I/Q Normalization

Transforms the IQ stream to have zero mean and unit variance by subtracting the mean and dividing by the standard deviation. This centers the constellation and scales amplitude to a consistent statistical range.

• Formula: x_norm = (x - μ) / σ
• Preserves relative amplitude relationships between samples
• Robust to outliers when computed over large sample windows
• Default choice for complex-valued neural network inputs

Linearly maps IQ sample amplitudes to a fixed interval, typically [-1, 1] or [0, 1], by dividing by the maximum absolute value in the segment.

• Formula: x_norm = x / max(|x|)
• Guarantees bounded input range for activation functions like tanh
• Sensitive to impulsive noise spikes that compress dynamic range
• Common in FPGA deployments where fixed-point arithmetic is used

Per-segment normalization computes statistics independently for each IQ segment, adapting to varying receiver gain. Global normalization uses fixed statistics derived from the entire training dataset.

• Per-segment: Handles dynamic range variation but removes absolute power information
• Global: Preserves power as a discriminative feature but assumes consistent gain
• Hybrid approaches compute running statistics with exponential moving averages for streaming inference

Treats the I and Q components as a single complex number rather than two independent real channels, preserving the circular symmetry of the constellation.

• Computes variance as E[|x|²] rather than separate I and Q variances
• Avoids elliptical distortion of the constellation shape
• Essential for phase-modulated signals like QPSK and 8-PSK
• Requires complex-aware neural network layers downstream

Scales the IQ segment to a target average power level, typically unit power (1.0) or 0 dB, by dividing by the root mean square (RMS) amplitude.

• Formula: x_norm = x / sqrt(E[|x|²])
• Preserves the constellation shape exactly while normalizing energy
• Critical for modulation classifiers that use amplitude-sensitive features
• Often combined with centering to remove DC offset simultaneously

Unnormalized IQ samples with large dynamic range cause exploding gradients and saturated activations during backpropagation, stalling neural network training.

• Prevents ReLU death from excessively large positive inputs
• Keeps softmax inputs within numerically stable ranges
• Enables consistent learning rates across different datasets and SNR conditions
• Reduces sensitivity to weight initialization schemes

A specific amplitude scaling technique that compensates for variable receiver gain settings. Unlike standard Z-score normalization, gain normalization explicitly divides the IQ stream by an estimate of the receiver's linear gain factor, ensuring the classifier focuses on the modulation's structural features rather than absolute signal power. This is critical when training data is collected from heterogeneous hardware with different automatic gain control (AGC) profiles.

A preprocessing operation that shifts the complex baseband signal to exactly zero mean frequency by removing residual Carrier Frequency Offset (CFO). Without centering, the constellation rotates continuously, destroying the static geometric features classifiers rely on. Centering estimates the residual frequency by analyzing the phase rotation rate and applies a counter-rotation to each IQ sample, centering the constellation in the complex plane before normalization.

A constant voltage bias added to the true signal in the analog front-end, manifesting as a non-zero mean in the IQ sample stream. This offset shifts the entire constellation away from the origin, distorting amplitude-based modulation features. Removal is typically performed by subtracting the empirical mean of the IQ segment before normalization, ensuring zero-centered data for subsequent Z-score scaling.

A hardware impairment in direct-conversion receivers where the gain or phase relationship between the I and Q paths deviates from perfect orthogonality. This causes constellation warping—circles become ellipses, squares become parallelograms. Correction applies an inverse mixing matrix to restore orthogonality. This step must precede normalization; otherwise, the scaling will lock in the distorted constellation shape.

A data regularization technique that applies realistic channel impairments—including phase rotation, noise addition, and fading—to synthetic or collected IQ samples. Normalization interacts critically with augmentation: applying random gain variations before normalization teaches the classifier to be invariant to amplitude, while phase rotations applied after centering enforce rotational invariance. The ordering of augmentation and normalization in the pipeline directly impacts model robustness.

The division of a continuous IQ stream into fixed-length, non-overlapping or overlapping segments to form individual inference examples. Normalization is typically applied per-segment rather than globally across the entire stream. This local normalization ensures that each inference example is independently scaled, preventing slow power drifts over time from degrading classification accuracy on later segments.