Underwater Target Recognition Based on Line Spectrum and Support Vector Machine

Jian Liu, He Yang, Zhong Liu, Ying XiongAdvances 2014 · DOI

Abstract

To increase the classification of underwater target recognition by ship radiated noise signals, the feature extraction method of line spectrum and classification algorithm of support vector machine (SVM) are adopted in the paper. The basic principles of power spectrum and SVM are briefly introduced at first. Then the signals of ship radiated are processed by the power spectrum estimation and the nonlinear least-square polynomial to fitting the continuous spectrum. The ultimate line spectrum is obtained through selecting peak and removes side lobe, and then extracts the feature of number and intensity of line spectrum from it. At last, the characteristic vector is established with normalization process and the signals are classified by SVM with parameter optimization. The simulation experiment results demonstrate that the ship radiated signals can be classified or recognized availably based on line spectrum feature and SVM. This paper presents a method of underwater target recognition and provides guidance for the research on multi-classes recognition algorithms.

Introduction

Identifying specific vessels in the ocean's complex acoustic environment is a critical task for maritime security and research. Every ship produces a unique "acoustic fingerprint" composed of broadband continuous noise and discrete line spectra. While continuous noise arises from hydrodynamic flow over the hull, line spectra are generated by periodic machinery vibrations, such as those from engines and propellers. These line spectra are highly stable and characteristic of specific ship types, making them ideal features for target recognition.

Feature Extraction Methodology

The proposed recognition system relies on isolating these stable line spectra from the masking background noise. The process begins with power spectrum estimation using a standard Fast Fourier Transform (FFT) periodogram to convert time-domain noise into the frequency domain.

Continuous Spectrum Fitting

A crucial step in this pipeline is modeling the continuous spectrum to separate it from the line components. The researchers employ a nonlinear least-square polynomial fitting method. By applying a 3rd to 5th-order polynomial, a smooth curve is generated that represents the broadband background noise.

Continuous spectrum fitting using nonlinear least-square polynomial to isolate background noise.

Line Spectrum Extraction and Vectorization

Once the continuous spectrum is modeled, it is subtracted from the original power spectrum, effectively "flattening" the data. A peak-selection algorithm then identifies the true line spectra while eliminating side lobes. To create a feature vector suitable for machine learning, the frequency range is divided into $K$ intervals. For each interval, the number of line spectra and their cumulative intensity are recorded, resulting in a $2K$-dimensional characteristic vector.

Classification and Results

The study utilized measured ship-radiated noise sampled at 22.05 kHz. Two distinct ship types were analyzed, providing a dataset for binary classification.

Power spectra of two different ship types (Type 1 and Type 2) highlighting the distinct distribution of line spectra.

The extracted feature vectors were fed into a Support Vector Machine (SVM) utilizing a Radial Basis Function (RBF) kernel. SVM was chosen for its robustness with small sample sizes and its ability to handle non-linear decision boundaries.

Experimental results showed that the choice of polynomial order and frequency intervals significantly impacts performance. An initial baseline SVM achieved 82.4% accuracy. However, by optimizing the SVM parameters (specifically the penalty factor $C$ and kernel parameter $\gamma$), the recognition accuracy reached a peak of 94.9%. This performance outperformed a hybrid KNN-SVM approach (84.7%), validating the effectiveness of the line spectrum feature extraction combined with optimized SVM classification.

Conclusion

This research demonstrates that modeling and subtracting the continuous background noise allows for the reliable extraction of ship "fingerprints" via line spectra. The high accuracy achieved with optimized SVM suggests that this method is a viable path forward for underwater target recognition. Future work will focus on expanding the system to multi-class recognition and improving real-time processing capabilities for sonar applications.

Summary updated

Abstract

To increase the classification of underwater target recognition by ship radiated noise signals, the feature extraction method of line spectrum and classification algorithm of support vector machine (SVM) are adopted in the paper. The basic principles of power spectrum and SVM are briefly introduced at first. Then the signals of ship radiated are processed by the power spectrum estimation and the nonlinear least-square polynomial to fitting the continuous spectrum. The ultimate line spectrum is obtained through selecting peak and removes side lobe, and then extracts the feature of number and intensity of line spectrum from it. At last, the characteristic vector is established with normalization process and the signals are classified by SVM with parameter optimization. The simulation experiment results demonstrate that the ship radiated signals can be classified or recognized availably based on line spectrum feature and SVM. This paper presents a method of underwater target recognition and provides guidance for the research on multi-classes recognition algorithms.
BibTeX
@inproceedings{liu2014underwater,
  author = {Jian Liu and He Yang and Zhong Liu and Ying Xiong},
  title = {{Underwater Target Recognition Based on Line Spectrum and Support Vector Machine}},
  booktitle = {Advances},
  year = {2014},
  doi = {10.2991/mce-14.2014.17},
}
Figure