Synthetic Signal Generation

Synthetic data must replicate real-world propagation physics to be useful. Key impairments include:

• Additive White Gaussian Noise (AWGN): Thermal noise with a flat power spectrum, the baseline impairment for testing at varying Signal-to-Noise Ratios (SNR)
• Multipath Fading: Rayleigh and Rician fading models simulate signals arriving via multiple paths with different delays, causing inter-symbol interference
• Carrier Frequency Offset (CFO): Simulates oscillator mismatch between transmitter and receiver, typically modeled as a phase rotation in the complex baseband
• Sample Clock Offset: Mimics timing drift between ADC sampling clocks, critical for testing real-world receiver robustness

Without these impairments, classifiers learn unrealistic signal features and fail in deployment.

The synthetic generation pipeline produces perfectly labeled training data, eliminating the costly and error-prone manual annotation required for real captured signals. The process follows:

• Modulation Scheme Selection: Define the target classes (BPSK, QPSK, 16-QAM, 64-QAM, GMSK, etc.)
• Bit Sequence Generation: Create random or pseudorandom binary data streams as the information payload
• Baseband Waveform Synthesis: Apply the mathematical modulation mapping to produce IQ samples in complex baseband
• Channel Impairment Application: Pass the clean signal through simulated channel models with configurable SNR, fading, and offset parameters
• Dataset Serialization: Store as labeled tensors in formats like HDF5 or TFRecord for efficient training pipeline ingestion

This pipeline can generate millions of labeled examples in hours, covering SNR ranges and channel conditions rarely captured in the field.

Synthetic data generation often pairs with Software-Defined Radio (SDR) hardware for hybrid workflows:

• Over-the-Air (OTA) Transmission: Generated waveforms are transmitted through actual SDR hardware (e.g., Ettus USRP, HackRF) and captured by a receiver, introducing real hardware impairments like IQ imbalance and phase noise
• Channel Sounding Integration: Real channel impulse responses captured in target environments (urban, indoor, vehicular) are convolved with synthetic waveforms for site-specific training data
• Hardware-in-the-Loop Testing: Classifiers trained on purely synthetic data are validated against OTA-captured signals to measure the sim-to-real transfer gap

This hybrid approach bridges the gap between mathematical models and real-world hardware behavior.

Two distinct approaches exist for expanding training datasets:

Full Synthetic Generation:

• Creates waveforms entirely from mathematical models
• Provides complete control over all signal parameters
• Ideal for exploring edge cases and rare modulation schemes
• Risk: May not capture all real-world hardware artifacts

Data Augmentation:

• Applies label-preserving transformations to existing real or synthetic samples
• Common augmentations: phase rotation, time shifting, amplitude scaling, frequency translation
• Preserves underlying signal structure while increasing diversity
• Acts as a regularizer to reduce overfitting

Most production pipelines combine both: full synthesis for volume, augmentation for robustness.

Domain randomization deliberately varies simulation parameters beyond realistic ranges to force classifiers to learn invariant features:

• SNR Randomization: Train across an extreme range (-30 dB to +50 dB) so the model learns modulation structure independent of noise level
• Symbol Rate Variation: Randomize the samples-per-symbol parameter to prevent the classifier from overfitting to a specific timing
• Pulse Shaping Diversity: Mix raised-cosine, Gaussian, and rectangular pulse shapes to force the model to ignore filter-specific artifacts
• Phase and Frequency Randomization: Apply uniform random phase offsets and frequency shifts to make the classifier invariant to synchronization state

This technique, borrowed from robotics sim-to-real transfer, produces classifiers that generalize to unseen channel conditions and receiver hardware.

Real-world signal environments exhibit severe class imbalance—common modulations like QPSK appear far more frequently than rare schemes like 32-APSK. Synthetic generation addresses this:

• Balanced Dataset Creation: Generate equal numbers of examples for every modulation class, preventing the classifier from developing bias toward common schemes
• Rare Signal Oversampling: For emerging or threat-specific modulations with no real-world captures, synthesis is the only viable data source
• Open Set Training: Generate synthetic 'unknown' signal classes by applying random modulation parameters, training the classifier to recognize when it encounters a truly novel signal
• Curriculum Learning: Start training with high-SNR, easy examples, then progressively introduce harder, lower-SNR samples

This control over class distribution is impossible with purely captured datasets.

The mathematical simulation of real-world propagation effects to ensure synthetic data mirrors physical reality. Key impairments include:

• Additive White Gaussian Noise (AWGN): Models thermal noise with a flat power spectrum.
• Multipath Fading: Simulates signal reflections using tapped-delay-line models like Rayleigh or Rician fading.
• Carrier Frequency Offset (CFO): Introduces a mismatch between transmitter and receiver oscillators.
• Phase Noise: Models the short-term random frequency fluctuations of a local oscillator.

The use of programmable hardware or software libraries to generate and capture complex baseband signals. Tools like GNU Radio and MATLAB's Communications Toolbox allow for the precise construction of transmission pipelines. This approach enables the creation of deterministic, repeatable datasets where every parameter—from pulse shaping filter roll-off to sample rate—is known and controlled, providing perfect ground truth labels for supervised learning.

A regularization strategy that applies label-preserving transformations to synthetically generated or real signals to artificially expand dataset diversity. Common augmentations include:

• Phase Rotation: Multiplying the complex IQ signal by a random complex exponential.
• Time Shifting: Applying a random circular shift to the sample sequence.
• Amplitude Scaling: Varying the signal gain to simulate different path losses.
• Additive Interference: Injecting narrowband tones or adjacent channel signals.

A structured approach to building datasets that systematically varies parameters to prevent model overfitting to a single condition. A generation script typically loops through:

1. Modulation Schemes: BPSK, QPSK, 16-QAM, 64-QAM, etc.
2. SNR Ranges: From -20 dB to +30 dB in discrete steps.
3. Channel Profiles: Static, pedestrian, vehicular fading models. This creates a combinatorial grid of labeled signal examples, ensuring comprehensive coverage of the operational domain.

An advanced technique where the parameters of the synthetic signal generation environment are randomized within wide, realistic bounds during training. By exposing the deep learning classifier to extreme variability in symbol rate, pulse shaping, and clock skew, the model is forced to learn the fundamental geometric structure of the modulation rather than spurious correlations with a specific simulator configuration. This significantly improves transfer learning to real-world hardware.