⚡ electrical engineering

Vision-Intelligence-Enabled Beam Tracking for Cross-Interface Optical Wireless Communication between Underwater and Low-Altitude Platforms

This paper addresses the challenge of beam misalignment in underwater-to-low-altitude optical wireless communication caused by dynamic sea surface refraction by proposing a vision-intelligence-enabled tracking algorithm that combines convolutional neural networks, bidirectional LSTMs, and attention mechanisms to achieve robust real-time transceiver alignment.

Original authors: Jiayue Liu, Tianqi Mao, Leyu Cao, Weijie Liu, Dezhi Zheng, Julian Cheng, Zhaocheng Wang

Published 2026-05-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jiayue Liu, Tianqi Mao, Leyu Cao, Weijie Liu, Dezhi Zheng, Julian Cheng, Zhaocheng Wang

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 Picture: Trying to Talk Through a Wobbly Window

Imagine you are underwater in a submarine (or a robot drone), and you need to send a massive amount of data—like high-definition video or sensor readings—to a drone flying in the sky.

Normally, underwater robots use sound (like sonar) to talk, but that's slow and narrow, like trying to send a whole movie through a tiny straw. The paper suggests using light instead, which is like sending that movie through a wide-open highway. This is called "Optical Wireless Communication."

The Problem:
The trouble is the surface of the ocean. It's not a flat, calm mirror; it's a wobbly, moving wall of water.

The Analogy: Imagine trying to shine a laser pointer from the bottom of a swimming pool to a friend standing on the deck. If the water is perfectly still, the beam goes straight up. But if the water is choppy with waves, the surface acts like a wobbly, curved mirror. It bends (refracts) the light in crazy directions.
The Result: By the time the light hits the air, it's pointing in the wrong direction. The drone in the sky misses the beam, and the connection breaks. The paper calls this "beam misalignment."

The Solution: Giving the System "Eyes" and a "Brain"

The authors propose a new way to fix this. Instead of just guessing where to point the laser, they give the system vision and artificial intelligence (AI) to track the beam in real-time.

1. The "Wobbly Window" Model (The Math Part)

First, the team built a mathematical model of the ocean.

The Analogy: Think of this like a weather forecast for the ocean surface. They used complex math (wave spectrum theory) to predict exactly how the waves will look and how they will bend the light at any given second. They calculated the "shortest path" the light should take, even when the water is moving.

2. The "Virtual Camera" (The Simulation)

To train their AI, they didn't just use real cameras; they built a virtual one.

The Analogy: Imagine a video game where you can see exactly how light bounces off every single wave. They used a "Ray Tracing" algorithm (the same tech used in high-end video games) to simulate what a camera would see if it were looking at the laser beam through choppy water. This created a library of thousands of "what-if" scenarios with different waves and light distortions.

3. The "Super-Brain" (The AI Algorithm)

This is the core of their invention. They created a special AI brain to look at the camera images and tell the laser where to point.

The Analogy: Imagine a goalkeeper trying to catch a ball.
- Old Methods: The goalkeeper just looks at where the ball is right now and jumps. If the ball is moving fast or the wind is weird, they miss.
- This Paper's Method: The goalkeeper has a "Super-Brain" that does three things:
  1. CNN (The Eyes): It looks at the picture and spots the blurry, distorted light beam, ignoring the "noise" (like sun glare or water bubbles).
  2. Bi-LSTM (The Memory): It remembers what the beam was doing a split second ago and predicts where it will be a split second in the future. It understands the flow of the waves, not just a single snapshot.
  3. Attention Mechanism (The Focus): This is the secret sauce. It's like the goalkeeper squinting to focus only on the most important part of the image (the actual beam) and ignoring the distracting background noise.

The Results: Catching the Ball Every Time

The authors ran simulations to see if their "Super-Brain" worked better than other methods.

The Test: They simulated a stormy ocean and asked different algorithms to keep the laser beam locked onto the drone.
The Winner: Their AI method (CNN + Bi-LSTM + Attention) kept the signal strong and steady much better than the old methods.
- The Analogy: While other methods were like a shaky hand trying to hold a flashlight on a moving target, the new AI method was like a steady, robotic arm that anticipated the target's movement perfectly.
The Proof: The new method received a stronger signal (less data loss) and made fewer errors (lower "Bit Error Rate") even when the ocean was rough and the "vision" was noisy.

Summary

In short, this paper solves the problem of "shaky connections" between underwater and sky drones. They did this by:

Modeling how ocean waves bend light.
Creating a virtual training ground for an AI.
Building an AI that uses memory (to predict movement) and focus (to ignore noise) to steer a laser beam perfectly through a wobbly ocean surface.

The result is a much more reliable way to send data from the deep ocean to the sky.

