Self-supervised acoustic representation learning via acoustic-embedding memory unit modified space autoencoder for...

Area of Science:

• Underwater acoustics
• Machine learning
• Signal processing

Background:

• High-quality sonar data annotation is costly.
• Single acoustic features lack generalization in ocean environments.
• Need for robust underwater target recognition methods.

Purpose of the Study:

• To propose a self-supervised acoustic representation learning method for underwater target recognition.
• To enhance feature generalization and anti-noise robustness.
• To address limitations of existing acoustic features.

Main Methods:

• Developed a space autoencoder (SAE) merging Mel filter-bank (FBank) and gammatone filter-bank (GBank) into SAE spectrogram (SAE Spec).
• Introduced an acoustic-embedding memory unit (AEMU) for adversarial enhancement.
• Utilized an improved contrastive loss function with negative samples for ASAE spectrogram (ASAE Spec) generation.

Main Results:

• ASAE Spec demonstrated over 0.96% improvement in accuracy compared to mainstream acoustic features.
• ASAE Spec showed enhanced convergence rate and anti-noise robustness.
• Experiments conducted on two underwater datasets validated the proposed method.

Conclusions:

• The proposed ASAE method offers significant improvements in underwater target recognition.
• ASAE Spec exhibits strong potential for practical applications in marine acoustics.
• Self-supervised learning with adversarial enhancement is effective for acoustic feature representation.