Engineering

Abstract: A computational launch vehicle guidance, navigation, and control simulator is evaluated through adispersed campaign of 1,000 runs. Thee model combines six-degree-of-freedom dynamics, inertial navigation,simulated GPS and barometric aiding, and thrust-vector attiitude control. A 0.01 s integration step and a 600s maximum duration were used. All runs produced numerical impacts, but 100 remained below 100 mapogee. Median apogee was 84.98 km and mean apogee was 76.28 km, revealing a distinct low-altitudepopulation. A standardized linear surrogate identifieed thrust misalignment as the dominant predictor ofimpact displacement within the confiegured dispersion ranges. Mean per-run normalized estimation errorsquared was 12.21 for the six-state position and velocity subspace; normalized GPS and barometerinnovation statistics averaged 0.73 and 0.97. Epoch-wise diagnostics revealed systematic departuresobscured by campaign averages. Theese results demonstrate how trajectory summaries, covariance checks,and sensitivity screening identify diffeerent limitations of a coupled GNC implementation. Atmosphericsimplifiecations, incomplete variable-mass rotational modeling, and absent independent trajectory validationrestrict physical interpretation. Thee fiendings provide specifiec targets for model verifiecation, estimatorcalibration, and investigation of early impacts.

Abstract: Small fixed-wing unmanned aerial vehicles are almost always driven by exposed propellers, whose slipstream washes the airframe. This paper discloses an alternative in which a multi-bladed micro electric ducted fan is embedded in the centreline body of a delta wing, with a row of guide vanes behind the rotor and the exhaust discharged behind the airframe, so that no rotating component is exposed to the external flow. Two variants with 13-bladed, high-solidity rotors of 68 mm and 46 mm tip diameter, driven at 30,000 revolutions per minute by a 6300 KV motor, are defined and assessed by computational fluid dynamics against the same airframe fitted with an APC 10x7 puller propeller, after independent validation of the airframe and propeller models. At 10 degrees angle of attack both embedded variants produce a net propulsive force at their design speeds of 30 m/s and 10 m/s, and velocity vectors confirm that the guide vanes remove the rotor swirl. The 46 mm variant retains the lift coefficient of the clean wing, 0.574 against 0.570, while the larger 68 mm duct reduces it to 0.427 but raises the lift-to-drag ratio from 9.8 with the puller propeller to 10.4. A 3D-printed prototype of the engine was built and run, and its rotor was tested on a thrust stand.

Abstract: The increase in the reliability and safety levels of electrical power devices plays a key role in the transition toward aviation electrification. Safety-critical equipment, such as propulsion motors, must comply with stringent requirements to obtain airworthiness certification. Permanent Magnet Synchronous Motors (PMSMs) have emerged as a suitable technological solution for high-performance flight control actuation systems, and their extension to propulsive functions appears promising. Nevertheless, the reliability of PMSM power electronics converters is still far from the levels required by certification; therefore, motors must incorporate fault-tolerant capabilities in order to be considered flightworthy. In this context, the development of prompt and effective Fault Detection and Isolation (FDI) algorithms is essential. This paper presents the Model-in-the-Loop (MIL) validation of a Neural Network (NN) algorithm for the detection and isolation of MOSFET faults and open-phase faults in PMSM power drives. The approach is based on a feedforward NN trained on simulation-generated data for a multi-class classification problem, combined with a counter-based decision logic to ensure fast, robust, and computationally efficient operation. A feature selection analysis is also carried out to reduce the network size without significant degradation in the classification accuracy. For the MIL validation, the NN-based algorithm is integrated within a high-fidelity dynamic model of an electrically driven UAV propeller, including the simulation of the three-phase PMSM with Field-Oriented Control (FOC) of the motor currents, digital signal processing, together with a model of the aerodynamic loads derived from Blade Element Momentum Theory (BEMT). Faults are injected at different operating speeds and under uncertain environmental conditions, to evaluate the FDI performance in terms of detection latency, isolation capability, and robustness. The results demonstrate the effectiveness of the approach, which, owing to its straightforward generalization, can be extended to other fault categories and leveraged to support the development of PMSMs with fault-tolerant control capabilities.

