Radar One-dimensional Range Profile Target Recognition Based On Deep Neural Network

The range resolution of high-resolution broadband radar is much smaller than the target size.The radar echo signal is called the target's one-dimensional high-resolution range profile(HRRP).The diversity of HRRP and its sensitivity to small changes in the target make it difficult to capture and distinguish the characteristics of the target.Therefore,improving the performance of the HRRP recognition method has important practical significance.The idealized thinking of independent modeling of each frame in traditional recognition methods limits the generalization of the algorithm,resulting in a sharp decline in the performance of the model under the conditions of small samples,unbalanced data sets and low signal-to-noise ratio.Based on the above problems,this thesis proposes a HRRP recognition method based on deep embedded neural network.The main contents are:(1)We have done research on the recognition performance of traditional methods and analyzed the average recognition performance of traditional methods under small samples or low signal-to-noise ratio conditions.Experiments have proved the shortcomings of traditional methods in practical application scenarios and laid the foundation for the later research.(2)Recognition based on HRRP envelope characteristics.Aiming at the problem of the loss of valuable feature information in the feature extraction process caused by the traditional method using sub-frame independent modeling,a HRRP recognition method based on multi-scale convolutional neural networks(CNN)is proposed.Aiming at the problem of poor robustness of traditional methods in low signal-to-noise ratio environments,an HRRP recognition method based on Denoising Auto Encoder(DAE)is implemented,which uses pre-training to extract features and then fine-tune parameters for classification.(3)Recognition based on the structural characteristics of HRRP.Aiming at the problem that CNN and encoder-decoder structures cannot extract HRRP timing features,a HRRP recognition method based on Recurrent Neural Network(RNN)combined with self-attention mechanism is proposed.(4)Propose a HRRP recognition method based on deep embedded neural network.Aiming at the mutual restriction between the input sequence structure and time step in the RNN modeling process and the network easily neglecting some distance units that contain valuable information but small amplitude,it is proposed to use a CNN module that combines dynamic adjustment layers,convolutional layers,and Squeeze-and-Excitation layers to extract high-level features of HRRP,and then embed them into a stacked bidirectional RNN recognition network The convolutional neural network extracts high-level features of HRRP.This thesis also proposes a multi-level attention mechanism based on stacked bidirectional RNN to adjust the importance of the local structure of the target reflected at different levels.The experimental results based on the measured data verify that the method proposed in this thesis has good recognition performance in the case of a large amount of data,and has advantages over other methods when the training data set is small or the signal-to-noise ratio is low.