Detecting Out-of-Order Control Messages in O-RAN: Dataset, Benchmarks, and Early-Warning Models

The Open Radio Access Network (O-RAN) architecture introduces openness, programmability, and multi-vendor xApp innovation through the Near-Real-Time RAN Intelligent Controller (Near-RT RIC). However, this flexibility also increases the risk of control-plane inconsistencies, particularly out-of-order control messages (OOM) that disrupt state synchronization and degrade control efficiency. Despite their impact, OOM behaviors remain insufficiently examined, and no datasets exist for benchmarking detection mechanisms. In this work, we present the first reproducible framework for generating, tracing, and analyzing OOM events in O-RAN using an instrumented FlexRIC platform. We modify the E2 Application Protocol (E2AP) stack, message handlers, and routing logic to record fine-grained sequencing metadata, enabling construction of a ground-truth dataset containing over one hundred thousand OOM events across diverse stress scenarios involving multi-xApp, multi-node, and subscription-conflict conditions. Our Key Performance Indicator (KPI)-driven evaluation reveals consistent degradation patterns such as increasing processing cost, Stream Control Transmission Protocol (SCTP) backpressure, and RIC state corruption leading to termination under heavy load. We further benchmark classical baselines against sequence-aware models for automated OOM detection and early warning. We introduce a Temporal-Aware Attention Model with Hierarchical Feature Fusion (TAAM-HFF), which achieves high recall across both rare and common OOM scenarios, underscoring the importance of temporal modeling for detecting emerging control-plane instability. The resulting dataset and benchmarking framework provide a foundation for resilient xApp operation and robust O-RAN control-plane security.