Research On High Resolution Range Profile Air Target Recognition Technology Based On Semi-supervised Learning

With the urgent need of refined air situational awareness,radar automatic target recognition has become a hot topic in the research of radar system technology.Low SNR in the robust feature extraction and under the small sample training dataset which are both bottlenecks in the research and application of radar target recognition technology.In this paper,combined with the deep learning on the extraction of effective feature and the excellent performance of environmental conditions adaptation and high resolution range profile(HRRP)to obtain target information accuracy to carry out research work on radar automatic target recognition technology,obtained under the condition of low SNR and small sample identification of a more stable and reliable results.The main content and innovation points of this paper can be summarized as follows:(1)The scattering model of radar and the basic principle and model design of convolutional neural network are studied.The translation robustness recognition model RSRNet of HRRP is designed,and the experiment verifies that RSRNet has a strong robustness to the translation sensitivity of HRRP target,that is,the translation invariance of convolutional neural network in image feature extraction is still valid on HRRP.(2)A semi-supervised deep U blind denoising full convolution HRRP identification network SMTRNet is proposed.SMTRNet consists of a deep blind denoising convolutional neural network DUBDNet and an identification network RSRNet in series.DUBDNet can use the two frames of noise-containing HRRP adjacent to the acquired target as the input and output of the model to learn to achieve blind denoising.Experiments verify that DURDNet can increase the HRRP SNR by more than 10 dB when the HRRP SNR is lower than 5 dB.Among them,when the HRRP SNR is-15 dB,SMTRNet can increase the HRRP SNR by 15 dB,and the recognition rate is 15% higher than that of RSRNet.(3)The spatial distribution of the target HRRP is visualized using the manifold learning algorithm t-SNE.Combining the sampling visualization of HRRP with the recognition performance of RSRNet,it is concluded that for HRRP recognition,the completeness of attitude sampling is the key factor compared with the number of samples.(4)In order to solve the problem of incomplete HRRP attitude sampling for noncooperative targets,two direct push semi-supervised HRRP recognition algorithms are adopted.Semi-supervised convolutional self-encoding(Semi-CAE)performs embedded learning on the target feature by pre-training,and then trains the recognition classifier.The experiment verifies that the Semi-CAE compared with strong supervised learning algorithm RSRNet alleviates the problem of sparse attitude sampling.According to the popular assumption of semi-supervised learning,a simple and efficient deep label reconstruction(PL)algorithm is adopted to improve the existing depth recognition algorithm.In the case of electromagnetic simulation data with equal interval sampling recognition experiments with different azimuth sampling intervals,the recognition performance of Semi-CAE and Semi-CAE(PL)under a very small number of training samples can reach 70%.