ENHANCING BUILDING FAULT DETECTION, DIAGNOSTICS, AND PROGNOSTICS THROUGH A HYBRID PHYSICS-INFORMED MODELING FRAMEWORK FOR OVERCOMING QUANTITY AND TEMPORAL DATA SCARCITY

Effective Fault Detection, Diagnosis, and Prognosis (FDDP) in buildings is severely hindered by the scarcity of real-world fault data. This thesis proposes a hybrid physics-informed modeling framework to overcome both data quantity scarcity and temporal sparsity. First, Hybrid Conditional Generative Adversarial Network (HCGAN) is developed. By leveraging physics-based simulations as conditional priors, it generates high-fidelity synthetic fault data, enabling zero-shot diagnosis without requiring real fault samples. Second, to address temporal sparsity in prognosis, Dual-Calibrated Particle Filter (DC-PF) is proposed. It fuses uncalibrated surrogate models with sparse observations to accurately estimate Remaining Useful Life (RUL). Experimental results demonstrate significant performance gains: HCGAN improves data fidelity by 50% over baselines, while DC-PF reduces prognosis error by roughly 57%. Integrated into a unified pipeline, this framework transforms sparse operational data into reliable, proactive maintenance decisions, providing a practical path for scalable smart building management.

Author dc:creator

• HAN JINTONG

• Building Automation
• Remaining Useful Life
• Generative Data Augmentation
• Predictive Maintenance
• Fault Prognosis
• Fault Detection and Diagnosis

Licence

• https://scholarbank.nus.edu.sg/bitstreams/cfe3d8b3-2146-4634-be56-4d7b271796c1/download