Neural-Based Classifier Evaluation for Maneuver Type Identification of Flying Target

Nowadays, operational support systems are rapidly advancing with the integration of machine learning methods. Specifically, the surface-to-air engagement decision-making process requires an accurate maneuver type identification for flying targets. In this paper, various neural-based classifier methods are utilized and compared to improve the performance of flying target maneuver classification. Furthermore, the relation between input maneuver lengths and classifier performances is analyzed. For this purpose, four neural-based classification methods are selected: convolutional neural network (CNN), recurrent neural networks (RNN), long short-term memory (LSTM), and gated recurrent units (GRU). Our experiments are conducted on a time-varying maneuver trajectory dataset that is syntactically generated using Newton-Euler equations in three dimensions (3D). Our dataset comprises three common maneuver types: level acceleration, loop, and snake. We introduce trajectory tilting with varying interference levels to enhance its diversity and robustness. Furthermore, we augment the dataset by injecting white Gaussian noise, allowing us to analyze the impact of methods on real scenarios. Throughout our experimental process, we employ a five-fold cross-validation approach. We report each classifier's performance in terms of accuracy, precision, and recall. The results show that LSTM is a promising method even when dealing with a noisy and time-varying maneuver trajectory.