空间特征映射与时序 Transformer 协同驱动的低轨星座干扰预测与检测算法研究

杨博宇; 邱堃; 陈哲; 赵进; 高跃

doi:10.11999/JEIT260368

空间特征映射与时序 Transformer 协同驱动的低轨星座干扰预测与检测算法研究

doi: 10.11999/JEIT260368 cstr: 32379.14.JEIT260368

杨博宇¹,
邱堃^2, ,,
陈哲²,
赵进¹,
高跃²

1.
复旦大学计算与智能创新学院空间互联网研究院上海 200438
2.
复旦大学空间互联网研究院上海 200438

基金项目: 国家自然科学基金 (U25A20396)

详细信息

作者简介:
杨博宇：男，博士生，研究方向为卫星干扰检测、机器学习、人工智能等

邱堃：男，助理研究员，研究方向为机器学习、网络路由等

陈哲：男，副研究员，研究方向为卫星互联网、边缘智能系统等

赵进：男，教授，研究方向为网络体系结构、网络功能虚拟化等

高跃：男，教授，研究方向为卫星互联网、压缩感知、机器学习等

通讯作者:
邱堃　qkun@fudan.edu.cn

中图分类号: TN927.2
计量
- 文章访问数: 41
- HTML全文浏览量: 6
- PDF下载量: 5
- 被引次数: 0
出版历程
- 修回日期: 2026-07-29
- 录用日期: 2026-07-29
- 网络出版日期: 2026-08-08

Research on LEO Constellation Interference Prediction and Detection Algorithm Driven by Collaborative Spatial Feature Mapping and Temporal Transformer

YANG Boyu¹,
QIU Kun^{2
, ,},
CHEN Zhe²,
ZHAO Jin¹,
GAO Yue²

1.
College of Computer Science and Artificial Intelligence, Institute of Space Internet, Fudan University, Shanghai 200438, China
2.
Institute of Space Internet, Fudan University, Shanghai 200438, China

Funds: The Natural Science Foundation of China (U25A20396)

