Self-supervised learning-based underwater acoustical signal classification via mask modeling.

Area of Science:

• Signal Processing
• Machine Learning
• Acoustics

Background:

• Underwater acoustic signal classification is vital for military and civilian applications.
• Deep neural networks excel at classification, but signal representation is underexplored.
• Large-scale dataset annotation for deep learning is costly and difficult.

Purpose of the Study:

• To develop a novel self-supervised representation learning method for underwater acoustic signal classification.
• To address challenges in signal representation and data annotation for deep learning models.
• To improve classification accuracy and robustness in various acoustic environments.

Main Methods:

• A two-stage approach: pretext learning with unlabeled data and fine-tuning with limited labeled data.
• Utilizing Swin Transformer architecture for reconstructing masked log Mel spectrograms.
• Self-supervised learning to extract general acoustic signal representations.

Main Results:

• Achieved 80.22% classification accuracy on the DeepShip dataset, surpassing previous methods.
• Demonstrated strong performance in low signal-to-noise ratio (SNR) conditions.
• Showcased effectiveness in few-shot learning scenarios.

Conclusions:

• The proposed self-supervised method offers an effective solution for underwater acoustic signal classification.
• The approach reduces reliance on large labeled datasets, mitigating annotation costs.
• The method provides robust classification performance, adaptable to diverse and challenging acoustic environments.