Data Augmentation

The most fundamental augmentation technique involves adding controlled levels of synthetic thermal noise to pristine signals. By training on samples with a dynamic range of Signal-to-Noise Ratios (SNRs)—typically from -10 dB to +30 dB—the classifier learns to extract modulation features that are robust to background interference. This directly prevents the model from relying on high-SNR artifacts that disappear in real-world, noisy deployments. The noise power is scaled relative to the signal power to preserve the desired SNR distribution.

Beyond simple noise, realistic channel models are applied to simulate the physics of wireless propagation. Key transformations include:

• Multipath Fading: Applying Rayleigh or Rician fading profiles to mimic reflections and signal scattering.
• Frequency Offset: Introducing a slight carrier frequency offset (CFO) to simulate oscillator mismatch between transmitter and receiver.
• Phase Rotation: Adding a constant or slowly varying phase shift to simulate the random phase of the local oscillator. These combined impairments force the network to learn representations invariant to the stochastic channel state.

Label-preserving transformations are applied directly to the IQ sample sequence or the constellation diagram. For raw IQ streams, techniques include small random time shifts (simulating synchronization errors) and amplitude scaling (simulating automatic gain control variance). For constellation-based classifiers, geometric augmentations like slight rotations, translations, and scaling of the point cloud are used. These teach the model that the relative geometry of the symbol states, not their absolute position, defines the modulation scheme.

To harden classifiers against congested spectrum environments, augmentation pipelines mix the target signal with synthetic interference. This includes:

• Co-Channel Interference: Adding a second, lower-power modulated signal on the same frequency.
• Adjacent Channel Interference: Simulating spectral leakage from a nearby carrier.
• Tone Interference: Injecting narrowband continuous wave (CW) tones. Training on these composite signals prevents the model from catastrophically failing when encountering real-world spectral overlap.

A critical augmentation step for bridging the sim-to-real gap involves modeling non-ideal hardware effects. This includes adding IQ imbalance (gain and phase mismatch between the I and Q branches), power amplifier non-linearity (clipping and spectral regrowth), and phase noise from local oscillators. By training on signals distorted with these specific hardware fingerprints, the classifier learns to ignore device-specific artifacts and focus on the invariant modulation format, improving cross-receiver generalization.

Advanced interpolation-based augmentation strategies adapted from computer vision. Mixup creates new training samples by taking a convex combination of two random IQ signal vectors and their corresponding one-hot encoded labels. CutMix replaces a contiguous time segment of one signal with a segment from another, blending the labels proportionally. These techniques act as strong regularizers, encouraging the network to behave linearly between training examples and reducing overfitting to spurious correlations in the limited original dataset.

The programmatic creation of labeled RF training data using mathematical channel models and software-defined radio simulations. This is the primary source of augmented samples, enabling the generation of infinite signal variations.

• Simulates AWGN, multipath fading, and carrier frequency offset
• Produces perfectly labeled IQ samples without costly over-the-air collection
• Enables training on rare or classified modulation schemes

A modeling failure where a neural network memorizes the noise and specific artifacts of the training set rather than learning the underlying modulation signatures. Data augmentation is the primary defense against this.

• Manifests as high training accuracy but poor validation performance
• Particularly dangerous with small, real-world RF datasets
• Augmentation injects label-preserving variance to force generalization

The fundamental channel impairment model representing thermal noise with a flat power spectral density. Injecting varying levels of AWGN is the most common data augmentation technique for training robust classifiers.

• Trains models to operate across a wide SNR range
• Prevents the classifier from relying on unrealistic, noise-free signal features
• Combined with phase rotation and frequency offset for comprehensive augmentation

A transfer learning technique that bridges the distribution gap between synthetic training data and real-world hardware captures. Augmentation strategies are tuned to mimic the specific impairments of target receiver hardware.

• Addresses mismatches in IQ imbalance, DC offset, and sampling rate
• Augmentation acts as a crude form of domain randomization
• Reduces the need for expensive labeled data from each deployment environment

A self-supervised framework where augmented versions of the same signal are pulled together in the embedding space while different signals are pushed apart. Data augmentation is the engine that creates the positive pairs.

• Learns robust representations without explicit modulation labels
• Augmentations define the invariance properties the model will acquire
• Enables few-shot classification of rare signal types

Preprocessing and model-based techniques for mitigating fading, noise, and offset effects before classification. Augmentation during training makes the classifier inherently robust, reducing the need for explicit compensation blocks.

• Blind equalization and carrier recovery as alternatives to augmentation
• Augmentation-trained models learn to disentangle modulation from channel effects
• Critical for real-time systems where preprocessing latency is constrained