Evaluating learning strategies for fault diagnosis in HVAC systems under labeled data scarcity: toward a metamodel-based transfer learning approach

Maintenance strategies in tertiary buildings, particularly for Heating, Ventilation and Air Conditioning (HVAC) systems, are still predominantly based on preventive approaches, with fixed intervention schedules aimed at minimizing unexpected failures and reducing corrective actions. In critical infrastructures such as hospitals and data centers, system redundancy is commonly employed to ensure operational continuity. While these strategies provide acceptable reliability, they are not optimized for resource efficiency, prompting the need for predictive maintenance solutions supported by fault detection and diagnosis (FDD) models. FDD functionalities have traditionally been developed using physical models, expert systems, and supervised machine learning (ML) techniques. However, physical models are costly to calibrate and often impractical for complex equipment, while supervised ML approaches require large volumes of labeled data, particularly for faulty states, which are rarely available in real-world conditions. To address these limitations, recent research has explored data augmentation techniques (e.g., Synthetic Minority Over-sampling Technique, Generative Adversarial Networks), semi-supervised learning paradigms (e.g., self-training), and transfer learning strategies. Despite their potential, these methods still face challenges in achieving robust performance under severe data imbalance and scarcity. To address these challenges, a novel methodology is introduced that combines physical modeling with instance-based transfer learning to enable scalable FDD development. Unlike previous approaches that require calibrated physical models of the target equipment to generate synthetic data, the proposed method leverages physical models of similar equipment (source domain) to simulate fault scenarios. These simulations are then used to train metamodels for the target equipment (target domain) via the proposed transfer learning approach, eliminating the need for extensive calibration and labeled data. Preliminary results using a publicly available constant air volume air handling unit dataset show promising operational advantages and diagnostic improvements with balanced accuracy increases of 0.35% in relation to semi-supervised and data augmentation methods under severe data scarcity conditions. Additionally, and as a necessary benchmark to contextualize and motivate the need for this new approach, the study presents an experimental evaluation of predictive modeling strategies under varying levels of labeled data availability. Using three publicly available HVAC datasets (constant and variable air volume air handling units and rooftop unit), seven supervised ML algorithms were assessed alongside data augmentation techniques, semi-supervised learning, and a hybrid approach combining both paradigms. The results indicate that, for the datasets considered, balanced accuracy values become acceptable for most algorithms when the number of labeled fault instances per class exceeds 100 (AHU datasets). Moreover, depending on the initial data availability, improvements in balanced accuracy of up to 15.24% (data augmentation), 13.08% (semi-supervised learning), and 19.98% (hybrid approach) were observed.