How do I implement automatic modulation recognition on an SDR receiver?

Implementing automatic modulation recognition (AMR) on an SDR receiver involves extracting discriminative features from the received I/Q samples and applying a classification algorithm to identify the modulation type without prior knowledge of the transmission parameters. The standard AMR pipeline captures a segment of I/Q samples (typically 256-2048 samples at the symbol rate or higher), preprocesses them (normalize power, estimate and remove frequency/timing offsets), extracts features that distinguish different modulation types, and applies a trained classifier to output the modulation identification. The most effective feature sets include: instantaneous signal statistics (mean, variance, kurtosis of the amplitude, phase, and frequency; kurtosis of the amplitude distinguishes constant-envelope from non-constant-envelope modulations), spectral features (spectral symmetry, number of peaks, spectral flatness), higher-order statistics and cumulants (4th-order and 6th-order cumulants have unique values for different modulations: C40 = 0 for Gaussian noise, -1 for BPSK, -1 for QPSK multiplied by different coefficients), and cyclostationary features (the cyclic spectrum has features at specific cycle frequencies determined by the symbol rate and carrier frequency). Classification algorithms range from decision trees (simple, fast, interpretable) to support vector machines (robust, good generalization) to deep neural networks (highest accuracy, learning features from raw I/Q).