I/Q Preprocessing

Scales the amplitude of an IQ sample stream to a standard range to prevent numerical instability during neural network training. Common techniques include Z-score normalization (zero mean, unit variance) and min-max scaling (mapping to [-1, 1]). Without normalization, variable receiver gain settings and path loss differences cause large dynamic range variations that dominate the loss function, forcing the network to waste capacity learning gain compensation rather than modulation features.

• Per-segment normalization computes statistics over each individual inference segment
• Global normalization uses statistics derived from the entire training corpus
• Essential for batch normalization layers to function effectively

Shifts the complex baseband signal to exactly zero mean frequency by removing residual Carrier Frequency Offset (CFO). CFO causes the received IQ constellation to rotate continuously over time, smearing the distinct phase states that define modulation schemes like QPSK and 16-QAM. Centering algorithms estimate the residual rotation rate and apply a counter-rotating complex exponential to each sample.

• Enables the classifier to observe stationary constellation points
• Typically performed after coarse frequency correction in the RF front-end
• Critical for phase-sensitive modulation classification (PSK, QAM)

A digital signal processing block that applies inverse filtering to compensate for hardware non-idealities in direct-conversion receivers. The two primary impairments are I/Q imbalance—where the gain or phase relationship between the I and Q paths deviates from perfect orthogonality—and DC offset, a constant voltage bias that manifests as a non-zero mean. Correction algorithms estimate the impairment parameters from the received signal and apply a compensatory matrix transformation.

• Restores constellation symmetry distorted by gain mismatch
• Eliminates the LO leakage spike at DC caused by offset
• Typically implemented as a complex FIR filter with adaptively updated coefficients

Divides a continuous IQ stream into fixed-length, non-overlapping or overlapping segments to form individual inference examples for a modulation recognition model. Segment length is a critical hyperparameter: too short and the segment lacks sufficient symbols for reliable classification; too long and it introduces latency and may span channel coherence time boundaries.

• Typical segment lengths range from 128 to 1024 complex samples
• Overlapping segments (50% overlap) increase training dataset size and provide temporal context
• Must align with the input tensor dimensions expected by the neural network architecture

A data regularization technique that applies realistic channel impairments to synthetic or collected IQ samples to expand training dataset diversity. Augmentations include phase rotation (teaching rotational invariance), additive white Gaussian noise (AWGN) at varying SNR levels, frequency offset injection, and fading profile application. This prevents overfitting to specific channel conditions present in the training set and improves generalization to unseen environments.

• Phase rotation augmentation: applies random angular shifts to teach the classifier that modulation identity is rotation-invariant
• SNR augmentation: mixes samples with calibrated noise levels to cover the expected operational SNR range
• Enables single-dataset training that generalizes across multiple receiver deployments

Changes the sample rate of an IQ stream through decimation (reducing rate) or interpolation (increasing rate) to match the native input requirements of a downstream neural network classifier. Mismatched sample rates between the receiver front-end and the model's expected input cause dimensional errors or, worse, misrepresent the signal's spectral content. Resampling typically uses polyphase filter banks to preserve signal fidelity while avoiding aliasing artifacts.

• Decimation includes an anti-aliasing low-pass filter stage before downsampling
• Interpolation inserts zeros and applies a reconstruction filter
• Enables a single trained model to accept inputs from heterogeneous receiver hardware with different native sample rates

The process of scaling raw IQ sample amplitudes to a standard range to ensure numerical stability during neural network training. Without normalization, models become biased toward high-power signals rather than learning modulation structure.

• Z-score normalization: Centers data to zero mean and unit variance
• Min-max scaling: Constrains values to a fixed range like [-1, 1]
• Per-segment normalization: Applied independently to each inference window
• Prevents gradient explosion in deep networks during backpropagation

A critical preprocessing step that removes residual frequency differences between transmitter and receiver local oscillators. Uncorrected CFO causes the entire IQ constellation to rotate continuously, making modulation classification nearly impossible.

• I/Q Centering shifts the signal to exactly zero mean frequency
• CFO estimation algorithms analyze phase rotation rate across samples
• Essential for higher-order QAM schemes with tight constellation spacing
• Residual offset above 1% of symbol rate significantly degrades classifier accuracy

Digital correction for hardware impairments in direct-conversion receivers where the I and Q paths deviate from perfect orthogonality. Gain imbalance and phase imbalance create elliptical constellation distortions that confuse classifiers.

• Blind estimation techniques derive correction coefficients from the signal itself
• Pilot-based methods use known reference symbols for precise calibration
• Imbalance creates mirror-frequency interference in the baseband spectrum
• Modern deep learning classifiers can learn to be robust to mild imbalance without explicit correction

The division of continuous IQ streams into fixed-length segments suitable for neural network input tensors. Segmentation defines the temporal context window, while windowing controls spectral leakage during frequency-domain transformations.

• Typical segment lengths: 128, 256, 512, or 1024 complex samples
• Overlapping segments increase training data density at inference cost
• Hamming and Blackman windows suppress sidelobes in spectrogram generation
• Segment duration must capture at least several symbol periods for reliable classification

A data regularization technique that applies realistic channel impairments to training samples, dramatically expanding dataset diversity without additional collection. Phase rotation, AWGN addition, and fading simulation teach classifiers invariance to common distortions.

• Phase rotation augmentation teaches rotational invariance to the classifier
• Noise augmentation improves robustness across SNR ranges
• Rayleigh and Rician fading models simulate real-world multipath
• Critical for few-shot learning scenarios with limited real-world captures

The rate conversion process that matches incoming IQ streams to the native sample rate expected by downstream neural networks. Decimation reduces sample rate through low-pass filtering and downsampling; interpolation increases it through upsampling and filtering.

• Rational resampling factors handle non-integer rate conversions
• Polyphase filter banks enable efficient hardware implementation
• Mismatched sample rates cause symbol timing errors and classification failures
• Modern SDR frontends often provide configurable output rates to avoid resampling