Aiming at the problems of false alarms, missing alarms and abnormal working status of sensor nodes in high temperature environment, this paper proposes a fire escape path planning research method combining LSTM and improved A ∗ algorithm. According to the LSTM, the real-time fire situation information was adaptive learned, and the abnormal node data prediction model was established to predict the threat situation of abnormal nodes, such as temperature and carbon monoxide concentration. Based on the real-time situation information of indoor fire, the fire threat situation spread model was built, and the improved A ∗ algorithm was used to dynamically plan the escape path to obtain the best safe escape path under abnormal conditions. The results show that this method can plan the best escape path in different fire periods, and gain valuable time for the evacuation of personnel, which has practical application value.