Intelligent Fault Diagnosis with a Deep Transfer Network based on Wasserstein Distance

Abstract

Abstract Intelligent fault-diagnosis methods based on deep-learning technology have been very successful for complex industrial systems. The deep learning based fault classification model requires a large number of labeled data. Moreover, the probability distribution of training set and test data should be the same. These two conditions are often not satisfied in practical working conditions. Thereby an intelligent fault-diagnosis method based on a deep adversarial transfer network is proposed, when the target domain only has unlabeled samples. The Wasserstein distance is used as a metric to learn a domain-independent feature through the adversarial training between the generator and the domain discriminator. Meanwhile, a reasonable loss function of fault classification is designed, which can ensure that the learned feature does not contain domain information, but also contains fault classification information. Finally the cross-domain fault classification can be solved, even if there is no labeled vibration data in the target domain. The experimental results show that in transfer tasks under different working conditions, the fault classification accuracy exceeds 90%, which is approximately 10% higher than that of the comparison method.