💻 computer science

Enhancing INS/GNSS Performance with Smart Learning Estimator

This paper proposes a "Smart Learning Estimator" (SLE), a deep learning-augmented Kalman filter combining TGAN, LSTM, and RWFNN architectures, which significantly enhances the accuracy and robustness of low-cost INS/GNSS navigation systems during extended GNSS signal outages by overcoming the limitations of traditional linear filters.

Original authors: Reza Safvat, Jafar Keighobadi, Morteza H. Sadeghi, Hossein Nourmohammadi

Published 2026-07-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Reza Safvat, Jafar Keighobadi, Morteza H. Sadeghi, Hossein Nourmohammadi

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

The Big Picture: Navigating Without a Map

Imagine you are driving a car in a thick fog. You have a GPS (like a map on your phone) that tells you exactly where you are, but the signal keeps dropping out because of the fog or tall buildings. You also have a dead reckoning system (like a pedometer and a compass) that counts your steps and tracks your turns.

The Problem: If the GPS signal disappears, your step-counter is great for a few seconds, but it starts to make tiny mistakes. Because it has to add up every single step, those tiny mistakes pile up quickly. After a minute, your "step-counter" might think you are in a different city than you actually are. This is the problem with low-cost sensors (called MEMS) used in drones and small robots.
The Goal: The authors wanted to build a "smart brain" that can guess where the drone is even when the GPS signal is completely gone, by learning from the drone's past movements.

The Solution: The "Smart Learning Estimator" (SLE)

The researchers created a new system called the Smart Learning Estimator (SLE). Think of this system as a highly trained co-pilot who has memorized thousands of flight patterns. Instead of just reacting to the current moment, this co-pilot looks at the history of the flight to predict the future.

The SLE is a "hybrid" engine made of three distinct parts working together, like a three-person pit crew:

1. The Wavelet Fuzzy Filter (The Noise-Canceling Headphones)

What it does: Real-world sensors are messy. They pick up vibrations from the drone's motors, wind gusts, and electrical static. This is like trying to hear a conversation in a loud rock concert.
The Analogy: This part acts like noise-canceling headphones. It uses "wavelets" (a mathematical way to look at sound waves) to separate the useful signal (the drone's actual movement) from the garbage noise (vibrations). It also uses "fuzzy logic," which is like a human saying, "It's not exactly 50 degrees, it's kind of hot," to handle uncertainty better than rigid math.

2. The LSTM (The Long-Term Memory)

What it does: Standard computer programs often forget what happened a few seconds ago. But to know where you are now, you need to remember where you were a minute ago and how you got there.
The Analogy: This is the long-term memory of the co-pilot. While a normal computer might forget a turn you made 10 seconds ago, this "Long Short-Term Memory" (LSTM) network keeps a running diary. It remembers the sequence of events: "We turned left, then accelerated, then dipped." This helps it predict where the drone should be, even if the GPS is silent.

3. The TGAN (The Creative Simulator)

What it does: This is the most advanced part. It's a "Generative Adversarial Network" designed for time-series data.
The Analogy: Imagine a master forger and a detective playing a game.
- The Generator tries to create a fake history of the flight that looks so real, it's indistinguishable from the real thing.
- The Discriminator tries to spot the fakes.
- They play this game over and over until the Generator gets so good at creating realistic flight paths that the Discriminator can't tell the difference.
- Why this matters: When the GPS signal is lost, this system doesn't just guess; it generates a highly probable, realistic path based on what the drone usually does in that situation. It fills in the missing map data with a "best guess" that is mathematically very close to reality.

How They Tested It

The researchers didn't just run this on a computer; they flew a real drone (UAV) equipped with these low-cost sensors.

The Setup: They flew the drone for about 23 minutes (1,400 seconds).
The Trick: They intentionally turned off the GPS signal for different periods (like 480 seconds, 580 seconds, and 800 seconds) to see how well the system could navigate without it.
The Competition: They compared their new "Smart Learning Estimator" against older methods like standard Kalman Filters (the traditional math approach) and other AI models.

The Results

The paper claims that the Smart Learning Estimator (SLE) was the clear winner.

Accuracy: When the GPS was cut, the SLE kept the drone's position and speed much more accurate than the other methods.
The Numbers: In one test, the standard method (Kalman Filter) was off by a huge margin (102 meters in error), while the SLE was only off by about 4.8 meters. That is a massive improvement (roughly 40-50% better).
Complex Maneuvers: The system shined when the drone was doing tricky moves (turning, speeding up, slowing down). The "smart brain" could predict these complex movements better than the rigid math of older systems.

The Trade-Off

The paper admits there is a cost. This "smart brain" is heavy. It requires more computer power to run than the simple math used in older systems.

The Verdict: The authors argue that for high-stakes jobs—like precision farming drones (spraying crops), delivery drones in cities, or military guidance—the extra computer power is worth it to ensure the drone doesn't get lost.

Summary

In short, the authors built a navigation system that combines noise-canceling, long-term memory, and creative simulation to help low-cost drones find their way when the GPS signal disappears. It's like giving a drone a co-pilot who has memorized the rules of flight so well that it can navigate through a storm even if the map is gone.

