Satellite Life Prediction Based On Deep Learning Method

As a large aerospace country,we have never been lazy in technological innovation in the aerospace field.As a product of sophisticated technology,the long-term and stable operation and use of satellites are vital.Lithium batteries are the core components of satellite energy storage systems.It conducts research on health diagnosis,detecting working conditions,predicting the future decline of battery performance,and building a complete battery management system.It can effectively prevent satellite failure caused by battery failure,shorten the life of the satellite,and avoid causing irreparable major losses.This article focuses on the prediction of lithium battery life.First,it introduces the existing research methods,which can be roughly divided into two categories.It analyzes the decline of the battery from the perspective of chemical experiments or based on the measured data such as capacitance,voltage,and temperature of the battery.This paper briefly introduces the existing algorithms used in battery life research-recursive filtering,Kalman filtering,etc.But the practicability is not ideal,and the demand for data is large.Secondly,the working principle and main components of the lithium battery are introduced.According to the experimental data of the 18650 lithium battery in the NASA open source database,the changes of the battery voltage and capacitance during the degradation of the lithium battery and how environmental factors affect the battery are analyzed.Impact,whether it will affect the rate of battery degradation.What kind of changes will occur inside the battery due to high or low temperature.Then,the prediction algorithm applied in this paper,the long-short-term memory network,is explained in detail.Starting from the simple loop structure at the beginning,the simple loop structure will accumulate indefinitely and eventually cause the gradient to explode or disappear.Based on this,an optimization method is proposed,that is,introducing a "gating" unit into the loop structure to make the data flow dynamic,avoiding the infinite accumulation of errors,and achieving a good prediction effect on the data.Finally,after the algorithm is selected,the algorithm is optimized.Through data experiments,the network layer structure,length,and loss function are optimized and set to build the most suitable network architecture.After performing dimensionless processing on the data,the input network is used to train the network,and finally the prediction function is implemented.Compared with the traditional ARIMA model prediction algorithm,the prediction effect of long-term and short-term memory networks is very satisfactory.Under the premise of ensuring reliability,the service life of the satellite is greatly extended,which has biggish practical application value and economic value.