Research On Modulation Classification Algorithm Based On Deep Learning

Blind Modulation Classification(BMC)is to determine the modulation method of a signal without prior knowledge of the channel.Blind modulation recognition is widely used in both civil and military communication systems.In civil systems,blind modulation identification is specially applied to software radios to deal with various communication systems.In military communication systems,signal interception requires modulation identification techniques,which are critical for electronic warfare decision-making.Existing modulation recognition algorithms are roughly divided into two categories: methods based on maximum likelihood(ML)and methods based on feature extraction(FB).Maximum likelihood-based methods require explicit signal models,which are difficult to achieve in non-cooperative blind modulation recognition where multiple signals are present.Simultaneously calculating the likelihood function requires clear channel parameters and noise variance,which has extremely high computational complexity and is difficult to use in resource-constrained communication devices and scenarios.Therefore,the method based on maximum likelihood is difficult to apply in reality.The modulation recognition algorithm based on feature extraction has the characteristics of avoiding the calculation of likelihood and using a well-trained model for recognition,and is widely used due to its high computational efficiency.However,the traditional method based on feature extraction has the following defects: manual selection of various features,difficult design,and only effective for certain types of modulation methods,unable to achieve universal identification of multi-type modulation methods.Therefore,many scholars propose to use deep learning,which can effectively extract features,to solve the above problems.In recent years,modulation recognition algorithms based on deep learning have significantly improved in recognition accuracy.However,feature-based methods,especially deep learning-based methods,require clean,large,and outlier-free datasets,which are often not readily available in the real world.Therefore,this thesis starts from the following three aspects to solve the problems caused by data.Aiming at the problem that the data set contains noisy data and the modulation recognition accuracy decreases rapidly under low signal-to-noise ratio,this thesis proposes a joint denoising modulation recognition method based on Multitask Learning(MTL),in which the denoising network and classification The networks are trained simultaneously in an end-to-end fashion.Furthermore,a focal loss function is employed to highlight the importance of hard-to-classify samples during training.Numerical experiments show that the proposed method can effectively improve the modulation recognition accuracy under low SNR and outperform the state-of-the-art Res Net method.Under the condition of 0d B SNR,the recognition accuracy of our proposed Multitask＿Res Net method is 12% higher than that of the Res Net method.Aiming at the problem that well-labeled data is very scarce and deep learning is difficult to achieve good performance with small samples,this thesis proposes a smallsample modulation recognition method based on Generative Adversarial Nets(GAN)to generate signal datasets.Using the Transformer-based generative adversarial network to generate fake data,more accurate modulation recognition can be achieved with only a small amount of sample data.Under the condition of 0d B SNR,the recognition accuracy of our proposed data augmentation method is 10% higher than that of the Res Net method.Aiming at the problem of out-of-class data that is not included in the data set during actual classification,that is,there are outliers,this thesis proposes an outlier detection method based on autoencoder and K-means clustering algorithm.Distinguish in-class data from outlier data.The algorithm first trains an auto-encoder network,then removes the decoder and retains the encoder,inputs the output of the encoder as the extracted feature into the K-means clustering algorithm to realize outlier detection,and then modulates the data after outlier detection.identify.It is verified by experiments that the detection accuracy of outliers can reach 88%.Under the condition of 0d B SNR,the modulation recognition accuracy after integrated outlier detection is 8% higher than that of the Res Net method.