Silhouette Coefficient Regularization Based Adversarial Domain Adaptation Network for Fault Diagnosis of Rotating Machinery

Despite the fact that machine learning approaches have been proven to be useful in the diagnosis of rotating machinery, there is a key assumption that all the data are distributed in a same manner. However, given the varying working conditions, this assumption is difficult to maintain in real-world settings, and it will result in unsupervised and domain shift issues, increasing the difficulty of fault diagnosis. Aimed at these issues, the adversarial domain adaptation method has provided a way to solve these problems caused by changeable working parameters. Due to the good embeddedness with deep learning network, and applicability of unsupervised learning, the adversarial domain adaptation has been widely concerned. Most of the adversarial domain adaptation based methods focus on how to adapt the source and target domain data to improve the fault diagnosis performance under variable working conditions. However, these methods only focus on the adaptation of inter-domain data, and ignore the separation of inter-class data, thereby limiting the improvement of fault diagnosis performance. In this study, we employ the silhouette coefficient to regularize the adversarial domain adaptation network in order to extract domain-invariant features while maintaining inter-class data separability. To enhance fault diagnostic performance, the proposed method has the ability to obtain a better classification boundary while achieving the alignment of the cross-domain in an unsupervised way. The experiments are designed via the CWRU dataset. The results suggest that the proposed method is effective to improve the adversarial domain adaptation-based framework in fault diagnosis.