Abstract: Green hypergolic hybrid rocket propellants are being investigated as a potential solution to the ignition challenges of hybrid rockets and as safer alternatives to toxic hypergolic systems. This review summarizes recent progress in hypergolic hybrid rocket propellants, focusing on ignition behavior, thermochemical performance and motor-scale combustion and propulsion characteristics. Drop-test studies are reviewed to identify promising propellant combinations and to examine the effects of a wide range of chemical, material, drop-dynamic and environmental parameters on ignition characteristics. Thermochemical calculations are used to assess theoretical propulsion performance as a function of additive type and additive loading in a suitable fuel binder. Finally, full-motor firing studies are reviewed to evaluate stable ignition, sustained combustion, regression-rate behavior and combustion efficiency. The literature shows that several green hypergolic propellant families can achieve rapid ignition delays in drop tests, and motor-scale propulsion performance was broadly comparable to conventional hybrid rocket motors. However, the technology is still immature as gaps in the literature remain related to ignition repeatability, safe operating envelopes, pressurization delay times and mechanical integrity of additive-loaded fuel grains. Future work should therefore focus on resolving these gaps to enable wider application of this otherwise promising propulsion technology.

Abstract: Fiber-reinforced polymer composites are attractive for load-bearing aeronautical structures, but their use near fire zones is limited by matrix pyrolysis and the loss of through-thickness integrity. Validating thermo-structural models requires temperature data from inside the laminate rather than at its boundaries. This paper reports in situ through-thickness monitoring of cross-ply carbon/epoxy (AS4/8552) laminates under one-sided flame exposure at approximately 1000 °C. Fiber Bragg Grating sensors were embedded during lay-up at 10, 50, 90 and 100 % of the thickness; temperature gratings were sheathed in 0.5 mm brass capillaries to decouple them from matrix strain, with bare gratings alongside to recover the mechanical component. Two configurations were tested: a 50-ply simply supported plate and a 30-ply plate clamped on four sides. Results are compared with a three-dimensional transient COMSOL model in which transverse conductivity and moduli switch at 673 K to represent degradation. The model reproduces the ordering, magnitude and strongly non-uniform spacing of the measured plateaus (520, 375, 290 and 290 °C), though it approaches steady state more slowly than the experiment. The simply supported plate degraded to roughly 70 % of its thickness without delamination, whereas restrained expansion in the clamped plate produced interlaminar cracking through the full thickness.

Abstract: Computational Fluid Dynamics (CFD) simulations based on Reynolds-Averaged Navier-Stokes (RANS) models are widely used for spacecraft cabin ventilation design; however, they primarily provide mean-flow quantities and often underestimate velocity extremes relevant to crew comfort, ventilation effectiveness, and stagnation-zone assessment. This study presents a simple empirical post-processing framework for estimating extreme air velocities from CFD mean-flow statistics. The proposed approach expresses the upper velocity envelope as a function of the local mean velocity and standard deviation through a single empirical coefficient, Kc. The method is calibrated using published velocity-field data from International Space Station (ISS) Crew Quarters studies and subsequently examined against independent ISS ventilation datasets, including the Columbus module and Crew Alternative Sleeping Area (CASA) configurations. Analysis of digitized experimental and CFD-derived velocity distributions yields a robust median coefficient of approximately Kc = 4.5, while external datasets indicate p99-equivalent values generally within the range Kc ≈ 4–6. Independent validation evidence from Columbus ventilation studies demonstrates that accounting for velocity fluctuations substantially improves agreement with experimental observations, reducing distribution mismatch by approximately 83% compared with uncorrected mean-velocity predictions. The results suggest that a simple statistical correction applied to standard CFD outputs can provide practical estimates of extreme ventilation velocities without the computational cost of Large Eddy Simulation (LES). While the proposed framework should be regarded as a conservative upper-envelope estimator rather than a universal turbulence closure, the consistency observed across multiple ISS ventilation datasets indicates that it offers a useful engineering tool for preliminary spacecraft cabin ventilation assessment and design.

Abstract: Magnetic measurements acquired by UAVs are affected by onboard propulsion, power, control, and electronic systems. This study evaluates sensor placement effects on UAV-induced magnetic disturbance using three RM3100 magnetometers installed at two wing locations and one tail location on a puller fixed-wing UAV. UAV OFF/ON, throttle, servo, heading, sensor-position crossover, reference, and flight-validation tests were performed using common calibration, filtering, and statistical procedures. Throttle produced the largest disturbances, reaching 430, 485, and 185 nT for RM3100-A, RM3100-B, and RM3100-C, respectively. RM3100-C generally showed the lowest disturbance, and crossover results indicated that installation position strongly influenced the observed response. These results show that calibration alone is insufficient and that sensor placement, wiring, servo proximity, electrical loading, and installation geometry should be considered jointly.

