💻 computer science

SENTINEL-GNSS: Proactive Multi-Horizon Classification of GNSS Signal Degradation Using a Transformer-LSTM Network

SENTINEL-GNSS is a Transformer-LSTM network that proactively classifies GNSS signal degradation 5, 15, and 30 seconds into the future with high accuracy and low latency, significantly outperforming existing baselines in both in-domain and zero-shot evaluations to provide actionable advance warnings for autonomous navigation integrity.

Original authors: Joel Segun Ojerinde, Ibrahim Ishola Mustapha, Joshua Oluwaseun Ogunlade, Claudine Ingeriyinkindi, Md Harun Or Rashid, Wasiu Akande Ahmed

Published 2026-10-08

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Joel Segun Ojerinde, Ibrahim Ishola Mustapha, Joshua Oluwaseun Ogunlade, Claudine Ingeriyinkindi, Md Harun Or Rashid, Wasiu Akande Ahmed

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Autonomous vehicles, from self-driving cars to delivery drones, rely heavily on a global network of satellites to know exactly where they are. In open spaces under a clear sky, this system works with remarkable precision, guiding machines with meter-level accuracy. However, the moment these vehicles enter a dense city, the environment turns hostile. Tall buildings block direct lines of sight to the satellites, causing signals to bounce off concrete and glass before reaching the receiver. This phenomenon, known as multipath, combined with signals that arrive only after reflecting off surfaces, creates a chaotic mix of data that can confuse the navigation system. When a vehicle's computer accepts these corrupted signals, it can suddenly believe it has jumped several meters off its intended path, a dangerous error for any machine operating at speed.

For years, the standard approach to this problem has been reactive. Navigation systems wait until a bad signal is detected, then immediately discard it or switch to backup sensors like gyroscopes. The flaw in this method is timing: by the time the system realizes the data is bad, it has already used that data to update its position, potentially causing a sudden, jarring jump in the vehicle's reported location. A more effective strategy would be to predict the trouble before it happens, allowing the system to prepare in advance. This is the central challenge addressed by a new study from researchers at Beihang University, who have developed a method to foresee signal degradation seconds before it occurs.

The researchers, led by Joel Segun Ojerinde and colleagues, created a system they call SENTINEL-GNSS. Instead of simply checking the quality of a signal at the exact moment it is received, this system looks at a thirty-second history of data to predict what will happen in the near future. It analyzes a specific set of thirty-seven different measurements that a standard satellite receiver can provide, such as the strength of the signal, the number of visible satellites, and the precision of the current position estimate. By feeding this historical data into a sophisticated computer model, the system generates a forecast for three distinct future moments: five seconds, fifteen seconds, and thirty seconds ahead.

The core of their innovation lies in the architecture of the model itself. The researchers combined two powerful types of artificial intelligence. The first part acts like a wide-angle lens, scanning the entire thirty-second window to understand the big picture of how the signals are behaving. The second part acts like a careful observer, tracking the sequence of events in order to understand how the signal quality is changing over time. Together, these components allow the system to spot subtle patterns that indicate a signal is about to degrade, such as a slow decline in signal strength or a sudden drop in the number of available satellites, long before the navigation filter would typically notice the problem.

To train this system, the team used a massive dataset of real-world driving routes collected in the dense urban canyons of Hong Kong. They labeled millions of seconds of data into three categories: clean signals, warning signals that are getting worse, and degraded signals that are unreliable. The model was taught to distinguish between these states, with a special emphasis on correctly identifying the degraded signals, as missing a warning in this category could lead to a safety failure. The training process involved a specific technique that forced the model to pay extra attention to the rare, dangerous moments when signals fail, ensuring it does not simply ignore them in favor of the more common, good signals.

The results of their testing were revealing. When evaluated on the same city where it was trained, the system successfully predicted degraded signals five seconds in advance with a high degree of reliability. Crucially, it caught these dangerous events significantly more often than traditional tree-based computer models, which are commonly used for similar tasks. While the traditional models were slightly better at getting the overall average score right, they frequently missed the specific moments when the signal was about to fail. The new system, by contrast, prioritized catching those failures, which is the most important metric for safety.

