Deep Learning Methods for Underwater Target Feature Extraction and Recognition.

Area of Science:

• Signal Processing
• Machine Learning
• Underwater Acoustics

Background:

• Underwater acoustic signal classification is crucial for signal processing.
• Current feature extraction methods like wavelet transform, Hilbert-Huang transform, and Mel frequency cepstral coefficients have limitations.
• Convolution Neural Networks (CNNs) offer automatic feature extraction but can have limited generalization.

Purpose of the Study:

• To propose a new method for underwater acoustic signal feature extraction and identification.
• To enhance the classification performance of underwater noise data.
• To leverage the strengths of CNNs for feature extraction and ELMs for robust classification.

Main Methods:

• Utilized a deep convolution network (CNN) for automatic and robust feature extraction from underwater acoustic signals.
• Removed the fully connected layers of the CNN to focus solely on feature learning.
• Employed an extreme learning machine (ELM) as the classifier, trained on features extracted by the CNN.

Main Results:

• Achieved a 93.04% recognition rate on a dataset of civil ships.
• Demonstrated significant improvement in recognition rates compared to traditional methods (Mel frequency cepstral coefficients, Hilbert-Huang transform).
• The proposed CNN-ELM hybrid model proved effective for underwater target recognition.

Conclusions:

• The combination of CNN for feature extraction and ELM for classification offers a powerful approach for underwater acoustic signal recognition.
• This method overcomes the generalization limitations of traditional CNNs by using ELM for the classification stage.
• The proposed technique shows great promise for improving the accuracy and efficiency of underwater acoustic signal processing.