Radar Compound Jamming Recognition Based On Multilabel Learning

The jamming environment faced by modern radar is becoming more and more complex,and compound jamming of various types and parameters are difficult to identify and suppress through the additive combination,which has become the main electromagnetic threat to radar.The traditional compound jamming recognition method is based on manual analysis and feature extraction,which has a low recognition rate,difficult operation and poor scalability,and can only recognize specific compound jamming patterns.For this problem,the classification of compound jamming by deep learning parties becomes a new solution with good application prospects.The multi-label classification model based on residual network（ML-Res Net）is proposed for the classification and recognition of composite interference in this thesis,which achieves accurate classification of compound jammings with better results than traditional feature extraction and single-label classification methods.For the problem of semantic segmentation of compound jammings,the compound jamming full convolutional network semantic segmentation model（CJ-FCN）is proposed to achieve the segmentation and estimation of jamming time-frequency distribution based on jamming type classification.The main content of the full thesis is summarized as follows:1.Design and creation of the compound jamming dataset.Modeling of 7 types of radar jamming,using additive combinations to generate multiple compound jamming.The compound jamming signal samples in uniform format are obtained by means of signal and image processing.Multiple sets of parameters are set at different noisejamming ratio points to construct compound jammings with different time-frequency overlap methods,and finally the compound jamming classification dataset and the semantic segmentation dataset are established.2.Compound jamming classification based on feature extraction.The time-domain,frequency-domain and time-frequency-domain features of compound jammings under different noise-jamming ratio points are analysed,and the features with high discrimination are selected for support vector machine classifier to achieve simple classification of compound jamming,which can be used as comparison results for multilabel learning method.3.Single tag/multi-tag recognition of compound jamming.Firstly,a unique label encoding is used for each compound jamming,and the classification recognition of compound jamming is achieved on the basis of residual network（Res Net）,and the results show that the accuracy of the model reaches 90.8%.Secondly,Multi-label learning is introduced on the basis of Res Net,and Multi-label compound jamming recognition algorithm（ML-Res Net）based on residual network is proposed.Using non-reciprocal label set annotation for compound jamming,the absolute accuracy of compound jamming classification can reach up to 95.5% with different label judgment thresholds,and the accuracy of label level exceeds 97.5%.Compared with the single-label classification model Res Net,the ML-Res Net model has higher recognition rate,faster convergence,classify more types of jamming,and is more scalable.4.Semantic segmentation in the time-frequency domain for compound jamming.In practical application environment,there are various combinations of compound jamming,the key information such as relative position information in the time-frequency domain needs to be estimating effectively.In this thesis,fully convolutional network FCN and improved model CJ-FCN are used to achieve segmentation of the compound jamming.The pixel-by-pixel prediction accuracy of the jamming reaches 98%.The relative position information in the time-frequency domain of the compound jamming is also obtained to provide support for the subsequent anti-interference processing.