MambaVision Model for UAV Propeller Damage Identification Driven by Acoustic and Vibration Flight Data

Source abstract

Small unmanned aerial vehicles increasingly operate near people, infrastructure, and dense sensor networks, yet propeller damage remains difficult to identify during flight because acoustic radiation and structural vibration vary with rotor speed, payload, maneuver mode, and air disturbance. This paper proposes an acoustic and vibration flight-data-driven MambaVision model for propeller damage identification. Synchronized microphone and accelerometer measurements are transformed into acoustic spectrogram tokens and vibration order-map tokens, while rotor speed and operating descriptors are retained to reduce flight-condition interference. The model combines local visual attention, selective state-space scanning, and gated cross-modal fusion so that short-range time-frequency damage patterns and longer flight-window dependencies can be learned jointly. Five propeller states are considered, including healthy blades, leading-edge nicking, surface erosion, root cracking, and imbalance deformation. The proposed method achieved 97.1% macro F1, 96.8% macro recall, and 97.0% overall accuracy on held-out flight sessions. Under added payload and crosswind conditions, the recognition accuracy remained above 94%, and ablation results showed that multimodal fusion and state scanning improved both accuracy and convergence stability. These results indicate that acoustic-vibration fusion with MambaVision can support practical UAV propeller health monitoring, especially when in-flight damage signatures are weak, coupled, and affected by changing operating conditions. The framework is suitable for onboard maintenance screening.