Article title: BIF-hosted deposit unit differentiation using multivariate Gaussian processes on measure while drilling data.
Authors: Silversides, K. L., Ball, A., & Melkumyan, A.
Journal: Applied Earth Science
DOI: https://doi.org/10.1080/25726838.2020.1829253
The following section was missed in the above published article. It is now included and article is re-published online.
4.3.2. Experiment 2
In the second experiment, the trained GP model was applied to all MWD points within 5m of the existing, exploration based contact surface. This experiment included points that were outside the intervals either side of the contact that were used for training. Therefore, this data can contain points that have MWD properties that are dissimilar to the training data, or similar to the wrong training data. For example, a point in a unit 1 shale band may have properties similar to the unit 2 training data. Therefore additional cleaning steps were required. These rules compared the points labelled in Rule 1 to nearby points. The GP model for this experiment used the first training data set from the last experiment, i.e. the data from G1 and G2.
Rule 2 was performed to reduce the number of points with labels inconsistent with those of the surrounding points. Any points that did not have a set number (β) of other points of the same label within 0.5m were removed. In this study we use β values of 1, 2, and 3.
Rule 3 was designed to reduce inconsistencies in down-hole classifications. The classification for each point was considered in relation to the other points within the same hole. If a point was inconsistent with the other points and the known stratigraphy, it was discarded. For example, a point labelled as the higher unit was discarded if it is below a significant number of points labelled as the lower unit. In practice, if a point was labelled as unit 1 (the top unit), the points located above it were considered. If less than a predefined threshold (γ) of these points were of the same unit the point was discarded. Similarly, if a point was labelled as unit 2 the points located below it were considered. In this study γ values of 70%, 80%, and 90% were used.
The three defined rules were varied to balance the number of identifications against the accuracy of the overall classifications. As this experiment used all MWD data, not just data from production holes where a label can be inferred from exploration data, ground truth labels were not available for the majority of the dataset. Therefore, the results were visually checked for consistency in the labelling within each hole and between nearby holes.