Support vector machines for characterising Whipple shield performance

Support Vector Machines (SVMs) are a classification technique used in data mining and machine learning that are particularly well suited for application with sparse data sets. A database of over 1100 hypervelocity impact tests using spherical aluminium projectiles against spaced aluminium armour (i.e. Whipple shield) was compiled and used to train four different SVMs. The SVMs were developed using a variety of input-attributes and Principal Component Analysis (PCA). Initially, a maximum accuracy of 75% was obtained for an SVM when applied to predict the perforated/not-perforated outcome of impact events not included in the training process. A number of tests were identified which were inconsistent with the pattern observed for other training data. By removing this conflicting data (<5% of the total number of entries), significant increases in the training and generalization accuracy (83%) were achieved. The qualitative outputs of the SVMs were investigated through comparison with classical ballistic limit curves and test data. Within a velocity range of ∼3-8 km/s the SVMs demonstrated a good level of agreement with the classical ballistic limit curves and test data. The application of machine learning methods, including SVM, to predict impact outcomes is limited by the statistical quality of the training dataset. A broader and more homogenous distribution of test conditions, target geometries, materials, and outcomes (i.e. from well above to well below the ballistic limit) is required for machine learning to provide a high level of quantitative accuracy with consistent qualitatively output. Improvements to the training data set may be best achieved via a process in which the current SVMs are applied to identify the most valuable test conditions for future analysis.