How do I implement real-time signal classification using machine learning on an SDR platform?

Implementing real-time signal classification using machine learning on an SDR platform involves capturing RF signals, extracting discriminative features from the I/Q samples, and applying a trained ML model to classify the signal type (modulation scheme, protocol, emitter identity) in real time. The signal processing pipeline is: (1) SDR captures wideband I/Q samples and the DDC extracts the signal of interest. (2) Feature extraction computes signal characteristics that discriminate between signal types, including time-domain features (higher-order statistics: variance, kurtosis, skewness of amplitude and phase), frequency-domain features (spectral flatness, bandwidth, peak-to-average ratio), cyclostationary features (cyclic autocorrelation at specific cycle frequencies unique to each modulation), and transform-domain features (constellation shape, wavelet coefficients). (3) The extracted features are input to a trained classifier: traditional ML models (SVM, Random Forest, k-NN) work well with hand-crafted features, while deep learning models (CNN, LSTM, ResNet) can learn features directly from raw I/Q samples or spectrograms, eliminating manual feature engineering. (4) The model outputs a classification decision (e.g., AM, FM, BPSK, QPSK, 16-QAM, OFDM, FHSS) with a confidence score. For real-time operation, the ML model must execute within the signal's coherence time, typically requiring inference latency below 1-10 milliseconds.