Projectile Launch Point Prediction via Multiple LSTM Networks

Abstract

Launch point prediction of projectile targets is the main task of ground-based air surveillance radars in military services. Traditionally, the radar captures the portion of the projectile’s trajectory to obtain time series data of position and velocity, and then uses an extrapolation method to determine the launch point. This work aims to enhance the accuracy of launch point prediction from the radar measurement data through the development of multiple Long-Short Term Memory (LSTM) networks. The proposed method consists of three stages. First, an LSTM-based filtering model is built to filter the noisy radar data and estimate the flight trajectory states. Second, an LSTM-based classifier model is created to identify the class of munitions based on their flight trajectories. Finally, multiple LSTM networks are developed to serve as the launch point predictor, in which each network is trained on different sets of munition classes. A switching law based on the classification result was established to select the best launch point predictor network for a specific class of munitions. The prediction results demonstrated the promising performance of the proposed method in launch point prediction. Furthermore, the correlation between the data distribution and the predictive performance of the network was explored.