Machine Learning Approaches to Anomaly Detection of Top Drive Torque Causing Drill Pipe Failure

Abstract

The scientific drillship Chikyu, operated by JAMSTEC has faced the problems with drill pipe damage and failure. After a failure, we examined the surface-measured drilling data, including Fourier transform analysis and spectrogram, and conducted an outlier investigation with an autoregressive model. We found some possible indications of top drive torque anomalies among the surface-measured drilling data. The purpose of this study is to attempt anomaly detection of top drive torque by applying machine learning approaches to such pre-examinations. Data acquired during the drilling operations of the Chikyu are first analyzed to observe the operations and determine the characteristics of the drilling data. Several hundred sets of surface drilling data are extracted and prepared as labeled learning data. Machine learning is then performed using several algorithms to build predictive models for anomaly detection of top drive torque. Simulations using these predictive models are then conducted for other sets of drilling data, including the anomaly conditions acquired during the operations of the Chikyu. The results demonstrate the validity of the predictive models. This study applies machine learning technique to the anomaly detection of the top drive torque with aim of preventing the drill pipe damage and failure. This paper describes our approaches in detail and discuss the results.