Session: 04-06-02 Underwater Vehicles and Subsea Communications II

Submission Number: 176633

Context-Aware Deep Learning for Underwater Acoustic Communication 

Underwater robotics—particularly the use of artificial intelligence in autonomous underwater vehicles (AUVs)—plays a critical role in remotely sensing and exploring the ocean’s vast and economically significant resources. This paper focuses on integrating deep learning techniques into the acoustic communication and positioning systems of AUVs, offering a highly promising subsea technology that can accelerate advancements in the global underwater robotics market. A Context-Aware Neural Network (CANN) is investigated as part of an integrated acoustic communication and sensing framework to optimize AUV navigation. In this model, contextual domain knowledge of the complex and dynamic underwater environment is incorporated into the network’s training process. By jointly leveraging acoustic communication and signal sensing, the proposed approach eliminates the need for the computationally expensive channel estimation typically required in underwater acoustic systems. The objective of this work is to conduct numerical simulations to investigate the effects of chaotic underwater conditions on the accuracy of acoustic communication and signal sensing for subsea technology.

To train the CANN, Quadrature Phase-Shift Keying (QPSK) modulation is employed to embed contextual information within the long short-term memory (LSTM) network architecture. The model learns by extracting discriminative features among different signal targets, which are then used to minimize the loss in predicting the labels of QPSK-modulated signals. Through this context-aware approach, acoustic communication and signal sensing are performed simultaneously, enabling adaptive control of acoustic packet transmission and minimizing interference between transmission and reception. Experimental results using Watermark sea trial data demonstrate that training a context-aware neural network allows accurate identification of the nonlinear mapping between transmitted and received acoustic packets.

The proposed methodology was verified through two experiments. First, we evaluated the accuracy of detecting stealthily-moving underwater vehicles. Tests involving six types of stealthily operating vehicles demonstrated that CANN effectively identifies and tracks such targets, highlighting the potential application in submarine warfares and search-and-rescue missions. In the second experiment, we deployed two AUVs operating simultaneously in a hostile underwater environment. The developed system function effectively as a minimally viable sensing and communication framework, establishing a foundation for a more reliable and efficient subsea technology.

We further assess the CANN's performance in tackling the problem of the hostile underwater environment, such as the chaotic marine dynamics, multipath interference, channel time-variability, and Doppler effects. Numerical experiments involved the decoupling the temporal evolution of acoustic signals from their spectral content using Fourier transform to ensure that the spectral content of the sound is properly extracted. This is done by comparing the temporal learning performance of memory-enhanced neural networks with sequence learning performance, showing that the accuracy of the signal detection and classification increases when spectral context is considered in CANN. Finally, the performance of the proposed model was evaluated against the theoretical error bounds and benchmarked against conventional methods, demonstrating superior adaptability and accuracy under complex acoustic conditions. In addition to simulation-based experiments, a sea trial was considered using the Watermark dataset to validate the model’s practical performance under real oceanic conditions. The results demonstrated that the proposed approach outperforms the classical channel equalizer utilizing complete channel statistics, confirming its robustness for underwater target localization and tracking. Furthermore, the CANN’s experimental outcomes closely matched the theoretical error bounds, indicating reliable real-world applicability and superior adaptability under complex acoustic environments. Although the results of this study are based on numerical simulations, this paper outlines the main concept of context-aware neural networks for integrating acoustic communication and sensing for subsea technology.

Presenting Author: Shahbad Alam College of the North Atlantic

Presenting Author Biography: Shahbad Alam is an AI engineer and an AI instructor at the School of Business & Information Technology, College of the North Atlantic (CNA), Canada. He recently graduated from the Department of Electrical and Computer Engineering, Memorial University, where he wrote an M.Eng thesis on the application of artificial intelligence in underwater robotics. He is currently developing integrated sensing and communication methodologies using context-aware neural networks and physics-informed machine learning for autonomous underwater vehicles.

Authors:

Shahbad Alam College of the North Atlantic
Cheng Li Simon Fraser University