A Novel Trajectory Prediction Model Based on Kolmogorov-Arnold Networks (KANs) and Knowledge Embedding

This paper proposes a target trajectory prediction algorithm based on the Kolmogorov-Arnold Network(KAN), aiming to improve trajectory prediction accuracy,reduce model complexity,and extract target motion features.The model is trained in two phases:an offline phase where KAN learns general,nonlinear features from a large train setand an online phase where it quickly adapts to new,linear features from real-time data to make predictions.First,the KAN is used to learn nonlinear trajectory features from the train set,which may contain the flight path information of the aircraft and noise information,then extract and store these features in a knowledge base.This knowledge is embedded into an integrated KAN.Subsequently,linear trajectory features are trained online using the test set,ultimately enabling trajectory prediction.The train set and test set are derived from radar tracking data of two types of aircraft.The results show that the integrated KAN reduces loss by more than an order of magnitude under L1 norm and achieves stronger generalization capabilities compared to traditional LSTM,which is selected as the baseline model for experiments,at a radar pulse frequency of 20 Hz.