⚡ electrical engineering

Multi-Modal Robust Fusion Navigation for RLG-INS/GNSS Integrated Systems with Adaptive Fault-Tolerant Capabilities

This paper proposes a multi-modal robust fusion navigation framework for RLG-INS/GNSS systems that integrates fuzzy-based mode switching, attention-enhanced LSTM prediction for signal outages, and an adaptive maximum correntropy unscented Kalman filter with dual-layer fault detection to ensure high accuracy and reliability in complex, noisy environments.

Original authors: Jinsong XU, Manrong Wang, Xintong Zhang, Xin Wang, Zihao Li

Published 2026-07-15

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Jinsong XU, Manrong Wang, Xintong Zhang, Xin Wang, Zihao Li

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

Imagine you are driving a super-smart robot car through a city that is constantly trying to trick it. Sometimes, tall buildings block the sky (so the car can't see the stars it uses for GPS). Sometimes, the radio signals bounce off walls and lie about where the car is. And sometimes, the car's internal sensors start to drift slowly, like a watch that gains a second every hour without you noticing.

This paper presents a new "brain" for that robot car, designed to handle all those tricks at once. It's called a Multi-Modal Robust Fusion Navigation system. Think of it not as a single tool, but as a team of four specialists working together, constantly checking each other's work.

The Four Specialists

1. The "Smart Switch" (Mode Switching)
Usually, navigation systems are stubborn; they stick to one plan even when the environment changes. This new system is like a savvy traffic cop who constantly checks the weather and road conditions. It uses a "fuzzy" logic system (a bit like a human making a judgment call rather than a strict yes/no) to rate how good the GPS signal is.

The Trick: If the signal is perfect, it uses a tight, high-precision mode. If the signal gets shaky, it smoothly switches to a looser mode. If the signal disappears completely (like driving into a tunnel), it switches to a "Pure INS" mode (using only the car's internal sensors).
The Safety Net: To stop the system from panicking and switching back and forth too fast (like a flickering light), it uses a "hysteresis" strategy. This is like a thermostat that waits until the room is really cold before turning on the heat, and waits until it's really warm before turning it off. This keeps the car's path smooth.

2. The "Time-Traveling Predictor" (LSTM with Attention)
When the car enters a tunnel and loses GPS, the internal sensors (Ring Laser Gyros) start to drift. Old systems would just guess and hope. This system uses a special AI brain called an Attention-Enhanced LSTM.

The Analogy: Imagine you are walking blindfolded. You remember your steps. A normal walker just counts steps. This AI is like a walker who remembers how you turned, how fast you spun, and when you accelerated. It uses "attention" to focus on the most important moments (like a sharp turn) rather than treating every second of history equally.
The Prediction: It predicts where the car should be in terms of both position and speed.
The Catch: The system knows that predictions get less trustworthy the longer you are in the dark. So, it uses a "time-decay factor." Imagine inflating a balloon around the prediction; the longer the GPS is gone, the bigger the balloon gets, telling the car, "I'm not 100% sure of this guess anymore." This prevents the car from trusting a bad guess too much.

3. The "Outlier Filter" (Adaptive Maximum Correntropy UKF)
Sometimes, the GPS signal gets hit by "heavy-tailed noise"—basically, wild, crazy errors that look like outliers (like a signal bouncing off a skyscraper and saying the car is on the roof). Standard filters get confused by these wild errors.

The Solution: This system uses a special filter that acts like a bouncer at a club. If a data point is behaving normally, it gets in. If it's acting crazy (an outlier), the filter shrinks its importance.
The Adaptation: The "bouncer" changes their strictness based on how much they trust the GPS. If the GPS is confident, the bouncer is strict about accuracy. If the GPS is messy, the bouncer becomes super strict about ignoring the crazy data. This balance is controlled automatically by the "Smart Switch" from step 1.

4. The "Slow-Drift Detective" (Adaptive SPRT)
Some problems aren't sudden crashes; they are slow, sneaky drifts (soft faults). A sensor might slowly start lying by 0.5 meters every minute. Standard detectors often miss these because they are looking for big, sudden jumps.

