Study On Deep Features Selection Method And Its Application In Emitter Signal Recognition

In the field of the intelligent Radar Emitter Signal(RES)recognition,feature extraction is the focus of scholars.With years of accumulation,the size of artificial feature space is expanding.These artificial features can get better recognition effect in specific environment.In contrast,the importance of feature selection is ignored.In fact,how to select an effective feature subset from the artificial feature space has become a great concern for the experts.Cascading multiple artificial features to reconstruct the radar signal may not improve the recognition performance,and also increase the time cost of the recognition process.Research shows that feature selection can effectively improve the prediction performance of the algorithm.Thus,the research on feature selection algorithms has practical significance for the recognition of RES.There are two ways of feature selection: feature ranking and subset filtering.In the former,each feature is ranked according to an evaluation function,and then the first some features with the highest score are selected to be feature subsets.In the context of improving the performance of learning algorithm,the famous criticism of feature ranking is that combining features with high scores does not always provide better performance.The sparse learningbased method is the representative research direction of the subset filtering.The basic idea of the method is to make the score of the feature tend to zero,and then select the features with nonzero score to be a subset.However,the traditional sparse feature selection models are usually linear,which can't deal with the nonlinear problem(for the classification problem,that is,the linear inseparable problem).Unfortunately,nonlinear problems are common in practical applications.Rapid advances in deep learning provide a research direction for feature selection.Deep Neural Network(DNN)which is the basic component of deep learning,and is regarded as a very effective model to solve the nonlinear problem.Thus,the restructured deep network with the additional technologies(structure)can recognition features.Here,the feature selection algorithm based on DNN and deep learning technology is called Deep Feature Selection(DFS).For the topological structure,a new DFS with paired connection layer structure is built in this paper.The construction idea of the new DFS is as follows: build a fully connected network with the same size of the input and output layer,which is used to output the feature scores;a decision layer is added to make decisions on samples and calculate training errors in the supervised learning scenario;a pairwise connected layer is constructed to connect the former and the latter structures.To get sparse feature scores form the new DFS,the hard threshold function is applied to the output layer of the fully connected network,which are regarded as the activation function.This makes the new DFS have two advantages: the new DFS does not add additional sparse technologies or structures;the sparse feature scores can be obtained by designing the activation function flexibly.Further,an DFS model based on the threshold value adaptive-learning is established.By embedding the soft threshold and garrote threshold function,the threshold value can participate in the learning of DNN.As the deficiencies of hard,soft,and garrote threshold function,an adaptive soft threshold function(AST)is designed.Finally,design a loss function with redundant terms to reduce the redundancy of feature subsets,so that the new DFS can recognize redundant features.Thirdly,as feature selection is a typical data mining technology,it requires strict consistency of data.Therefore,adding data cleaning steps can effectively improve the performance of the algorithm.Data cleaning usually needs to remove the samples that deviate from the data distribution seriously(these samples are called anomalies)without prior knowledge.Thus,an unsupervised anomaly detection algorithm based on principal component analysis(PCA)is proposed in this paper.The proposed method maps the original feature space to the new principal component feature space based PCA technology,and the outliers are often easy to detect in the new space.Thus,a new anomaly evaluation model is established in the principal component space to score the anomaly samples,and anomaly detection is realized based on the scores.Finally,this paper focuses on the application problem of DFS,and identify RES by using the proposed DFS.The existing RES recognition methods stay within the framework of“feature extraction + classifier design”,in which feature extraction is the key.With decades of research,a lot of effective radar signal features have been proposed.Therefore,the accumulated artificial features are faced with the problem of feature selection,that is,how to select a feature subset to identify RES effectively.This is the problem of artificial feature selection in the feature layer.As DNN is good at handling the unstructured data,selecting features directly on the signal level is proposed in this paper.The features selection on signal layer does not need to build complex artificial features and retains a large amount of information of the signal.At present,there are few researches on this way,because the traditional feature selection algorithm is difficult to work on high-dimensional radar signals.To sum up,a supervised DFS model is built in this paper,including the establishment of the topological structure,the design of new activation function and loss function.The new DFS is an end-to-end model and the threshold value can be can learned adaptively.Experimental results on synthetic datasets and benchmark datasets show that the proposed method has Better feature recognition performance than the traditional algorithms.The proposed method can effectively improve the recognition performance of RES.