An efficient neural-network-based image processing method for water quantification in a transparent proton exchange membrane fuel cell

Water and thermal management are pivotal to the performance of proton exchange membrane fuel cells. This paper presents the design of a transparent single fuel cell with an active area of 25 cm2 for characterising water distribution under various operating conditions. In the design and assembly of batteries, the proposed design overcomes the challenges in battery sealing. Water was quantified using a neural network that analysed videos recorded under various operational conditions frame-by-frame. For comparative analysis, a threshold processing method was employed, and its advantages and disadvantages were discussed in detail. A high-quality training set comprising 137 frames derived from the threshold processing results was employed for the neural network training. This study investigated the impacts of temperature, voltage, and flow field design on water accumulation. The neural-network-based semantic segmentation method demonstrated superior recognition, adaptability, and sensitivity to liquid water under complex operating conditions. It was found that a square bender was more likely to accumulate water than a semicircular corner bender in the serpentine flow channel in the early stage, whereas the difference in the flow channel had almost no effect on the steady stage. Furthermore, there was no evident linear relationship between cell performance and water cover ratio.