Research On Blind Estimation Of Signal Modulation Parameters Based On Transformer

In civil-communication,automatic modulation recognition can reduce signaling transmission overhead and improve the spectral efficiency and transmission reliability of communication links;in military communication,the modulation type of hostile signals can be obtained to formulate corresponding interference or anti-jamming strategies.The modulation recognition neural network based on deep learning integrates the feature extractor and classifier to fully extract the signal features and improve the classification ability.The modulation recognition performance far exceeds the traditional algorithm;and the good recognition performance on different databases reflects its Strong learning and adaptability.Therefore,with the vigorous development of neural networks,automatic modulation recognition technology has become a research hotspot.First,in view of the limited receptive field and high time complexity of classical neural networks,this paper proposes a new Transformer-Encoder modulation recognition model based on the parallel computing structure of Transformer and its excellent modeling ability for long-term dependencies.On the Radio ML2018.01 a database,the average recognition performance is improved by 6.2% compared with the optimal classical model MC-Net,and the time complexity is reduced by 54%.Since the Transformer-Encoder model requires a lot of training data to reflect its excellent recognition performance,the average recognition performance of the Radio ML2016.10 a database is 1.9% lower than the optimal classical model MCLDNN.Therefore,this paper proposes a combination of CNN and Transformer.The C-TNet model improves the recognition performance of the network model through the innate bias induction ability of CNN.The average recognition performance of the MCLDNN model is improved by 0.2% on the Radio ML2016.10 a database.Further,the LSTM network based on RNN can better model the long-term dependence of temporal features.A CL-TNet model combining the characteristics of RNN,CNN and Transformer is proposed,which makes up for the shortcomings of the CTNet model in extracting temporal features and makes The advantages of different network structures are complementary.On the Radio ML2016.10 a database,the CL-TNet model not only improves the average recognition performance of the MCLDNN model by 0.5%,but also improves the average recognition performance of the MC-Net model by 7.7% on the Radio ML2018.01 a database.The TNet modulation recognition algorithm model has better recognition performance and generalization ability.Finally,in the field of modulation recognition,there is a shortage of label data,which leads to the problem that the neural network cannot learn effective features.In the field of computer vision,this problem is effectively solved by various pre-training methods,so the pre-training method is effectively migrated to modulation recognition.In the field,the CL-TNet model and Transformer-Encoder model proposed in this paper were pre-trained under fully supervised and unsupervised conditions,respectively,and the average recognition performance of the model without pre-training was improved by 0.6% and0.5%,respectively.The pre-training mode can not only improve the recognition performance of the model,but also unsupervised pre-training improves the practicability of the model.