Abstract: The fire source location and heat release rate （HRR） are crucial information guiding emergency firefighting and rescue perations during tunnel fires. However, in practice, the information that can be obtained about the fire is limited. It is difficult to get the fire source key parameters directly. Therefore, we researched the deep learning-based method for inversing thefire source location and HRR in tunnel fires. Firstly, a tunnel fire dataset under different boundary conditions is established based on numerical simulations validated by experimental data. Based on convolutional neural network （CNN） and long short-term memory network （LSTM）, the mapping relationship between fixed temperature sensor data and fire source location and HRR was established. The inversion effectiveness of the model for fire source parameters was evaluated. And the effect of time series length and sensor spacing on the inversion effectiveness were evaluated. The results demonstrate that the model has good inversion performance for both HRR and fire source location. When the time series length was 20 s and the sensor spacing was 30 m, the R2 values of the model inversion for HRR and fire location are 0.97 and 0.99, respectively.

Key words: tunnel fire, numerical simulation, fire tests, long and short-term memory network, heat release rate, deep learning