Machine Learning Algorithms for Smart City Applications


Student thesis: Doctoral Thesis

View graph of relations


Related Research Unit(s)


Awarding Institution
Award date9 Jun 2020


Smart City is a sophisticated and multidisciplinary infrastructure that seamlessly incorporates a tremendous number of devices, networks, systems, and services to achieve sustainable city development and improve our quality of life. The Smart City development will render a remarkable impact on essential areas, such as economy, environment, mobility, and living. Internet-of-things (IoT) is one of the essential building blocks in the development of Smart Cities. It renders cost-effective and high-scalability solutions for establishing Smart City networks. There will be hundreds of billions of IoT connections in the coming decade and will generate a vast amount of data. Thus, Machine Learning is another crucial technology that can be exploited to efficiently and scientifically analyse the IoT data and make applications smart. Nonetheless, there are no universal Machine Learning algorithms that can cater for all kinds of applications. Therefore, in this thesis, four contributions have been developed to demonstrate the feasibility of exploiting Machine Learning in Smart City applications.

First, a new supply chain solution, namely Supply-chain-of-things (SCoT), is proposed to integrate IoT and Machine Learning into supply chain management, aiming to increase the efficiency and productivity of the supply chain in the Smart Economy. The SCoT defines high-level connection guidance, namely Reliable and Ubiquitous Connectivity (RUC), to categorise different types of connections in the Smart Supply Chain Network (SSCN) and discuss the requirements for each connection category. Since the SSCN will generate a vast amount of data, a new SCoT platform is proposed to analyse the data and coordinate various supply chain operations by incorporating Machine Learning and emerging IoT standards, such as the IEEE P2668. As a result, the SCoT will deliver five vital functions, including (I) adaptive quality management, (II) systematic solution integration, (III) intelligent logistics, (IV) analytical decision-making, and (V) transparent end-to-end traceability.

The second contribution aims to render an efficient tree health evaluation scheme for the Smart Environment. The developed evaluation scheme defines a new tree health indicator, namely Interne-to-things Tree Health Index (ITHL), and it incorporates urban trees with various IoT sensors, i.e., the IoT trees, to enable automated data collection. Afterwards, a new Machine Learning algorithm, namely Heterogeneous Neural Network (HNN), is designed for the ITHL classification. The HNN-based evaluation scheme analyses seven Dynamic Attributes (DAs) and seven Static Attributes (SAs), which are collected from the IoT sensors and related to the environmental impact on tree health. Since there are two distinct types of input attributes, the HNN embraces two distinct learning models for identifying the underlying correlations between the attributes and tree health. Experimental results show that the classification accuracy of the HNN is 35% – 65% higher than that of other Machine Learning approaches. The HNN-based tree health evaluation scheme achieves a low misclassification rate of 4.7%.

In the third contribution, an Electrocardiogram (ECG) based Drunk Driving Detection (DDD) scheme is developed to cater to the safety requirement in the Smart Mobility. After the analysis of the ECG patterns under normal and drunk conditions, ten attributes are extracted to construct the DDD classifier. Besides, a weighted kernel is designed to customise the importance level of each attribute into the DDD classifier for improving detection accuracy. The results verify that the designed ECG-based DDD achieves 18% improvement of detection accuracy compared to other existing DDD schemes.

Finally, a Multiclass and Assembled Classifier (MAC) is developed for ECG-based Stress Level Classification (SLC) in an attempt to provide an efficient method for remote mental healthcare in the Smart Living. The designed SLC considers three classes, namely Low-Stress Level (LSL), Mid-Stress Level (MSL), and High-Stress Level (HSL). The MAC is an ensemble learning structure that customises and assembles fiducial and non-fiducial classifications to achieve a compromise between classification accuracy and practicability. Evaluation results show that compared with other Machine Learning algorithms, the MAC achieves 11% – 41% improvement of model fitness and 26% improvement of classification accuracy. Moreover, the classification accuracy of the MAC-based SLC is 30% higher than that of other state-of-the-art works.