Deep Learning Few-shot Electromagnetic Signal Classification

The core task of electromagnetic signal classification is to effectively distinguish different types or individuals of electromagnetic signals by mining the characteristics of electromagnetic signals and designing reasonable classifiers.With the rapid development of information technology,electromagnetic signal classification technology plays an increasingly important role in the fields of electromagnetic spectrum control,cyberspace security,and cognitive radio.However,with the emergence of a large number of radio equipment,the electromagnetic environment is increasingly complex,and the spectrum range,pulse frequency and working time of various types of equipment overlap intensively,which makes the traditional electromagnetic signal classification technology based on artificial design features gradually difficult to cope with.In recent years,deep learning has been applied to electromagnetic signal classification tasks with its powerful automatic feature extraction ability,which has significantly improved the classification effect.However,the fully supervised learning method represented by convolutional neural networks needs a large number of labeled data for training and processing to achieve high-precision classification result of electromagnetic signals.In specific fields such as spectrum sensing and spectrum control,the difficulty of intercepting electromagnetic signals and the variety of electromagnetic signals with fast update change rate lead to the extreme scarcity of high-quality labeled signal data that can be used for research.How to improve the cognitive performance of electromagnetic signals under few-shot conditions through efficient characterization and deep learning of a small number of labeled or even unlabeled signal samples is an urgent challenge for electromagnetic signal classification tasks.Meanwhile,limited by storage resources,the number of old class signal samples that can be utilized when dealing with incremental target classification problems in open environments is relatively limited,which makes it very difficult to achieve effective learning of both old and new class signals.Starting from the above-mentioned problems,this thesis deeply investigates the electromagnetic signal learning characterization and intelligent classification technology under few-shot conditions,provides theoretical and technology support for the research of solving the bottleneck problem of electromagnetic signal classification technology,and analyzes and verifies the engineering practicality of the algorithm,which has certain research significance and application value.Details of the work are as follows:(1)For the situation that there are few labeled signal samples but have knowledge in similar domains,a few-shot electromagnetic signal classification method based on selective knowledge transfer is proposed.The method compares the transferability of feature knowledge by singular value decomposition of feature matrix in source and target domain.The negative transfer during knowledge transfer is suppressed by directly suppressing the small singular value of the signal characteristic matrix,and the selective transfer of knowledge in the source domain is realized.At the same time,the Stochastic normalization layer is used to replace the batch normalization layer in the depth neural network to transfer more parameters to avoid over-fitting.A large number of experiments have been carried out on the RML2016.04 c,RML2016.10 a and ACARS dataset.The experimental results show that the method based on selective knowledge transfer can greatly improve the classification performance compared with direct training.(2)For the situation that there are few labeled signal samples and abundant unlabeled samples,a novel semi-supervised learning framework based on generative adversarial networks is proposed for electromagnetic signal data identification,including modulation type identification and emitter individual identification.The framework can directly process raw IQ data of electromagnetic signals,make full use of unlabeled sample resources,and achieve end-to-end classification of electromagnetic signals by training effective classifiers.According to the characteristics of electromagnetic signals,a weighted loss function is designed,which combines binary cross entropy and feature matching in the network training process to achieve more efficient signal feature extraction.The experimental results on RML2016.04 c and ACARS dataset show that the proposed method can achieve end-to-end accurate classification of electromagnetic signals under semi-supervised conditions.(3)For the situation that there are few labeled signal samples and no background knowledge such as similar domain knowledge and unlabeled samples,a data union augmentation method of electromagnetic signal based on generative adversarial networks is proposed for few-shot electromagnetic signal classification.In order to effectively augment the electromagnetic signal data under limited sample conditions,an electromagnetic signal generation method based on generative adversarial networks is firstly designed.In order to improve the quality of the generated data,an electromagnetic signal screening mechanism based on the similarity principle is established to eliminate the generated signal samples with large errors from the real signal.In view of the limitation of the distribution state that can be expressed by limited samples,in order to further improve the feature extraction ability in the case of small samples,combined with the physical mechanism of electromagnetic signals,a data union augmentation mechanism of signal data is designed,which introduces the spatiotemporally flipped shapes of the signal samples,and further enriches the training set.The experimental results on RML2016.04 c and ACARS data sets show that the proposed method can effectively augment the electromagnetic signal data,thus significantly improving the performance of few-shot electromagnetic signal classification.(4)For the situation that there is no labeled signal samples at all,an autoencoder-based deep clustering method for electromagnetic signals is proposed.The combined loss function composed of reconstruction loss and deep clustering loss is used to train the autoencoder,which makes the deep features of the same class samples more clustered in the feature space,thus significantly improving the clustering effect.In addition,a feature visualization method for signal clustering is proposed based on Grad-CAM technique,and the interpretability of the autoencoder is analyzed by saliency maps.Extensive experiments have been conducted on a modulated signal dataset,the results show that the proposed method outperforms other clustering algorithms.In addition,the high separability of the features extracted by the clustering model is verified by visualizing the salient features of different modulation signals through the interpretability analysis of the clustering model.(5)For the situation that the deep learning model is difficult to realize the effective learning of new and old class signals at the same time due to the limited samples of old class signals in incremental signal classification under open electromagnetic environment,an incremental classification model for electromagnetic signals based on sample selection and multi-objective linear programming is proposed.The proposed model includes an adaptive class sample selection algorithm considering normalized mutual information and a multi-objective linear programming classifier.The former is used to maintain the model’s ability to recognize previous classes by selecting key samples,and the latter is used to enable the model to quickly adapt to new classes.At the same time,distillation loss and cross entropy loss are used to fine-tune the incremental model,which further alleviates the catastrophic forgetting problem of the incremental model.Through extensive experiments on RML2016.04 c dataset and ACARS signal dataset,the results of the experiments show that the proposed method achieves accurate classification of new class targets with few available old class samples,while maintaining the classification ability of old class signals.(6)Aiming at the engineering application of few-shot electromagnetic signal algorithms,a lightweight weight-variable scattering convolution networks is designed to improve the application efficiency of the deep learning method in real electromagnetic scenarios.The designed networks organically combines wavelet transform and convolutional neural networks to efficiently utilize wavelet scattering features and deep networks to achieve lightweight classification models while maintaining classification accuracy.Extensive experiments are conducted on ACARS signal dataset and ADS-B signal dataset,and the experimental simulation results show that the traditional classification methods based on wavelet scattering convolution networks and support vector machine classifiers have complicated parameter selection and low classification efficiency,while the proposed weight-variable scattering convolution networks can greatly reduce the computational complexity of the classification model while maintaining a high classification accuracy,and has good engineering application value.