Cross-Domain Anomaly Detection using Unsupervised Representation Learning and Domain Adaption

Aiming at the urgent demands of industrial fault detection, cross-domain detection is a promising strategy for overcoming the obstacle of the premise of data identical-distribution. This paper proposes a cross-domain anomaly detection method based on unsupervised representation learning and domain adaptation. In order to learn effective features from the original signals, the multidimensional scale loss and an improved instance-based discriminative loss are combined. The first one is for retaining structural information of the data and the second one is for obtaining domain-invariant characteristic. The proposed method is validated in two detection cases including manipulator and bearing. Detection results show that the proposed method has superior performance than several widely used detection methods.