IoT traffic modelling and QoS prediction framework using advanced deep learning

The Internet of Things (IoT) refers to the billions of physical devices around the world that are connected to the internet, all collecting and sharing data. IoT includes a wide variety of devices such as smart home devices like thermostats, refrigerators, and security systems, industrial machines, wearable health monitors, and even smart city technologies. The expansion of IoT devices brings numerous benefits, yet it also poses challenges arising from the resource-limited characteristics of many such devices. These limitations, including constrained computing power, memory, and energy efficiency, can impact performance and responsiveness, particularly in the context of crucial, time-sensitive applications. The advent of Edge Computing (EC) offers a promising solution to these challenges. By situating computational resources in proximity to IoT devices, EC can minimize latency and enable efficient data processing. Nonetheless, the unpredictable nature of IoT data generation, given the diversity of device types and their activity cycles, introduces its own set of complexities. Moreover, the employment of wireless communication and unlicensed frequency bands compounds issues related to data transmission efficiency, interference, and limited capacity, potentially leading to increased latency within IoT networks. Adding to this intricate communication framework is the variable network connectivity of mobile IoT devices such as drones or robots, which contributes to the existing complications. Therefore, it becomes imperative to predict the time-varying characteristics of IoT devices to better manage resources and ensure Quality of Service (QoS). Grouping similar IoT devices can improve workload estimates and facilitate resource allocation. Moreover, QoS prediction models that incorporate time series data of packet generation and network characteristics can contribute to a better understanding of the IoT platform’s requirements and limitations. \n \nIn this thesis, we first consider the categorization of traffic from IoT devices into distinct classes. This approach is intended to facilitate the managing and understanding of the diverse range of data coming from various IoT devices within a smart city environment. We model the IoT devices traffic identification problem as a multi-classification learning problem, and propose a two stage learning model as a solution. In order to capture the nuances of IoT device behavior, we suggest an extended feature set, which includes flow, packet, and device-level features. These features describe the characteristics and behavior of IoT devices in a smart city context. We further propose a custom weighting pre-processing algorithm to determine the contribution of traffic data features to the classification process. Subsequently, an ANOVA-based feature selection mechanism and the Pearson’s coefficient-based correlation method are employed for a more profound understanding of the traffic characteristics (features). Lastly, we propose an innovative, two-stage learning algorithm, which utilizes logistic regression at stage 0 to initially categorize the data, and a Multi-Layer Perceptron (MLP) network at stage 1 to refine the classification. This two-stage process allows for more accurate classification of IoT devices, as demonstrated with two different real IoT traffic datasets. \n \nThe second aspect of the work presented in this thesis involves the efficient prediction of various QoS metrics, such as throughput, Packet Delivery ratio (PDR), Packet Loss Ratio (PLR) and latency for a diverse set of IoT applications. We propose a temporal transformer model within a unified system, designed to predict typical QoS metrics. This prediction is model as a time series forecasting problem, utilizing both univariate and multivariate setups. We generate a dataset comprised of real-time traffic information from five distinct IoT applications: Heating, Ventilation, and Air Conditioning (HVAC), smart lighting, Voice over Internet Protocol (VoIP) for a virtual assistant, surveillance, and emergency response. The proposed temporal transformer model leverages an attention module to predict both short-term and long-term QoS sequences, thereby more effectively extracting time dependencies. This second aspect of the thesis focuses solely on predicting QoS metrics per IoT application with static sensor nodes. \n \nThe third part of this thesis extends the aforementioned direction and concerns the prediction of QoS metrics in a setting where the IoT gateways are mobile robots. To this end, we implement three IoT applications in a real-world environment under eight distinctive network configurations that consider robot mobility, IoT device transmission power, and the application-specific channel frequencies used. To predict the QoS behavior of various metrics in such a dynamic environment, we propose a Federated Temporal Sparse Transformer (FeD-TST) framework. This framework enables local clients to train individual models with their unique QoS dataset for each network configuration, subsequently updating an associated global model through the amalgamation of local models. For each client, a Temporal Sparse Transformer model is trained, which comprises of several components, including the encoder and decoder modules, the multi-head sparse attention module, and the masked sparse attention module. Finally, the framework generates univariate or multivariate forecasts, and the effectiveness of which is then evaluated based on a test dataset. \n \nHence, the major contribution of this particular thesis can be summarized as providing IoT devices with traffic classification and QoS prediction mechanisms for an array of IoT applications. Our proposed methodologies take into account the varied patterns of IoT data generation, device mobility, and access networks within an IoT ecosystem, while extensive experimental evaluations indicate that our proposed approaches outperform current state-of-the-art techniques in terms of performance and efficiency. Additionally, our proposed approaches can provide tangible insights into the resource requirements and limitations encountered at the access level of the IoT network and within the edge infrastructure.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.