Environment Classification for Global Navigation Satellite Systems Using Attention-Based Recurrent Neural Networks

In this paper, an environment classification method for Global Navigation Satellite System (GNSS) is presented. The goal of the study is to characterize the statistical properties of the historical GNSS data in certain typical environments, so that appropriate localization or navigation algorithms can be chosen to achieve better performances once any environments are recognized in real practice. We extract Dilute of Precision (DOP) value, Carrier-to-Noise Ratio (C/N) and Number of Satellite in View from NMEA-0183 data collected in three real typical environments to characterize the environments. Further, an attention-based Recurrent Neural Network (RNN) is constructed; the historical characteristics extracted above are fed into the RNN. Attention values are then calculated using real-time characteristics and the RNN output in each time steps. High dimensional features are then constructed by soft attention and are used as the input of a fully connected network for classification. The performance of proposed method on the classification task of three typical environments has significantly improvement compared to recurrent neural networks without attention mechanism, and achieves an average accuracy of 94% on the testing set.