Abstract: Quadrotors attract broad military and civilian interest, yet their underactuated, highly coupled, and nonlinear dynamics make precise control difficult. While Sliding Mode Control (SMC) offers robust trajectory tracking, tuning its parameters—such as sliding surfaces and switching gains—remains a tedious manual task. To overcome this, we present an adaptive SMC framework integrated with an Actor-Critic Deep Reinforcement Learning (DRL) algorithm for automated gain synthesis. A full 6-DoF quadrotor model is derived via Euler-Lagrange mechanics, with aerodynamic coefficients estimated using Blade Element Momentum theory. The DRL agent dynamically adjusts SMC gains online, reducing chattering while ensuring accurate altitude and trajectory tracking. Lyapunov's direct method rigorously establishes asymptotic stability of the closed-loop system. Comparative simulations show that the proposed DRL-SMC approach consistently outperforms conventional fixed-gain baselines, achieving lower Root Mean Squared Error across complex flight profiles.

Abstract: The increasing complexity of general aviation flight training poses significant challenges to tower controllers, who must reconcile safety separation standards with training efficiency under diverse and often unexpected conflict scenarios. This study proposes a BERT-BiLSTM-XGBoost multimodal feature fusion ensemble learning model that automatically matches optimal allocation schemes based on predicted conflict characteristics, thereby providing real-time decision support for tower controllers. The framework operates in three stages. First, a BERT model extracts high-level semantic features from textual data—including aircraft status descriptions, controller characteristics, pilot profiles, and airspace constraints—by leveraging bidirectional contextual encoding. Second, a Bidirectional Long Short-Term Memory (BiLSTM) network captures temporal dynamic features from sequential aircraft operation data. Third, the extracted semantic and temporal feature vectors are fused and fed into an XGBoost classifier, which performs the final allocation scheme matching. The XGBoost algorithm additionally provides feature importance scores, supporting the traceability requirements of tower control decision-making. Experimental results demonstrate that the proposed model achieves a matching accuracy of 97.2%, substantially outperforming baseline methods including Naive Bayes, linear SVM, and LightGBM. Moreover, the model reduces controller decision-making time by 75–80% in simulated tower control tests, validating its practical utility as an effective decision-support tool for general aviation tower operations.

Abstract: This paper presents a design methodology for the automatic control system of an airship, accounting for atmospheric disturbances and parametric modeling uncertainties. The derived design model is a linear stochastic system with multiplicative noise incorporating both the airship’s uncertain dynamics and the wind model. First, an H_∞ state feedback control law is derived for the stochastic system to ensure robust stability and trajectory tracking performance. Subsequently, a robust Kalman filter is designed to estimate the turbulence model states based on available measurements. It is demonstrated that the optimal gain of this robust filter depends on the solution to a coupled system of specific Riccati and Lyapunov equations. Numerical results highlight a significant improvement in robustness and tracking performance when the control law incorporates wind gust velocity estimation.

Abstract: In the process of finite element fatigue analysis, the undamaged fatigue life will be used. The undamaged fatigue life can be obtained by fatigue tests of standard specimens, in which the stress level is assumed to be constant. In fact, stress level gradually increases during fatigue process, as the sectional area of specimen decreases with increase of fatigue cycle, but the loading remains constant. As a result, the undamaged fatigue life obtained by tests is usually less than that corresponding to the certain stress level, so that finite element fatigue analysis using such undamaged fatigue life will largely underestimate fatigue life. Therefore, it is desired to construct a real fatigue life model (RFLM) through considering change of stress level caused by decrease of material properties. To this end, a novel RFLM has been constructed for the first time by carefully studying the relationship between the RFLM and the existing fatigue life model. Then, the fatigue progressive damage model (FPDM) based on RFLM has been implemented into finite element model. By fatigue analysis within one element, it is found that the proposed RFLM is accurate and effective. To further evaluate performance of the RFLM, fatigue failure behaviors of the flexible composite structures (FCSs) has been analyzed by using the proposed model. To verify the accuracy of the RFLM, the fatigue experiments have been conducted to explore the failure modes and stiffness degradation process. Numerical results show that the fatigue stiffness degradation and failure modes obtained by the proposed model are in good agreement with experiments. It shows that the RFLM has ability to more accurately predict real fatigue performance of FCSs than the traditional FLM, even in wide range of stress levels under different stress states.

