Research On Bearing Fault Diagnosis Based On Extreme Learning Machine

Rolling bearings,extensively applied in a plenty of fields,are one of the vital part of mechanical equipments.Their conditions are directly affect machines’ working position.Because rolling bearings are in a bad working environment for a long time,so they are easily damaged by external factors.Therefore,in order to guarantee not only the safety and reliability of large machines,but also the steady growth of social economy,conducting a research about the technology of fault diagnosis is of great concern.The reserach content of the paper will focus on the following two aspects,the extraction of feature and the classification of fault.One of the difficulties of the rolling bearing fault diagnosis research is that the extraction of nonstationary and nonlinear properties should sufficiently contains the information of the fault.Due to the time domain characteristics can better reflect the working condition of bearing,wavelet packet transform is sensitive to non-stationary and nonlinear signals enough.So combining the time domain analysis and wavelet packet transform to extract mixed fault features of signals.During the course of extracting features,redundant features will not only increase the computational load of mixed feature sets,but also lead to the accuracy and efficiency of later fault classification and diagnosis.In order to avoid the influence of unrelated features on the diagnosis results,the locality preserving projection is used to reduce the dimension of the mixed features.And then,we can obtain the more targeted features as the input of classification model.Extreme learning machine is unsteady because its input weight and threshold of hidden layer built with randomness.The shortcomings makes a bad influence on the stability of network’s performacne.So,using good optimization ability of whales,setting up fitness function and iterative optimization to improve the network performance of ELM,and building WOA-ELM diagnosis model.And then using the experimental datas of Case Western Reserve University to diagnosis model for training and testing.The result shows that the whale algorithm can improve the diagnostic accuracy of the ELM network.By comparing with BP,SVM,ELM models,the validity and reliability of WOA-ELM diagnostic model are verified.In the interative process of running of whale algorithm,the WOA is easy to fall into the local optimal solution.The mutation operator can increase the population diversity of the algorithm,and the ergodicity and randomness of chaos theory can help WOA jump out of the local optimal solution.So using the mutation operator and chaotic dynamic inertia weight to improve the whale algorithm.12 groups of classical test functions are used to test the performance of IWOA algorithm and compared with GA,PSO and WOA algorithm.As the pictures of the result imply,IWOA algorithm has the best convergence precision and speed among the four algorithms.Finally,building IWOA-ELM diagnosis network for fault diagnosis experiment and compared with GA-ELM,PSO-ELM,WOA-ELM diagnosis model.The final result shows that the IWOA-ELM diagnosis model not only can better improve the diagnosis accuracy of ELM,diagnosis results is 98.33% on average,but also has a advantage of fast convergence,high giagnostic accuracy and good stability.