Against the backdrop of the global energy transition, zero-emission hydrogen fuel cell vehicles are critical for achieving deep decarbonization in the transportation sector. However, the potential safety hazards associated with their onboard high-pressure hydrogen storage systems cannot be overlooked. This issue is particularly critical in confined spaces with poor ventilation, such as underground parking garages. Hydrogen leaks pose a significant safety hazard owing to the risk of combustion. Current hydrogen detection technologies for large enclosed spaces suffer from response delays, low positioning accuracy, and susceptibility to environmental noise, which makes it difficult to meet the stringent requirements for rapid and precise leak localization in emergency rescue operations. Therefore, improved hydrogen leak detection and early warning systems must be urgently developed to ensure the safety of hydrogen fuel cell vehicles in underground parking facilities.

This study developed a hydrogen leak localization and early warning platform for underground parking garages by integrating computational fluid dynamics numerical simulations, deep learning algorithms, and visualization technology. First, 240 typical scenarios covering different leakage rates and locations were simulated using Fluent, and a high-fidelity and high-volume dataset of original signals was constructed. Second, an innovative deep learning localization model integrating complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), convolutional neural networks (CNNs), and bidirectional long short-term memory (BiLSTM) networks (hereafter known as "the CEEMDAN–CNN–BiLSTM deep learning model") was developed. In this architecture, the CEEMDAN algorithm performs multiscale decomposition and denoising on nonstationary concentration signals, effectively eliminating interference from environmental noise. CNNs are used to extract the spatial distribution features of the sensor array data, whereas BiLSTM networks are used to deeply mine the temporal sequence dependencies of gas diffusion. Finally, an early warning platform integrating data import, intelligent hierarchical alarming, and visual leakage source localization was developed using MATLAB App Designer.

The results indicated that the combustible hydrogen cloud primarily accumulated beneath the vehicle and around the side door areas, stabilizing at 25 s as the leakage and hydrogen diffusion rates reached dynamic equilibrium. The CEEMDAN–CNN–BiLSTM deep learning model demonstrated outstanding performance. Allowing for an error margin of one parking space, the row deep learning model achieved 100% accuracy, precision, recall, and F1 scores, whereas the column deep learning model attained an accuracy of 99%, with precision, recall, and F1 score reaching 99.5%. Using two sets of row and column deep learning models, preliminary localization was achieved 5 s after an alarm was triggered, meeting the strict timeliness requirements for emergency response. Furthermore, high-precision localization was accomplished after 20 s of data accumulation, ensuring accurate results. These findings provide technical support for balancing the efficiency and precision of hydrogen safety monitoring in underground parking garages. The proposed platform enables real-time data import, hierarchical alarming, and leak localization, provides an intuitive display of concentration changes at the sensors, and precisely locates the leak.

The simulation–data–model–platform workflow fosters the interdisciplinary competencies and safety risk awareness of students as well as their ability to solve real-world engineering problems. This initiative offers a practical, results-oriented approach to hydrogen safety and related professional education.