Open Radio Access Network (RAN) was designed with native Artificial Intelligence (AI) as a core pillar, enabling AI- driven xApps and rApps to dynamically optimize network performance. However, the independent ICP adjustments made by these applications can inadvertently create conflicts- direct, indirect, and implicit, which lead to network instability and KPI degradation. Traditional rule-based conflict management becomes increasingly impractical as Open RAN scales in terms of xApps, associated ICPs, and relevant KPIs, struggling to handle the complexity of multi-xApp interactions. This highlights the necessity for AI-driven solutions that can efficiently detect, classify, and mitigate conflicts in real-time. This paper proposes an AI-powered framework for conflict detection, classification, and mitigation in Open RAN. We introduce GenC, a synthetic conflict generation framework for large-scale labeled datasets with controlled parameter sharing and realistic class imbalance, enabling robust training and evaluation of AI models. Our classification pipeline leverages GNNs, Bi-LSTM, and SMOTE-enhanced GNNs, with results demonstrating SMOTE-GNN's superior robustness in handling imbalanced data. Experimental validation using both synthetic datasets (5-50 xApps) and realistic ns3-oran simulations with OpenCellID-derived Dublin topology shows that AI-based methods achieve 3.2x faster classification than rule-based approaches while maintaining near-perfect accuracy. Our framework successfully addresses Energy Saving (ES)/Mobility Robustness Optimization (MRO) conflict scenarios using realistic ns3-oran and scales efficiently to large-scale xApp environments. By embedding this workflow into Open RAN's AI-driven architecture, our solution ensures autonomous and self-optimizing conflict management, paving the way for resilient, ultra-low-latency, and energy-efficient 6G networks.

翻译：开放式无线接入网络（Open RAN）在设计之初便将原生人工智能（AI）作为核心支柱，使得AI驱动的xApp和rApp能够动态优化网络性能。然而，这些应用独立进行的ICP调整可能无意中引发直接、间接和隐式冲突，从而导致网络不稳定和KPI下降。随着Open RAN在xApp数量、相关ICP及对应KPI方面的扩展，传统的基于规则的冲突管理方法在处理多xApp交互的复杂性时日益显得不切实际。这凸显了需要AI驱动的解决方案来实时高效地检测、分类和缓解冲突。本文提出了一种用于Open RAN中冲突检测、分类与缓解的AI驱动框架。我们引入了GenC，一个用于生成大规模带标签数据集的合成冲突生成框架，该框架通过控制参数共享和模拟真实的类别不平衡，为AI模型的鲁棒训练与评估提供了支持。我们的分类流程利用了图神经网络（GNN）、双向长短期记忆网络（Bi-LSTM）以及SMOTE增强的GNN，结果表明SMOTE-GNN在处理不平衡数据时表现出更优的鲁棒性。使用合成数据集（5-50个xApp）以及基于OpenCellID都柏林拓扑的真实ns3-oran仿真进行的实验验证表明，基于AI的方法在保持近乎完美准确率的同时，其分类速度比基于规则的方法快3.2倍。我们的框架成功解决了使用真实ns3-oran模拟的节能（ES）/移动鲁棒性优化（MRO）冲突场景，并能高效扩展到大规模xApp环境。通过将此工作流嵌入到Open RAN的AI驱动架构中，我们的解决方案确保了自主且自优化的冲突管理，为构建弹性、超低延迟和节能的6G网络铺平了道路。