Study On Convolutional Neural Network Based Domain Adaptive Machine Diagnosis Methods

As equipment evolves toward automation and intellectualization,Prognostic and Health Management plays a significant role in the high-precision and high-reliability electromechanical systems.At the same time,convolutional neural network and deep learning technology have greatly promoted the development of data-driven fault diagnosis approaches.However,the following problems seriously restrict the application of convolution network in practical engineering: the high cost of data labeling,the everchanging working conditions,and the differences between data-collected and testing machines.In this paper,domain adaptive theory and transfer learning are introduced into the intelligent fault diagnosis algorithms,and the fault diagnosis under the conditions of lack label,variable working situations and cross-equipment are studied respectively.The theoretical basis of the convolutional neural network is explained.Meanwhile,the concepts of transfer learning and domain adaptation are introduced into the field of fault diagnosis.The dataset adopted in this paper is interpreted in detail.A platform for ball screw fault signal acquisition is built to verify the accuracy of the proposed algorithm.Frist,the model-based transfer learning fault diagnosis algorithm is studied.This paper improves the existing generative adversarial network,adds features to the simulation signal generated by the physical model,obtains the generated signal tending to the distribution of acquisition signal,and trains the classifier for fault diagnosis.The experimental results show that the proposed algorithm can greatly reduce the problems of the traditional generative adversarial network such as mode collapse,training nonconvergence.And it can maintain the fault information while generating data.It can also use unlabeled data to diagnose faults and verify the domain adaptive ability of the proposed algorithm.Second,the fault diagnosis method for ICN cross-working condition is studied.Since the existing methods are difficult to adapt to the changeable working conditions,this paper uses capsule network and Inception block to improve the non-linearity of the network and extract more abstract features.experiments on bearing and ball screw datasets are carried out to verify the performance of the proposed method in crosscondition fault diagnosis.Last but not least,the feature regularization method for cross equipment based on W-distance is studied,which overcomes the low accuracy problem caused by equipment differences in existing fault diagnosis algorithms.In transfer learning,W-distance,as a common measure of distribution difference,is difficult to obtain exact solutions.In this paper,an approximate simplification theorem of W-distance is proposed and proved.In addition,results on several open datasets and self-collected datasets indicate that the proposed method can handle cross-equipment fault diagnosis with essential domain adaptive ability.