Second PRIN UTMOST FDD project newsletter: Transferring hybrid fault detection and diagnosis strategies in HVAC Systems
Transferring hybrid fault detection and diagnosis strategies in HVAC Systems
As part of the PRIN 2022 MUR - UTMOST FDD project, a recent study investigated how transfer learning techniques can support the deployment of hybrid Fault Detection and Diagnosis (FDD) strategies in real Heating, Ventilation and Air Conditioning (HVAC) systems. The work focuses on improving the transferability of diagnostic models between simulated environments and real operational installations, a key challenge for the practical adoption of data-driven maintenance tools in buildings.
The proposed framework follows the hybrid philosophy adopted throughout the PRIN 2022 MUR - UTMOST FDD project. Data-driven models are used to estimate the expected behavior of the system, while expert knowledge is embedded in a probabilistic reasoning structure to support fault diagnosis. In the first stage, ensembles of Random Forest regression models predict key operational variables of an Air Handling Unit, such as temperatures, humidity levels, fan power consumption and control signals. The deviations between predicted and measured values generate residual signals that indicate potential anomalies (Paolini, 2025).
In the second stage, these residuals are analyzed by a Bayesian Network, which incorporates expert knowledge about the relationships between system components and fault mechanisms. By combining multiple residual signals, the Bayesian Network estimates the probability of different fault conditions and supports component-level fault isolation. This hybrid structure enables the diagnostic framework to combine the adaptability of data-driven methods with the interpretability of knowledge-based reasoning.
A central aspect of the study concerns the transferability of the regression models used in the first stage of the diagnostic chain. In many real applications, only limited monitoring data are available when a system is first deployed. At the same time, simulation environments can provide large datasets representing nominal operating conditions. However, differences between simulated and real environments often lead to performance degradation when models trained on simulated data are applied directly to real systems.
To address this issue, the study compares three deployment strategies: direct application of models trained on simulation data (Approach 1), model training performed exclusively using real operational data (approach 2) and transfer learning through domain adaptation (approach 3). The analysis shows that boosting-based transfer learning methods, in particular the TrAdaBoostR2 algorithm, improve prediction accuracy especially when only limited target-domain data are available.
A sensitivity analysis further evaluates how the amount of real operational data influences diagnostic performance. The results show that transfer learning provides a clear advantage during the initial deployment phase, when only a few days of monitoring data are available. As the volume of real data increases, models trained directly on target-domain data progressively improve and can eventually reach comparable or higher diagnostic accuracy.
Figure: Performance gap in fault detection and isolation accuracy between Domain Adaptation (Approach 2) and Cumulative Training (Approach 3) as a function of the available fine-tuning data volume, evaluated under summer operating conditions.
These findings provide useful indications for the deployment of hybrid FDD systems in real buildings. Transfer learning can support reliable diagnostic performance during the early stages of system monitoring, while the accumulation of operational data allows the gradual transition toward models trained directly on real system behavior. The study therefore contributes to the development of scalable hybrid diagnostic approaches capable of supporting predictive maintenance in HVAC systems.
Reference
Acknowledgements
The project is funded by the European Union in the framework of the initiatives "Next Generation EU"
