Deep Sequential Feature Learning And Recognition Of Signal

Signal feature extraction methods are the key technology in the field of signal recognition.At present,the methods of signal feature extraction and recognition based on deep learning invove some problems in practical application:Firstly,the objects of signal recognition based on deep learning are mostly simple modulation of signal,which rarely involves the coding of signal and radio signal.The popular signal feature extraction methods use stack convolution neural network to extract the structural features of signal,which ignores the sequential characteristics of signal.Secondly,the signal recognition methods based on deep learning can not extract the correlation between the complex orthogonal channels of radio signal.Thirdly,the signal recognition methods based on deep learning are difficult to fully mine the structure features of signal under a small number of labeled samples.In view of the above issues,this paper designs three methods of signal recognition based on deep learning,which can intelligently extract the sequential features of signal and complete the task of signal recognition.The main research contents and achievements of this paper are as follows:1.A modulation/coding recognition method based on coordinate convolution attention sequential network is designed.Firstly,the coordinate convolution is introduced into the shallow feature extraction module to extract the primary sequential features of the signal.Secondly,the multi-scale coordinate convolution module composed of multiple channels is used to extract and fuse the features of different scales of the signal,then the gate current neural unit is used to further extract the sequential features of the signal,finally the attention module is used to select and enhance the important features of the extracted sequential features.The experimental results show that the method can complete the task of signal recognition with more than 98%accuracy on the two data sets of signal,which verify effectiveness of the method on signal recognition.2.Aiming at the problem that the traditional network can not mine the correlation between the complex channels of radio signal,a radio signal recognition method based on the lightweight complex hash sequential network is designed.Firstly,the complex convolution is introduced into the signal feature extraction module to extract the correlation of orthogonal channels.Secondly,considering the similarity of the signal feature map extracted by the convolution operation,a simple linear transformation can be used to obtain the feature map with similar structure,so as to avoid the complexity of the network structure caused by feature maps directly generated in the traditional convolution operation.Then,the enhanced timing sequential module is used to extract the enhanced sequential feature.The experimental results show that the method can complete the task of radio signal recognition with the lowest accuracy of 98%on the two data sets of radio signal,which verify the effectiveness of the method for radio signal recognition.3.In order to solve the problem that the deep learning method is difficult to fully mine the signal structure features under a small number of label samples,a radio signal recognition method based on the deep feature fusion self-training sequential network is designed.Firstly,the self-training algorithm is introduced into the neural network model,allowing a large number of unlabeled radio signals to assist a small number of labeled signals to train the basic classification network;Secondly,the feature fusion structure of adding points by bits is adopted in the basic classification network,which fuses the primary features of the signal extracted by the multi-scale setting convolution designed in Chapter 2 and the lightweight complex convolution in Chapter 3,then the primary features are enhanced by the timing sequential unit to extract the timing sequential features of the signal,which enhances the ability of the model to extract the features of the signal under a small number of labeled samples.The experimental results show that the method can complete the task of two kinds of radio signal data sets with 90%recognition accuracy under the premise that the labeled samples account for 10%of the total samples.