⚡ electrical engineering

Deep Learning Based Multi-Step Channel Prediction for Adaptive Underwater Acoustic OFDM Systems

This paper proposes an adaptive OFDM framework for underwater acoustic communications that utilizes a Transformer-based PatchCSI-T model for multi-step channel prediction, combined with a greedy resource allocation scheme, to achieve accurate low-latency CSI forecasting and improved bit error rate and spectral efficiency on real-world datasets.

Original authors: Tian Tian, Ying Zhang, Agastya Raj, Fei-Yun Wu, Marco Ruffini

Published 2026-06-04

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Tian Tian, Ying Zhang, Agastya Raj, Fei-Yun Wu, Marco Ruffini

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

Imagine trying to have a conversation with a friend who is swimming deep underwater. The ocean is a tricky place for sound: it bounces off the surface and the bottom, gets distorted by moving water, and travels so slowly that by the time your friend hears you, the water conditions have already changed.

In the world of underwater communication, this "conversation" is a digital signal sent via sound waves. The paper you provided describes a new, smarter way to send these messages so they arrive clearly, even when the ocean is chaotic.

Here is the breakdown of their solution, using simple analogies:

The Problem: The "Outdated Map"

To send a message efficiently, a radio (or in this case, a sound transmitter) needs to know the current state of the "road" (the water channel). This is called Channel State Information (CSI).

The Old Way: Usually, the receiver checks the road, sends a note back to the transmitter saying, "Hey, the road is bumpy here, slow down!" But because sound travels slowly underwater, that note takes seconds to arrive. By the time the transmitter gets the note, the road has changed again. It's like trying to drive using a map that is five minutes old in a city where traffic lights change every ten seconds.
The Consequence: To be safe, systems often just send data very slowly and simply, wasting potential speed.

The Solution: The "Crystal Ball" (PatchCSI-T)

The authors built a system that doesn't wait for a note from the receiver. Instead, it uses a Deep Learning AI (called PatchCSI-T) to act as a crystal ball. It looks at the recent history of the water's behavior and predicts what the road will look like in the future.

Think of it like a weather forecaster. Instead of just telling you it's raining now, it predicts it will rain in 30 minutes, so you can grab your umbrella before you step outside.

How the AI works (The "Patch" Trick):

Breaking it into chunks: Instead of looking at every single second of data individually (which is slow and confusing), the AI breaks the history into small overlapping "patches" or chunks, like looking at a movie frame-by-frame rather than pixel-by-pixel.
Learning patterns: It uses a special type of AI (a Transformer, similar to the tech behind modern chatbots) to learn how these chunks relate to each other.
No "Feature Mixing": Older AI models tried to mix all the different data points together, which often led to "overfitting" (memorizing the training data too well and failing in real life). This new model treats each data stream independently but learns from them all at once. It's like having a team of specialists who all use the same rulebook but focus on their own specific job, rather than one generalist trying to do everything at once.

The Strategy: The "Smart Driver" (Adaptive OFDM)

Once the AI predicts the future road conditions, the transmitter uses a Greedy Algorithm to decide how to drive.

The Car: The data is split into many small lanes (subcarriers).
The Decision:
- If the AI predicts a lane is smooth and clear, the system loads it up with a high-speed, complex message (like a sports car going fast).
- If the AI predicts a lane is bumpy or foggy, it slows down or switches to a simpler, more robust message (like a truck driving slowly).
- If a lane is too dangerous, it turns it off completely and uses that fuel (power) to boost the other lanes.

This happens in real-time, ensuring the message gets through with the fewest errors possible.

The Results: Faster and Smoother

The team tested this on real underwater data collected from a reservoir in China.

Prediction Accuracy: The AI predicted the future water conditions much better than older methods. It was significantly more accurate (about 1.4 to 4.9 dB better in technical terms) and did it much faster.
Speed: It was up to 50 times faster at making predictions than some of the previous top models.
Real-World Performance: When they put this AI into the communication system, they could send more data per second (higher spectral efficiency) while keeping the error rate very low. At a specific signal strength, they achieved a speed of 5.14 bits per second per Hertz, with nearly 90% of the messages arriving perfectly.

In Summary

The paper presents a system that stops underwater communication from reacting to the past and starts it anticipating the future. By using a smart AI to predict how the water will behave and a flexible strategy to adjust the message speed accordingly, they can send data faster and more reliably through the noisy, slow, and changing underwater environment.

