Cyclostationary Feature Extraction

The periodic autocorrelation function is the foundational mathematical tool for cyclostationary analysis. Unlike stationary signals whose autocorrelation depends only on the time lag, a cyclostationary signal's autocorrelation is a periodic function of time. This periodicity arises from the coupling between the carrier frequency and the symbol rate in modulated signals.

• Quadratic transformation: Computed by squaring or delaying-and-multiplying the signal to reveal hidden periodicities
• Fourier series expansion: The periodic autocorrelation is decomposed into cyclic autocorrelation functions at discrete cycle frequencies
• Cycle frequencies: Typically occur at multiples of the symbol rate, carrier frequency offsets, and combinations thereof (e.g., α = k/T_symbol ± 2f_c)
• Noise suppression: Stationary noise exhibits no cyclostationarity, making this feature inherently robust to additive white Gaussian noise (AWGN)

The Spectral Correlation Density is the frequency-domain representation of cyclostationarity, obtained by taking the 2D Fourier transform of the periodic autocorrelation. It reveals the correlation between spectral components separated by specific frequency shifts.

• Bifrequency plane: The SCD is a 2D function S_x^α(f) where α is the cycle frequency and f is the spectral frequency
• Spectral redundancy: Modulated signals exhibit correlation between spectral components at f + α/2 and f - α/2 due to the periodic structure of the pulse-shaping filter
• Modulation-specific patterns: BPSK, QPSK, and QAM each produce distinct SCD signatures with peaks at unique cycle frequencies
• Interference separation: Overlapping signals with different symbol rates or carrier frequencies can be separated in the bifrequency plane even when they occupy the same bandwidth

Cyclic cumulants extend cyclostationary analysis beyond second-order statistics to higher orders, capturing non-Gaussian signal properties that are critical for distinguishing between modulation formats with identical power spectra.

• nth-order cyclic cumulants: Capture phase and frequency coupling at orders n ≥ 3, revealing nonlinear signal structure
• Modulation classification: Higher-order cyclic cumulants can uniquely identify modulation types (e.g., distinguishing 16-QAM from 64-QAM) because each constellation imposes distinct higher-order moment patterns
• Gaussian noise immunity: Gaussian processes have zero cumulants above second order, so higher-order cyclic cumulants are theoretically immune to colored Gaussian interference
• Hardware fingerprinting: Subtle amplifier non-linearities generate unique higher-order cyclostationary signatures that serve as device-specific identifiers

In OFDM systems, the cyclic prefix (CP) intentionally introduces a structured periodicity that creates strong cyclostationary features exploitable for both signal identification and emitter fingerprinting.

• CP structure: The cyclic prefix is a copy of the end of each OFDM symbol prepended to its beginning, creating a repeating pattern with period equal to the useful symbol length T_u
• Cycle frequencies: CP-induced cyclostationarity appears at α = k/T_s where T_s = T_u + T_cp is the total OFDM symbol duration
• Blind parameter estimation: The CP length, useful symbol duration, and FFT size can be blindly estimated by detecting these cycle frequencies without prior knowledge of the transmitter configuration
• Device-specific CP variations: Slight timing errors, clock drift, and filter imperfections in the CP insertion process create unique, hardware-dependent cyclostationary signatures

The conjugate cyclic autocorrelation function captures the correlation between a signal and its complex conjugate, revealing cyclostationarity that is invisible to conventional non-conjugate analysis. This is essential for detecting impropriety in complex-valued signals.

• Improper signals: A complex signal is improper if it is correlated with its conjugate, a property exhibited by many real-world modulated signals (e.g., BPSK, GMSK)
• Complementary cycle frequencies: Conjugate analysis reveals cycle frequencies at α = ±2f_c + k/T_symbol, where f_c is the carrier frequency
• I/Q imbalance detection: Hardware impairments like gain and phase mismatch between I and Q branches generate strong conjugate cyclostationary features that serve as robust device fingerprints
• Carrier frequency offset estimation: The conjugate cyclic autocorrelation provides a high-precision estimate of the carrier frequency offset independent of the symbol timing

The Strip Spectral Correlation Analyzer is a computationally efficient algorithm for estimating the spectral correlation density in real-time, making cyclostationary feature extraction practical for embedded and edge deployment.

