Automated classification of drill string vibrations using machine learning algorithms

Abstract

Downhole vibration can cause drill string failures and significant economic losses. Recognizing harmful vibrations in real time is highly significant for optimizing drilling operations. This work aims to identify whirls and stick-slip using machine learning (ML) from downhole measured accelerations. A rotor dynamics model is utilized to simulate whirl and a coupling method between whirl and stick-slip is presented. 448 sets of simulated acceleration data are used for the training of ML classification models. 14 features for whirl identification and 8 for stick-slip are defined in each acceleration. The generalization abilities of the trained whirl and stick-slip classifiers are validated using ABAQUS simulation data and field-measured data, respectively. Decision Tree, Naive Bayes, Support Vector Machine (SVM), and Neural Network algorithms are tested. The classification accuracy is highest when whirl is divided into synchronous forward whirl, synchronous backward whirl, and discontinuous contact whirl. Both SVM and Neural Network algorithms can achieve nearly 100% accuracy in rotor dynamics data and ABAQUS data. The highest accuracy is attained by utilizing both tangential and radial acceleration. However, using only tangential acceleration can still fulfill engineering requirements. The stick-slip classifier, when using SVM and Neural Network, can achieve 100% recognition accuracy on all test data. The proposed methods can lessen sensor need for downhole measurements and speed up data processing, and may be applied to other problems involving rotating soft shafts.