Abstract
Tracing the source of air pollution presents a significant challenge, especially in densely populated urban areas, because of the unpredictable and complex nature of aerodynamics. To address this issue, intelligent lamp posts have been developed with smart sensors and edge computing capabilities. These lamp posts serve as nodes in the EIPA (Edge Intelligent Perception Agent) network within urban campuses. These lamp posts aim to track air pollutants by employing a tracking algorithm that utilizes big data learning and Gaussian diffusion models. This approach focuses on monitoring the quality of urban air and identifying pollution sources, rather than relying solely on traditional CFD simulations for air pollution dispersion. The algorithm comprises three primary components: (1) the Federated Learning framework built on the EIPA system; (2) the LSTM model implemented on the edge nodes of the EIPA system; and (3) a genetic algorithm utilized for optimizing the model parameters. By using CFD simulations in a simulated city park, training data on air dynamic movements is gathered. The usefulness of the method for tracing air pollutants based on federated learning of edge intelligent perception agents is demonstrated by the outcomes of algorithm training. Experimental results show that, compared to the traditional genetic algorithm (GA) and LSTM + genetic algorithm, the proposed FL + LSTM + GA method significantly improves the pollution source positioning accuracy to 99.5% and reduces the average absolute error (MAE) of Gaussian model parameter estimation to 0.20.