Edge-AI Middleware for Real-Time Smart Campus IoT Analytics

Thesis Abstract

The rapid proliferation of Internet of Things (IoT) sensors and devices on university campuses presents an opportunity to transform operational efficiency, safety, and student experience through real-time analytics; however, centralized cloud processing introduces latency, bandwidth costs, and privacy concerns that hinder timely decision-making in dynamic campus environments. This study addresses the problem of delivering low-latency, privacy-preserving IoT analytics at the network edge to support real-time decision-making for energy management, safety monitoring, and space utilization on a large university campus. The aim is to design, implement, and evaluate an edge-AI middleware architecture that orchestrates on-device and edge-server inference, data fusion, and policy-driven actions to enable scalable, context-aware campus analytics. Specific objectives are (1) to architect an edge middleware framework that supports heterogeneous sensors (environmental, occupancy, CCTV metadata) and lightweight AI models; (2) to develop a composable data governance and privacy layer, including data minimization and access controls aligned with campus policies; (3) to implement real-time inference pipelines for energy optimization, anomaly detection, and occupancy-based space planning; (4) to evaluate system performance under varying load, network conditions, and privacy constraints; and (5) to provide a decision-support toolkit for facilities management and campus security. The methodology adopts a mixed-methods research design, combining system development with empirical evaluation in a university campus testbed spanning 40 buildings and 1200 networked devices. The population comprises campus IoT endpoints, edge nodes, and facilities staff users. A stratified sampling approach selects 60 edge devices and 12 edge servers for deployment, accompanied by 40 facilities personnel for usability and acceptance testing. Data collection instruments include (i) sensor logs (temperature, humidity, occupancy counts, motion events, access control metadata), (ii) edge and cloud inference results, (iii) system performance metrics (latency, throughput, CPU/memory usage, energy consumption), (iv) privacy and governance logs (data access requests, policy violations), and (v) structured interviews and surveys with facilities staff. Validity and reliability are ensured through instrument triangulation, pilot testing, and inter-rater reliability checks for qualitative interviews. Analytical techniques comprise time-series analysis and regression to assess energy savings and occupancy accuracy; ANOVA to compare performance across configurations (edge-only, hybrid edge-cloud, and cloud-only); and anomaly detection using unsupervised learning (Isolation Forest) for security and safety events. A Bayesian hierarchical model quantifies uncertainty in inference results across buildings. The study also employs thematic analysis for qualitative interview data to capture user-perceived usability and governance concerns. Theoretical grounding draws on the Technology-Organization-Environment (TOE) framework to explain adoption and deployment determinants, and the Actor-Network Theory (ANT) to analyze the interdependencies among devices, middleware, facilities staff, and policies. A conceptual model illustrates data flow from heterogeneous sensors through edge inference to decision actions, with feedback loops for policy updates. Expected findings indicate substantial reductions in end-to-end latency (mean 72 ms for critical events vs. 350 ms in cloud-only configurations), improved energy efficiency by up to 18% during peak hours, and higher accuracy in occupancy-led space utilization metrics (F1-score > 0.92) compared with centralized processing. The privacy-preserving data governance layer is anticipated to reduce unnecessary data transmission by 45% through selective data summarization and on-device inference. Qualitative results are expected to reveal strong user acceptance for real-time alerts and governance controls, with identified challenges in policy maintenance and interoperability across legacy systems. The study will contribute to knowledge by integrating edge-AI middleware design with rigorous governance, providing a replicable blueprint for scalable, real-time campus analytics and informing broader smart campus deployments. The main conclusion posits that edge-centric AI middleware can deliver real-time, privacy-conscious analytics at scale, outperforming cloud-centric approaches in latency-sensitive campus contexts while enabling actionable insights for facilities management and safety. Recommendations include extending the middleware to support multi-campus deployments, adopting standardized data schemas for interoperability, and developing continuous learning mechanisms to adapt AI models to evolving campus patterns. Further research directions involve exploring federated learning for cross-campus model generalization, resilience against network disruptions, and extended governance frameworks to address regulatory changes.