Disaster Recognition Method by Time Series Model Based on Deep Learning for ERESS

Since ancient times, disasters have caused many victims around the world. In particular, sudden disasters such as fires cause the greatest damage due to delayed escape. To solve this problem, we are developing an Emergency Rescue Evacuation Support System (ERESS) that immediately provides information for evacuators with mobile devices. ERESS recognizes disaster outbreaks by analyzing the behavior of device holders. The behavior analysis has been used machine learning such as Support Vector Machine (SVM). However, there is a problem that it mistakenly judges sudden movements as emergency situations. In this paper, we propose a method by time series model to improve conventional disaster recognition method. The proposed method classifiers time-series changes in sensor information obtained from mobile devices based on deep learning. For this analysis, we use Long Short-Term Memory (LSTM) networks, a category of Recurrent Neural Networks (RNN) that are suitable for pattern recognition in time-series data. We demonstrated through a disaster simulation that this method can recognize the occurrence of a disaster about $46 \%$ more quickly than conventional methods.