Improving GNSS Positioning Using Neural-Network-Based Corrections

Abstract Deep neural networks (DNNs) are a promising tool for global navigation satellite system (GNSS) positioning in the presence of multipath and non-line-of-sight errors, owing to their ability model complex errors using data. However, developing DNN GNSS presents various challenges, such as (a) poor numerical conditioning caused by large variations measurements position values across globe, (b) varying number order within set due changing visibility, (c) overfitting available In this work, we address aforementioned challenges propose an approach applying DNN-based corrections initial guess. Our learns output correction pseudorange residuals line-of-sight vectors inputs. The limited variation these input improves our DNN. We design architecture combine information from measurements, which vary both order, leveraging recent advancements set-based deep learning methods. Furthermore, present data augmentation strategy reduce randomizing guesses. We, first, perform simulations show improvement error when applied. After this, demonstrate that outperforms weighted least squares (WLS) baseline on real-world implementation is at github.com/Stanford-NavLab/deep_gnss.