Cyclostationary Feature Extraction

The Spectral Correlation Density is the fundamental two-dimensional transform that reveals cyclostationarity. It measures the correlation between spectral components separated by a cyclic frequency (α).

• Domain: Bi-frequency plane (f, α)
• Key Property: Stationary noise exhibits correlation only at α = 0, while modulated signals show distinct peaks at non-zero α corresponding to symbol rate, carrier frequency, and guard intervals.
• Computation: Typically estimated via the FAM (FFT Accumulation Method) or SSCA (Strip Spectral Correlation Analyzer) for computational efficiency.

The Cyclic Autocorrelation Function is the time-domain counterpart to the SCD, defined as the Fourier coefficient of the time-varying autocorrelation. It quantifies the correlation between a signal and a frequency-shifted version of itself.

• Quadratic Transformation: Converts a signal into a function of lag (τ) and cyclic frequency (α).
• Blind Estimation: Peaks in the CAF magnitude directly reveal the symbol rate and carrier frequency offset without prior demodulation.
• Robustness: CAF-based features are inherently resilient to stationary Gaussian noise and narrowband interference.

Higher-order cyclic cumulants capture the non-Gaussian statistical behavior of modulated signals, extending analysis beyond second-order statistics. They are critical for classifying signals with identical power spectra.

• Order Selection: 4th-order cumulants differentiate QPSK from 16-QAM; 6th-order cumulants separate 16-QAM from 64-QAM.
• Phase Sensitivity: Cyclic cumulants preserve phase information, enabling discrimination between modulation families with identical cyclic frequencies.
• Noise Immunity: Gaussian noise has zero cumulants above 2nd order, making higher-order cyclic cumulants theoretically immune to colored Gaussian interference.

A Cyclic Domain Profile is a compressed, one-dimensional feature vector derived by integrating the SCD magnitude along the spectral frequency axis for each cyclic frequency. It serves as a compact, highly discriminative signature.

• Dimensionality Reduction: Collapses the 2D SCD into a 1D vector indexed by α, suitable for lightweight classifiers.
• Key Peaks: Distinct peaks appear at α = k/T_s (symbol rate harmonics) and α = 2f_c ± k/T_s (carrier-related features).
• Application: Widely used in Specific Emitter Identification (SEI) as a hardware fingerprint that captures unique transmitter imperfections.

Cyclostationary signals exhibit two distinct types of correlation, and exploiting both is essential for complete feature extraction.

• Non-Conjugate CAF: Standard autocorrelation E{x(t)x*(t-τ)}. Reveals features at cyclic frequencies related to the symbol rate (k/T_s).
• Conjugate CAF: Involves the product E{x(t)x(t-τ)} without conjugation. Reveals features at cyclic frequencies related to doubled carrier frequency (2f_c + k/T_s).
• Discrimination Power: Signals with identical non-conjugate profiles (e.g., BPSK and QPSK at the same symbol rate) can be distinguished by their conjugate cyclostationary signatures.

Practical extraction requires efficient algorithms to estimate cyclostationary features from finite, discrete-time samples.

• FAM (FFT Accumulation Method): A computationally efficient SCD estimator using a channelizer and FFT. Complexity: O(N² log N).
• SSCA (Strip Spectral Correlation Analyzer): An alternative estimator optimized for real-time processing with lower memory requirements.
• Time-Smoothing: Direct averaging of cyclic periodograms over time, suitable for streaming architectures.
• Deep Learning Integration: Modern approaches use Complex-Valued Neural Networks to learn cyclostationary feature extractors directly from raw I/Q samples, bypassing explicit SCD computation.

The fundamental analysis plane for cyclostationary signals. The SCD is a two-dimensional function, S_x^α(f), that measures the spectral correlation between frequency components separated by α/2.

• α (Cycle Frequency): The separation between correlated spectral components, directly linked to the signal's symbol rate, carrier frequency, or pulse shaping.
• f (Spectral Frequency): The standard frequency axis.
• Noise Rejection: Stationary noise exhibits no spectral correlation (S_x^α(f) = 0 for α ≠ 0), making SCD inherently robust to interference.

The time-domain equivalent of the SCD, defined as R_x^α(τ) = ⟨x(t+τ/2)x(t-τ/2)e^-j2παt⟩*. It quantifies the periodic time-varying autocorrelation of a signal.

• Quadratic Transformation: Converts a modulated signal into a sine wave at the cycle frequency α, enabling detection with simple Fourier analysis.
• Delay Parameter (τ): Reveals the correlation structure of the pulse-shaping filter.
• Conjugate CAF: An alternative form, R_xx^α(τ)*, is critical for detecting signals with non-circular constellations like BPSK.

The set of cycle frequencies α forms a unique fingerprint for each modulation scheme. Extracting these discrete periodicities is the primary goal of cyclostationary analysis.

• Symbol Rate: For linearly modulated signals (QAM, PSK), a strong cycle frequency appears at α = 1/T_s.
• Carrier Frequency: A cycle frequency at α = 2f_c is generated by the squaring operation, useful for blind carrier recovery.
• Chip Rate: For spread-spectrum signals (DSSS), α equals the chip rate, enabling detection below the noise floor.

The most computationally efficient algorithm for estimating the SCD. The FAM transforms the brute-force 2D correlation into a series of 1D FFT operations.

• Channelization: Input data is first processed by a bank of bandpass filters.
• Decimation: Downsampling after channelization drastically reduces the data rate for subsequent FFTs.
• Trade-off: Balances cycle frequency resolution (Δα) against spectral frequency resolution (Δf) via the channelizer design, governed by the uncertainty principle.

Extends analysis beyond second-order moments to kth-order cumulants and polyspectra. HOCS is essential when second-order features are absent or identical between signals.

• Gaussian Noise Suppression: Third-order and higher cumulants are theoretically zero for Gaussian processes, providing extreme noise immunity.
• Non-Linearity Detection: Detects signal features generated by power amplifier compression that are invisible to the SCD.
• Modulation Differentiation: Can separate QPSK from 16QAM, which share identical second-order cycle frequencies but differ in their fourth-order cumulant cyclic spectra.

A specific application of cyclostationarity for Orthogonal Frequency Division Multiplexing signals. The intentional repetition of the cyclic prefix (CP) induces a strong correlation peak.

• Correlation Lag: The peak occurs at a lag τ equal to the useful symbol duration T_u.
• Cycle Frequency: The periodicity is at α = 1/(T_u + T_cp), the full OFDM symbol rate.
• Blind Parameter Estimation: This allows an interceptor to estimate the FFT size and CP length of an unknown OFDM transmitter without demodulation.