Abstract
Fixed-wing unmanned aerial vehicles (FW-UAVs) are strategically important in several fields, but are prone to failures during various missions, causing serious damage. Deep learning has brought new opportunities for fault diagnosis techniques for FW-UAVs, but the available fault samples are very limited. In addition, multiple components or multiple modes of faults may occur simultaneously. General methods usually have difficulty dealing with these problems, which poses a significant challenge for FW-UAV fault diagnosis. Therefore, this article proposes a multilevel task knowledge supplement (MTKS) network for FW-UAV fault diagnosis under small samples, enabling simultaneous handling of the faulty component diagnosis (FCD) task (level 1) and the fault severity diagnosis (FSD) task (level 2), while effectively utilizing the common knowledge of the two tasks to improve the model's performance under small samples. Specifically, we first extract features for the FCD and FSD tasks through two task-specific networks. These features contain abundant shared knowledge. Second, we design a novel adaptive knowledge supplementation strategy, which aims to store the public knowledge of the two tasks in a public knowledge pool (PKP) and supplement additional knowledge to the two task-specific networks when needed to improve the performance of the networks. Moreover, a dynamic weighted average (DWA) approach is used to fully train the two task-specific networks. Finally, extensive experimental results show that the proposed MTKS exhibits an excellent performance and is of significant competitiveness with currently popular methods.