Detection of tonal frequency of underwater radiated noise via atomic norm minimization

Junhan Kim, Jin‐Hong Kim, Byonghyo Shim, Jungpyo Hong, Seongil Kim, Wooyoung HongThe Journal of the Acoustical Society of Korea 2019 · DOI

Introduction

Identifying the radiated noise of underwater vessels, such as ships and submarines, is a cornerstone of maritime surveillance and target classification. This noise often contains "tonal" components—discrete frequencies generated by mechanical systems like engines and gearboxes. Detecting these signals is challenging due to the high background noise of the ocean and the limitations of traditional signal processing techniques.

Traditional methods like the Discrete Fourier Transform (DFT) suffer from resolution issues over short observation windows. Subspace methods, such as MUSIC and ESPRIT, require prior knowledge of the number of signal components, which is rarely available in practice. While Compressive Sensing (CS) offers a sparse solution, it relies on a discrete frequency grid. When a vessel's actual frequency falls between grid points, a "basis mismatch" occurs, leading to significant estimation errors. This paper proposes a continuous-domain approach using Atomic Norm Minimization (ANM) to eliminate these grid-related inaccuracies.

Methodology

The researchers developed a robust pipeline designed to handle the high-dimensional nature of sonar data while maintaining the accuracy of continuous-domain optimization.

Preprocessing

To make the optimization feasible, the raw signal $s(t)$ undergoes three critical steps:

  1. Low-pass Filtering: Isolates the low-frequency band where tonal components are typically found.
  2. Decimation: Reduces the data volume, which is essential because the subsequent optimization solvers are computationally intensive.
  3. Auto-correlation: Enhances the periodic tonal signals relative to the stochastic background noise.

The ANM Framework

Instead of searching a predefined grid, the ANM approach treats the signal as a combination of "atoms" from a continuous dictionary of complex exponentials. The problem is formulated as a convex optimization task—specifically, a Semi-Definite Programming (SDP) problem. This framework finds the sparsest set of continuous frequencies that best fits the observed data, inherently performing denoising in the process.

The proposed ANM-based frequency estimation pipeline, including preprocessing and optimization stages.

Experimental Results

The proposed method was evaluated using simulated data modeled after underwater acoustic environments, featuring 10 random tonal frequencies under 500 Hz and varying levels of white Gaussian noise (SNR from -10 dB to 10 dB). It was benchmarked against industry standards: root-MUSIC, ESPRIT, and Matrix Pencil.

Performance Metrics

  • Exact Recovery Ratio (ERR): The ANM method demonstrated superior reliability in low-SNR conditions. At -8 dB SNR, the proposed method achieved an ERR of approximately 0.5, whereas traditional subspace methods fell below 0.2.
  • Mean Square Error (MSE): The method consistently maintained a lower MSE across the entire SNR range, indicating higher precision in frequency estimation compared to root-MUSIC and ESPRIT.
Performance comparison across SNR levels: (a) Exact Recovery Ratio and (b) Mean Square Error.

Conclusion

By moving frequency estimation into the continuous domain, the ANM-based approach successfully bypasses the basis mismatch errors that plague discrete methods. While the computational complexity of SDP solvers remains a challenge for real-time deployment, the significant gains in detection accuracy and noise robustness provide a strong foundation for next-generation sonar signal processing. Future research will focus on developing faster, optimization-free continuous domain techniques to bridge the gap between high accuracy and real-time performance.

Summary updated

BibTeX
@article{kim2019detection,
  author = {Junhan Kim and Jin‐Hong Kim and Byonghyo Shim and Jungpyo Hong and Seongil Kim and Wooyoung Hong},
  title = {{Detection of tonal frequency of underwater radiated noise via atomic norm minimization}},
  journal = {The Journal of the Acoustical Society of Korea},
  year = {2019},
  doi = {10.7776/ask.2019.38.5.543},
}
Figure