Spatiotemporal IoT Data Analysis and Prediction using OGC Open Standards in Digital Contact Tracing and Air Quality Prediction Applications

Abstract

Internet of Things (IoT) devices have become more and more integral to our everyday lives by the advent of sensors, computing and communications technology. The increasing use of IoT devices has caused a demand for the processing and analysis of spatial and temporal data. IoT devices such as smartphones, Bluetooth Low Energy beacons, and air quality sensing systems collect various types of spatiotemporal data used by various applications. This M.Sc. thesis focuses on two research applications of spatiotemporal IoT data analysis. First, an IoT system is designed to automatically trace human contacts with contaminated places and diagnosed carriers in indoor environments. This designed digital contact tracing system aims to find and notify possibly exposed people at the earliest possible stage to "flatten the curve" of spreading coronavirus. This research proposes a hierarchical graph-based data model that provides contact tracers with different granularity levels in spatial, temporal, and contextual dimensions. The spatial dimension of our data model is designed based on the Open Geospatial Consortium (OGC) IndoorGML standard. In comparison to other digital contact tracing systems, some key differences can be noted for this IoT-based contact tracing system. First, most digital contact tracing applications have been focused on person-to-person interactions. However, both person-to-person and person-to-place types of spreading infectious diseases have been considered in our system. Second, using OGC IndoorGML standard and considering topological relationships between indoor spaces, a validation procedure is designed to detect noisy indoor trajectory points. Third, different types of low-price BLE beacons have been utilized in this application. Finally, various user contexts (e.g., activity type and vulnerability to be exposed by coronavirus) are considered in an enhanced digital contact tracing application. Evaluating the functionality of our proposed system proved that the proposed contact tracing system had recognized 44.98% fewer possible contacts in comparison to traditional contact tracing applications. Also, 73.53% of noisy indoor trajectory points have been recognized using the proposed validation procedure. In all, the proposed data model and designed IoT architecture provide more flexibility in considering additional spatiotemporal information for indoor applications. Second, a mixed edge-based and cloud-based framework have been designed to predict air quality using IoT sensors based on deep learning methods. In this application, the most challenging problem is dealing with inconsistent data provided by IoT sensors. The term "inconsistent" refers to the fact that the actual time series from IoT sensors are generally incomplete. Thus, performing predictions on inconsistent spatiotemporal data will provide lower quality results. Hence, data preprocessing comprising filling up missing data should be of equal importance and sometimes more important than providing a prediction model. In this application, an edge-based data preprocessing approach is proposed to predict missing values and aggregate the raw PM2.5 measurements collected by PM2.5 IoT sensors. Both spatial and temporal information has been considered in edge-based preprocessing to improve the estimated values for missing measurements. Also, three different aggregation techniques have been applied by the edge component to reduce data granularity from minutes to an hour. The preprocessed data will then be combined with meteorological data and feed into the Long Short Term Memory (LSTM) technique representing the multivariate deep learning-based prediction model. To evaluate the significance of the proposed preprocessing technique, the prediction model's performance was evaluated on both preprocessed and unprocessed datasets. The proposed preprocessing technique improved air quality prediction accuracy by 40.18 percent on average.