Sound source ranging using a feed-forward neural network trained with fitting-based early stopping.

Area of Science:

• Ocean acoustics
• Machine learning
• Signal processing

Background:

• Accurate localization of acoustic sources in ocean environments is crucial for various applications.
• Evaluating the performance of trained neural networks on unseen data, particularly for ranging, presents a significant challenge.
• Traditional methods struggle with determining the ranging accuracy of feed-forward neural networks (FNNs) on unlabeled test data.

Purpose of the Study:

• To introduce a novel method for evaluating and improving the ranging accuracy of FNNs in ocean waveguides.
• To address the challenge of assessing FNN performance on unlabeled test data where the source-receiver distance is unknown.
• To enhance the practical applicability of FNNs for acoustic source ranging tasks.

Main Methods:

• Development of a fitting-based early stopping (FEAST) method tailored for FNNs in acoustic ranging.
• Utilizing FEAST to monitor and minimize the evaluated ranging error on unlabeled test data during FNN training.
• Stopping FNN training when the minimum ranging error on the test data is achieved.

Main Results:

• The FEAST method effectively evaluates the ranging error of FNNs on unlabeled test data.
• Implementing FEAST leads to improved ranging accuracy of FNNs on the test dataset.
• The FEAST method demonstrated successful application on both simulated and experimental acoustic data.

Conclusions:

• FEAST provides a robust mechanism for optimizing FNN training for acoustic source ranging.
• The proposed method enhances the reliability and accuracy of underwater acoustic localization using FNNs.
• FEAST is a valuable technique for improving the performance of machine learning models in complex acoustic environments.