摘要

摘要: 针对低轨(LEO)卫星星座高动态拓扑导致的计算开销暴增问题，现有方法主要基于国际电信联盟(ITU)标准的全量物理迭代方法。然而，在轨LEO卫星数量的急剧增加导致干扰预测与检测的计算复杂度阶跃式上升。全量物理迭代仿真下，单个地面站完成单周期干扰评估需超过30个小时。为此，本文提出一种空间特征映射与时序Transformer协同驱动的低轨星座干扰预测与检测算法。首先，建立多星座共存场景下的干扰物理模型，解析出干扰函数的置换不变性、空间稀疏性和时间连续性。调整构建结合可变注意力机制的空间特征映射模块，通过自适应聚焦和对称聚合实现无序变长结构的高效降维压缩。最后，依托时序Transformer精准捕获干扰轨迹的全局演化规律，通过预测未来干扰噪声比(I/N)的时序轨迹，实现干扰事件的敏捷检测。模拟LEO卫星星座场景并以ITU标准为基准进行仿真验证。结果表明：在干扰预测方面，本文算法在20 s长程预测下的均方根误差为0.45 dB；在干扰检测方面，模型能精准刻画物理边界并维持较高的干扰召回率。单次推理耗时稳定在14.2 ms，仅为物理迭代的4.1%。与基线相比，本文算法在多尺度预测步长和动态开销下具有更强自适应性。
- 低轨卫星星座 /
- 干扰预测与检测 /
- 空间特征映射 /
- 时序Transformer /
- 干扰噪声比
Abstract: Objective The continuous putting into use of large-scale low Earth orbit (LEO) satellite constellation groups has caused a serious crowding of orbital space resources. The wide employment of frequency repeating technique by satellite operation persons has seriously brought challenges to the frequency spectrum resources of current LEO satellite network systems. At the same time, the inner quick motion and changing topological structure of LEO satellites bring about violent changes in satellite electric power lead to serious co-frequency interference between satellites. Existing detection methods for interference between satellites mainly depend on compulsory rules formulated by the International Telecommunication Union (ITU), which employ the Interference-to-Noise Ratio (I/N) to serve as the primary indicator. Nevertheless, the link attenuation formulae and satellite position computations that are contained in the calculation procedure bring about a sharp rise in total calculation cost, hence it is hard to satisfy the real-time interference check demands of large-scale satellite constellation working. The recent interference detection algorithms which are based on deep learning are still limited to taking all visible satellites as input, thus they cannot carry out constraint on the spatial sparsity of interference feature distribution. For solving these problems, this paper puts forward a LEO constellation interference prediction and detection method which is based on spatial feature mapping and temporal Transformer, therefore it greatly decreases calculation complexity while it realizes high-efficiency interference prediction and detection. Methods The proposed spatial-temporal Transformer framework efficiently models dynamic co-frequency interference in large-scale heterogeneous multi-constellation LEO networks. By analyzing ITU interference formulations, we extract three core physical properties: permutation invariance, spatial sparsity, and temporal continuity. Consequently, the traditional physical iteration is reframed as a spatio-temporally decoupled feature mapping problem. A dedicated spatial mapping module, leveraging attention and parallel pooling mechanisms, compresses raw satellite parameters into robust permutation-invariant features. This effectively highlights key interference sources while filtering redundant background nodes. Finally, a temporal Transformer captures the dynamics of interference trajectories, achieving high-accuracy prediction and significantly reducing computational complexity for real-time monitoring applications. Results and Discussions The interference prediction and detection algorithm for LEO constellation uses space feature mapping and a time Transformer to carry out interference detection and forecast by making use of high-fidelity data. According to ITU regulations, one high-mobility LEO satellite scene has been simulated, which includes thousands of real satellites from Starlink and OneWeb constellations. Parameter analysis results show that setting reasonable feature dimensions and encoder layers can achieve an optimal balance between feature extraction accuracy and computational cost (Fig. 4 and Fig. 5). By setting model parameters such as sampling step size, prediction accuracy and engineering real-time performance can be comprehensively considered (Fig. 6 and Fig. 7). Simulation results show that compared to baseline methods, the predicted trajectory of the proposed method closely matches the baseline ground truth (Fig. 8), with the absolute deviation controlled within 0.5 dB at a 90% cumulative probability (Fig. 9). Furthermore, the algorithm maintains an extremely high interference detection rate with low false alarms (Fig. 10), and the Root Mean Square Error (RMSE) remains stable at 0.45 dB within a 20.0 s prediction window, significantly outperforming traditional regression and neural network models (Fig. 11). In terms of computational efficiency, the algorithm in this paper takes 14.2 ms for a single inference, which is only 4.1% of the time required by the traditional ITU-based physical iteration method (Table 3), effectively decoupling computational overhead and constellation scale. Conclusions For the problem that LEO satellite network topology has extremely strong dynamic property, and traditional physical iteration interference checking methods have limitations brought by sharply risen computation complexity, this paper puts forward a low-Earth orbit constellation interference checking algorithm which combines spatial feature mapping and temporal Transformer, therefore it effectively makes computation cost and constellation size decouple. The results of simulation show that, when the comparison is made with baseline methods, this algorithm possesses fine long-term dynamic tracking capabilities, and at the same time it keeps conformity with ITU evaluation rules. Experiments which carry on equipment analysis with different parameters have proven that the algorithm we put forward can obtain a balance between feature extraction accuracy and computation spending. By utilizing the long-distance dependence modeling ability of time-domain Transformer, the RMSE still keeps steady at 0.45 dB at identical time step. Through the method of eliminating unnecessary background satellite nodes, the single inference time has been reduced to 14.2 ms. By means of manifold experiments, which contain error distribution assessment and model parameter analysis, the algorithm put forward by us obtains the best precision and full decoupling of calculation cost from constellation scale, hence it gives a highly flexible engineering scheme for future super-large constellation arrangements.
- Low Earth Orbit satellite constellations /
- Co-frequency interference /
- Spatial feature mapping /
- Transformer architecture /
- Interference-to-Noise Ratio

