质子交换膜燃料电池以其无污染、高效率、低噪音等优点，受到车辆、船舶以及航空航天等领域的重点关注，然而其可靠性问题阻碍了燃料电池的大规模商业化推广。为进一步提升燃料电池的可靠性，该文提出了一种基于深度学习的燃料电池故障预测方法。首先，针对电压、电流、湿度、温度等状态监测数据，根据燃料电池故障机理选取用于故障诊断的特征参数，以降低数据维度、抑制冗余信息，提高故障预测模型的运算效率，并采用标准化、时间滑窗等预处理方法，消除监测参数之间量纲不同的影响；然后，构建基于长短时记忆网络的燃料电池状态预测模型，其输入为预处理后的多维特征序列，输出为未来T个时间步长的燃料电池状态；最后，将预测的状态数据输入至卷积神经网络故障辨识模型，实现燃料电池故障状态预测。使用燃料电池实验故障数据对所提出的方法进行验证，结果表明，所提出的故障预测模型能够提前预测到故障。该文所提基于深度学习的燃料电池故障预测方法通过有效的数据预处理、基于长短时记忆网络的未来状态预测及基于卷积神经网络的故障识别，实现了对质子交换膜燃料电池运行异常的提前预测。

Proton exchange membrane fuel cells (PEMFCs) have attracted significant attention in the fields of transportation, marine engineering, and aerospace due to their advantages of pollution-free operation, high efficiency, and low noise. However, reliability issues hinder their large-scale commercialization. To further enhance fuel cell reliability, this paper proposed a fault prediction method based on deep learning. First, for operational monitoring data including voltage, current, humidity, and temperature, feature parameters for fault diagnosis were selected based on fuel cell failure mechanisms. This approach reduces data dimensionality, suppresses redundant information, and improves the computational efficiency of the prediction model. Additionally, pre-processing techniques such as normalization and sliding time windows were employed to eliminate the effects of differing dimensions among monitoring parameters. Then, a fuel cell state prediction model based on the long short-term memory (LSTM) network was constructed. Its inputs were preprocessed multidimensional feature sequences, and its output predicts the fuel cell state for the next T time steps. Finally, the predicted state data was fed into a convolutional neural network (CNN)-based fault identification model to achieve fuel cell fault state prediction. The proposed method was validated using experimental fault data from fuel cell tests, and the results show that the model can predict failures in advance. By virtue of effective data preprocessing, future state prediction via LSTM, and fault recognition through CNN, this deep learning-based approach enables early prediction of operational anomalies in proton exchange membrane fuel cells.