Parameter Estimation Of Aircraft Rotor Based On Multi-Task Convolutional Neural Network

Radar target recognition technology can obtain information about the attributes,types and models of aircrafts,and it is important for our army.The recognition task can be completed by traditional radar target recognition technology through template matching and other methods based on the template database.However,due to the non-cooperative nature of the target and the emergence of new aircraft,it is difficult to establish a complete template database,which will influence the traditional recognition technology.In addition,combined with prior knowledge of the target,the method of calculating aircraft rotor parameters from radar echoes can achieve model recognition without relying on template libraries.This method has clear mechanism and it is robust.It is difficult to obtain analytical solutions by the existing methods due to the highly non-linear relationship between rotor parameters and echoes.Therefore,searching methods are usually used to solve this problem.The application of the existing methods may be limited due to the large calculation amount and the high parameter coupling.It is difficult to meet real-time requirements.The deep networks have strong nonlinear fitting ability to describe the nonlinear relationship between rotor parameters and echoes.And they also have high computational efficiency.Therefore,this thesis studies aircraft rotor parameter estimation methods based on multi-task convolutional neural networks.The main work is as follows:(1)The rotor echo model is studied and simulated.Firstly,the echo model of the aircraft rotor is derived.Then,by analyzing the echo signal in the time domain,Doppler domain,and time-frequency domain,we obtain the characteristic information in different dimensions for rotor parameter estimation.Finally,the traditional rotor parameter estimation methods are introduced,including Hough transform,inverse Radon transform,and single-range Doppler interference.And the performance of these three methods is briefly compared and analyzed.(2)To make up for the shortness of traditional estimation methods in terms of real-time and independence,a multi-task convolutional neural network is proposed to realize rapid and independent estimation of multiple parameters at the same time.The multi-task convolutional neural network is composed of a multi-task shared module and three nonshared modules.For different tasks,the shared module is used to extract coarse-grained features,and the non-shared modules are used to extract fine-grained features.The proposed network can reduce the coupling between the parameters in the estimation process,and avoid the risk of estimation error transmission.Finally,it is verified by experiments that the proposed method can realize the real-time and independent estimation of rotor parameters ensuring a high rate of successful estimation.(3)In order to further improve performance of the rotor parameter estimation,a multi-task convolutional neural network based on the attention mechanism is proposed for rotor parameter estimation.The attention mechanism module SENet and SKNet are introduced in the multi-task shared module and the non-shared module,so that the features can be weighted by the network according to the characteristics of different rotor parameters.Therefore,it can enhance the useful feature channels and suppress the invalid feature channels.Finally,it is verified by experiments that the proposed method can further improve the successful rate of parameter estimation.