融合注意力机制的航天器重建信号异常检测

Anomaly Detection of Spacecraft Reconstructed Signals Based on Attention Mechanism

针对航天器遥测数据异常检测时先验知识缺失、难以进行有监督条件下机器学习的问题,提出一种融合注意力机制的航天器信号智能异常检测算法。首先,通过注意力机制捕捉航天器遥测数据长距离特征,分析注意力关系矩阵为异常溯源提供指导。其次,采用堆叠自动编码器压缩数据维度并基于此重建输入信号,利用输入信号与重建信号间的残差获取误差重构序列。然后,基于窗口阈值法标记误差重构序列异常索引,实现航天器遥测信号异常检测。最后,通过多通道航天器遥测信号算例验证算法在提高航天器遥测信号异常检测性能与可解释能力的有效性。

Anomaly detection of spacecraft telemetry data by supervised machine learning is a challenging task due to the lack of priori knowledge.A spacecraft signal intelligent anomaly detection algorithm integra⁃ting with attention mechanism is proposed for solution.Firstly,the long⁃distance characteristics of space⁃craft telemetry data are captured by attention mechanism,and an instruction is provided for anomaly trac⁃ing in the attention relationship matrix.Then,the stacked auto⁃encoder compresses the data dimension and reassembles the input signal to obtain the error reconstruction sequence.Furthermore,the anomaly indexes of the error reconstruction sequence are marked by the window threshold method to realize the anomaly detec⁃tion of the spacecraft telemetry signal.Finally,the effectiveness of the proposed algorithm in terms of impro⁃ving the anomaly detection performance and interpretability of spacecraft telemetry signals is verified by u⁃sing a multi⁃channel spacecraft telemetry signal example.