Coupled Noise-Aware UAV Fault Diagnosis Based on Learnable Wavelet Packet Transform and Scale-Graph Label Enhancement

As a critical component in the Internet of Things (IoT) ecosystem, unmanned aerial vehicles (UAVs) play a pivotal role in low-altitude economy applications, where flight safety is essential for reliable and efficient operations. Although significant progress has been made in UAV intelligent fault diagnosis, it is challenging to effectively address the noise issues in complex flight environments, especially when sample noise and label noise coexist. To this end, this paper first investigates the sample-label noise coupling fault diagnosis (SLNFD) problem in UAVs, and proposes a novel two-stage framework called Learnable Wavelet Packet Transform and Scale Graph Label Enhancement (LWPT-SGLE). Specifically, the LWPT subnetwork integrates wavelet packet transform and deep learning models, and a convolutional layer is employed to replace the traditional filter component to adaptively learn denoising parameters for efficient noisy sample reconstruction. The SGLE subnetwork is designed for label noise, which extracts multi-scale data embeddings, identifies clean samples using a small loss criterion, and applies graph embedding learning to relabel the remaining samples. Comprehensive experiments on the RflyMAD dataset demonstrate the superiority of the proposed framework, achieving outstanding denoising performance, with a reconstructed signal <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">R</sup>² reaching 0.9974, and outperforming existing methods by at least 7% when handling label noise.