HTML全文

图 1 低轨卫星星座同频干扰示意图

下载: 全尺寸图片幻灯片

图 2 LEO星座时空同频干扰联合预测与检测架构

下载: 全尺寸图片幻灯片

图 3 混合星座干扰态势演化轨迹

下载: 全尺寸图片幻灯片

图 4 不同卫星星座部署场景下潜在干扰源规模$ K $对预测精度的影响

下载: 全尺寸图片幻灯片

图 5 空间映射特征维度$ {d}_{\text{z}} $与编码器层数$ {N}_{\text{layer}} $对算法性能的影响

下载: 全尺寸图片幻灯片

图 6 不同采样步长$ \Delta t $条件下时序滑动窗口长度$ L $与推理耗时的影响

下载: 全尺寸图片幻灯片

图 7 多头自注意力头数$ H $对预测RMSE与推理耗时的影响

下载: 全尺寸图片幻灯片

图 8 不同算法下的干扰噪声比预测输出

下载: 全尺寸图片幻灯片

图 9 不同算法下预测偏差绝对值的累积分布函数

下载: 全尺寸图片幻灯片

图 10 不同算法在干扰检测任务中的查准率-查全率曲线

下载: 全尺寸图片幻灯片

图 11 不同算法在不同预测步长下的性能对比

下载: 全尺寸图片幻灯片

表 1 基本符号及其含义

符号	定义
$ N $	LEO 星座系统的在轨卫星总数量
$ {d}_{i,m}(t) $	$ t $时刻卫星$ i $与地面站$ m $间的空间直线距离
$ {\alpha }_{i,m}(t) $	$ t $时刻卫星$ i $相对于地面站$ m $的仰角
$ {\boldsymbol{V}}_{m}(t) $	$ t $时刻地面站$ m $的可视卫星集合
$ {\boldsymbol{I}}_{m}(t) $	$ t $时刻地面站$ m $的潜在干扰卫星集合
$ {L}_{\text{fs},i,m}(t) $	$ t $时刻第$ i $颗卫星与地面站间的自由空间损耗
$ {L}_{i,m}(t) $	$ t $时刻第$ i $颗卫星与地面站$ m $间的综合链路损耗
$ {I}_{i,m}(t) $	$ t $时刻地面站$ m $接收到卫星$ i $的单星干扰功率
$ {\boldsymbol{h}}_{i,t} $	$ t $时刻潜在干扰卫星集合中各元素的特征向量
$ {\boldsymbol{x}}_{i,t} $	$ t $时刻干扰卫星$ i $经归一化处理后的输入特征向量
$ {\boldsymbol{z}}_{t} $	$ t $时刻经过空间重构算子降维后的空间态势向量
$ {\boldsymbol{P}}_{\text{E}} $	输入特征序列的位置编码

下载: 导出CSV

1 基于空间特征映射与时序Transformer的干扰预测与检测

输入：时间窗口$ {L}_{w} $，干扰卫星集合$ {\boldsymbol{I}}_{m}(t) $及其特征$ {\boldsymbol{h}}_{i,t} $，ITU标　签$ {y}_{\text{ITU}} $
输出：训练完成的网络权重参数$ \theta $
(1)初始化网络权重参数$ \theta $与特征维度$ {d}_{\text{z}} $
(2) 　for epoch = 1 to $ {E}_{\max } $ do
(3) 　　抽取批次大小为$ B $的序列样本
(4) 　　 for $ t=1 $ to $ {L}_{w} $ do
(5) 　　　提取集合$ {\boldsymbol{I}}_{m}(t) $的物理特征$ {\boldsymbol{h}}_{i,t} $并映射获取高维嵌入　　　　　特征矩阵$ {\boldsymbol{X}}_{t} $
(7) 　　　基于$ {\boldsymbol{X}}_{t} $生成$ {\boldsymbol{Q}}_{t} $、$ {\boldsymbol{K}}_{t} $、$ {\boldsymbol{V}}_{t} $，计算空间注意力得分　　　　　 $ {\boldsymbol{A}}_{t} $并结合$ {\boldsymbol{V}}_{t} $经对称池化生成$ {\boldsymbol{z}}_{t} $
(10) 　　end for
(11)　　拼接态势向量$ {\boldsymbol{z}}_{t} $构建特征序列$ \boldsymbol{Z} $，并引入频率基数$ \gamma $注　　　　　入绝对位置编码$ {\boldsymbol{P}}_{\text{E}} $
(12)　　将序列投影至$ H $个子空间，利用$ {\boldsymbol{Q}}_{h} $与$ \boldsymbol{K} $计算多头注意　　　　力$ {\text{head}}_{h} $以获取编码器输出$ {\boldsymbol{Z}}_{\text{output}} $
(13)　　利用向量化算子重构$ {\boldsymbol{Z}}_{\text{output}} $，并通过多层感知机预测　　　　下一时刻指标$ {\hat{y}}_{t+1} $
(14)　　引入正则化系数$ \lambda $构建总体损失$ {\mathcal{L}}_{\text{total}} $，并利用反向传　　　　播更新网络参数$ \theta $
(18) 　end for
(19)返回网络权重参数$ \theta $

