A novel method for rotor fault diagnosis based on deep transfer learning with simulated samples

For solving the problem of data without failure samples, a novel transfer unsupervised learning method called classifier constrained domain adaptation network (CCDAN) is proposed for extracting transfer characteristics from simulated samples by theory model for experimental rotor fault diagnosis. Firstly, a dynamical model of Jeffcott rotor with crack fault is established and the achieved vibration responses of the system are used as generated simulation samples. Then, the multilayer convolutional network and multiple-kernel maximum mean discrepancy (MK-MMD) are adopted to extract similar characteristics between source and target domains. Finally, two independent classifiers are designed to constrain the features which have fallen near the decision boundary in network learning, which could effectively increase the accuracy and promote the fault identifying generalization. The proposed CCDAN method is verified through two unsupervised transfer tasks of crack rotor fault datasets. The comprehensive experiment analyses conclude the proposed method can learn the transferable characteristics of rotor crack fault diagnosis and its classification accuracy is superior to the existing intelligent transfer network of fault diagnosis.