A Data-Driven Prediagnosis Framework for Early Detection of Water Imbalance in PEMFC

This paper presents a pre-diagnosis approach for identifying water imbalance faults in proton exchange membrane fuel cells (PEMFCs) system, aiming to provide early warning alarms to support proactive protection measures. Unlike traditional fixed-model methods, the proposed approach introduces a dynamic framework to handle the complexity of PEMFC fault conditions with improved efficiency and accuracy. First, the minimal redundancy maximum relevance (mRMR) algorithm is employed for feature selection to optimally balance input data quality and quantity. Subsequently, multiple encoder layers enhanced by a mixture of experts mechanism are utilized to extract key information. By dynamically activating only the most relevant experts, the model’s adaptability to diverse input is improved, and computational load is reduced. Next, an adaptive multi-task learning (MTL) framework is developed to train time-series forecasting and fault classification tasks simultaneously. Finally, the proposed method is validated on a 5kW PEMFC stack under various operating conditions. Experimental results demonstrate that the proposed method achieves rapid and accurate identification of diverse PEMFC water management faults, detecting anomalies 25 s before occurrence with approximately 98.0% diagnostic accuracy at a fast response time of 1.15 ms.