Continuous Wavelet Transform (CWT)

The CWT uses a variable window that adapts its shape based on frequency content. High frequencies (low scales) are analyzed with a narrow time window for good temporal localization, while low frequencies (high scales) use a wide time window for high spectral resolution. This directly addresses the Heisenberg uncertainty principle by trading off time and frequency precision dynamically, making it superior to the fixed-resolution Short-Time Fourier Transform (STFT) for signals with both short bursts and long oscillations.

Unlike the orthogonal Discrete Wavelet Transform (DWT), the CWT produces a highly redundant representation. By continuously varying the scale and translation parameters, it maps a 1D time series into a 2D time-scale image. This overcompleteness preserves phase information and provides robustness to small time shifts, making the resulting scalogram visually intuitive and rich in features for machine learning models, particularly for identifying subtle hardware impairments in RF fingerprinting.

The CWT's performance hinges on the choice of the mother wavelet. The transform computes the inner product between the signal and scaled, shifted versions of this prototype function. Effective analysis requires selecting a wavelet that morphologically resembles the target signal features:

• Morlet wavelet: Optimal for oscillatory, transient RF pulses due to its Gaussian-enveloped complex sinusoid.
• Mexican Hat wavelet: Proportional to the second derivative of a Gaussian, ideal for detecting sharp edges and singularities.
• Haar wavelet: A simple step function useful for detecting abrupt changes in signal state.

The CWT output is typically visualized as a scalogram, a 2D heatmap where the x-axis represents time, the y-axis represents scale (inversely proportional to frequency), and color intensity represents the magnitude of the wavelet coefficients. Unlike a spectrogram's linear frequency axis, a scalogram's logarithmic scale axis provides a natural multi-resolution view. Energy ridges in the scalogram directly correspond to the instantaneous frequency trajectories of signal components, a critical feature for emitter identification.

Directly computing the CWT integral for every scale and translation is computationally prohibitive. In practice, the CWT is efficiently implemented using the Convolution Theorem. The transform is calculated as a series of inverse Fast Fourier Transforms (IFFTs) of the product of the signal's FFT and the complex conjugate of the scaled wavelet's FFT. This frequency-domain implementation dramatically accelerates computation, enabling real-time or near-real-time analysis of long signal recordings.

CWT scales do not directly map to Fourier frequencies. The conversion depends on the center frequency of the mother wavelet. The pseudo-frequency F_a for a given scale a is calculated as F_a = F_c / (a * Δ), where F_c is the wavelet's center frequency and Δ is the signal's sampling period. This relationship is non-linear, meaning the CWT provides a logarithmic frequency resolution that naturally matches the perceptual and physical characteristics of many real-world signals.

The scalogram is the visual representation of the absolute value of CWT coefficients plotted as a function of time and scale. Since scale is inversely related to frequency, it provides a multi-resolution heatmap of signal energy distribution.

• Time resolution: Fine at high frequencies (small scales)
• Frequency resolution: Fine at low frequencies (large scales)
• Unlike a spectrogram, the scalogram uses a logarithmic frequency axis that mirrors human perception

The Morlet wavelet is the most widely used mother wavelet for CWT analysis, consisting of a complex plane wave modulated by a Gaussian envelope. Its shape closely resembles a windowed sinusoid.

• Provides optimal joint time-frequency localization
• The complex nature yields both magnitude and phase information
• Directly related to the Fourier transform, making frequency interpretation intuitive
• Central frequency and bandwidth are adjustable via a single parameter

Multiresolution Analysis is the mathematical framework that underpins wavelet theory, decomposing a function space into a sequence of nested subspaces. It formalizes how CWT analyzes signals at different frequencies with different resolutions.

• Coarse scales capture global trends and slow variations
• Fine scales capture transient events and sharp discontinuities
• Provides the theoretical link between CWT and DWT
• Enables perfect reconstruction from decomposed components

The Synchrosqueezing Transform is a post-processing technique that sharpens a CWT scalogram by reallocating coefficients along the frequency axis based on instantaneous frequency estimates. It concentrates diffused energy onto the true time-frequency ridges.

• Improves readability of closely spaced components
• Retains the invertibility of the original CWT
• Particularly effective for signals with strong frequency modulation
• Bridges the gap between wavelet and empirical mode decomposition methods

The STFT is the linear predecessor to the CWT, computing the Fourier transform over windowed signal segments. The critical distinction lies in the fixed window size, which imposes a uniform time-frequency resolution trade-off across all frequencies.

• CWT's variable window provides superior analysis of transients
• STFT uses a linear frequency scale; CWT uses logarithmic
• The Gabor transform is a special case using a Gaussian window
• STFT is computationally lighter but less adaptive than CWT

The Wigner-Ville Distribution is a quadratic time-frequency representation offering the highest possible joint resolution. Unlike the linear CWT, it computes the Fourier transform of the signal's instantaneous autocorrelation.

• Achieves maximum theoretical sharpness for mono-component signals
• Suffers from cross-term interference for multi-component signals
• CWT avoids cross-terms entirely due to its linear nature
• Cohen's class distributions apply kernels to suppress WVD artifacts