Perhaps the most impressive finding came when the researchers tested the system in a completely different city, Tokyo, without showing it any data from that location beforehand. This is a rigorous test of whether the model has learned general rules about how signals behave in cities, or if it has simply memorized the specific streets of Hong Kong. The traditional models failed dramatically in this new environment, essentially losing their ability to detect bad signals. The new system, however, retained its ability to predict degradation, proving that it had learned the underlying physics of signal reflection and blockage rather than just the geography of one city. When combined with a traditional model in a voting system, the prediction accuracy in Tokyo improved even further, reaching a level of reliability that outperformed any single model tested.

The practical value of this advance was demonstrated by integrating the predictions into a navigation filter used for a vehicle in Tokyo. When the system predicted a bad signal was coming five seconds later, the navigation filter automatically reduced its trust in the satellite data and relied more heavily on its internal motion sensors before the corruption actually arrived. This proactive adjustment prevented the large position jumps that usually occur in such scenarios. The result was a reduction in positioning errors by nearly forty percent during the most difficult parts of the drive, a significant improvement over standard methods that wait for the error to happen before reacting.

The study also highlighted the efficiency of their approach. The entire model is small enough to fit on a standard navigation computer and runs incredibly fast, producing predictions for all three future time horizons in a single instant. This speed means the system can be deployed in real-time without slowing down the vehicle's computer. While the researchers noted that their model is not perfect and still requires some fine-tuning when moving between different types of satellite receivers, the ability to foresee signal trouble and act before it happens represents a meaningful step forward. By shifting from a reactive stance to a proactive one, this work offers a path toward safer, more reliable autonomous navigation in the complex and unpredictable environments of modern cities.

Technical Summary: SENTINEL-GNSS

Problem Statement

Autonomous systems relying on Global Navigation Satellite Systems (GNSS) face significant challenges in dense urban environments ("urban canyons"). Signal degradation caused by multipath effects and non-line-of-sight (NLOS) reception leads to abrupt drops in carrier-to-noise density ( $C/N_0$ ) and spikes in horizontal dilution of precision (HDOP). These events introduce meter-level position jumps that corrupt navigation filters.

Current mitigation strategies are reactive: they detect degradation only at the epoch it occurs. By the time a degradation event is confirmed, the navigation filter (e.g., Extended Kalman Filter) has already processed the corrupted measurement, updating the state and covariance with bad data. Reverting to a consistent state requires complex backward smoothing or integrity resets. Furthermore, existing machine learning approaches for GNSS signal characterization (e.g., Random Forests, CNN-LSTMs) primarily focus on current-epoch detection or regression, failing to provide proactive warnings that allow a system to prepare for degradation before it arrives.

Methodology

Dataset and Feature Engineering

The study utilizes the open-source UrbanNav dataset, comprising driving routes in Hong Kong (training) and Tokyo Shinjuku (zero-shot evaluation).

Input Data: Raw GNSS observations processed via RTKLIB (Single-Point Positioning) to generate 1-second epochs.
Feature Vector: A 37-dimensional feature set is extracted per epoch, including positioning solution quality, $C/N_0$ statistics, satellite availability, dilution of precision (PDOP, HDOP, etc.), receiver status, signal dynamics, and atmospheric residuals.
Labeling: Three mutually exclusive classes are assigned based on position error ( $\rho$ $ρ$ ), mean $C/N_0$ $C / N_{0}$ , and satellite count ( $N_{sv}$ $N_{s v}$ ):
- CLEAN: $\rho < 2.0$ m, $C/N_0 > 35$ dB-Hz, $N_{sv} \ge 6$ .
- DEGRADED: $\rho > 5.0$ m OR $C/N_0 < 30$ dB-Hz OR $N_{sv} < 4$ .
- WARNING: All other cases.
Windowing: 30-second sliding windows ( $W=30$ ) of the 37 features are used as input.

Architecture: SENTINEL-GNSS

The proposed SENTINEL-GNSS is a unified deep learning architecture designed for multi-horizon classification. It simultaneously predicts signal quality at three future time steps: +5s, +15s, and +30s.

Input Projection: The $30 \times 37$ input tensor is projected to a model dimension of 128 via a linear layer with batch normalization, followed by fixed sinusoidal positional encodings.
Transformer Encoder: A two-layer Transformer encoder (8 attention heads, feed-forward dimension 512) processes the sequence to capture global temporal dependencies within the 30-second window.
LSTM Decoder: The Transformer output is fed into a two-layer unidirectional LSTM (hidden dimension 256). This component provides causal state-tracking, accumulating signal history directionally to suppress false alarms.
Classification Heads: Four independent linear heads branch from the final LSTM hidden state:
- Three primary heads for horizons $h \in \{5, 15, 30\}$ seconds.
- One auxiliary head for the current epoch ( $h=0$ ) to regularize the shared representation.
Training Objective: The model is trained using class-balanced focal loss to address the class imbalance (CLEAN:WARNING:DEGRADED $\approx$ 18%:60%:22%). The loss prioritizes the minority "DEGRADED" class with specific weights ( $\alpha = [1.0, 2.0, 5.0]$ ) and a focal exponent $\gamma=1.0$ .

