Federated learning based atmospheric source term estimation in urban environments

Abstract

Establishing an accurate and efficient source term estimation (STE) system is of great significance for identifying unknown gas leakage sources in urban environments. Many successful STE methods have been proposed, while most of them assume there is only one point source. However, the complicated urban atmospheric dispersion and the massive sensor data in distributed edge devices pose new challenges. To address these issues, this paper proposes a method to convert measured concentrations into visual features, which retains the characteristics of diffusion and the layout of obstacles. Then, a federated STE (FL-STE) framework is proposed to extract knowledge from local models collaboratively without collecting all privacy data, in which a deep neural network is used to recognize the relationship between visual features and source terms. Furthermore, we construct an urban dispersion dataset with multiple obstacles and sources by FDS simulation. Various empirical studies prove the efficiency of the proposed method.