Dual-Channel Input

Each component stream—In-Phase (I) and Quadrature (Q)—is fed into the network as a separate input channel, preserving the distinct amplitude and phase information without premature mixing. This allows the first convolutional layer to learn cross-channel correlations through its filter kernels, discovering features such as the circular symmetry of QPSK or the amplitude variations of 16-QAM directly from the raw data. The approach avoids the information loss that can occur when collapsing the complex signal into a single real-valued magnitude or phase representation before feature extraction.

This architecture directly borrows from computer vision, where a color image is represented as three separate Red, Green, and Blue channels. In the same way that a CNN learns edge detectors that combine information across RGB channels, a dual-channel modulation classifier learns I/Q cross-correlation filters that detect phase transitions and amplitude shifts. This design choice enables the reuse of highly optimized, pre-existing 2D CNN backbones—such as ResNet or EfficientNet—without requiring custom complex-valued layer implementations.

By separating I and Q into two real-valued tensors, the network operates entirely in the real number domain, avoiding the need for complex-valued weights, complex backpropagation, or specialized complex activation functions. This simplifies implementation in standard deep learning frameworks like PyTorch and TensorFlow, and ensures compatibility with common hardware accelerators. The trade-off is that the network must learn to implicitly reconstruct the phase relationship between I and Q from the paired channels rather than having it encoded natively in a complex data type.

A raw IQ segment of N samples is reshaped into a tensor of shape [2, N] or [2, H, W] if further structured into a 2D grid. The first dimension represents the two channels: index 0 for the I component and index 1 for the Q component. This format is directly ingestible by standard 2D convolutional layers where the in_channels parameter is set to 2. For batch processing, the full tensor shape becomes [Batch, 2, N], enabling efficient parallel inference on GPU hardware.

The primary advantage of dual-channel input is that the network autonomously learns the optimal feature representations directly from the raw time-domain samples, bypassing manual feature engineering. The first layer's filters typically converge to detect instantaneous phase transitions, amplitude shifts, and cross-channel patterns that correspond to specific modulation constellations. This end-to-end learning approach often surpasses handcrafted feature-based methods, especially in low-SNR conditions where traditional statistical features degrade.

Unlike a complex-valued neural network that processes I and Q as a single complex number I + jQ, the dual-channel approach treats them as independent but paired real values. The complex-valued approach natively preserves phase rotation equivariance and requires fewer parameters to model circular symmetries. However, the dual-channel method benefits from a mature ecosystem of pre-trained models, optimization techniques, and deployment tooling. The choice between the two often depends on the target hardware and the availability of complex-valued operation support.

An alternative neural network design that processes IQ data natively as complex numbers rather than splitting them into two real-valued channels. This approach uses complex-valued weights and activation functions, such as the complex ReLU, to inherently preserve the phase relationships and algebraic structure of the signal. While mathematically elegant, it requires specialized deep learning frameworks and can be more computationally intensive than the pragmatic dual-channel approach.

A critical hardware impairment in direct-conversion receivers where the gain or phase relationship between the I and Q signal paths deviates from perfect orthogonality. This creates a mirror-image interference in the constellation. For a dual-channel input model, uncorrected I/Q imbalance presents a distorted geometric structure to the network, forcing the classifier to learn to compensate for a hardware-specific artifact rather than focusing solely on the modulation scheme.

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 where one axis is time and the other is frequency. Unlike the dual-channel input, which feeds raw samples to a 1D CNN or transformer, the spectrogram approach allows the use of standard 2D Convolutional Neural Networks pre-trained on visual tasks, trading temporal precision for frequency-domain clarity.

A preprocessing step that scales the amplitude of an IQ stream to a standard range, typically using Z-score or min-max scaling. For dual-channel inputs, this is essential to prevent numerical instability during training. Normalization ensures that the I and Q channels have zero mean and unit variance, preventing the network's weights from saturating and allowing the optimizer to converge on the true modulation structure rather than being biased by absolute signal power.

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

• Phase rotation: Multiplying by a complex exponential to teach rotational invariance
• Additive White Gaussian Noise (AWGN): Adding thermal noise to improve low-SNR robustness
• Fading simulation: Applying Rayleigh or Rician fading profiles These transformations are applied identically to both the I and Q channels to maintain their physical correlation.

The residual frequency difference between the transmitter and receiver local oscillators, causing the received IQ constellation to rotate continuously over time. For a dual-channel input classifier, an uncompensated CFO introduces a time-varying phase shift that corrupts the static geometric features the network relies on. Robust models must either learn CFO invariance through augmentation or rely on a preprocessing I/Q Centering block to remove the offset.