Learning of a tracker model from multi-radar data for performance prediction of air surveillance system

A valid model of the air surveillance system performance is highly valued when making decisions related to the optimal control of the system. We formulate a model for a multi-radar tracker system by combining a radar performance model with a tracker performance model. A tracker as a complex software system is hard to model mathematically and physically. Our novel approach is to utilize machine learning to create a tracker model based on measurement data from which the input and target output for the model are calculated. The measured data comprises the time series of 3D coordinates of cooperative aircraft flights, the corresponding target detection recordings from multiple radars, and the related multi-radar track recordings. The collected data is used to calculate performance measures for the radars and the tracker at specific locations in the air space. We apply genetic programming to learning such rules from radar performance measures that explain tracker performance. The easily interpretable rules are intended to reveal the real behavior of the system providing comprehension for its control and further development. The learned rules allow predicting tracker performance level for the system control in all radar geometries, modes, and conditions at any location. In the experiments, we show the feasibility of our approach to learning a tracker model and compare our rule learner with two tree classifiers, another rule learner, a neural network, and an instance-based classifier using the real air surveillance data. The tracker model created by our rule learner outperforms the models by the other methods except for the neural network whose prediction performance is equal.Peer reviewe