Smart Building Diagnostics via Edge AI for Fault Detection and Maintenance

Thesis Abstract

The rapid urbanization and aging infrastructure of modern buildings intensify the need for proactive fault detection and maintenance to ensure occupant comfort, safety, and energy efficiency, yet traditional schedules and remote diagnostic methods fail to deliver timely, scalable insights in complex building systems. This study investigates an edge AI-enabled fault detection and maintenance framework for smart buildings, addressing the gap between centralized cloud-centric analytics and real-time on-site decision support. The aim is to design, implement, and validate a deployable edge intelligence solution that detects faults in building automation systems (BAS), HVAC, electrical networks, and envelope diagnostics, and prescribes maintenance actions with minimum human intervention. The specific objectives are (1) to develop an edge-based sensor fusion architecture integrating HVAC, electrical, IoT, and envelope sensors for real-time anomaly detection; (2) to formulate and validate lightweight machine learning models (including LSTM networks, gradient boosting, and autoencoders) optimized for on-device inference under constrained processing power and energy budgets; (3) to compare rule-based, statistical, and data-driven approaches for fault classification and remaining useful life (RUL) estimation; (4) to evaluate maintenance decision support via a cost-benefit analysis and lifecycle performance indicators; (5) to assess the framework’s generalizability across commercial and residential building typologies; and (6) to establish governance and security protocols for edge deployment. The research adopts a pragmatic, mixed-methods design combining quantitative model development with qualitative expert validation. The population comprises ten representative mid-rise commercial buildings and five large residential complexes within a metropolitan district, with a total of 1,200 deployed sensors and 40,000 hours of historical BAS logs. A stratified sampling approach yields a dataset of 600,000 labeled events for fault instances across equipment classes, complemented by 24 expert interviews and three focus groups with facilities managers. Data collection instruments include standardized sensor readings, fault incident logs, maintenance records, and interview protocols; anomalies are labeled using domain-specific ontologies and a semi-supervised labeling process to address data sparsity. Quantitative analysis employs time-series preprocessing, feature engineering (temporal windows, spectral features, and cross-domain correlations), and model evaluation using precision, recall, F1-score, ROC-AUC, and calibration curves. Edge deployment is implemented on device-grade microcontrollers and gateways with TensorFlow Lite and PyTorch Mobile, emphasizing real-time inference latency below 500 ms per inference and energy consumption under 2.5 W per node. Model selection and validation leverage nested cross-validation, ablation studies, and SHAP value interpretation for explainability. A comparative analysis against cloud-based baselines examines latency, reliability, and total cost of ownership. Theoretical grounding draws on the Activity Theory for human-system interaction and the Maintenance Theory of Reliability, supplemented by the Diffusion of Innovations framework to evaluate adoption dynamics. Anticipated findings indicate that edge-based models achieve comparable fault detection accuracy with significant reductions in data transmission, leading to faster maintenance recommendations and lower energy use. Specifically, an ensemble model combining an autoencoder for anomaly detection and a gradient-boosted tree classifier for fault categorization is expected to yield F1-scores above 0.92 for common faults (sensor drift, actuator sticking, and compressor cycling) and a 15–25% improvement in maintenance scheduling efficiency relative to preventive-only regimes. The study contributes to knowledge by (i) advancing a practical, scalable edge AI architecture for multipurpose building diagnostics; (ii) providing a rigorously validated methodology for on-device fault detection and RUL estimation across diverse systems; and (iii) offering a framework for cost-effective maintenance planning with security and governance considerations for edge deployments. The main conclusion anticipated is that edge AI-enabled diagnostics can deliver timely fault detection with reduced cloud dependency, enabling proactive maintenance and higher building performance, while the recommendations propose standardized data schemas, model governance, protocol for incremental rollout across building portfolios, and policy guidelines for data privacy, cybersecurity, and interoperability standards. Further studies are suggested to explore transfer learning across climates, integration with occupant comfort models, and long-term impact assessment on resilience and sustainability metrics.