Convolutional neural networks (CNNs) have demonstrated exceptional effectiveness in the field of intelligent fault diagnosis. However, due to the varying working conditions of rotating machinery, the training and testing data in practical engineering applications often follow different distributions, leading to a decline in the performance of CNNs. To address this issue, we propose a novel network that incorporates domain-specific hybrid normalization, aimed at enhancing a CNN’s modeling capabilities within one domain while improving its generalization ability in another domain. Specifically, batch normalization and instance normalization are integrated into a unified building block to replace the batch normalization in CNNs. To further boost generalization, we have developed a domain-specific dynamic adjustment technique for hybrid normalization. It can achieve deep adaptation effect for cross domain fault diagnosis by modulating the statistics from the source domain to the target domain in all hybrid normalization layers. The domain-specific hybrid normalization has been validated in aeronautical bearing fault diagnosis experiments under various speeds and loads conditions. Experimental results show that our domain-specific hybrid normalization can be applied to many deep architectures, significantly improving performance without increasing computational cost.
 

fault diagnosis, normalization, transfer learning, convolutional neural network, deep learning