Deep learning based actuator fault identification for hypersonic vehicles: A zero-shot case

Hypersonic vehicles play a crucial role in various applications and are complex systems that integrate aviation, electronics, computer control, electrical information, and sensing technologies. Owing to this complexity and their harsh working environment, hypersonic vehicles frequently face various faults or failures. Furthermore, these vehicles face a more challenging flight environment and complex dynamic characteristics than traditional aircraft. Building an accurate system model for hypersonic vehicles is considerably difficult. In recent years, extensive research has been conducted on fault diagnosis using deep learning and large datasets. However, the substantial success of deep learning techniques relies on the assumption that sufficient labeled training samples are available. In practical scenarios, problems such as data imbalance, insufficient labeled data, or even the absence of data are frequently encountered. This study investigates zero-shot fault identification for ultra-hypersonic aircraft. In particular, this study focuses on the diagnosis of faults in the flight control system actuators. This study aims to employ deep learning techniques to distinguish whether a specific fault is a loss-of-effectiveness (LoE) or locked-in-place (LiP) fault. In the context of this study, “zero-shot” indicates that no sample data related to the target faults has been included or introduced during the construction of the deep learning model for fault diagnosis. Therefore, the model must rely on alternative methods and features to infer and accurately identify unknown faults for effective fault recognition. To address this problem, artificial descriptions of faults are employed to characterize unknown faults. In particular, a relational network is used to compare the definitions of known fault samples with the descriptions of unknown faults. Furthermore, a deep neural network structure is built by combining convolutional neural networks with long short-term memory networks for feature extraction. Finally, zero-shot fault identification experiments are conducted on a high-hypersonic aircraft with a Winged-cone configuration. Fourteen types of faults, including seven types of LoE faults and seven types of LiP faults, are considered. The highest accuracies range from 81.44% to 89.92% over different types of faults. This demonstrates that it is possible to diagnose and classify different types of faults without training samples, realizing the initial objectives of the fault description-based method. This diagnosis is based on human-defined fault descriptions that allow for fault classification. The proposed zero-shot fault identification method aircraft can mitigate risks, enhance operational reliability, and improve safety in high-hypersonic aircraft operations.