Multitask convolutional neural network for acoustic localization of a transiting broadband source using a hydrophone...

Area of Science:

• Underwater acoustics
• Signal processing
• Machine learning

Background:

• Accurate localization of underwater acoustic sources is crucial for various applications.
• Traditional methods face challenges in multipath environments and near endfire directions.

Purpose of the Study:

• To develop and evaluate a multitask convolutional neural network (CNN) for real-time motorboat localization.
• To compare the performance of cepstrum-based, correlation-based, and combined CNN approaches.

Main Methods:

• A multitask CNN was trained using cepstrogram and cross-correlogram databases from hydrophone arrays.
• Feature vectors derived from cepstrum and cross-correlation were used as input for CNNs.
• The CNNs predicted instantaneous source range and bearing for unseen motorboat transits.

Main Results:

• CNNs utilizing multi-sensor cepstrum-based feature maps accurately predicted motorboat range and bearing, even near endfire directions.
• Multi-sensor generalized cross-correlation feature maps enabled accurate localization in the presence of multipath arrivals.
• The combined cepstrum-cross correlation CNN demonstrated superior source localization performance compared to individual methods and conventional techniques.

Conclusions:

• Multitask CNNs offer a robust solution for underwater acoustic source localization.
• Combining cepstrum and cross-correlation features enhances localization accuracy in challenging environments.
• The proposed CNN approach outperforms traditional passive ranging methods in multipath conditions.