IoT and Digital Technologies for Quality-Oriented Cleaning and Integrated Services

Nowadays, people spend up to 90% of their time in enclosed spaces, so indoor hygiene is a critical yet often overlooked factor in public health. Despite this, the management of sanitization in built environments is still predominantly based on static schedules, visual inspections, and manual reporting, offering limited protection against invisible biological risks. This thesis addresses this gap by proposing integrated technologies for quality-oriented cleaning for intelligent hygiene management in smart buildings. The research is structured around three complementary pillars. The first tackles the problem of biological risk invisibility by introducing a virtual sensor. Instead of relying on slow and costly microbiological sampling, low-cost IoT sensors are used to continuously monitor environmental proxies such as CO2, volatile organic compounds, temperature, humidity, and particulate matter. Through machine learning models, these quantities are mapped to airborne bacterial concentrations, transforming standard IoT nodes into virtual microbiological sensors. Experimental results demonstrate high predictive fidelity (R^2 up to 0.92) and sufficient accuracy to trigger preventive actions well before critical contamination thresholds, while deep learning–based power management strategies reduce wireless transmissions and extend battery life by up to 89%. The second research pillar is based on computer vision. Human occupancy is estimated by a virtual sensor extracting people counts, flows, and dwell times from video streams processed directly on edge devices. A lightweight tracking-by-detection pipeline, combined with a Dynamic Inference Power Manager, enables adaptive frame skipping based on scene dynamics. This approach achieves energy savings up to 36% without significantly degrading tracking accuracy, proving that visual sensing can be both sustainable and reliable for continuous deployment. The third pillar places human behavior at the center of hygiene dynamics through sensor-based Human Activity Recognition. A hierarchical architecture, named Lightweight Accurate Trigger, dramatically reduces wearable energy consumption by activating complex classifiers only when relevant motion patterns occur, achieving savings up to 95%. Generalized Zero-Shot Activity Recognition based on Siamese Neural Networks enables recognition of unseen activities without retraining, while Transformer-based models demonstrate robust signal reconstruction and real-time inference on resource-constrained wearable hardware. Overall, the results confirm that the proposed technologies are practical and scalable solutions, capable of combining high analytical accuracy with strict energy constraints. By integrating environmental, spatial, and behavioral intelligence, this thesis lays the foundation for predictive, adaptive, and verifiable hygiene management in future smart buildings.