I/Q Filtering

Unlike real-valued filters, I/Q filters use complex coefficients to process the analytic signal directly. This enables asymmetric frequency responses around DC, allowing the filter to selectively pass the upper or lower sideband of a modulated signal without introducing image frequencies. The complex multiplication preserves the phase relationships critical for modulation recognition, making it essential for isolating signals in congested spectrum environments.

A specialized filter whose impulse response is the time-reversed, complex-conjugated version of the transmitted pulse shape. When applied to the received IQ stream, it maximizes the signal-to-noise ratio (SNR) at the optimal sampling instant. Common pulse shapes include:

• Root-Raised Cosine (RRC): Minimizes inter-symbol interference
• Gaussian: Used in GMSK modulation for Bluetooth and GSM
• Rectangular: Simplest form, used in OFDM systems

Band-pass or low-pass filtering applied to suppress energy from neighboring frequency channels that would otherwise alias into the signal of interest during decimation. A typical specification requires 60-80 dB of rejection in the adjacent channel. This is critical in software-defined radio front-ends where a wideband ADC digitizes multiple channels simultaneously, and the digital down-converter must isolate a single carrier for demodulation and classification.

Before reducing the sample rate of an IQ stream through decimation, a digital anti-aliasing filter must attenuate all frequency components above the new Nyquist rate. This is typically implemented as a CIC (Cascaded Integrator-Comb) filter followed by FIR compensation stages. The CIC provides efficient high-rate decimation, while the FIR stages correct the passband droop and sharpen the transition band, ensuring no out-of-band energy folds into the classifier's input bandwidth.

A narrow-band notch filter centered at 0 Hz in the complex baseband domain removes the DC offset caused by local oscillator leakage and mixer self-mixing. This prevents the classifier from learning spurious features from a static constellation offset. The notch bandwidth must be carefully designed to avoid attenuating legitimate low-frequency modulation components, particularly in schemes with significant DC energy like unbalanced QPSK or pilot tones.

An advanced filtering architecture that uses an LMS (Least Mean Squares) or RLS (Recursive Least Squares) algorithm to dynamically adapt filter coefficients in response to non-stationary interference. The adaptive filter takes a reference input from an auxiliary antenna or a delayed version of the primary signal to estimate and subtract the interference component. This is vital for classifying weak signals in the presence of strong co-channel interferers or intentional jamming.

A hardware impairment in direct-conversion receivers where the gain or phase relationship between the I and Q paths deviates from perfect orthogonality.

• Gain Imbalance: Amplitude mismatch between I and Q branches
• Phase Imbalance: Deviation from the ideal 90-degree phase offset
• Effect: Causes constellation warping that degrades classifier accuracy

Correction typically involves applying inverse filtering during preprocessing to restore signal orthogonality before classification.

The process of scaling IQ sample amplitudes to a standard range to prevent numerical instability during neural network training.

• Z-score Normalization: Centers data to zero mean with unit variance
• Min-Max Scaling: Maps values to a fixed range like [-1, 1]
• Per-Segment Normalization: Normalizes each inference window independently

Gain normalization is a specific variant that compensates for variable receiver gain settings, ensuring the classifier focuses on modulation structure rather than absolute signal power.

The residual frequency difference between transmitter and receiver local oscillators that causes the received IQ constellation to rotate continuously over time.

• Manifestation: A linear phase ramp across the IQ sample stream
• Impact: Destroys static constellation patterns used by classifiers
• Mitigation: I/Q centering applies a counter-rotation to shift the signal to exactly zero mean frequency

Blind CFO estimation algorithms can recover the offset without known pilot symbols, critical for non-cooperative classification scenarios.

The division of a continuous IQ stream into fixed-length, non-overlapping or overlapping segments to form individual inference examples.

• Segment Length: Typically 1024 or 2048 samples to capture sufficient modulation statistics
• Overlap Ratio: 50% overlap increases temporal resolution and dataset size
• Windowing: Applying a Hamming or Blackman taper before transformation reduces spectral leakage

Proper segmentation ensures each input tensor presented to the classifier contains a statistically meaningful snapshot of the signal.

A data regularization technique that applies realistic channel impairments to synthetic or collected IQ samples to expand training dataset diversity.

• Phase Rotation: Teaches rotational invariance by applying arbitrary angular shifts
• Noise Addition: Varies the SNR to improve robustness across channel conditions
• Fading Simulation: Applies Rayleigh or Rician multipath profiles

Synthetic I/Q generation combined with augmentation creates perfectly labeled, diverse datasets for rare or difficult-to-capture signal types.

A neural network design that processes IQ data natively as complex numbers, preserving the phase relationships inherent in the signal.

• Complex Weights: Both real and imaginary components are learned during training
• Complex Activation: Functions like modReLU operate on magnitude while preserving phase
• Advantage: Captures non-linear interactions between I and Q components that dual-channel real-valued networks may miss

Contrasts with dual-channel input where I and Q are treated as separate real-valued channels analogous to image color planes.