Pre-diagnosis of flooding and drying in proton exchange membrane fuel cells by bagging ensemble deep learning models using long short-term memory and convolutional neural networks

Polymer electrolyte membrane fuel cells (PEMFC) are a prevalent power source in transportation because of their ability to generate energy at low temperatures without harmful emissions. However, problems related to water management cause performance and durability degradation. The main faults are flooding, whereby the performance suffers owing to the stagnated water in gas diffusion paths and catalyst layers, and drying, which increases the ohmic loss owing to water evaporation in the electrolyte or insufficient water supply. Difficulties in recognizing these faults and normalizing operations impair the PEMFC stability. However, detecting errors in advance contributes to maintaining normal operation. Therefore, a system that diagnoses flooding and drying of the PEMFC before they occur is developed in this study using deep learning. The characteristics of flooding and drying are analyzed through preliminary experiments. Experimental data in the form of a time series are accumulated through a full-scale single-cell test. A pre-diagnosis system, developed using long short-term memory (LSTM) and a convolutional neural network (CNN), is reinforced through the bagging ensemble method. The expandability of the target future time and real-time system applicability are discussed. The detection rates achieved by the proposed system for flooding and drying that occur after 30 s are 98.52% and 95.36%, respectively.