Research On Modulation Signal Recognition Based On Deep Learning

With the development of communication technology,various new modulation methods emerge endlessly.In order to adapt to different recognition requirements,modulation recognition technology needs to be developed continuously.In recent years,modulation signal recognition technology based on deep learning has attracted much attention because of its strong self-learning ability.However,there are still some problems with the technology.Therefore,this paper mainly studies two problems:Firstly,to solve the problem that the traditional modulation recognition method is difficult to recognize the complex modulation mode of signals,this paper proposes a multi-mode and multi-channel modulation signal recognition algorithm based on attention.The algorithm converts the original data into AP format,FFT format and IQ format as the input of the model,making full use of the interaction and diversity of different features.Multi-channel models in MCLDNN were used for feature fusion to avoid the deterioration of orthogonal relations caused by parameter imbalance.In the feature extraction module,the residual network based on depthseparable convolution and BiLSTM network based on attention mechanism are used to extract the temporal features of signals.Meanwhile,the introduction of self-attention mechanism and multi-head attention mechanism can capture the dependency relationship between features.The experimental results show that the algorithm can improve the recognition accuracy of the model,reduce the complexity of the model,and improve the training efficiency of the model to some extent.Secondly,the adversarial migration network is introduced to realize the transfer of knowledge from the source domain to the target domain in order to improve the recognition performance of the target domain model,aiming at the small sample data in the actual environment and the modulated signal data in the low SNR environment.We propose two antitransfer learning models:one based on information decoupling and the other based on multimodal mechanism.Among them,the model based on information decoupling introduces private components and cross-domain shared components for modeling,which improves the generalization and performance of the model.Private feature extractors are used in each domain to extract private features,while shared feature extractors are added to extract common features,which can avoid negative migration to a certain extent and improve target regionality.In addition,the feature fusion of IQ format and AP format signals can be carried out by introducing a multi-modal mechanism,which effectively improves the model recognition accuracy of the target domain.Experimental results show that the proposed method can effectively improve the recognition accuracy and model performance of small sample data and low signal-to-noise ratio data sets.