欢迎访问中国科学院大学学报,今天是

中国科学院大学学报 ›› 2026, Vol. 43 ›› Issue (6): 823-833.DOI: 10.7523/j.ucas.2025.025

• 电子信息与计算机科学 • 上一篇    下一篇

基于扩散模型和Informer的卫星转台故障数据生成方法

兰青, 林宝军, 朱野()   

  1. 中国科学院微小卫星创新研究院,上海 201304
    中国科学院大学,北京 100049
    上海微小卫星工程中心,上海 201304
  • 收稿日期:2025-01-10 修回日期:2025-04-17 发布日期:2025-07-17
  • 通讯作者: 朱野
  • 基金资助:
    中国科学院青年创新促进会优秀会员(Y2022078)

Fault data generation method for satellite turntable based on diffusion model and Informer

Qing LAN, Baojun LIN, Ye ZHU()   

  1. Innovation Academy for Microsatellite,Chinese Academy of Sciences,Shanghai 201304,China
    University of Chinese Academy of Sciences,Beijing 100049,China
    Shanghai Engineering Center for Microsatellites,Shanghai 201304,China
  • Received:2025-01-10 Revised:2025-04-17 Published:2025-07-17
  • Contact: Ye ZHU

摘要:

针对卫星遥测数据中故障数据样本少,故障检测和诊断模型表现受限的问题,提出一种基于扩散模型和Informer架构的卫星转台故障数据生成方法。该方法结合时间序列分解技术与深度学习模型,旨在生成高质量且具备可解释性的时间序列数据。在扩散正向过程,通过数据分解实现成分解耦,提升了生成数据的多样性和准确性。在逆向生成阶段,利用Informer解码器的稀疏自注意力机制,有效捕捉长时依赖关系和复杂的时序模式。该方法适用于无条件和有条件的故障数据生成,包括数据缺失的场景。实验结果表明,该方法在均方误差、相关评分和相关系数等指标上显著优于生成对抗网络和变分自编码器等方法。

关键词: 卫星转台, 遥测数据, 扩散模型, 时间序列分解, 故障数据生成

Abstract:

This paper proposes a fault data generation method for satellite turntables based on a diffusion model and the Informer architecture, to address the issue of limited fault data samples in satellite telemetry, which restricts the performance of fault detection and diagnosis models. By integrating time series decomposition with deep learning models, the proposed method aims to generate high quality and interpretable time series data. Through data decomposition, the forward diffusion process achieves component decoupling, improving the diversity and accuracy of the generated data. During the reverse generation phase, the Informer decoder leverages a sparse self-attention mechanism to effectively capture long-term dependencies and complex temporal patterns. The method is suitable for both unconditional and conditional fault data generation, including scenarios with missing data. Experimental results demonstrate that the proposed method significantly outperforms generative adversarial networks and variational autoencoders in terms of metrics such as mean squared error, correlation score, and correlation coefficient.

Key words: satellite turntable, telemetry data, diffusion model, time series decomposition, fault data generation

中图分类号: