深層学習を用いた屋外環境のガス源探索

Abstract

The aim of this research project is to attain accurate gas-source localization in outdoor environments with large wind fluctuations. For this purpose, we propose to use a long short-term memory deep-learning framework to time-series data collected by a sensor network consisting of multiple gas sensors and an anemometer. This paper describes impacts of the length of time-series data and smoothing of wind data provided to a deep neural network model. We have collected three datasets by placing 30 semiconductor gas sensors and one ultrasonic anemometer in an outdoor field in different seasons. We have found that the success rate of gas-source location estimation can be effectively increased by removing high frequency fluctuations in the time-series data of the wind velocity vector by taking moving average before applying the data to the neural network. By adjusting the data length provided to the neural network and smoothing the wind data, the success rate of gas-source location estimation has been increased from 82.5% to 86.7%. A success rate of 78.8% has been obtained even when half of the gas sensors have been removed.