Abstract:Unmanned aerial vehicles (UAVs) have demonstrated remarkable versatility across a spectrum of missions, and their autonomous flight and real-time control heavily rely on onboard sensors. However, UAV sensors are susceptible to external cyberattacks and internal failures in complex environments, and the scarcity of real sensor fault data poses challenges for developing accurate detection models. This article is the first to investigate the zero-shot fault detection (ZSFD) problem for UAV sensors, which considers the most challenging scenario where there are no fault data provided for model training. Therefore, a novel convolutional variational autoencoder-generative adversarial network (CVAE-GAN) is proposed, which is designed with a discriminator, and a generator consisting of two encoders and one decoder. First, a fixed-length sliding window is used on normal data of multiple sensors to construct multidimensional input samples, and the generator first learns the latent distribution of fault-free data in adversarial training. Then, the CVAE-GAN generates large reconstruction errors for unseen fault data in testing, which are used as anomaly scores to be compared with adaptive thresholds to achieve fine-grained detection of different sensor faults. Finally, comprehensive experimental results based on real sensor data from a UAV demonstrate the effectiveness of our method in performing ZSFD tasks with different sensor faults, especially in detecting bias, stuck, and impact faults in magnetometer, where the GM metrics outperform the compared models by approximately 7% on average, and the test time efficiency opens up the possibility of realizing online fault detection.

A Zero-Shot Fault Detection Method for UAV Sensors Based on a Novel CVAE-GAN Model