下载: 导出CSV

表 2 主要参数设置

系统参数	参数设置
卫星总数$ N $	7450 (6800 Starlink + 650 OneWeb)
载波频率$ {f}_{\text{c}} $	12.5 GHz (Ku 频段)
参考带宽$ {B}_{\text{ref}} $	20 MHz
最小可视仰角 $ {\alpha }_{\min } $	10°
卫星最大等效全向辐射功率$ {E}_{\text{max}} $	34 dBW
天线峰值增益$ {G}_{\text{max}} $	45 dBi
系统噪声温度$ {T}_{\text{sys}} $	150 K
玻尔兹曼常数$ {k}_{B} $	1.38×10^–23 J/K
潜在干扰源规模$ K $	24
时序滑动窗口长度$ {L}_{w} $	20
空间映射特征维度$ {d}_{\text{z}} $	128
编码器层数$ {N}_{\text{layer}} $	4
频率基数参数$ \gamma $	10000
正则化系数$ \lambda $	10^–4
训练批次大小$ {B}_{\text{batch}} $	64
多头自注意力头数$ H $	8
采样步长$ \Delta t $	1s
最大训练轮数$ {E}_{\max } $	200
干扰判定阈值$ \eta $	–6 dB

下载: 导出CSV

表 3 不同算法的计算复杂度与推理开销

系统参数	理论计算复杂度	模型参数量 (M)	单次推理耗时 (ms)	RMSE (dB)
ITU	$ \mathcal{O}({N}_{v}(t){C}_{\text{phy}}) $	-	342.50	-
MLP	$ \mathcal{O}({N}_{v}(t)d_{z}^{2}) $	1.01	8.54	2.30
LSTM	$ \mathcal{O}({L}_{w}{N}_{v}(t)d_{z}^{2}) $	2.14	45.27	1.80
TCN	$ \mathcal{O}({L}_{w}{N}_{v}(t)d_{z}^{2}) $	1.76	38.63	1.15
Transformer	$ \mathcal{O}({N}_{v}{(t)}^{2}{d}_{z}+L_{w}^{2}{d}_{z}+{L}_{w}d_{z}^{2}) $	2.32	85.46	0.95
本文	$ \mathcal{O}(Kd_{z}^{2}+{K}^{2}{d}_{z}+L_{w}^{2}{d}_{z}+{L}_{w}d_{z}^{2}) $	2.18	14.20	0.45

下载: 导出CSV

参考文献(28)

