Increasing transferability for automated fault detection and diagnosis in HVAC systems through a hybrid AI methodology

In this study, we present a methodology combining a rule-based system and machine learning (ML) models for an automated residual generating fault detection and diagnosis (FDD) process in Heating, Ventilation, and Air Conditioning (HVAC) systems. This hybrid approach reduces implementation efforts and enhances adaptation across various building systems. Residual-generating FDD models nominal system behaviour and detects faults by analysing differences between estimated and observed data. For residual generation, we apply a bank of Multilayer Perceptron (MLP) models, with hyperparameters tuned through Bayesian optimisation. To capture the fault-free behaviour, we first filter the historical operating data using the rule-based system. For reliable fault detection (FD), residuals were evaluated using a variable tolerance band and scoring method. We demonstrated our methodology on nine air handling units (AHUs) in three different buildings. Our results demonstrate that filtering faulty states enhances detection sensitivity by increasing residual magnitude during faults. However, this may reduce model generalisation and accuracy for underrepresented operating modes in the training data. During application, both methods operate in parallel, reducing the need for extensive rule sets while improving FDD quality, as the residual generation method can detect unknown faults. Compared to individual application, the hybrid approach achieves higher FD rates. Despite its higher sensitivity, the residual-generating method shows a vulnerability to false alarms, especially when the nominal behaviour of the systems is insufficiently learned. Nevertheless, false alarm rates remain below 1% when the models are well-trained. Combining both approaches improved the overall FDD performance by leveraging the strengths of each method. Additionally, the integration of ML-Models into a residual generating FDD process enables a better interpretation of results, thereby enhancing transparency and trust in AI solutions in practice. The proposed methodology offers a fully automated, data-driven FDD approach that requires only minimal parameterisation of rules, enabling efficient fault detection in HVAC systems.