Underwater Acoustic Communication Channel Modeling Using Reservoir Computing

Oluwaseyi Onasami, Ming Feng, Hao Xu, Mulugeta Haile, Lijun QianIEEE Access 2022 · DOI

TL;DR

This paper introduces a data-driven approach to modeling underwater acoustic (UWA) communication channels using Reservoir Computing (RC), specifically Echo State Networks (ESNs). The proposed ESN models achieve superior accuracy and significantly faster training times compared to traditional deep learning methods, especially in challenging, chaotic UWA environments.

What Problem Does This Paper Address?

Underwater acoustic (UWA) communication is crucial for various applications, but it's notoriously difficult to model due to the complex and unpredictable nature of the underwater environment. Factors like varying sound velocity, rough seabeds, multipath propagation, and ambient noise create significant challenges. Existing physics-based models often rely on simplifying assumptions that don't fully capture real-world complexities, and there's a lack of high-quality, real-world UWA data for training machine learning models. This paper aims to bridge this gap by developing accurate, data-driven UWA channel models using a powerful machine learning technique.

Methods

The core of this research lies in the application of Reservoir Computing (RC), specifically the Echo State Network (ESN), for UWA channel modeling. ESNs are a type of recurrent neural network (RNN) where the recurrent part (the "reservoir") has randomly initialized, fixed weights. This significantly simplifies training, as only the output layer needs to be trained. The authors leverage ESNs' ability to model dynamical systems and predict chaotic behavior.

Key methodological contributions include:

  • Data-Driven Approach: Utilizing real-world UWA data instead of relying solely on physics-based simulations.
  • ESN for UWA Channel Modeling: Applying ESNs to accurately predict the channel's behavior.
  • Transfer Learning Exploration: Investigating how knowledge gained from other domains (like RF communications or simulated UWA data) can be transferred to improve ESN performance.
  • Novel Reservoir Design: Proposing the use of pre-trained deep learning models (DNNs, LSTMs) as the reservoir within an ESN architecture.
  • In-depth Hyperparameter Analysis: Systematically evaluating the impact of various ESN configurations (initialization methods, spectral radius, reservoir size, activation functions, regression methods) on model performance.

Datasets

The paper utilizes real-world UWA communication data collected from two primary sources:

  • Water Tank Experiment: Data collected in a controlled water tank environment with no external disturbance.
  • Lake Tahoe Experiment: Data collected from Lake Tahoe, Nevada, under varying levels of disturbance (none, mild, strong). This dataset is crucial as it captures more realistic, chaotic underwater conditions.

Four distinct categories of data were collected, differing in their channel (water tank vs. lake) and the level of external disturbance. These datasets were used for training and evaluating the developed models.

Key Insights and Ideas

  • ESNs Excel in Chaotic Environments: The most significant insight is that ESNs demonstrate superior performance in modeling chaotic UWA channels compared to traditional deep learning models. This highlights RC's inherent strength in handling complex, dynamic systems.
  • Training Efficiency: ESNs offer a substantial advantage in training time, being orders of magnitude faster than DNNs and LSTMs, making them more practical for real-world applications where rapid model development is needed.
  • Echo State Property is Crucial: The paper emphasizes the importance of satisfying the echo state condition (spectral radius < 1) for ESNs to achieve good performance. Normalizing reservoir matrices to meet this condition significantly improves accuracy.
  • Reservoir Design Matters: While ESNs are known for their random reservoir initialization, the study shows that using pre-trained deep learning models as reservoirs can offer some improvements, though not always outperforming a well-tuned conventional ESN.
  • Transfer Learning Nuances: Transfer learning from vastly different domains (like RF to UWA) yields poor results. However, transferring knowledge from simulated UWA data (like Bellhop) to real-world UWA data shows promise, especially when the pre-trained models are fine-tuned with the target data.

Key Results

  • Superior Accuracy: ESN models consistently outperformed DNN and LSTM models in terms of Mean Absolute Percentage Error (MAPE). Specifically, ESN achieved a 2% improvement over DNN in benign conditions and up to a 40% improvement in chaotic UWA scenarios.
  • Faster Training: ESN training times were dramatically shorter, averaging around 6.7 minutes compared to 2.3 hours for LSTM on the same dataset and hardware.
  • Robustness: ESNs showed robustness across various spectral radii, activation functions, and regression methods, provided the echo state property was maintained.
  • Optimal Reservoir Size: The best performance was observed when the reservoir size matched the input sequence length, suggesting a synergistic relationship between input data structure and reservoir capacity.

Limitations and Future Work

  • Generalization to Ocean Environments: While Lake Tahoe data offers realistic conditions, the authors acknowledge that direct transfer to ocean environments might require modifications due to differences like salinity. Future work aims to refine the model for such environments.
  • Pre-trained Reservoir Performance: Using pre-trained deep learning models as ESN reservoirs showed some improvement but did not consistently outperform conventional ESNs. Further research could explore more sophisticated methods for integrating pre-trained models.
  • Networked UWA Systems: The current model is for a single Tx/Rx pair. Future work includes generalizing the model to handle network settings with multiple transceivers, including non-orthogonal channels and mutual interference.
  • Software-Defined Receiver: The developed ESN model could be adapted for software-based receivers to recover transmitted signals, which is identified as a future research direction.

Summary updated

Abstract

Underwater acoustic (UWA) communications have been widely used but greatly impaired due to the complicated nature of the underwater environment. In order to improve UWA communications, modeling and understanding the UWA channel is indispensable. However, there exist many challenges due to the high uncertainties of the underwater environment and the lack of real-world measurement data. In this work, the capability of reservoir computing and deep learning has been explored for modeling the UWA communication channel accurately using real underwater data collected from a water tank with disturbance and from Lake Tahoe. We leverage the capability of reservoir computing for modeling dynamical systems and provided a data-driven approach to modeling the UWA channel using Echo State Network (ESN). In addition, the potential application of transfer learning to reservoir computing has been examined. Experimental results show that ESN is able to model chaotic UWA channels with better performance compared to popular deep learning models in terms of mean absolute percentage error (MAPE), specifically, ESN has outperformed deep neural network by 2% and as much as 40% in benign and chaotic UWA respectively.
BibTeX
@inproceedings{onasami2022underwater,
  author = {Oluwaseyi Onasami and Ming Feng and Hao Xu and Mulugeta Haile and Lijun Qian},
  title = {{Underwater Acoustic Communication Channel Modeling Using Reservoir Computing}},
  booktitle = {IEEE Access},
  year = {2022},
  doi = {10.1109/access.2022.3177728},
}
Figure