针对质子交换膜燃料电池气体扩散层(gas diffusion layer composition，GDL)形貌划分与制备工艺改进问题，提出了一种基于金字塔池化网络(pyramid scene parsing network，PSPNet)与多层感知器(multi-layer perception，MLP)的气体扩散层组分识别与比例推理方法：首先将带标签的气体扩散层扫描电镜(scanning electron microscope，SEM)图片输入神经网络，得到特征图；得到的图像特征层进入金字塔池化模块后，获取SEM图像的深层和浅层特征；随后将深层和浅层特征图层融合输入全卷积网络(fully convolutional network，FCN)模块，得到预测图像；最后统计各个组分上的像素点比例，通过MLP完成组分比例推理。结果表明：所提方法组分识别像素准确率达81.24%；在5%偏差范围内，比例推理准确率为88.89%。该方法解决了气体扩散层多组分无法区分、比例无法获知的问题，可有效应用于气体扩散层的质检、数值重构以及制备工艺改进。

To rapidly determine the morphology of the gas diffusion layer for proton exchange membrane fuel cell (PEMFC) and improve its fabrication process, a method of gas diffusion layer (GDL) component identification and proportional reasoning based on a combination of pyramid scene parsing network (PSPNet) and multilayer perceptron (MLP) is proposed. First, labeled GDL scanning electron microscope （SEM) images are input into the neural network to obtain a feature extraction map. This map is used in the pyramid pooling module to extract both deep and shallow features of the SEM images. Subsequently, these feature layers are input into the fully convolutional network (FCN) module to produce a predicted image of the same size. Finally, the proportion of pixels for each component is calculated, and the inference of component proportion is achieved by using the MLP. The accuracy of the proposed method is 81.24%, with an accuracy of proportional reasoning reaching 88.89% within a 5% deviation range. The proposed method can be effectively used for gas diffusion layer quality detection, numerical reconstruction, and process improvement.

TK91

国家自然科学基金青年项目（12102188）；广东省重点领域研发计划项目（2019B090909003）；先进能源科学与技术广东省实验室佛山分中心（佛山仙湖实验室）开放基金（XHD2020-004）。