Anomaly Detection And Radar Recognition Based On Variational Autoencoder

The development of deep learning is in full swing,and many fields are trying to use deep learning to solve current challenges.Variational autoencoders occupy a place in the field of deep learning because they can construct nonlinear models of neural networks and achieve random sampling.Relying on the ability to expand the hidden space and the characteristics of being sensitive to external changes,the variational autoencoder can show its own unique ability in many important research directions.This paper focuses on the anomaly detection technology based on the variational autoencoder and the radar target recognition technology to solve the problems of anomaly detection and radar target recognition.Aiming at the problem of separation between feature extraction and anomaly detection in anomaly detection,this paper proposes a deep support vector data description model based on variational autoencoders.The model maps the features to the latent space through the encoder of the variational autoencoder,and performs spherical constraints on the features in the latent space,so that the model learns the description of the features,and at the same time,the decoder of the variational self-encoder ensures the reconstruction effect of the input data.The model proposed in this paper can jointly optimize the two processes of feature extraction and detection of anomalies,avoiding the mismatch between the extracted features and the back-end classifier.It is ensured that all inputs will not be mapped to the same point in the hidden space in the description of the depth support vector data,and the phenomenon that the radius of the hypersphere is zero,that is,the collapse of the hypersphere is avoided.The anomaly detection experiments on the handwritten data set and the CIFAR-10 data set prove that the model proposed in this paper has a good anomaly detection effect,and the German traffic sign data set verifies that the model has high anti-robustness at the same time.Through the ablation experiment,the situation of replacing the variational autoencoder with other autoencoders was analyzed to prove the unique performance of the variational autoencoder in the whole model.Finally,the experiment proved that the model has good parameters robustness.In order to solve the problem that the shallow model in radar target recognition cannot extract deeper classification information,this paper proposes a frame classification model based on a variational autoencoder.The model includes encoding and decoding.The encoder projects the observation data into the deep probability hidden space to obtain the hidden features of the data;the decoding network describes the generation process from the hidden features to the observation data according to the number of frames,and each decoder determines the frame to which it belongs.The data has a better description of the data from the same frame,but the ability to represent data of other frames is weaker.The test sample category attribute can be judged based on the minimum reconstruction error criterion.Experiments based on the handwritten data set MNIST and the measured one-dimensional high-resolution range profile data set verify the performance of the proposed model.On this basis,the K-nearest neighbor constraint is added to avoid the unreasonable classification situation that occurs when only the minimum reconstruction error is used for classification.The experimental results of the measured one-dimensional high-resolution range profile data set prove the effectiveness of increasing the K-nearest neighbor constraint.