Adaptive Bayesian Spatio-Temporal Models for IoT Sensor Networks

Thesis Abstract

The proliferation of Internet of Things (IoT) sensor networks across smart cities and industrial environments generates high-velocity, spatially structured data streams that challenge conventional statistical approaches to real-time inference and decision making. This study addresses the problem of robust, scalable inference for IoT sensor networks under non-stationary conditions, missing data, and evolving spatial dependencies. The aim is to develop adaptive Bayesian spatio-temporal models that can learn from streaming data, quantify uncertainty, and deliver timely predictions for anomaly detection, fault diagnosis, and environmental monitoring. Specific objectives are to (1) formulate a hierarchical Bayesian framework that integrates spatial Gaussian processes with temporal dynamic models to capture evolving dependencies among heterogeneous sensors; (2) develop an online updating scheme based on sequential Monte Carlo and variational Bayes to accommodate non-stationarities and concept drift; (3) implement an adaptive model selection mechanism leveraging Bayes factors and information criteria to switch among latent structure specifications as network conditions change; (4) validate the framework with simulated IoT networks and a real-world urban monitoring dataset comprising 5000 sensors across five districts, with ground-truth events for anomaly evaluation; and (5) assess computational performance, scalability, and robustness to missing data and sensor failures. The methodology employs a hybrid research design combining simulation experiments and empirical validation. The population comprises heterogeneous IoT sensor networks deployed in smart-city corridors and industrial facilities. A stratified sampling approach yields a synthetic network of 4000 spatially distributed sensors for controlled simulations and a real-world dataset of 5000 sensors from a metropolitan environmental monitoring program. Data collection instruments include simulated sensor streams generated under controlled non-stationarity, drift, and bursty missingness, complemented by open-access urban air quality and traffic sensor feeds. The analysis uses a Bayesian hierarchical model that couples a spatial Gaussian process with a temporal state-space component, implemented within an online learning framework. Key analytical techniques include sequential Monte Carlo (particle filtering) for posterior updating, variational Bayes for scalable approximation, Gaussian process priors with non-stationary kernel adaptations, and Bayes factor-based model selection. Model validation leverages predictive accuracy, continuous ranked probability score (CRPS), and coverage of predictive intervals. The research also incorporates a theoretical grounding in hierarchical Bayesian inference and adaptive learning theory, drawing on the concepts of non-stationary spatial-temporal processes and concept drift in streaming data. Expected findings indicate that the adaptive Bayesian spatio-temporal models will deliver improved predictive accuracy and calibrated uncertainty quantification relative to baseline spatio-temporal models under non-stationary conditions. In particular, it is anticipated that online updating will rapidly detect regime changes, maintain robust performance during sensor outages, and accurately localize anomalies through posterior anomaly scores. The framework should demonstrate superior CRPS and reliable 95% predictive interval coverage for both synthetic and real-world datasets. Computational experiments are expected to reveal favorable scalability properties, with sub-linear growth in inference time as the network size increases and effective handling of up to 20% random missing observations without substantial degradation. The study contributes to knowledge by advancing adaptive Bayesian methods for high-dimensional, non-stationary spatio-temporal data in IoT contexts, offering a unified approach that integrates online inference, model selection, and uncertainty quantification. It provides a practical blueprint for deploying real-time analytics in smart-city and industrial IoT environments, including guidelines for kernel selection, prior specification, and computational resource planning. The research also extends theory on non-stationary spatial processes and dynamic hierarchical modeling with streaming data, contributing to both methodological and empirical literatures. The main conclusion is that adaptive Bayesian spatio-temporal modeling, with online updating and data-driven structure adaptation, substantially enhances real-time decision support in IoT sensor networks. Recommendations include extending the framework to multi-modal sensor fusion, exploring distributed implementations for edge-computing architectures, and developing domain-specific benchmarks to standardize evaluation across diverse IoT deployments.