【作者】 刘云；

【导师】 尹传环；

【作者基本信息】 北京交通大学 ， 计算机科学与技术， 2020， 硕士

【摘要】 近年来,随着我国航天事业的发展,通信卫星结构日趋复杂,伴随的是卫星故障种类与数量的剧增。目前,工业上对卫星故障的检测以阈值法和专家经验法为主,如何实现通信卫星故障检测的智能化,成为了当前航天领域研究的热点问题。本文基于某航天机构提供的某颗通信卫星的大量遥测数据,针对基于循环神经网络的故障检测算法进行研究,提出了相应的解决方法。并将其与该领域各已有的算法进行实验对比分析,验证了本文所提出故障检测算法的有效性。本文的主要工作内容与创新点如下:(1)利用某航天机构提供的某通信卫星时长为2年的24维遥测数据,针对每个遥测参数,训练相应的长短时记忆(Long Short Term Memory,LSTM)网络进行时序数据的预测。在此基础上,定义了一种时序数据偏离度,将其用于计算各时间点的加权欧式距离。进一步得到各时间点的故障分数,然后选定合适的阈值进行多维故障检测。通过与现有故障检测算法的实验对比,验证了模型的有效性。(2)利用训练的各个遥测参数的LSTM预测模型,对各个单参数进行预测。提出了一种阈值化方式,为各个遥测参数选取不同的故障判定阈值,从而进行各个单参数的故障检测。在各不同时刻可以获取发生故障的具体参数,得到故障判定矩阵。(3)在内容(2)中单参数故障检测的基础上,对各个卫星遥测参数进行相关性分析,将具有强相关性的遥测参数进行分组。若某时刻仅有个别参数发生故障,但与之具有强相关性的其它参数均未发生故障,则认定其为非故障点,否则为故障点,从而进行系统级的故障检测。通过与现有故障检测算法进行实验对比,展现出了该模型的良好性能。(4)根据以上两个故障检测算法的特点,同时结合上述提出的多维故障检测与基于单维检测与相关性分析的故障检测算法,进行系统级的故障检测。通过实验对比,该模型可以有效的区分故障检测时可能存在的误判,减小故障误报率。更多还原

【Abstract】 In recent years,with the development of China ’s aerospace industry,the structure of communication satellites has become increasingly complex,accompanied by a sharp increase in the types and number of faults.At present,the industrial fault detection of satellite telemetry data is mainly based on the threshold method and expert experience method.How to realize the intelligence of fault detection has become a hot topic in the current aerospace research.Based on a large amount of telemetry data from a communication satellite provided by a space agency,this paper studies the fault detection algorithm based on recurrent neural network,proposes corresponding solutions,and compares it with various existing algorithms in the field.The experimental comparison proves the effectiveness of the proposed fault detection algorithm.The main contents and innovations of this paper are as follows:(1)This paper uses the 24-dimensional telemetry data of a communication satellite provided by the space agency for research.For each telemetry parameter,a corresponding LSTM model is trained to predict time series data.Based on the sequence prediction of LSTM,the concept of time series data deviation is designed and used to calculate the weighted Euclidean distance at each time point.Thus,the fault scores at each time point are obtained,and then a suitable threshold is selected for multi-dimensional fault detection.The experimental comparison with the existing fault detection algorithm verifies the effectiveness of the model.(2)This paper uses the trained LSTM prediction model of each telemetry parameter to predict each single parameter data.A thresholding method is proposed to select different fault judgment thresholds for each telemetry parameter,so as to perform fault detection for each single parameter.The specific parameters of the fault can be obtained at different times,and a failure determination matrix can be obtained.(3)Based on single parameter fault detection in content(2),this paper analyzes the correlation of each satellite telemetry parameter and groups the telemetry parameters.If only some parameters are alarmed at a time,but other parameters with strong correlation are not alarmed,it is considered as a non-fault point.In this way,this paper conducts system-level fault detection.Through experimental comparison with the existing fault detection algorithm,it shows the good performance of the model.(4)According to the characteristics of the above two fault detection algorithms,this paper combines the multi-dimensional fault detection and the fault detection algorithm based on single-dimensional detection and correlation analysis proposed above to perform system-level fault detection.Through experimental comparison,the model can effectively distinguish the possible misjudgments in fault detection and reduce the false alarm rate.更多还原