Wenliang Zhang

With the advancement of unmanned aerial vehicle (UAV) technologies, there has been growing demand for enhanced reliability and safety, leading to increased research focus on accurate fault diagnosis in UAV systems. However, existing fault diagnosis methods primarily rely on conventional multivariate time series models, which often fail to account for the unique characteristics of UAV data, such as limited data availability, varying sampling frequencies, sensor data loss caused by faults, and the inherent non-stationarity of flight data. To resolve these challenges, we propose an Unstable Transformer with Imbalanced Data Handling (UTIDH) framework for UAV multivariate time series fault diagnosis. To our knowledge, UTIDH is the first fault diagnosis method that considers the varying sampling frequencies of UAV sensor data. It employs a Multi-task Gaussian Process Imputation (MTGPI) to capture the correlations between UAV components, enabling alignment and supplementation of multivariate time-series data. Additionally, We design a novel Unstable Attention (US-Attention) mechanism to adaptively model the non-stationary characteristics of UAV data, thereby enhancing the detection of genuine fault patterns. To address the issue of limited data for fixed-wing UAVs, we have created and released the first-ever fixed-wing UAV simulation dataset (FUSD), which contains 1104 UAV simulation flight records, with 228 features across 4 flight states, with validity confirmed through transfer learning experiments bridging simulation and real-world flight data. Extensive evaluations on two UAV datasets demonstrate that UTIDH achieves state-of-the-art performance in multivariate time series fault diagnosis for UAVs. Compared to baseline methods, UTIDH improves classification accuracy by 17.79% and detection accuracy by 4.49% on the FUSD dataset. On the publicly available ALFA dataset, it enhances classification accuracy by 8.34% and detection accuracy by 8.33%.