Data-based flooding fault diagnosis of proton exchange membrane fuel cell systems using LSTM networks

Abstract Flooding fault diagnosis is critical to the stable and efficient operation of fuel cells, while the on-board embedded controller has limited computing power and sensors, making it difficult to incorporate the complex gas-liquid two-phase flow models. Then in fuel cell system for cars, the neural network modeling is usually regarded as an appropriate tool for the on-line diagnosis of water status. Traditional neural network classifiers are not good at processing time series data, so in this paper, Long Short-Term Memory (LSTM) network model is developed and applied to the flooding fault diagnosis based on the embedded platform. Moreover, the fuel cell auxiliary system statuses are adopted as the inputs of the diagnosis network, which avoids installing a large number of sensors in the fuel cell system, and contributes to reduce the total system cost. The bench test on the 92 kW vehicle fuel cell system proved that this model can effectively diagnose/pre-diagnose the fuel cell flooding, and thus help optimize the water management under vehicle conditions.