Multiphysics Field Virtual Sensors for PEM Fuel Cells Based on a Deep Learning Method

Abstract

In transportation and stationary power generation applications, an accurate and reliable fuel-cell condition monitoring system is essential to achieve efficient and long-life fuel-cell system operation. However, due to the structure of bipolar plates and membrane electrodes, it is difficult to directly arrange sensors to obtain internal information. In this article, a deep learning-based virtual sensor construction method for fuel cells is proposed, which utilizes only external sensors to reconstruct the physical distribution inside the fuel cell. The U-Net is employed to encode and decode the original signals, while the spatial attention mechanism enables the deep neural network to focus on the effective information in the spatial distribution. In addition, the virtual sensor responds well to flooding in the flow channels at high current densities. The results show that the structural similarity (SSIM) and peak signal-to-noise ratio (PSNR) of the multiphysics field reconstruction are 0.9599 and 37.24, respectively. In the real fuel-cell test, the SSIM for multiphysics field reconstruction reached 0.9621. Additionally, the model’s effectiveness was further validated through comparisons with existing deep learning models. This work effectively advances the design of fuel-cell condition monitoring systems and contributes to the construction of high-precision fuel-cell digital twins.