Deep learning classification for improved bicoherence feature based on cyclic modulation and cross-correlation.

Area of Science:

• Signal Processing
• Machine Learning
• Underwater Acoustics

Background:

• Cyclic modulation spectrum (CMS) is used for hydrophone signal analysis.
• Challenges include spectrum leakage and interference from non-periodic signals, especially at low signal-to-noise ratios (SNRs).
• Deep learning (DL) models require robust features for effective hydrophone signal classification.

Purpose of the Study:

• To develop an improved bicoherence spectrum (IBS) for enhanced hydrophone signal classification.
• To combine IBS with deep learning (DL) models for improved performance.
• To evaluate the effectiveness of IBS in detecting periodic harmonics and suppressing interference.

Main Methods:

• Utilized all-phase fast Fourier transform to mitigate spectrum leakage in CMS.
• Employed cross-correlation and bispectrum analysis to suppress non-periodic line spectra interference.
• Integrated the proposed IBS feature with deep belief network (DBN)-based classifiers (DBN-softmax, DBN-SVM, DBN-Random Forest).

Main Results:

• IBS demonstrated higher-precision detection of periodic harmonics without single-line interference compared to CMS and conventional bispectrum.
• IBS exhibited superior robustness under low SNR conditions.
• DBN-based models utilizing IBS achieved classification accuracy generally exceeding 80% across various experimental scenarios.

Conclusions:

• The improved bicoherence spectrum (IBS) is a robust and effective feature for hydrophone signal classification.
• Combining IBS with deep belief networks (DBNs) significantly enhances classification performance, particularly in challenging low SNR environments.
• The proposed method offers a promising approach for underwater acoustic signal analysis and identification.