Research On Radar Modulated Signal Recognition Method Under The Condition Of Sample Deficiency

Radar signal modulation type recognition plays a crucial role in electronic reconnaissance.It provides precise identification results for electronic reconnaissance,thereby helping one’s own side gain certain information advantages in battlefield games.With the growing sophisticated electromagnetic environment,the modulation methods of radar signals are becoming more diverse.Currently,most of the existing deep learning based radar signal modulation recognition methods are based on offline training.Once the training is completed,the specific categories that can be recognized are clear.When the receiver intercepts a new radar signal that has not appeared before,the model will classify it as the signal with the closest feature in the sample library.At the same time,the proportion of new system radars appearing on the battlefield is gradually increasing.As a typical non cooperative signal,radar signals are difficult to accumulate and obtain samples.However,the widely used deep learning based radar signal modulation type recognition method has a strong dependence on the number of signal sample libraries.This article uses the ideas of metric learning and Few-shot learning to design and solve the above problems.The main work is as follows:1.Firstly,this article provides a unified mathematical model for the radar modulation signal intercepted by the receiver,and divides it into different modulation types,combining them with mathematical models for simulation analysis.Because radar signals belong to nonstationary signals,they are transformed into time-frequency domain for joint analysis to better reflect the relationship between frequency and time.Subsequently,the advantages and disadvantages of several commonly used time-frequency analysis methods in time-frequency analysis were discussed from factors such as time-frequency clustering and cross term interference,and a smooth pseudo Wigner Ville distribution time-frequency analysis method was introduced.The simulation in subsequent chapters will perform smooth pseudo Wigner Ville distribution on the signal as preprocessing.2.A new modulation type radar signal discrimination algorithm based on SER block and triplet loss is proposed to address the misclassification problem of unknown radar modulation signals.Firstly,a network is constructed using SE structure and residual basic unit blocks.Then,samples from eight common radar modulation signals are preprocessed and inputted into the network after SPWVD time-frequency analysis.The network is trained and the samples are mapped to a specific 128-dimensional embedding space.And use the triple loss function to mine semi difficult samples to calculate the loss adjustment network parameters,so that the samples within the class are more clustered and the samples between classes are more dispersed in the embedded space.Take a certain sample from the training set that is closest to the center feature vector of each type of radar modulation signal to form their own sample library,and use the method of "different categories are pseudo unknown,and the same category is pseudo known" to determine the size of the distance threshold using the signals in the training set.In the testing phase,a distance-threshold based discrimination method is used to determine whether the tested sample belongs to a certain type of signal in the sample library.If it belongs to a certain type of signal in the sample library,the recognition result is output.If it belongs to a new modulation type signal,it is placed in the unknown radar modulation signal sample library.Algorithm simulation shows that this method has excellent recognition performance for8 known radar modulation signals in low signal-to-noise ratio environments,and the discrimination accuracy for 6 new modulation types of radar signals reaches over 89% at-8d B.3.A radar modulation signal recognition algorithm based on Few-shot learning is proposed to address the issue of existing deep learning based radar modulation signal recognition algorithms having a strong dependence on the training sample library,resulting in severe performance degradation when the training samples are extremely insufficient.By utilizing twin networks and comparative learning to design network structures,images of eight common radar modulation signals subjected to SPWVD time-frequency analysis are input to an iterator.The iterator trains the input network based on a balanced number of positive and negative samples composed of labels,effectively reducing the recognition algorithm’s requirement for the number of target samples to be classified by fully utilizing the class samples included in the training set.The recognition results indicate that this method can achieve high recognition accuracy when only providing a few samples.