Study On Fault-embedded Modeling And Data-driven Fault Diagnosis For Proton Exchange Membrane Fuel Cell

Proton exchange membrane fuel cell has attracted much interest due to its advantages of high power density and low operating temperature.To ensure the safety and stability of the fuel cell system,it is important to develop a complete process monitoring system and fault diagnosis system.As a nonlinear,dynamic,and timevarying system,the fuel cell system involves the fuel cell stack and multiple auxiliary subsystems including reactant gas supply,water and heat management subsystems.This increases the system complexity and failure possibility,and also increases the difficulty of the online real-time diagnosis.In the thesis,a fault-embedded model is developed,and an end-to-end fault diagnosis method framework is proposed based on a data-driven approach.Firstly,this paper develops a fault embedding model based on the quasi-twodimensional transient model of the proton exchange membrane fuel cell,including local resistance increased,inlet pressure decreased,insufficient reactant flow,insufficient cooling flow,and insufficient inlet humidity.Dynamic performance response analysis with the help of this model for system prediction and health management.At the same time,The model can help accumulate f data to address the problems of limited data samples and unbalanced data types.Besides,the model is utilized for sensitivity analysis under different fault modes to help the development of fault diagnosis method framework below.Secondly,two sensor optimization selection methods are proposed in this paper to solve the problem of high-dimensional data generated by multiple sensors in complex systems.The one is based on the fuel cell model developed in the previous chapter.The effect of various common failure modes on the system is considered,and the health parameters are selected based on failure mechanism.Then the sensitivity matrix of sensors to the system health parameters are calculated.The optimal sensor set is determined with the help of adaptive network-based fuzzy inference system.While the sensor selection based on signal analysis relies on the system historical data.The features of time domain and frequency domain are extracted for data information analysis.Signal processing methods such as wavelet packet transform to complete sonsor selection.The results show that the optimization results of the two methods are consistent.In practical application,we can choose from two methods according to the degree of data accumulation or model development.Finally,the experimental data monitored by the remaining sensors are utilized to establish a high-precision fault diagnosis model.In this paper,the LevenbergMarquardt algorithm is used to train the artificial neural network classification model,which demonstrates faster convergence speed and smaller computational error than scaled conjugate gradient and bayesian regulation.It is verified that the final diagnosis accuracy rate reaches 99.2%,precision rate reaches 99.59%,and the recall rate reaches98.3%,which is significantly better than the two commonly used methods of support vector machine and logistic regression.Meanwhile,the diagnostic results is better than the neural network model without sensor selection.It confirms that the proposed sensor selection method is effective;with an optimal sensor set,different fuel cell states can be discriminated with better quality.In addition,the factors affecting the diagnosis results in the data-driven approach are discussed and analyzed,as well as their practical impact.The important impact of data balance on the recall rater is emphasized.In conclusion,the fault diagnosis method proposed in this paper ensures the reliability of diagnosis results,and reduces the computational cost significantly.