Blind Source Separation (BSS)

The cornerstone statistical technique for BSS that assumes source signals are statistically independent and non-Gaussian. ICA searches for a linear transformation that maximizes the non-Gaussianity of the estimated sources, often using kurtosis or negentropy as objective functions. The FastICA algorithm is a widely used fixed-point iteration scheme that converges quickly. ICA is highly effective for separating co-channel interference and identifying rogue emitters when the mixing environment is linear and instantaneous.

An algebraic BSS method that exploits fourth-order cumulants (higher-order statistics) to separate sources. JADE constructs a set of cumulant matrices and simultaneously diagonalizes them to find the unmixing matrix. It is particularly robust when source signals have distinct spectral kurtosis profiles, making it suitable for separating communication signals with different modulation formats. Unlike iterative ICA methods, JADE provides a closed-form solution, offering deterministic performance without convergence concerns.

A decomposition technique that constrains source signals and mixing coefficients to be non-negative, making it ideal for power spectrogram data where energy values cannot be negative. NMF factorizes a time-frequency representation into a basis matrix (spectral templates of individual sources) and an activation matrix (when each source is active). This approach excels at separating intermittent transmissions and frequency-hopping signals in dense spectrum environments by learning recurring spectral patterns.

A BSS paradigm that exploits the property that sources are sparse in some transform domain (e.g., time-frequency). When signals rarely overlap in the transformed space, clustering algorithms can identify single-source points where only one source is active, enabling estimation of the mixing matrix. SCA can solve the underdetermined case where there are more sources than sensors, a critical capability for separating multiple anomalous emitters with a limited antenna array.

A BSS algorithm that relies on second-order statistics (correlation matrices) rather than higher-order moments. SOBI jointly diagonalizes multiple time-lagged covariance matrices, exploiting the temporal structure and stationarity of source signals. This makes it particularly effective for separating colored sources with distinct autocorrelation functions, such as modulated carriers with different symbol rates. SOBI is computationally efficient and robust to Gaussian noise.

Modern neural network architectures that learn to perform BSS directly from data, bypassing explicit statistical assumptions. TasNet and Conv-TasNet use temporal convolutional networks to estimate separation masks in the time domain. Deep Clustering embeds time-frequency bins into a high-dimensional space where bins belonging to the same source cluster together. These methods excel at separating complex, non-linear mixtures and can adapt to convolutive mixing environments with multipath propagation.

The foundational statistical technique for BSS that separates multivariate signals into additive subcomponents by maximizing their statistical independence. ICA assumes source signals are non-Gaussian and mutually independent, making it ideal for separating co-channel emitters in spectrum monitoring. Common algorithms include FastICA and Infomax, which iteratively optimize a contrast function to unmix the observed signals.

The process of recovering the unknown mixing matrix A that describes how source signals combine at sensor arrays. In RF applications, this matrix represents the complex channel coefficients between emitters and receiver antennas. Accurate estimation is critical for spatial filtering and source localization. Techniques include joint diagonalization of time-lagged covariance matrices and maximum likelihood approaches.

A BSS alternative that decomposes a non-negative data matrix into two lower-rank non-negative matrices. In spectrum analysis, NMF excels at separating power spectral densities where negative values have no physical meaning. Particularly effective for identifying intermittent emitters whose spectral footprints overlap in time and frequency but have distinct activation patterns.

A BSS approach exploiting the property that source signals are sparse in some transform domain (e.g., time-frequency). When sources rarely overlap in the transform domain, single-source points can be identified to estimate the mixing matrix. This technique is robust for separating pulsed radar signals or frequency-hopping transmissions where sources occupy disjoint time-frequency bins.

A BSS algorithm that exploits temporal correlations in source signals rather than higher-order statistics. SOBI jointly diagonalizes multiple time-lagged covariance matrices, making it computationally efficient and well-suited for separating colored sources like modulated communications signals. It requires only second-order statistics, unlike ICA which depends on non-Gaussianity.

An inherent limitation in BSS where the ordering of recovered sources is arbitrary. In spectrum monitoring, this means separated emitters cannot be assigned persistent identities across time windows without additional fingerprinting. Solutions include using RF fingerprinting or tracking spectral features to maintain consistent source labeling across consecutive separation operations.