Physics-Based Simulation Framework for Multi-Sensor Fusion Benchmarking in Unmanned Aerial Vehicle Concrete Inspection: A Monte Carlo Reproducibility Study
Article
Saripalli, HHB, Saripalli, JL, Rayhan, MM et al. (2026). Physics-Based Simulation Framework for Multi-Sensor Fusion Benchmarking in Unmanned Aerial Vehicle Concrete Inspection: A Monte Carlo Reproducibility Study
. IEEE SENSORS JOURNAL, 10.1109/JSEN.2026.3703331
Saripalli, HHB, Saripalli, JL, Rayhan, MM et al. (2026). Physics-Based Simulation Framework for Multi-Sensor Fusion Benchmarking in Unmanned Aerial Vehicle Concrete Inspection: A Monte Carlo Reproducibility Study
. IEEE SENSORS JOURNAL, 10.1109/JSEN.2026.3703331
Autonomous inspection in nuclear concrete environments requires robust anomaly detection. Real-world validation is limited because operational nuclear facilities impose access restrictions and radiation hazards, forcing most research to rely on mock-ups or simulation. This paper presents a physics-based simulation framework capable of executing comprehensive Unmanned Aerial Vehicle (UAV) inspection scenarios in 18.1 ± 0.6 seconds per seed on a standard desktop workstation. Traditional electromagnetic workflows require hours per scenario, representing a 200–800× speedup under the simulation conditions described herein. Light Detection and Ranging (LiDAR) model uses time-of-flight ranging with signal-to-noise-ratio-(SNR)-dependent detection probabilities; the Ground Penetrating Radar (GPR) model uses electromagnetic wave propagation in lossy media with depth-dependent dielectric properties. Sensor parameters are based on commercial equipment specifications, such as the Trimble X7 LiDAR and Noggin 250 MHz radar (simulated at 600 MHz for near-surface resolution). Framework consistency is assessed via 100 Monte Carlo simulations, yielding a coefficient of variation of 0.39% for fusion detection rates. For 137 synthetic anomalies across 11 material categories, the baseline Boolean OR fusion rule achieves a mean detection rate of 99.75% ± 0.39% under idealized simulation conditions. Because OR fusion admits any single-sensor positive, this high recall comes at the cost of limited per-event precision (0.56) and F1 (0.71); the 99.75% value should therefore be interpreted as an upper-bound recall reference for benchmarking more selective fusion strategies, not as an operational detection-performance estimate. Quadrotor dynamics includes stochastic wind disturbances that are based on Freefly Alta X hover tests. This framework enables systematic comparison of multi-sensor fusion algorithms such as probabilistic, evidence theory, and machine learning approaches in controlled and repeatable settings, eliminating the need for access to nuclear facilities. Idealized noise models and surface interactions are to be further developed.
