Severity-based fault diagnostic method for polymer electrolyte membrane fuel cell systems

This research suggests a severity-based multi-stage fault diagnostic method for polymer electrolyte membrane fuel cell systems. By separating the diagnostic stage depending on the fault severity, robustness and sensitivity of the diagnosis algorithm on each stage can be designed with flexibility. The details of the fault diagnosis algorithm development process are described and validated with fault experimental data using a 1 kW class fuel cell system. First, a nominal model is developed to generate a residual between the predicted normal state and the observed state. Second, expected fault responses of the system are organized in the form of residual patterns. These residual patterns are used for training neural networks that diagnose critical faults, significant faults, and minor faults. Third, the generated residuals are standardized and moving averaged to be used as inputs for the neural networks at each stage. Lastly, diagnosis results from the neural network-based algorithm are compared with the fault experimental data. As a result, 17 different faults are all successfully diagnosed. More specifically, five critical faults, seven significant faults, and 13 minor faults are diagnosed. In addition, a diagnosis method for multi-faults is suggested. Double faults and triple faults are experimentally simulated and diagnosed with the diagnosis algorithm.