• Time-smoothing method: The SSCA computes the SCD by averaging the complex demodulates of narrowband frequency channels over time, trading resolution for computational tractability
• Channelizer-based architecture: Uses an FFT-based filter bank to decompose the wideband input into parallel narrowband channels, each processed independently
• Real-time operation: Enables continuous cyclostationary feature extraction on streaming IQ data without requiring batch processing
• FPGA implementation: The SSCA's parallel structure maps efficiently to FPGA and GPU architectures, enabling deployment on software-defined radios for tactical emitter identification

The fundamental two-dimensional transform that reveals cyclostationarity by measuring the correlation between a signal's spectral components at frequencies separated by the cycle frequency. Unlike the power spectral density, the SCD exposes hidden periodicities in the signal's statistics, making it the primary tool for distinguishing between modulated signals that share the same PSD but have different symbol rates or carrier offsets.

• Computed as the Fourier transform of the cyclic autocorrelation function
• Peaks in the SCD occur at cycle frequencies corresponding to the symbol rate, carrier frequency, and their harmonics
• Robust to stationary noise and interference, which exhibit no spectral correlation

A time-domain transformation that quantifies the periodic statistical structure of a signal by computing the autocorrelation as a function of both time lag and cyclic frequency. For a cyclostationary signal, this function is non-zero only at discrete cycle frequencies that are integer multiples of the signal's fundamental periodicities, such as the symbol rate.

• Defined as the Fourier coefficient of the time-varying autocorrelation
• Directly isolates the hidden periodicity of modulated signals
• Serves as the time-domain counterpart to the spectral correlation density

A one-dimensional slice of the spectral correlation density at a fixed spectral frequency, revealing the unique cycle frequency signature of a specific modulation scheme. Each modulation format—BPSK, QPSK, 16-QAM—produces a distinct pattern of cycle frequency peaks that can be used as a robust feature vector for automatic modulation classification and emitter identification.

• Peaks at the symbol rate and its harmonics are universal across digital modulations
• Carrier frequency offsets create device-specific shifts in the cycle frequency profile
• Provides a low-dimensional, highly discriminative feature set for deep learning classifiers

An extension of cyclostationary analysis that exploits higher-order statistics to extract features from signals where second-order cyclostationarity is weak or absent. Cyclic cumulants are sensitive to the non-Gaussian nature of modulated signals and can uniquely identify modulation formats that share identical second-order cyclic features.

• Fourth-order cyclic cumulants distinguish between QAM constellations of different orders
• Robust to Gaussian noise, which has zero higher-order cumulants
• Enables hierarchical classification by first detecting cyclostationarity, then identifying the specific modulation

The FFT Accumulation Method (FAM) is a computationally efficient algorithm for estimating the spectral correlation density. By extracting specific slices of the SCD at known cycle frequencies, this technique generates a compact, device-specific feature vector that captures the unique hardware impairments and modulation parameters of a transmitter.

• Reduces the full SCD to a manageable set of discriminative features
• FAM algorithm uses channelization to trade off resolution and compute time
• Extracted slices serve as input to convolutional neural networks for emitter classification

A variant of cyclic autocorrelation that exploits the complex-valued nature of baseband IQ signals. For non-circular modulations like BPSK and GMSK, the conjugate autocorrelation reveals additional cyclostationary features that are invisible to standard cyclic analysis, providing a richer feature set for distinguishing between emitters using similar modulation schemes.

• Detects impropriety in complex-valued signals
• Essential for identifying real-valued modulations after downconversion
• Doubles the available feature dimensions for fingerprinting algorithms