Research On Fault Diagnosis Method Based On Deep Transfer Learning

Now we are in an era of Big Numbers.Every day,all walks of life will generate a huge amount of information,informing the continuous accumulation of hardware technology in recent years,which makes machine learning have better conditions and be widely used.Traditional machine learning methods generally rely on the basic assumption that the mechanism of data generation does not change with the environment.However,in various existing application fields,such as machine vision,character recognition,Natural language processing,retrieval and recommendation services,the above assumptions are often changed due to changes in the environment.In the actual industrial production,the data of mechanical equipment under different working conditions are inconsistent in the general spatial and temporal distribution or unbalanced,which affects the accuracy and generalization ability of the diagnostic model.As a learning method that can effectively model related domain problems,transfer learning solves the problem of inconsistent data distribution and small-scale data set model training in traditional machine learning.In order to reduce the problem of time and energy consumption in data labeling,these is first proposes a deep transfer learning bearing fault diagnosis algorithm based on unsupervised multilayer fitting.Firstly,the vibration signals converted into grayscale images are input into the deep residual network to improve the extraction of transferable features.Then,the vibration features are mapped into a high-dimensional space,and then multi-layer adaptation is carried out on the basis of the original one-layer adaptation.Multiple kernel functions are used to minimize the probability distribution distance,so as to achieve unsupervised depth adaptive diagnosis.In order to further improve the generalization ability of predictive models to unlabeled data in fault diagnosis,an unsupervised depth adaptive fault diagnosis algorithm based on second-order statistical feature alignment was proposed.First,the residual network is used to extract the features,and then the extracted features are transformed into the same subspace to establish a unified model.Then,through nonlinear transformation,the covariance of the domain distribution is aligned to continuously reduce the distribution difference.Finally,the joint loss function is optimized to realize unsupervised adaptive diagnosis.Finally,based on the above theoretical research,the selection of case western reserve university bearing fault diagnosis experiment center data experiment research,we validate MK-UDAN and DCA-UDAN method,and contrast experiment design,for other different learning methods and network structure,the record in the process of experimental training precision,loss and confusion matrix,and finally through the combination of theory and experiment prove that the proposed algorithm has good generalization ability and application prospect.