기하 편향 트랜스포머를 이용한 GNSS 의사거리 잡음·다중경로 보정과 측위 성능 개선 = Geometry-Biased Transformer for GNSS Pseudorange Correction and Positioning Performance Enhancement

다국어 초록 (Multilingual Abstract)

In dense urban environments, Global Navigation Satellite System (GNSS) Single-Point Positioning(SPP) is frequently compromised by severe multipath effects and Non-Line-of-Sight(NLOS) signals, often resulting in meter-level horizontal errors. Given the hardware constraints of mass-market receivers and the unavailability of high-cost correction infrastructures such as RTK or PPP, there is a critical need for lightweight, software-driven error mitigation solutions. To address this challenge, this dissertation proposes a novel deep learning framework, termed GeoFormer, which utilizes a Geometry-Biased Transformer to estimate satellite -specific signal quality and correct pseudorange errors using only raw measurements.First, the proposed method extracts feature vectors—including Code-Minus-Carrier(CMC), Time-Differenced Code(TDC), and SNR indices—from raw logs(e.g., UBX) of commercial receivers. These features are processed as set-structured time-series tokens. To overcome the scarcity of labeled data, we introduce a self-supervised learning scheme named PIMSM(Pseudorange/Indicator Masked Signal Modeling). This approach trains the model to reconstruct masked signal features, enabling it to learn latent signal patterns without ground truth labels. A key innovation of this work is the Geometry-Biased Attention mechanism, which explicitly injects geometric contexts—derived from satellite elevation, azimuth, and Line-of-Sight(LOS) probability—into the attention layers.The trained model outputs satellite-specific quality masks(m), weights(w), and pseudorange corrections (), which are directly integrated into a standard Weighted Least Squares (WLS) solver. Furthermore, a RINEX-Optional safe mode is designed to ensure robust operation using trigonometric features and empirical models even when external navigation messages are unavailable, thereby improving deployability in real-world scenarios.Extensive field experiments in urban driving scenarios demonstrate the efficacy of the proposed method. Compared to standard SPP, the GeoFormer -corrected SPP reduced the overall horizontal Root Mean Square Error(RMSE) from 6.4m to 5.5m, with consistent improvements in p50 and p90 error statistics. Trajectory analysis confirms that the corrected path adheres closely to road geometry in urban canyons, effectively suppressing the erratic drifts and altitude spikes observed in raw SPP solutions. Additional analysis reveals that the predicted weights and masks behave conservatively for low-SNR and low-elevation satellites, validating that the geometric bias and self-supervised training successfully discriminate signal quality.The contributions of this work are threefold: (1) designing a lightweight deep learning correction module that operates solely on commercial receiver measurements; (2) introducing a Geometry-Biased Transformer with PIMSM self-supervised learning to learn signal quality and correction terms without extra labels; and (3) demonstrating significant reduction in meter-level errors on real-world urban data. As a software-updatable solution, the proposed model offers a practical and scalable pathway to enhance the positioning reliability of mass-market GNSS devices.

목차 (Table of Contents)

• <제목 차례>
• 1. 서론 1
• 1.1. 연구 동기 2
• 1.2. 주요 기여 4
• 1.3. 논문 구성 6
• 2. 연구 배경 7
• 2.1. 전통적인 GNSS 측위 기술 7
• 2.2. GNSS 신호 구조와 관측 데이터 12
• 2.3. GNSS 측위 오차 원인 14
• 2.4. GNSS 데이터 포맷 18
• 2.5. GNSS 보정 방법 20
• 2.6. 딥러닝 기반 GNSS 보정 동향 28
• 2.7. 딥러닝을 활용한 GNSS 품질 평가마스킹보정 30
• 2.8. 트랜스포머와 자기지도 학습 33
• 2.9. 기하정보 기반 학습 36
• 2.10. 본 연구와의 차별점 38
• 3. 기하 편향 트랜스포머 기반의 측위 성능 개선 알고리즘 41
• 3.1. GeoFormer 입력 데이터 전처리 44
• 3.2. 트랜스포머 시계열 입력을 위한 윈도우링 49
• 3.3. GeoFormer 사전학습을 위한 PIMSM 방법 52
• 3.4. 기하 편향 트랜스포머 모델 57
• 3.5. GeoFormer 기반의 측위 보정 알고리즘 64
• 4. 실험환경 및 결과분석 74
• 4.1. 실험 환경 74
• 4.2. 실험 결과 80
• 5. 결론 92
• 5.1. 연구 요약 92
• 5.2. 연구 의의 93