⚡ electrical engineering

Underwater MEMS Gyrocompassing: A Virtual Testing Ground

This paper proposes a machine learning-based framework that refines disturbed inertial signals to enable accurate and resilient underwater gyrocompassing for unmanned underwater vehicles (UUVs) despite environmental challenges like ocean currents.

Original authors: Daniel Engelsman, Itzik Klein

Published 2026-05-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Daniel Engelsman, Itzik Klein

Original paper licensed under CC BY 4.0 (http://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 Problem: Finding North Under the Waves

Imagine you are driving a car, but you are blindfolded, and the GPS is broken. To know where you are going, you need to know which way is "North." On land, you might use a compass (which points to magnetic north) or look at the stars.

But underwater, things are different:

Magnetic compasses fail: The metal in the submarine (Unmanned Underwater Vehicle or UUV) and the electronics inside it mess up the magnetic signal, like trying to hear a whisper in a rock concert.
GPS doesn't work: Radio waves can't travel through water, so the satellite signal is lost.
The "True North" trick: The Earth spins. If you have a very sensitive gyroscope (a spinning wheel that resists changing direction), you can feel the Earth rotating underneath you. By measuring this tiny spin, you can figure out exactly where North is. This process is called gyrocompassing.

The Catch: The Earth spins very slowly (one full turn every 24 hours). The signal is incredibly faint. It's like trying to hear a single drop of water falling in a quiet library.

The New Obstacle: The Ocean is Not Quiet

The paper points out a major problem: The ocean is rarely quiet. Currents, waves, and turbulence push the submarine around.

The Analogy: Imagine trying to listen to that single drop of water (the Earth's spin) while someone is shaking the table the library is sitting on. The shaking (ocean currents) creates so much "noise" that the tiny signal gets lost.
The Consequence: Traditional methods that try to filter out this noise usually fail because they assume the submarine is sitting perfectly still. Once the submarine starts moving or rocking, the math breaks down.

The Solution: A "Virtual Training Ground"

The authors, Daniel Engelsman and Itzik Klein, propose a new way to solve this using Machine Learning (AI). Instead of trying to write a perfect math equation to cancel out the noise, they taught a computer to learn how to find the signal.

Here is how they did it, step-by-step:

1. Building a Digital Twin

They couldn't just test this on a real submarine in the ocean immediately because it's expensive and risky. So, they built a virtual simulator.

The Analogy: Think of a flight simulator for pilots. They didn't just fly a plane; they created a computer program that mimics exactly how a plane reacts to wind, turbulence, and gravity.
What they did: They created a digital model of a submarine and programmed it to react to three types of "shaking":
- Impulse: A sudden bump (like hitting a rock).
- Step: A sudden push that stays (like a strong current hitting it).
- Sine Wave: A rhythmic rocking (like ocean waves).

2. The "Fake Noise" Recipe

They took real data from a stationary sensor (where the Earth's spin is easy to see) and mixed it with the "fake noise" from their simulator.

The Analogy: Imagine you have a recording of a singer's voice (the Earth's signal). You then record yourself banging pots and pans (the ocean noise). You mix them together to create a messy recording.
The Goal: They fed thousands of these "messy recordings" into a computer program (a Deep Learning model) and told it: "Here is the messy noise. Can you figure out where the singer is?"

3. The Learning Process

The computer tried to guess the direction of North. Every time it was wrong, it adjusted its internal "brain" (its parameters) to do better next time.

The Result: The AI learned to ignore the specific patterns of the "pot-banging" (ocean currents) and focus only on the "singer" (the Earth's rotation).

The Results: Beating the Old Ways

The authors tested their new AI against old-school methods (like standard filters used in engineering).

The Competition: They compared their AI to four traditional "noise-canceling" techniques (Wavelet, Wiener, Savitzky-Golay, and FIR filters).
The Scoreboard: The traditional methods struggled when the ocean was rough. Their error rates went up significantly.
The Winner: The AI model was much better. It reduced the error by 30% to 60% compared to the old methods, even when the "shaking" was very intense.

The Bottom Line

This paper doesn't claim to have built a new submarine or solved every navigation problem in the world. Instead, it proves a specific concept:

Virtual Testing Works: You can simulate underwater chaos on a computer to train AI.
AI is Stronger: Machine learning can learn to separate the Earth's tiny spin signal from the loud noise of ocean currents better than traditional math formulas can.
Cheaper Sensors: Because this method is so good at cleaning up the signal, it means we might be able to use cheaper, off-the-shelf gyroscopes (MEMS) for underwater navigation, rather than needing incredibly expensive, high-end sensors.

In short: They taught a computer to find North in a stormy ocean by training it in a virtual storm, proving that AI can do a better job than old-school math at keeping a submarine on course.

