Radar High Resolution Range Profile Target Recognition In Class Unbalanced Conditions

High-resolution range profile(HRRP)is widely used in the field of radar target recognition due to its advantages of including fine target feature information,easy acquisition and simple calculation.As a research hotspot in recent years,deep neural network has been widely used in HRRP target recognition.It can not only extract the deep separability features of target data,but also the model has strong scalability.It has been achieved significant results in tasks such as classification and detection.However,the classification and recognition based on deep neural networks are based on a large amount of complete data for repeated training of the model to obtain the best performance.In actual scenarios,because the acquisition of data is affected by many factors,the acquired training data is often incomplete,the number of categories varies greatly,the model is biased toward the majority category,and the generalization ability is poor on the minority category.Under the condition of unbalanced categories,the performance of traditional classification and recognition algorithms will be greatly reduced.In order to avoid problems such as the decline in the recognition performance of the classifier due to insufficient data and other reasons under unbalanced conditions,it effectively improves the generalization performance of the classifier in different categories.This thesis focuses on radar range profile target recognition under unbalanced conditions.The main work and innovation are summarized as follows:1.Introduce the related concepts of HRRP,the current research status of HRRP target recognition and the sensitivity of data,and introduce traditional classification and recognition algorithms based on template matching such as MCC and AGC,SVM classifiers,and algorithms based on deep neural networks such as MLP and CNN.Introduce three different unbalanced oversampling algorithms,SMOTE,Borderline-SMOTE and ADASYN,as well as evaluation indicators for unbalanced problems.2.From the data level,research on unbalanced radar HRRP target recognition based on Gaussian mixture-convolutional neural network.Aiming at the problem that the incomplete training samples in the unbalanced condition lead to the heavy classifier,a Gaussian mixtureconvolutional neural network is proposed.Because the traditional unbalanced sampling algorithm only performs random sampling in the feature space,the quality of the generated samples is different.In order to generate samples that are more in line with the original data distribution and improve the range profile recognition under unbalanced conditions,the Gaussian mixture is innovatively mixed The model is combined with the convolutional neural network,and at the same time,it uses the Gaussian mixture's ability to fit the original data and the powerful feature extraction ability of the neural network to perform feature extraction and classification.In addition,in order to improve the feature extraction ability of the convolutional neural network,a distance profile relationship weight matrix is introduced,and each distance unit of the distance profile is given different weights according to its importance,which effectively improves the feature extraction capability.Experiments with different unbalanced settings on the measured data,through the introduction of different evaluation indicators and visual display of the generated samples,fully demonstrate the effectiveness of this method in solving the unbalanced problem.3.From the algorithm level,research on unbalanced radar HRRP target recognition based on relational memory network.Aiming at the problem of the model's poor recognition ability for minority samples,and improving the model's preference for minority samples,a relational memory network is proposed.The method based on sample sampling needs to manually evaluate the quality of the generated data.Under the condition of large imbalance,the number of samples needs to be generated is large.The second-stage task increases the time complexity of the recognition task.The relational memory network innovatively uses additional memory storage modules to store the sample features that are difficult to classify in the model,which simplifies the process of identification and classification tasks.In the subsequent recognition process,the memory module is used for auxiliary classification,and the relational network is used to calculate the memory module.The correlation between the data in and the current sample uses prior knowledge to alleviate the imbalance problem due to missing samples.The algorithm has strong scalability,can be effectively combined with deep networks,and does not require pre-processing operations such as data expansion in advance.Experiments on different non-equilibrium conditions are performed on the measured data to verify the effectiveness of the method,and the parameters that affect the performance of the model are analyzed experimentally.