A Hybrid GP Regression and Clustering Approach for Characterizing Rock Properties from Drilling Data
Hang Zhou, Peter Hatherly, Sildomar Monteiro, Fabio Ramos, Florian Oppolzer, Eric Nettleton
Technical Report: ACFR-TR-2011-001
Abstract: Supervised learning (including regression and classification) produces a function mapping inputs to outputs given some training data. It is difficult for supervised learning to learn models without a clear input-output pair connection. This is the case when characterizing rock types from drill performance data. There is not an obvious one to one correspondence between the Measurement While Drilling (MWD) data and the rock types due to the changing geology from site to site. In this paper, a hybrid classification approach is proposed by combining Gaussian Process (GP) regression with unsupervised clustering. A rock hardness characterizing measure - Adjusted Penetration Rate (APR) is extracted from the raw MWD data. GP regression is then applied on APR, followed by an unsupervised clustering which produces discrete class labels. The intermediate characteristic measure APR decomposes the complexity of the classification problem into a regression followed by an unsupervised clustering, where both are easily tractable. No initial labeling is needed. Comparisons have been made with alternative measurement Specific Energy of Drilling (SED) as well as the state of the art GP classification. Experimental results on real world mining site data have shown the effectiveness of our proposed approach.