Fault Diagnosis Based On Information Fusion And Deep Learning For Fuel Cell In Tram

In order to facilitate the commercial application of high-power proton exchange membrane fuel cell,it is very important to ensure the safe and reliable operation of fuel cell system.Therefore,it is of great significance to explore effective fault diagnosis methods to improve the safety and reliability of high-power proton exchange membrane fuel cell.In this paper,the research object is the proton exchange membrane fuel cell system of the fuel cell/ultracapacitor hybrid 100% low floor tram jointly developed by Southwest Jiaotong University and CRRC Tangshan Co.,Ltd.Based on the historical data of the tram,the datadriven fault diagnosis method is studied to realize the fault diagnosis of proton exchange membrane fuel cell in tram.The specific research contents are as follows:Firstly,this paper introduces the appearance and parameters of the hybrid tram,as well as the structure and principle of the proton exchange membrane fuel cell system in tram.Then the electric and non-electric parameters that can be monitored by the data acquisition system for proton exchange membrane fuel cell are explained.The characteristics of different typical monitored quantities are analyzed based on the variation of the typical monitored quantities acquired in the course of running of the hybrid tram.Then,the fault levels and fault types of the proton exchange membrane fuel cell system in tram are described in detail,and the typical monitored quantities of fault samples are compared with those of normal samples.The coupling and randomness among fault sample data has been verified from the results.Secondly,a fault diagnosis framework based on GRU based on information fusion in data layer is designed for complex fault sample data set.The t-SNE dimensionality reduction visualization algorithm is used to illustrate the complexity of the sample data set,and the model parameters such as appropriate GRU layers,optimizer type and data preprocessing strategy are selected through comparative analysis.The results show that the method can identify five types of sample data including normal state well.In addition,the diagnostic results with 28,14 and 11 dimensions as the inputs of the model are compared,and it is verified that the proposed method has better diagnostic accuracy with more dimensional inputs.The proposed method is compared with the test results of SVM algorithm,which further demonstrates the feasibility of the proposed method.Thirdly,a fault diagnosis framework based on BPNN-Inception Net based on information fusion in feature layer is proposed for unbalanced fault sample data set.In this framework,a BPNN is first used to raise the dimension of input data into higher-dimensional abstract features,and then feature reconstruction technology is applied to transform the abstract features into feature maps.Finally,a convolutional network based on Inception Net is used to classify feature maps.The diagnostic results show that the kappa coefficient of the proposed method is higher than that of BPNN-VGG,GRU and SVM models for the unbalanced data sets with seven state classes,which proves the diagnosing effectiveness of the proposed method for the unbalanced data sets.Finally,a multi-model fusion fault diagnosis framework based on D-S evidence theory based on information fusion in decision level is introduced for a coexisting fault data sets containing a few samples.In this framework,the prediction outputs of two BPNNInception Net models are fused by using D-S evidence theory combination rules,and the results of an example show that the proposed method is feasible in obtaining comprehensive diagnosis decisions.