Research On Digital Modulation Recognition

This paper discussed the method of automatic modulation recognition of digital communication signals. It has decision theory method and statistical pattern method, the decision theory method, analyzing the statistics characteristic of signals, because it has considered the influence of noise jamming, has a better perform anceunder the low SNR. But this kind of method usually aims at some kind of concrete modulation signals’statistical characteristic which the analysis obtains; therefore the recognition scope is narrow. The statistical pattern recognition method usually extract features based on the non-noise jamming signals, therefore under high SNR condition, recognition performance is generally good, under low SNR condition, the recognition performance is generally bad.Compare to the tradition technology neural networks have the ability of resolving the complicated classification problems rapidly. Addition to its fault-tolerant ability insensitive to noise and incomplete data though its train or self-taught, neural networks are chosen to solve the problem of automatic digital modulation.In accordance with decision theory method and statistical pattern method these two approaches. This paper mainly proposed two kinds of feature parameters:one are five Nandi feature parameter that are derived from the instantaneous amplitude, instantaneous phase, instantaneous frequency and power spectrum of modulated signal. These five feature parameters are based on decision-theoretic;another are four feature parameters which are picked up from two-order, four-order and six-order cumulants, these four feature parameters are based on statistical pattern.Direct at2ASK、4ASK、2PSK、4PSK、2FSK、4FSK six signals,BP neural network classifier was studied to identify modulation with these two kinds of feature parameters, in order to conquer the disadvantages of standard back propagation (BP)algorithm,such as the low speed of convergence and the problem of local minimum points, this classifier used Levenberg-Marquardt algorithm, the tan-sigmoid function was used as the exciting function in the hidden layer and the log-sigmoid function was used as the exciting function in the output layer. There are ten nodes of hidden layer in the Nandi feature parameter classifier and eight nodes of hidden layer in the HOC feature parameter classifier. Through several simulations, this paper obtained simulation results under2-20dB SNR. Simulation results show that recognition rate of Nandi feature parameters is more than HOC feature parameters, when Signal to Noise Ratios (SNR) larger than6dB,the recognition rate of Nandi larger than97%. HOC feature parameters can not separate2PSK and2ASK signal, because when the signal transformed into the baseband, signal2PSK and2ASK is equivalent.