Effective fire detection provides early warnings and key information for first responders and people trapped insides. The idea of integrating sensor data and fire modeling presents a general framework for fire source parameter estimation. However, most methods fail to achieve a real-time accurate estimation due to complex building structures and high computational requirements. Inspired by the capability of deep learning in data mining, a model based on Gated recurrent unit (GRU) is proposed to determine fire locations and intensity. First, a series of fire scenarios is simulated to form the dataset. Second, GRU is applied to learn representations from sensor data. Third, fire source parameters are estimated by the trained GRU with sequential sensor measurements. Multiple configurations and data are used to assess the inverse model. The results show that this model performs well and achieves a high test accuracy. The estimation of fire location is not influenced by the precision of fire simulations, while the intensity inversion is sensitive to the deviations. In addition, reliability, efficiency, and robustness of the inverse model are studied. This study is a fundamental step towards a credible and applicable deep learning-based model for fire source parameter inversion that assists in building fire protection.

有效的火灾探测为急救人员和困在里面的人提供预警和关键信息。集成传感器数据和火模型的想法提出了火源参数估计的通用框架。但是，由于复杂的建筑结构和较高的计算要求，大多数方法无法实现实时准确的估算。受数据挖掘中深度学习能力的启发，提出了一种基于门控循环单元（GRU）的模型来确定火灾的位置和强度。首先，模拟一系列火灾场景以形成数据集。其次，将GRU应用于从传感器数据中学习表示形式。第三，由训练有素的GRU通过连续的传感器测量来估算火源参数。多种配置和数据用于评估逆模型。结果表明，该模型性能良好，测试精度较高。火灾位置的估计不受火灾模拟精度的影响，而强度反演对偏差敏感。此外，还研究了逆模型的可靠性，效率和鲁棒性。这项研究是朝着可靠和适用的基于深度学习的火源参数反演模型迈出的重要一步，该模型有助于建筑防火。