Hybrid Autoencoder-Driven Machine Learning Framework for Intelligent and Scalable Performance Monitoring in 5G Network

Abstract

The advent of fifth generation (5G) cellular networks marks a revolutionary leap in communication technology, promising ultra-low latency, enhanced bandwidth, and massive device connectivity. With the exponential growth of mobile applications, IoT devices, and high-throughput services such as autonomous driving and augmented reality, 5G is positioned as the backbone of future digital ecosystems. Historically, mobile networks have evolved from 1G analog voice systems through 4G LTE, each generation improving speed and efficiency but still facing critical challenges in spectrum utilization, energy efficiency, and dynamic resource management. Despite these advances, traditional cellular systems struggle to meet the stringent requirements of real-time applications in terms of reliability, security, and scalability. Conventional systems rely on static resource allocation, rule-based scheduling, and limited traffic prediction models. These methods often fail under highly dynamic environments, leading to issues such as network congestion, poor Quality of Service (QoS), and inefficient spectrum usage. Moreover, the heterogeneity of devices and applications in 5G networks makes centralized control mechanisms less effective, thereby creating a need for intelligent, adaptive, and data-driven solutions. The proposed system leverages machine learning and data-driven analytics applied to a comprehensive 5G dataset containing key performance indicators (KPIs) such as throughput, latency, jitter, signal strength, and spectrum usage. The proposed system employs a deep autoencoder to automatically learn compact and meaningful latent representations from highdimensional 5G network performance data, eliminating noise and redundant information. These deep latent features are then classified using advanced machine learning models including K-Nearest Neighbours (KNN), Categorical Boosting (Cat Boost), and Extreme Gradient Boosting (XGBoost) for baseline comparison, while a Random Forest classifier (RFC) integrated with the autoencoder forms the proposed Deep Latent-Forest (DLF) model. By combining deep feature learning with ensemblebased classification, the proposed system significantly improves prediction accuracy, robustness, and scalability in detecting dropped connection events compared to traditional manual monitoring systems, making it suitable for data-driven and real-time 5G network performance management.