引入自监督机制的燃料电池水管理系统健康状态检测方法  

A Health State Detection Method of Fuel Cell Water Management System Withthe Introduction of Self-supervision Mechanism

作　　者：张译丹 李奇[1] 尹良震 刘述奎 陈晓雯 陈维荣[1] ZHANG Yidan;LI Qi;YIN Liangzhen;LIU Shukui;CHEN Xiaowen;CHEN Weirong(School of Electrical Engineeringm,Southwest Jiaotong University,Chengdu 610031,Sichuan Province,China;State Grid Sichuan Electric Power Company Chengdu Power Supply Company,Chengdu 610000,Sichuan Province,China)

机构地区：[1]西南交通大学电气工程学院,四川省成都市610031 [2]国网四川省电力公司成都供电公司,四川省成都市610000

出　　处：《中国电机工程学报》2023年第13期5025-5036,共12页Proceedings of the CSEE

基　　金：国家自然科学基金项目(51977181,52077180);四川省自然科学基金(2022NSFSC0027);霍英东教育基金会高等院校青年教师基金(171104)。

摘　　要：为保证质子交换膜燃料电池(proton exchange membrane fuel cell,PEMFC)运行过程中的安全性和平稳性,该文提出一种引入自监督机制的燃料电池水管理系统健康状态检测方法。该方法的计算网络引入自监督机制挖掘数据特征,具体通过堆叠两层稀疏收缩自编码器(sparse contractive autoencoder,SCAE)实现功能。该自编码网络以逐层贪婪训练的方式完成参数初始化,避免因初值取值不佳使得网络参数训练陷入局部最优解。同时,改进的自编码网络也能有效避免模型过拟合,从而达到良好的特征提取效果。用Softmax分类器替换最后一个自编码器的输出层,然后运用Adadelta算法进行基于自适应学习率的自编码网络权值微调,从而完成系统识别网络的搭建。该算法根据梯度动态调节学习率大小,使得网络权值快速逼近符合数据特点的最佳取值。实验结果表明,该方法能快速准确地识别PEMFC正常、水淹和膜干3种状态,检测正确率高达98.5%,检测时间为3.24s。与线性判别分析-概率神经网络(linear discriminant analysis-probabilistic neural network,LDAPNN)、稀疏自编码器-支持向量机(sparse autoencodersupport vector machine,SAE-SVM)和主成分分析-向后传播神经网络(principal component analysis-backpropagation neural network,PCA-BPNN)方法相比,所提方法计算时间分别减少3.08、5.38和7.15s,准确率分别提高7.92、25.08和9.08%。普适性验证表明,该方法对于多节电池的健康状态检测同样适用。To ensure the safety and stability of the proton exchange membrane fuel cell(PEMFC)operation process,this paper proposes a health state detection method of fuel cell water management system that introduces a self-supervision mechanism.The self-supervision mechanism is introduced by the computational network of this method to mine data features,specifically by stacking two layers of sparse contractive autoencoders(SCAE)to achieve the function.The self-encoding network completes the parameter initialization in the way of layer-by-layer greedy training,which avoid the network parameter training falling into the local optimal solution due to the poor initial value.At the same time,the improved autoencoding network could also effectively avoid model overfitting to achieve a good feature extraction effect.The output layer of the last autoencoder is replaced with the softmax classifier,and then the adadelta algorithm is used to fine-tune the weights of the autoencoder network based on the adaptive learning rate to complete the system identification network construction.The algorithm dynamically adjust the learning rate according to the gradient,so the network weight quickly approximates the best value that meets the characteristics of fault data.The results show that the method could quickly and accurately identify the three states of PEMFC normal,flooded and membrane dry.The recognition accuracy rate is 98.5%,and the calculation time is 3.24s.Compared with the LDA-PNN,SAE-SVM and PCA-BPNNmethods,the calculation time of the method described in thispaper is reduced by 3.08,5.38 and 7.15s,respectively,and theaccuracy rate is increased by 7.92%,25.08%and 9.08%,respectively.The universality verification shows that themethod is also applicable to the health state detection ofmulti-chip fuel cells.

关 键 词：燃料电池 健康状态 自监督机制 稀疏收缩自编码器 自适应学习率 

分 类 号：TM911[电气工程—电力电子与电力传动]