Research On The Vibration Data Fault Diagnosis Based On Deep Domain Adaptation

Deep learning theory,with its powerful modeling and representing capabilities,has become one of the most active frontiers in data-driven intelligent fault diagnosis,however,the two basic prerequisites for the accuracy of the fault classification trained by the deep learning model are: training and test vibration data are independently distributed and the number of labeled fault samples is sufficient.In practice,these two conditions are often unable to meet.First of all,the labeling of data is an expensive and time-consuming operation,and the mechanical equipment will work under different working conditions,resulting in a difference in the distribution of collected vibration data.This raises an important question,how to use a small amount of labeled data or auxiliary domain data to build a reliable model to diagnose target domain with different data distribution.Transfer learning can be used to solve such problems.Therefore,this paper combines deep learning and transfer learning for intelligent fault diagnosis of mechanical equipment,which can save training costs while improving model diagnostic accuracy.The main research work of this dissertation includes the following points:(1)We first introduced the research background and existing problems of data-driven fault diagnosis of rotating bearings.Then studied the basic principles of deep learning and transfer learning,and the idea of combining transfer learning and deep learning is proposed for intelligent fault diagnosis of rotating bearings.(2)An intelligent fault diagnosis method based on convolutional neural network and domain adaptation is proposed for the case when there are only a small number of labeled samples in the target domain.First,samples from source domain and target domain are constructed as paired samples for training.Then a convolution siamese adaptation network is designed to map the data in the two domains into a common feature space,and a softmax classifier is used to transform the data from the feature space to the class space.(3)An intelligent fault diagnosis method based on deep adversarial transfer network is proposed for the case where only unlabeled samples are included in the target domain.the feature extractor and the domain discriminator are optimized through adversarial training,and the Wasserstein distance is used as a metric to learn a domain-invariant feature to achieve the purpose of domain adaptation.At the meantime,the classification error is also minimized.Ultimately the cross-domain fault diagnosis can be achieved.