一种卫星遥测数据多参数预测方法

A Multi-parameter Prediction Method for Satellite Telemetry Data

针对目前遥测数据多参数预测精度不足的问题,文章提出一种基于图注意力网络和时域卷积网络的预测方法。首先,采用多尺度时域卷积残差网络组件提取遥测时序数据在不同时间跨度下的时间依赖关系,以捕捉时间模式;随后,利用图结构学习组件自动获取遥测数据变量之间的空间依赖关系,以捕捉空间模式;最后,将图节点特征表示与数据嵌入表示进行融合,增强图注意力网络在信息聚合和消息传递过程中的学习能力。在某卫星遥测数据集上的应用结果表明:该方法比双向长短期记忆网络(LSTM)模型的平均绝对误差(MAE)降低62.38%,显著提高了遥测数据多参数的预测精度,为保障在轨卫星正常运行提供了更多决策分析支持。

In response to the current problem of insufficient accuracy in multi-parameter prediction of telemetry data,a prediction method based on graph attention networks and time-domain convolutional networks is proposed in the paper.Firstly,a multi-scale time-domain convolutional residual network component is used to extract the time dependencies of telemetry temporal data under different time spans,in order to capture temporal patterns.Subsequently,a graph structure learning component is used to automatically acquire spatial dependencies between telemetry data variables,in order to capture spatial patterns.Finally,the graph node feature representation is fused with the data embedding representation to enhance the learning capability of the graph attention networks in the process of information aggregation and message transmission.The application results on a certain satellite telemetry dataset show that the method reduces the MAE by 62.38%compared to the LSTM model,significantly improving the prediction accuracy of multiple parameters in the telemetry data,and providing more decision-making and analysis support for guaranteeing the normal operation of satellites in orbit.