DUAL-TRAP: A Dual-Branch Temporal Convolutional Network for Planner-Based Anomaly Detection in Heterogeneous Unmanned Vehicles

초록

This article presents dual-branch temporal convolutional network for planner-based anomaly detection (DUAL-TRAP), a dual-branch temporal convolutional network (TCN) for planner residual anomaly detection in heterogeneous unmanned vehicles. The proposed method formulates anomaly detection on residual sequences defined by the discrepancy between planned trajectory references and executed vehicle motion. To characterize nominal residual behavior, Dual-TRAP jointly reconstructs planner residual patterns and predicts their short-term evolution, enabling abnormal behavior to be identified. The model is designed for onboard execution, while residual-based diagnostic metrics are transmitted to the ground station interface for monitoring and analysis. Experimental validation on real-world unmanned aerial vehicle (UAV) and unmanned ground vehicle (UGV) platforms shows that Dual-TRAP achieves anomaly indication, with response times ranging from 0.5 to 7.6 s. The results further show that Dual-TRAP achieves 8.8 & times; lower inference latency and 3.3 & times; fewer parameters, while maintaining false alarm ratios below 6 %. These results indicate that planner residual modeling provides a suitable strategy for onboard condition monitoring in heterogeneous unmanned vehicles.

키워드