[1]	LIU Jiajia, SHI Yongpeng, FADLULLAH Z M, et al. Space-air-ground integrated network: A survey[J]. IEEE Communications Surveys & Tutorials, 2018, 20(4): 2714–2741. doi: 10.1109/COMST.2018.2841996.
[2]	YUE Pingyue, AN Jianping, ZHANG Jiankang, et al. Low earth orbit satellite security and reliability: Issues, solutions, and the road ahead[J]. IEEE Communications Surveys & Tutorials, 2023, 25(3): 1604–1652. doi: 10.1109/COMST.2023.3296160.
[3]	苏昭阳, 刘留, 艾渤, 等. 面向低轨卫星的星地信道模型综述[J]. 电子与信息学报, 2024, 46(5): 1684–1702. doi: 10.11999/JEIT230941. SU Zhaoyang, LIU Liu, AI Bo, et al. Survey of satellite-ground channel models for low earth orbit satellites[J]. Journal of Electronics & Information Technology, 2024, 46(5): 1684–1702. doi: 10.11999/JEIT230941.
[4]	杜星葵, 束妮娜, 刘春生, 等. 低轨卫星网络安全问题及防御技术综述[J]. 电子与信息学报, 2025, 47(6): 1609–1622. doi: 10.11999/JEIT240957. DU Xingkui, SHU Nina, LIU Chunsheng, et al. Overview of security issues and defense technologies for low earth orbit satellite network[J]. Journal of Electronics & Information Technology, 2025, 47(6): 1609–1622. doi: 10.11999/JEIT240957.
[5]	International Telecommunication Union (ITU). Radio Regulations, Edition of 2024[M]. Geneva: ITU, 2024. (查阅网上资料, 未找到本条文献页码信息, 请补充).
[6]	International Telecommunication Union. Recommendation ITU-R S. 1432–1 Apportionment of the allowable error performance degradations to fixed-satellite service (FSS) hypothetical reference digital paths arising from time invariant interference for systems operating below 30 GHz[S]. Geneva: ITU, 2006.
[7]	武燕燕, 吴松, 邓伟. 6G天地一体通感算智能协同网络资源管理技术综述[J]. 电子与信息学报, 2025, 47(8): 2448–2472. doi: 10.11999/JEIT250140. WU Yanyan, WU Song, and DENG Wei. An overview of resource management technology of 6G integrated communication, sensing, and computation enabled satellite-terrestrial intelligent network[J]. Journal of Electronics & Information Technology, 2025, 47(8): 2448–2472. doi: 10.11999/JEIT250140.
[8]	International Telecommunication Union. Space networks and related software[EB/OL]. https://www.itu.int/en/ITU-R/software/Pages/space-network-software.aspx, 2026.
[9]	刘畅, 魏文康, 李伟. 对地静止轨道卫星网络间的频率干扰分析计算问题[J]. 天地一体化信息网络, 2021, 2(1): 52–59,68. doi: 10.11959/j.issn.2096-8930.2021007. LIU Chang, WEI Wenkang, and LI Wei. Frequency interference analysis and calculation of geostationary-orbit satellite network[J]. Space-Integrated-Ground Information Networks, 2021, 2(1): 52–59,68. doi: 10.11959/j.issn.2096-8930.2021007.
[10]	KIM D, PARK J, CHOI J, et al. Spectrum sharing between low earth orbit satellite and terrestrial networks: A stochastic geometry perspective analysis[J]. IEEE Transactions on Wireless Communications, 2026, 25: 8905–8921. doi: 10.1109/TWC.2025.3646185.
[11]	RU Juanjuan, WANG Ruibo, and ALOUINI M S. Coverage and rate analysis of follower-based LEO satellite networks: A stochastic geometry approach[J]. IEEE Transactions on Wireless Communications, 2026, 25: 12662–12675. doi: 10.1109/TWC.2026.3666549.
[12]	李伟, 严康, 耿静茹, 等. NGSO通信星座系统间同频干扰场景与建模研究[J]. 天地一体化信息网络, 2021, 2(1): 20–27. doi: 10.11959/j.issn.2096-8930.2021003. LI Wei, YAN Kang, GENG Jingru, et al. Research on scenarios and modeling of co-frequency interference between NGSO communication constellation systems[J]. Space-Integrated-Ground Information Networks, 2021, 2(1): 20–27. doi: 10.11959/j.issn.2096-8930.2021003.
[13]	HEYDARISHAHREZA N, HAN Tao, and ANSARI N. Spectrum sharing and interference management for 6G LEO satellite-terrestrial network integration[J]. IEEE Communications Surveys & Tutorials, 2025, 27(5): 2794–2825. doi: 10.1109/COMST.2024.3507019.
