DustAI: Monitor de Material Particulado de Baixo Custo com Calibração via Aprendizado de Máquina

• Ano de defesa: 2023
• Autor(a) principal: Pastório, André Francisco
• Orientador(a): Camargo, Edson Tavares de
• Banca de defesa: Brun, André Luiz, Pfrimer, Felipe Walter Dafico, Sousa Junior, Wilson Cabral de
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Estadual do Oeste do Paraná; Cascavel
• Programa de Pós-Graduação: Programa de Pós-Graduação em Ciência da Computação
• Departamento: Centro de Ciências Exatas e Tecnológicas
• País: Brasil
• Palavras-chave em Português: Material particulado; Sensores de baixo custo; Sensores; Qualidade do ar; Calibração; Aprendizado de máquina
• Palavras-chave em Inglês: Particulate matter; Low-cost sensors; Sensors; Air quality; Calibration; Machine learning
• Área do conhecimento CNPq: SISTEMAS DE COMPUTAÇÃO
• Link de acesso: https://tede.unioeste.br/handle/tede/7102

Low-Cost Particulate Matter (LC-PM) sensors have been investigated worldwide as an alternative to expensive reference stations for air quality monitoring. However, LC-PM sensors are inaccurate and subject to uncertainties depending on the conditions of the environment in which they operate. Calibration of these sensors can be performed using different methods where the sensor is placed in a real environment exposed to local environmental conditions and its measurement is compared to a reference equipment. This work proposes the development of the DustAI device, a low-cost particulate matter monitor with calibration in the cloud. The calibration is performed using Machine Learning (ML) models. Hosting the ML model in the cloud provides flexibility and scalability. The measurements obtained by DustAI are transmitted to the Internet in real time via a long-range, low-power wireless network using LoRaWAN technology. The LM models were created after comparing the MP-BC sensors with reference devices at three different points in time over 4, 5 and 1 month. The results obtained from the collections showed that calibration is required and that temperature and humidity affect the sensor measurements. The models obtained were able to increase the linear correlation (Pearson) and decrease the mean absolute error (MAE) and the squared error (MSE), focusing on the following ML algorithms: XGBoost, LGBM and CatBoost. When creating a calibration model with two weeks of data collected by DustAI, the linear correlation for PM10 increased from 0.48 to 0.74 and for PM2.5 from 0.69 to 0.79. The MAE of PM10 decreased from 16.56 to 6.88 and the MSE from 472.25 to 126.66. And for PM2.5 from 3.39 to 2.25 (MAE) and from 22.27 to 10.69 (MSE)