Electronic Thesis and Dissertation Repository

Thesis Format

Monograph

Degree

Doctor of Philosophy

Program

Electrical and Computer Engineering

Supervisor

Wang, Xianbin

2nd Supervisor

Shami, Abdallah

Co-Supervisor

Abstract

Based on the evolving communications, computing and embedded systems technologies, Internet of Things (IoT) systems can interconnect not only physical users and devices but also virtual services and objects, which have already been applied to many different application scenarios, such as smart home, smart healthcare, and intelligent transportation. With the rapid development, the number of involving devices increases tremendously. The huge number of devices and correspondingly generated data bring critical challenges to the IoT systems. To enhance the overall performance, this thesis aims to address the related technical issues on IoT data processing and physical topology discovery of the subnets self-organized by IoT devices.

First of all, the issues on outlier detection and data aggregation are addressed through the development of recursive principal component analysis (R-PCA) based data analysis framework. The framework is developed in a cluster-based structure to fully exploit the spatial correlation of IoT data. Specifically, the sensing devices are gathered into clusters based on spatial data correlation. Edge devices are assigned to the clusters for the R-PCA based outlier detection and data aggregation. The outlier-free and aggregated data are forwarded to the remote cloud server for data reconstruction and storage. Moreover, a data reduction scheme is further proposed to relieve the burden on the trunk link for data uploading by utilizing the temporal data correlation. Kalman filters (KFs) with identical parameters are maintained at the edge and cloud for data prediction. The amount of data uploading is reduced by using the data predicted by the KF in the cloud instead of uploading all the practically measured data.

Furthermore, an unmanned aerial vehicle (UAV) assisted IoT system is particularly designed for large-scale monitoring. Wireless sensor nodes are flexibly deployed for environmental sensing and self-organized into wireless sensor networks (WSNs). A physical topology discovery scheme is proposed to construct the physical topology of WSNs in the cloud server to facilitate performance optimization, where the physical topology indicates both the logical connectivity statuses of WSNs and the physical locations of WSN nodes. The physical topology discovery scheme is implemented through the newly developed parallel Metropolis-Hastings random walk based information sampling and network-wide 3D localization algorithms, where UAVs are served as the mobile edge devices and anchor nodes. Based on the physical topology constructed in the cloud, a UAV-enabled spatial data sampling scheme is further proposed to efficiently sample data from the monitoring area by using denoising autoencoder (DAE). By deploying the encoder of DAE at the UAV and decoder in the cloud, the data can be partially sampled from the sensing field and accurately reconstructed in the cloud.

In the final part of the thesis, a novel autoencoder (AE) neural network based data outlier detection algorithm is proposed, where both encoder and decoder of AE are deployed at the edge devices. Data outliers can be accurately detected by the large fluctuations in the squared error generated by the data passing through the encoder and decoder of the AE.

Summary for Lay Audience

Based on the evolving communications, computing and embedded systems technologies, the Internet of Things (IoT) can interconnect not only physical users and devices but also virtual services and objects, which have already been pervasively deployed. With the rapid development, the number of involving devices increases tremendously. The huge number of devices and generated data bring critical challenges. To enhance the overall performance, this thesis aims to address the related issues on IoT data processing and physical topology discovery of the subnets self-organized by IoT devices.

Firstly, the issues on outlier detection and data aggregation are addressed through the development of recursive principal component analysis based data analysis framework. The framework is developed in a cluster-based structure to fully exploit the spatial data correlation. Moreover, a temporal data correlation based reduction scheme is further proposed to reduce the amount of data uploading, which is implemented by using the data predicted by the Kalman filters in the cloud instead of uploading all the practically measured data.

Furthermore, an unmanned aerial vehicle (UAV) assisted IoT system is designed for large-scale monitoring, where UAVs are served as the mobile edge devices. Specifically, wireless sensor nodes are flexibly deployed for environmental sensing and self-organized into wireless sensor networks (WSNs). The physical topology of WSNs unveils the logical connectivity statuses of WSNs and the physical locations of nodes, which can facilitate system performance optimization. Thus, a physical topology discovery scheme is proposed to construct the physical topology in the cloud. Moreover, a UAV-enabled spatial data sampling scheme is further proposed to efficiently sample data from the monitoring area by using denoising autoencoder (DAE). By deploying the encoder of DAE at the UAV and decoder in the cloud, the data can be partially sampled from the area and accurately reconstructed in the cloud.

In the final part, a novel autoencoder based data outlier detection algorithm is proposed, where both encoder and decoder of autoencoder are deployed at the edge devices. Data outliers can be accurately detected by the large fluctuations in the squared error generated by the data passing through the encoder and decoder.

Download

Share

COinS