LAUSR

Convolutional neural networks (CNNs) have demonstrated promising effectiveness in vibration-based fault diagnosis. However, the faulty characteristics are usually distributed on different scales and contaminated by noises from various sources. Therefore, it is still a challenging task for traditional CNNs to efficiently extract multiscale features and suppress unrelated noises in vibrational signals. In this article, a novel fault diagnosis framework called multiscale residual attention CNN (MRA-CNN) is proposed to learn discriminative multiscale features from vibrational signals and fully utilize the multiscale features to reduce noises. First, the raw vibrational signals are fed into the multiscale learning module (MLMod) to obtain multichannel feature maps generated with different kernel sizes. Second, the feature maps are passed through an efficient residual attention module (RAMod) to get the attention mask for weighing all locations of different channels of the multiscale feature maps to denoise. Third, to mitigate the information loss in attention, a new strategy called residual attention learning (RAL) is proposed to improve the feature extraction ability of RAMod, in which the learned attention mask itself is also regarded as feature maps by a shortcut connection. Experimental validation is conducted on two bearing datasets. The results show that the proposed method can learn more effective features from vibrational signals and deliver much higher accuracy than the seven state-of-the-art methods under highly noisy environments.