ARTIFICIAL NEURAL NETWORK ANALYSIS OF SPACE DEBRIS

摘要

Models of the current and future space debris environment typically rely on simple isotropic break-up models for computational efficiency. However, historical break-up events (explosions and hypervelocity collisions) often result in an asymmetric dispersion of fragments that cannot be described by these simple models. This paper presents the investigation of new strategies that use artificial neural networks for analysing asymmetric break-up events in three scenarios. The first scenario was 'explosion classification'. A neural network pattern recognition algorithm is developed to classify fragment orbital data into one of three explosion categories (isotropic, radially enhanced or tangentially enhanced). This had a correct classification performance of 97% using the mean and standard deviation of fragments' apogees, perigees, inclinations, eccentricities and periods as inputs. In the second and third scenarios, continuous neural network models of (a) the ejection vector in an anisotropic explosion and (b) the relative velocity vector of a projectile in a hypervelocity collision were developed. The components of the ejection vector (scenario 2) were predicted with an average root mean square error of 0.43 and the components of the projectile velocity vector (scenario 3) were predicted with an average root mean square error of 2.45 km/s. For all three scenarios, break-ups were simulated and fragment orbits evolved using the Space Debris Simulation (SDS) software.