Spectral Feature Selection for Automated Rock Recognition using Gaussian Process Classification

A spectral feature selection scheme is proposed for multi-class automated rock recognition from real world drilling data using Gaussian Process classification. This work is part of a larger project aimed at surface mine automation. The motivation for this research is to investigate which combination of drilling data measurements is most relevant for rock recognition. We conduct feature selection in the frequency domain where characteristics are more distinguishable. In particular, we extended the spectral feature selection method from binary classification to multi-class classification by decomposing the multi-class classification dataset into a series of one versus one binary classification datasets. A non-uniform discrete Fourier transform (NDFT) is then applied to data on each of the binary classification features, where the features with the most consistent “major bandwidth” across all decomposed binary classifications are selected. The approach has been applied on multi-class rock recognition (based on drilling data) and results are presented on real world drilling data.