FALCON++: A Semi-Supervised Anomaly Detection and Localization Framework for Resilient O-RAN Deployments

Fault detection is a critical problem in Open Radio Access Network (O-RAN) observability, but the availability of labeled fault data is extremely limited, and enumerating all possible fault scenarios is costly and often infeasible. Therefore, it is essential to rely on normal network behavior to enable effective fault detection. In this work, we propose FALCON++, a tailored semi-supervised anomaly detection and localization framework. Specifically, we use a specialized contextual autoencoder that fuses feature-wise gating and hierarchical temporal encoding for the effective reconstruction of normal network behavior. Furthermore, we combine a causal decoder with residual skip connections to improve robustness to data deemed anomalous due to their higher reconstruction error. FALCON++ also addresses scalability issues by shifting from a monolithic system model to CU-DU pairwise inferencing. Through experimental results on synthetic faults generated using STRESS-NG and TC, we show that FALCON++ achieves a high-precision and high- recall operating point with an F1-score of 0.77, thereby enabling anomaly detection and localization.