Blind Source Separation (BSS)

The canonical BSS algorithm that recovers statistically independent source signals from linear mixtures by maximizing non-Gaussianity or minimizing mutual information. ICA operates on the central assumption that source signals are non-Gaussian and mutually independent. In RF applications, ICA separates co-channel interfering signals by iteratively estimating an unmixing matrix W such that the output components are maximally independent. Common optimization approaches include FastICA, which uses a fixed-point iteration scheme to maximize negentropy, and Infomax, which maximizes the entropy of a non-linear transformation of the outputs. ICA is particularly effective for separating constant-modulus communication signals like FM, QPSK, and FSK from unknown mixtures.

A higher-order statistics BSS algorithm that exploits fourth-order cumulants to separate source signals. JADE constructs a set of cumulant matrices from the whitened observed data and simultaneously diagonalizes them using a Jacobi rotation technique. Unlike basic ICA, JADE explicitly leverages the algebraic structure of cumulant tensors, making it analytically elegant and requiring no iterative learning rate tuning. The algorithm is particularly robust for separating mixtures where sources exhibit distinct kurtosis signatures. In spectrum sensing, JADE excels at separating signals with different modulation formats that produce unique higher-order statistical fingerprints, even when second-order methods fail.

A BSS method that exploits the temporal structure of source signals through their time-lagged covariance matrices rather than higher-order statistics. SOBI constructs multiple covariance matrices at different time delays and jointly diagonalizes them to recover the mixing matrix. This approach is effective when source signals exhibit temporal correlation or distinct power spectra—common in modulated RF signals with different symbol rates. SOBI requires only second-order statistics, making it computationally lighter than cumulant-based methods and more sample-efficient. It is particularly suited for separating narrowband communication signals with different bandwidths or carrier frequencies in spectrum monitoring applications.

A constrained BSS approach that decomposes a non-negative data matrix into two lower-rank non-negative matrices representing spectral bases and their activation patterns. In RF spectrogram processing, NMF separates overlapping signals by modeling the time-frequency representation as a sum of distinct spectral components. The non-negativity constraint produces parts-based representations that align naturally with power spectral densities. NMF is particularly effective for separating signals with sparse or repetitive spectral signatures, such as frequency-hopping patterns or radar pulses. Multiplicative update rules guarantee convergence while maintaining non-negativity, making NMF a robust tool for blind signal decomposition in dense electromagnetic environments.

A BSS paradigm that exploits signal sparsity in a transformed domain rather than statistical independence. SCA assumes that source signals are sparsely represented in some basis—such as the time-frequency domain via the Short-Time Fourier Transform (STFT)—where only one source is active at each point. This transforms the separation problem into a clustering task on scatter plots of mixture coefficients. SCA can solve the underdetermined case where there are more sources than sensors, a scenario where traditional ICA fails. In RF applications, SCA separates frequency-hopping signals, radar pulses, and burst transmissions that exhibit sparse time-frequency occupancy patterns.

Modern neural network architectures that learn to perform source separation directly from data, bypassing explicit statistical assumptions. Convolutional autoencoders and U-Net architectures trained on spectrogram mixtures can learn end-to-end separation mappings. TasNet and Conv-TasNet architectures use temporal convolutional networks to estimate separation masks in the time domain. These methods excel when the mixing process is non-linear or convolutive, violating classical ICA assumptions. In RFML, deep BSS models trained on synthetic mixtures of known modulation types can generalize to separate real-world co-channel interference, learning complex spectral and temporal separation cues that hand-crafted algorithms cannot capture.

The central optimization principle driving ICA. By the Central Limit Theorem, a mixture of independent random variables is always more Gaussian than its constituent sources. ICA algorithms iteratively search for a demixing matrix that produces outputs with maximal non-Gaussianity.

• Kurtosis: A classical measure of the 'peakedness' of a distribution. Super-Gaussian (leptokurtic) distributions have heavy tails; sub-Gaussian (platykurtic) distributions are flatter.
• Negentropy: A more robust, information-theoretic measure based on the differential entropy difference between a distribution and a Gaussian of the same variance. Approximated in FastICA for computational efficiency.

An alternative but equivalent formulation of ICA. Mutual information quantifies the amount of information one random variable contains about another. The goal is to find a linear transformation that minimizes the mutual information between output components, rendering them statistically independent.

• Kullback-Leibler Divergence: Used to measure the distance between the joint probability density function of the outputs and the product of their marginal densities.
• Infomax Algorithm: Maximizes the joint entropy of a non-linear neural network's outputs, which implicitly minimizes mutual information between inputs in a BSS context.

A critical preprocessing step before applying ICA. Whitening linearly transforms the observed mixed data so that its components are uncorrelated and have unit variance. This reduces the number of parameters the ICA algorithm must estimate, simplifying the problem from finding an arbitrary matrix to finding an orthogonal rotation.

• Process: Typically achieved via eigenvalue decomposition (PCA) of the data covariance matrix.
• Effect: The demixing matrix becomes constrained to the space of orthogonal matrices, halving the degrees of freedom and accelerating convergence.

An algebraic approach to ICA that operates on fourth-order cumulant tensors. Instead of iterative optimization, JADE computes a set of eigenmatrices from the data's cumulant tensor and then finds a rotation matrix that simultaneously diagonalizes them.

• Advantage: Provides a closed-form solution without convergence issues, making it deterministic and reliable for small to medium-dimensional problems.
• Limitation: Computational complexity scales as O(n⁴), making it impractical for high-dimensional data where iterative methods like FastICA are preferred.

An extension of the instantaneous mixing model to account for time delays and multipath reflections. In real acoustic environments, each source reaches each sensor with different delays and reverberation, requiring a demixing filter instead of a scalar matrix.

• Frequency-Domain Approach: Transform signals to the time-frequency domain via Short-Time Fourier Transform (STFT), then apply instantaneous ICA independently in each frequency bin. Requires solving the permutation problem—aligning separated components across bins.
• Applications: Real-world speech separation in reverberant rooms, convolutive radio frequency interference cancellation.