| Radar radiation source identification,as a crucial component in electronic warfare,can greatly enhance battlefield situational awareness and provide necessary intelligence support for electromagnetic interference and protection.However,traditional methods for radar radiation source identification are often limited by manual expertise,resulting in poor adaptability and difficulty in meeting the demands of modern battlefield environments.In recent years,deep learning has shown remarkable advancements in the field of signal processing,leveraging its powerful self-learning and generalization abilities to offer a new breakthrough in overcoming the limitations of traditional radiation source identification methods.This thesis focuses on deep learning-based radar radiation source identification and conducts the following research:(1)The current convolutional neural networks used for radiation source identification employ two different approaches to process the In-phase and Quadrature-phase(IQ)signals.The first approach involves transforming the IQ signals into images,which increases computational complexity,while the second approach extracts shallow features from the IQ signals,leading to lower identification accuracy.To address these issues,a method called Multi-scale Features Extraction and Features Selection Network(MFESNet)is proposed for radar radiation source identification.This method takes IQ signals as input and employs a multi-scale feature extraction network to extract both shallow features and multi-scale features from the IQ signals.A feature selection network is utilized to reduce the dimensionality of the multi-scale features,and effective transmission of shallow features is achieved through skip connections.Finally,a Softmax classifier is employed to classify the radiation sources.Experimental results demonstrate that compared to Residual Network-18(Res Net18),One Dimensional Convolutional Neural Network Long Short Term Memory(1D-CNN-LSTM),and Long Short-Term Memory(LSTM),the proposed method has lower computational complexity and exhibits superior noise resistance.(2)When training samples are imbalanced,training neural networks may lead to deteriorated performance in radiation source identification.To address this issue,a method called Auxiliary Classification-Conditional Variational Auto-Encoder-Generative Adversarial Network(AC-CVAE-GAN)is proposed for radar radiation source identification.This method utilizes an improved CVAE network to generate high-quality samples,thereby resolving the problem of limited samples in certain categories.By training the discriminator and auxiliary classifier,the network’s recognition accuracy for radiation sources is improved.Experimental results demonstrate that the proposed method is applicable for radiation source identification under varying degrees of sample imbalance.In comparison to Synthetic Minority Over-sampling Technique(SMOTE)+ deep learning network,Stacked Variational Auto-Encoder(VAE)+ deep learning network,and Auxiliary Classifier Generative Adversarial Network(ACGAN),the AC-CVAE-GAN network exhibits significant advantages in terms of noise resistance.(3)To address the issues of manual parameter tuning and additional computational resources required by traditional signal denoising methods,a combined approach of Sparse Convolutional Auto-Encoder(SCAE)and MFESNet is studied for radar radiation source identification.The network utilizes the Sparse Convolutional Auto-Encoder to perform denoising on radiation source signals,followed by the utilization of MFESNet for radiation source identification.Experimental results demonstrate that compared to improved wavelet threshold denoising + deep learning network and variational mode decomposition+ deep learning network,the proposed network can improve the recognition performance of the identification network to a certain extent. |
