A joint artificial and semantic feature space mixup for deep-model-based passive underwater acoustic multi-target...

Area of Science:

• Underwater Acoustics
• Signal Processing
• Machine Learning

Background:

• Passive sonar limitations cause overlapping signals, creating multi-target recognition challenges.
• Ocean acoustic complexity leads to significant intra-class diversity and distribution shifts, especially with limited data.
• Nonlinear interactions in underwater channels further complicate multi-target radiated noise analysis.

Purpose of the Study:

• To propose a joint artificial and semantic feature space mixup method for underwater acoustic multi-target recognition.
• To enhance deep network learning of target diversity using limited data.
• To mitigate distribution shifts and nonlinear interactions in complex acoustic environments.

Main Methods:

• A novel mixup strategy is applied across original signal, artificial feature (spectrograms), and semantic feature spaces.
• Multi-target data is constructed across different feature spaces to guide deep network learning.
• Theoretical proofs validate the method's rationality in addressing distribution shifts and nonlinear interactions.

Main Results:

• The proposed joint feature space mixup method demonstrates consistent efficacy.
• The approach effectively guides deep networks to learn target diversity from limited datasets.
• Experimental results confirm the method's robustness across different deep models and artificial features.

Conclusions:

• The joint artificial and semantic feature space mixup is a promising approach for underwater acoustic multi-target recognition.
• This method effectively tackles challenges posed by signal overlap, environmental complexity, and data limitations.
• The technique offers a viable solution for improving the accuracy and robustness of underwater target identification systems.