Robust Acoustic Imaging Based on Bregman Iteration and Fast Iterative Shrinkage-Thresholding Algorithm

Linsen Huang, Shaoyu Song, Zhongming Xu, Zhifei Zhang, Yansong He2020 · DOI

TL;DR

This paper introduces a novel acoustic imaging (AI) method, BI-AI, that combines Bregman iteration with the Fast Iterative Shrinkage-Thresholding Algorithm (FISTA) to improve the accuracy and resolution of sound source localization, especially in challenging low-frequency scenarios. By reformulating AI as a sparse representation problem and refining the computational grid using wavelets, BI-AI demonstrates superior performance over existing methods in both simulated and experimental settings.

What Problem Does This Paper Address?

Acoustic imaging (AI) is crucial for pinpointing the location and strength of sound sources in various applications, from machinery diagnostics to transportation noise analysis. Existing methods like far-field beamforming (BF) and near-field acoustic holography (NAH) have limitations, particularly at low frequencies where they struggle with poor spatial resolution and inaccurate source strength estimation. While compressive sensing (CS) techniques, especially those using the L1 norm, have shown promise in improving resolution, effectively applying them to acoustic imaging, particularly for separating coherent sources and handling noise, remains an active area of research. This paper aims to address the performance degradation of conventional methods in low-frequency scenarios and improve the overall accuracy and robustness of acoustic imaging.

Methods

The core of the proposed method, BI-AI, lies in reformulating the acoustic imaging problem as an L1-norm minimization problem, leveraging the sparsity of sound sources. This is achieved through:

  • Bregman Iteration (BI): This iterative optimization technique is employed to solve the L1-regularized inverse problem. BI decomposes the complex optimization problem into a sequence of simpler sub-problems, making it computationally more tractable.
  • Fast Iterative Shrinkage-Thresholding Algorithm (FISTA): For each sub-problem generated by Bregman iteration, FISTA is used to efficiently solve the L1-norm minimization. FISTA is a first-order method known for its speed in handling smooth and non-smooth optimization terms.
  • Interpolating Wavelet Method: To enhance performance in the low-frequency range, a wavelet-based method is introduced to refine the computational grid. This approach leverages the localized nature of wavelets in both spatial and wave-number domains to extract essential source information and create a more optimized grid for the BI-AI algorithm. This helps in better representing source distributions and preventing spurious sources.
  • "Warm Start" Initialization: For frequencies above 3000 Hz, the algorithm utilizes the output of conventional beamforming as an initial guess, which is shown to speed up convergence.

Datasets

The paper does not introduce new datasets. Instead, it relies on:

  • Simulated Data: Numerical simulations are used to generate sound fields from single and two coherent monopole sources. These simulations allow for controlled testing and quantitative comparison with other methods.
  • Experimental Data: Real-world acoustic measurements are acquired using a HBK 36-microphone Combo Array. This experimental data is used to validate the performance of the proposed BI-AI method in a practical scenario.

Key Insights and Ideas

  • L1-norm for Sparsity: Framing the acoustic imaging problem as an L1-norm minimization problem is key to achieving sparse source representations, which directly translates to better resolution and separation of coherent sources.
  • Bregman Iteration + FISTA Synergy: The combination of Bregman iteration for problem decomposition and FISTA for efficient sub-problem solving provides a powerful and convergent framework for acoustic imaging.
  • Wavelets for Grid Refinement: Using wavelets to refine the computational grid, especially in low-frequency scenarios, is a novel approach that significantly improves the ability to capture source details and reduce artifacts.
  • "Warm Start" for Efficiency: Leveraging existing beamforming results as an initial guess for higher frequencies demonstrates a practical strategy to accelerate the convergence of the iterative process.
  • Robustness Across Frequencies: BI-AI shows consistent performance across a wide frequency range, outperforming traditional methods that often degrade at lower frequencies.

Key Results

  • Improved Resolution: BI-AI consistently provides sharper and narrower main lobes compared to SOAP and L1-GIB, indicating superior resolution quality.
  • Accurate Strength Estimation: The method demonstrates more accurate estimation of source strength, especially for coherent sources, and is less susceptible to noise contamination compared to other methods.
  • Low-Frequency Performance: BI-AI significantly outperforms conventional methods like CESM and WBH in reconstructing sound fields at low frequencies (e.g., 500 Hz), where traditional methods struggle.
  • Coherent Source Separation: BI-AI effectively separates closely spaced coherent sources, a task that proves challenging for other techniques.
  • Quantitative Performance: Relative errors in sound field reconstruction are generally lower for BI-AI compared to CESM and WBH, often below -40 dB. For example, at 1500 Hz, the relative error for BI-AI was 0.46%, compared to higher errors for other methods.

Limitations and Future Work

  • Computational Cost: The authors acknowledge that the proposed method can be time-consuming, especially in complex environments or for large grids.
  • Retreat Distance Determination: Determining the optimal "retreat distance" (related to the holographic distance) can be challenging.
  • Array Form Influence: The study used a specific 36-channel random array; further investigation into the influence of different array forms on identification accuracy is needed.
  • Low-Frequency Computation Time: The computational time for source identification increases with decreasing frequency due to the growing complexity of the transfer matrix.
  • Source Complexity: The current study focused on simple source models (monopoles). Future work should explore more complex scenarios involving transient, mobile, or rotating sources in three-dimensional space.
  • Optimization: Further optimization of computing time for large grids and adaptation of wavelets for 3D applications are suggested for future research.

Summary updated

BibTeX
@misc{huang2020robust,
  author = {Linsen Huang and Shaoyu Song and Zhongming Xu and Zhifei Zhang and Yansong He},
  title = {{Robust Acoustic Imaging Based on Bregman Iteration and Fast Iterative Shrinkage-Thresholding Algorithm}},
  year = {2020},
  doi = {10.3390/s20247298},
}

Loading related content...

Figure