Estimation of Water Coverage Ratio in Low Temperature PEM-Fuel Cell Using Deep Neural Network

Abstract

Proton exchange membrane fuel cell (PEMFC) is a rich source of renewable energy. A non-destructive prediction method is needed to determine the content of water in the PEMFC. In the gas channel of a transparent PEMFC, water is detected with image processing. This method has a high computational cost and is sensitive to the initial position of the camera and ambient lighting. In this paper, the deep neural network (DNN) has been trained to learn the transparent PEMFC’s labeled images as a way to determine the content of water, limit human interference and employed in a real-time process. This DNN model is a virtual sensor for measuring the water coverage ratio. To produce the label of images, all data are divided into 6 classes based on the percentage of water coverage ratio. Through analyzing the number of each class, the imbalance data is unfolded. To overcome this problem, random oversampling and undersampling techniques are used. The images and the classes are considered as the input and output of the DNN, respectively. Also, the region of water accumulation in the gas channel can be recognized with robustness to the environmental conditions. Final results of 4-41-64-120-99-6 nodes for DNN layers were derived due to GA optimization. Accuracy of 96.77% and 94.23% in train and test data have been achieved. This DNN, processes the images up to 10 times faster than image processing. Also, the region of water accumulation in the gas channel can be recognized.