Abstract
The introduction of commercial 5G networks, coupled with the anticipated beyond
5G networks that were once merely theoretical and laboratory-tested, has
introduced numerous new possibilities. These developments have enabled applications
that require substantial data usage, high speed, low latency, and
massive connectivity, with the added benefit of use case customization. The
projected need for enhanced mobile broadband (eMBB), for example, is increasing
optimism about adopting 5G. As the number of connected users and
devices exponentially augments in this era of the multiverse and the Internet
of Things (IoT), the need for intelligent functions in 5G increases. Such intelligence
is crucial to continuously monitor network behavior and automate specific
actions to preserve service levels above a predetermined standard for quality of
experience (QoE) optimization. In addition, the change in the radio access
network (RAN), the modification of the architecture, and the introduction of
advanced technologies such as network slicing, network functions virtualization
(NFV), and software-defined networking (SDN) have introduced new challenges.
These changes bring complexity in terms of monitoring the quality of experience
(QoE) and service (QoS) on user, network, and application levels. Hence,
the integration of machine learning (ML), as a subset of artificial intelligence
(AI), is critically needed to advance network automation and to optimize the
quality of experience (QoE) for end-users. This research explores the conceptual
design, development, and optimization of ML-based intelligent analytics to effectively
monitor users’ 5G experience and network cells’ behavior toward eMBB
applications. The approach transcends conventional network management and
operation paradigms, centering on the adaptability of service behavior monitoring
and performance evaluation, real-time QoE awareness, and network cells’
segmentation.
It employs a holistic methodology that integrates theoretical modeling and practical
experimentation using live 5G network data as opposed to most simulated
lab results’ research works. A state-of-the-art, cost-effective analytical framework,
hybrid in nature is presented in this work. The framework intersects
legacy techniques and ML to perform service awareness, compute the global
user experience using the concept of user quality index (UQI), and optimally
segment 5G symmetric cells based on performance behavior patterns in a standalone
(SA) network. The approach explores advanced analysis in the user
domain and network domain. In the first context, we begin with the detection,
4
classification, and analysis of eMBB service applications from raw user traffic
through flow-level packet inspection and machine learning. The latter is used to
classify network traffic as eMBB or non-eMBB. A conceptually tested statistical
method is then employed to calculate the UQI. An advanced clustering methodology
employing a heterogeneous ensemble technique is developed in the network
domain to organize 5G cells based on their behavior, characterized by defined
key performance indicators (KPIs). The method involves a coordinated application
of k-means and hierarchical clustering algorithms, enabling the dynamic
segmentation of cells. An XGBoost classifier is then trained on the clustered cell
data to facilitate the identification of cells with performance deficiencies. Tested
on 5G raw user traffic and cell data, our findings reveal significant improvements
in users’ eMBB QoE insights, assessment, and network self-awareness, providing
Mobile Network Operators (MNOs) with an analytical ML-based 5G QoE and
cell behavior tracking framework for better traffic policies adjustment, resource
allocation management, and coverage optimization. Our research demonstrates
notable advancements in the detection of eMBB traffic, measuring user quality,
and improving service-awareness. The findings present substantial qualitative
improvements in user and network performance evaluation.
In addition, this experiment presents an ML-based imputation technique that
addresses the issue of missing RAN metrics because preserving the overall integrity
of RAN-related information is critical to our study. The approach involves
applying more dynamic and adaptive correlation analysis in an effort
to enhance the utilization of the feature extraction, maintaining variability between
metrics. The output of this approach has shown superior performance
when compared to statistical techniques such as the mean, mode, and median.
Several algorithms are trained, and the best-performing model is automatically
selected for the imputation task in the analytical pipeline.
Our research adds to the critical role of ML in shaping and advancing the future
communication generations (5G specifically and to a certain extent beyond 5G).
We present both conceptual and practical perspectives on how these technologies
can be leveraged to enhance the operational performance and efficiency of
intelligent RANs.