Blind Source Separation Research Based on the Feature Distance Using Evolutionary Algorithms

Yang Yang, Xiuqin Wang, Di ZhangThe International Journal of Acoustics and Vibration 2014 · DOI

Abstract

Without any information on the mixing system, the blind source separation (BSS) technique efficiently separates mixed signals. The approach called evolutionary algorithms was used for the BSS problem in this paper. A fitness function based on the feature distance and kurtosis was proposed to measure the degree of the separated signals. Compared with traditional algorithms in the BSS problem, both the mathematical calculation and the physical significance of the separated signals are taken into consideration in the proposed method. Therefore, the separated signals could have great correlation with the original individual signal and could be used in additional signal processing steps with good signal property. Experimental results on mixed spoken signals indicated that the established evolutionary algorithm of particle swarm optimization (PSO) and genetic algorithm (GA) could effectively solve the BSS problem from the signal feature distance and independence measurement. The study in this paper was implemented with MATLAB language.

Introduction

The "Cocktail Party Problem" remains a fundamental challenge in signal processing: how can an automated system isolate a single voice from a cacophony of overlapping sounds without prior knowledge of the sources or the mixing environment? This task, known as Blind Source Separation (BSS), is critical for applications ranging from telecommunications to medical imaging.

Traditionally, BSS has relied on Independent Component Analysis (ICA) using gradient-based optimization. However, these methods often suffer from two primary drawbacks: they frequently converge on "local optima" (suboptimal solutions) and they prioritize mathematical independence at the expense of the signal's physical characteristics. This research addresses these gaps by employing evolutionary algorithms and a multi-faceted fitness function.

Methodology

To overcome the limitations of gradient descent, this study utilizes two robust evolutionary algorithms: Genetic Algorithms (GA) and Particle Swarm Optimization (PSO). Unlike traditional calculus-based methods, these algorithms explore the solution space globally, making them significantly more resilient against local optima.

The Hybrid Fitness Function

The core innovation of this research lies in the evaluation criteria used to guide the optimization process. The authors propose a fitness function that combines two distinct metrics:

  1. Kurtosis: A higher-order statistical measure that identifies non-Gaussianity. Since independent source signals are typically non-Gaussian, maximizing kurtosis helps ensure statistical independence.
  2. Feature Distance: This metric measures the distance between the features of the separated signals and the expected characteristics of the source. By incorporating feature distance, the algorithm ensures that the output maintains its physical significance and high correlation with the original source.

By optimizing for both statistical independence and physical fidelity, the algorithm produces signals that are not just mathematically distinct, but practically useful.

Experimental Results

The proposed framework was implemented in MATLAB and tested using mixed spoken signals. The experimental setup simulated a linear mixing environment where multiple speech sources were combined into a single dataset.

The results demonstrated that both GA and PSO, when guided by the hybrid fitness function, successfully separated the mixed audio. Key findings included:

  • High Fidelity: The separated signals exhibited a high degree of correlation with the original individual sources.
  • Global Convergence: The evolutionary approach effectively navigated the complex search space to find optimal separating matrices where traditional ICA might have failed.
  • Downstream Compatibility: Because the separated signals retained their physical properties, they were found to be better suited for subsequent signal processing tasks compared to those processed by purely statistical methods.

Discussion and Limitations

While the use of GA and PSO provides a more thorough search of the solution landscape, it comes with an increased computational cost. Evolutionary algorithms require more processing power and time than fast gradient-based alternatives. Additionally, this study assumes a linear mixing model. In real-world scenarios involving echoes or reverberation (convolutive mixing), the complexity of the BSS problem increases significantly.

Conclusion

This research demonstrates that incorporating physical "feature distance" into the optimization process for Blind Source Separation yields superior results compared to methods relying solely on statistical independence. By leveraging the global search capabilities of Genetic Algorithms and Particle Swarm Optimization, the proposed method provides a robust solution for recovering high-fidelity signals from complex mixtures. Future work will focus on adapting these evolutionary strategies for more complex, non-linear, and convolutive mixing environments.

Summary updated

Abstract

Without any information on the mixing system, the blind source separation (BSS) technique efficiently separates mixed signals. The approach called evolutionary algorithms was used for the BSS problem in this paper. The fitness function based on the feature distance and kurtosis was proposed to measure the degree of the separated signals in this paper. Compared with the traditional algorithm in the BSS problem, the mathematical calculation and the physical significance of the separated signals are both taken into consideration in the proposed method. Therefore, the separated signals could have great correlation with the original individual signal and could be used in the additional signal processing step with good signal property. Experimental results on mixed spoken signals indicated that the established evolutionary algorithm of particle swarm optimization (PSO) and genetic algorithm (GA) could effectively solve the BSS problem from the signal feature distance and independence measurement. The study in this paper was implemented with MATLAB language.
BibTeX
@article{yang2014blind,
  author = {Yang Yang and Xiuqin Wang and Di Zhang},
  title = {{Blind Source Separation Research Based on the Feature Distance Using Evolutionary Algorithms}},
  journal = {The International Journal of Acoustics and Vibration},
  year = {2014},
  doi = {10.20855/ijav.2014.19.4360},
}
Figure