Data Augmentation for AMC

The foundational augmentation that injects controlled levels of thermal noise into clean I/Q samples. By training across a range of signal-to-noise ratios (SNRs) , the model learns to disregard background noise floor variations.

• Mechanism: Adds a complex-valued noise vector with zero mean and configurable variance to the original signal.
• Benefit: Prevents the classifier from relying on unrealistic, noise-free signal representations.
• Implementation: Typically sweeps from -20 dB to +30 dB SNR during training to cover both harsh and pristine channel conditions.

Simulates the carrier frequency offset (CFO) and phase asynchrony inherent between independent transmitter and receiver local oscillators. This forces the model to learn modulation features that are invariant to constellation rotation.

• Mechanism: Multiplies the complex I/Q samples by e^(j*θ), where θ is a random phase drawn uniformly or linearly increasing over time.
• Critical for Blind AMC: Without this, a model will catastrophically fail when encountering a signal with even a slight frequency mismatch.
• Combined Effect: Often applied simultaneously with a small frequency shift to mimic Doppler effects in mobile environments.

Emulates the destructive and constructive interference patterns caused by signal reflections in urban or indoor environments. This augmentation teaches the model to handle severe frequency-selective fading.

• Mechanism: Convolves the signal with a time-varying channel impulse response modeled as a complex Gaussian process.
• Realism: Transforms a simple line-of-sight training sample into a realistic, distorted waveform with deep spectral nulls.
• Generalization: Critical for deploying AMC in non-stationary environments where the channel coherence time is short.

Randomly shifts the I/Q sequence in time to simulate the unknown burst start time of intercepted signals. Combined with slight resampling, it mimics symbol timing offset and hardware clock inaccuracies.

• Mechanism: Applies a random circular or zero-padded shift to the sample buffer, and optionally resamples the signal at a slightly different rate.
• Robustness: Prevents the model from overfitting to a specific transient preamble or a fixed sampling instant.
• Practical Use: Essential for transitioning from simulated datasets like RadioML to real-world captures where perfect synchronization is impossible.

Models the saturation effects of a power amplifier (PA) operating near its compression point. This augmentation introduces realistic hardware impairments that distort the outer constellation points.

• Mechanism: Applies a Rapp or Saleh model to the signal amplitude, introducing AM/AM and AM/PM distortion.
• Feature Preservation: Forces the classifier to rely on phase transitions and cyclostationary features rather than perfect constellation geometry.
• EW Context: Critical for electronic warfare applications where intercepted signals often originate from low-cost, non-linear transmitters.

A regularization technique that creates virtual training examples by linearly interpolating between two random I/Q samples and their corresponding labels. This encourages smoother decision boundaries in the high-dimensional signal space.

• Mechanism: Generates a new sample x̃ = λx_i + (1-λ)x_j, where λ is sampled from a Beta distribution.
• Adversarial Robustness: Improves model stability against small, adversarial perturbations that might otherwise flip the classification result.
• Implementation: Works best when combined with domain-specific augmentations like fading rather than as a standalone technique.

A geometric transformation that multiplies complex I/Q samples by a unit-magnitude complex exponential, effectively rotating the constellation diagram by a random angle. This simulates the effect of an unresolved Carrier Frequency Offset (CFO) and phase jitter from local oscillator drift. By applying uniform random rotations between 0 and 2π, the model learns to recognize modulation schemes independently of their absolute phase alignment, a critical requirement for blind modulation recognition systems that lack carrier synchronization.

A composite augmentation pipeline that applies realistic multipath fading profiles (e.g., Rayleigh, Rician) and Doppler shifts to training samples. This transforms static, lab-generated signals into dynamic representations that mimic real-world mobile environments. The technique often uses tapped-delay-line models to simulate frequency-selective fading, teaching the deep learning AMC model to handle inter-symbol interference without requiring explicit blind equalization as a separate preprocessing step during inference.

A spectral augmentation technique that applies random masks to the time-frequency representation of a signal, such as a spectrogram or cyclostationary profile. By zeroing out contiguous blocks of time steps or frequency bins, the model is forced to rely on partial, non-contiguous evidence for classification. This prevents overfitting to narrowband artifacts or transient interference patterns and is particularly effective for improving open-set recognition performance when encountering novel jamming waveforms or unknown interference types.