Introduction
Acoustic source imaging is a critical tool in modern engineering, used for applications ranging from urban noise mapping to the diagnosis of industrial machinery faults. These systems typically rely on phased microphone arrays to localize sound sources. However, a persistent challenge in the field is achieving high spatial resolution at low frequencies (e.g., 500–1500 Hz). At these wavelengths, the physical aperture of standard microphone arrays is often insufficient, leading to blurred intensity maps where distinct sound sources appear as a single, ambiguous mass.
This paper introduces Bi-NAM (Bayesian inference based on Non-synchronous Array Measurements), a novel method designed to "sharpen" these acoustic images. By combining synthetic aperture expansion with a robust Bayesian inference framework, Bi-NAM achieves high-resolution beamforming that surpasses the hardware limitations of traditional arrays.
The Resolution Challenge
The primary obstacle to low-frequency imaging is the Point Spread Function (PSF). In an ideal system, the PSF would be a sharp spike (a Dirac delta function), perfectly localizing a source. In practice, limited array size and microphone density cause the PSF to become "fat and smooth," resulting in poor spatial resolution.
To resolve this without building prohibitively large and expensive microphone arrays, researchers have turned to signal processing techniques that can simulate a larger aperture through sequential measurements.
The Bi-NAM Methodology
The Bi-NAM approach consists of three integrated stages:
1. Synthetic Aperture Expansion
Instead of using one massive array, the researchers move a smaller "prototype" array to multiple sequential positions. This process creates a large synthetic aperture with high microphone density, effectively increasing the system's ability to distinguish between closely spaced sources.
2. Matrix Completion via FISTA
Because the measurements at different positions are taken at different times (non-synchronous), the phase relationships between them are initially unknown. This results in a synthetic cross-spectrum matrix (CSM) with missing elements. To solve this, the researchers employ the Fast Iterative Shrinkage Thresholding Algorithm (FISTA) to complete the matrix, reconstructing the missing phase information.
3. Bayesian Inference with Student-t Priors
To solve the resulting inverse problem and localize the sources, the framework uses Joint Maximum A Posterior (JMAP) estimation. A key innovation is the use of the Student-t distribution as a prior for both source strength and background noise.
Unlike standard Gaussian priors, which often produce overly smooth results, the Student-t distribution has "heavy tails" that promote sparsity. This allows the model to accurately identify distinct, sharp sources while adaptively handling various noise profiles, including non-Gaussian Laplace noise.
Experimental Validation and Results
The Bi-NAM method was validated in an anechoic chamber using four loudspeakers acting as monopole sources. A 56-channel planar spiral array was moved through nine positions to create the synthetic aperture.
Key findings include:
- Resolution Breakthrough: At 800 Hz and 1000 Hz, Bi-NAM successfully resolved all four sources, whereas Conventional Beamforming (CBF) produced a single blurred blob.
- Robustness: The method maintained high accuracy even at a Signal-to-Noise Ratio (SNR) of -5 dB under challenging Laplace noise conditions.
- Efficiency: The computational cost remained practical, with images generated in approximately 3 seconds, making it suitable for real-world industrial applications.
Conclusion
Bi-NAM represents a significant step forward in acoustic imaging, providing a flexible and cost-effective way to achieve high-resolution results at low frequencies. By shifting the burden from hardware (larger arrays) to software (Bayesian inference and matrix completion), it opens new possibilities for precise noise localization in complex environments. Future work will look to extend this 2D planar framework into 3D source localization and optimize the movement paths of the prototype arrays.
Summary updated