I/Q Augmentation

Applies realistic physical-layer distortions directly to the complex baseband samples. This forces the classifier to learn invariances to channel effects rather than memorizing pristine signal representations.

• Additive White Gaussian Noise (AWGN): Adds thermal noise across a range of SNR values
• Phase Rotation: Multiplies samples by e^(jθ) to simulate Carrier Frequency Offset (CFO) and oscillator drift
• Fading Profiles: Convolves the signal with Rayleigh or Rician impulse responses to simulate multipath
• I/Q Imbalance: Introduces gain mismatch (ε) and phase error (φ) between I and Q branches

Augmentation is applied on-the-fly during training using randomized parameter ranges, ensuring the model never sees the exact same example twice. This acts as a powerful regularizer against overfitting.

• Random SNR Sampling: Uniformly draws noise power from a defined range per batch
• Random Phase Offset: Applies a uniform random rotation between 0 and 2π to each segment
• Time Stretching: Resamples the IQ stream with small random variations to simulate sample clock offset
• Random Cropping: Extracts variable-length segments from longer recordings to improve temporal generalization

Augmentation is the critical bridge between purely synthetic I/Q training data and real-world deployment. By applying measured channel characteristics to simulated waveforms, models trained entirely in simulation can generalize to over-the-air signals.

• Hardware-in-the-Loop Profiling: Extracts real impairment parameters from target receivers to replicate in augmentation
• Non-Linear Distortion: Models power amplifier compression curves and applies them to synthetic samples
• Adjacent Channel Interference: Injects filtered, frequency-shifted copies of other signals to simulate congested spectrum

All augmentations are mathematically constrained to preserve the underlying modulation label. A QPSK signal with phase rotation and noise remains a QPSK signal. This distinguishes augmentation from adversarial perturbation.

• Modulation Order Invariance: Phase shifts do not alter the modulation family
• Amplitude Scaling: Uniform gain changes preserve the relative constellation geometry
• Frequency Offset Limits: CFO is bounded to prevent symbol boundary collapse
• Noise Floor Ceiling: SNR is kept above a threshold where the modulation becomes unrecognizable

When using complex-valued neural networks, augmentations must operate in the complex domain to preserve the algebraic relationships between I and Q components. Real-valued augmentations applied separately to I and Q channels can destroy phase coherence.

• Complex Gaussian Noise: Adds circularly symmetric complex noise n ~ CN(0, σ²)
• Unitary Transformations: Applies complex rotation matrices that preserve inner products
• Complex Convolution: Performs channel simulation using complex-valued FIR filters
• Polar Coordinate Augmentation: Operates directly on instantaneous amplitude and phase rather than Cartesian I/Q

The intensity of augmentation is dynamically adjusted throughout training. Early epochs use mild augmentation to learn basic features; later epochs introduce severe distortions to force robust representations.

• SNR Curriculum: Starts at high SNR (30dB) and progressively decreases to low SNR (0dB)
• Impairment Stacking: Begins with single impairments and gradually combines multiple effects
• Difficulty Sampling: Actively selects augmentation parameters that produce higher loss, focusing training on hard examples
• Validation on Clean Data: Always evaluates on unaugmented holdout sets to measure true generalization

The process of applying mathematical models of fading, multipath, and noise to a clean synthetic IQ signal to replicate real-world wireless propagation distortions. This is the engine behind realistic augmentation.

• Rayleigh/Rician Fading: Models multi-path scattering
• Doppler Shift: Simulates relative motion between Tx/Rx
• Delay Spread: Introduces inter-symbol interference

Without accurate channel simulation, augmented datasets fail to represent the statistical properties of over-the-air signals, leading to brittle classifiers.

A deliberate or channel-induced angular shift applied uniformly to all IQ samples in a segment. It is one of the most fundamental augmentation operations.

• Teaches classifiers rotational invariance
• Compensates for lack of carrier synchronization in real-world captures
• Applied as: samples * exp(j * theta) where theta is drawn from a uniform distribution

A model that cannot handle arbitrary phase rotation will fail immediately on any signal not perfectly centered in the complex plane.

A fundamental channel model that adds a random noise signal with a flat spectral density and Gaussian amplitude distribution to the IQ stream, simulating thermal noise in the receiver.

• Parameterized by Signal-to-Noise Ratio (SNR) in dB
• Augmentation sweeps across an SNR range (e.g., -10 dB to +30 dB)
• Teaches the classifier to extract modulation features buried in noise

Training on a single SNR produces a model that overfits to that noise floor and fails at other operating points.

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

• Gain Imbalance: I and Q branches have different amplification
• Phase Imbalance: The 90-degree offset between I and Q is not exact
• Manifests as an elliptical stretching of the constellation

Augmenting with controlled I/Q imbalance forces the classifier to become robust to low-cost receiver front-ends.

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

• Measured in parts-per-million (ppm) of the carrier frequency
• Causes a linear phase ramp across the sample segment
• Distinct from a static phase rotation; CFO is a time-varying effect

Augmentation with random CFO values teaches the model to recognize modulation structure independent of absolute frequency alignment.

Artificially generated IQ samples created through software simulation of modulation and channel models, providing a cost-effective source of perfectly labeled training data for rare signal types.

• Enables infinite dataset scaling for uncommon modulations
• Labels are deterministic and error-free
• Can generate edge cases that are dangerous or impossible to capture in the field

Synthetic data is the raw material; augmentation is the process that makes it statistically representative of real-world conditions.