xApp Empowered Resource Management for Non-Terrestrial Users in 5G O-RAN Networks

Technical Summary: xApp Empowered Resource Management for Non-Terrestrial Users in 5G O-RAN Networks

Problem Statement
The integration of Unmanned Aerial Vehicles (UAVs) into existing cellular networks presents distinct challenges compared to terrestrial user equipment (UE). UAVs operate with high mobility, three-dimensional (3D) movement patterns, and elevated altitudes, leading to frequent handover events, increased signaling overhead, and potential radio link failures. Traditional handover management schemes are largely reactive, relying on instantaneous signal measurements and predefined thresholds. These approaches are inadequate for UAV scenarios where rapid mobility and predictable flight paths offer opportunities for proactive optimization. Furthermore, existing mobility management solutions have not fully leveraged the intelligent capabilities of the Open Radio Access Network (O-RAN) architecture, particularly in scenarios involving predetermined flight paths where predictive optimization is feasible. The core challenge lies in balancing the minimization of outage probability with the reduction of handover frequency to ensure seamless connectivity for mission-critical aerial operations.

Methodology
The authors propose a proactive UAV mobility management framework implemented as an xApp within the O-RAN Near Real-Time Radio Intelligent Controller (Near-RT RIC). The system employs a Double Deep Q-Network (DDQN) reinforcement learning (RL) algorithm to optimize handover decisions for UAVs operating along predetermined flight trajectories.

• System Model: The framework considers a cellular network with multiple terrestrial base stations (BSs) and a UAV following a 3D trajectory. The optimization objectives are twofold: minimizing the average outage probability (ensuring Signal-to-Interference-plus-Noise Ratio (SINR) remains above a threshold) and minimizing the frequency of handover events.
• Reinforcement Learning Framework:
  • State Space: Includes the UAV's current and next 3D positions, SINR measurements from the three most relevant base stations, and the identity of the serving and candidate cells.
  • Action Space: Discrete actions allowing the UAV to maintain the current connection or switch to one of two candidate base stations.
  • Reward Function: Designed to penalize both outage events and unnecessary handovers. A sigmoid-based smooth approximation is used for the handover penalty to balance continuous outage values with discrete handover events.
  • Algorithm Selection: The authors selected DDQN over standard DQN and policy gradient methods (PPO, A2C). Experiments showed that policy gradient methods suffered from premature convergence to "no-handover" policies due to high gradient variance and short-term penalty bias. DDQN was chosen for its ability to mitigate overestimation bias in Q-value updates, providing more stable learning in dynamic environments.
• Transfer Learning: To address the challenge of deploying models across diverse flight scenarios without extensive retraining, the authors propose a centralized weight averaging approach. Multiple DDQN models are trained independently on distinct flight paths. Their weights are then averaged to create a global model. This global model is subsequently fine-tuned on new, unseen flight paths using a reduced learning rate and layer-wise parameter freezing (freezing lower layers to preserve general patterns while adapting upper layers).
• O-RAN Integration: The xApp operates within the Near-RT RIC, utilizing the E2 interface to receive Key Performance Metrics (KPMs) like SINR from Distributed Units (DUs) and the A1 interface to receive Enrichment Information (EI), specifically UAV flight path data from an Unmanned Aircraft System Traffic Management (UTM) system. This allows the xApp to anticipate network conditions and execute control decisions with sub-second latency.

Key Contributions

1. Proactive xApp for O-RAN: The development of a DDQN-based xApp that transforms reactive handover decisions into predictive optimization by leveraging predetermined flight trajectories.
2. Transfer Learning Framework: A novel approach using centralized weight averaging to consolidate knowledge from multiple flight scenarios into a global model, enabling generalization to unseen environments with minimal retraining.
3. Comprehensive Evaluation: A thorough assessment demonstrating that the proposed framework achieves a superior trade-off between handover frequency reduction and connectivity reliability compared to traditional schemes.

Results
The framework was evaluated in a simulation environment featuring a 2000 × 2000 m² area with five base stations and ten distinct UAV flight paths. The results were compared against a Greedy scheme, a Hysteresis-based scheme, a Minimum Outage Probability (MOP) scheme, and a standard DQN.

• Handover Reduction: The proposed DDQN approach reduced handover events by 54.6% compared to the greedy baseline (2.45 events vs. 5.40 events). It also outperformed the Hysteresis scheme (2.45 vs. 4.30 events) and the MOP scheme (2.45 vs. 4.00 events).
• Outage Probability: While the greedy scheme achieved the lowest outage probability (0.021%), the DDQN approach maintained outage probability at a practically negligible level (0.048%), representing only a marginal increase of 0.027 percentage points. This is considered operationally insignificant compared to the significant reduction in signaling overhead.
• Stability: The DDQN approach demonstrated lower variance in handover events compared to baseline methods, indicating more consistent performance.
• Transfer Learning Efficacy: Cosine similarity and Euclidean distance analyses confirmed that the global model successfully generalized handover decision patterns across different flight paths, with similarity scores improving after fine-tuning.

Significance and Claims
The paper claims that the proposed framework validates the effectiveness of intelligent, learning-based approaches for UAV mobility management in next-generation O-RAN architectures. By anticipating network conditions through predetermined flight paths, the system minimizes both outage probability and handover frequency, addressing the limitations of reactive mechanisms. The successful implementation of transfer learning suggests that robust global models can be deployed across diverse operational environments without the computational burden of retraining for every new scenario. The authors assert that this work contributes to the seamless integration of aerial user equipment into cellular networks by providing a scalable, proactive solution that balances resource efficiency with connectivity reliability. The study acknowledges that while the approach shows effective generalization within tested conditions, real-world validation on O-RAN testbeds and exploration of multi-UAV scenarios remain future objectives.