Estimating Downhole Vibration via Machine Learning Techniques Using Only Surface Drilling Parameters

Abstract Drillstring vibration can be divided into three types: axial, lateral and torsional. All three can cause significant wear and tear in drilling equipment, which leads to increased failures, non-productive time, and poor drilling performance. It also causes wasted mechanical energy and wellbore instabilities. Access to real-time, high-frequency downhole vibration data while drilling remains prohibitively expensive; however, it may be estimated via machine learning (ML) techniques using only surface drilling parameters. The task of predicting the severity of downhole vibration using surface parameters was approached as a supervised classification ML problem. Five basic, traditional techniques were investigated: the nearest neighbour, logistic regression, naïve Bayes, discriminant analysis, and decision trees. Drilling data was obtained from multiple bottom hole assemblies (BHAs) from several wells in North America. The learning tasks were separated into inter-BHA runs (where the learner is trained on data from one BHA and tested with data from a different BHA) and intra-BHA runs (where the learner is trained and tested with data from the same BHA). Severity of vibration was assessed primarily through the time-weighted average of root mean square amplitude and then classed into severity levels. Performance of the classification results was assessed using the predictive accuracy and weighted macro-average of precision obtained using cross validation and presented as confusion matrices for specific iterations of the cross validation. The classification ML for the intra-BHA runs produced overall predictive accuracies that averaged between 50% and 85%. Of particular concern is the misprediction of certain vibration levels as either lower or higher levels, even when overall predictive accuracy is high. The results show that these simple ML techniques can achieve considerable accuracy in the prediction of vibration levels for intra-BHA runs. For inter-BHA runs, predictive performance was reduced. This demonstration of the viability of ML in predicting bottom hole vibration motivates the application of more advance ML techniques, including deep learning estimators, and it signals the potential benefits that can be reaped.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.