Abstract
This study addresses the challenges of Blind Source Separation (BSS) in convolutive mixtures where source signals typically lack independent identical distribution. We propose a class of Jacobi-type Joint Block Diagonalization (JBD) algorithms based on LU and QR factorizations. By decomposing high-dimensional minimization problems into sequences of one-dimensional sub-problems, these methods eliminate the need for restrictive orthogonality constraints or preliminary whitening stages. Experimental results from numerical simulations and vibration tests demonstrate that the proposed LUJBD and QUJBD algorithms offer superior convergence speed, higher precision, and improved success rates compared to state-of-the-art methods.
Introduction
Blind Source Separation (BSS) is a fundamental problem in signal processing, aiming to recover original source signals from observed mixtures without prior knowledge of the mixing process. In practical applications like acoustics and mechanical engineering, signals are often convolutive—characterized by echoes and delays—making separation significantly more complex.
Joint Block Diagonalization (JBD) has emerged as a robust framework for BSS. However, traditional JBD methods often rely on two limiting assumptions: the requirement for an orthogonal unmixing matrix and the necessity of a "pre-whitening" stage. Pre-whitening, while simplifying the problem, can introduce irreversible errors. Furthermore, many existing gradient-based solvers are computationally intensive and prone to local minima.
Proposed Methodology: LUJBD and QUJBD
To overcome these limitations, we introduce two non-orthogonal JBD algorithms: LUJBD (based on Lower-Upper factorization) and QUJBD (based on Orthogonal-Triangular factorization).
Jacobi-type Optimization
The core innovation lies in the use of a Jacobi-type scheme. Instead of tackling a high-dimensional optimization problem directly, the algorithm breaks it down into a series of one-dimensional sub-optimizations. This approach iteratively updates the parameters of the LU or QR components to minimize the off-diagonal block energy of the target matrices.
Computational Efficiency
Unlike previous methods that required solving degree-4 polynomials for each update, the proposed LUJBD and QUJBD algorithms only require solving degree-2 polynomials. This reduction in algebraic complexity leads to significantly faster computation and enhanced robustness against local optima.
Experimental Validation
The performance of the proposed algorithms was evaluated through two primary methods:
- Numerical Simulations: Synthetic datasets with varying Signal-to-Noise Ratios (SNR) and matrix dimensions were used to benchmark the algorithms against existing state-of-the-art JBD methods (e.g., JBD-OG and JBD-ORG).
- Underwater Vibration Experiment: A real-world test was conducted using a double-stiffened cylindrical shell submerged in an anechoic pool. Three exciters generated distinct frequency sources (3kHz, 4kHz, and 5kHz), and 29 sensors captured the resulting convolutive mixtures.
Results and Discussion
The experimental results highlight several key advantages of the LUJBD and QUJBD approaches:
- Convergence and Precision: Both algorithms demonstrated a higher success rate in reaching the global minimum compared to gradient-based alternatives. The Performance Index (PI), a metric for separation quality, was consistently lower for the proposed methods across different SNR levels.
- Speed: LUJBD, in particular, showed the shortest average running time per successful convergence, making it highly suitable for time-sensitive applications.
- Robustness: By avoiding the pre-whitening stage, the algorithms proved more effective for ill-conditioned mixtures where sensor placement or signal correlation might otherwise degrade performance.
- Practical Application: In the vibration experiment, the algorithms successfully isolated the three source frequencies from the complex, aliased sensor data, confirming their utility in structural health monitoring and mechanical diagnosis.
Conclusion
The proposed LUJBD and QUJBD algorithms provide a more general and efficient framework for convolutive blind source separation. By leveraging Jacobi-type updates and non-orthogonal factorizations, they achieve superior precision and speed without the drawbacks of traditional whitening or orthogonality constraints. Future work will focus on refining termination criteria to further optimize the balance between computational cost and accuracy in diverse industrial environments.
Summary updated