数据驱动的卫星关键部件剩余寿命预测模型

针对目前基于统计学方法对卫星及其关键部件进行剩余寿命预测时普遍存在的建模困难、预测精度不高等问题，为更快速、更精确地预测在轨运行卫星关键零部件的剩余使用寿命（RUL），选取时序数据特征提取能力较强的门控循环单元（GRU）网络构建RUL预测模型。在模型构建时，除了利用卫星遥测数据之外，还将反映卫星通信质量的统计类数据添加到模型中；同时，为进一步提高GRU模型的预测精度，将卷积神经网络（CNN）与GRU融合。最后，以某型号卫星的天线转发器这一关键部件作为研究对象，通过模型预测结果的评价对比，验证CNN-GRU预测模型的RUL预测精度相比GRU模型的有明显提升。

Data-driven model for predicting residual life of key components of satellite

The prediction of the residual useful life (RUL) of satellites and their key components with statistical methods, such as those based on the reliability, involves some problems like modeling difficulty and low prediction accuracy. For seeking a more fast and accurate way to predict the RUL of key components of the satellite in orbit, the gated recurrent unit (GRU) network, with excellent ability of feature extraction of the time series data, is selected to build the RUL prediction model of satellite key components. In addition to use the satellite telemetry data, the satellite operation environmental data are also introduced to the model. At the same time, to improve the prediction accuracy, the method of convolutional neural network (CNN) is combined with the GRU. The proposed new model enjoys the respective advantages of the CNN in the feature extraction and the GRU in the time series prediction in improving the prediction accuracy. Finally, with the whole lifetime telemetry data of antenna transponder of a certain satellite as an example, the prediction accuracy is shown to be significantly improved with the combined CNN-GRU model for RUL prediction.