[14]	LIM B and VU M. Interference analysis for coexistence of terrestrial networks with satellite services[J]. IEEE Transactions on Wireless Communications, 2024, 23(4): 3146–3161. doi: 10.1109/TWC.2023.3306010.
[15]	TORRENS S A, PETROV V, and JORNET J M. Modeling interference from millimeter wave and terahertz bands cross-links in low earth orbit satellite networks for 6G and beyond[J]. IEEE Journal on Selected Areas in Communications, 2024, 42(5): 1371–1386. doi: 10.1109/JSAC.2024.3365894.
[16]	LIU Xiangnan, ZHANG Haijun, SHENG Min, et al. Ultra dense satellite-enabled 6G networks: Resource optimization and interference management[J]. China Communications, 2023, 20(10): 262–275. doi: 10.23919/JCC.ea.2021-0740.202302.
[17]	ZHANG Bo, GAO Ronghao, GAO Pengyu, et al. Interference-suppressed joint channel and power allocation for downlinks in large-scale satellite networks: A dynamic hypergraph neural network approach[J]. IEEE Transactions on Wireless Communications, 2026, 25: 728–742. doi: 10.1109/TWC.2025.3586230.
[18]	SAIFALDAWLA A, ORTIZ F, LAGUNAS E, et al. Convolutional autoencoders for non-geostationary satellite interference detection[C]. Proceedings of 2024 IEEE International Conference on Communications Workshops (ICC Workshops), Denver, USA, 2024: 1334–1339. doi: 10.1109/ICCWorkshops59551.2024.10615457.
[19]	SAIFALDAWLA A, ORTIZ F, LAGUNAS E, et al. GenAI-based models for NGSO satellites interference detection[J]. IEEE Transactions on Machine Learning in Communications and Networking, 2024, 2: 904–924. doi: 10.1109/TMLCN.2024.3418933.
[20]	YANG Chunyu, YANG Boyu, QIU Kun, et al. DualAttWaveNet: Multiscale attention networks for satellite interference detection[C]. Proceedings of 2025 IEEE/CIC International Conference on Communications in China (ICCC), Shanghai, China, 2025: 1–6. doi: 10.1109/ICCC65529.2025.11148638.
[21]	International Telecommunication Union. Non-geostationary-satellite networks (Non-GSO): ITU regulatory procedures[EB/OL]. https://www.itu.int/en/ITU-R/space/support/nonGSO/Pages/default.aspx, 2026.
[22]	VALLADO D A, CRAWFORD P, HUJSAK R, et al. Revisiting spacetrack report #3[C]. Proceedings of AIAA/AAS Astrodynamics Specialist Conference and Exhibit, Keystone, USA, 2006: AIAA-2006–6753. doi: 10.2514/6.2006-6753.
[23]	International Telecommunication Union. Recommendation ITU-R S. 1503-3 Functional description to be used in developing software tools for determining conformity of non-geostationary-satellite orbit fixed-satellite service systems or networks with limits contained in article 22 of the radio regulations[S]. Geneva: ITU, 2018.
[24]	International Telecommunication Union. Recommendation ITU-R S. 580-6 Radiation diagrams for use as design objectives for antennas of earth stations operating with geostationary satellites[S]. Geneva: ITU, 2004.
[25]	LECUN Y, BENGIO Y, and HINTON G. Deep learning[J]. Nature, 2015, 521(7553): 436–444. doi: 10.1038/nature14539.
[26]	GREFF K, SRIVASTAVA R K, KOUTNIK J, et al. LSTM: A search space odyssey[J]. IEEE Transactions on Neural Networks and Learning Systems, 2017, 28(10): 2222–2232. doi: 10.1109/TNNLS.2016.2582924.
[27]	PELLETIER C, WEBB G I, and PETITJEAN F. Temporal convolutional neural network for the classification of satellite image time series[J]. Remote Sensing, 2019, 11(5): 523. doi: 10.3390/rs11050523.
[28]	VASWANI A, SHAZEER N, PARMAR N, et al. Attention is all you need[C]. Proceedings of the 31st International Conference on Neural Information Processing Systems, Long Beach, USA, 2017: 6000–6010.