Investigating AI-based approaches for data processing on IOT edge devices

Abstract

The unprecedented dissemination of sensor technologies and their applications in different aspects of contemporary life has significantly contributed to the emergence of the Internet of Things (IoT). IoT envisions seamless integration of digital and physical worlds fabricating the ever witnessed stupendous network of connected devices. These devices have been harnessed to serve a diversity of applications including smart buildings, transportation, health, and agriculture. Fundamentally, collecting data is the main objective of such deployed networks. These data are constantly being generated leading to enormous volumes of raw sensory data that require further manipulation to convert it to valuable insight. Typically, IoT data are transmitted to the cloud for processing and to deduce a meaningful outcome. Additionally, several IoT use-cases require the cloud to send back some instruction to fulfill actuation purposes. Considering the ever-increasing number of IoT connected devices and the limited network capacity, several challenges have arisen. First, blindly transmitting data to the cloud has amplified the stress in the network, leading to network bottlenecks. In addition, the reliance on the central cloud topology has raised the cost of IoT data manipulation. This cost can be envisioned in the bandwidth cost, in addition to the cloud operational cost. Furthermore, several time-sensitive IoT applications might not tolerate the latency imposed by the round-trip from sensor nodes to the cloud. All aforementioned issues have stimulated investigating alternative solutions that push such computing capabilities to the network edge and close to data sources. Auspiciously, this has become feasible thanks to the recent diversity and advancement that has been seen in the IoT edge devices endowing them the capability to handle a variety of use-cases. This also has been simultaneously accompanied with a broad investigation of the applicability of several AI-based techniques to perform IoT data manipulation tasks. More specifically, several AI-based techniques including semantic web, machine learning, and fuzzy logic have been widely investigated and accepted as key enabler to reveal the ambiguity of raw sensor data and to support decision-making tasks. Although the advantages introduced by these techniques have been widely acknowledged in the IoT domain, the growing tendency to extend such intelligence to the network edge has revealed several challenges. This includes their applicability to perform under resource-restricted platforms, in addition to, their consistency in unstable environments that the IoT edge based solutions are highly exposed to. The main objective of this dissertation is to examine through several case studies the applicability of AI-based techniques in maintaining balanced efficiency in data processing and analysis tasks. This efficiency is obtained taking into consideration the resource-constrained aspect that is featured in a considerable portion of IoT edge devices. In addition, each case study represents a research problem that has been addressed individually in this thesis. Thus, the contributions introduced by this work correspond to the presented case studies and are structured sequentially interrelated objectives. The first contribution is proposed to assess the feasibility of edge-based solutions in performing timely and efficient event detection tasks based on the semantic web and complex event processing techniques. The feasibility of the semantic annotation process in performing efficient event detection and modeling, in addition to data filtration tasks on IoT edge devices was examined. Second, in order to support IoT edge solutions to not only detect events on time but also to predict events and contemplate them in advance, we have examined the combination of Machine Learning and fuzzy logic techniques. More specifically, we have investigated and applied best practices for efficient ML and FL-based solution deployment in resource-limited environments. Third, in view of the challenges of using ML models in real-time, we have investigated a fuzzy logic based approach to keep track of model performance. The proposed mechanism is intended to maintain an acceptable real-time model performance in accordance with balanced resource utilization on IoT edge devices. In all analyzed cases, we conducted a deep exploration of the related subjects of research and supported our assumptions by a thorough analysis of real-life use cases. In addition, we have validated our assumptions with several proof-of-concept implementations. Such methodology demonstrated progress in highlighting a variety of perspectives and recognizing open issues and potential research issues within the scope of the presented research.

Date	3 Jun 2021
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Abdelouahed Gherbi (Supervisor)