Return to Session B1a

Session B1a: GNSS Integrity and Augmentation

Dual-Branch Transformer with Series Decomposition for Long-Term Orbit Correction
Ziyu Liu, Jingran Wang, School of Automation, Guangdong University of Technology, Guangzhou; Shaolong Zheng, School of Automation, Guangdong University of Technology, Guangdong-HongKong-Macao Joint Laboratory for Smart Discrete Manufacturing; Shiyi Wei, School of Automation, Guangdong University of Technology, Guangzhou; Zhenni Li, School of Automation, Guangdong University of Technology, Center for Intelligent Batch Manufacturing Based on LoT Technology; Shengli Xie, Guangdong Key Laboratory of LoT Information Technology; Mingwei Wang, School of Automation, Guangdong University of Technology, Techtotop Microelectronics Technology Co. Ltd.
Location: Beacon A

Peer Reviewed

Time to First Fix (TTFF) is one of the most important measures of GNSS receiver performance. Under the hot-start mode, the receiver needs only 1 to 2 seconds for positioning. However, under the cold start mode, it takes more than 20 seconds to receive the ephemeris to obtain the accurate satellite position. This is due to the fact that current satellite orbit prediction method cannot provide position information accurately based on the old ephemeris, which is detrimental to fast positioning. To address the above challenge, we propose a novel Transformer-based model for long-term orbit corrections. The model makes full use of the time-series information of historical orbit errors and is able to capture long-term dependencies from orbit error series in different lengths. The predicted orbits can be corrected according to the predicted error components, thereby achieving accurate satellite orbit prediction. Due to the temporal patterns in the orbit error series, the features and information hidden in their complex temporal patterns are often difficult to be extracted and utilised. To tackle the intricate temporal patterns, we introduce an inner series decomposition block to deconstruct the seasonal part and trend-cyclical part from orbit errors series, which also empowers the model with immanent decomposition capacity. In addition, in order to extract the cross-time dependency within the curves as well as the cross-dimension dependency between the curves, we propose the Dual-Branch Attention (DBA) layer to capture the cross-time and cross-dimension dependency. Integrating the proposed series decomposition block and the Dual-Branch Attention(DBA) layer into Transformer, we build a Dual-branch Transformer with series decomposition (DBSD-Transformer) for long-term orbit corrections. Experimental results on satellite orbit error data in 2022 from the BeiDou MEO-10 satellite show the effectiveness of our method. Compared to the prediction method based on the Kepler model, the accuracy increased by more than 80% in prediction for the following 5, 7 and 14 days.



Return to Session B1a