Automated Fault Detection and Diagnosis (AFDD) for Air Handling Units (AHUs) has largely relied on supervised learning, which is difficult to deploy when labeled data are scarce and fault classes are imbalanced. Existing label-efficient AFDD studies often evaluate self-supervised schemes on simulated or laboratory datasets, frequently in tabular form, and therefore do not fully capture the temporal structure and noise characteristics of real operational logs. This study proposes Class-Aware Temporal and Contextual Contrasting (CA-TCC) for label-efficient AHU AFDD on real buildings. CA-TCC is a semi-supervised framework that first performs self-supervised temporal/contextual contrastive pretraining on unlabeled operational logs and then performs class-aware semi-supervised refinement using limited labels together with unlabeled data via confidence-filtered pseudo-labels; a lightweight classifier head is subsequently trained for fault classification. Experiments on six datasets—three synthetic AHU benchmarks and three real operational datasets from an auditorium, a hospital, and an office building—show that CA-TCC consistently outperforms alternative self-supervised backbones across label budgets. With only 5% labeled data, CA-TCC improves macro F1 score and accuracy by approximately 5–10 and 5–9 points, respectively, while remaining within about 1–1.5 points of strong fully supervised models Cross-building transfer experiments demonstrate reliable source-to-target generalization across all building pairs under consistent label budgets. Additional analyses evaluate backbone fine-tuning policies, sensitivity to CA-TCC-specific hyperparameters, per-class behavior under different label fractions, and ablations of temporal, contextual, and class-aware components. Comparisons with representative baselines and inference-speed measurements further characterize suitability for near real-time AFDD.
Seunghyeon Wang (Wed,) studied this question.