Research On Fault Diagnosis Of Rolling Bearing Based On Transfer Learning

With the rapid development of big data technology,the fault monitoring of mechanical equipment has entered the intelligent era.As a key component of rotating machinery,once a fault occurs,production efficiency will be affected at least,and major safety accidents will be caused at most.Therefore,the fault diagnosis of rolling bearing is of great significance.The traditional data-driven diagnosis method needs the prior knowledge of experts,and most deep learning algorithms need to assume that the characteristic distribution of training set and test set is the same,so they can not achieve good results in cross domain diagnosis.Transfer learning can apply the diagnostic model of source domain data training to different but related target domains,and solve the problem of different data distribution between target domain and source domain.In this paper,the fault diagnosis of mechanical rolling bearing will be carried out based on transfer learning.The main work is as follows:(1)For domain adaptation problems in rolling bearing migration fault diagnosis,a domain adaptation algorithm based on high-order moment momentum matching is proposed to match the distance from the high dimension.In order to avoid the computational explosion caused by high-dimensional matching algorithm,a random matching sampling principle is proposed to greatly reduce the complexity of the algorithm.By introducing the dynamic exponential penalty term,the problem of large difference in convergence speed and value of various optimization terms is solved.The collected multi-channel data are fused to obtain more representative characteristic data,and the data are expanded by sliding window sampling.Finally,the two-dimensional gray image is transformed as the input of convolutional neural network.(2)For the problem of transfer fault diagnosis under small samples,a transfer convolution neural network model based on time-frequency transform is proposed.Through the feature extraction of the original signal by short-time Fourier transform,the spectrum with obvious feature representation can be obtained.In the convolution neural network,hierarchical regularization is used to avoid over fitting,and the maximum mean difference minimization is used for domain adaptation.By adjusting the parameters to obtain time-frequency maps of different sizes,we can use less data to obtain large-scale spectrum maps,which can reduce the demand for data in the target domain.(3)For the real-time problems in transfer fault diagnosis,a rolling bearing fault diagnosis model based on on-line transfer convolution neural network is proposed,and a pair of off-line convolution network and on-line convolution network with the same model parameters are constructed.The model parameters and the source domain feature data in the fully connected layer are obtained by pre training the offline network.The online network uses the parameters of the pre training model for parameter initialization,realizes domain adaptation with the trained source domain data,and fine tunes the network model through a small number of labeled samples to realize online transfer diagnosis.