高压氢气泄漏是加氢站常见的安全隐患之一，严重影响加氢站的安全运行。准确且及时地识别泄漏位置，密切检测氢气体积分数，对于加氢站的燃爆安全防控至关重要。该文提出一种基于深度神经网络的氢气泄漏智能定位检测模型，为加氢站提供智能且及时的检测方案。首先利用计算流体动力学(CFD)模拟构建了不同泄漏位置、泄漏强度、风向的加氢站高压氢气泄漏的专有数据库，分析发现风向对氢气泄漏的影响最大；比较了6种现有深度学习的检测模型针对氢气泄漏位置的预测效果，发现基于卷积双向长短期记忆神经网络(CNN-BiLSTM)的模型准确率和F1分数均超过98%，展现了优异的检测性能和显著的鲁棒性。该研究为加氢站的氢泄漏检测提供了理论基础，也为智慧消防技术的实现提供了行之有效的解决方案。

Objective: High-pressure hydrogen leakage is a common safety concern in hydrogen refueling stations, significantly affecting the safe operation of these facilities. Accurate and timely identification of the leak source location and continuous monitoring of hydrogen concentration are essential for preventing explosions and ensuring the safety of refueling operations. Methods: In this study, we propose a novel deep learning-based hydrogen leak detection model that provides a smart, real-time detection solution. The model leverages computational fluid dynamics (CFD) simulations to construct a proprietary database of high-pressure hydrogen leaks under various conditions, including different leak locations, flow rates, and wind directions. Results: Our analysis revealed that wind direction plays the most significant role in influencing hydrogen dispersion patterns, which is crucial for accurately identifying leak sources and predicting the affected area. We compared six deep learning models: a backpropagation neural network (BPNN) based on a multi-layer perceptron (MLP), convolutional neural network (CNN), long short-term memory (LSTM) network, convolutional long short-term memory (CNN-LSTM), bidirectional long short-term memory (BiLSTM), and convolutional neural network bidirectional long short-term memory (CNN-BiLSTM). Among these models, the CNN-BiLSTM model exhibited the highest performance in leak detection tasks. By combining the strong local feature extraction capabilities of CNN with the long-term dependency capturing abilities of BiLSTM, this hybrid model significantly outperformed other models, achieving an accuracy and F1 score exceeding 98%. These results highlight the model's ability to handle complex temporal data efficiently, making it particularly effective for identifying hydrogen leaks in real-time industrial environments. The study also explores key factors influencing the performance of the detection models. We conducted sensitivity analyses on two critical hyperparameters: batch size and the number of training iterations. We found that a batch size of 16 and 400 iterations provided the optimal trade-off between convergence speed and detection accuracy. In addition, the robustness of the model has been demonstrated, maintaining high accuracy even in the face of complex conditions such as changes in wind direction and leak strength. The model has a high localization accuracy of more than 98.00% when detecting most leakage sources. The detection accuracy is slightly lower only in the hydrogen unloading region and the hydrogen storage region, mainly due to the limitation of the sensor layout. In addition, the research introduced data preprocessing techniques, including normalization, data dimension reduction, and feature selection, which significantly improved the efficiency of the detection process. By minimizing the dimensionality of input data, the computational load was reduced, enabling faster detection without sacrificing accuracy. Notably, the CNN-BiLSTM model also excelled in detecting rare but dangerous leak events, enhancing the overall safety monitoring capabilities of hydrogen refueling stations. Conclusions: This study's findings not only provide a theoretical foundation for hydrogen leak detection in refueling stations but also present a practical solution that improves both detection accuracy and operational efficiency. The proposed CNN-BiLSTM model offers a robust and intelligent approach for monitoring hydrogen leaks, significantly enhancing real-time safety measures in complex industrial settings. Future work will focus on expanding the model's generalizability to broader industrial applications and exploring further optimization of the feature extraction and classification processes to support the development of intelligent safety monitoring systems.