Multi-label Classification And Recognition Method Of Radar Radiation Source Signal Based On Implicit Feature Extraction

Recently,electronic warfare has become the core of modern high-tech warfare.Radar reconnaissance has the kinds of applications,high concealment and long reconnaissance distance,and is the main form of electronic reconnaissance.Radar reconnaissance,also known as radar signal identification or radar radiation source identification,includes two identification modes: radar radiation source type identification and radar radiation source individual identification.With the increasing number of types of radiation sources,the working mode,signal modulation method and electromagnetic environment of radars are becoming more and more complex,and the recognition of radar signal is becoming difficult.In addition,radiation sources often need to perform model identification and individual identification at the same time,but no multi-tag identification algorithm for radar signals has been proposed yet.Aiming at the above two problems,this paper firstly constructs a radar radiation source dataset,and uses three methods for extracting implicit features of radar radiation sources.The paper uses three multi-label classification algorithms to construct a multi-label radar radiation source based on implicit feature extraction.Classification and good experimental results are obtained on the radar radiation source dataset.The construction of radar radiation source dataset includes three steps: data acquisition,data preprocessing and data labeling.Among them,data preprocessing includes data preliminary screening and data normalization.Data preliminary screening deletes pulse data with serious missing content.Data normalization enables the data to be uniformly mapped to a uniform [0,1] interval;data annotation includes single There are two labeling methods: label labeling and multi label labeling.Implicit features are the internal implicit features in radar pulses,which contain the most representative and effective information of the radiation source.In this paper,time-frequency analysis and variational information bottleneck are firstly applied to the extraction of implicit features of radar radiation sources,and then a feature extraction method based on pulse amplitude prediction is proposed?Finally,the three feature extraction methods are fused to construct the best implicit feature extraction model of radar radiation source:1)The paper proposes implicit feature extraction method based on pulse amplitude prediction.This method masks a certain or part of the pulse points,and then trains the network model to predict the masked points according to the data around the pulse points,so that the model learns the sequence of the hidden points.Intravascular correlation characteristics between2)The paper proposes integrate the time-frequency analysis and variational information bottleneck with the pulse amplitude prediction method to construct a radar radiation source implicit feature model based on pulse amplitude prediction,time-frequency analysis and variational information bottleneck.The three methods are integrated and supplemented with each other.Experiments show that the feature information extracted by the fusion model is the most effective.The radar radiation source multi-label data set contains individual label and model label information.In order to realize the multi-label classification and recognition of radar radiation sources,this paper applies label fusion and multi-label loss to the radar signal multi-label classification network,and proposes a hierarchical fine-tuning based method.A neural network multi-label recognition method.Compared with the existing advanced radar radiation source classification and identification models,the multi-label loss and hierarchical fine-tuning model achieves better experimental results:1)The paper uses multi-label recognition based on multi-label loss and proposes method of transforming hierarchical multi-label problems into parallel multilabel problems.The algorithm achieves an excellent multi-label classification effect under a relatively simple network structure,and the model training speed is fast and consumes less resources.The experimental results show that the classification accuracy of the model is about 95.7%.2)The paper proposes multi-label classification method of neural network based on hierarchical fine-tuning.This method belongs to the algorithm adaptation method.By constructing a hierarchical deep neural network,the model is adapted to hierarchical multi-label data.The algorithm does not require any label transformation on the data,and the model reconstruction will train and predict the labels of the two levels separately.The experimental results show that the classification accuracy of the model is as high as 96.1%...