First PRIN UTMOST FDD project newsletter: Hybrid artificial intelligence for automated fault detection in HVAC systems
Hybrid artificial intelligence for automated fault detection in HVAC systems
Heating, ventilation and air conditioning (HVAC) systems represent one of the largest energy consumers in modern buildings. In commercial buildings alone, HVAC operation can account for up to 50–60% of the total building energy demand, while faulty operation or improper control strategies may increase energy consumption by 15–30%.
To address these inefficiencies, researchers involved in the PRIN 2022 MUR - UTMOST FDD have recently developed a novel hybrid methodology for the automatic detection and diagnosis of faults in HVAC systems. The proposed approach combines machine learning techniques with expert engineering knowledge in order to create a diagnostic framework that is both accurate and applicable in real operational conditions.
Traditional Fault Detection and Diagnosis (FDD) methods typically follow either a data-driven or a knowledge-based approach. Data-driven techniques rely on machine learning models trained on historical data, while knowledge-based methods reproduce expert reasoning through diagnostic rules. Although both approaches offer advantages, they also present significant limitations when used individually. Data-driven methods often require large labelled datasets that are rarely available in real buildings, while knowledge-based systems may lack adaptability to different system configurations.
The research conducted in the framework of the project (Paolini, 2025) introduces a hybrid framework based on Bayesian Networks, which integrates both perspectives. In the methodology, Random Forest models are first used to establish baseline predictions of key operational variables. Deviations from this baseline are interpreted as residuals and converted into “virtual evidence”. These signals are then combined with expert-defined rules and processed through Bayesian Network models that represent the relationships between system components and operational variables.
Figure: Developed methodological framework
An important advantage of this framework is that it relies only on variables that are typically available in building management systems. This makes the approach suitable for practical implementation without requiring extensive sensor deployments or manually labelled fault datasets.
Another innovative aspect of the study is the use of semantic metadata models to define the structure of the Bayesian Network. By representing the relationships between sensors, components and control signals through standardized metadata schemas, the model structure can be generated more efficiently and potentially transferred across similar HVAC systems.
The proposed diagnostic framework was evaluated on two different HVAC systems: a Single Duct Air Handling Unit (SDAHU) and a Fan Coil Unit (FCU). The results demonstrated high diagnostic performance, with fault detection and isolation accuracies of approximately 91% for the SDAHU and 87% for the FCU, confirming the robustness and adaptability of the proposed method.
Beyond the methodological contribution, the research also highlights several future research directions. These include the development of FDD approaches that can operate without labelled training data, the deployment of diagnostic models in real building environments, and the integration of advanced technologies such as transfer learning and large language models to improve usability and automation.
The results of this study represent an important step toward the development of scalable and transparent diagnostic tools capable of improving building energy performance and supporting predictive maintenance strategies. By enabling earlier identification of operational faults, these technologies could significantly reduce energy waste while enhancing indoor environmental quality.
Further developments within the PRIN 2022 MUR - UTMOST FDD will focus on extending the methodology to additional HVAC systems and on validating the approach using experimental datasets and digital twin simulations.
Reference
Acknowledgements
The project is funded by the European Union in the framework of the initiatives "Next Generation EU"