Technical Summary: Underwater MEMS Gyrocompassing: A Virtual Testing Ground

Problem Statement
Accurate heading determination is fundamental for tracking trajectories during extended underwater missions. While roll and pitch can be reliably computed using accelerometers and local gravity, determining the initial heading (yaw) underwater presents significant challenges. Traditional methods relying on magnetometers are susceptible to electromagnetic interference from the Unmanned Underwater Vehicle's (UUV) own electronics and the complex underwater environment. Satellite and RF-based navigation systems are ineffective underwater. Consequently, gyrocompassing—utilizing the Earth's rotation rate ( $\omega_{ie}$ )—is the preferred standalone solution. However, the Earth's rotation signal is extremely weak ( $\approx 73 \times 10^{-6}$ rad/sec). While Micro-Electro-Mechanical Systems (MEMS) gyroscopes have reached a sensitivity level capable of detecting this signal, their performance is severely degraded in dynamic underwater conditions. Ocean currents and wave-induced disturbances introduce noise and dynamic motions that obscure the subtle Earth rotation vector, rendering conventional stationary-based gyrocompassing algorithms ineffective for moving or disturbed UUVs.

Methodology
The authors propose a learning-based framework designed to mitigate environmental disturbances and refine disturbed inertial signals to extract the Earth's rotation vector. The methodology integrates theoretical dynamics with deep learning (DL) in a virtual testing environment:

Dynamic Modeling: The paper establishes a 6-DoF rigid body dynamic model for the UUV. Unlike kinematic models, this approach incorporates Newton-Euler equations to account for external forces and moments (torques) caused by ocean currents. The model simplifies the system to focus on rotational coordinates ( $\dot{\nu}_2$ ), treating external moments ( $\tau_2$ ) as forcing terms.
Disturbance Synthesis: To simulate realistic underwater conditions, the authors generate synthetic dynamics by superimposing three distinct disturbance profiles onto real-world stationary sensor data:
- Impulse: Instantaneous perturbations.
- Step: Sudden changes transitioning to steady-state.
- Sinusoidal: Periodic wave-like patterns mimicking ocean currents.
  These disturbances are solved using the system's differential equations, creating a "mixed-mode" dataset that combines real sensor noise with simulated dynamic effects.
Learning Framework: A supervised deep learning model is trained to map the synthesized, disturbed input signals ( $\check{\omega}_b$ ) to the ground truth heading angles ( $\psi$ ). The model learns to distinguish the meaningful Earth rotation signal buried beneath instrumental errors and complex hydrodynamic dynamics. The training objective minimizes the quadratic loss between predicted and actual headings.

Key Contributions

Virtual Testing Ground: The paper introduces a novel pipeline for generating synthetic underwater dynamic data. This allows for the training and evaluation of gyrocompassing algorithms under controlled, variable disturbance intensities without the need for extensive and costly physical sea trials.
Dynamic-Aware Learning: Unlike conventional noise removal methods (e.g., wavelet denoising, Wiener filters) that often assume stationary conditions, the proposed framework explicitly learns to handle time-varying dynamics and non-stationary disturbances.
Integration of Physics and AI: The approach bridges the gap between analytical rigid body dynamics and data-driven machine learning, using physical models to generate training data that respects the laws of motion and hydrodynamics.

Results
Empirical simulations were conducted to assess the framework's adaptability to varying disturbance intensities, quantified by the torque-to-inertia ratio ( $\gamma$ ).

Performance Comparison: The proposed learning model was benchmarked against four state-of-the-art filtering techniques: Wavelet denoising, Wiener filter, Savitzky-Golay (SG) filter, and Finite Impulse Response (FIR) filter.
Error Reduction: The learning-based model demonstrated superior performance across all tested disturbance levels. It achieved Root Mean Squared Error (RMSE) improvements ranging from 30% to 60% compared to the best-performing conventional filters.
Robustness: While conventional filters struggled as the intensity of dynamic disturbances increased (higher $\gamma$ values), the learning model maintained lower error rates, effectively "ignoring" additive dynamics to focus on the underlying rotation signal. In high-disturbance scenarios (e.g., $\gamma = 10$ ), where Signal-to-Noise Ratio (SNR) dropped significantly, the model successfully extracted the heading information where traditional methods failed.

Significance and Claims
The paper claims that its primary contribution is providing a resilient gyrocompassing solution for UUVs operating in dynamic underwater environments. By leveraging recent advancements in machine learning, the framework addresses the limitations of model-based procedures that degrade under ocean currents. The authors posit that this approach redefines the error limits for underwater navigation, making off-the-shelf MEMS gyroscopes competent for precise heading determination even when subjected to significant environmental disturbances. The work serves as a realistic proof of concept, demonstrating that learning-based approaches can effectively mitigate the complex interplay between vehicle dynamics and sensor noise, thereby enabling accurate initial heading alignment without reliance on external magnetic or satellite signals.