A Novel Approach to Flooding Fault Detection and Risk Assessment in PEM Fuel Cells Using Data-Driven Models

The reliable and efficient operation of polymer electrolyte membrane fuel cells (PEMFCs) necessitates the implementation of preventive strategies and maintenance protocols to minimize the likelihood of failures. To address this, the study identifies effective diagnostic techniques for detecting faults in PEMFCs. A data-driven approach leveraging machine learning methods is proposed to enhance the detection of flooding faults under varying operational conditions. This approach enables the automatic extraction of fault-related features directly from raw data. Experimental data obtained from an 80 W PEMFC, widely used in the literature for comparability, was utilized in the study. Various machine learning classification algorithms were applied, and their performance metrics were analysed. Among these, the random subspace k-nearest neighbor algorithm demonstrated superior accuracy and the shortest training time, leading to the development of a novel model. The study evaluated 22 variables associated with PEMFCs, performed fault diagnosis, and assessed fault severity. Furthermore, a risk analysis was conducted using the proposed model, enabling the prediction of both the risk level and the probability of fault occurrence as percentages. Key performance metrics, including accuracy, sensitivity, precision, and specificity, were calculated as 99.97%, 99.98%, 99.90%, and 99.98%, respectively, during model validation. During testing, these metrics were recorded as 99.45%, 100%, 98.42%, and 99.16%, respectively.