Joint sparsity models for wideband array processing

Petros T. Boufounos, Paris Smaragdis, Bhiksha RajProceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE 2011 · DOI

TL;DR

This paper introduces a novel approach to localizing and reconstructing multiple broadband sound sources using sensor arrays by reformulating the problem as a "joint sparsity" problem in a combined frequency-space domain. This method leverages compressive sensing techniques to robustly identify source locations and their signals, even when the number of sources exceeds the number of sensors.

What Problem Does This Paper Address?

Traditional array processing for source localization and signal recovery often struggles with broadband signals and the ambiguity that arises when multiple sources are present, especially when the number of sources is large. Narrowband approaches treat each frequency independently, losing valuable information about the broadband nature of the sources. While sparsity-based methods have shown promise in signal processing, their application to broadband array processing, particularly for localizing and recovering multiple sources simultaneously, presents challenges. The core problem is to effectively exploit the shared structure of broadband signals across different frequencies to overcome the inherent ambiguities in array measurements and achieve robust localization and reconstruction.

Methods

The paper's key contribution is the formulation of the broadband array processing problem as a joint/group sparsity problem. Instead of treating each frequency independently, the authors model the broadband signal acquisition as a single system where the spatial sparsity (i.e., the locations of the sources) is common across all frequencies. This means that if a source is present at a particular location, it emits signals at all frequencies, and this spatial support is consistent.

They achieve this by:

  • Frequency-Space Domain Reformulation: The wideband signal acquisition is viewed in a combined frequency-space domain.
  • Joint Sparsity Model: The assumption is that the signal is sparse in the spatial domain (only a few locations have sources) and this sparsity pattern is shared across all frequencies. This is mathematically represented using a mixed $\ell_1/\ell_2$ norm.
  • Compressive Sensing Techniques: They adapt model-based compressive sensing algorithms, specifically a modified version of the CoSaMP (Compressive Sampling Matching Pursuit) algorithm, to solve the resulting joint sparsity optimization problem. The modification involves using a line search for efficiency.

Datasets

Not discussed in the paper. The experiments use simulated data based on a defined scenario.

Key Insights and Ideas

  • Broadband Structure is Key: The core insight is that the shared spatial support of broadband sources across frequencies provides a crucial structural constraint that can resolve ambiguities that plague narrowband approaches.
  • Joint Sparsity Unifies: By framing the problem as joint sparsity, the authors elegantly unify the analysis of multiple sources across the entire bandwidth, rather than treating each frequency bin in isolation.
  • Beyond Narrowband Limitations: The joint sparsity model allows for localization of more sources than sensors, a feat often impossible with traditional narrowband methods.
  • Coherence vs. Joint Sparsity: While coherence analysis is useful for understanding narrowband limitations, joint sparsity offers a more powerful framework for broadband scenarios, where lower frequency information can help resolve ambiguities present at higher frequencies.

Key Results

  • Robust Localization: The proposed joint sparsity approach demonstrates accurate localization of sources, even when the number of sources exceeds the number of sensors (e.g., M=4 sensors, K=8 sources).
  • Improved Performance with Longer Snapshots: Experimental results show that using longer snapshots (more data points per frequency) improves localization accuracy, providing more information and diversity.
  • Reconstruction Limitations: While localization is robust, the paper acknowledges that reconstructing the actual signals of all sources is fundamentally limited by the number of sensors, unless additional assumptions about source structure are made.

Limitations and Future Work

  • Reconstruction Limits: The most significant limitation acknowledged is that the number of sources that can be reconstructed is generally limited by the number of sensors. To overcome this, the authors suggest that incorporating additional assumptions about source structure (e.g., sparsity in a different basis) could enable reconstruction of more sources.
  • Worst-Case vs. Probabilistic Analysis: The paper notes that traditional worst-case analysis for sparsity models is pessimistic for broadband array processing. More optimistic results can be achieved using probabilistic analysis, which the authors suggest as a promising avenue for future research.
  • Model Extensions: The authors hint at future work involving extensions to models that provide additional source structure, which could improve reconstruction capabilities.

Summary updated

Abstract

Recent work has demonstrated the power of sparse models and representations in signal processing applications and has provided the community with computational tools to use it. In this paper we explore the use of sparsity in localization and beamforming when capturing multiple broadband sources using a sensor array. Specifically, we reformulate the wideband signal acquisition as a joint/group sparsity problem in a combined frequency-space domain. Under this formulation the signal is sparse in the spatial domain but has common support in all frequencies. Using techniques from the model-based compressive sensing literature we demonstrate that it is possible to robustly capture, localize and often reconstruct multiple signals present in the scene.
BibTeX
@inproceedings{boufounos2011joint,
  author = {Petros T. Boufounos and Paris Smaragdis and Bhiksha Raj},
  title = {{Joint sparsity models for wideband array processing}},
  booktitle = {Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE},
  year = {2011},
  doi = {10.1117/12.893870},
}
Figure