Hilbert-Huang Transform

The foundational, adaptive algorithm of the HHT. EMD sifts a complex signal into a finite set of Intrinsic Mode Functions (IMFs) without predefined basis functions.

• Data-driven: Decomposition is based on the signal's local time scale, not fixed kernels.
• Sifting process: Iteratively identifies and subtracts local mean envelopes from extrema.
• Completeness: The original signal can be perfectly reconstructed by summing all IMFs and the final residual trend.

The fundamental building blocks output by EMD. An IMF is a function satisfying two strict conditions:

• Zero mean symmetry: The number of extrema and zero-crossings must differ by at most one.
• Local mean is zero: The mean value of the upper and lower envelopes, defined by local maxima and minima, is zero at any point. These properties ensure each IMF represents a simple oscillatory mode amenable to instantaneous frequency analysis.

Applying the Hilbert transform to each IMF yields analytic signals, from which physically meaningful instantaneous amplitude and frequency can be derived.

• Hilbert Spectral Analysis: Constructs a time-frequency-energy distribution without the uncertainty principle limitations of wavelet or Fourier methods.
• Sharp localization: Captures transient events like power amplifier memory effects or turn-on spikes with high temporal precision.
• Non-stationary analysis: Ideal for tracking how a transmitter's carrier frequency offset drifts during a burst.

Unlike Fourier or wavelet transforms, HHT is not constrained by assumptions of linearity. EMD naturally isolates non-linear harmonic distortions.

• Intra-wave frequency modulation: The Hilbert transform can reveal frequency changes within a single cycle of an IMF, directly quantifying amplifier non-linearity.
• Bispectrum alternative: HHT can separate non-linear coupling effects without higher-order statistics, providing a computationally efficient path to fingerprinting non-linear hardware impairments.

HHT excels at isolating the non-repeating, transient components of a signal that are critical for transient analysis in RF fingerprinting.

• Residual trend separation: The final EMD residual captures the slow-varying trend, leaving IMFs that cleanly represent the fast turn-on amplitude ramp.
• No cross-term interference: Unlike the Wigner-Ville Distribution, HHT avoids artificial interference terms, providing a clean time-frequency view of a transmitter's start-up signature.

A known drawback of standard EMD is mode mixing, where a single IMF contains signals of disparate scales or a single scale appears across multiple IMFs.

• Cause: Often triggered by intermittent noise or signal fragmentation.
• Mitigation: The Ensemble EMD (EEMD) variant adds white noise to the signal before decomposition, averaging out the mixing across multiple trials to improve stability.
• Relevance: For RF fingerprinting, EEMD ensures that a specific hardware impairment, like local oscillator leakage, is consistently isolated in the same IMF across captures.