Abstract: The vertical wind profile within the Atmospheric Boundary Layer (ABL) is a critical variable 2 for a wide range of applications, including wind energy, air traffic and unmanned aerial 3 system (UAS) management, pollutant dispersion, and urban air mobility. However, its 4 accurate characterization typically requires expensive vertical-profiling LiDAR instruments, 5 which limits the number of locations that can be simultaneously monitored. Building upon 6 previous work in which machine-learning algorithms were trained on LiDAR data from the 7 University of León (Spain) to reconstruct full ABL wind profiles from a single near-surface 8 measurement, this study replaces the traditional data-driven neural network approach with 9 Physics-Informed Neural Networks (PINNs), embedding a physical constraint directly into 10 the loss function as evaluated by automatic differentiation, and tests two such constraints 11 under architectural control: the Ekman boundary-layer momentum balance, closed with a 12 Monin-Obukhov eddy viscosity, and a shear-constrained formulation based on a power-law 13 profile with an inferred exponent. With dense supervision at the target heights, neither 14 constraint improves on an architecture-matched network without a physical term (RMSE 15 1.35-1.36 m/s, direction MAE 20.1-20.4◦ across all three, Kruskal-Wallis p = 0.0125), and 16 in a vertical-extrapolation protocol, training with gates up to 111 m, selecting models 17 by their skill at 140 m, and testing blind on 170-300 m, the Ekman constraint shows 18 no advantage over its own architectural control, but the shear-constrained model does, 19 reaching 1.80 ± 0.01 m/s speed RMSE and 27.0◦ ± 0.3◦ direction MAE against 2.66 ± 0.10 20 m/s and 34.2◦ ± 1.8◦ for the Ekman-constrained arm and 2.61 ± 0.12 m/s and 33.9◦ ± 21 2.0◦ for the physics-free arm, which are statistically indistinguishable from one another; 22 the shear-constrained gain is confirmed against its own architectural control (2.64 ± 0.08 23 m/s, 33.0◦ ± 1.8◦, matching the physics-free arm), ruling out the additional output as 24 the explanation, and converts an erratic extrapolator into a stable one. The benefit of 25 embedding physics into the loss is therefore specific to the extrapolation regime and to the 26 formulation tested, not a generic property of physics-informed learning. By eliminating the 27 need for continuous LiDAR operation after the training phase, the proposed methodology 28 aims to enable cost-effective monitoring of multiple locations with a single instrument.

Abstract: In modern society, air travel plays an important role in connecting people and economies across large distances. Electric aircraft are becoming more realistic as battery and motor technology keeps improving, but it's still unclear whether they could actually work on a large commercial scale. However, there is increasing pressure to make the aviation industry more sustainable, as air travel contributes largely to greenhouse gas emissions. It is important to understand to what extent do current battery and power electronics technological systems restrict the widespread adoption of electric commercial aircraft. The focus of this review is on the technical challenges involved, not the environmental benefits that are often talked about with electric aviation. A deeper understanding of the battery technology, including energy density, weight, charging, degradation, and safety, along with the power electronics and propulsion systems that convert electrical energy into thrust are explored. Other issues including efficiency losses, heat generation, high-voltage systems, and how reliable these systems are also discussed. The research suggests that battery and power-electronics limitations are still major barriers for large-scale all-electric commercial aviation, but continued development could slowly widen the range of aircraft where electric propulsion actually makes sense.

