Aplicação de TinyML em sensores virtuais para monitoramento da qualidade do ar em ambientes industriais

Abstract

Continuous air quality monitoring, especially in the mining sector, is essential to ensure that pollutant levels remain within regulatory standards. Traditional solutions for this monitoring face technical and economic limitations. In this context, Machine Learning (ML) techniques applied to embedded systems (TinyML) offer a promising alternative for forecasting and analyzing atmospheric pollutants. This work aims to develop a low-cost system with limited computational resources for continuous air quality monitoring that uses artificial intelligence (AI) techniques to predict chlorine concentration in the air based on easily accessible input variables: potassium and sodium. The AI model was trained, optimized, and embedded in an ESP32-S3 microcontroller. Model quantization reduced the model size without compromising accuracy, ensuring processing feasibility on resource-limited devices, with an average inference process time of approximately 49.01 µs. Regarding chlorine prediction accuracy, the model presented a mean squared error of 3.96, a mean absolute error of 1.14, and a determination coefficient of 0.95, demonstrating high efficiency and confirming the viability of the proposed system.