ML-based LOS/NLOS/multipath signal classifiers for GNSS in simulated multipath environment

Abstract

The position accuracy of GNSS is limited by several errors including multipath error. The multipath error is well known as one of the dominant error sources in most of the high-precision GNSS applications, as its fast-changing and site-dependent nature make it challenging to model and mitigate. The Non-Line-of-Sight (NLOS) signals in combination with the original Line-of-Sight (LOS) signal lead to multipath (MP), which results in erroneous range estimation. To mitigate the effect of multipath, detecting the presence of NLOS/multipath signals plays a vital role. In this paper, GPS and IRNSS signals are considered in simulated multipath environment and in open-sky conditions. A machine learning (ML) approach for classification of LOS/NLOS/multipath is presented in both the environments. In this paper, two classifiers are proposed. The proposed classifiers are trained with signal strength, elevation angle, Doppler shift, delta pseudorange, and pseudorange residuals as attributes. The accuracies of these models are computed and compared and it is found that, among all the algorithms, K-Nearest Neighbors, Decision Tree, and its ensemble functions have demonstrated superior performance. Experimental results are presented using GPS L1, IRNSS L5, and S1 data. A comparative analysis on both the classifiers is also presented. Further, to substantiate these results, another experiment is conducted in a complex real-time dynamic multipath environment and the obtained results are also presented.

Appendix-A

1.1 Salient features of prominent supervised classification machine learning (ML) algorithms

The following is the brief introduction to all the supervised ML algorithms implemented in this research:

Logistic Regression (LR) Logistic Regression stands as a foundational linear model renowned for its application in binary classification tasks. It estimates the probability of an instance belonging to a specific class, relying on a linear combination of features.

LR with K-fold cross-validation Integrating Logistic Regression with K-fold cross-validation augments its reliability by systematically dividing the dataset into ten subsets. The model undergoes training on nine subsets while validating on the remaining one in a cyclical manner. This rigorous approach enhances the model's generalization performance, providing a robust estimate of its capabilities. In this research, a tenfold cross-validation is employed.

Naïve Bayes Naïve Bayes, a probabilistic classification algorithm grounded in Bayes' theorem, retains its appeal despite its assumption of feature independence. Its simplicity is particularly advantageous in various applications, such as classification problems, where it often outperforms more complex models.

K-Nearest Neighbors K-Nearest Neighbors (KNN), an instance-based algorithm, classifies data points based on the majority class among their three nearest neighbors. This approach proves flexible and effective, especially in scenarios where local patterns play a crucial role, although it may be sensitive to outliers. In this research K = 3 is considered and analyzed.

Support Vector Machine (SVM)—linear Linear Support Vector Machines seek to identify the optimal hyperplane that separates classes in binary classification problems. While effective for linearly separable data, SVM's true power emerges when combined with non-linear kernels, allowing it to capture intricate relationships in more complex datasets.

SVM with K-fold cross-validation Utilizing Support Vector Machines with K-fold cross-validation extends the model evaluation process, offering a more comprehensive assessment of its generalization performance. Through iterative training and testing across K different subsets of the data, the model's robustness and reliability are thoroughly examined. In this research, K = 20 is chosen and the same is analyzed.

SVM with kernels Enhancing SVM with non-linear kernels, such as Sigmoid, Polynomial, and Radial Basis Function (RBF), empowers it to handle non-linear datasets by transforming the input space. The choice of kernel type depends on the underlying characteristics of the data, allowing for flexibility and adaptability to diverse problem domains.

Decision Tree (DT) Decision Trees are intuitive models that make decisions by recursively partitioning the data based on feature values. While they offer interpretability and the ability to model complex relationships, they may be prone to overfitting, especially in situations with noise or outliers.

DT with ensemble functions (Voting, Random forest, Bagging, AdaBoost) Decision Tree ensembles, including Voting, Random forest, Bagging, and AdaBoost, address the limitations of individual trees. These ensemble methods combine or modify multiple Decision Trees to enhance overall predictive performance, offering robustness and adaptability across various data scenarios.