Technical Summary: Enhancing INS/GNSS Performance with Smart Learning Estimator

Problem Statement
The integration of Global Navigation Satellite Systems (GNSS) with Inertial Navigation Systems (INS) is a standard strategy for mitigating the inherent drift of position, velocity, and orientation estimates. However, this reliance becomes critical when using low-cost Micro-Electro-Mechanical Systems (MEMS) sensors, where uncompensated biases and stochastic noise cause errors to accumulate exponentially. While GNSS provides absolute positioning, it is susceptible to signal obstructions, blockage, and low refresh rates (typically up to 5 Hz). During GNSS outages, traditional integration methods—primarily Kalman Filters (KF) and their variants (EKF, UKF)—struggle to maintain accuracy. These conventional filters often rely on linearized dynamics and Gaussian noise assumptions, which fail to capture the complex, nonlinear dynamics and long-term temporal dependencies inherent in MEMS sensor errors during extended signal interruptions. Furthermore, standard Recurrent Neural Networks (RNNs) suffer from vanishing or exploding gradients, limiting their ability to model long-range sequential data.

Methodology: The Smart Learning Estimator (SLE)
To address these limitations, the authors propose a "Smart Learning Estimator" (SLE), a deep learning-augmented Kalman filtering paradigm. The core of the SLE is a fourteen-layer hybrid architecture designated as TGAN-LSTM-RFW-KF. This architecture integrates three primary components to model nonlinear dynamics and long-term dependencies:

Time-series Generative Adversarial Network (TGAN): This module is designed to learn the underlying distribution of INS/GNSS data and reconstruct latent features. It utilizes a generator, discriminator, embedding, and recovery network to capture complex, multi-modal probability distributions of sensor errors. It employs three loss functions (supervised, reconstruction, and unsupervised) to ensure the model learns both the temporal dynamics and the structural representation of the data.
Long Short-Term Memory (LSTM): Integrated to overcome the limitations of standard RNNs, the LSTM layers utilize gating mechanisms (input, forget, and output gates) to selectively retain salient information over extended sequences. This allows the system to capture long-range temporal dependencies in the vehicle's motion history.
Recurrent Fuzzy Wavelet Neural Network (RFWNN): This component combines fuzzy logic systems with wavelet neural networks. The fuzzy inference mechanism manages intricate nonlinear dynamics, while the wavelet decomposition (using Mexican hat basis functions) provides precise estimation of localized variations and noise suppression. The RFWNN acts as an intelligent preprocessor, denoising the signal and handling uncertainty before it reaches the deep learning layers.

Operational Workflow
The system operates in two distinct modes:

Training Mode (GNSS Available): When GNSS signals are robust, the system uses a Kalman Filter to fuse INS and GNSS data. The resulting error states (differences between INS and GNSS) serve as ground truth targets. The TGAN-LSTM-RFW network is trained on batched data, using INS-derived velocity, position, and specific force as inputs to predict the corresponding errors. The model updates its synaptic weights to encapsulate navigation knowledge.
Prediction Mode (GNSS Outage): During signal interruptions, the system switches to forecasting. The converged TGAN-LSTM-RFW model takes current and historical INS data as input to synthesize simulated geocoordinates and velocity errors. These predicted errors are then fed back into the Kalman Filter to correct the navigation solution, effectively bridging the gap until GNSS signals are restored.

Key Contributions

Hybrid Architecture: The paper introduces a novel 14-layer architecture that synergistically combines TGAN, LSTM, and RFWNN within a Kalman filtering framework. This addresses the specific constraints of standard KFs (linearity and Gaussian assumptions) under high noise and low update frequencies.
Long-Range Dependency Modeling: By utilizing LSTM and TGAN, the method effectively mitigates the inability of traditional filters and simple RNNs to model long-term sequential dependencies in INS drift.
Noise Robustness: The inclusion of the Wavelet component allows for the decomposition of signals into frequency bands, separating high-frequency sensor noise (thermal, vibrational) from useful motion data, thereby enhancing performance in turbulent conditions.
Real-World Validation: The methodology is validated using real flight data from an Unmanned Aerial Vehicle (UAV) equipped with a MEMS IMU and a Garmin GNSS receiver, rather than relying solely on synthetic data.

Experimental Results
The SLE was evaluated against several benchmark algorithms, including standard KF, Feed-forward Time Delay Neural Networks (FTDNN), NARX, RFWNN, GRU, and standalone LSTM-RFW-KF models. The evaluation utilized Root Mean Square Error (RMSE) metrics for position (latitude/longitude) and velocity (North/East) errors during artificial GNSS outages of varying durations (starting at 480s, 580s, and 800s).

Performance: The SLE consistently achieved the lowest RMSE values across all test scenarios. For instance, in a 480-second outage, the SLE reduced velocity error RMSE to $1.07 \times 10^1$ , significantly outperforming the standard KF ( $1.77 \times 10^2$ ) and other neural network variants.
Maneuver Sensitivity: The superiority of the SLE was particularly evident during complex, dynamic, and rotational maneuvers where traditional filters and simpler neural networks showed higher error accumulation.
Accuracy Gain: The authors report an approximate 40–50% improvement in accuracy during extended GNSS outages compared to baseline methods.

Significance and Claims
The paper claims that the Smart Learning Estimator successfully shifts the estimation paradigm from "instantaneous reaction" to "prediction based on recent motion context." By leveraging deep learning to model latent data dynamics and long-term dependencies, the SLE enables precise trajectory reconstruction and sustained accuracy during extended GNSS signal interruptions. The authors assert that while the model incurs a higher computational burden, modern edge-computing hardware (such as the Texas Instruments AM69A processor) makes real-time implementation viable. The work positions this approach as a robust solution for applications requiring high safety and reliability, specifically citing precision farming UAVs, autonomous navigation, and urban package delivery, where maintaining navigation fidelity in GNSS-denied environments is critical.