Abstract: Metallic thermal protection structures of reusable high-speed vehicles experience aerodynamic heating, high-temperature hold, and cooling in each flight mission, and the hot and cold faces accumulate different temperature and plasticity histories; after multiple missions, the elastic modulus, yield strength, cyclic hardening parameters, and thermal conductivity no longer vary only with mission count, but form a spatial field varying jointly with position and mission count. What effects such mission-to-mission, spatially non-uniform property evolution brings to structural responses, and what is missed in design when material time-dependence is not considered, are questions to be answered in the design of reusable thermal protection structures. This paper establishes a mission-to-mission, spatially local property update method: each mission solves the thermo-mechanical response with the current property field; at the end of the mission, the thermal exposure and plasticity accumulation at each integration point are extracted, and the material properties are updated through an evolution relation driven by the dual histories, for use in the next mission; the method is embedded in a non-isothermal Chaboche–Perzyna viscoplasticity program. Under a unified mission profile, a thin plate, a perforated plate, and a thick plate are computed for 50 missions each, with pointwise update (Case C) as the reference and no evolution (Case A) and uniform update (Case B) as comparisons. The results show that ignoring evolution gives higher structural stresses than the evolution-tracking case, with the strength-check deviation not exceeding about 4% within the 50-mission window; this is conservative for strength checks, while it continuously underestimates the damage accumulation, with the maximum accumulated equivalent plastic strain underestimated by about 12% at the thin-plate hot face and about 5.4% at the perforated-plate hole edge, biasing life assessment toward the non-conservative side; uniform update deviates by only 0.4% in the thin plate with a mild degradation distribution, but produces 14%–19% stress deviations in the perforated plate with concentrated degradation and alters the temperature field itself. The degradation amplitude assumed in this paper is weak (strength-type parameters decreasing by no more than about 11% within 50 missions), and the above conclusions hold under the current assumption and case-study conditions. The results provide a basis for the thickness design, in-service inspection location determination, and reuse count assessment of reusable thermal protection structures.

Abstract: This paper proposes a novel inertial sensor based on a solid diamagnetic cubic test mass passively suspended by a symmetric configuration of six magnetic quadrupoles. Designed specifically for microgravity environments, this system eliminates the mechanical noise and hysteresis inherent in elastic suspensions, as well as the complexity associated with electrostatic suspension systems. We provide an analysis of the magnetic intrinsic noise floor imposed by eddy current damping, alongside ground-based preliminary experimental results demonstrating potential sensitivities in the order of the \(10^{-10} \text{ m/s}^2/\sqrt{\text{Hz}}\) regime.

Abstract: Telemetry data provide essential information on a spacecraft’s in-orbit operating state and its evolution, supporting system health-state awareness, assessment, and judgment. However, existing data-driven health-state modeling methods may provide unstable representations of healthy operating patterns and have difficulty distinguishing local disturbances from state deviations in complex spacecraft telemetry. To address these problems, this paper proposes a Health Prototype Memory Transformer (HPM-Former) for telemetry health-state baseline modeling. HPM-Former uses convolutional structures to extract local fluctuation features and a Transformer to model global temporal dependencies among multivariate telemetry parameters. Health-memory prototypes learn representative healthy patterns under normal operating conditions, enabling stable representations of healthy spacecraft operation. A health-consistency deviation is constructed from reconstruction deviation, health-prototype matching, and uncertainty in memory addressing to characterize input deviation from the health-state baseline, thereby supporting comprehensive state assessment. Experiments on the European Space Agency Anomalies Dataset (ESA-AD) show that HPM-Former stably represents healthy operating patterns and improves the reliability of current-state assessment against the health-state baseline. It outperforms the comparison methods in corrected event-wise F0.5 and event-wise alarming precision, providing an effective data-driven approach for operational-state assessment of complex spacecraft.

Abstract: This note documents, at a preliminary stage, a propeller concept for small expendable interceptor UAVs cruising near 300 km/h (83 m/s). The concept derives from the DA4052 (9 × 6.75 in variant), an open-geometry small-scale propeller designed at the University of Illinois at Urbana-Champaign and characterised by wind-tunnel measurement in the UIUC Propeller Database. Four geometric modifications are proposed and have been implemented in CAD and realised as a three-blade fused-filament (PLA) article: truncation of both tip and root, giving a modified rotor diameter of 170 mm; an enlarged root chord in place of the parent’s root dip; replacement of the 12 % PROFOIL sections by the 14.6 % thick MH 113 propeller section; and a large increase in blade twist, raising pitch-to-diameter ratio from 0.75 to 2.23. A minimal cylindrical shroud, without lip contouring, diffuser or nozzle, is proposed as a subsequent noise-mitigation step. A single static bench run produced 6 N of thrust at an unrecorded rotational speed. The purpose of this note is to place the concept and its hypotheses on the public record. Computational, structural and experimental characterisation are ongoing. No validated performance claim is made here.

