Research On Intelligent Fault Diagnosis Method For Vehicle Traction Transformer Based On Information Fusion

The on-board traction transformer is a key equipment in the train traction power supply system and is widely used in the train traction power supply system.During long-term operation,due to factors such as operating conditions and environment,various defects or faults may occur in the on-board traction transformer due to aging,leading to sudden temperature rise and abnormal vibration of the transformer.In severe cases,it may even lead to train shutdown.To ensure the safe and reliable operation of on-board traction transformers,based on the idea of information fusion and various status monitoring data of on-board traction transformers,through data analysis and processing,machine learning algorithms were systematically studied for fault diagnosis of on-board traction transformers.The accuracy and practicality of the constructed fault diagnosis model and algorithm were verified through practical cases.The main innovative research work carried out in this thesis is as follows:Firstly,a method based on local mean decomposition for extracting feature information of on-board traction transformers is proposed to address the issue of multiple types of status detection data and difficulty in accurately extracting key fault features.By analyzing the internal fault mechanism of on-board traction transformers,a fault feature sample dataset was established,which includes dissolved gases in oil,insulation oil test data,and main insulation aging status.By combining association rule analysis with long-term and short-term memory networks,abnormal data in the fault feature sample dataset was identified and interpolated.After preprocessing the fault feature data,reliable data guarantee is provided for the verification of subsequent fault diagnosis algorithms.Then,aiming at the problem that there are many types of condition monitoring data and fault characteristics of vehicle mounted traction transformer,which leads to inaccurate classification of traditional fault algorithms,a traction transformer diagnosis algorithm based on the combination of kernel principal component analysis(KPCA)and fuzzy clustering is proposed.Based on various condition monitoring data of on-board traction transformer,the KPCA is used to extract the feature quantity that can effectively represent the fault status of onboard traction transformer,and finally effectively reduces the dimension of fault feature and the complexity of diagnosis process.On this basis,the fault state of vehicle mounted traction transformer is diagnosed by Fuzzy clustering algorithm.Compared with the traditional methods,the proposed method based on KPCA and fuzzy clustering not only considers the feature differences of traction transformers under different working conditions,but also can realize the rapid diagnosis of fault states.At the same time,a hybrid diagnosis method based on support vector and improved seagull optimization algorithm(ISOA)is proposed to meet the accuracy requirements for fault diagnosis of on-board traction transformers.Firstly,the KPCA is used to preprocess and extract features of dissolved gas data in oil,electrical measurement parameters and insulation oil quality detection data to reduce feature dimensions and data complexity.On this basis,in order to optimize the algorithm parameters of support vector machines to solve overfitting problems,an improved seagull optimization algorithm was proposed.By simulating the foraging process of seagulls,the optimal combination of support vector machine(SVM)parameters was found,which improved the classification accuracy and robustness of SVM.Compared with Fuzzy clustering method,this method has higher diagnostic accuracy under different sample numbers,and can effectively improve the reliability of fault diagnosis of vehicle mounted traction transformer.Finally,in order to balance the accuracy and speed of fault diagnosis for on-board traction transformers,a comprehensive fault diagnosis method based on the combination of timefrequency entropy and probabilistic neural network is proposed to address the shortcomings of a single state monitoring data feature that cannot effectively represent traction transformer faults.Using the correlation between condition monitoring information and train mileage,fault features are effectively selected,and time-frequency entropy is used to extract features of frequency domain feature signals.Through the strong adaptability and nonlinear mapping ability of probabilistic neural network,time-frequency entropy features are fused and fault classification is carried out,thus improving the accuracy and rapidity of fault diagnosis.The proposed method can meet the operation and maintenance needs of on-board traction transformers under different working conditions.In scenarios with fourth level maintenance or above,the KPCA-ISOA-SVM method with high diagnostic accuracy can be used;In the second/third level repair,the method of combining KPCA with Fuzzy clustering,which can diagnose more quickly,can be used;In first-level maintenance scenarios where there is an urgent need for maintenance time and high accuracy requirements,the proposed singular spectrum analysis and approximate entropy based probabilistic neural network(SSA-Ap EnPNN)method can be used.The research work will provide a fast and reliable diagnostic method for the fault diagnosis of on-board traction transformers,effectively improving the efficiency of equipment operation and inspection.