Integrating modeled environmental variability into neural network training for underwater source localization.

Area of Science:

• Underwater acoustics
• Machine learning applications
• Signal processing

Background:

• Supervised machine learning (ML) is increasingly used in underwater acoustics, particularly for acoustic inversion and source localization.
• Obtaining extensive labeled datasets for underwater source localization is challenging, leading to potential issues like model mismatch in feed-forward neural networks (FNNs).
• Environmental variability between training data and real-world conditions can degrade ML model performance.

Purpose of the Study:

• To investigate the effectiveness of using modeled data for training FNNs in underwater acoustics.
• To enhance the robustness of FNNs for underwater source localization by addressing data limitations.
• To compare the performance of FNNs trained with synthetic data against traditional matched field processing (MFP).

Main Methods:

• Utilizing physical and numerical propagation models as data augmentation tools to create diverse training datasets.
• Training feed-forward neural networks (FNNs) on a combination of real and synthetically generated acoustic data.
• Conducting mismatch tests comparing FNN and matched field processing (MFP) outputs under various environmental conditions.

Main Results:

• FNNs trained with diverse, modeled environments demonstrate increased robustness to environmental mismatches compared to those trained on limited data.
• Networks trained using synthetic data exhibit superior and more consistent localization performance than conventional MFP when environmental variability is considered.
• Systematic analysis confirms that training dataset variability significantly impacts FNN localization accuracy on experimental data.

Conclusions:

• Modeled data is an effective strategy for data augmentation in training FNNs for underwater acoustic applications.
• Incorporating environmental variability into training datasets significantly improves the robustness and performance of ML-based source localization.
• FNNs trained with diverse synthetic data offer a promising alternative to traditional methods like MFP for underwater source localization in complex environments.