Abstract: Water is a scarce commodity on Mars, yet large amounts of water are needed for crew life support, and in most human mission scenarios, even larger amounts of water are needed to produce propellants for departing Mars for the return trip to Earth. There is evidence that significant amounts of water occur as mineral hydration of magnesium sulfates in the accessible upper layer of Mars regolith at various scattered equatorial locations. Several such magnesium sulfates occur on Mars with water content 20% to 50% of the sulfate mass. Several forms of hydrated MgSO₄ are known to provide a significant share of observed water-equivalent hydrogen in the upper meter of Mars regolith. These include “Gypsum” (MgSO₄⋅2H₂O) containing 20.9% H₂O by weight, and “Epsomite” (MgSO4·7H₂O) (commonly known as “epsom salts”) containing 51% H₂O by weight, as well as hexahydrite (MgSO₄·6H₂O) and starkeyite (MgSO₄·4H₂O). Scans using the neutron spectrometer from orbit show a remarkable correlation between occurrence of S and H in the equatorial region, indicating that hydrated sulfates are a primary source of H₂O there. Recent higher resolution scans using the collimated neutron spectrometer show significant pockets of higher H₂O content. This implies that even higher local concentrations of H₂O almost surely exist within those areas. The power requirements to evolve H₂O from a range of potential hydrated magnesium sulfates are moderate. We suggest that a human mission to Mars at equatorial latitudes based on hydrated sulfates as a source of water is at least as attractive as a mission to higher latitudes based on putative accessible ice.

Abstract: Over the past three decades, composite materials have transitioned from niche aerospace components to primary load-bearing structures in modern aircraft and increasingly critical subsystems in spacecraft. This review synthesizes the state of the art in aerospace composites across polymer, metal, and ceramic-matrix systems, with an emphasis on how material selection, manufacturing routes, and structural integrity considerations jointly determine in-service performance. We discuss the evolution and aerospace rationale of carbon fiber-reinforced polymers (CFRPs), ceramic matrix composites (CMCs), metal matrix composites (MMCs), and emerging bio-inspired hybrid architectures. The review connects materials and processing (prepreg/autoclave, resin transfer molding, automated fiber placement, and additive manufacturing) to dominant damage mechanisms and durability drivers, including barely visible impact damage, delamination, fatigue, moisture/thermal cycling, and defect formation during fabrication. We then examine representative aerospace case studies, including composite-intensive commercial aircraft (e.g., Boeing 787 and Airbus A350) and high-temperature spacecraft/propulsion systems (thermal protection and ablatives). Finally, we highlight enabling directions that are shaping next-generation aerospace composites: self-healing chemistries, multifunctional structural energy storage, fire-resistant systems, corrosion protection strategies, and structural health monitoring supported by data-driven analytics. Despite major progress, challenges persist in scalable defect detection, environmental durability, recyclability, and cost-effective qualification. The review concludes with a research roadmap focused on intelligent, multifunctional, and sustainability-aligned composites capable of reliable operation in extreme aero-thermal environments.

Abstract: Wing icing alters the aerodynamic shape of lifting surfaces and threatens aviation safety. Image-based semantic segmentation provides spatial information on visible wing-surface icing, yet most existing studies rely on wind-tunnel, bench-test, or simulated images, and image-level random splits overestimate generalization for video-derived data. This study builds a complete workflow from flight-test data collection to onboard prototype verification for wing icing segmentation under real flight conditions. A real UAV flight-test dataset of 13,894 pixel-level annotated images was built from backward-view wing videos over the Qinghai–Tibet Plateau, with flight-wise split as the primary protocol to assess generalization to unseen sorties. To address boundary under-segmentation in early-stage icing, an ROI boundary-weighted loss is introduced, combining ROI spatial masking and boundary pixel weighting, supplemented by optical-flow-guided inter-frame smoothing. Relative to the SENet+CBAM attention baseline, this loss raises Recall in the Rice<3% group by up to 6.03 percentage points in the best single run (mean +2.27 pp over three runs) and reduces inter-frame Dice variance by 71% in the single-run comparison (approximately 51% on three-run means). The final scheme achieves 91.34% Dice and 84.07% IoU (three-run means), and a bench-level prototype evaluation on an embedded GPU platform indicates compute feasibility for a 2 s/frame detection cycle. The primary contributions are a real UAV flight-test benchmark with rigorous flight-wise generalization assessment, an ROI boundary-weighted loss that improves early-stage icing recall relative to the attention baseline and reduces inter-frame prediction variance, and a bench-level onboard prototype evaluation on an embedded GPU platform indicating compute feasibility for a 2 s/frame detection cycle.