Key Contributions

Unified Multi-Horizon Framework: Unlike prior works that focus on single-epoch detection or regression, SENTINEL-GNSS presents a unified model that simultaneously classifies signal quality at multiple future horizons (5s, 15s, 30s), addressing a gap where previous works did not address simultaneous multi-horizon signal quality classification from a unified model.
Proactive Safety Metric Optimization: By employing focal loss and class weighting, the model explicitly prioritizes the recall of the safety-critical "DEGRADED" class over aggregate accuracy, achieving a recall of 0.847 at the 5s horizon.
Zero-Shot Cross-City Generalization: The study demonstrates that the model retains predictive capability on a completely unseen city (Tokyo) and a different receiver hardware (Trimble vs. u-blox/NovAtel), a scenario where traditional tree-based models fail.
Systematic Baseline Inclusion: The paper rigorously includes a "persistence" baseline (copying the current label to future horizons), establishing that any proactive model must outperform simple autocorrelation to claim genuine forecasting ability.

Experimental Results

In-Domain Performance (Hong Kong)

Aggregate Metrics: Tree-based models (Random Forest, XGBoost) outperform SENTINEL-GNSS on aggregate Macro-F1 (XGBoost: 0.919 vs. SENTINEL-GNSS: 0.821). This is attributed to the dataset's reliance on instantaneous features where shallow decision boundaries suffice.
Safety-Critical Metrics: SENTINEL-GNSS significantly outperforms tree baselines on DEGRADED-class recall (0.847 vs. ~0.722 for trees), a 12 percentage point improvement. This indicates superior ability to detect impending failures.
Efficiency: The model runs at 0.045 ms per window on an NVIDIA Tesla T4 GPU, delivering all three horizon outputs in a single forward pass, which is 9.5× faster than running three separate tree models.

Cross-City Generalization (Tokyo Shinjuku)

Model Robustness: In a zero-shot evaluation on Tokyo data:
- Random Forest collapsed, with DEGRADED F1 dropping from 0.822 (in-domain) to 0.148.
- XGBoost remained robust (DEGRADED F1 = 0.784).
- SENTINEL-GNSS retained strong performance with a DEGRADED F1 of 0.753.
Ensemble Performance: A soft-vote ensemble of SENTINEL-GNSS and XGBoost achieved the highest cross-city performance, with a Macro-F1 of 0.892 and DEGRADED F1 of 0.896.

Navigation Integration

The study integrated SENTINEL-GNSS predictions into an adaptive Extended Kalman Filter (EKF) on the Tokyo route.

Mechanism: When the model forecasts a high probability of degradation at +5s, the EKF increases the measurement noise covariance ( $R_t$ ), down-weighting the GNSS measurement and trusting the IMU before the degradation occurs.
Outcome: The proactive approach reduced the positioning RMSE in degraded segments by 38.7% (from 47.4 m to 29.0 m) compared to raw GNSS, outperforming robust baselines like Huber EKF and Student-t Particle Filters.

Significance and Claims

The paper claims that proactive multi-horizon classification is a feasible and necessary step for navigation integrity management.

Actionable Warning: The system provides a 5–30 second advance warning, allowing navigation filters to prepare (e.g., downweighting GNSS, activating alternative sensors) before corrupted data is processed.
Generalization: The Transformer-LSTM architecture captures signal dynamics (e.g., patterns of $C/N_0$ decline) that are more environment-agnostic than the feature-specific memorization of tree-based models, enabling transfer across cities and receiver types.
Design Trade-off: The authors acknowledge that while the model has a lower aggregate Macro-F1 than tree baselines in-domain, this is a deliberate design choice to maximize safety-critical recall. The model's true value lies in its ability to detect the minority "DEGRADED" class and generalize to unseen environments, rather than maximizing overall accuracy.

The study concludes that while calibration and further multi-city validation are needed, SENTINEL-GNSS successfully demonstrates that deep learning can provide the lead time required to maintain navigation integrity in complex urban environments.