I/Q Segmentation

The raw I/Q stream is partitioned into segments of N complex samples, where N is typically a power of two (e.g., 1024, 2048) to align with FFT-based feature extraction pipelines. Each segment becomes an independent inference example.

• Deterministic sizing: All segments have identical dimensionality, enabling efficient batch processing on GPUs
• Symbol coverage: Segment length must span multiple symbol periods to capture the modulation's statistical fingerprint
• Trade-off: Longer segments improve classification confidence at low SNR but increase latency and memory consumption

Segments can be extracted with or without overlap between consecutive windows.

• Non-overlapping: Each sample belongs to exactly one segment. Maximizes throughput and minimizes redundant computation, but may miss transient signal events occurring at segment boundaries
• Overlapping (sliding window): Adjacent segments share a fraction of samples, typically 25-50%. Increases temporal resolution and ensures transient events are fully captured within at least one window, at the cost of higher computational load
• Hop length: The number of samples the window advances between segments; a smaller hop length increases overlap and total segment count

Each extracted segment inherits a modulation label for supervised training. Labeling strategies depend on the data source.

• Synthetic data: Labels are known deterministically from the simulation parameters, providing perfect ground truth
• Over-the-air captures: Segments may require manual annotation or metadata extraction from protocol headers
• Multi-label segments: In dynamic spectrum access scenarios, a single segment may contain multiple modulation types due to interference or spectrum sharing, requiring multi-label classification architectures
• Boundary alignment: Labels must be temporally synchronized with the segment; misalignment introduces label noise that degrades classifier performance

In real-time deployment, segmentation operates as a continuous streaming process where new samples are buffered and segments are produced at a fixed stride interval.

• Stride: The number of new samples required before the next segment is emitted; equal to the hop length in sliding window implementations
• Buffer management: A circular buffer accumulates incoming I/Q samples, and a segment is extracted once the buffer contains sufficient new data
• Latency budget: The stride directly controls inference latency; a stride of 512 samples at 1 MSPS yields a new classification every 0.512 ms
• Drop policies: When the classifier cannot keep up with the incoming sample rate, segments may be dropped or the stride increased dynamically

In multi-antenna receivers, segmentation must preserve spatial coherence across channels.

• Spatial alignment: Segments from all antenna elements must correspond to the exact same time interval to maintain phase coherence for spatial processing
• Tensor shaping: A MIMO segment is typically shaped as [N_samples × N_antennas × 2] (I/Q as separate channels) or [N_samples × N_antennas] (complex-valued)
• Synchronization requirement: All receiver chains must share a common time reference; sample-level misalignment between channels destroys the spatial signature used by the classifier
• Throughput scaling: The aggregate sample rate scales linearly with the number of antennas, demanding proportionally higher segmentation throughput

Abruptly truncating a continuous I/Q stream at segment boundaries introduces spectral artifacts that can mislead classifiers.

• Spectral leakage: The rectangular window implicit in segmentation spreads energy into adjacent frequency bins, distorting the signal's true spectrum
• Windowing functions: Applying a Hamming, Hann, or Blackman window to each segment tapers the edges to zero, reducing leakage at the cost of slightly attenuating samples near segment boundaries
• Overlap-add reconstruction: In applications requiring signal reconstruction after classification, overlapping segments with appropriate windowing enable artifact-free reassembly
• Classifier robustness: Deep learning models trained on windowed segments learn to compensate for the amplitude roll-off at segment edges

The critical signal conditioning pipeline applied to raw IQ samples before segmentation and inference. This includes:

• I/Q Normalization: Scaling amplitude to a standard range (e.g., Z-score) to prevent numerical instability
• I/Q Centering: Removing residual Carrier Frequency Offset (CFO) to center the constellation
• I/Q Correction: Compensating for hardware impairments like I/Q Imbalance and DC Offset

Proper preprocessing ensures that the fixed-length segments fed to the classifier contain a clean, standardized representation of the modulation structure.

The application of a tapering function to each segmented IQ frame to reduce spectral leakage before transformation. Common window functions include:

• Hamming Window: Optimized for side-lobe suppression
• Blackman Window: Superior stopband attenuation
• Hann Window: Balanced trade-off for general use

Windowing is essential when segments will be converted to I/Q Spectrograms via the Short-Time Fourier Transform (STFT), ensuring clean time-frequency representations for CNN-based classifiers.

A data regularization technique that applies realistic channel impairments to segmented IQ examples to expand training dataset diversity. Key augmentations include:

• Phase Rotation: Teaching classifiers rotational invariance by applying uniform angular shifts
• Additive White Gaussian Noise (AWGN): Simulating thermal noise across a range of SNR levels
• Fading Profiles: Applying Rayleigh or Rician fading to mimic multipath propagation

Augmentation is typically applied after segmentation, creating multiple impaired variants of each clean segment to improve model robustness.

The process of changing the sample rate of an IQ stream through decimation or interpolation to match the native input requirements of a downstream neural network. This step is often tightly coupled with segmentation:

• Decimation: Reducing the sample rate by an integer factor to lower computational load
• Interpolation: Increasing the sample rate to capture finer temporal details
• Rational Resampling: Using fractional rate changes for precise alignment

Resampling ensures each fixed-length segment contains the correct number of samples per symbol for optimal classification performance.

The systematic curation of labeled IQ recordings used to train and benchmark modulation classifiers. A robust dataset requires:

• Diverse Modulation Schemes: BPSK, QPSK, 16-QAM, 64-QAM, and beyond
• Varied SNR Levels: From -20 dB to +30 dB to ensure real-world applicability
• Channel Impairments: Multipath fading, frequency offsets, and phase noise
• Synthetic I/Q Generation: Software-simulated signals providing perfectly labeled data for rare or classified modulation types

Proper segmentation parameters (segment length, overlap ratio) are critical metadata in any I/Q dataset specification.

The process of recovering the optimal sampling instant from a continuous waveform before segmentation occurs. Key considerations:

• Symbol Timing Recovery: Aligning discrete IQ samples precisely with transmitted symbol centers
• Gardner Algorithm: A popular feedback-based timing error detector
• Inter-Symbol Interference (ISI): Minimized when sampling occurs at the point of maximum eye opening

Without proper synchronization, segmented frames will contain samples at suboptimal points, degrading the classifier's ability to discriminate between modulation schemes.