A novel densely connected neural network for proton exchange membrane fuel cell fault diagnosis. (5th December 2022)

• Record Type: Journal Article
• Authors: Liu, Zhongyong; Mao, Lei; Hu, Zhiyong; Huang, Weiguo; Wu, Qiang; Jackson, Lisa
• Is Part Of: International journal of hydrogen energy. Volume 47:Number 94(2022)
• Page Start: 40041
• Page End: 40053
• Publication Date: 2022-12-05
• Subjects: PEMFC -- Deep learning -- Fault diagnosis -- Water management faults -- Air starvation; Hydrogen as fuel -- Periodicals; Hydrogène (Combustible) -- Périodiques; Hydrogen as fuel; Periodicals; 665.81
• DOI: 10.1016/j.ijhydene.2022.09.158
• Languages: English
• ISSNs: 0360-3199

Abstract

Abstract: It is of great significance to perform proton exchange membrane fuel cell (PEMFC) fault diagnosis and take action timely to mitigate or even eliminate the faults, which can strengthen PEMFC reliability and durability. In previous studies, cell voltage is extensively used for PEMFC fault diagnosis. However, there exists similar cell voltage drop phenomenon as different PEMFC faults occur, especially for faults like flooding and air starvation having extremely similar voltage dynamic variation, which makes it difficult to capture the features sensitive to faults. Moreover, cell voltages collected from different MEAs follow different distributions even in the same operation condition, which challenges the diagnosis consistency of fault diagnosis methods. In this paper, in order to break through the hindrances, a novel densely connected neural network codenamed Inc-DenseNet is proposed for PEMFC fault diagnosis, which integrates advantages of InceptionNet and DenseNet to extract more specific and robust features from cell voltage. In the analysis, the collected PEMFC voltage signal is transformed into 2D image data, which is then used to train the Inc-DenseNet. Results demonstrate that with the trained Inc-DenseNet, the diagnostic accuracy for four PEMFC states of health (normal, flooding, dehydration, air starvation) can reach 95.3%, especially for flooding and air starvation. In addition, by using the voltage datasets collected from two different MEAs, the generalization capacity of the Inc-DenseNet is proved. With the findings, the proposed network Inc-DenseNet can not only achieve high-precision fault diagnosis, but also has a high computing efficiency, which makes it promising in real-time PEMFC fault diagnosis in the future. Highlights: A novel densely connected neural network codenamed Inc-DenseNet is developed for PEMFC fault diagnosis. More abstract and discriminant features can be extracted for PEMFC fault diagnosis. Voltage datasets from two different MEAs are collected to verify the generalization capacity of the Inc-DenseNet. High computational efficiency and high-precision diagnosis accuracy can be achieved.