Modulation Recognition: Low-SNR Instability
Uses normalization, feature checks and confusion matrices to explain classification instability at low SNR.
Project: MATLAB Digital Modulation Recognition Simulation
Symptom: Phase and frequency-difference features became unstable at low SNR, while class amplitude scales biased decisions.
Hypothesis: The model could be relying too heavily on energy, with individual features becoming unreliable under noise.
Investigation: Inspected waveforms, spectra, constellations, feature NaN values and ranges, compared low- and high-SNR confusion matrices, and ruled out data leakage.
Root cause: Unequal scale and noise-sensitive individual features.
Fix: Normalized each frame and combined amplitude, phase and higher-order statistics in an ECOC RBF-SVM.
Lesson: Accuracy requires context from normalization, feature stability and confusion matrices rather than one aggregate score.