Research On Fault Diagnosis Of Rolling Bearing Based On Deep Belief Network

Due to the complex working conditions of the rolling bearing,the collected vibration signals are often submerged in strong background noise,so it is difficult to extract effective fault features.Therefore,the fault type and location can hardly be determined through accurate physical model based on traditional vibration signal.The development of Deep Belief Networks(DBN)makes up for the shortcomings of the traditional fault diagnosis methods in dealing with non-stationary signals.The deep Belief Networks can reduce the complexity of the fault diagnosis of the rolling bearing to the greatest extent by means of the RBM stacking,mining the internal features of data layer-by-layer,suppressing the noise,compressing and reconstructing the signal,visualization of fault features and fast recognition of fault types.Firstly,aiming at the problem of decreasing the accuracy of Fault diagnosis caused by stochastic determination of the layers of DBN,a method for determining the number of implicit layers based on dynamic adding algorithm is proposed in this paper.The purposed method is to compensate the number of layers by error and decline rate of error,to improve the intelligent degree of network structure selection,and to ensure that DBN will always complete the precision requirement in fault diagnosis process with the simplest structure.Then number of hidden layer neuron nodes of the model is set according to the principle of layer-by-layer decline,which not only realizes the extraction of high-level features,but also compresses and simplifies data to facilitate classification.Next,the BP algorithm adopted by the traditional depth architecture is retained at the supervisory fine tuning stage of DBN.When the parameter space optimization is carried out,the recursive speed of each layer is inevitably limited by the gradient of the activation function itself,which leads to the increase of training difficulty and duration,and it is easy to fall into the local optimal solution,which leads to the reduction of the generalization of the network.In view of the above problems,a new optimization algorithm based on adaptive learning rate L-M DBN is proposed,and the speed of parameter optimization can be adjusted by variable learning rate.When the error of the initial iteration is large,the gradient is large to ensure a satisfactory speed of parameter optimization and improves the sensitivity of the model.With the increase of iterations inthe later stage,the error gradually decreases and the speed of parameter updating slows down to prevent the miss of the optimal value and the non-convergence of the network,which greatly improves the efficiency of the training of DBN.Finally,the model is tested from several aspects,such as fault separation ability,stability and sensitivity.The diagnosis results are compared with the traditional DBN model,it is verified that the proposed method can not only ensure the high learning efficiency of DBN,but also effectively avoid the occurrence of model over-fitting and the recognition accuracy based on the rolling bearing fault verification set is as high as97.96%.