| High-voltage circuit breakers(HVCBs)are key components in a power system,which play a role in control(switching load)and protection(cutting off fault).Once the HVCB fails,it will directly endanger the safety and stability of the entire power system.Numerous investigations on the reliability of HVCBs in China and abroad show that mechanical fault is the main factor for the failure of HVCBs.Therefore,it is of great significance to improve the reliability of a power system by carrying out the research on mechanical fault monitoring and diagnosis methods of HVCBs and formulating reasonable maintenance and repair strategies accordingly.The traditional periodic overhaul-based maintenance strategy can not meet the requirements of the intelligent development of HVCBs.In recent years,intelligent fault diagnosis methods of HVCBs based on machine learning algorithms(MLAs)have acquired extensive research and attention.This kind of method has generally achieved good diagnostic results under ideal diagnostic conditions,but there are still some problems to be solved in actual fault diagnosis situations.In this thesis,the HVCB is taken as the research object,and the vibration signal is taken as the medium.The research is carried out from the aspects of vibration signal feature extraction and fault identification.Aiming at the limitations of conventional MLAs in the actual fault diagnosis of HVCBs,the corresponding solutions are proposed.The main contents and innovations of this thesis are as follows:(1)Aiming at the problem of vibration signal feature extraction of HVCBs,a feature extraction method based on mechanism action time parameters(MATPs)is proposed.First,based on the short-term Teager energy and the short-term quadratic energy,the action event enhancement parameter is designed,then MATPs are extracted from the vibration signal accordingly.Afterwards,the vibration signal is segmented by using the extracted MATPs.Finally,the energy entropy of each segment signal is calculated to form the feature vector of MLAs.Compared with the feature vector calculated based on the equal-time segmentation and equal-energy segmentation,the feature vector calculated based on the MATPs segmentation shows a better classification effect between different classes in space.(2)The influences of imbalanced data(the number of normal samples is more than that of fault samples)and unlabeled fault data(the monitoring data has only normal samples or unknown faults occur)on the performance of the diagnostic models established by conventional MLAs are analyzed.The experimental results show that imbalanced data and unlabeled fault data will degrade the performance of the diagnostic models or even disable the diagnostic models.Imbalanced data makes the diagnostic model biased toward the normal state,resulting in low fault identification accuracy,and this problem becomes more serious with the aggravation of data imbalance.The unlabeled fault data makes it difficult for MLAs to establish an effective diagnostic model to identify unlabeled faults.Based on the analysis of the above two actual fault diagnosis situations,the subsequent chapters are drawn.(3)Aiming at the imbalanced phenomenon of actual monitoring data of HVCBs,an oversampling algorithm is proposed to rebalance the data.The oversampling algorithm increases the number of samples of the minority class by synthesizing new samples so that the number of samples of different classes tends to be balanced.However,the existing oversampling algorithms have some blindness in sample synthesis,namely,they do not fully consider the distribution characteristics of the original data,which may lead to invalid synthesis or wrong synthesis.In order to alleviate this problem,this thesis proposes a new oversampling algorithm,called density-weighted minority oversampling(DWMO).According to the distribution characteristics of the original data,DWMO sets different oversampling weights for the data of different areas,realizing high-quality synthesis of new samples,and effectively alleviating the classification deviation caused by imbalanced data.The experimental results show that DWMO can effectively improve the diagnostic performance of conventional MLAs in imbalanced data fault diagnosis of HVCBs.(4)Aiming at the problem of low diagnostic accuracy of conventional MLAs in the imbalanced data fault diagnosis of HVCBs,an imbalanced data classification algorithm based on one-class extreme learning machine(OCELM)ensemble is proposed,namely MC-OCELM.The MC-OCELM algorithm ensembles multiple OCELM models and the number of OCELM models is adaptively adjusted according to the number of classes in the training set to ensure that each class corresponds to an OCELM model.During training,each OCELM model in the MC-OCELM algorithm is trained separately based on its corresponding class.Because of this training characteristic,MC-OCELM effectively avoids the influence of imbalanced data.The experimental results show that MC-OCELM achieves better diagnostic results than conventional MLAs in the imbalanced data fault diagnosis of HVCBs.(5)Aiming at the problem of unlabeled fault identification of HVCBs,a method based on the improved OCELM algorithm is proposed.The problem of unlabeled fault identification is regarded as an outlier detection problem,and a one-class classification algorithm(OCCA)is used to solve it.Considering that the existing OCCAs generally ignore the influence of the density of the sample area on the decision boundary,the density weight is introduced into OCELM algorithm,thus an improved OCELM algorithm is proposed,namely.density-weighted one-class extreme learning machine(DW-OCELM).The DW-OCELM algorithm assigns high weights to high-density area samples,which makes the diagnostic model tend to reject low-density area samples and accept high-density area samples as much as possible.The experimental results show that DW-OCELM effectively solves the problem of unlabeled fault identification of HVCBs,and achieves a better identification result than other commonly used OCCAs. |
PDF Full Text Request