Abstract
Passive sonar systems rely on detecting tonal signals from marine vessels amidst significant ambient and hydrodynamic noise. Traditional methods like Discrete Fourier Transform (DFT) and Compressed Sensing (CS) suffer from "basis mismatch" due to frequency discretization. This paper presents a grid-less frequency estimation approach using Atomic Norm Minimization (ANM). By incorporating a noise-suppressing pre-filter and a modified orthonormal basis for multi-channel data, the proposed method significantly enhances detection capacity and accuracy. Experimental results on synthetic and real-world SWellEx-96 data demonstrate that the modified ANM approach achieves up to three times the detection capacity of conventional methods in highly correlated sensor environments.
Introduction
Passive sonar systems identify marine objects by detecting specific sinusoidal frequencies, known as tonal signals, radiated by machinery. A major challenge in this field is the "basis mismatch" error. Traditional spectral estimation techniques like DFT or CS-based methods operate on a discrete frequency grid. When a target's actual frequency falls between grid points, spectral leakage occurs, degrading detection performance. While high-resolution methods like MUSIC or ESPRIT exist, they often require prior knowledge of the signal component count and a large number of samples. This research proposes a continuous-domain frequency estimation technique using Atomic Norm Minimization to overcome these limitations.
Proposed Methodology
The proposed system utilizes a two-stage pipeline designed for robust frequency estimation in noisy underwater environments.
First, raw signals undergo noise suppression via low-pass filtering and decimation. An auto-correlation filter is then applied to enhance periodic tonal signals while suppressing stochastic noise components like hydrodynamic and ambient noise. This significantly improves the Signal-to-Noise Ratio (SNR) before the estimation stage.
Second, the system employs a Modified Atomic Norm Minimization (ANM). Standard ANM treats signals as combinations of "atoms" from a continuous set, formulating estimation as a convex optimization problem (specifically, a Semi-Definite Program). However, standard multi-measurement ANM is less effective when sensor data is highly correlated. The authors address this by performing Singular Value Decomposition (SVD) to extract an orthonormal basis of the signal subspace. Solving the dual problem of ANM using this basis allows for more efficient information extraction from redundant sensor arrays.
Results and Discussion
The method was evaluated using synthetic data and the SWellEx-96 real-world dataset. A key metric was "critical sparsity" — the maximum number of detectable tones with a success ratio greater than 90%.
As shown in the results, the modified ANM approach significantly outperforms the conventional version. For a 5-sensor configuration ($L=5$), the proposed method successfully handled 24 tones, whereas the conventional ANM could only manage 8. In real-world testing with the SWellEx-96 dataset, the proposed method achieved a Mean Square Error (MSE) approximately 1.2 times smaller and a success ratio 2.5 times higher than conventional ANM-based approaches. Furthermore, in low SNR conditions (0 dB), the scheme maintained an MSE 1.5 times lower than standard line spectrum estimation methods.
Conclusion
By moving to a grid-less, continuous frequency domain and optimizing for multi-channel redundancy, this research provides a more robust framework for passive sonar detection. While the computational complexity of solving Semi-Definite Programs remains a challenge for real-time applications, the significant gains in detection capacity and accuracy mark a substantial improvement over traditional discrete-grid methods. Future work will focus on optimizing solvers to enable real-time implementation in embedded sonar systems.
Summary updated