Application Of Incremental Learning In Radar HRRP Target Recognition

Radar has the capability of radio detection,which can realize monitoring,tracking target and judgment of the target type.It plays an important role in judging enemy conditions on the battlefield.Due to the needs of information warfare,radar automatic target recognition(RATR)has developed rapidly.High resolution range profile(HRRP)contains rich details of the target and is easy to acquire and process,which has gradually become a research hotspot in the field of RATR.In recent years,deep learning theory has developed rapidly,and the work of combining neural network models with HRRP recognition has also made a lot of progress.The premise of network model identification is to establish a complete radar database before training.After the training,the model parameters will be fixed and will not be updated.In practical applications,due to difficulties in obtaining non-cooperative target samples,it is often impossible to establish a complete database in advance.To this end,the model must be able to continuously learn with new samples during use.However,when a new sample arrives,if the new sample is used to update the model parameters directly,the model will often gradually forget what has been learned.To solve this kind of problem,people have proposed an incremental learning method,which enables the model to learn new knowledge while still retaining the old knowledge that has been learned.This thesis studies the incremental learning problem in HRRP recognition.The main research contents are as follows:(1)Based on the scattering point model,the characteristics,sensitivity issues and pretreatment methods of HRRP are introduced.The classic statistical recognition model and the currently popular deep neural network model are introduced,including the maximum correlation coefficient,adaptive Gaussian classifier,and multi-layer perceptron and convolutional neural network.Using the above method,the HRRP recognition experiment with a complete database was conducted,and the performance differences between the various methods were compared and analyzed.(2)For the application scenarios where the number of target classes in the incomplete database is fixed,a study on the fixed-class incremental learning problem is carried out.Dynamically expandable networks(DEN)has some shortcomings,including the unrepresentative selection of trained neurons and the simple use of L2 distance to calculate the amount of neuron drift.To this end,this thesis improves DEN and proposes an expandable networks with mask(ENWM).The innovation of the network is:(a)In order to make the network better use new knowledge and reduce the forgetting of old knowledge,the threshold mask is obtained by using the training mask matrix,and it is applied to the network weight of the initial task to obtain the task subnet required for subsequent network expansion;(b)In order to better assess the impact of neuron changes on incremental learning tasks,a neuron drift calculation method based on Fisher information matrix is used.Simulation experiment results show that ENWM can well learn new knowledge while retaining old knowledge,which improves DEN's recognition rate of HRRP in a fixed-class incremental learning scenario.(3)For the application scenarios where the number of target classes in the incomplete database is gradually increasing,a multi-class incremental learning problem is researched.end-to-end incremental learning(E2E)has some shortcomings,including the unreasonable setting of the distillation coefficient and the selection of representative samples,and the imbalance of the number of samples in the new and old classes,which leads to the classification preference of the network output.To this end,this thesis improves the E2 E method and proposes an reinforced end to end incremental learning(RE2E)method.The innovation of this method is:(a)To solve the problem of the proportion of distillation losses in the loss function,use the distillation coefficient that changes with the ratio of the number of new and old classes;(b)To select the most representative samples of target classes,use the probability value ranking list formed by the softmax layer output results to select samples;(c)To solve the classification preference problem,add bias correction layer to the network completes the correction of the output of the classification layer.The simulation experiment results show that the RE2 E method can learn and retain the knowledge of the old and new classes well and alleviate the classification preference problem,which improves the recognition rate of HRRP by the E2 E method in the multi-class incremental learning scenario.