Raw I/Q Input

Raw I/Q input retains the complete complex baseband representation of the signal, preserving both magnitude and phase relationships that are often destroyed by feature extraction. Each sample is a complex number I + jQ, where the In-Phase (I) and Quadrature (Q) components capture the signal's instantaneous state. This allows the neural network to learn phase-sensitive patterns such as:

• Rotational symmetries in QPSK and QAM constellations
• Phase transitions in CPM and GMSK signals
• Cross-coupling between I and Q channels caused by hardware impairments Unlike magnitude-only or spectrogram-based inputs, complex-valued preservation enables the classifier to exploit the full information content of the received waveform.

Feeding raw I/Q samples directly to a neural network eliminates the need for hand-crafted feature engineering. The model learns hierarchical representations automatically:

• Early layers typically learn bandpass filters and transient detectors
• Middle layers discover modulation-specific temporal patterns
• Deeper layers construct invariant representations robust to noise and channel impairments This approach outperforms traditional feature-based methods when the optimal discriminative features are unknown or when channel conditions deviate from textbook models. The network discovers data-driven features that human engineers might overlook, particularly for complex or compound modulation schemes.

Raw I/Q input can be processed through two fundamentally different neural network paradigms:

Dual-Channel Real-Valued Input:

• I and Q components treated as separate real-valued channels
• Analogous to RGB image channels in computer vision
• Compatible with standard real-valued CNN and ResNet architectures
• Requires the network to implicitly learn I/Q relationships

Complex-Valued Neural Networks (CVNNs):

• Process I/Q data natively as complex numbers
• Use complex-valued weights, activations, and backpropagation
• Naturally preserve phase information through complex multiplication
• Often achieve superior performance with fewer parameters
• More computationally demanding and less supported by standard frameworks

Raw I/Q input imposes specific constraints on data preparation:

Sample Rate Considerations:

• Must satisfy Nyquist criterion for the signal bandwidth
• Typically 2-4× the symbol rate for adequate temporal resolution
• Higher sample rates capture finer transient details but increase computational load

Segmentation Strategy:

• Continuous I/Q streams must be divided into fixed-length segments for batch processing
• Segment length determines the temporal context window available to the classifier
• Typical lengths range from 128 to 2048 samples per inference example
• Overlapping segments can increase temporal coverage at the cost of redundant computation
• Segment boundaries should ideally align with symbol boundaries to avoid truncation artifacts

Raw I/Q input exposes the neural network to real-world receiver non-idealities that must be learned or compensated:

I/Q Imbalance:

• Gain mismatch between I and Q paths causes elliptical constellation distortion
• Phase error destroys orthogonality, creating cross-talk

DC Offset:

• Manifests as a shifted constellation center
• Can saturate subsequent network layers if not addressed

Carrier Frequency Offset (CFO):

• Causes continuous constellation rotation over time
• The network must learn rotational invariance or rely on preprocessing

Phase Noise:

• Introduces random phase jitter from local oscillator instability
• Particularly problematic for high-order QAM classification

Networks trained on clean synthetic data often fail when deployed on real hardware unless these impairments are included in training augmentation.

Training with raw I/Q input requires aggressive channel-aware data augmentation to achieve real-world robustness:

• Phase rotation: Random angular shifts teach rotational invariance for CFO resilience
• Additive White Gaussian Noise (AWGN): Variable SNR levels simulate thermal noise conditions
• Amplitude scaling: Random gain variations mimic automatic gain control (AGC) fluctuations
• Frequency offset injection: Synthetic CFO teaches the network to handle rotating constellations
• Multipath fading: Rayleigh and Rician fading models simulate mobile channel conditions
• Time stretching/compression: Small sample rate variations mimic clock drift

These augmentations effectively multiply the training dataset size and force the network to learn impairment-invariant features rather than memorizing clean signal templates.