Non-Terrestrial Networks (NTNs) are expected to become an essential component of future sixth-generation (6G) mobile networks by extending coverage, improving service continuity, and supporting resilient connectivity beyond the reach of terrestrial infrastructures. At the same time, Open Radio Access Network (O-RAN) architectures provide functional disaggregation, open interfaces, and multi-time-scale control mechanisms that enable flexible and programmable RAN operation. Extending O-RAN principles to NTN environments, however, is not a direct transposition of terrestrial deployments. Space–ground propagation delay, intermittent feeder-link availability, limited onboard processing and energy resources, and the overhead of telemetry, control signaling, and model updates strongly constrain where RAN functions and Artificial Intelligence/Machine Learning (AI/ML) components can be placed. This paper analyzes intelligence deployment in NTN-based O-RAN architectures from a system-level and network-design perspective. We organize the design space by identifying representative deployment options for Radio Unit (RU), Distributed Unit (DU), and Central Unit (CU) functions across terrestrial and non-terrestrial infrastructures. We then examine how O-RAN control entities and AI/ML workflows can be mapped onto these deployments, distinguishing between long-term orchestration, near-real-time control, and local inference. The discussion explicitly connects each deployment choice to practical feasibility constraints, including functional split requirements, RU–DU co-location, E2/O1 telemetry overhead, model and policy distribution traffic, onboard compute limitations, and intermittent connectivity. To support the architectural discussion, we provide an architecture-level control-loop feasibility analysis showing how NTN transport and processing constraints restrict the applicability of centralized, distributed, and hybrid intelligence placement strategies. The resulting analysis provides design guidelines for selecting suitable O-RAN and ML deployment configurations according to service requirements, control-loop time scales, and NTN infrastructure capabilities.
Optimizing the performance of a complex system, such as a 5G radio cell, is a challenging, costly, time-consuming, and subjective task. Current data-driven methods for predicting 5G cell performance do not effectively leverage domain expert knowledge and fail to accurately predict the effects of fine-tuned cell configurations, rendering them unreliable for guiding reconfiguration decisions. We present a methodology for representing 5G cell configurations and domain-expert knowledge as a graph, and propose two model architectures designed for cell-level performance prediction. The models are evaluated using benchmark 5G data and a state-of-the-art predictor. Our approach achieves consistently lower prediction errors than the reference models in sample-wise and cell configuration state evaluations, with reductions of 12% and 33% in MAE, respectively, while requiring 40% fewer trainable parameters. Reduced prediction errors enable the identification of configuration changes that most significantly impact the predicted Key Performance Indicators, providing a solid foundation for optimizing the performance of 5G radio cells.
Cold start latency in Function-as-a-Service (FaaS) platforms remains a critical barrier to predictable performance, especially in geographically distributed edge environments where resource constraints and network heterogeneity exacerbate the problem. This paper introduces a two-phase analytical framework that integrates machine learning with geography-aware simulation to mitigate cold-start latency. The framework does not propose a fundamentally novel learning algorithm. Rather, its primary contribution lies in the systematic integration of existing methods — specifically XGBoost, LightGBM, and rule-based routing — into a design that prioritizes deployability in resource-constrained edge environments. In Phase 1, we conduct an empirical analysis of the Alibaba cluster trace (200 K invocations, 180 unique functions), evaluating four ML architectures: LSTM, XGBoost, LightGBM, and proposed Context-Aware Hybrid Ensemble (CAHE) model, which dynamically selects between XGBoost and LightGBM based on real-time pattern recognition. In Phase 2, we develop and validate a geographic simulation framework that generates 800 K invocation across a realistic 5-region edge network with 12 servers, incorporating OpenWhisk function profiles and workload patterns. CAHE model achieves 29.6% cold start reduction on real-world data and 33.2% in simulated environments, demonstrating statistically significant improvements over individual models. Simulation reveals strong geographic dependencies; notably, while baseline sequence models show reduced accuracy under the high variance of cross-regional traffic, augmenting LSTM with spatial features yields a relative improvement (up to 171%), alongside clear regional variations in cold start rates. This work provides both a validated ML solution for cold start mitigation and a reproducible simulation platform, offering a practical approach for predictive scaling in distributed FaaS environments.
In 5G and beyond networks, establishing radio communication between User Equipment (UE) and base stations (gNodeB, gNB) relies on the Radio Resource Control (RRC) layer, which is vulnerable to signaling storms: high volumes of connection requests can overwhelm the control plane and prevent new UEs from establishing connections. In this work, we present a comparative evaluation of advanced detection approaches for RRC signaling storms. We consider an Extreme Value Theory (EVT)-based adaptive thresholding system alongside Machine Learning (ML) approaches, including probabilistic forecasting models and supervised classifiers. Using real-world RRC traffic from an operator network augmented with realistic threat scenarios, we assess detection accuracy, precision, recall, and latency for each approach. Our results show that EVT-based adaptive thresholding delivers robust and low-overhead detection that is comparable to more complex ML models, achieving metrics above 97% in the single-cell scenarios, and that it generalizes across sites and traffic patterns without retraining. Tree-based classifiers reach the highest detection scores, but at the cost of requiring reliable labeled training data and model transfer. These findings offer practical guidance for selecting effective detection methods under diverse traffic and threat conditions in operational 5G/6G deployments.
UAV-assisted urban traffic monitoring provides flexible viewpoints, rapid deployment, and adaptive coverage beyond what fixed roadside cameras can offer. Its practical utility depends not only on onboard result generation, but also on timely and reliable delivery of the generated outputs over a dynamic multi-hop aerial network. This paper therefore studies end-to-end result acquisition at the network layer, rather than visual detection accuracy. Instead of transmitting raw video streams, provider UAVs continuously generate compact time-indexed result objects, and consumers issue requests for objects derived from the ongoing visual process. We propose a reception-determined result-delivery framework in which eligible receivers coordinate forwarding according to current packet receptions and historical forwarding relevance. The framework maintains per-request reverse-path state and consumer–provider forwarding-priority state, applies a time-index-aware update rule so that stale observations cannot overwrite newer forwarding knowledge, and uses localized repair when a recorded return hop becomes unavailable. Under the evaluated OMNeT++ settings, aggregate simulation results indicate a better reliability–timeliness balance than AODV, NDNF, and IHCR: the framework attains higher result-acquisition success and effective-result return rate, lower completion latency than AODV and IHCR, completion latency comparable to NDNF, and lower post-generation delivery delay.
