A Multi-Stack PEMFC Fault Diagnosis Algorithm Based on Multi-Head Cross-Attention

This study focused on prospective research of the yet-to-be-commercialized Multi-Stack Fuel Cell System (MFCS) under specific topological structures and stack configurations. A proficient fault diagnosis system holds the potential to enhance the operational lifespan of MFCS. Furthermore, membrane drying faults represent one of the highest priority issues among typical failures in fuel cell systems. So this study delved into the multi-level membrane drying fault diagnosis issue within MFCS. Utilizing data-driven fault diagnosis methods, we constructed a simulation MFCS model using AVL CRUISE M software. We designed corresponding boundary conditions for multi-level membrane drying faults to generate simulated datasets. To address this diagnostic challenge, we introduced an innovative approach called the Multi-Head Cross-Attention Fault Diagnosis Algorithm. In this method, sensor and actuator signals from multiple stacks were transmitted to a multi-head cross-attention module for attention calculation. The integrated output was then sent to three back propagation (BP) neural networks (one for each stack). These three BP sub-diagnosis modules were interconnected in a chain structure, collectively producing the fault diagnosis results. In this study, we conducted a detailed comparative analysis of the number of attention heads and tested the optimal combination of hidden layer nodes for each diagnosis unit in the chain model to obtain the optimal model structure for our specific research scenario. Following testing on simulated data, the Multi-Stack Multi-Head Cross-Attention Fault Diagnosis Algorithm achieved a diagnostic accuracy of 0.945 and a recall rate of 0.954, outperforming back propagation neural networks (BPNN), LSTM combined with BPNN (LSTM+BPNN), and support vector machine (SVM) algorithms trained on the same dataset.