利用大型自监督时间序列模型实现跨飞机型 bleed air system 诊断的可迁移性
摘要
Bleed Air System (BAS) 是维持 flight safety and operational efficiency 至关重要的系统，支持 cabin pressurization、air conditioning 和 engine anti-icing 等功能。然而，BAS 的 malfunction 包括 overpressure、low pressure 和 overheating 等，都 pose significant risks，如 cabin depressurization、equipment failure 或 engine damage。当前的 diagnostic 方法在应用于不同 aircraft type 时面临 notable limitations，尤其是对于缺乏 sufficient operational data 的 newer models。为解决这些挑战，本文提出了基于 self-supervised learning 的 foundation model， enables the transfer of diagnostic knowledge 从 mature aircraft (如 A320、A330) to newer ones (如 C919)。通过 self-supervised pretraining，该 model 从 flight signals 中 learn universal feature representations，无需 labeled data， 在 data-scarce scenarios 中效果显著。这一 model 不仅 enhances anomaly detection 和 baseline signal prediction，从而 improve system reliability。本文引入了 cross-model dataset、self-supervised learning framework for BAS diagnostics，以及 a novel Joint Baseline and Anomaly Detection Loss Function，专为 real-world flight data 设计。这些 innovation facilitate the transfer of diagnostic knowledge across aircraft types， ensure robust support for new models 的 early operational stages。此外，本文探讨了 model capacity 与 transferability 之间的关系，为 future research on large-scale flight signal models 提供了基础。
引用
@article{arxiv.2504.09090,
title = {Leveraging Large Self-Supervised Time-Series Models for Transferable Diagnosis in Cross-Aircraft Type Bleed Air System},
author = {Yilin Wang and Peixuan Lei and Xuyang Wang and Liangliang Jiang and Liming Xuan and Wei Cheng and Honghua Zhao and Yuanxiang Li},
journal= {arXiv preprint arXiv:2504.09090},
year = {2025}
}