Classifier Design And Implementation Of Automatic Digital Modulation Recognition Of Communication Signal

Pattern recognition, which has a broad application prospects, now has been widely used in character recognition, speech recognition, fingerprint recognition and so on. The automatic recognition of modulated communication signals also belongs to the scope of pattern recognition. The major research problems of pattern recognition are selection and optimization of classification feature and design of classifier. Therefore, the automatic recognition of modulated communication signals is also aimed at these two issues. But this paper focuses on the research of classifier which can be applied to the automatic recognition of digital modulated communication signals, in order to identify the signal's modulations effectively.Aiming at six kinds of digital modulation signals which commonly used in communication, this paper first introduces five classification parameters which are used to identify these six kinds of signals. Including the significance of these characteristic parameters and how to extract them. Then decision tree classifier provided with fixed threshold is adopted to identify these signals, the results show that the effect is not good at low SNR (Signal Noise Ration), unless at high SNR. When SNR is 25dB, the recognition rate percent is 94.44%.Then, neural network is used as a classifier.Firstly, the paper introduces the architecture of BP (Back-Propagation) neural network and the training method which amend the weights and thresholds by mean square error back propagation. Then a neural network architecture which is suitable for automatic modulation recognition of digital signals is designed, utilizing adaptive learning rate training method to train the BP network. After considering the effect of hidden layer neurons number and samples on recognition rate, concluding that the optimal number of hidden layer neurons is 10. So we verified the correctness of the empirical formula through which selecting the number of hidden layer neurons. Overall, when SNR is 15dB the recognition rate achieves over 90%, so the neural network classifier performs well at automatic modulation recognition of digital communication signals.In order to overcome the shortcomings of neural network such as local extremum, uncontrollable training process, requiring large training samples, we research the multi-class classification algorithm of support vector machine which possesses the advantages such as requiring small training samples, good generalization, global optimal. In this paper, three kind multi-class classification algorithms of support vector machine have been adopted, they are one versus many, one versus one, decision directed acyclic graph respectively. Through simulation comparison, the conclusion that one versus one algorithm is more suitable for automatic recognition of digital modulated communication signals is reached. This is not just because of one versus one algorithm has higher accuracy which achieves 87.17% at 10dB and 95.83% at 15 dB, but also because of it's superiority in training and classification time.Finally, the paper briefly introduces the hardware which would be used in automatic modulation recognition. Also the multi-class algorithm of one versus one is implemented on DSP, achieving the effect of identifying the signal's modulation in real-time. And thus validates the effectiveness of the designed classifier.