?
Разработка микросервиса ADP для идентификации источников выбросов на основе машинного обучения с подкреплением
The results of the development of a software microservice embedded in atmospheric air quality monitoring systems to support the identification of industrial pollution sources are presented. The emission and subsequent spread of harmful substances in the lower layers of the atmosphere is dynamic and characterized by high uncertainty due to the specific features of technological installations, their operating modes, the influence of terrain relief, buildings and meteorological factors. The relationship between the location of the emission source and the information from sensors installed in central areas of the city or on the boundaries of sanitary protection zones of large industrial facilities cannot be described analytically, Therefore, formalizing the knowledge of environmentalists and subsequently automating the detection of objects responsible for the formation of hazardous concentration levels at control points is a pressing task. The aim of the study is to develop an algorithm for the continuous optimization of search strategies using Approximate Dynamic Programming technology. This article proposes implementing the ADP mechanism based on Q-Learning, which in turn is performed in simulation mode through interaction with the Lagrange model describing the physical processes of pollution dispersion. The developed model learns to select the best search steps (actions) on a marked map of the terrain, considering the cost function approximated by a neural network, meteorological factors and terrain relief, which is a new technological solution. The design of basic information processes was carried out, including the consideration of processes for collecting and pre-processing data on the measurement of harmful substance concentrations and meteorological data at control points, the preparation of a table for Q-Learning and its use for training a neural network model, and the application of the model to solve the problem of determining the source of an emergency release. The results of experimental testing showed that the microservice developed and integrated into the digital ecomonitoring platform accurately captures the characteristics of industrial pollution dispersion processes in the atmosphere and can be used for automated identification of emission sources in dynamics. The average values of the contribution of the emergency release source to the formation of pollution in a given territory differ from the values calculated using the UPRZA example by no more than 15%, which allows us to conclude that the results are highly reliable and can be compared with GOST methods that operate in static conditions.