The Detective: This system uses a "Sequential Probability Ratio Test" (SPRT) with a special "fading factor." It's like a detective who notices that the suspect is acting slightly suspicious every day. Instead of waiting for a crime to happen, the detective accumulates these small clues.
The Result: This allows the system to catch these slow drifts much faster than before.

What the Simulations Showed

The authors tested this system in a computer simulation (a virtual world) with three specific scenarios:

The Tunnel Test: The car drove into a tunnel for 60 seconds with no GPS.
- Old Way: The car drifted off by 28 meters.
- New Way: The car stayed within 2.0 meters of the true path.
- Best Case: In good conditions, the car was accurate to within 0.08 meters (less than 3 inches).
The "Lying Signal" Test: They injected crazy, non-Gaussian noise (fake signals).
- Standard filters failed, with errors jumping to 0.45 meters.
- The new system kept errors down to 0.15 meters, a massive improvement.
The "Slow Drift" Test: They introduced a slow, growing error.
- Old detectors took about 15 seconds to notice.
- The new "Detective" spotted it in just 5 seconds.
- It caught 95% of these slow faults, with very few false alarms.

What This System is NOT

It is important to note what this paper does not claim:

It does not claim to solve the problem of the car's sensors failing completely or the car driving off a cliff.
It does not claim to work perfectly in every possible real-world environment without testing. The authors admit that the "slow drift" and "crazy noise" scenarios were simulated, and while they tested the system on a real car for short periods, the extreme fault scenarios were not fully induced in the real world yet.
It does not use a "Factor Graph" (a different, more complex mathematical method) or "Vision-based" systems that rely on cameras seeing features (like walls or trees) to navigate, as those were explicitly ruled out or compared as different approaches. This system relies on the Ring Laser Gyro, the GPS, and the AI prediction.

The Bottom Line

This paper suggests that by combining a smart switching system, a predictive AI that knows when to trust itself, a filter that ignores crazy data, and a detective that spots slow lies, you can build a navigation system that is much tougher and more accurate than current methods.

In the simulations, this team approach worked beautifully, keeping the car on track even when the GPS went silent or started lying. The authors are confident enough to say this method significantly improves adaptability and reliability, but they also admit that more real-world testing with actual sensor failures is needed to prove it works everywhere. It's a very promising step forward, but not quite a magic wand yet.

Technical Summary: Multi-Modal Robust Fusion Navigation for RLG-INS/GNSS Integrated Systems

1. Problem Statement

The paper addresses critical reliability and accuracy challenges in Ring Laser Gyro (RLG)-based Inertial Navigation Systems (INS) integrated with Global Navigation Satellite Systems (GNSS) when operating in complex dynamic environments. Specifically, the authors identify four fundamental constraints limiting current integrated navigation systems:

Inflexible Fusion Strategies: Conventional systems often rely on fixed fusion structures (e.g., single-mode loosely or tightly coupled) that fail to adaptively adjust to varying GNSS signal qualities, vehicle dynamics, or accuracy requirements, leading to rapid accuracy degradation during signal blockage or multipath interference.
Error Divergence During Outages: In the event of complete GNSS signal loss (e.g., in tunnels), the system degrades to pure INS mode, causing errors to accumulate and diverge rapidly. Existing AI-assisted methods often output only position increments, failing to provide complete velocity observation information or effectively utilizing prediction covariance for uncertainty propagation.
Non-Gaussian Noise Sensitivity: Traditional Kalman filters assume Gaussian noise. In urban canyons, measurement noise often exhibits non-Gaussian heavy-tailed characteristics (e.g., impulsive interference from multipath), causing standard filters to suffer severe performance degradation. Furthermore, existing robust filters often use empirically fixed kernel bandwidths, lacking adaptive mechanisms to balance accuracy and robustness dynamically.
Delayed Soft Fault Detection: Soft faults (slowly drifting biases) are difficult to detect due to their concealment. Existing detection methods, such as standard Sequential Probability Ratio Tests (SPRT) or residual chi-square tests, often rely on empirical forgetting factors or lack sensitivity to small-amplitude, slowly varying anomalies, resulting in significant detection delays.

