Research On Air Target Recognition Technology Of Narrowband Radar

Based on the project "Narrowband Radar Target Recognition Verification and Evaluation System",this thesis studies the GPU-accelerated radar flying target recognition technology and its engineering implementation.Taking the specific requirements of the application scenario as the starting point,respectively studied the methods of flying target recognition based on traditional classifier and neural network.According to the index requirements of the neural network identification method in the project,designed a multi-channel fusion feature(Multi-channel Feature Fusion,MFF)network.And for the problem that the recognition algorithm requires a large amount of computation,completed the algorithm acceleration processing implementation scheme of feature extraction on GPU,which meets the requirement that the running time of the feature extraction algorithm does not exceed0.1ms.designed and developed the engine software architecture of the radar signal processing algorithm in C++ language,which provides a stable operation platform for the scheduling and driving of various algorithms in the project.The main research contents of this thesis are as follows:1.Firstly,this thesis established the narrow-band radar echo model of aircraft fretting parts,and deduced the analytical expression of the radar echo signal under ideal conditions.Secondly,respectively studied two kinds of traditional radar target classification and recognition methods based on feature extraction of the simulated aircraft target and the measured warhead target in the Doppler domain and time-frequency domain,compared and analyzed the recognition performance of the extracted features under different classifiers.Thirdly,in response to meet the project’s need to accelerate the effective features,designed and implemented the parallel processing scheme of the Doppler domain waveform entropy,the second-order central moment and the fourth-order central moment feature extraction algorithm on the GPU platform.After running the designed algorithm for many times,the average running time under GPU is 0.039 ms,which fully verifies that the designed scheme has the advantage of high real-time performance and meets the project requirements.2.Designed a MFF network according to the project requirements of the neural network identification method to ensure that the recognition rate is above 95%,and the network calculation amount is not more than 1GFLOPs.The MFF network is formed by adding crossresidual block connections to the residual neural network(Residual Neural Network,Res Net)and building a quadratic residual direct connection structure internally.The network with this structure incorporates multi-channel features to make it has better learning ability.Adding branches reduces the dimension of the directly connected part,reduces the complexity by half,and makes the network more lightweight.Finally,the network training set is constructed with the real and imaginary parts of the flying target echo and sent it to the network for training.The experimental results show that the recognition rate of the MFF model for three types of aircraft reaches 98.5 percent,The identification rate of the three types of warheads reached 97.4 percent,and the calculation amount is 0.92 GFLOPs.the experiments verify that the MFF network has good generalization and robustness.3.Designed and implemented the algorithm engine software of radar target recognition system.Firstly,introduced the whole target recognition system,and introduced the overall system design scheme of the algorithm engine software(Signal Processing Algorithm,SPA),including function analysis and software design architecture.Secondly introduces the software sub-modules of SPA in detail,and focuses on the algorithm processing module.Finally,introduced the software and hardware environment and algorithm effect of SPA software running.As the control platform of the scheduling algorithm component library,SPA is the core part of the target recognition system.The effect of engineering implementation shows that the software can stably call algorithm component libraries with different functions in sequence according to the algorithm flow table,it solved the problem of unified algorithm scheduling and path planning,and provided a reference solution for the engineering implementation of the system.