摘要： 针对航天器遥测数据面临的标签高度稀缺与多变量复杂耦合难题，传统方法往往陷入重架构轻监督或重监督轻拓扑的孤立困境。为此，该研究提出一种架构设计与数据生成深度协同的自监督异常检测范式（DAS-Net）。一方面，构建了包含时序软替换与变量空间拓扑破坏的混合数据降级模块（Hybrid Data Degradation Module，HDDM），以合成物理意义上的多维逻辑异常；另一方面，为有效解耦这种高维伪异常，提出了一种严格的“先时间、后变量”的串行双轴注意力编码器。两者相辅相成：空间降级强制激活了变量轴的拓扑感知潜能，而串行级联机制则确保了在时序先验下精准解耦变量耦合。在火星科学实验室（Mars Science Laboratory，MSL）数据集上的实验与统计显著性检验表明，该范式显著优于现有基准模型，多次独立实验的F1平均期望极显著地提升了约4.4%，基于点调整（Point Adjustment, PA）策略的曲线面积积分（Area Under the Curve，AUC）提升了约1.2%，证明了该协同机制在复杂航天遥测数据中具有卓越的有效性。
中图分类号：TP391.41文献标识码：ADOI:10.19339/j.issn.1674-2583.2026.01.008 中文引用格式：龚旭，李林峰,张凤芳，等. 基于双轴注意力网络的航天器遥测多维时间序列异常检测［J］.集成电路应用，2026，43（01）：52-60. 英文引用格式：Gong Xu，Li Linfeng，Zhang Fengfang，et al. Dual-axis attention network for multivariate time series anomaly detection in spacecraft telemetry［J］.Application of IC，2026，43（01）：52-60.
Dual-axis attention network for multivariate time series anomaly detection in spacecraft telemetry
Gong Xu，Li Linfeng，Zhang Fengfang，Lyu Lin
National Computer System Engineering Research Institute of China
Abstract： To address the challenges of extreme label scarcity and complex multivariate coupling in spacecraft telemetry data, traditional methods often fall into the isolated dilemma of either overemphasizing architecture at the expense of supervision, or prioritizing supervision while neglecting spatial topology. To this end, this paper proposes a novel self-supervised anomaly detection paradigm, DAS-Net, which features a deep synergy between architecture design and data generation. On the one hand, a Hybrid Data Degradation Module (HDDM) incorporating temporal soft replacement and spatial topological disruption of variables is constructed to synthesize multidimensional logical anomalies with clear physical significance. On the other hand, to effectively decouple these highdimensional pseudo-anomalies, a serial dual-axis attention encoder with a strict “time-first, variable-second” order is proposed. These two components are mutually reinforcing: the spatial degradation forcibly activates the topology-aware potential of the variable axis, while the serial cascaded mechanism ensures the precise decoupling of multivariate correlations under temporal priors. Extensive experiments and statistical significance tests on the Mars Science Laboratory (MSL) dataset demonstrate that the proposed paradigm significantly outperforms existing baseline models. Specifically, the average expected F1-score over multiple independent runs achieve a significant absolute improvement of approximately 4.4 percentage points, and the Area Under the Curve (AUC) based on the Point Adjustment (PA) protocol increases by about 1.2 percentage points. These results strongly substantiate the superior effectiveness of the proposed synergistic mechanism in processing complex aerospace telemetry data.
Key words : multivariate time series anomaly detection; self-supervised learning; dual-axis attention mechanism; data
引言
随着航天任务的复杂度不断提升，航天器系统产生的时间序列数据呈现出海量、高维且非线性的特征。这些遥测数据不仅包含反映系统状态的连续读数，还隐含着不同子系统间复杂的耦合关系。如何从这些多变量时间序列（Multivariable Time Series，MTS）中检测出潜在的异常模式，是保障航天器在轨安全与降低地面运维成本的关键挑战［1］。传统的基于阈值监测或简单的统计模型方法，在面对航天器复杂的工况变化时，往往因无法捕捉变量间的动态相关性而导致高误报率或漏报率［2］。
