Variational Mode Decomposition (VMD)

Unlike recursive sifting in Empirical Mode Decomposition (EMD), VMD formulates mode extraction as a constrained variational problem. It minimizes the sum of the estimated bandwidths of each mode, subject to the constraint that the modes collectively reconstruct the original signal. This is achieved by finding the analytic signal via the Hilbert transform, shifting it to baseband, and estimating bandwidth through H¹ Gaussian smoothness.

VMD solves the variational optimization problem using the ADMM algorithm. This splits the complex minimization into simpler sub-problems, iteratively updating:

• Each Intrinsic Mode Function (IMF) in the frequency domain via Wiener filtering.
• The center frequencies as the power spectrum centroid of each mode. This ensures concurrent, non-recursive extraction and convergence to a saddle point of the augmented Lagrangian.

VMD redefines an Intrinsic Mode Function as an amplitude-modulated-frequency-modulated (AM-FM) signal with a specific sparsity property: it is compact around a central frequency.

• Mathematical definition: u_k(t) = A_k(t) cos(φ_k(t))
• Bandwidth constraint: The spectral support is limited, preventing mode mixing. This contrasts with EMD's heuristic definition based on extrema and zero-crossings.

The variational formulation provides inherent noise robustness and resistance to mode mixing. Key advantages include:

• Wiener filtering embedded in the update step directly denoises each mode.
• Non-recursive nature prevents error propagation between modes.
• Down-sampling immunity: VMD accurately recovers center frequencies even when the signal is sampled below the Nyquist rate of the highest component, a property known as sub-Nyquist tone detection.

VMD requires the user to predefine the number of modes, K. This is both a limitation and a feature:

• Advantage: Direct control over the decomposition granularity for specific applications like RF fingerprinting where the number of expected signal components may be known.
• Challenge: Over-specifying K leads to mode duplication; under-specifying causes mode mixing.
• Mitigation: Detrended fluctuation analysis or center frequency inspection is used to optimize K.

To enforce the reconstruction constraint (sum of modes equals original signal), VMD uses a quadratic penalty and Lagrangian multipliers (λ). The dual ascent step updates λ to drive the residual toward zero:

• Quadratic penalty: Accelerates convergence via convexity.
• Lagrangian multiplier: Enforces strict constraint satisfaction.
• Update rule: λ̂ⁿ⁺¹ = λ̂ⁿ + τ(f̂ - Σ û_k), where τ is the dual ascent step size. This combination ensures the final decomposition accurately reconstructs the input signal.

A data-driven, recursive algorithm that decomposes a signal into Intrinsic Mode Functions (IMFs) without requiring a predefined basis. Unlike VMD's variational optimization framework, EMD uses an iterative sifting process to extract local mean envelopes.

• Key difference: EMD is heuristic and lacks a rigorous mathematical foundation, making it sensitive to noise and sampling.
• Mode mixing: A common failure where a single IMF contains disparate frequency components, a problem VMD's band-limited optimization explicitly avoids.
• Applications: Still widely used in biomedical signal processing and geophysical data analysis where computational simplicity is prioritized.

An adaptive wavelet-based decomposition that segments the Fourier spectrum into compact supports and builds a bank of wavelet filters tailored to each segment. Like VMD, EWT extracts amplitude-modulated and frequency-modulated (AM-FM) components.

• Spectrum segmentation: Requires detecting boundaries between modes in the frequency domain, which can be challenging in noisy environments.
• Comparison to VMD: EWT is non-iterative and computationally lighter, but VMD's variational formulation provides superior robustness to noise and sampling artifacts.
• Use case: Effective for signals with clearly separated spectral supports, such as rotating machinery fault diagnosis.

A time-frequency reassignment technique that sharpens spectrograms by reallocating coefficients along the frequency axis based on instantaneous frequency estimates. SST concentrates diffuse energy onto sharp, readable ridges.

• Mode retrieval: Individual components can be reconstructed by integrating around identified ridges, enabling signal decomposition similar to VMD's output.
• Mathematical rigor: SST provides strong theoretical guarantees for mode separation under well-separated instantaneous frequency conditions.
• Limitation: Struggles with modes that cross in the time-frequency plane, whereas VMD's non-recursive nature handles overlapping spectra more gracefully.

A two-stage adaptive analysis framework combining Empirical Mode Decomposition with the Hilbert spectral analysis. EMD first extracts IMFs, then the Hilbert transform computes instantaneous frequencies and amplitudes for each mode.

• Output: Produces a high-resolution time-frequency-energy representation without the uncertainty principle limitations of linear transforms.
• Nonlinear capability: Designed specifically for non-stationary and nonlinear signals, making it popular in structural health monitoring.
• VMD advantage: VMD replaces EMD's heuristic sifting with a principled variational optimization, resulting in modes that are more spectrally compact and less prone to mode mixing.

A greedy sparse approximation algorithm that iteratively decomposes a signal into a linear combination of waveforms (atoms) selected from a redundant dictionary. Each iteration selects the atom with the highest inner product with the current residual.

• Dictionary flexibility: Unlike VMD's implicit sinusoidal basis, matching pursuit can use Gabor atoms, chirplets, or custom waveforms tailored to specific signal morphologies.
• Sparsity: Naturally produces a parsimonious representation, useful for compression and feature extraction.
• Computational cost: The iterative search over large dictionaries is computationally intensive compared to VMD's efficient alternating direction method of multipliers (ADMM) optimization.

An optimization framework that decomposes a signal into a sparse superposition of dictionary atoms by minimizing a least-squares error term subject to an L1-norm penalty on the coefficients. This convex relaxation promotes sparsity while denoising.

• Convexity guarantee: Unlike matching pursuit's greedy approach, BPDN solves a convex optimization problem with a guaranteed global minimum.
• Relation to VMD: Both use variational optimization, but BPDN operates on fixed dictionaries while VMD jointly estimates modes and their center frequencies adaptively.
• Parameter tuning: The regularization parameter balances fidelity and sparsity, analogous to VMD's penalty parameter controlling mode bandwidth compactness.