Exploiting data-driven processes for detecting and diagnosing the occurrence of faults in HVAC systems

According to the most recent IEA study from 2022, the building sector is a significant global energy consumer and one of the biggest CO2 producers, accounting for 30% of global final energy consumption and 26% of global energy-related emissions (11% from direct emissions in buildings and more than 20% from the production of electricity and heat used in buildings). Up to 50% of the energy used in commercial and institutional buildings can be attributed to the HVAC system alone. It is crucial for these systems to operate effectively and efficiently because malfunctioning sensors and component faults can lead to significant energy waste, increased emissions, and negatively impact occupants' comfort. Therefore, a substantial amount of energy can be saved by effectively applying fault detection and diagnosis (FDD) techniques. Despite extensive research efforts in HVAC system fault detection, a major challenge remains in making these methods scalable and adaptable to real-world applications. A notable gap in the literature is the portability of these fault detection methodologies, especially when applied to practical, real-world data, which often consists of both labeled and unlabeled faulty datasets. When system data is unlabeled and faulty operations are difficult to distinguish, the need for further investigation becomes clear. Even with labeled datasets, significant challenges persist, underscoring the need for continued focus and refinement in this area. This thesis addresses these challenges by developing a hybrid FDD system tailored for HVAC systems, with a particular focus on AHUs. The proposed methodology combines data-driven statistical analysis with rule-based processes specifically designed for single duct air handling units (SDAHU). By introducing a novel approach that employs a portable framework for energy management and information systems (EMIS), this research aims to fill the gap in scalability and adaptability. The framework leverages data-driven analysis and is built on a Metadata schema, making it modular and scalable. With the aid of the Brick ontology, the tools developed can be easily applied across various building systems and environments. Moreover, this approach enables the detection of faults in both labeled and unlabeled datasets, addressing the challenges posed by real-world data variability.