Machine Learning–Based Trigger Detection of Drilling Events Based on Drilling Data
- Jie Zhao (Schlumberger) | Yuelin Shen (Schlumberger) | Wei Chen (Schlumberger) | Zhengxin Zhang (Schlumberger) | Sonny Johnston (Schlumberger)
- Document ID
- Society of Petroleum Engineers
- SPE Eastern Regional Meeting, 4-6 October , Lexington, Kentucky, USA
- Publication Date
- Document Type
- Conference Paper
- 2017. Society of Petroleum Engineers
- 7.6.6 Artificial Intelligence, 1.10 Drilling Equipment, 1.6 Drilling Operations, 7 Management and Information, 7.6.4 Data Mining, 7.6 Information Management and Systems, 1.12.6 Drilling Data Management and Standards, 1.12 Drilling Measurement, Data Acquisition and Automation, 1.10 Drilling Equipment
- Pattern Recognition, Machine Learning, Trigger detection, Stick slip, Drilling events
- 71 in the last 30 days
- 139 since 2007
- Show more detail
- View rights & permissions
|SPE Member Price:||USD 8.50|
|SPE Non-Member Price:||USD 25.00|
A method is developed to detect the precursors of drilling events based on drilling data such as surface data, wellbore geometry data, lithology (formation characteristics), and downhole measurements from various downhole tools. The drilling events refer to interesting behavior of the drilling system detected or recorded, such as severe vibration, stuck pipe, fluid loss, sudden equivalent circulating density (ECD) changes, etc.
The method is based on various machine learning techniques to learn the changing trend of drilling parameters when the drilling events happen. Specifically, the drilling events are first extracted from massive drilling data using defined thresholds and/or criteria. Then, the time series of drilling parameters are represented by symbolic aggregate approximation (SAX). The patterns of these SAX strings are clustered by unsupervised learning and then used for pattern recognition with dynamic time warping (DTW). Finally, the searching pattern recognition is proposed to classify the changing trend of drilling parameters.
The traditional SAX method is not suitable for drilling data processing because it assumes that the data should be Gaussian distributed. With the modified SAX, two sets of SAX parameters are used to cluster the time series by unsupervised learning with dynamic time warping distance as measure. It is found that one set of SAX parameters (alphabet size of 5 and 15-data-point window) could yield more reasonable patterns, including flat, ramp up, ramp down, step up, step down, pulse up, and pulse down. The changing trends of drilling parameters are classified to the predefined patterns with shortest DTW distance by using searching pattern recognition method. Finally, several experiments are conducted to demonstrate the effectiveness of proposed method.
This is an innovative work to apply the data mining and machine learning techniques to drilling interpretation. It provides a useful way for the remote center to monitor the onset of abnormal drilling events and inform the drillers taking actions to optimize the drilling operation. In addition, it helps the drilling community to understand the triggers of challenging drilling events such as stick/slip, whirling, etc.
|File Size||1 MB||Number of Pages||11|