⚡ electrical engineering

Energy-Efficient and Classification-Aided Adaptive Channel Estimation for UAV-FSO Links via Mixture-of-Experts Approach

This paper proposes a Classification-Aided Adaptive Channel Estimation (CA-ACE) framework for UAV-FSO links that utilizes a Mixture-of-Experts approach to dynamically select a single energy-efficient estimator based on real-time channel classification, thereby reducing computational complexity by approximately 98.9% and extending UAV operational lifetime.

Original authors: Yamuna Tumma, Mahesh Miriyala, Jayashree Pradhan, Vinod Kiran Kappala, Goutham Veerapu

Published 2026-10-08

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yamuna Tumma, Mahesh Miriyala, Jayashree Pradhan, Vinod Kiran Kappala, Goutham Veerapu

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

In the high, thin air above our cities and landscapes, a new kind of network is taking shape, one that does not rely on radio waves but on beams of light. These are unmanned aerial vehicles, or drones, equipped with free-space optical communication systems. Instead of sending data through the air as invisible radio signals, these drones shoot invisible laser beams at ground stations or other aircraft. This method offers a massive advantage: it can carry far more data than radio, is immune to the interference that plagues wireless networks, and is inherently secure because the light travels in a straight line that is hard to intercept. However, this high-speed link faces a unique and relentless enemy. The atmosphere is never still. It churns with invisible currents of heat and pressure that bend and distort the laser beam, a phenomenon known as turbulence. Compounding this, the drones themselves are not perfectly steady; they vibrate and wobble as they hover, causing the laser to miss its target slightly. This combination of a shaking platform and a turbulent sky makes the signal flicker and fade, threatening to break the connection entirely. To keep the link alive, the receiver must constantly guess the current state of the channel, a process called channel estimation, to correct the data before it is lost.

The challenge for engineers is that drones have strict limits on their size, weight, and power. They cannot carry heavy computers or large batteries. Traditional methods for guessing the channel state often rely on simple math that fails when the weather gets rough, while the most accurate methods use complex artificial intelligence models that are too heavy and power-hungry for a small drone to run in real time. A team of researchers has now proposed a new way to solve this problem, one that balances the need for extreme accuracy with the need for extreme energy efficiency. They call their system "Classification-Aided Adaptive Channel Estimation." Instead of trying to build one giant, super-smart computer brain to handle every possible condition the drone might face, they built a system that acts more like a skilled dispatcher. This system first takes a quick, low-energy look at the signal to determine exactly what kind of trouble it is facing. Is the air calm, or is it churning with strong turbulence? Is the drone shaking slightly, or is it drifting significantly?

Once the system identifies the specific situation, it does not run a massive calculation. Instead, it instantly switches on a single, tiny, pre-calculated tool designed specifically for that exact scenario. The researchers designed sixteen different tools, or "experts," each one a simple mathematical formula optimized for a specific combination of weather and vibration. When the system detects a mild breeze and a steady drone, it uses the tool for that condition. If the wind picks up and the drone starts to jitter, it switches to a different tool instantly. This approach, known as a Mixture-of-Experts, allows the drone to avoid the heavy energy cost of running a complex artificial intelligence model for every single moment of flight. The researchers tested this idea through detailed computer simulations, modeling a drone flying two kilometers away from a ground station. They found that their new system could track the channel with a level of precision that was previously only possible with the heavy, power-draining models. In fact, the new system reduced the amount of computing work required by nearly 99 percent compared to the complex models, while still achieving the same high quality of data correction.

The results of these simulations show that the new method is not just a theoretical possibility but a practical solution for the future of aerial networks. By breaking the problem down into small, manageable pieces and only using the right tool for the job at hand, the researchers demonstrated that a drone can maintain a reliable, high-speed connection without draining its battery. The system proved capable of keeping the error rate of the data transmission extremely low, even when the atmospheric conditions were severe and the drone was struggling to stay steady. This means that in the future, drones could stay in the air longer, carrying more data for disaster recovery, surveillance, or emergency communications, without needing to land frequently to recharge. The work suggests that the key to efficient communication in the sky is not to build a smarter, heavier brain, but to build a lighter, more adaptable one that knows exactly when to use its strength and when to conserve its energy.

