Unmanned aerial vehicles (UAVs) have demonstrated remarkable versatility across a spectrum of missions, and their autonomous flight and real-time control heavily rely on on-board 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 paper 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 an 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.

Carbon fiber-reinforced polymer (CFRP) composites are widely used in aerospace due to their outstanding mechanical properties. However, composite damage detection and localization techniques based on minimal actuator–receiver sensing pairs are still challenging. Based on Lamb wave (LW), current damage detection and localization methods rely on numerous sensors and baseline signals to extract damage information. This study introduces a novel approach for damage detection and localization in CFRP composites using a multimodal gated recurrent unit neural network (MGNN) model. The local maximum energy of the LW is obtained through the continuous wavelet transform (CWT) using complex Morlet wavelets. MGNN combines the time-domain LW-based and CWT-energy signals as an input, dramatically improving the anti-interference and damage feature extraction accuracy. In the MGNN framework, multiple feature mappings are aggregated to establish correlations between damage coordinates and features. This enables damage assessment of composite panels even with a limited dataset, utilizing a minimal sensing actuator–receiver pair configuration. Numerous experiments demonstrate the superior performance of the proposed method compared to other current models. Additionally, anti-interference and ablation experiments substantiate the effectiveness and more robust anti-interference capability of the proposed method. These findings reduce the complexity of sensing networks and detection costs for LW-based methods, thereby significantly improving the accuracy and reliability of composite damage detection.