Design and implementation of AI-based thermal comfort sensing and control for adaptable indoor environments : a thesis by publications presented in partial fulfillment of the requirements for the degree of Doctor of Philosophy in Engineering, Massey University, Auckland, New Zealand

Abstract

Thermal comfort is a critical determinant of occupant well-being, productivity, and energy efficiency in modern buildings. Despite advances in building management systems, current control strategies often lack real-time adaptability and remain largely rule-based or model-driven, failing to account for the dynamic, individualized physiological needs of occupants. This research addresses these challenges by developing a comprehensive, AI-driven framework for thermal comfort sensing and adaptive control with a strong emphasis on occupant-centric intelligence and real-time responsiveness.

The work begins by identifying the limitations of conventional control models (such as schedule-based and PMV-driven approaches) and the pressing need for occupant-centric solutions in contemporary indoor environments. To bridge this gap, a novel low-cost RGB-T (Red, Green, Blue, Thermal) imaging system was designed and implemented, enabling non-contact, real-time measurement of facial skin temperature as a direct physiological proxy for thermal comfort. This system was further integrated with an open-source Raspberry Pi platform, transforming it into a BACnet-enabled controller and sensor node capable of closed-loop operation and seamless communication within real-time building automation networks. The integration demonstrates the practical feasibility of using affordable hardware not only for sensing, but also for adaptive control, within a standards-based automation environment.

Building on this foundation, the research advances the field by moving away from traditional face detection methods. Instead, it employs a deep learning approach and applies transfer learning with MobileNetV2 to enable lightweight, edge-deployable inference for robust and accurate facial temperature estimation. This methodological innovation significantly enhances the reliability and scalability of occupant comfort assessment, particularly in dynamic and heterogeneous indoor environments, paving the way for intelligent, real-time HVAC control based on physiological data.

Experimental validation in both simulated and operational settings using comparative baselines against conventional control strategies confirms the effectiveness of the proposed framework, with consistently improved comfort prediction accuracy and measurable gains in energy efficiency over traditional methods. The findings underscore the novelty and practical impact of integrating AI, low-cost sensing, and open communication protocols to realize adaptive, occupant-centric building management. This work lays a robust foundation for future research and deployment of smart, responsive indoor environments that are both energy-efficient and human-centered.