Computationally efficient UAV fault diagnosis with adaptive vibration denoising: A signal processing approach for rotorcraft systems

Abstract

Reliable diagnosis of incipient faults in unmanned aerial vehicles (UAVs) requires denoising methods capable of suppressing mixed noise while preserving short-duration, low-SNR fault signatures. This study presents an Adaptive Dual-Tree Complex Wavelet Transform (ADTCWT) with optimized thresholds via K-means Crowding Differentiated Creative Search (KCDCS), coupled with a Dynamic Attention-Weighted Gradient Boosting Ensemble (DAW-GBE) for fault classification. Validation employed the ICUAS-2023 UAV Fault Diagnosis (UAV-FD) benchmark under Gaussian–impulsive noise, and a DJI Phantom 4 Pro dataset with induced motor looseness, blade damage, and foreign-object debris imbalance. On UAV-FD, ADTCWT surpassed baseline DTCWT, a literature wavelet, and improved SVD + VMD, yielding SNR 16.30 dB, PSNR 20.68 dB, and correlation 0.9548 under 10 % fault severity. Time–frequency analysis confirmed the concentration of fault energy in 50–500  Hz, without introducing spurious components. On DJI data (100  Hz sampling), ADTCWT increased SNR from − 4.25 dB to 23.92 dB, achieved correlation 0.997 with MAE 0.023, and maintained stability across − 6.0 to + 6.0 dB input SNR. KCDCS converged within 40 iterations at 0.032  s/iteration. With ADTCWT preprocessing, DAW-GBE attained 96.1 % accuracy, reduced misclassification between blade damage and looseness by 10.8 %, and reached AUC 0.998. All performance gains were statistically significant at p < 0.01, underlining the framework’s suitability for reliable, on-board UAV health monitoring.