• Home
  • Current congress
  • Public Website
  • My papers
  • root
  • browse
  • IAC-19
  • A6
  • 1
  • paper
  • Development of a High Fidelity Simulator for Generalised Photometric Based Space Object Classification using Machine Learning

    Paper number

    IAC-19,A6,1,5,x52530

    Author

    Mr. James Allworth, Australia, The University of Sydney

    Coauthor

    Dr. Lloyd Windrim, Australia, The University of Sydney

    Coauthor

    Mr. Jeffrey Wardman, Australia, EOS Space Systems Pty Ltd

    Coauthor

    Dr. Daniel Kucharski, Australia

    Coauthor

    Dr. James Bennett, Australia, Space Environment Research Centre Ltd. (SERC)

    Coauthor

    Dr. Mitch Bryson, Australia, The University of Sydney

    Year

    2019

    Abstract
    This paper presents the initial stages in the development of deep learning classifier for generalised Real Space Object (RSO) characterisation that combines high-fidelity simulated light curves and transfer learning to improve the performance of object characterisation models that are trained on real data. The classification and characterisation of RSOs, is a significant goal in Space Situational Awareness (SSA) due to the coupling between an object’s characteristics and the external forces acting on it.
    
    Deep learning techniques have been successful in a variety of domains, and it is theorised that a data-driven deep learning approach would enable quick determination of RSO classes made directly from observational data.  Deep learning models have been shown to outperform traditional methods when training sets are large as this enables them to learn complex nonlinear features that are difficult for humans to recognize and extract.  Observational RSO light curve data however is both difficult to obtain and label, consequently, the size and quality of the dataset limits the achievable performance and robustness of the neural network.  
    
    A high-fidelity light curve simulation environment enables the possibility of creating a large, well-labelled dataset that could be used to augment RSO light curve datasets used for training neural networks.  Provided the simulated light curves encapsulate similar features to those found in a real light curve dataset, part of this pre-trained model could then be retrained on a small real light curve dataset in a process known as transfer learning.  In a similar manner to data augmentation, transfer learning is known to be an effective way of increasing training performance, particularly in cases where the real dataset is small.  
    
    In the literature neural networks have been applied to light curves simulated using a fixed starting epoch/orbit, initial rotation, initial angular rotation, observing site, sample rate and observation length.  These restrictions significantly simplify the classification problem by removing some of the variations that are present in real light curve data.  
    
    Initial results applying a convolutional neural network to a dataset of simulated light curves with the above restrictions varied, show that neural networks are able to perform classification with an accuracy of 88% on a test set with 7 classes of object. Further work will be performed improving the simulation environment and investigating alternative machine learning techniques.  Ongoing work with industry partner Electro Optic Systems will enable validation and testing on a dataset of real light curves.
    Abstract document

    IAC-19,A6,1,5,x52530.brief.pdf

    Manuscript document

    IAC-19,A6,1,5,x52530.pdf (🔒 authorized access only).

    To get the manuscript, please contact IAF Secretariat.