Improvement of satellite orbit prediction accuracy and quality with deep learning and spectral analysis
Pihlajasalo, Jaakko Mikael (2017)
2017
Teknis-luonnontieteellinen
Teknis-luonnontieteellinen tiedekunta - Faculty of Natural Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2017-11-08
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tty-201710262067
https://urn.fi/URN:NBN:fi:tty-201710262067
Tiivistelmä
In this study, we consider methods of spectral analysis and deep learning to improve accuracy and quality of GNSS satellite orbit predictions. The quality of predictions decreases when satellite’s orbit is fixed with maneuvers, which causes satellite orbit predictions to fail. In our previous research orbit accuracy has been improved using analytical and data-driven models.
The goal of this study is to improve quality of BeiDou satellite predictions with spectral analysis and to improve accuracy of GNSS satellite orbit predictions with deep learning. Both methods are used to improve the existing model for satellite orbit predictions. The improvement in quality is done by predicting unhealthy time periods of BeiDou’s GEO and IGSO satellites. Periodogram was used to estimate a period of health parameter from broadcast. The improvement in accuracy was done with convolutional neural networks. The convolutional neural network was used to predict RTN errors of existing models orbit predictions. The errors were predited from RTN errors from the beginning of the orbit prediction. With these error predictions the predicted orbit can be corrected.
The main results of this study were that unhealthy periods can be predicted from the broadcast data and even a simple convolutional neural network can improve orbit prediction accuracy significantly. Health prediction algorithm could be created from the estimated periods to predict health successfully. Health prediction worked with BeiDou’s GEO and IGSO satellites and predicted over 70% of unhealthy periods in testing. Two different methods for error correction with convolutional neural networks were created and tested with GPS satellites. The better method was also tested with BeiDou satellites of all orbit types. The method improves two week orbit prediction accuracy over 40% on average for all GPS and BeiDou satellites. The best improvements in accuracy were achieved with GEO and IGSO satellites.
The goal of this study is to improve quality of BeiDou satellite predictions with spectral analysis and to improve accuracy of GNSS satellite orbit predictions with deep learning. Both methods are used to improve the existing model for satellite orbit predictions. The improvement in quality is done by predicting unhealthy time periods of BeiDou’s GEO and IGSO satellites. Periodogram was used to estimate a period of health parameter from broadcast. The improvement in accuracy was done with convolutional neural networks. The convolutional neural network was used to predict RTN errors of existing models orbit predictions. The errors were predited from RTN errors from the beginning of the orbit prediction. With these error predictions the predicted orbit can be corrected.
The main results of this study were that unhealthy periods can be predicted from the broadcast data and even a simple convolutional neural network can improve orbit prediction accuracy significantly. Health prediction algorithm could be created from the estimated periods to predict health successfully. Health prediction worked with BeiDou’s GEO and IGSO satellites and predicted over 70% of unhealthy periods in testing. Two different methods for error correction with convolutional neural networks were created and tested with GPS satellites. The better method was also tested with BeiDou satellites of all orbit types. The method improves two week orbit prediction accuracy over 40% on average for all GPS and BeiDou satellites. The best improvements in accuracy were achieved with GEO and IGSO satellites.