I/Q Dataset

The dataset must contain a balanced representation of target modulation classes, from simple BPSK to high-order 256-QAM. Class imbalance leads to biased classifiers that over-predict common schemes. A robust dataset includes both linear (PSK, QAM) and non-linear (CPM, GMSK) modulations to prevent the model from learning spurious correlations.

Samples must span a wide SNR range, typically from -20 dB to +30 dB, to train classifiers that are robust in both noisy and clean channel conditions. Low-SNR samples teach the model to extract features from noise-like signals, while high-SNR samples provide clean reference constellations. A common pitfall is training only on high-SNR data, causing catastrophic failure in real-world low-SNR environments.

Realistic datasets apply physics-based channel simulations to synthetic or collected IQ streams. Key impairments include:

• Multipath fading: Rician and Rayleigh fading profiles that cause frequency-selective distortion
• Carrier Frequency Offset (CFO): Residual rotation of the constellation
• Sample Timing Offset: Misalignment of the optimal sampling instant
• Phase Noise: Random phase jitter from local oscillators Without these, a classifier learns an idealized version of the signal that does not exist in the field.

Each IQ segment must be paired with a ground-truth label and rich metadata. Beyond the modulation type, metadata should capture the exact SNR, sample rate, center frequency, and channel model parameters used during generation or collection. This enables stratified evaluation, where performance is analyzed per impairment condition rather than as a single aggregate metric.

Dataset splits must be stratified by modulation type and SNR to ensure each subset contains a proportional representation of all classes. A naive random split can create a test set that is easier or harder than the training set, invalidating benchmark results. For channel-impaired data, group-based splitting by recording session prevents data leakage where nearly identical samples appear in both training and test sets.

Synthetic datasets are generated via software simulation, offering perfect labels and infinite scalability but often lack the hardware-specific impairments of real receivers. Over-the-Air (OTA) datasets are captured from physical transmitters and receivers, providing authentic hardware fingerprints and environmental effects but requiring costly labeling. Production-grade datasets often blend both, using synthetic data for pre-training and OTA data for fine-tuning.

The sequence of signal conditioning steps applied to raw IQ samples to create a standardized input tensor. Critical operations include:

• I/Q Normalization: Scaling amplitude to prevent numerical instability during training
• I/Q Centering: Removing residual Carrier Frequency Offset (CFO) to center the constellation
• I/Q Correction: Applying inverse filtering to compensate for hardware impairments like I/Q imbalance and DC offset

Without proper preprocessing, classifiers learn hardware artifacts rather than modulation structure.

A data regularization technique that applies realistic channel impairments to expand training dataset diversity. Common augmentations include:

• Phase Rotation: Teaching classifiers rotational invariance by applying uniform angular shifts
• Noise Addition: Injecting Additive White Gaussian Noise (AWGN) at varying SNR levels
• Fading Simulation: Applying multipath and Doppler models to mimic mobile environments

Augmentation is essential when real-world collected data is scarce or lacks diversity across channel conditions.

Artificially generated IQ samples created through software simulation of modulation and channel models. This approach provides:

• Perfect labels: No human annotation errors or ambiguity
• Rare signal coverage: Generate examples of uncommon or hostile modulations
• Controlled SNR sweeps: Produce datasets with precise, graduated noise levels for benchmarking

Combined with channel simulation, synthetic I/Q enables training on signal types that are difficult or dangerous to capture in the field.

The division of a continuous IQ stream into fixed-length segments to form individual inference examples. Key considerations:

• Segment length: Must capture enough symbols for modulation identification while fitting model input constraints
• Overlap strategy: Overlapping segments increase dataset size but introduce correlation between training examples
• I/Q Windowing: Applying tapering functions like Hamming or Blackman windows reduces spectral leakage before transformation to spectrograms

Proper segmentation ensures each training example contains a complete, coherent signal burst.

Two architectural strategies for feeding IQ data into neural networks:

• Dual-Channel Input: Treats I and Q as separate real-valued channels, analogous to RGB image channels. Compatible with standard CNN architectures
• Complex-Valued Input: Processes IQ natively as complex numbers using complex weights and activation functions, preserving phase relationships inherent in the signal

The choice impacts how the network learns representations—complex-valued networks often require fewer parameters to achieve equivalent performance on modulation tasks.

A time-frequency representation generated by applying the Short-Time Fourier Transform (STFT) to an IQ stream. This converts raw time-domain samples into a 2D image suitable for Convolutional Neural Networks. Benefits include:

• Revealing instantaneous frequency patterns that discriminate FM and FSK modulations
• Making temporal-spectral structures visually interpretable for debugging
• Leveraging pre-trained image model architectures through transfer learning

The trade-off is the loss of fine-grained phase information present in the raw complex samples.