برآورد زمان عملیاتی سلول سوختی برای استراتژی های نگهداری پیشگویانه
|کد مقاله||سال انتشار||مقاله انگلیسی||ترجمه فارسی||تعداد کلمات|
|21852||2010||8 صفحه PDF||سفارش دهید||5057 کلمه|
Publisher : Elsevier - Science Direct (الزویر - ساینس دایرکت)
Journal : International Journal of Hydrogen Energy, Volume 35, Issue 15, August 2010, Pages 8022–8029
Durability is one of the limiting factors for spreading and commercialization of fuel cell technology. That is why research to extend fuel cell durability is being conducted world wide. A pattern-recognition approach aiming to estimate fuel cell operating time based on electrochemical impedance spectroscopy measurements is presented here. It is based on extracting the features from the impedance spectra. For that purpose, two approaches have been investigated. In the first one, particular points of the spectrum are empirically extracted as features. In the second approach, a parametric modeling is performed to extract features from both the real and the imaginary parts of the impedance spectrum. In particular, a latent regression model is used to automatically split the spectrum into several segments that are approximated by polynomials. The number of segments is adjusted taking into account the a priori knowledge about the physical behavior of the fuel cell components. Then, a linear regression model using different subsets of extracted features is employed for an estimate of the fuel cell operating time. The effectiveness of the proposed approach is evaluated on an experimental dataset. Allowing the estimation of the fuel cell operating time, and consequently its remaining duration life, these results could lead to interesting perspectives for predictive fuel cells maintenance policy.
Fuel Cells (FCs) appear to be a promising and environmentally friendly energy conversion technology for the future, especially for transport applications. However, the economical viability of FC systems, especially in the transportation sector, depends notably on improving the stack durability and reliability. Indeed, the stack is prone to material degradation (e.g. poisoning of the catalyst sites, loss of proton conductivity in the membrane, corrosion of plates, etc.) and the performance decay induced is strongly linked to the operating conditions (i.e. pollutants in the reactants, insufficient amounts of reactive gas flows versus the load current demand, operating temperature, mechanical constraints on the membrane electrode assemblies, etc.) , , ,  and . Typical life requirements range from at least 5000 h for car applications to 20 000 operating hours for bus applications. Moreover, when dealing with durability and reliability, the efficient diagnosis of FC stack and system appears as a major issue. The development of diagnostic schemes can help evaluating the FC state-of-health, and thus speed up the development cycle and deployment of new FC vehicles. Various diagnosis approaches for FC stacks and systems have been developed. They include model-based methods , , ,  and , gray or black box model approaches using fuzzy logic , design of experiment methods , neural networks , or non-parametric identification by Markov parameters . Recent FC stack diagnosis approaches based on fuzzy clustering  and Bayesian networks have been proposed . This paper presents a pattern-recognition-based diagnosis aiming to estimate the FC operating time from Electrochemical Impedance Spectroscopy (EIS) measurements done at approximately regular time intervals throughout two different Proton Exchange Membrane (PEM) FC durability tests (conducted on two stacks, noted here as FC1 and FC2). Our analysis uses data collected on FC1 and FC2 to ensure the robustness of the obtained results. Furthermore, with the proposed approach, we are also able to estimate the remaining FC lifetime. For predictive maintenance purposes, it is obviously important to know how FC performances evolve throughout its working. The complete diagnosis system consists of several steps, shown in Fig. 1. From each recorded impedance spectrum (high dimensional data), a feature extraction is performed to generate the features (low dimensional data). The goal of extraction and selection is to find a small number of features among the original ones that are particularly informative for the problem to be solved. Furthermore, dimensionality reduction is also essential when the available training dataset is small as it is in our case, otherwise the well known phenomenon of curse of dimensionality could inevitably appear, leading to over-fitting  and . Finally, linear regression between a meaningful subset of features and the considered output (operating time) is achieved. The paper is organized as follows. Section 2 describes the two ageing tests and highlights the link between FC ageing and EIS measurements. Section 3 focuses on the two different feature extraction methods that have been used. Section 4 presents the different solutions for operating time estimation based on linear regression and compares the different results obtained with the two feature extraction approaches. Experimental results are reported in this section. Section 5 concludes the paper with some perspectives.
نتیجه گیری انگلیسی
A method to estimate the operating time of FCs, using a pattern-recognition-based approach on EIS measurements, has been presented in this article. It includes two methods of feature extraction from the impedance spectrum, a feature selection procedure to keep only relevant descriptors and a regression model. While the first feature extraction approach uses hyperparameters extracted from the impedance spectrum, the second approach is based on parameterization of the real part with an external model fitting, and then in the parameterization of the imaginary part with a specific regression model incorporating a discrete hidden logistic process. Because of the small size of the available dataset, a particular attention has been paid to the feature extraction and selection steps. Evaluations on real dataset have shown that FC operating time (respectively lifetime) can be estimated with a mean error of 214 h over a global operating duration of 1000 h, when one uses the first feature extraction method, whereas the mean error equals 142 h when using the second method and falls to 95 h, when using features extracted using both methods. These are encouraging results, particularly in predictive maintenance context. Improvements can be achieved either by the use of additional features extracted from another characterization measurement (such as polarization curve) and/or more exhaustive dataset.