Residual State Space Learning for Early Detect and Avoid Hazard Prediction in Maneuvering UAS Encounters

• Hassan Emad

  Department of Information Systems, Sinai Higher Institute for Engineering and Technology, El-Arish Road, Arish, North Sinai Governorate, Egypt

  Author

• Mostafa Adel

  Department of Electrical Engineering, Modern Academy for Engineering and Technology, Industrial Area, Maadi, Cairo 11728, Egypt

  Author

• Youssef Samir

  Department of Computer Engineering, Thebes Higher Institute of Engineering, Thebes Academy, South El-Academia District, New Cairo, Cairo Governorate, Egypt

  Author

Detect and Avoid prediction for unmanned aircraft requires timely recognition of hazardous encounters under noisy sensing, maneuvering intruders, and incomplete knowledge of future intent. Conventional extrapolation can become unstable when the intruder turns, accelerates, climbs, or descends, while unconstrained machine learning can produce accurate rankings without preserving physically interpretable encounter geometry. This paper develops and evaluates an empirical residual state space learning method for early hazard prediction in non-cooperative unmanned aircraft encounters. The method combines a stochastic kinematic predictor with a compact recurrent residual learner trained on simulated encounter histories, continuous closest-approach quantities, and binary loss-of-well-clear labels. A dataset of 240000 maneuvering encounters is generated with strict separation between training, validation, and test seeds. The proposed model is compared against constant-velocity extrapolation, coordinated-turn prediction, interacting multiple-model filtering, gradient-boosted hazard prediction, and an unconstrained recurrent neural baseline. In the held-out test set, the residual state space model reduces horizontal miss-distance root mean square error by 31.7\% relative to the best filtering baseline and by 22.8\% relative to the direct recurrent baseline. The model improves loss-of-well-clear area under the receiver operating characteristic curve from 0.941 to 0.973 relative to the interacting multiple-model predictor and increases 35-second warning recall at an approximately fixed 5\% false-alert rate from 81.4\% to 89.6\%. Bootstrap intervals, McNemar tests, and encounter-level loss comparisons indicate that these gains are statistically stable across maneuver intensity, sensor noise, and prediction horizon strata.

• PDF

Copyright (c) 2026 International Journal of Scientific and Applied Computing Research

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.