Technical Summary: Energy-Efficient and Classification-Aided Adaptive Channel Estimation for UAV-FSO Links via Mixture-of-Experts Approach

Problem Statement
Unmanned Aerial Vehicle (UAV) Free Space Optical (FSO) communication links face significant challenges due to the Size, Weight, and Power (SWaP) constraints inherent to aerial platforms. While Deep Learning (DL) models, such as Convolutional Neural Networks (CNNs), offer robust channel estimation under atmospheric turbulence and pointing errors, their high computational complexity leads to excessive energy consumption and latency, rapidly draining UAV batteries. Conversely, traditional linear estimators (e.g., Least Squares, MMSE) are energy-efficient but fail to adapt to the non-stationary, composite nature of UAV-FSO channels, which are subject to rapid switching between turbulence regimes and mechanical pointing jitter. The core problem addressed is the trade-off between estimation accuracy and energy efficiency: how to achieve high-precision channel state information (CSI) without the prohibitive computational cost of universal deep learning models.

Methodology: The CA-ACE Framework
The authors propose the Classification-Aided Adaptive Channel Estimation (CA-ACE) framework, which utilizes a Mixture-of-Experts (MoE) architecture to decouple regime identification from coefficient estimation. The system operates in three distinct stages:

Statistical Feature Extraction: Instead of processing raw high-dimensional pilot sequences, the system compresses the received signal into a low-dimensional vector (8 real-valued scalars). This vector consists of higher-order statistical descriptors, including mean, standard deviation, extrema, skewness, kurtosis, average power, and a logarithmic term for dynamic range compression.
Regime Classification (Gating Network): A lightweight Multi-Layer Perceptron (MLP) acts as a "gatekeeper." It classifies the instantaneous channel state into one of 16 unique regimes. These regimes are defined as the Cartesian product of four turbulence levels (No, Weak, Moderate, Strong) and four pointing error scenarios (No PE, Non-Identical Jitter, Identical Jitter, Boresight Error). This classification is performed using a hard decision rule to select a single active path.
Adaptive Estimation (Mixture of Experts): Based on the classification, the system activates only one specific "expert" from a bank of 16 pre-trained linear estimators. These experts are not deep neural networks but specialized linear filters with scalar weights ( $w_k$ ) derived offline via ground-based training. The weight for a specific regime $k$ is calculated to minimize the Mean Square Error (MSE) given the empirical channel variance and noise variance for that specific regime. During operation, the channel estimate is computed as a simple affine transformation ( $\hat{h} = w_{\hat{k}} \cdot \hat{h}_{avg}$ ) without the need for explicit pilot tones during flight.

Key Contributions

Regime-Aware Architecture: The paper introduces a framework that leverages physical statistical properties (skewness and kurtosis) to classify channel regimes, moving away from end-to-end black-box regression toward a "divide-and-conquer" strategy.
SWaP-Optimized Design: By replacing complex deep generative models with a bank of lightweight linear experts controlled by a small MLP, the system achieves DL-level accuracy with minimal latency and power overhead.
Offline Training, Online Efficiency: The computationally intensive training of expert weights is performed offline on the ground. The UAV receiver only executes low-complexity feature extraction, classification, and linear filtering during flight.

Simulation Results
The framework was evaluated against benchmarks including Least Squares (LS), MMSE, LMMSE, and standard CNN estimators under composite fading conditions (varying turbulence and pointing errors) over a 2 km link at 1550 nm.

Estimation Accuracy: The CA-ACE framework achieved a Mean Square Error (MSE) of approximately $10^{-4}$ at 30 dB SNR. This represents an improvement of more than three orders of magnitude over traditional linear estimators (which hit an error floor around $5 \times 10^{-1}$ ) and performance comparable to high-complexity CNNs.
Computational Efficiency: The proposed method requires approximately 5,633 Floating Point Operations (FLOPs) per block. In contrast, a standard CNN requires roughly 548,000 FLOPs. This results in a 98.9% reduction in computational complexity.
Link Reliability: Bit Error Rate (BER) analysis demonstrated that CA-ACE maintains robust link availability even under strong turbulence and pointing errors, achieving a BER of $< 10^{-2}$ at 25 dB SNR.

Significance and Claims
The paper claims that CA-ACE successfully bridges the gap between high estimation fidelity and energy efficiency for UAV-FSO networks. By dynamically routing signals to specialized, low-complexity linear experts based on real-time channel classification, the framework eliminates the need for energy-prohibitive deep learning models on board the UAV. The authors assert that this massive reduction in computational load (approx. 99%) directly translates to extended flight endurance and operational time for SWaP-constrained aerial platforms, enabling reliable high-speed non-terrestrial communications without sacrificing data detection accuracy.

Technical Summary: Energy-Efficient and Classification-Aided Adaptive Channel Estimation for UAV-FSO Links via Mixture-of-Experts Approach

More like this