Field detection of indoor fire threat situation based on LSTM-Kriging network

Abstract

Obtaining the fire development information quickly and accurately is an important part of indoor fire emergency evacuation and rescue. However, at present, buildings only use fire sensors to monitor whether a fire is occurring, and there is a lack of means to obtain real-time threat situation information on the development of fires. An LSTM-Kriging neural network model is proposed in this paper, which can quickly generate an indoor fire threat field using the temporal data of fire sensors when an indoor fire occurs. Firstly, the Kriging algorithm was improved by replacing the Euclidean distance used in the variational function with the nearest passable distance to obtain a more accurate representation of the indoor threat distribution trend, and combined it with the Long Short-Term Memory (LSTM) network to construct the LSTM-Kriging network. Secondly, indoor fire scenarios were constructed in PyroSim, and fire parameters were simulated randomly. After data augmentation, the indoor fire threat situation dataset was generated. Finally, the optimized LSTM-Kriging model was trained and compared with Transformer, LSTM, and Recurrent Neural Network (RNN) on the indoor fire hazardous situation dataset. The experimental results show that LSTM-Kriging is the most effective model compared with several models mentioned above and the prediction of threatening situation reaches an accuracy of 97.8 %, with an average computation time of 0.19s on the GPU, which can meet the requirement of real-time fire situation monitoring. The indoor fire threat situation field detection method proposed in this paper achieves intelligent indoor fire detection and improves the building fire safety.