Deep-Learning-End-to-End-Digital-Modulation-Classification-and-Demodulation-System

Conducted data processing using MATLAB signal toolbox, which calculate cummulants and gen- erate labeled QPSK, 8PSK, 16QAM train and test dataset (including different signal-to-noise ratio)

Built modulation recognition model using Multilayer Perceptron, for data with SNR above 15db, the accuracy rate is close to 100%.

Built demodulation model using Convolutional Neural Network and generated confusion matrix, for data with SNR above 20db, the accuracy rate is close to 100%.

Parameter tuning, using relu activation function, softmax crossentropy loss function, adam opti- mization algorithm and adjusting learning rate, which increase the convergence speed and improve classification accuracy.

Cascaded two models to achieve end-to-end modulation and demodulation

modulation_accuracy

8PSK_End_to_End_Modulation_Classification_Demodulation_accuracy

16QAM_End_to_End_Modulation_Classification_Demodulation_accuracy

QPSK_End_to_End_Modulation_Classification_Demodulation_accuracy

8PSK_Demodulation_accuracy

16QAM_Demodulatio_accuracy

QPSK_Demodulatio_accuracy

By using covolution layer we can improve the performance of 16QAM