Automatic Subtractive Clustering Algorithm Model (ASCAM) for Blind Identification of QAM Constellations

Blind System Identification (BSI) is a major pattern recognition task in many fields, including digital communication systems. Recently, many machine learning techniques have been applied to identify modulation types of signals solely by the receiver without prior knowledge of the transmission protocols. In this paper, we propose Automatic Subtractive Clustering Algorithm Model (ASCAM), a novel hybrid approach between subtractive clustering algorithm (SCA) and deep neural network for not only classifying the order of Quadrature Amplitude Modulation (QAM) from the corrupted baseband constellations, but also estimating the original symbols' locations in their constellations. For this model, a deep neural network based on parallel convolution is used in conjunction with a negative straight-through estimator to formulate this problem as a supervised task, regardless of the non-differentiability of SCA. Our simulation demonstrates that the model works well for QAMs with Rayleigh fading and additive white Gaussian noise (AWGN), and achieves better performance than other existing methods. It also generalizes its feature extraction well, as it can maintain its performance even when some symbols in the observed constellations are not present.