Gas Source Localization in Outdoor Environment Using Deep Learning: Data Augmentation by Gas Dispersion Simulator

This paper presents an application of a gas dispersion simulator that uses recorded wind data to reproduce gas distribution observed in an open outdoor field. Outdoor gas source localization is highly challenging because of the large fluctuations of the wind. Combination of a gas sensor network and a deep-learning pattern recognition model is expected to provide high accuracy in estimating a gas source location. However, collecting sufficient amount of gas sensor data for training a neural network model can be extremely laborious. Here we propose to use the simulator to generate the training data. Gas distributions trailing from its source in various directions are simulated by rotating the wind data measured in an outdoor environment. The temporal waveform of each sensor response in a stationary gas sensor network is calculated by feeding the simulated gas concentration values to a second-order sensor response model. The sensor response data is then used to train a long short-term memory neural network (LSTM) to estimate the location of the gas source. The high-accuracy results demonstrate the effectiveness of augmenting the training data using the gas dispersion simulator.