卷积与长短期记忆网络在火灾源强实时预测中的应用

Application of Convolutional Neural Networks and Long Short-Term Memory networks in real-time prediction of fire source intensity

针对火灾场景中火源位置及强度实时、准确识别困难的问题,利用卷积神经网络(Convolutional Neural Networks,CNN)与长短期记忆(Long Short-Term Memory,LSTM)网络的算法优势,构建一种火灾源强实时预测模型,该模型通过建筑内温度传感器接收的序列数据,实时、准确地预测火灾场景中的火源位置及强度信息。采用火灾动力学模拟软件(Fire Dynamics Simulator,FDS)模拟火灾场景,获得温度传感器实时接收的序列数据,建立火灾场景数据库,进行火灾场景数据分析并对火灾源强实时预测模型完成训练,通过实例验证该模型的准确性、及时性和鲁棒性。结果表明,该模型能够通过较短长度样本数据实时、准确预测火灾场景中火源位置及强度,预测准确率为99.18%,在温度传感器间隔损坏且损坏率不高于70%时,预测准确率仍可达到95.10%以上。

To solve the prohlem of real-time and accurate identification of fire location and intensity in fire scenes,a realtime fire source intensity prediction model was built by integrating the algorithm advantages of Convolutional Neural Networks(CNN)and Long Short-Term Memory(LSTM)networks.The CNN is good at local special feature extraction,and the LSTM is good at time series capture.The model can accurately predict the fire location and intensity in fire scenes through the real-time sequence data received by the temperature sensor in the building.A Fire Dynamics Simulator(FDS)was used to recreate the fire scenes and collect real-time sequence data by temperature sensors.The temperature sequence data was analyzed and used to establish the fire scenes database.The prediction model was trained to mine the mapping relationship between fire source intensity and fire scene data,and evaluated for accuracy,timeliness,and robustness with a case.The results show that the convergence speed of fire location prediction is faster,compared with fire intensity prediction,fire location,and intensity prediction.Through short-length sample data,the trained model can accurately predict the fire location and intensity in fire scenes in real-time.The prediction accuracy is higher than 94.04%with more than 2-length of the sample.When the sample length is 10,the prediction accuracy of fire location is 99.39%,the prediction accuracy of fire intensity is 99.70%,and the prediction accuracy of both location and intensity is 99.18%.Moreover,the prediction accuracy of both location and intensity can still exceed 95.10%when the temperature sensor is damaged at intervals and the damage rate is less than 70%.If the temperature sensor is constantly damaged at one end or both ends of the corridor,the damage rate shouldn't be higher than 30%or 50%,respectively,to guarantee greater than 90%accuracy.The fire source intensity prediction model provides a reference for obtaining fire scene information in real-time.