摘要
为解决卫星遥测数据异常检测面临的数据不平衡且缺乏有标签样本的问题,提出一种基于一维卷积神经网络(1dCNN)迁移学习的异常检测方法。首先利用源域卫星的遥测数据对1dCNN进行预训练,使得模型的卷积层具有卫星状态特征的提取能力;然后将训练好的模型迁移到缺乏标签数据的目标域卫星中;利用目标域有标签样本对预训练模型进行微调,从而实现了对目标域测试集样本的异常检测。为了使1dCNN能够适应遥测数据样本的不平衡性,引入了代价敏感训练策略,建立动态损失函数,从而提升代价敏感一维卷积神经网络(cs-1dCNN)对于异常样本的识别能力。以某两个卫星的电源分系统遥测数据进行了验证,实验结果表明该异常检测迁移方法具有较好的有效性和鲁棒性。
In order to solve the problems of imbalanced data and lack of labeled samples faced by satellite telemetry data anomaly detection,a method of anomaly detection based on one-dimensional convolutional neural network(1dCNN)transfer learning is proposed.Firstly,the source domain satellite data are used to pre-train the 1dCNN to make the model’s convolutional layer have satellite state feature extraction capabilities;then the pre-trained model is transferred to the target domain satellite lacking labeled data;finally,the labeled samples of the target domain are used to fine-tune the pre-trained model so as to achieve anomaly detection of the samples of the target domain test set.In order to make the 1dCNN adapt to the imbalance of satellite telemetry data samples,a cost-sensitive training strategy is introduced.By establishing a dynamic loss function,the cs-1dCNN can improve the capacity of recognizing the abnormal samples.Experiments with the power subsystem data of two anonymous satellites verify the effectiveness and robustness of the proposed anomaly detection transfer method.
作者
陈俊夫
皮德常
张强
CHEN Jun-fu;PI De-chang;ZHANG Qiang(College of Computer Science and Technology,Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China;Beijing Aerospace Control Center,Beijing 100094,China)
出处
《宇航学报》
EI
CAS
CSCD
北大核心
2021年第4期522-530,共9页
Journal of Astronautics
基金
国家自然科学基金(U1433116)
南京航空航天大学研究生创新基地(实验室)开放基金(kfjj20191603)。
关键词
卫星遥测数据
迁移学习
深度学习
异常检测
Satellite telemetry data
Transfer learning
Deep learning
Anomaly detection
