Research On Feature Selection In Diagnosis Of Rotor Faults

With the rapid development of science and technology,mechanical equipment will produce a large amount of data at every moment in practical work.However,only part of the data can be used by us.Therefore,how to extract sensitive features from these data is currently a research hotspot in the field of machine learning.Feature selection algorithm can select the features with better classification performance from massive data,which provides a new idea for obtaining the optimal feature subset with high classification accuracy.Therefore,in order to effectively reduce the dimension of fault data set and extract the optimal feature subset sensitive to classifier,this paper studies the feature selection method.The main research contents and results are as follows:(1)The development of feature selection methods is briefly introduced,and summarized several common methods.Aiming at the problem of extracting the optimal feature subset which can effectively represent fault information,proposed a hybrid feature selection algorithm based on Relief F algorithm and Particle Swarm Optimization(PSO).The performance of the proposed method is verified on a two-span rotor test bench.This method can obtain a simplified feature subset.The accuracy of fault type identification is improved obviously.(2)To improve the classification accuracy of fault data sets,proposed a feature selection method based on hybrid Relief F and Quantum Particle Swarm Optimization(QPSO).Firstly,Relief F algorithm is used to screen the fea tures of the original fault data set.Then,QPSO is used to filter feature set for the second time.Finally,the optimal feature subset is obtained and the parameters of support vector machine(SVM)are optimized.The method can achieve synchronous optimiz ation of feature subset and SVM parameters,and obtain the highest classification accuracy.(3)In order to effectively reduce the dimension of original fault data sets,proposed an integrated feature extraction model based on empirical wavelet transform(EWT),multi-scale fuzzy entropy(MFE)and t-distributed stochastic neighborhood embedding(t-SNE).Firstly,a series of amplitude-frequency modulation(AM-FM)components are obtained by EWT adaptive decomposition of vibration signals,and AM-FM with larger correlation coefficient is selected to reconstruct the signals.Then,calculated MFE of reconstructed signals to form a high-dimensional feature set to characterize the fault state,and the dimension of high-dimensional feature information is reduced by t-SNE.Finally,the sensitive fault data set is input into SVM for pattern recognition and classification.The diagnosis model can greatly reduce the dimension of the original data set and significantly improve the visualization in low-dimensional space.(4)The methods proposed in this paper are all to study the double-span rotor testbed at the same frequency and speed,without changing the working environment.In the next study,we should try to do further research under different frequencies,different environments and different rotational speeds.