Technical Summary: Deep Learning Based Multi-Step Channel Prediction for Adaptive Underwater Acoustic OFDM Systems

Problem Statement
Underwater acoustic (UWA) communications face significant challenges due to frequency-dependent attenuation, severe time-varying multipath fading, motion-induced Doppler shifting, and long propagation delays caused by low sound speed. These factors make timely acquisition of Channel State Information (CSI) difficult; round-trip delays over kilometer-scale links can reach seconds, rendering traditional feedback-based link adaptation impractical when the channel coherence time is of a similar order. While channel prediction offers a solution by enabling proactive link adaptation, existing methods have limitations. Conventional adaptive filtering (e.g., Recursive Least Squares) struggles with the nonlinear and rapidly time-varying nature of UWA channels. Furthermore, existing Deep Learning (DL) approaches, such as RNN-based architectures, suffer from error accumulation in multi-step prediction, while models employing cross-dimension feature mixing are prone to overfitting given the scarcity of at-sea measurement data.

Methodology
The authors propose an adaptive Orthogonal Frequency-Division Multiplexing (OFDM) framework comprising two primary components:

PatchCSI-T (Channel Prediction Model):
- Architecture: A Transformer-based model inspired by the time-series forecasting model PatchTST.
- Patching: To reduce computational complexity and capture segment-level dependencies, input CSI sequences are partitioned into overlapping patches rather than processing individual time samples. This reduces the self-attention complexity from $O(T_{win}^2)$ to $O(N_{patch}^2)$ .
- Feature-Independent Modeling: Unlike many prior DL methods that mix features (e.g., across subcarriers or taps) to capture correlations, PatchCSI-T treats features independently. The feature dimension is merged into the batch dimension, allowing all features to be processed in parallel with shared model parameters. This approach aims to mitigate overfitting, particularly when training data is limited, and avoids the complexity of explicit cross-feature modeling.
- Joint Multi-Step Prediction: The model generates the entire future prediction sequence ( $T_{pred}$ steps) in a single forward pass, avoiding the error propagation inherent in autoregressive recursive methods.
Greedy Adaptive Modulation and Power Allocation:
- The system utilizes the predicted CSI to jointly select modulation orders and transmit power per subcarrier.
- A closed-form minimum power threshold is derived based on the predicted channel gain and prediction error variance (modeled as Gaussian) to satisfy a target Bit Error Rate (BER).
- A low-complexity greedy algorithm iteratively upgrades the modulation order of the subcarrier requiring the smallest incremental power until the total power budget is exhausted or the target BER constraint is violated. If channel conditions deteriorate, the algorithm downgrades subcarriers with the largest marginal cost.

Key Contributions

PatchCSI-T Model: Introduction of a Transformer-based architecture specifically adapted for UWA channels using patching and feature-independent modeling to balance accuracy, inference speed, and robustness against overfitting.
Integrated Framework: Development of a complete adaptive OFDM system that couples the multi-step prediction model with a greedy resource allocation scheme.
Empirical Validation: Evaluation on real-world, field-measured UWA channel datasets (China-Wanlu Reservoir), demonstrating improvements over state-of-the-art baselines.

Experimental Results
Experiments were conducted using the China-Wanlu Reservoir dataset (2,137 CIR measurements, 256 subcarriers, 6 kHz bandwidth). The PatchCSI-T model was compared against RLS, BiGRU, CNN1d-LSTM, and MTL-LSTM.

Prediction Accuracy: At a prediction horizon of $T_{pred} = 32$ , PatchCSI-T achieved a Normalized Mean-Square Error (NMSE) of -14.76 dB. This represents a gain of 1.4 to 4.9 dB over the baselines. Notably, CNN1d-LSTM and BiGRU exhibited NMSE saturation near -10 dB for larger horizons.
Inference Efficiency: PatchCSI-T demonstrated significantly faster inference times (14.7 ms at $T_{pred}=32$ ), which is approximately 6 times faster than CNN1d-LSTM/BiGRU and more than 50 times faster than MTL-LSTM.
System Performance: When integrated into the adaptive OFDM system at an SNR of 15 dB:
- PatchCSI-T maintained 89.7% of transmitted symbols with a BER $\le 10^{-3}$ .
- It achieved a spectral efficiency of 5.14 bps/Hz.
- This outperformed RLS (72.9% symbols, 4.15 bps/Hz) and MTL-LSTM (75.8% symbols, 4.34 bps/Hz).
Adaptation Behavior: The greedy allocation algorithm successfully tracked channel variations, activating more subcarriers and selecting higher-order modulations during channel gains, while reducing resources during fades to maintain the target BER.

Significance
The paper claims that the proposed framework addresses the critical latency and accuracy trade-offs in UWA communications. By leveraging feature-independent modeling and patching, PatchCSI-T effectively mitigates overfitting and error accumulation, enabling accurate, low-latency CSI forecasting. The integration of this predictor with a greedy allocation scheme allows adaptive systems to operate on predicted rather than outdated channel conditions, thereby improving end-to-end Bit Error Rate (BER) and spectral efficiency in real-world UWA environments. The authors note that future work will extend validation to at-sea data with stronger Doppler effects and longer delays, as well as real-time closed-loop experiments.