Technical Summary: Vision-Intelligence-Enabled Beam Tracking for Cross-Interface Optical Wireless Communication

Problem Statement
The rapid expansion of oceanic applications, such as underwater surveillance and mineral exploration, necessitates real-time wireless backhaul for massive observational data. Traditional narrowband acoustic approaches are insufficient for these high-volume demands. While Optical Wireless Communication (OWC) between underwater platforms (e.g., UUVs) and Low-Altitude Platforms (LAPs, e.g., UAVs) offers a promising broadband alternative, its practical implementation faces significant challenges. The primary obstacle is beam misalignment caused by the dynamic fluctuations of the water-air interface. These fluctuations induce complex refraction and path perturbations, leading to severe signal attenuation and instability. Existing solutions, such as wide-beam transmitters or traditional mechanical alignment, often lack the precision to adapt to rapid oceanic dynamics or fail to effectively sense and predict interface variations for real-time adjustment.

Methodology
To address these challenges, the authors propose a comprehensive framework combining a rigorous mathematical channel model with a vision-based deep learning algorithm.

Mathematical Channel Modeling:
- Ocean Wave Model: The study models the water-air interface using ocean-wave spectrum theory, constructing a shape function $W(x, y, t)$ based on harmonic wave components to simulate dynamic sea surfaces.
- Optical Path Derivation: The propagation trajectory is determined using the Minimum Optical Path Length (OPL) theory. The refraction point on the dynamic interface is calculated by minimizing the OPL ( $n_1d_w + n_2d_a$ ) subject to the surface geometry constraints.
- Path Gain Model: The overall channel gain is formulated by integrating directional attenuation at the transmitter (Laser Diode) and receiver (Avalanche Photodiode), distance and media attenuation (following Beer-Lambert law), and refraction attenuation derived from Fresnel's equations.
- Vision Simulation: A Ray-Tracing (RT) algorithm is employed to simulate the visual data perceived by photosensitive screens at the transceivers. This algorithm backtracks rays from pixel positions through the dynamic interface to the light source, computing intensity patterns that serve as the input for the AI system.
Vision-Intelligence Beam Tracking Algorithm:
- Architecture: The core of the proposed solution is a deep learning network combining Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory (Bi-LSTM) layers, enhanced with an attention mechanism.
- Feature Extraction: The CNN component extracts spatial features from the raw visual input (grayscale images of the beacon light), reducing dimensionality while increasing feature depth.
- Temporal Processing: The Bi-LSTM structure processes the sequence of visual frames to capture temporal dependencies, considering both past and future contexts to predict the optimal beam direction.
- Attention Mechanism: An attention module assigns importance weights to different time steps, allowing the network to focus on critical directional information and suppress noise.
- Output: The network outputs a 3D directional vector to guide the rotation motors of the transceivers, ensuring alignment with the minimum OPL path.
- Training: The model is trained using a combined loss function (Mean Squared Error and L1 loss) with data augmentation techniques (random rotation, translation, Gaussian noise) to enhance generalization.

Key Contributions
The paper highlights three primary contributions:

Channel Modeling: The establishment of a water-air interface refraction model based on wave spectrum theory and the demonstration of optical path uniqueness using minimum OPL theory, enabling a complete channel model that incorporates refraction and attenuation.
Vision Modeling: The proposal of a ray-tracing-based method to construct a vision model for directional sensing, where pixel intensities are computed by accumulating light contributions from sources and refractive interfaces during backtracking.
Algorithm Development: The development of a deep learning-based beam tracking algorithm that integrates CNN, Bi-LSTM, and an attention mechanism. This approach is designed to extract spatio-temporal features from vision data, demonstrating superior performance in complex environments compared to classical methods.

Results
Numerical simulations were conducted in a simulated oceanic environment to validate the proposed algorithm against benchmarks, including Mean Shift, 3D-CNN, and CNN-LSTM without attention.

Received Signal Strength (RSS): The proposed algorithm achieved an average RSS 15.7% higher than Mean Shift and 32.5% higher than 3D-CNN. It also demonstrated significantly lower RSS variance ( $1.4 \times 10^{-5}$ ), indicating greater stability.
Alignment Accuracy: The algorithm exhibited a lower average angle difference from the theoretical minimum OPL path, particularly under high vision noise conditions (low Signal-to-Noise Ratio of the beacon light).
Noise Resilience: The attention mechanism allowed for clearer discrimination of effective signal directions from noise, resulting in lower Bit Error Rates (BER) compared to the counterparts.
Convergence: Training history showed that the proposed model converged with lower validation loss than the 3D-CNN and standard CNN-LSTM architectures.

Significance
The paper posits that its work addresses the critical gap in systematic application of advanced AI methods to cross-interface communication technologies. By moving beyond traditional sensing and reactive alignment, the proposed vision-intelligence framework enables proactive, real-time adaptation to the complex dynamics of the ocean surface. The results suggest that this approach significantly enhances the reliability and stability of water-air OWC links, making it a viable solution for high-data-rate maritime communication networks where maintaining a line-of-sight optical path is otherwise compromised by environmental fluctuations.