133 / 2021-03-31 12:13:36

Intelligent Cross-Domain Fault Diagnosis Method with Domain Alignment and Discriminative Feature Learning

Fault diagnosis; Domain adaptation; Maximum mean discrepancy; Classifier alignment; Convolutional neural network

Abstract: Rotating machinery is a key component of modern machinery and equipment, any failure of which may cause equipment damage or serious safety accidents. For this reason, it is essential to ensure the normal operation of rotating machinery through accurate fault diagnosis. Due to the changeable operating conditions of rotating machinery, the feature distributions of fault are usually changed, most current cross-domain intelligent fault diagnosis methods only achieve global domain alignment. In this paper, a novel joint domain alignment and discriminative feature learning method (DADFL) is proposed for cross-domain fault diagnosis of rotating machinery. In DADFL, a novel strategy of synchronously implementing global domain alignment and class alignment is innovatively proposed, in which a feature extractor based on CNN is designed to extract high-level features and two different structure classifiers are designed to improve the ability to capture discrepancy features of target samples near the class boundaries. First, a feature extractor and two discrepant classifiers are established to extract high-level features and output predicted results. Then, the maximum mean discrepancy loss is used to reduce the marginal distribution discrepancy of high-level features between the source domain and target domain. Finally, the classifier discrepancy loss and the contrastive loss is creatively combined for class alignment learning, which can effectively reduce the conditional probability discrepancy between the two domains. Moreover, two experimental cases are performed to demonstrate the effectiveness of the proposed method. Besides, comparative study is conducted by several advanced cross-domain diagnosis methods and the experimental result illustrates superiority of the proposed method.