LSTM-based comprehensive evaluation method of GNSS observation quality for robust RTK positioning

Abstract Real-time and accurate GNSS signal quality analysis is a prerequisite for improving the performance of GNSS RTK positioning in complex scenarios. Its core lies in ensuring the reliability and stability of high-precision positioning through a multi-dimensional signal characteristic evaluation system and RTK robust technology. In view of the fact that the evaluation of GNSS data quality in dynamic and complex scenarios depends mostly on regular threshold determination or independent index analysis, which leads to problems such as static modeling defects, isolated analysis misjudgment, and insufficient real-time, a multi-index fusion evaluation method of GNSS satellite observation quality in urban dynamic and complex environments based on LSTM is proposed. On the basis of adaptive unsupervised clustering, the labels of each cluster data are reordered according to the RTK positioning accuracy, and the LSTM neural network is used to accurately classify the unmodeled data to achieve a comprehensive evaluation and classification of GNSS satellite observation quality. When compared with other machine learning algorithms, LSTM's performance in GPS and BDS data quality classification is stable at the best, and the accuracy of GPS and BDS in the test set is 96.97% and 98.51%, respectively. The advantages and disadvantages of the classification are introduced into the execution judgment of the robust Kalman filter. The experimental results show that the RTK positioning accuracy that takes into account the LSTM data quality classification is better than that is not taken into account.