Telemetry Data Prediction Method Based On Attention Mec Hanism Neural Network

Spacecraft telemetry data is the only basis for the satellite ground applicati on system to interpret the status of spaceborne equipment.It is an effective m ethod for spacecraft anomaly detection to accurately predict the future change t rend of telemetry data.The traditional telemetry data prediction method based on sliding autoregressive average model and machine learning has shortcomings in the following aspects:(1)Other telemetry data related to the predicted tele metry parameters are not comprehensively considered;(2)Most research focuses on the telemetry time series Point prediction,and the existing interval predictio n methods do not consider the performance indicators such as interval width a nd interval coverage;(3)The prediction effect on stationary data is better,and the telemetry prediction with catastrophic characteristics is not good.Based on the full investigation of the point prediction and interval predicti on methods of the existing telemetry time series,the paper proposes a self-codi ng neural network method based on the attention mechanism for the above-me ntioned defects of telemetry data prediction.Prediction and interval prediction,and experiments were conducted using pilot special ground support system qua ntum satellite telemetry data.The main research contents of the paper are as f ollows:1)According to the characteristics of high density and slow change of sate llite telemetry data,preprocess the telemetry data.Including compressing the tra ining data to construct new feature values ??to meet the input and output re quirements of the proposed prediction model;using data visualization analysis methods to visually analyze the data,drawing a box plot to analyze the charac teristics of the original data distribution;Normalize the data with non-uniform outlines;use the maximum information coefficient to analyze the correlation of telemetry data,and select the telemetry variables associated with the predicted target;2)Aiming at the problem of multi-step prediction of mutant satellite telem etry data,a self-encoder model of compound long-short-term memory neural ne twork combined with attention mechanism is proposed.This model learns telem etry multivariate data with a certain degree of correlation to complete automati c feature extraction and multi-step prediction of telemetry sequences.Using qua ntum satellite data for experimental verification,and comparing the vector autoregression model and the support vector regression model,it is verified that thi s method has higher prediction accuracy on the root mean square error index;Aiming at the problem of interval prediction of telemetry data,a prediction method is proposed that combines a self-encoder model of a composite long-t erm and short-term memory neural network combined with an attention mechan ism and an improved quality-driven interval prediction method.This method dir ectly learns the telemetry data with a certain degree of correlation through a self-encoding model,and outputs the upper and lower bounds and point predicti on results of the prediction interval.Using quantum satellite mission telemetry data for verification experiments,the results show that compared with the tradit ional vector autoregression,correlation vector machine,Gaussian process regress ion and other models,the method in this paper has the root mean square error,interval coverage,prediction interval width,and interval distance indicators.Be tter prediction performance.