Empirical Mode Decomposition

A two-step adaptive time-frequency analysis method that pairs Empirical Mode Decomposition with the Hilbert Spectral Analysis. After EMD decomposes a signal into Intrinsic Mode Functions (IMFs), the Hilbert transform is applied to each IMF to derive instantaneous frequency and amplitude. This yields a high-resolution time-frequency-energy distribution called the Hilbert spectrum, which is particularly effective for non-linear and non-stationary RF signals where Fourier-based methods fail. The combination reveals transient hardware-induced oscillatory signatures that are invisible to traditional spectrograms.

The fundamental oscillatory components produced by EMD. An IMF must satisfy two conditions: (1) the number of extrema and zero-crossings must differ by at most one, and (2) the mean value of the envelope defined by local maxima and minima must be zero at any point. Unlike Fourier harmonics, IMFs are data-driven and adaptive, meaning they can capture the asymmetric, time-varying waveforms characteristic of power amplifier memory effects and transient turn-on signatures. Each IMF isolates a specific physical mechanism within the transmitter's hardware chain.

The iterative algorithmic core of EMD that extracts IMFs from a signal. The process works by:

• Identifying all local maxima and minima of the signal
• Fitting cubic spline envelopes through these extrema
• Computing the mean envelope and subtracting it from the signal
• Repeating until the residual meets the IMF stopping criterion This sifting isolates the finest oscillatory mode first, progressively extracting coarser components. For RF fingerprinting, the sifting process naturally separates carrier frequency offset effects from higher-order phase noise modulations without requiring predefined basis functions.

A deep convolutional network based on wavelet operators that produces stable, translation-invariant signal representations. While EMD decomposes adaptively in the time domain, wavelet scattering cascades through fixed wavelet filter banks followed by modulus non-linearities. Both techniques share the goal of extracting robust features from non-stationary signals. In RF fingerprinting, scattering networks can complement EMD by providing deformation-stable features that are invariant to small time-warping distortions caused by channel effects, while EMD captures the intrinsic oscillatory modes of hardware impairments.

A quadratic time-frequency representation that provides the highest possible joint resolution for analyzing instantaneous frequency and energy distribution. Unlike the Hilbert spectrum derived from EMD, the Wigner-Ville distribution can suffer from cross-term interference when analyzing multi-component signals. EMD's decomposition into mono-component IMFs before applying Hilbert analysis elegantly avoids this artifact. For transient signal analysis, the combination of EMD pre-processing followed by Wigner-Ville analysis on individual IMFs yields exceptionally clean time-frequency representations of device turn-on signatures.

The analysis of signals whose statistical properties—mean, autocorrelation—vary periodically with time. Communication signals exhibit cyclostationarity due to symbol rates, carrier frequencies, and guard intervals. While EMD operates on the raw waveform to extract adaptive oscillatory modes, cyclostationary processing exploits these hidden periodicities to compute the spectral correlation density function. The two approaches are complementary: EMD can isolate transient and non-periodic hardware impairments, while cyclostationary analysis extracts robust, modulation-specific features that are immune to stationary noise.