Fault-Aware Adaptive Temporal–Feature Fusion for Multi-Scale Transformer-Based Early Predictive Maintenance in Smart-Building HVAC Systems

Abstract

HVAC fault detection is essential for reliable smart-building operation and predictive maintenance, yet existing data-driven approaches often exhibit limited multi-scale temporal modeling and insufficient adaptation to fault-relevant sensor interactions. This study proposes a Multi-Scale Temporal Transformer with Adaptive Attention to address these limitations by jointly learning short-, medium-, and long-term HVAC operating patterns and dynamically fusing temporal and feature-level representations for early fault identification. The framework is implemented using Python with PyTorch and evaluated using a publicly available HVAC dataset containing 15-minute measurements from a non-residential building in Turin, Italy, including temperature, humidity, setpoint, fan power, and energy variables. The proposed model achieves 97.59% accuracy, 97.72% precision, 97.31% recall, 97.51% F1-score, and 0.987 AUC, outperforming the strongest baseline accuracy of 96.30% by 1.29 percentage points (1.34% relative improvement). The results demonstrate the potential of adaptive multi-scale temporal learning for reliable HVAC fault detection and early-warning predictive maintenance.