2. Methodology

The proposed solution is a Multi-Modal Robust Integrated Navigation Framework that integrates four synergistic modules:

A. Multi-Source Information Quality Assessment & Dynamic Mode Switching

The system employs a Fuzzy Comprehensive Evaluation framework to assess information quality across three dimensions: GNSS signal quality (satellite count, C/N₀, GDOP), INS accuracy (bias stability, divergence rates), and vehicle dynamic states.

Mechanism: Membership functions map these indicators to a [0,1] scale. A weighted comprehensive evaluation determines the confidence of four candidate modes: Tightly-Coupled, Loosely-Coupled, Pure INS, and Auxiliary Sensor-Integrated.
Stability: A hysteresis comparison strategy (with high/low thresholds of 0.75/0.60) prevents frequent mode oscillation at boundaries.
Coupling: The resulting GNSS confidence score is not just a switch signal; it drives the adaptive adjustment of the filter's kernel bandwidth (see Section C).

B. Attention-Enhanced LSTM for Signal Prediction

During GNSS outages, a Scaled Dot-Product Self-Attention-Enhanced LSTM network predicts navigation states.

Inputs: Gyro angular rate increments, acceleration increments, and INS attitude information from the preceding 10 seconds.
Architecture: The network utilizes a causal self-attention mechanism to dynamically weight historical hidden states, focusing on high-excitation moments (e.g., turns).
Dual-Output: Unlike prior single-output methods, this network simultaneously predicts position increments and velocity increments.
Uncertainty Propagation: A time-decay factor dynamically inflates the prediction covariance matrix based on outage duration ( $\gamma_k = \gamma_0 \exp(\lambda_{decay} \Delta t)$ ). This inflated covariance is injected into the filter's measurement noise matrix, ensuring the filter appropriately reduces trust in virtual observations as outage duration increases.

C. Adaptive Maximum Correntropy Unscented Kalman Filter (MCUKF)

To handle non-Gaussian heavy-tailed noise, the standard UKF is replaced with an Adaptive MCUKF.

Criterion: It utilizes the Maximum Correntropy Criterion (MCC) instead of Minimum Mean Square Error (MMSE), using a Gaussian kernel to down-weight large errors (outliers).
Adaptation: The kernel bandwidth ( $\sigma$ $σ$ ) is autonomously regulated by the GNSS confidence score ( $C_{GNSS}$ $C_{GN S S}$ ) derived from the mode switching module: $\sigma_k = \sigma_{max} - (\sigma_{max} - \sigma_{min})C_{GNSS}$ $σ_{k} = σ_{ma x} - (σ_{ma x} - σ_{min}) C_{GN S S}$ .
- High confidence $\rightarrow$ Small $\sigma$ (prioritizes accuracy).
- Low confidence/noise $\rightarrow$ Large $\sigma$ (prioritizes robustness).
Solution: A fixed-point iteration algorithm solves the non-linear cost function, with equivalent measurement noise covariance adjusted based on kernel function values.

D. Adaptive Fading Factor-Enhanced SPRT for Fault Detection

A dual-layer fault detection architecture is proposed to identify soft faults rapidly.

Mechanism: An adaptive fading factor ( $d_k$ ) is constructed from the statistics of the innovation sequence. This factor quantifies the deviation of actual innovation variance from theoretical covariance.
Integration: $d_k$ is embedded into the SPRT likelihood ratio recursion as a multiplicative weight. This amplifies the contribution of current innovations when faults are detected, accelerating the rise of the test statistic.
Architecture: This adaptive SPRT operates in parallel with a standard residual chi-square detector (for abrupt faults), forming an OR-logic dual-layer system.

