Deep convolution stack for waveform in underwater acoustic target recognition.

Area of Science:

• Marine acoustics
• Artificial intelligence
• Signal processing

Background:

• Deep learning effectively recognizes underwater acoustic targets from signal waveforms.
• Existing methods using large convolutional kernels create shallow, imbalanced networks, underutilizing deep learning's potential.
• Deep convolution stacks offer flexible, balanced structures but are underexplored in this domain.

Purpose of the Study:

• To introduce a novel deep convolution stack network for underwater acoustic target recognition.
• To propose a multiscale residual unit (MSRU) for constructing effective deep neural networks.
• To enhance the accuracy of classifying underwater acoustic targets.

Main Methods:

• Developed a multiscale residual unit (MSRU) to build a deep convolution stack network.
• Proposed the multiscale residual deep neural network (MSRDN) for underwater acoustic target classification.
• Validated the MSRU within Generative Adversarial Networks and tested MSRDN on real-world acoustic data.

Main Results:

• The MSRDN model achieved a top recognition accuracy of 83.15%.
• This represents a 6.99% improvement over related networks using original signal waveforms.
• It also shows a 4.48% improvement compared to networks using time-frequency representations.

Conclusions:

• The proposed MSRU is effective for building deep convolution stack networks in underwater acoustics.
• MSRDN demonstrates superior performance in underwater acoustic target recognition.
• The deep convolution stack approach offers a promising direction for improving acoustic target classification accuracy.