Research On Evolutionary Deep Learning And Application Of Communication Signal Identification

The current classic pattern of deep learning applied to target recognition is still the isolated learning paradigm,which relies on human screening and preparation of a given set of data.It has to run a machine learning algorithm and build a model before applying it to a real task.This learning paradigm excels at working in closed environments,revealing patterns in the data for specific tasks.However,practical application scenarios are usually open and dynamic.The application effect of deep learning will be significantly deteriorated with the change of data,the change of target mode,and the emergence of unknown targets.Aiming at the target recognition task in the actual complex electromagnetic environment,this paper explores a new evolutionary deep learning paradigm that can deal with the data evolution,task evolution,and model self-evolution.In this paper,a new evolutionary deep learning method with continuous learning capability at the level of data,task and model is designed.The designed methods are verified in the identification task of communication signal,and the results show that the methods are feasible and effective.Main research contents and achievements are as follows:(1)In view of the data evolution problem of deteriorating data quality,a communication signal identification method based on deep residual shrinkage network is designed.The residual structure can be used to train the deep network and find deeper features of signal.The residual shrinkage module is introduced to automatically learn the optimal filtering threshold in the training process,and the noise information is converted to the near-zero domain.Finally,based on the spatial attention mechanism,soft thresholding is carried out for each channel of extracted features to filter the redundancy and noise information in the signal,so as to improve the robustness of the model.The experimental analysis of individual recognition on the collected data sets shows that the method can achieve the overall recognition accuracy of 97%on the pure data sets.Compared with other commonly used deep learning models,the recognition accuracy of noisy data sets is improved by more than 10%.These experiments prove the robustness of the method.(2)Aiming at the task evolution problem,an open set identification method of communication signal based on deep siamese metric network is designed.Firstly,the siamese metric network is used to train the known classes signals to reduce the inner-class distance and increase the between-classes distance.After that,the test samples and the known classes samples are formed into sample pairs.By calculating the similarity between the sample pairs,the ones with lower similarity are identified as unknown classes.In order to improve the efficiency of the algorithm,the transfer learning method is used to fine-tune the pre-trained individual recognition network after detecting a small number of samples of unknown classes.The remaining samples of unknown classes can be identified by trained individual recognition network.The experimental results show that the method can effectively identify all unknown categories,and the accuracy of the preliminary detection results is greater than 95%,which proves the effectiveness of the method.(3)Aiming at the sorting problem of unknown classes in chapter 3,a deep self-evolving clustering algorithm is designed for the sorting of communication emitter signals.Firstly,the labeled known classes signals are used to pre-train the individual recognition network which is used to reduce the dimension of the unknown signals.This method can make full use of the supervision information of the known classes samples.Secondly,the self-evolving clustering algorithm is used to train the "simple" sample pairs and then the "complex" sample pairs are added gradually for training.Finally,the learned features are transformed into a specific one-hot vector.The method of locating the maximum activation response position is adopted to mark all unknown category samples,so as to simplify the process of category marking in the clustering algorithm.The experimental results show that the recognition accuracy of the method can reach 95%,which proves the effectiveness of the method.