• Temporal–contextual self-supervised backbone is proposed for AHU AFDD • Proposed method achieves ∼80% F1 score with 5% labels and >92% F1 score with 15% labels • Proposed method reaches ≈99% F1 score and 99.5% accuracy with 30% labels across three datasets • Benchmarking is conducted against supervised and self-supervised baselines • Semi-supervised learning baselines are included under limited-label settings 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. To reduce the labeling burden, recent work has explored self-supervised learning to leverage unlabeled operational logs; however, many studies treat sensor measurements as static tabular vectors rather than sequential time series. This study adopts Time-Series representation learning via Temporal and Contextual Contrasting (TS-TCC) as a backbone for AHU-level AFDD. TS-TCC pretrains a transformer-based encoder on unlabeled operational data using temporal and contextual contrastive objectives and then fine-tunes a lightweight classifier using limited labeled samples. Experiments on three benchmark AHU datasets with 5–30% label budgets compare TS-TCC with representative self-supervised baselines, supervised detectors, and semi-supervised methods. Across datasets, TS-TCC achieves 79–81% macro F1 score with 5% labels and reaches approximately 99% macro F1 score with 30% labels, outperforming self-supervised baselines and matching strong supervised detectors. Under a 30% labeled + 70% unlabeled setting, TS-TCC attains near-ceiling performance, while consistency-based semi-supervised methods provide only marginal F1 score gains in some cases. The resulting model remains suitable for near real-time deployment, with inference times of 20–30 ms per instance.
Seunghyeon Wang (Sat,) studied this question.