I/Q Correction

Corrects gain mismatch (amplitude difference between I and Q branches) and phase error (deviation from the ideal 90-degree separation). These impairments, common in direct-conversion receivers, cause constellation warping that degrades modulation recognition accuracy. The correction typically applies a complex finite impulse response filter that mixes a fraction of the Q signal into the I path and vice versa, mathematically restoring orthogonality. Without this step, a 16-QAM signal can appear as a distorted elliptical cluster, confusing downstream neural network classifiers.

Eliminates the constant voltage bias introduced by local oscillator self-mixing and component mismatches in the analog front-end. This offset manifests as a non-zero mean in the IQ sample stream, shifting the entire constellation away from the origin. The correction estimates the DC component by averaging a block of samples and subtracting it, centering the constellation. For modulation schemes like BPSK or QPSK, even a small DC offset can mimic a legitimate signal component, leading to systematic classification errors.

Specifically targets the phase imbalance where the I and Q local oscillators are not exactly 90 degrees apart. This error causes a correlation between the supposedly independent I and Q components, resulting in a skewed constellation. The correction algorithm estimates the phase error from the cross-correlation of the I and Q streams and applies a rotation matrix to decorrelate them. In severe cases, quadrature error can make a QPSK signal indistinguishable from an unbalanced 4-PAM signal.

Many modern I/Q correction algorithms operate blindly, without requiring a known training sequence. They exploit statistical properties of the received signal—such as the assumption that ideal I and Q components are uncorrelated and have equal variance—to iteratively estimate and correct impairments. Techniques like the Gram-Schmidt orthogonalization procedure or adaptive least-mean-squares filters continuously track time-varying imbalances caused by temperature drift or component aging, ensuring sustained correction during long-duration signal collection missions.

Addresses I/Q imbalance that varies across the signal bandwidth, a common issue in wideband receivers where analog filters in the I and Q paths have slightly different frequency responses. Simple gain and phase corrections are insufficient; instead, a complex-valued asymmetric frequency-domain filter is applied. This filter compensates for the frequency-selective nature of the imbalance by modeling the impairment as a mismatch between the desired signal and its complex conjugate image, suppressing the mirror-frequency interference that would otherwise corrupt wideband modulation classifiers.

Uncorrected I/Q impairments directly reduce the effective signal-to-noise ratio and distort the feature space used by automatic modulation classifiers. For deep learning models trained on pristine synthetic data, even moderate hardware imbalance creates a domain shift that can drop classification accuracy by 20-40%. Implementing robust I/Q correction in the preprocessing pipeline ensures that the statistical signatures learned during training—such as constellation geometry and cumulant values—remain valid when the model is deployed on real receiver hardware.