A label-free hybrid fault detection and diagnosis approach for HVAC systems using bayesian networks

• A Bayesian network-based hybrid FDD framework for HVAC systems was developed. • Random Forest models were used to define baselines and virtual fault evidence. • The framework requires standard operational variables and no labeled fault data. • Semantic metadata and expert knowledge enabled an efficient BN construction. • A fault detection and isolation accuracy of about 90 % was achieved on two systems. • Targeted statistical analyses were used to perform component-level fault diagnosis. The building sector accounts for up to 40 % of global energy demand, with HVAC systems responsible for nearly half of this consumption. Faulty HVAC operation can result in substantial energy waste, reduced equipment lifespan, and increased operational costs. To address these challenges, the scientific community has focused on developing Fault Detection and Diagnosis (FDD) strategies that are both accurate and applicable in real conditions. Although data-driven approaches have shown strong potential, their practical deployment remains limited by the need for labeled data and variables not commonly available in real buildings. This paper presents a hybrid FDD framework based on Bayesian Networks (BNs) that combines data-driven models with expert knowledge. A set of reference models, using Random Forest algorithms, was developed to predict key variable values and define a baseline. Deviations from this baseline, expressed as residuals, were converted into virtual evidence and combined with hard evidence derived from domain expertise. These inputs were fed into a set of BN models, one for each operational mode, whose parameters and system-level structure were informed by expert knowledge and constructed efficiently through semantic metadata schemas based on brick ontology. The BNs performed fault detection and component isolation while the diagnosis is supported by targeted statistical analyses tailored to each system component. The main advantage of the proposed framework is that it requires only variables typically available in building management systems and does not rely on the a-priori knowledge of fault labels. The approach was validated on a simulated dataset from the Single Duct Air Handling Unit developed by the LBNL and further tested on a Fan Coil Unit. It achieved detection and isolation accuracies of approximately 91 % and 87 %, respectively, confirming its robustness, adaptability, and practical relevance.