24 April 2023
13:00 Master's Defense IC3 Auditorium
Theme
Analysis of Air Pollutants using Virtual Sensors
Student
Gabriel Oliveira Campos
Advisor / Teacher
Leandro Aparecido Villas - Co-advisor: Felipe Domingos da Cunha
Brief summary
Virtual sensors have been increasingly used to generate synthetic data and provide complementary information for smart cities. This approach is important to analyze aspects, variables and events for which there are no physical sensors, or in cases where there are physical sensors but they are not present on site. The use of data from virtual sensors, for the analysis of air pollutants, can present results that help the population that has more sensitive health, given the importance of the presence of sensors able to capture levels of pollutants in the atmosphere. Allied to this, the creation of this type of sensor helps in city planning, as more and more smart cities are emerging and the dissemination and control of this data becomes necessary. However, there are cases in which physical sensors may be unavailable, either due to lack of infrastructure, access difficulties, or even incorrect data reading. Therefore, one of the biggest challenges of using physical sensors to analyze the concentration of pollutants in the air is the lack of infrastructure in cities, which are not sufficient for the analysis of pollutants in their large urban centers. Therefore, this work aims to create virtual sensors for detecting air pollutants through the use of machine learning or deep learning models, which are robust, accurate and replicable. The use of these virtual sensors is a good cost-effective alternative to replace the lack of physical sensors that analyze the concentration of these pollutants, or even to replace values when there is a loss of data or an incorrect reading. Therefore, a preliminary correlation analysis of data in different cities was carried out to verify the robustness of the model. The importance of each meteorological variable and pollutants was also evaluated to verify which ones have the most impact on the creation of virtual sensors. Finally, different learning methods were used to identify the best model for creating virtual sensors. In this work, we produced several models of virtual sensing, ranging from daily and hourly models, using several cities in completely different locations, and finally, we also made a Forecasting analysis to deal with the problem of complete loss of data from the monitoring network, using only historical information on the pollutant. After all the analyses, we were able to verify that the Boosted Trees obtained the best results, with a lower RMSE for all the evaluated scenarios, in addition, this model also obtained a faster processing time than most of the evaluated models.
Examination Board
Headlines:
| Leandro Aparecido Villas | IC / UNICAMP |
| Max do Val Machado | ICEI / PUC Minas |
| Luiz Fernando Bittencourt | IC / UNICAMP |
Substitutes:
| Esther Luna Colombini | IC / UNICAMP |
| Helder May Nunes da Silva Oliveira | CECS / UFABC |