[First Hit]

2019 AAPG Annual Convention and Exhibition:

Datapages, Inc.Print this page

An Application of Dynamic Time Warping in Rapid Depth Shifting to Improve the Quality of Machine Learning Training Datasets


The inherent differences between core depth and log depth require that core data be depth shifted to log depth in order to maintain data quality. In order to create log parameter and core analysis training datasets of sufficient sample size for machine learning the depth shifting workflow becomes human capital-intensive. This study automated core gamma normalization and applied Dynamic Time Warping (DTW) to the depth shift workflow with Python. DTW is an algorithm commonly used to align time-series datasets to find the optimum, or lowest distance, match. Properties of the wireline gamma logs were sampled near core collection depths so the tabulated core gamma data could be normalized for midline, maximum, and minimum. Core gamma data was then interpolated and re-sampled on the same sample interval as the well log to produce a core log in LAS format. For longer cores the core and well logs were smoothed and downsampled before using DTW to ensure efficient processing. Two rounds of shifting with DTW were utilized. The first round compared the core log to the well log inclusive of user-defined boundary conditions and shifted the core log on the median shift distance. The second application of DTW decreased the boundary conditions to the core length and compared only the peaks of both logs in order to fine-tune the shift at significant log features. These peaks were selected with a floating window mean ± standard deviation filter. The log curves were plotted for the user to visually confirm the shifts before the core log shift information was saved in LAS format. These shift files were then used to batch shift core analysis data prior to import into geologic mapping software. This Rapid Depth Shifting workflow provided a dramatic decrease in the time investment to depth shift core data while maintaining the quality of the machine learning training datasets by sampling well logs parameters at locations of core analysis accurately.