Previous Abstract Return to Session B3

Session B3: AI and Machine Learning for GNSS Signal Processing, Error Modeling, and Threat Detection

Time–Frequency-Based Classification and Mitigation of GNSS NLOS and Multipath Errors Using Neural Networks
Hyewon Jo, Min-Ji Kim, O-Jong Kim, Department of Aerospace Systems Engineering, Sejong University
Location: Windsong 7-8
Date/Time: Thursday, Sep. 17, 11:48 a.m.

Global Navigation Satellite System (GNSS) technology has become an essential tool for positioning and navigation in our modern society. Although numerous studies have addressed many common errors, multipath error remains a challenging issue due to the complex urban environments and the limitations of current technology in differentiating between direct and reflected signals. As cities grow more crowded, detecting and fixing these errors is vital for maintaining precision positioning system accuracy.
Multipath errors occur when a signal is reflected by surrounding structures before reaching the receiver antenna. This creates two main problems: the receiver might mistake a reflected signal for a direct one, or the signals might overlap and distort the correlation function. In dense urban areas, these distortions can cause position errors of several hundred meters.
This study introduces a novel approach that combines Short-Time Fourier Transform (STFT) and Neural Networks, offering a more efficient and accurate solution compared to traditional hardware and signal processing methods. The goal is to identify and separate signals corrupted by multipath error. Our method has two main parts. First, we apply STFT to the pseudorange residuals and Doppler measurements to extract their respective time-frequency characteristics. Specifically, we focus on how each measurement type reacts differently to multipath error. These unique patterns within the complex-valued STFT coefficients, which represent both magnitude and phase information, provide a clear indicator that helps the network identify whether a signal has been contaminated. Second, our approach employs a three-class signal comprising 'Clean LOS', 'Multipath-affected LOS', and 'NLOS', where 3D building models and ray-tracing are used solely during the training stage to label the LOS or NLOS. This precisely labeled dataset then enables the network to learn and understand the complex signal patterns characteristic of urban environments.
We used high-precision GNSS/INS tools to ensure our training data was accurate by removing clock and atmospheric errors. The distinction between 'Multipath-affected LOS' and 'NLOS' is crucial because 'Multipath-affected LOS' involves both direct and reflected signals, affecting accuracy differently than 'NLOS', which consists only of reflected signals when the direct path is blocked. While 'Multipath-affected LOS' involves a combination of direct and reflected signals, 'NLOS' consists solely of reflected signals when the direct path is completely obstructed. The neural network, trained on LOS and NLOS data labeled via 3D building models, exhibits significant improvement in classification performance. It is particularly effective at distinguishing between Multipath-affected LOS signals and actual NLOS signals. By categorizing these signals separately, the neural network can more effectively identify unique frequency signatures, leading to improved error mitigation in complex urban environments.
One of major advantages of this method is its computational efficiency, as it processes signals in fixed time windows using STFT, reducing the processing load and making it suitable for real-time applications. Furthermore, the neural network can analyze patterns from multiple satellites at the same time to catch errors more effectively. By learning these distinct error patterns, the network can simultaneously analyze the STFT signatures of multiple satellites to identify and isolate the most severely contaminated measurements. This integrated approach ensures that only high-quality data contributes to the final position fix, thereby significantly maintaining the reliability of the overall navigation solution.
We tested this system in real urban canyon environments in Seoul, South Korea. Experimental results showed that our method significantly reduced positioning errors and improved overall accuracy compared to traditional methods.

Previous Abstract Return to Session B3