Fault Diagnosis Of Fuel Cell System For Tram Based On Deep Learning

As the core of the hybrid tram,fuel cell system’s life,performance and running state have a direct impact on its commercial development.And effective fault diagnosis methods can locate and eliminate the fault timely to ensure the fuel cell system maintaining a good state.In this paper,the fuel cell/supercapacitor hybrid electric tram jointly developed by the research group and CRRC Tangshan Locomotive & Rolling Stock Co.Ltd is taken as the object,and the fault diagnosis of the fuel cell system is studied based on the data collected during the actual operation.Considering that the original data are large and has multiple features and strong feature correlation,deep learning is introduced into the fault diagnosis field.Firstly,the parameters and structure of the tram and the fuel cell system are introduced,and the characteristics through the collected data are analyzed.Since fault samples has influence on fault diagnosis,a variety of different fault samples are selected from more than200,000 sets of data,and are preprocessed by the characteristics.Sample 1 is used to study the performance of deep learning method,while Sample 2 and Sample 3 are unbalanced samples and highly correlated samples respectively,which are used to study the effect of the fault diagnosis of complex samples in order to make the results more reliable and comprehensive.Secondly,deep belief network and convolutional neural network are built respectively for sample 1 to analyze the influence of different network structure and parameters on the accuracy.And the best fault diagnosis models of the two methods are obtained by comparing the different effects.The results show that the two deep learning methods can achieve good results,and the deep belief network classification accuracy is higher than another method.In addition,the influence of different feature combinations on fault diagnosis results is analyzed compared with support vector machine.The analysis shows that deep learning methods are more applicable to the fault diagnosis field.Finally,in order to study the practicability of the proposed fault diagnosis model,the unbalanced samples are diagnosed by an improved method which can effectively improve the classification accuracy of unbalanced samples.And the fault diagnosis effect of this method is analyzed through the results.In addition,sample 3 is taken as input to study the fault diagnosis performance of the two fault diagnosis models for highly correlated samples.The practical significance of the two methods is compared and analyzed.