Comprehensive classification assessment of GNSS observation data quality by fusing k-means and KNN algorithms

Abstract

The observation data is the basis for the global navigation satellite system (GNSS) to provide positioning, navigation and timing (PNT) service, and the observation quality directly determines the performance level of PNT service. At present, the analysis of GNSS observations quality is partial and can only be based on a single index assessment. GNSS observation quality is difficult to analyze comprehensively by fusing multiple indicators. To solve the above problem, the supervised and unsupervised machine learning algorithms are applied, and a new comprehensive and classification method of GNSS observations quality based on the k-means clustering algorithm (k-means) and K-nearest neighbor algorithm (KNN) was proposed. The four core index features of GNSS observations, including data integrity rate, carrier-to-noise-density ratio (CNR), pseudorange multipath and the number of observations per slip, were selected to construct the sample dataset. The sample set was unsupervised clustered based on the k-means algorithm, and the classification label of GNSS observations quality was obtained. Then KNN algorithm was used to construct a comprehensive classification and evaluation model for GNSS observations quality. The data of 30 MGEX stations in the Asia-Pacific region in 2019 were selected for modeling analysis. The experiment results show that: (1) a strong correlation has been revealed between pseudorange multipath, CNR and the number of observations per slip. (2) The average classification correctness rate of the new model was over 90% by 𝑛-fold cross-validation. (3) The new model can effectively realize the automatic evaluation and classification of GNSS observations quality and easily distinguish the superiority and inferiority of the station observations. The relevant results provide a new idea for the automatic classification and assessment of GNSS observation quality.

PPP-AR reference satellite selection based on the observation quality factors

“…Li et al based on machine learning-based classification algorithms, using multiple GNSS signal quality metrics for signal subset selection and weighting schemes [16]. Li et al proposed a comprehensive classification and assessment method of GNSS observation quality, by incorporating machine learning algorithms, which are able to automatically differentiate between good and poor station observation quality [17]. The above comprehensive assessment methods and applications have initially realized the comprehensive assessment of the observational data quality, but the data quality index of different frequencies for each observation at each epoch is not considered.…”