Real-time network anomaly detection in O-RAN using deep learning on streaming big data

The growing adoption of Open Radio Access Network (O-RAN) architecture in next-generation wireless systems offers unprecedented flexibility and openness, enabling real-time control and intelligent network functions. However, the decentralized and multi-vendor nature of O-RAN also increases the surface for network anomalies, including service degradations, misconfigurations, cyber threats, and radio interference. Traditional monitoring tools lack the scalability and responsiveness to process heterogeneous, high-velocity telemetry from RAN Intelligent Controllers (RICs), Distributed Units (DUs), and User Equipment (UE). This research proposes a novel real-time anomaly detection framework that integrates streaming big data pipelines with deep learning-based temporal models to detect and localize anomalies within the O-RAN ecosystem. Leveraging platforms like Apache Kafka and Apache Flink for high-throughput data ingestion and processing, the system applies Long Short-Term Memory (LSTM) and Temporal Convolutional Networks (TCNs) to learn temporal behavior from real-time Key Performance Indicators (KPIs), signal logs, and E2 interface events. The model is trained using both real and synthetic datasets generated from O-RAN emulators and validated in an emulated 5G environment with open-source xApps. Additionally, the framework incorporates an adaptive feedback mechanism between the Near-Real-Time RIC and Non-Real-Time RIC via A1 and E2 interfaces, enabling continuous learning and dynamic policy adaptation. The proposed solution demonstrates over 95% anomaly detection accuracy with sub-second latency, outperforming traditional statistical baselines. This work provides a scalable, vendor-agnostic, and future-proof anomaly detection system for intelligent RAN management, marking a significant step toward self-healing and self-optimizing networks in 5G and beyond.