Abstract
This paper presents the Prewhitened Non-Cancellation Multistage (PNCMS) algorithms, a significant optimization for Blind Source Separation (BSS). By integrating prewhitening with a novel parallel processing architecture, the proposed method reduces computational complexity and processing latency while maintaining high signal separation quality. Experimental results demonstrate that PNCMS can extract multiple signal sources faster than traditional sequential methods and exhibits enhanced robustness in low signal-to-noise ratio (SNR) environments.
Introduction
Blind Source Separation (BSS) addresses the "cocktail party problem," where individual signal sources must be recovered from a set of mixed recordings without prior knowledge of the mixing process. While existing Non-Cancellation Multistage (NCMS) algorithms effectively mitigate error propagation—where errors in early extraction stages compromise subsequent ones—they typically suffer from high computational overhead and significant latency due to their sequential nature. As the number of sensors and sources increases, these limitations become more pronounced.
Proposed Methodology
The authors propose two primary enhancements to the NCMS framework, specifically targeting the Turbo Source Extraction (TSEA) and Fast Kurtosis Maximization (FKMA) variants:
- Prewhitening (PNCMS): Before the unmixing process, the input data undergoes a prewhitening transformation. This reduces dimensionality by ignoring noise-only components and transforms the unknown mixing matrix into a unitary matrix. This "unitary shortcut" allows complex projection operations, such as Singular Value Decomposition (SVD), to be replaced by simpler matrix subtractions, drastically reducing the required floating-point operations (FLOPs).
- Parallel Extraction: Instead of sequential processing, the PNCMS(p) versions utilize a bank of extraction modules operating in parallel. To prevent these modules from converging on the same source, they are initialized using distinct columns of a unitary matrix. A subsequent "Source Signal Classification" step identifies and merges duplicate signals based on correlation.
Experimental Results and Discussion
The performance of the PNCMS algorithms was evaluated using synthetic datasets with mixing matrices ranging from $5 \times 4$ to $8 \times 6$ (sensors $\times$ sources). Source signals were simulated as temporally colored i.i.d. binary sequences in the presence of white Gaussian noise.
Efficiency and Latency
A key finding is the efficiency of the parallel architecture. In a 10-source scenario, the PNCMS-TSEA(p) variant can typically extract all sources in fewer than three stages, whereas sequential NCMS requires ten. This is supported by the average number of distinct sources captured in the first stage of parallel extraction.
As shown in the table above, the parallel bank typically extracts over 80% of distinct sources in the very first stage. Furthermore, the average number of iterations required per source decreased significantly; for instance, in tests using mixing matrix $A_2$, iterations dropped from approximately 450 to 310.
Robustness
The PNCMS approach demonstrated superior performance at low SNR compared to original NCMS methods. The prewhitening step acts as a preliminary noise filter, cleaning the data before extraction begins. The Signal-to-Interference-plus-Noise Ratio (SINR) matched or exceeded benchmarks like FastICA and SOBI, particularly in challenging cases where sources shared identical power spectra.
Conclusion
The PNCMS framework provides a scalable and computationally efficient solution for Blind Source Separation. By simplifying the search space through prewhitening and leveraging parallel processing, it overcomes the latency and complexity hurdles of previous multistage methods. While currently limited to non-Gaussian sources, future work may adapt this non-cancellation structure to other separation criteria, broadening its applicability in complex signal environments.
Summary updated