Study on the general framework for real-time heat release rate inversion of tunnel fires with deep learning and transfer learning

The real-time heat release rate is an important parameter for evaluating fire risk and adjusting the emergency response plan. The fusion of numerical simulation and machine learning makes it possible to predict the tunnel fire state in real-time, but there are still limitations such as only applicable to the single-source fire scenarios of the specific tunnel and lack of training data. To further extend the applicability of the tunnel fire state inversion method, a general framework for tunnel fire heat release rate inversion based on deep learning and transfer learning was established in this study. By taking three different full-scale tunnel fire test data as the test datasets, the influence of training dataset size on the inversion accuracy of the tunnel fire state is analyzed. The transfer of fire state inversion models for single-source fire and double-source fire scenarios in different full-scale tunnels with small training databases is achieved, and the determination coefficient of the relevant models reaches up to 80.2% on the full-scale tunnel fire test dataset. The efficiency and reliability of the tunnel fire state inversion models based on transfer learning were evaluated. The results indicate that the time taken by the models to complete real-time tunnel fire state inversion for six fire scenarios is within 1 s, which has a significant advantage in computational efficiency and meets the timeliness requirements for emergency response to tunnel fire accidents. In the full-scale tunnel single-source fire and double-source fire scenarios, the inversion accuracy drops by only 8% on average when one temperature sensor near the fire source is completely damaged, demonstrating good reliability in complex tunnel fire scenarios.