Aristotle University Of Thessaloniki (AUTH)
Automated computational analysis and modeling for the operational improvement of air quality sensor networks in real time
Abstract
dc:descriptionThis dissertation advances the deployment and management of low-cost sensor networks(LCSNs) for air quality (AQ) monitoring through novel machine learning (ML)-based calibration techniques and operational frameworks. By combining empirical experiments and a comprehensive review of current methodologies, the research identifies and addresses critical factors impacting LCSN functionality, accuracy, and scalability. This work evaluated several multifaceted LCS apparatus, integrating temperature and humidity sensors alongside AQ measurements to mitigate environmental interferences, enhancing measurement precision. Key innovations include the development of the “distribution guided neural disaggregation” method, which improves temporal resolution through the separation of mixed-source AQ data, and “spatial online learning”, a real-time calibration approach that adapts to changing environmental patterns across diverse sites and the “global (network-wide) calibration” scheme for faster learning, adaptive insertion of nodes to the network, and increased calibration accuracy. It moreover, facilitates scalability, allowing calibration models to generalize across varied locations, reducing the need for frequent recalibration. A robust operational framework supports these innovations, emphasizing strategic sensor placement, and adaptive calibration. Using advanced frameworks such as “Genetic Algorithm Hybrid Stacking” (GAHS v2 and v3) for model integration, and “Spatio-Temporal Cross-Validation” (STCV), this research enables realistic model evaluation under dynamic spatiotemporal conditions. With a focus on model robustness, the study investigates techniques to handle unique AQ data patterns, including concept drift, spatiotemporal autocorrelation, and late-arriving reference data, enabling sustainable long-term LCSN operation. Incorporating graph neural network (GNN) imputation for missing data and GDL for multi-site forecasting, the research bridges ML methodologies with practical LCSN deployment. The outcomes present a foundational framework for urban-scale AQ management, establishing LCSNs as a scalable, cost-effective alternative to traditional monitoring networks. Key contributions include improved temporal resolution, adaptive calibration, and operational tools that enhance model longevity and accuracy across heterogeneous sensor networks.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2024
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Bagkis, Evangelos
- Μπαγκής, Ευάγγελος
Subjects
dc:subject × 6Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/58048