Research On Simulation Data-driven Fault Diagnosis Method Of Rolling Bearing Based On Deep Transfer Learning

With the improvement of precision and integration degree of industrial equipment,minor faults of bearing may cause equipment failure.Therefore,it is necessary to make fault state classification of rolling bearing.Although the improvement of equipment intelligence provides a large amount of monitoring data,the training of intelligent diagnosis model still needs data sets,which have rich fault types,sufficient samples and fault labels.The bearing fault data set can be obtained from the dynamic model.However,when applied to real fault diagnosis,the performance of intelligent diagnosis model trained by simulation data decreases.To solve these problems,this paper proposes to acquire complete simulation training data set through bearing fault dynamics model,then analyze the feasibility basis of simulation data used for actual fault diagnosis,and finally establish a deep transfer learning fault diagnosis model to realize the diagnosis of actual faults.For dynamic modeling of rolling bearing faults,firstly,the local defects of raceway are equivalent to rectangles.Then the interaction between the radial edge and bottom of the defect and the rolling body is analyzed.Finally,the piecewise displacement excitation function is derived when the rolling body passes the raceway with local defects of different angles.Based on Hertz contact theory,the nonlinear contact force of the rolling body is calculated,and the influence of raceway lubricating oil film,random sliding of the rolling body and other factors is considered,the dynamic model that can simulate local defects of the inner and outer raceway of the rolling bearing is obtained.The differences and similarities between virtual and real data are studied from time domain,frequency domain,time and frequency domain,probability distribution and so on.On the basis of analyzing the principle of fault excitation,the rationality of the established dynamic model is verified,and the movability of the simulation data used in the actual fault diagnosis is evaluated,which provides a theoretical basis for the deep transfer learning.According to the two ways in the development of deep learning and transfer learning that deepening network and multi-layer adaptation,the deep learning diagnosis model based on Alex Net and Res Net is established.The data preprocessing,loss function and optimization algorithm of the model are selected.The wavelet time-frequency diagram of simulated signal is taken as the model input,and the training and diagnosis performance of the two models are compared.Based on the above two deep learning diagnosis model is established,Multi-Kernel MMD is selected as the measurement of probability distribution difference,the train domain and diagnosis domain hidden features are mapped to Reproducing Kernel Hilbert Space.Then the differences of data set is calculated,and combined with training domain classification loss for the total loss function,achieve to the actual fault diagnosis.Finally,the performance of the deep transfer learning diagnosis model is verified by experiments,and the advantage of training the model with simulation data is verified.