Applications Of Particle Swarm Optimized Deep Auto-Encoder In Fault Diagnosis Of Rolling Element Bearings

As one of the key parts of vehicle transmission system,the condition of rolling element bearings directly affects the operation safety of the whole system.Therefore,it is of great significance to study the fault diagnosis of rolling element bearings.In recent years,such deep learning methods as deep auto-encoder(DAE)have been widely applied in this field with their powerful ability in mining and extracting the deep fault features.However,there are a lot of hyper-parameters in these deep learning models,which directly affects the diagnostic accuracy but selected mainly depending on expert experience at present.Thus,their adaptiveness need to be further improved.To solve this problem,particle swarm optimized DAE models are proposed in this thesis and applied to solve the intelligent diagnosis of rolling element bearings.Firstly,the framework of DAE as well as it training skills are studied.To verify its performance in feature extraction under unsupervised learning,a reconstruction test is conducted based on simulated signals.The results show that the similarity between the reconstructed data from DAE and the original one is more than 96%.In addition,its performance in feature classification are compared with principal component analysis.The results show that the diagnostic accuracy is improved by 4%-8%,which also proves its effectiveness.After that,the effects of these hyper-parameters including number of nodes,batchsize,momentum and sparse penalty term coefficient of each hidden layer on the performance of DAE are investigated.It is found that these hyper-parameters have different influences on the diagnostic accuracy and calculation time of DAE.Based on a comprehensive consideration,the selection range of each hyper-parameter is determined.Secondly,particle swarm optimization(PSO),as an excellent evolutionary algorithm,is utilized for the hyper-parameter optimization in DAE and a PSO-DAE method is proposed and applied in fault classification of rolling element bearings.The experimental results show that PSO-DAE can obtain the diagnosis accuracy more than 95% under several data sets,and can converge rapidly within 15 iterations,which verifies the effectiveness of this proposed method.At last,multi-objective particle swarm optimization(MOPSO),which can overcome the hyper-parameter searching falling into local optimum,is employed to enhance the capability of DAE in feature representation.The MOPSO-DAE method with classification accuracy and calculation time as two objective functions is proposed.The experimental results under multiple sets of data sets show that compared with PSO-DAE,the MOPSO-DAE can significantly reduce the computational cost between 10% to 20% while ensuring the classification accuracy at99%,which has a wider practical value.