Adaptive hybrid attention mechanism deep residual threshold networks for bearing fault diagnosis under noisy environments

Intelligent bearing fault diagnosis based on deep learning has immense potential. However, improving the noise immunity, generality, and accuracy of fault diagnosis methods is still challenging. This paper proposes a novel adaptive hybrid attention mechanism deep residual threshold network (AHA-RTN) for bearing fault diagnosis under various noise conditions. First, channel-wise and spatial attention were both integrated into residual blocks to capture multiscale information. Hybrid attention was obtained using the proposed adaptive attention module, which computes adjustment coefficients of each attentional mechanism. Next, a novel noise reduction activation function based on soft thresholding was incorporated to suppress noise. Finally, the method was validated on two distinct bearing datasets under various noise conditions. The results show that the proposed AHA-RTN has better noise immunity and accuracy than the other advanced multiscale convolutional neural networks.