Source localization for hazardous material release in an outdoor chemical plant via a combination of LSTM-RNN and CFD simulation

Chemical leak accidents not properly handled at the early stage can spread to major industrial disasters escalating through fire and explosion. Therefore, it is very important to develop a method that enables prompt and systematic response by identifying the location of leakage source quickly and accurately and informing on-site personnel of the probable location(s). In this study, a model that predicts the suspicious leak location(s) in real-time, using sensor data, is proposed. Feed-forward neural network and recurrent neural network with long short-term memory that learned the data gathered from the installed sensors are proposed to predict the Top-5 points in the order of highest likelihood. In order to train and verify the neural networks, the sensor data generated from computational fluid dynamics simulations for a real chemical plant are used. The model learns the inverse problem solving for accident scenarios and predicts the leak point with very high accuracy.