3. Key Contributions

The paper claims four primary contributions:

Closed-Loop Management: A framework linking multi-source quality assessment, intelligent mode switching, and adaptive filter parameter tuning (kernel bandwidth) into a synergistic closed loop, rather than cascading independent modules.
Dual-Output Prediction with Uncertainty Control: An attention-LSTM network that outputs both position and velocity increments, coupled with a time-decay covariance inflation mechanism to mathematically characterize and propagate prediction uncertainty during signal outages.
Adaptive Robust Filtering: An MCUKF where the kernel bandwidth is dynamically regulated by GNSS confidence, achieving an optimal balance between robustness against non-Gaussian noise and estimation accuracy under Gaussian conditions.
Rapid Soft Fault Detection: A novel SPRT variant embedding an innovation-derived adaptive fading factor, enabling a dual-layer detection architecture that significantly reduces detection delay for slowly varying soft faults compared to standard methods.

4. Experimental Results

The method was validated through simulations involving 50 Monte Carlo runs across three scenarios: GNSS outages, non-Gaussian noise, and soft fault injection.

GNSS Outage Performance: During a 60-second complete signal outage, the proposed method maintained horizontal positioning errors within 2.0 m. This was a significant improvement over pure INS (error > 28 m) and standard LSTM-aided methods (error ~2.3 m). The attention mechanism effectively focused on dynamic maneuvers, and the covariance inflation prevented error divergence.
Non-Gaussian Noise Performance: Under heavy-tailed noise (Student's t-distribution), the adaptive MCUKF achieved a position RMSE of 0.15 m, representing a 66.7% improvement over standard UKF (0.45 m) and a 28.6% improvement over fixed-bandwidth MCUKF (0.21 m). The adaptive bandwidth successfully shifted between accuracy and robustness modes based on signal quality.
Soft Fault Detection: The adaptive SPRT detected soft faults (0.5 m/min bias growth) with a delay of 5.2 seconds, a 50% reduction compared to standard SPRT (10.1 s) and a 67% reduction compared to residual chi-square detection (15.2 s). The detection rate exceeded 95% with a false alarm rate below 5%.
Comprehensive Performance: In a full trajectory simulation, the system achieved horizontal positioning errors below 0.08 m in GNSS-favorable regions (utilizing differential corrections) and maintained errors under 2.0 m during outages. The ablation study confirmed that mode switching contributed most to overall accuracy (36% RMSE reduction), while the dual-output LSTM improved outage performance by 27% over single-output variants.
Computational Efficiency: The total processing time was approximately 28.4 ms per update on an Intel i7 processor, confirming real-time feasibility for 1 Hz GNSS updates.

5. Significance and Claims

The paper asserts that the proposed method significantly enhances the adaptability, accuracy, and reliability of integrated navigation systems in complex environments. By integrating mode switching, time-series prediction, robust filtering, and fault detection within a unified framework coordinated by cross-module confidence coupling, the system achieves performance that exceeds the sum of its individual components.

The authors emphasize that the method provides a systematic solution to the interrelated constraints of signal denial, non-Gaussian noise, and sensor soft faults. They claim the approach is particularly suitable for high-precision applications requiring continuous, reliable positioning where traditional fixed-structure systems fail.

Limitations Acknowledged:
The paper modestly notes several limitations:

Parameter Sensitivity: Performance relies on careful tuning of parameters (e.g., membership boundaries, decay coefficients), though sensitivity analysis provides recommended ranges.
Computational Load: The attention-LSTM module (8.5 ms) may challenge resource-constrained embedded platforms without optimization (e.g., pruning).
Training Dependency: The LSTM model was trained on specific environments; performance in significantly different domains (e.g., marine, off-road) may require transfer learning.
Simulation vs. Reality: While real-world data was used for training and basic validation, the non-Gaussian noise and soft fault scenarios were simulated. Further field testing with controlled fault injection is planned.
Comparison Scope: Due to implementation complexity, quantitative comparisons with Factor Graph Optimization and Variational Bayesian methods were not included in this study.