Research On Fault Diagnosis Of Improved Extreme Learning Machine Based On Information Reconstruction

As a method to detect the stability of the signal sequence,the permutation entropy PE(Permutation Entropy)can amplify the minute changes of the time series to measure the stability of the signal.The advantage of PE is that the calculation and the speed is simple and fast.The disadvantage is that PE algorithm can’t perfectly deal with the multi-frequent signal.Extreme learning machine(ELM)is characterized by its fast training speed and go without interfere in parameter adjustment.Therefore,the generalization ability of ELM is strong.However,there are many inadequacies in the training process of ELM as a whole.On hand is the random selection of the parameters leads to the poor robustness of the model,and the result leads to the formation of many non-optimal parameters.Another reason is the number of hidden layer nodes ‘adjustment according to the artificial experience.With the increase of the number of hidden layers,the whole system could turn into morbidness.The quality of rolling bearing diagnosis leaves much to be desired.To solve this problem,an optimization algorithm — adaptive weighted permutation entropy(WPE),is proposed in this paper in order to overcome the shortcomings of PE.According to the idea of entropy weight,weight factors to be added if the current entropy arrangement vector is just in line with some kind.Current entropy vectors are standardized in the scale time,and the results after the processing are weighted.Finally,when the overall contribution rate is calculated,the influence of the entropy arrangement with weight is highlighted,which can more reflect the volatility of the original signal.The adaptivity of this topic can be introduced from two aspects: mutual information method(MI)and false nearest neighbor(FNN)method.he minimum delay time is determined according to MI method,and the number of embedding dimensions is determined according to the FNN method.As traditional ELM can not effectively discover its most influential acious vectors.In this paper,the WPE-FWELM algorithm—ELM based Filter & Wrapper is proposed.The core ideal is to use the filtering and packing(Filter+Warpper)algorithm.First,the data is arranged in the feature information,and the original data is divided into several sub data sets,and each data set is scored according to the F-score method in the filter process.In the process of Wrapper,training model is carried out according to the eigenvectors.K-fold cross validation as the validation set during the whole process.During each time of train each sub dataset,the error range is continuously updated to achieve the optimal error rate.The experimental test demonstrated that according to the criterion trained by the optimal feature,the minimum error diagnosis results can be obtained well.The platform of the experiment is MATLAB 2017 a.By comparing with the traditional ELM and new method of fault diagnosis algorithms,the results show that the algorithm—WPE-FWELM has higher classification accuracy and robustness.