文献类型：期刊文献

英文题名：Inverse tracing of fire source in a single room based on CFD simulation and deep learning

作者：Shen, Xiaobo[1,2];Cao, Zhaoyang[1];Liu, Haifeng[1];Cong, Beihua[3];Zhou, Feng[4];Ma, Yunsheng[5];Zou, Xiong[5];Wei, Shengke[5]

机构：[1]East China Univ Sci & Technol, Sch Resources & Environm Engn, Shanghai 200237, Peoples R China;[2]Shanghai Inst Pollut Control & Ecol Secur, Shanghai 200092, Peoples R China;[3]Tongji Univ, Shanghai Inst Disaster Prevent & Relief, Shanghai, Peoples R China;[4]Shanghai Fire Res Inst MEM, Shanghai 200032, Peoples R China;[5]Shandong Chambroad Holding Grp Co Ltd, Shandong 256500, Peoples R China

年份：2023

卷号：76

外文期刊名：JOURNAL OF BUILDING ENGINEERING

收录：;EI(收录号：20232714345473);WOS:【SCI-EXPANDED(收录号:WOS:001058883900001)】；

基金：This work was supported by the National Natural Science Foundation of China (Grant No. 22278135 and 22078095) and the Shanghai Science and Technology Committee (Grant No. 20dz1200903 and 21QC1400400) . The authors sincerely thank these supports.

语种：英文

外文关键词：Fire investigation; CFD simulation; Inverse model; BP network; Deep learning

摘要：Inverse tracing of fire source location is important for investigation after the fire accidents. In this work, the CFD simulations and deep learning were combined to explore a more efficient and intelligent tool for fire investigation. Firstly, a CFD model for single room was built using FDS. Then abundant simulations were performed by varying the initial conditions to collect massive data including temperature distribution and smoke layer heights in the room. The dataset was divided into two parts for the training and validation of BP network, respectively. The inverse model and forward model with the same data set and neural network parameters were established, which saved the time in adjusting models and producing data and improves the efficiency of research. The shared data sets and neural network parameters did not affect the final prediction results. The accuracy of forward and inverse models is still excellent, reaching a high accuracy. In particular, the accuracy of the inverse model had been improved to more than 99% compared with previous studies. The accuracy of the forward model is more than 80%. Besides, the inverse model's robustness was also examined, the model is still valid when some input features are lost.