NEC Develops Self-Supervised Underwater Acoustic Foundation Model

NEC Corporation has received a research contract under the Innovative Breakthrough Research program administered by the Defense Innovation Science and Technology Institute of Japan's Agency for Defense Equipment. The contract initiates research on underwater acoustic foundation models using self-supervised learning methods, aiming to complete model construction by fiscal year 2027.

Underwater Data Processing and Foundational Architecture
Because electromagnetic and light waves degrade rapidly in water, underwater monitoring relies heavily on acoustic signals captured through sonar arrays and hydrophone networks. Analyzing these sound waves has historically required manual signal processing by specialized analysts, creating throughput bottlenecks and limiting analytical precision across complex data volumes.

The project applies generative AI principles—analogous to large language models—to underwater acoustics, establishing a foundation model capable of interpreting multi-source acoustic inputs. The architecture is designed to support the construction of ocean digital twins and accelerate target recognition.

Self-Supervised Learning and Sonar Integration
Conventional supervised machine learning architectures require large volumes of human-annotated, labeled training data, which remain scarce in oceanic acoustic environments. To circumvent this limitation, the model uses self-supervised learning to train on vast unlabeled acoustic data sets gathered through public-private partnerships.

NEC is integrating its sonar engineering expertise with large-scale pre-training methodologies developed for its proprietary AI technology, cotomi.

Dual-Use Applications and Project Scope
The initiative targets dual-use deployment across national defense and commercial maritime domains. Defense applications include enhanced target classification and situational awareness, while civilian applications encompass marine life tracking, resource exploration, oceanographic environmental monitoring, and high-precision seismic event analysis.

NEC will conduct testing and validation in collaboration with the Defense Innovation Science and Technology Institute through fiscal year 2027.




Additional Context
This section details technical specifications not included in the original news release.

Underwater acoustic propagation is subject to dynamic transmission losses caused by geometric spreading, frequency-dependent absorption, multi-path boundary reverberations from the sea surface and seabed, and refraction along the deep sound channel (SOFAR channel). In self-supervised learning (SSL) audio frameworks, continuous raw acoustic waveforms recorded by passive hydrophones are converted into two-dimensional time-frequency representations—such as short-time Fourier transform (STFT) spectrograms or log-mel filterbank energies—or encoded directly via one-dimensional convolutional neural networks.

Pre-training employs masked spectrogram modeling or contrastive predictive coding objectives, where portions of the acoustic spectrogram or continuous latent feature frames are masked or corrupted. The neural network (typically a Vision Transformer or Conformer backbone) learns to reconstruct the missing time-frequency patches or predict future latent vectors based on surrounding temporal context without human annotations. This process allows the model to learn representations of hydrodynamic cavitation, mechanical propulsion harmonics, marine mammal vocalizations, and seismic rumble while filtering out ambient background noise. Downstream task-specific classifiers, such as passive sonar target recognition or marine mammal whistle classification, can subsequently be fine-tuned using smaller quantities of labeled target data.

Edited by Romila DSilva, Induportals Editor, with AI assistance.