Development of a hybrid process for fault detection and diagnosis in HVAC systems integrating data-driven and knowledge-driven approaches

Despite significant progress in improving the energy efficiency of Italy’s building stock, the building sector remains the country’s largest consumer of electricity, accounting for approximately 40% of national demand. HVAC systems are responsible for nearly half of this consumption. In this context, faults in HVAC operation can further increase energy use, accelerate equipment degradation, and negatively affect occupant comfort and indoor environmental quality. To improve reliability and performance, Fault Detection and Diagnosis (FDD) systems are receiving growing attention as a strategy to support more efficient and resilient HVAC operations. Traditional FDD methods, primarily based on rule sets or physical models, remain prevalent but are often limited in flexibility and scalability across different systems. In recent years, data-driven approaches have demonstrated strong potential by learning behavioral patterns directly from operational data. However, their deployment in practice remains limited, mainly due to the need for large volumes of labeled data that specify both the presence of faults and the components involved, a condition that is rarely met in real buildings. In this perspective, this work presents a label-free hybrid FDD approach that combines expert knowledge and data-driven modeling within a probabilistic framework based on Bayesian Networks. A Random Forest algorithm is used to estimate the baseline behavior of selected system variables. Deviations between predicted and actual values are introduced as virtual evidences in a Bayesian model, which also incorporates hard evidences derived from system setpoints and expert-defined rules. The Bayesian network structure is defined using domain expertise and informed by a semantic model of the HVAC system based on the Brick ontology, enabling transparent mapping between components and their relationships. The proposed framework is validated using real-world data from an experimental HVAC system at the SENS-i Lab of the University of Campania “Luigi Vanvitelli,” using a limited set of variables typically available in building management and automation systems. The objective is not only to detect the presence of a fault but also to perform fault isolation by identifying the most likely affected component. The proposed hybrid approach combines data-driven modeling with expert knowledge, offering the advantage of operating without relying on labeled fault data. In contrast, two fully data-driven baselines, namely a Random Forest Classifier and a Conditional Gaussian Network, were developed under the assumption that labeled fault information was available. Considering performance across both winter and summer periods and a total of 26 faults, the models achieved fault isolation accuracies of 91% and 86%, respectively. The hybrid method reached a slightly lower accuracy of approximately 80% but offers significant benefits in terms of interpretability, transparency, and robustness when labeled data are scarce or unavailable. These characteristics make it particularly suitable for deployment in real HVAC systems, where practical limitations and the need for explainable results often outweigh small gains in predictive performance.