Towards Explainable Artificial Intelligence for GNSS Multipath LSTM Training Models.

Area of Science:

• Geomatics Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Global Navigation Satellite Systems (GNSS) increasingly use deep learning for signal processing.
• Lack of interpretability in deep learning models poses risks for safety-critical GNSS applications.
• Multipath effects are a significant challenge in GNSS signal analysis.

Purpose of the Study:

• To develop an explainable deep learning framework for multipath effect detection in GNSS.
• To enhance the interpretability of deep learning models used in GNSS applications.
• To establish a novel method for anomaly detection in GNSS signals.

Main Methods:

• Developed an interpretable Long Short-Term Memory (LSTM) architecture for GNSS observables.
• Adapted Layer-wise Relevance Propagation (LRP) for GNSS signal analysis to attribute model decisions.
• Correlated LRP relevance scores with signal anomalies for anomaly detection.

Main Results:

• The LSTM model achieved high prediction accuracy across GNSS parameters with maintained interpretability.
• LRP relevance scores consistently increased during anomalous signal conditions (7.34%-32.48%).
• Specific increases observed: multipath parameters (7.34-8.81%), carrier-to-noise ratios (12.50-32.48%), elevation angles (16.10%).

Conclusions:

• The LRP-based analysis enhances GNSS signal quality monitoring and integrity assessment.
• The proposed approach provides a practical framework for detecting and analyzing GNSS signal anomalies.
• This work contributes to developing more reliable and trustworthy navigation systems.