Research On Condition Recognition Methods Of Rolling Bearings Based On Kernel Entropy Component Analysis

Bearing failure is one of the most common reasons for machine breakdown,and fault diagnosis is essentially a pattern recognition issue.How to extract sensitive,separable and regular features from the original complicated vibration signals is the focus of fault diagnosis of rolling bearings.The dissertation introduces kernel entropy component analysis(KECA)to recognize different operating conditions of rolling bearings,and several improved feature extraction and diagnosis methods are proposed based on KECA.The main content of the article is displayed as follows:High-dimensional feature sets constructed with mix-domain features are often used for fault diagnosis of rolling bearings,whose relevance and redundant may influence the performance of the recognition.Hence,a condition recognition method of rolling bearings based on KECA is proposed.A multi-domain high-dimensional feature set including time-domain,frequency-domain and time-frequency domain features of the original vibration signals are first constructed,which can reflect the operating condition of the rolling bearings.The most sensitive low-dimensional features of the high-dimensional feature space are extracted based on the information entropy,which can reflect the structure related to the Renyi entropy and maintain the maximum entropy of the input feature set.The extracted features are finally inputted into a classifier to recognize the different bearing faults.Aiming at the KECA algorithm is unsupervised,a supervised method called supervised KECA(SKECA)are proposed.SKECA are used to learn the multi-domain high-dimensional feature set,and extracted the low-dimensional sensitive features embedded in the high-dimensional space to reflect the operating condition of rolling bearings.The tagged information is introduced in SKECA to adjust the distance between the classes,which can largen the between-class distance and diminish the within-class distance.Then the intrinsic structures of the high-dimensional nonlinear fault data can be extracted from the vibration signals.Another condition recognition model based on a fisher discriminant KECA,combining KECA and LDA(KECA+LDA),is proposed to compare.KECA+LDA take advantages of those two methods: LDA can utilize the labelled information;while KECA can maintain the Renyi entropy of the input data and solve the small sample size problem encountered by LDA.In view of the SKECA algorithm cannot estimate label information of the testing sample,a class-information-incorporated KECA(CIKECA)is proposed.CIKECA can sufficiently utilize the class information of given data and still following the same simple mathematical formulation as KECA.A kernel training matrix containing the class information is constructed by combining the training samples and the labelled imformation.CIKECA is used for extracting the sensitive low-dimensional manifold features and the estimated label information to reflect the operating condition of the rolling bearing.In addition,a fusion nearest classifier based on those two features is used for classification the operating condition of rolling bearings.In view of the features leading to large class separability do not have to result in a high predictive accuracy,a weighted kernel entropy component analysis(WKECA)is proposed.WKECA can extract discriminative features to express the original clusters,and meanwhile to find a trade-off between maximizing the testing accuracy and minimizing the training error.WKECA constructes a modified Fisher criterion by using the labeled information and introduces a weight strategy in the feature extraction.The class-related weights are introduced to denote differences among the samples from different patterns.Genetic algorithm(GA)is implemented to seek out appropriate weights for optimizing the classification results,which can maximize the testing accuracy and minimize the training error.In order to verify the feasibility of the above algorithms,the fault data of three expriments are employed in this article.The experimental results demonstrated the feasibility and effectiveness of of the recoginition model of rolling bearings based on KECA,SKECA,CIKECA and WKECA.