Research On Fault Diagnosis And Residual Life Prediction Of Rotating Machery Based On DBN

Rolling bearing,gearbox and other rotating machinery units are extremely important parts in industrial production.In recent years,due to the rapid development of industrial automation,the requirements of equipment performance are also increasing.It is particularly important to accurately grasp the operation status of equipment and quickly identify faults.In recent years,thanks to its strong feature extraction ability,recognition ability and nonlinear fitting ability,the deep learning network has been widely used in engineering fault diagnosis and bearing residual life prediction.Based on the deep belief network,two fault diagnosis methods and a residual life prediction model are proposed in this paper.First of all,in order to deal with the problem that random noise affects the authenticity of data in the process of signal acquisition,this paper uses wavelet packet energy entropy transformation and multi-scale permutation entropy transformation on the original signal data and carries out data reconstruction.This greatly reduces the influence of random noise in the signal acquisition process and increases the diversity of different types of fault signals,making fault diagnosis less difficult.In the process of using deep belief network(DBN)for fault diagnosis,dynamic learning rate is used to avoid the phenomenon that the network convergence effect is poor or even non convergence caused by improper selection of learning rate.Dynamic learning rate strategy can gradually adapt to the size of learning rate in the training phase,so it can enhance the stability of the network to a certain extent.Secondly,the DBN is stacked by a number of restricted Boltzmann machines.The structure of the traditional DBN network and the values of hyper-parameters are determined based on previous experience or after a large number of repeated experiments,which is not only easy to fall into empiricism and but also consume a lot of time and affect the efficiency and accuracy of the experiment.This paper uses particle swarm optimization(PSO)to initialize the DBN network structure and uses the loss function value of the DBN network to construct the convergence function in the PSO adaptation process.The DBN structure is adjusted in real time according to the diagnostic ability of the network,and the dynamic learning rate is invoked during the training process.It has high accuracy and stability in the diagnosis of rotating machinery faults under multiple working conditions.Finally,in order to predict the remaining life of the bearing,a data fusion feedback deep belief echo state networks(DBESN)rolling bearing remaining life prediction model is proposed to realize the prediction of the remaining life of the rolling bearing.First,the relative root mean square of the life cycle data is obtained to get the characteristic matrix,which is smoothed by linear rectification technology to obtain the life cycle degradation curve.And then the degradation threshold and the failure threshold are obtained.Secondly,using the obtained degradation threshold and failure threshold to intercept the rolling bearing data of the two sensor channels,two sets of time series to be predicted are obtained.The multi-scale permutation entropy theory is used to extract the entropy value of the sequence to be predicted,and the extracted features are input into the data fusion feedback DBESN model to predict the remaining life.According to experimental verification,this model has good fitting ability and bearing life prediction ability.