| 摘要: |
| 多维遥测数据模式挖掘对卫星状态监测具有重要意义，但遥测参数多、数据量大，很难在短时间内得到精确解。针对这一问题，提出了一种基于矩阵轮廓的模式挖掘方法，利用随机思想搜索近似解，在误差允许的范围内替代精确解。首先对多维遥测数据进行频谱分析，根据模式特征频率计算得到模板长度。然后按照随机的原则循环使用聚类相似性搜索算法（Mueen搒 Algorithm for Similarity Search,MASS计算距离矩阵中的元素，并对主对角线附近的元素进行归零处理，形成多维距离矩阵。最后按列取最小值，生成多维距离矩阵轮廓曲线（Multi-dimensional Distance Matrix Profile,MDMP。在该曲线上，极大值和极小值分别对应于挖掘出的稀有模式和频繁模式位置。实验分析显示，在处理包含15万个采样点的三维遥测数据时，通过该方法在1%的挖掘程度下获得的近似解与精确解之间的位置误差控制在400个采样点以内。 |
| 关键词: 卫星遥测 频繁模式挖掘 稀有模式挖掘 异常检测 矩阵轮廓 |
| DOI：10.20079/j.issn.1001-893x.240226002 |
|
| 基金项目: |
|
| A Multi-dimensional Telemetry Data Pattern Mining Method Based on Matrix Profile |
| LOU Le,LIU Zhen |
| (1.Aerospace Science and Industry Hiwing Group Co.,Ltd.,Beijing 100071,China;2.China Satellite Network Innovation Co.,Ltd.,Beijing 100029,China) |
| Abstract: |
| Multi-dimensional telemetry data pattern mining holds significant importance for satellite status monitoring.However,the sheer volume of telemetry parameters and data poses a challenge in obtaining precise solutions within a short timeframe.To address this issue,the authors propose a matrix profile-based pattern mining approach that employs stochastic principles to search for approximate solutions,which can serve as surrogates for precise solutions within an acceptable error margin.Firstly,spectral analysis is performed on the multi-dimensional telemetry data to determine the template length based on the characteristic frequencies of the patterns.Subsequently,the Mueen搒 algorithm for similarity search(MASS is iteratively applied in a stochastic manner to compute elements within the distance matrix.A crucial step involves zeroing out elements near the main diagonal to form the multi-dimensional distance matrix.Finally,the minimum values are extracted from each column to generate the multi-dimensional distance matrix profile(MDMP.On this profile,the locations of the maximum and minimum values correspond to the identified rare and frequent patterns,respectively.Experimental analysis indicates that when processing three-dimensional telemetry data containing 150 000 sampling points,the proposed method,at a 1% mining depth,is able to constrain the positional error between the approximate and precise solutions within 400 sampling points. |
| Key words: satellite telemetry frequeat pattern mining rate pattern mining anomaly detection matrix profile |
