Recognizing Abnormal Shock Signatures During Drilling with Help of Machine Learning

Abstract

Drilling generated shocks and vibrations (torsional, axial, and lateral) are among the main causes of failures in the drilling industry; because they affect the rate of penetration, directional control, and wellbore quality. Rotary steerable system tools are equipped with measurement devices such as magnetometers, accelerometers, and shocks and vibration sensors from which statistical information is obtained, such as root-mean squared error, maximum peaks, and peak levels. From these statistics, whirl, bit bounce, and stick/slip severity are inferred. Often, the derived statistics are not enough to distinguish between normal drilling versus abnormal drilling for a location in the wellbore or to determine whether the shocks and vibrations are the result of poor drilling practice, formation disturbances, or mechanical failures of the bottomhole assembly, including the bit. Machine learning methods were used for analyzing the high-frequency radial shock burst data, which compresses and classifies the data; i.e., good drilling and abnormal drilling. The method is capable of further clustering the data into whirl or no whirl, bit-bounce or no bit-bounce, formation change or no change, and/or faulty equipment and parts; thus, assist in the robust post-failure analysis of existing data sets and prevent catastrophic failures in real time and improve the trajectory control.

Detecting downhole vibrations through drilling horizontal sections: machine learning study

“…Ignova et al 32 developed a classification model using K-Means cluster analysis algorithm and principal component analysis (PCA) pattern recognition technique. The model was trained by 2048 measurements data set for the classification of good or damaged drilling using high-frequency shock data for the model training process.…”

Real‐time prediction of horizontal drilling pressure based on convolutional Transformer

“…The training process drew from a dataset encompassing 1400 measurements that were harnessed for stick–slip classification. These datasets incorporated a comprehensive array of drilling‐related metrics, including drilling torque, tension, drilling speed, drilling pressure, and triaxial acceleration; Ignova et al 26 devised a classification model utilizing a combination of the K‐Means cluster analysis algorithm and the Principal Component Analysis (PCA) pattern recognition technique. This innovative approach effectively employed high‐frequency impact data to differentiate drill bits into categories of “good” or “damaged”; Wiktorski et al 27 devised a pair of models employing Support Vector Machines (SVM), one geared toward detecting stick–slip phenomena, and the other designed for identifying lateral vibration occurrences.…”

Machine Learning Solution for Predicting Vibrations while Drilling the Curve Section

“…A deep neural network was also implemented to develop a machine learning solution for detecting the torsional (stick–slip) vibration mode from the surface drilling data by using the classification technique for 1400 data points of drilling data and vibration measurements, and the classification accuracy reached 96% that was further enhanced by more data generalization to record 99% classification accuracy . K-Means cluster was also tested for the vibration classification algorithm in addition to the principal component analysis tool and training the model by 2048 measurements for the training data set that represents the shock data of high frequency . Another research applied the support vector machine (SVM) for building different separate prediction models for predicting the torsional and lateral vibration modes, while the models were trained by a few measurements as 260 data sets for the torsional vibration model and 865 data points for the lateral model.…”

“… 25 K-Means cluster was also tested for the vibration classification algorithm in addition to the principal component analysis tool and training the model by 2048 measurements for the training data set that represents the shock data of high frequency. 26 Another research applied the support vector machine (SVM) for building different separate prediction models for predicting the torsional and lateral vibration modes, while the models were trained by a few measurements as 260 data sets for the torsional vibration model and 865 data points for the lateral model. Meanwhile, the two models were trained by feeding drilling torque, drilling rotation speed, weight on bit (for the torsional vibration model), and hook load parameter (for the lateral vibration model).…”