Skip to Content
AerospaceAerospace
  • Article
  • Open Access

21 November 2026

21 Pages

Generalized GERT Network Modeling with Sensitivity-Guided Adaptive Planning for UAV Mission Reliability

and
1
Computer Engineering Center, National Technical University “Kharkiv Polytechnic Institute”, 61002 Kharkiv, Ukraine
2
Department of Computer Engineering and Cybersecurity, University of the National Education Commission, 30-084 Krakow, Poland
*
Authors to whom correspondence should be addressed.
Aerospace2026, 13(12), 1091;https://doi.org/10.3390/aerospace13121091 
(registering DOI)

Abstract

Graphical Evaluation and Review Technique (GERT) networks model unmanned aerial vehicle (UAV) missions whose flight segments may be retried, aborted, or restarted, yielding closed-form distributions of mission completion time. Even the treatment of a single mandatory detour—whether the reconnaissance loop is flown at least once (forced excursion) or may be cleared on the first pass (first passage)—is a modeling choice between physically different mission processes, and it alone shifts the window probability by 0.1016 in a reference network, motivating exact first-passage modeling. We (i) recast the model under exact first-passage semantics and prove that the equivalent W-function follows from a linear signal-flow system, verified symbolically against the Mason gain formula; (ii) give a two-level well-posedness analysis—local sub-stochasticity (residual mass read as abort) and global transience, verified numerically (ρ(P)<1 by eigenvalue routines)—under which the first-passage transform is a defective (sub-probability) measure on mission time; (iii) exploit the millisecond analytic evaluation of window probabilities on small networks (n≤ 20) as an in-the-loop objective replacing Monte Carlo inner loops; and (iv) demonstrate in a horizon-one adaptive loop that the analytic evaluator is a fast, deterministic, noise-free reward for single-branch rate corrections: the per-state optimum is available by enumerating 24 analytic evaluations, and a tabular Q-learning agent—a proof of concept of the in-the-loop objective rather than a methodological contribution—recovers the per-state oracle exactly (the best state-agnostic policy captures only 64% of the adaptation gain). Directional-wind and endurance-coupled physics extensions refine the scalar-wind baseline, and case studies on five-node and six-node topologies demonstrate topology-agnostic modeling and quantify the reliability benefit of a backup leg.

Article Metrics

Citations

Article Access Statistics

Article metric data becomes available approximately 24 hours after publication online.