Intelligent Building Energy Management via Edge-Computed Fault Diagnosis

Thesis Abstract

The study addresses the escalating energy inefficiencies and reliability challenges in contemporary commercial buildings, where centralized energy management systems struggle to promptly detect and isolate faults across heterogenous HVAC, lighting, and building automation components. The aim is to develop an edge-computed fault diagnosis framework integrated with intelligent energy management to enhance real-time decision-making, reduce energy waste, and improve occupant comfort. Specific objectives are (1) to design an edge-enabled sensing and analytics architecture capable of real-time fault detection for HVAC and lighting subsystems; (2) to formulate a hybrid analytical model combining machine learning, physics-informed reasoning, and signal processing for fault diagnosis; (3) to quantify energy savings, peak-demand reductions, and occupancy-utility trade-offs under varied operating scenarios; (4) to evaluate the robustness of the framework under cyber-physical disruptions and data privacy constraints; and (5) to validate the approach in a real-world building pilot with scalable deployment prospects. The methodology adopts a mixed-methods approach anchored in a pragmatic research design. The population comprises 12 commercial office buildings within a metropolitan campus network, with a pilot implementation in one building and a cross-validation sample of 11 others. A stratified sampling technique selects 3 office zones per building, yielding 36 monitored zones. Data collection employs a multi-sensor suite including IoT-enabled temperature, humidity, CO2, occupancy, power meters, and actuator state logs, complemented by archival energy consumption data spanning 12 months. Instruments include calibrated smart meters, sub-metering for HVAC and lighting circuits, and a fault-annotation protocol informed by facility maintenance records. The edge analytics framework processes streaming data at 1–5 Hz on micro-edge devices, while a cloud repository archives long-term data for retrospective analysis. The diagnostic model integrates supervised learning (Random Forests, Gradient Boosting) for fault classification, unsupervised anomaly detection (Isolation Forest, One-Class SVM) to identify unseen fault patterns, and physics-informed constraints derived from first-principles energy balance. A Bayesian decision layer fuses edge inferences with prior knowledge to produce robust fault hypotheses. The model is trained and validated with a 70/30 train-test split, augmented with k-fold cross-validation (k=5). Feature engineering emphasizes transient energy signatures, system duty cycles, and thermodynamic consistency checks. Validity and reliability are ensured through instrument calibration, test-retest procedures, and inter-raciner reliability checks of fault labels. Data analysis employs regression analysis to quantify energy impact, ANOVA to compare performance across scenarios, and time-series analysis for fault progression. Model performance is evaluated using confusion matrices, F1-scores, ROC-AUC, and energy-saving metrics. An ethical framework addresses data privacy, informed consent for monitoring occupants, and secure edge–cloud communications. Anticipated findings indicate that edge-computed fault diagnosis can achieve real-time fault detection with 92–96% precision for HVAC anomalies and 88–92% precision for lighting faults, enabling adaptive reconfiguration strategies that yield 12–18% reductions in residential-equivalent energy use and 8–12% peak demand reductions in the pilot building. The framework is expected to demonstrate resilience to network latency and data loss, maintaining stable performance with marginal degradation under partial sensor failure. The study contributes to knowledge by advancing a scalable, edge-centric methodology for integrated fault diagnosis and energy management, bridging data-driven analytics with physics-based reasoning, and providing a replicable blueprint for retrofit in existing buildings. The theoretical implications include a synthesis of edge intelligence with built-environment control theory, extending the applicability of ensemble and Bayesian fusion in real-time energy optimization. Practical implications encompass improved occupant comfort through proactive fault mitigation, enhanced maintenance scheduling via diagnostic confidence scoring, and quantified pathways to near-term energy and carbon reductions in commercial portfolios. The main conclusion posits that edge-computed fault diagnosis significantly enhances the responsiveness and effectiveness of intelligent building energy management, outperforming cloud-only baselines in detection speed, fault coverage, and energy savings. Recommendations include standardizing edge analytics interfaces for interoperability, developing a library of physics-informed feature templates for common building systems, implementing a staged retrofit roadmap prioritizing high-energy-impact zones, and extending the framework to incorporate occupant behavioral models for further optimization under real-world operating conditions.