Blind Source Separation with Compressively Sensed Linear Mixtures

Martin Kleinsteuber, Hao Shen

Abstract

This work studies the problem of simultaneously separating and reconstructing signals from compressively sensed linear mixtures. We assume that all source signals share a common sparse representation basis. The approach combines classical Compressive Sensing (CS) theory with a linear mixing model. It allows the mixtures to be sampled independently of each other. If samples are acquired in the time domain, this means that the sensors need not be synchronized. Since Blind Source Separation (BSS) from a linear mixture is only possible up to permutation and scaling, factoring out these ambiguities leads to a minimization problem on the so-called oblique manifold. We develop a geometric conjugate subgradient method that scales to large systems for solving the problem. Numerical results demonstrate the promising performance of the proposed algorithm compared to several state of the art methods.

Introduction

This paper introduces a framework for the simultaneous separation and reconstruction of signals from compressively sensed linear mixtures. By integrating Compressive Sensing (CS) with Blind Source Separation (BSS) and optimizing over the "oblique manifold," the proposed framework enables the recovery of original signals even when sensors are unsynchronized and sampling at sub-Nyquist rates.

The authors address two primary challenges in signal processing: Blind Source Separation (BSS) and excessive data volume. Standard BSS, often exemplified by the "cocktail party problem," typically requires high-density data sampling. While Compressive Sensing allows for reconstruction from sparse samples, its application to BSS is complicated by the requirement for sensor synchronization in existing methods. This research addresses BSS when mixtures are compressively sampled by independent, uncoordinated sensors.

Methodology

The core of this research is the CS-BSS model, which treats the problem as a joint optimization of the mixing matrix ($A$) and the sparse representation of the signals ($X$). Key technical components include:

  • The Oblique Manifold: To resolve the inherent scaling ambiguity in BSS, the authors constrain the columns of the mixing matrix to a unit norm, defining a geometric space known as the oblique manifold.
  • Riemannian Conjugate Subgradient (CSG) Method: Since the objective function utilizes an $\ell_1$-norm for sparsity, it is non-smooth. The authors adapted a conjugate gradient method to operate with subgradients on Riemannian manifolds.
  • Independent Sampling: The model supports unique sampling matrices ($\Phi_i$) for each mixture, removing the necessity for global sensor synchronization.

Experimental Setup

The algorithm was evaluated using synthetically generated signals. Three source signals of 1,000 samples each were generated with 100 non-zero entries following a Laplacian distribution. These mixtures were compressively sampled down to 300 samples, representing a 70% reduction in data volume. The focus of these experiments was to establish the theoretical framework and validate algorithmic performance against state-of-the-art methods.

Results

The proposed CSG algorithm was compared against several alternatives, including alternating thresholding methods (Alt-IST, Alt-IHT) and smoothing techniques.

SNR performance of various algorithms

As shown in Figure 1, the CSG method achieved significantly higher and more stable Signal-to-Noise Ratios (SNR), often peaking between 40-50 dB. In contrast, the "Alt-CG-CSG" approach performed poorly.

Amari error comparison across algorithms

Separation quality was further assessed using the Amari Error. As illustrated in Figure 2, the CSG method provided consistently lower error rates than alternating methods. Furthermore, the CSG-type algorithms exhibited much smaller variance across 50 experiments, indicating greater robustness to varying initial conditions compared to thresholding-based approaches.

Discussion and Conclusion

The research highlights that recognizing the geometric structure of the mixing matrix on an oblique manifold allows for more stable optimization than standard Euclidean approaches. Furthermore, sparsity serves a dual purpose: it facilitates both signal reconstruction and source separation. This efficiency, combined with the ability to handle independent sampling, offers significant advantages for hardware implementation in sensor networks.

While the current study focuses on the "determined" case (equal sources and sensors), future work may explore underdetermined BSS. Additionally, while the CSG method scales well, the geometric operations involved add mathematical complexity. Future iterations could also incorporate "dictionary learning" to automatically identify the sparse representation basis ($D$) rather than assuming it is known a priori.

Summary updated

Abstract

This work studies the problem of simultaneously separating and reconstructing signals from compressively sensed linear mixtures. We assume that all source signals share a common sparse representation basis. The approach combines classical Compressive Sensing (CS) theory with a linear mixing model. It allows the mixtures to be sampled independently of each other. If samples are acquired in the time domain, this means that the sensors need not be synchronized. Since Blind Source Separation (BSS) from a linear mixture is only possible up to permutation and scaling, factoring out these ambiguities leads to a minimization problem on the so-called oblique manifold. We develop a geometric conjugate subgradient method that scales to large systems for solving the problem. Numerical results demonstrate the promising performance of the proposed algorithm compared to several state of the art methods.
BibTeX
@article{kleinsteuber2012blind,
  author = {Martin Kleinsteuber and Hao Shen},
  title = {{Blind Source Separation with Compressively Sensed Linear Mixtures}},
  journal = {IEEE Signal Processing Letters},
  year = {2012},
  url = {https://openreview.net/forum?id=MHnVSulLIJd},
}
Figure