Research On Deep Learning Based Automatic Modulation Recognition Technology

Automatic Modulation Recognition(AMR)is a technique used by communication receivers to detect the modulation of signals in non-cooperative scenarios.It has been widely applied in various civil and military fields,such as cognitive radio and spectrum sensing.With the rapid development of wireless communication technology,the application scenarios have become more diverse,and the modulation of signals has become more varied.Therefore,there is an urgent need to design more effective methods in complex radio environments to further improve the performance of modulation recognition models.Due to the powerful feature extraction and classification capabilities of deep neural networks,deep learning(DL)-based automatic modulation recognition methods have higher recognition accuracy than traditional likelihood-based and feature-based methods,making them a research focus in the field of automatic modulation recognition.This thesis conducts research in two different scenarios: single-input single-output(SISO)and multiple-input multiple-output(MIMO).It focuses on the lightweight model and data augmentation issues in the SISO scenario and further extends DL-AMR research to the MIMO scenario,improving the modulation recognition accuracy in pre-coded MIMO systems.The specific work is as follows:Firstly,to address the problem that most current models focus only on recognition accuracy,resulting in large model sizes and high computational complexity,this thesis proposes a lightweight model based on phase parameter estimation and transformation.The model first processes the input data through the parameter estimation and transformation module,and then further extracts features using a Convolutional Neural Network(CNN)and Gated Recurrent Unit(GRU).This model can achieve equivalent recognition accuracy to existing models,but reduces the parameter quantity by more than 40%.Additionally,the proposed model has lower time cost than existing high-accuracy recognition models.Furthermore,this thesis compresses the model further through pruning,and the pruned model can maintain equivalent accuracy to the baseline model with less than 1/8of the parameter quantity.Secondly,a method based on autoencoder is proposed to enhance the I/Q data by addressing the issue of many existing models that directly input I/Q data into a real-valued model without fully utilizing the correlated information in the I/Q channel.First,the method uses a fully connected layer-based autoencoder to correlate the features of I/Q data and obtain the correlation features of the I/Q channels from the middle layer of the autoencoder.Then,the data is enhanced using the obtained interaction features.In order to apply this data enhancement method and adapt the model to the new dimension of the enhanced data,one can simply modify the convolution kernel size of the existing modulation recognition model or the feature dimension of the Recurrent Neural Networks(RNN).Experimental results show that this method can improve the recognition accuracy of the existing baseline model and has a smaller time cost compared to the complex neural network methods.Finally,this thesis extends the research on deep learning-based automatic modulation recognition to the precoding MIMO scenario and investigates the modulation recognition problem of the precoding MIMO system using singular value decomposition.In response to the instability of the performance of the existing DL-AMR model with changing antenna numbers in the system,an attention-aided convolutional gated recurrent unit deep neural network model is proposed.This model can balance and utilize effective features from multiple antenna-received data streams through attention mechanisms.Experimental results in both precoding MIMO scenarios with and without receiving processing show that this model can achieve the overall best recognition accuracy.