Research On Intelligent Fault Diagnosis Technology Of PEMFC Based On Mechanism Model And Data Driven

Countries around the world keep investigating approaches on reducing carbon dioxide emissions to limit rising temperatures.Among these measures,fuel cells have the advantage of fuel cells including short hydrogen refueling time and long driving mileage and appear promising for commercial road vehicles.Developing online fault diagnosis technologies of proton exchange membrane fuel cell(PEMFC)for vehicle applications can improve their durability and sustainability and hence to advance their applications in commercial vehicles.Conventional online fault diagnosis approaches apply cell voltage monitoring(CVM)or electrochemical impedance spectroscopy(EIS).However,CVM is unable to reflect the dehydration,flooding and gas starvation while EIS is vulnerable to external interference and the computation time for low-frequency impedance is relatively long which is not appropriate for real-time online diagnosis.A PEMFC system is a typical distributed parameter system,where developing transitive relationships between PEMFC system signals and internal faults ensures rapid,accurate and cost-effective fault diagnosis.Therefore,this thesis aims to focus on four aspects sequentially,including simulation of fault dataset,optimization of intelligent fault diagnosis algorithm,practical technologies for online fault diagnosis and self-adaptive fault diagnosis.First of all,a quasi-two-dimensional simulation model is developed for a typical PEMFC system to investigate the transitive relationship between system signals and internal states of PEMFC.Based on the novel application of the proposed simulation model and the fault tree analysis,the current signal and different fault types are simulated for vehicles under dynamic conditions.Besides,a multi-classification and multi-level fault dataset covering dehydration(mild/severe),flooding(mild /severe),and gas starvation(mild/severe)is built for the following data-driven fault diagnosis research.Secondly,the approaches for fault diagnosis are analysed and investigated according to the fault diagnosis simulation dataset.An improved one-vs-all classification algorithm based on support vector machine(SVM)with Gaussian kernels and penalty factors is designed for the nonlinear seven-type fault classification problem.This algorithm enables fast convergence and high precision and recall,both higher 85%.Besides,to improve accuracy and real-time performance of the classification algorithm,an improved back propagation(BP)algorithm based on cross-entropy classification loss functions is applied with one-hot encoding of input nodes and the regularization process before each layer of the network.In addition,to further improve the accuracy of diagnosis,a comparative analysis is conducted by the application of the wavelet packet energy algorithm.A three-layer wavelet packet decomposition structure combined with the BP full connection layer network structure is established by introducing logarithmic energy entropy to improve the discrimination of energy distribution.The precision and the recall reach 89% but the real-time performance decreases significantly.Therefore,to increase the accuracy and real-time performance,an optimised application of a 100-layer time series long short-term memory(LSTM)algorithm and the BP full connection layer network ensures the highest precision and recall up to 99% based on appropriate activation functions and network structures.The detailed analysis demonstrates that the optimized LSTM-BP algorithm for the PEMFC fault diagnosis problem entails dominant advantages in robustness,accuracy,real-time performance and storage space.Furthermore,to achieve online multi-classification and multi-level fault diagnosis,a high-power PEMFC text bench with an intelligent controller is newly developed for the validation and verification of the proposed online fault diagnosis approach.The efficiency of the LSTM algorithm is improved by using the pruning optimization of weights and the approximation optimization of low rank matrix.The experimental results suggest that the online multi-classification and multi-level fault diagnosis methods are reliable.The real-time performance and memory occupation are improved by 31.41% and 31.23% respectively.Lastly,based on the analysis of the aging mechanism of PEMFC physical parameters,a model is simulated for the dynamic PEMFC aging system.Accordingly,a self-adaptive fault diagnosis algorithm is developed based on the LSTM and the semisupervised learning method to improve the accuracy of classification by 5.1% and 7.7%from the simulation and experimental results respectively.The experimental results indicate that the proposed self-adaptive fault diagnosis algorithm maintains high accuracy of classification and thus has important practical application value for the PEMFC fault diagnosis system throughout the life cycle.