Convolutional neural network analysis of radiography images for rapid water quantification in PEM fuel cell

Abstract

Polymer electrolyte membrane (PEM) fuel cells produce water as a byproduct, which, if not properly managed, will cause electrode “flooding” and consequently performance loss. Nonintrusive methods, such as neutron or X-ray radiography, have been employed to obtain in-situ water images in PEM fuel cell. This study presents one of the first studies developing a machine learning approach to analyze neutron radiography images using the convolutional neural network (CNN), a deep learning model, to quantify liquid water content in PEM fuel cell. The CNN model is trained using the labeled radiography images constructed from the contour legend, which contains the information of the water areal mass density. Image enhancement is carried out to generate additional data for CNN training. The properly-trained CNN model is subsequently applied to the radiography images to obtain the average water areal mass densities under various current densities, relative humidity (RH), and flow fields. The results show that the water content can be significantly reduced by using low RH inlet flow and the counter-flow configuration renders the fuel cell a higher water content than the co-flow one. In addition, the quad-serpentine flow field increases the water content in the current density range of 0.4–0.8 A/cm2 compared with the single-serpentine one. The CNN model takes less than 0.1 s to analyze an image and its results agree well with the literature data from the conventional image processing method with an accuracy of 91%. The CNN method is applicable to other radiography methods and is suitable for the online/rapid monitoring of liquid water in PEM fuel cell.