Wavelet Transform

Unlike the Short-Time Fourier Transform (STFT) which uses a fixed window size, the wavelet transform employs a variable window that adapts to frequency. High frequencies are analyzed with narrow time windows for precise temporal localization, while low frequencies use wide windows for fine spectral resolution. This scale-dependent analysis reveals both the sharp transient edges of a signal's turn-on burst and the subtle, persistent modulation distortions in its steady-state payload, providing a richer feature set for Specific Emitter Identification (SEI).

The turn-on and turn-off transients of a transmitter are rich in hardware-specific information but are extremely brief and non-stationary. The Discrete Wavelet Transform (DWT) excels at localizing these events in time. By decomposing the signal into approximation and detail coefficients, the DWT isolates the sharp discontinuities caused by power amplifier ramp-up and oscillator settling. These coefficients serve as compact, highly discriminative feature vectors that directly encode the unique physical impulse response of the transmitter's analog front-end.

Wavelet-based denoising is a powerful preprocessing step for RF fingerprinting. The process involves:

• Decomposition: Applying the DWT to the noisy IQ signal to produce wavelet coefficients.
• Thresholding: Zeroing out coefficients below a calculated noise floor, based on the principle that signal energy is concentrated in a few large coefficients while noise is spread across many small ones.
• Reconstruction: Applying the inverse DWT to the thresholded coefficients. This technique removes additive white Gaussian noise without smearing the sharp transient edges that are vital for device discrimination, unlike linear filtering.

A scalogram is a visual representation of the Continuous Wavelet Transform (CWT) coefficients, plotting the magnitude of correlation between the signal and a chosen mother wavelet across time and scale. This time-scale image reveals the instantaneous frequency content of a signal. When fed into a Convolutional Neural Network (CNN), the scalogram acts as a highly detailed texture map of the transmitter's behavior. Subtle hardware impairments, such as I/Q imbalance and phase noise, manifest as distinct, repeatable patterns in the scalogram's texture.

The choice of mother wavelet is critical and must match the signal's morphology to maximize feature extraction. Common choices include:

• Morlet: Excellent for signals with oscillatory components, providing a good balance between time and frequency localization.
• Daubechies (dbN): Compact support and vanishing moments make them ideal for detecting discontinuities and singularities in transient signals.
• Haar: The simplest wavelet, effective for detecting sudden step-changes in amplitude. The optimal mother wavelet maximizes the Kullback-Leibler divergence between the coefficient distributions of different devices.

The Discrete Wavelet Transform (DWT) uses dyadic scales and shifts, producing a non-redundant, compact representation ideal for feature vector extraction and computational efficiency on edge hardware. In contrast, the Continuous Wavelet Transform (CWT) computes correlations at every possible scale, generating a highly redundant but information-rich scalogram. The CWT is preferred for generating detailed visual inputs for CNNs, while the DWT is the standard for extracting low-dimensional statistical features for lightweight classifiers.

A visual representation of the frequency spectrum of a signal as it varies over time, generated via the Short-Time Fourier Transform (STFT). Unlike the wavelet transform's variable resolution, the STFT uses a fixed window size, creating a uniform trade-off between time and frequency precision. Spectrograms are frequently used as 2D image inputs for Convolutional Neural Networks (CNNs) in deep learning-based signal identification, converting raw IQ data into a format suitable for standard computer vision architectures.

A signal processing technique that exploits the periodic statistical properties of modulated signals. While wavelet transforms excel at isolating transient events and non-stationary anomalies, cyclostationary analysis focuses on the inherent periodicities in a signal's mean and autocorrelation function. These cyclic features are robust to stationary noise and are highly effective for identifying the modulation scheme and specific hardware imperfections that repeat at symbol-rate intervals.

A higher-order statistic that suppresses Gaussian noise while preserving phase information, revealing non-linear coupling characteristics unique to specific transmitter hardware. Wavelet transforms provide multi-resolution time-frequency localization, whereas the bispectrum operates in the bifrequency domain to detect quadratic phase coupling. This makes it a powerful complementary tool for extracting features caused by non-linear amplifier distortions that a linear time-frequency transform might obscure.

A deep learning architecture that uses convolutional filters to automatically learn spatial hierarchies of features from grid-like data. When applied to RF fingerprinting, CNNs often process wavelet scalograms—2D visual representations of the wavelet transform's coefficients. The multi-resolution property of the scalogram provides the CNN with a rich, structured input that captures both fine-grained transient details and broader spectral trends, improving classification accuracy over raw time-series data.

The extraction of identifying features from the brief turn-on and turn-off periods of a transmitter's signal burst. The Discrete Wavelet Transform (DWT) is uniquely suited for this task because its variable window size provides high temporal resolution at high frequencies, precisely capturing the abrupt, non-stationary nature of transients. This allows the system to isolate hardware-dependent ringing and overshoot behaviors that are often masked in steady-state analysis.

The process of mapping high-dimensional signal data into a lower-dimensional vector space where semantically similar device signatures are clustered closely together. Wavelet coefficients often serve as the initial feature set for this process. By decomposing a signal into a set of approximation and detail coefficients, the wavelet transform creates a compact, informative representation that can be fed into Siamese Networks or Triplet Loss frameworks to learn highly discriminative, channel-robust embeddings for device authentication.