PSK Demodulation Algorithm Based On One-dimensional Convolutional Neural Network

In recent years,digital communication technology occupies a dominant position in the field of wireless communication.Modulation and demodulation is the key technology in digital communication.Phase Shift Keying(PSK)is the most commonly used digital modulation mode,which has advantages of high modulation efficiency,high utilization of transmission band and strong anti-noise capability.The conventional PSK demodulator is usually implemented by specific hardware platform.It has many shortcomings,such as high cost and long development cycle.In recent years,Conventional demodulator has a tendency to be replaced by software defined radio.In this thesis,a PSK demodulation algorithm based on One-Dimensional Convolutional Neural Network(1-D CNN)is proposed.The algorithm performs the demodulation function by detecting the position and type of phase jump in the PSK modulated signal.Firstly,the time window is employed to obtain the input vectors of 1-D CNN.Then the 1-D CNN is used to detect whether the phase jump is existing in each input vector.The outputs of 1-D CNN are made up to a sequence in time order.Finally,the demodulation results are obtained after the sequence is processed by the timing synchronization module.Compared with other demodulation algorithms based on neural network,the 1-D CNN structure proposed in this thesis can reduce the weight connection and the computational complexity.The phase information is extracted from the location of phase jump and the type of phase jump in PSK sampling signal,instead of directly mapping the sampling point to symbol,so that the PSK signal is avoided to be grouped according to the symbol period.With the use of timing synchronization module,the demodulation algorithm in this thesis can deal with the carrier frequency offset and sampling frequency error.In this thesis,we construct a 1-D CNN model,and generate BPSK and QPSK modulated signal to gain the training sets.1-D CNNs are trained to detect the phase jump in BPSK and QPSK modulated signal.The experimental results show that the 1-D CNNs trained by different signal-to-noise ratio(SNR)training sets have different demodulation performance,and the optimal 1-D CNN is achieved by the training set with SNR of-2d B.For BPSK signals,the algorithm proposed in this thesis can provide better demodulation performance than the theoretical value when SNR is higher than 4d B.For QPSK signals,the performance of the proposed algorithm is close to that of the conventional coherent demodulation method.Compared with the conventional demodulator,the demodulation algorithm proposed in this thesis has more flexibility as it is not necessary to use specific hardware.Because the function of the demodulation is realized by the neural network,the training set can be generated according to the channel environments to improve the adaptability of the algorithm.Moreover,the generalization ability of neural network can provide better anti-noise capability for demodulation.