaptive feature fusion (MRAFF-CNN) module, which is capable of suppressing noise and extracting multi-scale features. By adding the early stopping approach and self-attention mechanism to the system, the proposed model can not only increase computing efficiency and alleviate the issue of over-fitting but also automatically learn features from the unprocessed data stream. The validation of the proposed method is conducted using the Case Western Reserve University (CWRU) dataset. In addition, experimental results show that the proposed method performs better under noise interference with different signal-to-noise ratios compared to other methods.