Research On Target Track Recognition And Prediction Based On Deep Learning

Modern air combat is one of the main ways to achieve multi-target defense and attack in the complex electromagnetic environment,and it is a form of deep joint attack that gradually dominates the modern battlefield.Traditional target track recognition and prediction research algorithms are not enough to deal with the massive data under the current battlefield for effectively identifying the threat of air targets.In order to mine effective and useful information from the complex,massive,and diversified battlefield data,we conduct research on the recognition and prediction of radar target track based on deep learning technology.The main contents of the thesis are as follows.(1)Radar target track recognition based on space-time relationship of tracks.Considering the spatial and temporal relationship characteristics of track point data,we propose a radar target Track Recognition algorithm based on Spatial-Temporal relationship of track points(TRST).First,we selects the important attributes from the track point data,then we mines the space value characteristics of the track point data in the spatial relationship,and finally we constructs a recurrent neural network(RNN)to further capture the spatio-temporal relationship characteristics of the track point data to achieve the classification and recognition of the target track.Simulation results show that the proposed algorithm TRST can effectively improve the accuracy,precision,recall and balanced average performance of target track recognition.(2)Radar target track prediction based on time residual self-attention.Considering the different effects of different track points on the predicted track points and the different effects of the distance prediction time on the prediction results,we propose a selfattention mechanism based on time residual.The attention mechanism takes the time interval between different track points as time residuals,and introduces a time sequence mask composed of the time residuals between all the points in a track to focus on the information of the track points near the predicted time when calculating the probability of attention allocation.On the basis of the proposed elf-attention mechanism,we propose a Track prediction model based on time Residual Self-Attention(TRSA).The model uses multi-head attention to calculate the attention features between the track points,which is composed of a lot of time residual self-attention,and obtains the final track feature vector through feature fusion,fully connected neural network,and multi-head attention.Finally,the track prediction is completed by a full connection layer with an activation function of Softmax.Simulation results show that the model effectively realizes the function of track prediction and reduce the error of track prediction,the root mean square error,the average absolute error and the running time of the model.