classification of the trajectories
the method of pattern recognition
artificial neural network
the probability of misclassification
At the present time, due to the rapid development of hypersonic technologies, increases the urgency of developing methods for solving the problem of classification of the trajectories of hypersonic aircraft (HSA) for atmospheric flight segment. The complexity of the problem of classification is based on the following factors. First, it is the development of HSA flight models, corresponding to different points of defeat. And, secondly, the complexity of solving the problem of classification (recognition) itself, due to the impossibility of formalizing the relationship between the parameters of the trajectory and the point of destruction.
The problem is solved with help of the proposed methods of pattern recognition, using learning «the teacher» on the set of classified samples – HSA trajectories in atmospheric flight phase, aimed at a variety of ground targets. A technique classifying HSA trajectories is developed on the basis of the minimum distance to the standard class and based on artificial neural networks like multilayer perceptron. Here are shown the results of computer simulations and calculations of probability of misclassification HSA trajectories of these methods with different number of classes of trajectories and constraints on decision time.
The results of these studies have shown that the method of minimum distance to the reference class and an artificial neural network like multilayer perceptron allow detection of strongly overlapping classes of trajectories (up to 30%) with an average probability of an incorrect decision less than 0.1 for a number of recognizable classes of 3...5, and restriction on time, decision making up to 60 seconds. At the same time of an artificial neural network like multilayer perceptron can obtain higher quality data (ie, smaller values of the average probabilities of erroneous decisions) compared to the method of minimum distance to the reference class, however, it requires a substantially larger additional calculations to optimize the network parameters in the learning process.