Detecting drilling dysfunctions from surface data is not always easy as downhole vibrations tend to get damped before they reach surface sensors. Building machine learning models to recognize patterns in the surface data requires vibration signals captured by downhole sensors for training purposes. Such datasets are not widely available and therefore a methodology to expand these datasets is highly desirable. This work explores ways to utilize data augmentation to artificially diversify and increase datasets to build better models.
Stick-slip (including full-stick), bit bounce, whirl, and bit balling are the primary dysfunctions considered in this work. Bayesian networks are used as classifiers to keep the model intuitive, and address situations where some input data is missing or unavailable. Once the dysfunction events in the downhole dataset were labeled, data augmentation techniques were used to generate synthetic data for scenarios where data was sparse.
The dataset used in the project consisted of nine wells (with 19 bit runs). Most of the bit runs had a downhole vibration sensor at the bit, while some had sensors along the string as well. Of these 19 bit runs, 15 were used for training and four were used to test the models. Various data augmentations techniques were applied and validated manually as appropriate synthetic data. In the case of full-stick event detection, the saw tooth pattern in the surface torque signal was captured and provided as an input to the classifier. The classifiers thus trained were able to detect the dysfunctions using data from surface sensors to a high level of accuracy and with low false alarm rates.
This paper presents models to predict downhole dysfunctions from surface data alone. This paper also provides guidance on data augmentation techniques that use sparse downhole datasets to improve machine learning drilling advisory models. For identifying drilling dysfunction from surface data, the tortuosity of the well is also taken into account.