Improved Deep Learning Algorithm And Its Applicaton In Life Prediction Of Rolling Bearing

Rolling bearing is widely used in production and plays a significant role in the equipment.Thus,the failure of the rolling bearing will stop the whole equipment and reduce production efficiency.The real-time monitoring of the state,which can judge the fault in advance and predict the life of the rolling bearing,can comprehensively reflect the failure degree,improve the production efficiency,and reduce the rate of accident.Therefore,the study of rolling bearing' life prediction is of practical application value.This thesis,with the rolling bearing as the research objective,firstly discusses the traditional method of vibration signal processing,and then the feature extraction is performed according to the vibration signal.Subsequently,the thesis discusses,on the basis of Ensemble Empirical Mode Decomposition(EEMD),the life prediction of the rolling bearing,and analyses its life with the combination of Neural Network(NN),Support Vector Machine(SVM)and EEMD.Then,this thesis studies the life of the rolling bearing with much attention given to the usage of the Deep Learning(DL).The three common DL models used in this thesis are Deep Belief Networks(DBN),Stacked AutoEncoder(SAE),and Deep Extreme Learning Machine(DELM).On account of these three original models,this thesis improves the DL model from the aspects of data dimensionality reduction and parameter optimization.Finally,the improved model is used for rolling bearing life prediction.By comparing the experimental analysis results,the improved DL model is better than the original one in practical application.The main research work of this thesis is as follows:1.Three analytical methods of vibration signal are discussed,and trend analysis is performed in terms of characteristics.EEMD,a time-frequency domain analysis method,is used to decompose the original signal and extract features of decomposed signals.The extracted features,used as input data of the NN and SVM models,accomplish the life prediction of the rolling bearing with these two models.2.The original DBN model is used for rolling bearing life analysis.In order to reduce the training time,the ELM model,based on the DBN model's reverse adjustment process,is used as the DBN output layer to obtain the DBN-ELM model.According to the idea of dimensionality reduction,this thesis combines the Principal Component Analysis(PCA),the Locally Linear Embedding(LLE),a kind of manifold learning algorithm,and the DBN,thus obtaining the PCA-DBN and LLE-DBN model.The application of the three improved DBN models predict the life of the rolling bearing,thus verifying the feasibility of the models.Finally,the comparative analysis shows that the LLE-DBN model bears better prediction results.3.The original SAE model is used to predict the life of the rolling bearing.The parameters of the SAE model are determined by experience,so the SAE model has a weak adaptability.Considering this the fact,this thesis proposes an Adaptive Stacked AutoEncoder(ASAE)model.The model can automatically determinate the number of layers and nodes by error and entropy,which makes the ASAE model more adaptable.Being applied to the life prediction of the rolling bearing,the ASAE model turns to be feasible.Finally,the prediction result of the ASAE model is much better when compared with that of the SAE model,indicating that the ASAE model has certain superiority in prediction.4.The DELM model is used to analyze the life prediction of the rolling bearing.Based on the disadvantages of the DELM model,this thesis optimizes the number of nodes in the DELM model,and proposes the Dynamic Particle Swarm Optimization Deep Extreme Learning Machine(DPSO-DELM).Compared with the original model,DPSO-DELM can optimize the number of nodes in accordance with the actual situation and improve the prediction result.The experiments demonstrate that the DELM and DPSO-DELM models are feasible for rolling bearing life prediction.Finally,the comparative experiment also verifies that the DPSO-DELM model has a better application value.