Problem
Rogii builds steering software for horizontal oil wells. The drill must stay inside the oil layer as long as possible, but the layer shifts underground. To follow it, the gamma-ray log of the new well has to be matched to the log of the reference (exploration) well that found the layer. Plain comparison fails: logs are stretched or compressed, tools differ in signal power, and geology varies between wells. Labels are scarce, and the result must run fast in a C++ product.
Example from the article: the reference well shows a layer at 2153 m. The new well, at 3100 m of pipe, was drilled at an absolute depth of 2150 m, so the layer has moved up by 3 m and the trajectory must change.
Solution in brief
For the vertical part, an encoder trained with triplet loss on synthetic triplets (crops and distorted copies) maps signals into a space where the nearest reference window is the match. For the horizontal part, both logs become one three-channel image and a U-Net predicts a mask that gives the layer shift. Both ship via C++ or ONNX.
Questions this case answers
- How to match a new signal sequence to a reference sequence from another source without labels, by training an embedding on distorted copies of the same signal
- How to turn two one-dimensional logs into one image and predict an offset with a segmentation model, when labeled offsets are scarce and the result must run inside a desktop product
- How to match a freshly drilled well section to a reference well by gamma-ray signal when labels are missing
- How to train a signal-similarity model without labels by distorting the original signal
Tools and methods
- Metric learning with triplet loss — matches a new vertical section to a reference well
- Signal augmentations — build positives
- U-Net with MaxViT encoder — predicts the layer shift as a mask
- segmentation_models.pytorch, timm — training
- xtensor (C++), ONNX — deployment in the C++ product
Google or a one-time email code. No password.