Research On Inverter Overcurrent Fault Diagnosis Method Of Electric Locomotive Based On Deep Learning

With the rapid development of China’s railroad industry,the safety of electric locomotives has received more and more attention from researchers.Traction converter inverter overcurrent,as a high incidence fault of electric locomotive,has complex fault causes which is difficult to be investigated.Traditional fault diagnosis methods have high requirements on expert knowledge,and the fault feature extraction methods are tedious in the face of complex fault conditions.Deep learning provides a new direction for fault diagnosis technology development because of its strong data feature recognition capability.It is found that convolutional neural network has strong feature extraction function for two-dimensional data,and after two-dimensional processing of one-dimensional electrical signal,the fault diagnosis work with good results when using deep learning related methods.However,the correlation analysis in the existing studies are inadequate,and most of the diagnosis is based on high quality data sets and single signal sources,which has some limitations.Therefore,this thesis introduces deep learning into fault diagnosis technology,takes common traction converter inverter overcurrent fault as the research object,focuses on solving the problems of small samples and data imbalance in inverter overcurrent fault diagnosis,and achieves the accurate and reliable fault diagnosis effect.The main work of the thesis is as follows.(1)Signal processing methods are the key to achieve high-quality fault diagnosis.To this end,this thesis designs a two-dimensional data model which can fuse multiple signal sources to address the limitations of existing signal analysis methods and the characteristics of convolutional neural networks.Multiple electrical signals at fault are processed in two dimensions,which to improve the information content of fault data.(2)Aiming at the characteristics of less actual data samples for electric locomotive fault diagnosis,this thesis designs an intelligent fault diagnosis method based on twodimensional model with multiple data sources and pre-training model.The method takes the two-dimensional processed fault data as input and selects a suitable pretraining model to build a fault diagnosis network.It avoids the dependence of traditional fault diagnosis methods on expert knowledge and signal processing results,reduces the required number of samples and training time,which simplifying the fault diagnosis workflow and improving the diagnosis accuracy.(3)The DCGAN method is used for data generation of unbalanced sample sets to address the data imbalance that is prone to occur in the actual fault diagnosis work.In this thesis,according to the characteristics of inverter overcurrent fault signals,the DCGAN network is used to generate data with a smaller number of fault samples,so that the data volume of each fault category in the data set is equal and a balanced state is achieved.The expanded dataset is used for fault diagnosis experiments,which has a high fault diagnosis accuracy.In this paper,a systematic study of the electric locomotive inverter overcurrent fault diagnosis method is conducted,and the fault signal processing method and the difficult problems in practical application are analyzed in depth,aiming to expand the application ideas of deep learning in electric locomotive fault diagnosis and lay a good foundation for digital and intelligent rail transit fault diagnosis.