Data-driven Fault Diagnosis And Life Prediction Methods For PEMFC Power Generation System

Proton exchange membrane fuel cell(PEMFC)is a new type of clean energy that converts chemical energy into electrical energy.It has the advantages of environmental protection,pollution-free,high energy conversion rate,and low noise.It has gradually been used in domestic cars,buses,and other transportation fields.However,defects such as short life,easy failure,and high cost have been the limiting factors that hinder its large-scale commercialization.In order to ensure the reliable operation of the PEMFC system and increase its service life,it is urgent to study the method of PEMFC fault diagnosis and life prediction.The main findings of this paper are shown as follows:(1)Aiming at the problem of PEMFC water management subsystem fault diagnosis,a PEMFC water management fault diagnosis method based on probabilistic neural network(PNN)and linear discriminant analysis(LDA)is proposed,which can significantly improve the accuracy of diagnosis and reduce computational complexity.This method uses normalization and LDA to eliminate dimensions and reduce dimensions of the original experimental data,which can effectively extract the fault feature vector and reduce the computational complexity.Using PNN to implement PEMFC water fault diagnosis on the fault feature samples can significantly improve the fault diagnosis accuracy.17086 sets of PEMFC water management fault sample data sets(including normal state,water flood fault and membrane dry fault)were used to experimentally validate the viability of the novel approach.The BP-NN and LDA-BPNN are compared with the proposed method to moreover proof of the capability of the novel approach.(2)Aiming at the fault diagnosis of the PEMFC air supply subsystem,a novel method for fault diagnosis of the PEMFC system based on data fusion is proposed.Feature extraction of electrical and non-electrical quantities of the fuel cell system under different faults,respectively using the kernel extreme learning machine(K-ELM)algorithm and online sequential extreme learning machine(OS-ELM)algorithm to establish a fuel cell system fault diagnosis model based on electrical quantities and non-electrical quantities for preliminary failure diagnosis of fuel cell systems.The squeezing function is used to convert the diagnosis results of the two strategies into basic probability assignment(BPA)function values,and the D-S evidence theory algorithm is used to perform fusion diagnosis output at the decision level.The feasibility of the novel method is verified by 154 sets of measured sample data composed of four different degrees of high air excess coefficient faults.Compared with the traditional one-against-one SVM algorithm and BP neural network algorithm,the recognition results are analyzed and compared to verify the effectiveness of the novel algorithm.(3)Aiming at the problem of high-power water-cooled PEMFC system fault diagnosis,a discrete hidden Markov model(DHMM)fault diagnosis strategy for PEMFC system for trams based on K-means clustering is proposed.It has high recognition accuracy and good scalability for various health states.K-means clustering algorithm is used to eliminate singular sample points.The Lloyd method is used to quantify the sample vector set to obtain the discrete code combination of the training samples and the test samples.The Baum-Welch algorithm and the forward-backward algorithm are used to train and infer DHMM,respectively.The feasibility of this strategy is verified using the experimental data of trams,and the superiority of this method is further proved by comparison with the one-against-one method SVM.(4)Aiming at the steady-state life prediction problem of PEMFC stack,a PEMFC life prediction method of long short-term memory(LSTM)recurrent neural network(RNN)and locally weighted scatterplot smoothing(LOESS)can greatly shorten the complexity of computation while engaging the accuracy of the prediction.This method uses equally spaced sampling and LOESS to achieve data reconstruction and smoothing of the original aged samples,uses normalization to filter the samples,and uses LSTM-RNN to achieve life prediction on the test set.Under [0 h,1154 h],the validity of the novel method was verified by 143862 sets of 1 k W PEMFC system steady-state operation experimental data.Compared with the prediction results of the BP neural network algorithm to further verify the effectiveness of the novel method.(5)Aiming at the transient life prediction problem of PEMFC stack,a PEMFC life prediction method based on sparse autoencoder(SAE)and deep neural network(DNN)is proposed.This method extracts the data set from the original experimental data and implements data reconstruction at an hourly interval.A Gaussian weighted moving average filter is used to smooth the noisy data(stack output voltage and current).The smoothed and filtered output signal of the stack is used as an aging indicator.Use SAE to automatically extract prediction features,and use DNN to realize life prediction.The proposed method is experimentally verified using 127369 sets of experimental data.The effectiveness of the novel method is verified by using three different training and test set configurations.The superiority and effectiveness of the proposed method are further verified by comparison with the K-nearest neighbor algorithm and support vector regression machine algorithm.