Transfer learning to account for sound speed changes: Example for source ranging in underwater laboratory tank

Area of Science:

• Underwater acoustics
• Machine learning

Background:

• Accurate source-receiver range prediction is crucial in underwater acoustics.
• Environmental variations, like water temperature, can significantly impact acoustic signal propagation and model performance.

Purpose of the Study:

• To investigate the application of transfer learning (TL) for predicting source-receiver range in a laboratory setting with varying water temperatures.
• To assess TL's ability to improve model generalization under different environmental conditions.

Main Methods:

• Utilized single-hydrophone spectral levels from linear chirps (50-100 kHz) as input data.
• Trained one-dimensional convolutional neural networks on data from room temperature water.
• Applied transfer learning with a small dataset to adapt models for warmer water conditions.

Main Results:

• Models trained in room temperature water exhibited bias when applied to warmer water data.
• Transfer learning significantly improved generalization performance in warmer water conditions.
• Demonstrated the effectiveness of TL in mitigating environmental variability impacts.

Conclusions:

• Transfer learning shows significant potential for enhancing the robustness of acoustic range prediction models.
• TL can adapt pre-trained models to new environmental conditions with minimal data, improving performance.
• This approach offers a viable solution for real-world applications facing dynamic environmental factors.