A Machine Learning Long-endurance Predictive Association Algorithm Based on Target Feature Attribute Constraints

This paper proposes a machine learning-based long-endurance predictive association algorithm that incorporates target feature attribute constraints. By integrating target characteristic attributes (e.g., motion features, Radar Cross-Section (RCS) characteristics, and A/C codes) with Long Short-Term Memory (LSTM) deep neural networks, the algorithm achieves efficient predictive association using time-series air target trajectory data. First, we design a novel target feature attribute representation method to capture critical information from multi-source radar measurements. Subsequently, an adaptive correlation model is constructed, where maneuverability and probability constraints derived from target classifications guide the LSTM-based prediction process. Experimental results on both synthetic and real-world datasets demonstrate significant improvements: the overall handover success rate reaches 94.98% (synthetic data) and 89.66% (real-world data under $T=100 \mathrm{~s}, \sigma=2 \mathrm{~km}$), outperforming traditional Track Segment Association (TSA) methods by 2.23% and 3.58%, respectively. This work provides a systematic framework for long-endurance prediction in complex aerial surveillance scenarios. (Abstract)