Contribution to the development of Wi-Fi networks through machine learning based prediction and classification techniques
(English) The growing number of Wi-Fi users and the emergence of bandwidth-intensive services have necessitated an increase in Access Point (AP) density, resulting in more complex network configuration, optimization, and management tasks. Concurrently, advancements in data monitoring and analytics t...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis doctoral |
| Fecha de publicación: | 2024 |
| País: | España |
| Institución: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/452743 |
| Acceso en línea: | https://hdl.handle.net/2117/452743 https://dx.doi.org/10.5821/dissertation-2117-452743 |
| Access Level: | acceso abierto |
| Palabra clave: | 621.3 - Enginyeria elèctrica. Electrotècnia. Telecomunicacions Àrees temàtiques de la UPC::Enginyeria de la telecomunicació |
| Sumario: | (English) The growing number of Wi-Fi users and the emergence of bandwidth-intensive services have necessitated an increase in Access Point (AP) density, resulting in more complex network configuration, optimization, and management tasks. Concurrently, advancements in data monitoring and analytics technologies in wireless networks offer opportunities to extract valuable insights into network and user behavior, facilitating more efficient network management. In this thesis, we propose different Machine Learning based techniques to enhance Wi-Fi network management, focusing on three aspects: user connectivity prediction, Wi-Fi traffic prediction, and Wi-Fi traffic classification. The first aspect of our work focuses on predicting the next Access Point (AP) a user will connect to in a Wi-Fi network. We propose a methodology based on historical information of the AP to which a user has been connected, extracting connectivity patterns at different time scales (hourly, daily, weekly). Predictions are done using techniques based on Neural Networks and Random Forest algorithms. This approach is evaluated using real data from a university campus Wi-Fi network. Predicting the next AP of users allows for proactive network reconfiguration, enhancing the efficiency of techniques like Pairwise Master Key caching and Opportunistic Key Caching, which reduce re-authentication times. Additionally, this prediction helps to identify the geographical region of the User Equipment (UE) and can be used for commercial purposes, such as targeted advertising, by customizing messages based on locations of the users. Secondly, we propose a methodology for predicting the aggregated traffic at access points (APs) by leveraging spatial and temporal correlations from neighboring APs to enhance prediction accuracy. Using real measurements, we evaluate various Deep Learning methods, including Convolutional Neural Network (CNN), Simple Recurrent Neural Network (SRNN), Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM), and Transformer, and present a hybrid approach combining CNN for spatial processing and RNN for temporal prediction. This hybrid method improves accuracy with minimal additional training time and negligible impact on prediction time. Accurate traffic forecasting at each AP enables better load distribution and can inform resource management techniques such as admission control, congestion control, and load balancing. Additionally, predicting low traffic periods can aid in energy-saving strategies by allowing APs with minimal traffic to be switched off during specific times. Finally, traffic classification is essential for enhancing network performance by allowing better resource allocation and prioritization of services with stringent latency requirements. The increasing demand for Virtual Reality (VR) services poses a significant challenge for Wi-Fi networks to meet strict latency needs, crucial for VR to ensure immediate response and avoid user discomfort. To improve VR Quality of Service (QoS), distinguishing interactive VR traffic from Non-VR traffic is key. We propose a machine learning-based method to identify interactive VR traffic in a Cloud-Edge VR environment by analyzing downlink and uplink data correlations and extracting features from single-user traffic characteristics. Six classification techniques (i.e., Logistic Regression, Support Vector Machines, k-Nearest Neighbors, Decision Trees, Random Forest, and Naive Bayes) are compared. The result of the classification is used for the prioritization of VR traffic over Non VR traffic. We evaluate our method using datasets from various VR applications and Wi-Fi network simulations. Our results show a significant reduction in VR traffic delays with minimal impact on Non-VR service latency. |
|---|