Robust classification of modulation types using spectral features applied to HMM

A robust method for asynchronous modulation recognition without a priori knowledge about modulation and transmission parameters is proposed. The generated features are based on modulation dependent spectral lines observed in the estimated spectra of m-th law transformations of the received signal. A twofold robust normalisation scheme is introduced, which is adequate to build significant feature vectors for linear and nonlinear (FSK) digital modulation types. The series of spectral line occurrences is mapped by a vector quantiser to a series of discrete states. Following this the feature vectors are classified using separate hidden-Markov models (HMM) for each modulation type. The classification results for the modulation types ASK, BPSK, QPSK, 2FSK, MSK and CW, obtained from computer simulations, prove the robustness of the classifier. Furthermore, it clearly outperforms a decision-tree based classifier.