Machine Learning Lessons Learnt in Stick-Slip Prediction

Abstract

Downhole vibration remains a major challenge for drillers. Today, there is technology to look at the problem from a unique perspective. A novel look at the problem focuses on evaluation of machine learning algorithms to predict downhole vibrations. Prediction is the first step in a longer road map. The goal would be to find an optimal combination of revolutions per minute (RPM) and weight-on-bit (WOB) to remedy drilling vibration in real-time, hence closing the loop. Drilling mechanics data for thousands of wells, acquired over more than ten years was analyzed. Some preparation of the drilling mechanics data was required. Data cleaning was first performed. This included corrections for time-dependent nature of the data. Data imputing for missing values and handling of outliers and anomalies was also performed in this stage. This was followed by feature engineering which included adding variables based on company-wide drilling domain expertise. Variables to capture data patterns and variables for better capturing the time-series dependencies were also created in this stage. This paper will discuss methodologies and general rules that were tested for preparing unstructured drilling data. A few of the machine learning algorithms used as building blocks of our full solution are gradient boosting and random forest. Deep learning models were also tested and the value of these are compared. The results were compiled to decide the best algorithm which could further be used to fine-tune optimum performance. The time series aspect of the data is captured in a moving window. As the window increases, the performance of each algorithm also varied. Also, evaluation of the benefits and drawbacks of each algorithm for the drilling predictions is detailed. Ways to improve the accuracy of prediction for downhole vibrations is also suggested with reference to the results showing the logic behind all recommendations. There will be a summary of the details of each finding and a short discussion on the way forward for the industry.

Detecting downhole vibrations through drilling horizontal sections: machine learning study

“…Gupta et al 34 employed the random forest and gradient boosting ANN for the stick–slip classification model using T, ROP, and WOB as model inputs. The overall model accuracy is 62% for detecting the stick–slip pattern.…”

Machine Learning Solution for Predicting Vibrations while Drilling the Curve Section

“…Based on the training phase, the models’ accuracy presented a determination coefficient between the predicted and actual measurements of 98 and 93% for the torsional and lateral vibration models, respectively . A recent study for the vibration detection scope was done by using the gradient boosting and random forest techniques for classifying the torsional vibration mode as a prediction objective by feeding the model with the drilling parameters of weight on bit, torque, and rate of penetration while drilling, while the model classification accuracy showed 62% for torsional vibration classification detection. The three modes of drillstring vibration were achieved through the ANN machine learning regression model with an accuracy of 0.95 for the correlation coefficient between the actual and predicted vibration measurements through training, testing, and validating the model …”