Enhancing Well Kick Classification in Drilling Operations Using a Novel PCA-Based Machine Learning Approach

Abstract

Listen

Translate article

ABSTRACT: Early detection of kicks is imperative for ensuring efficient and safe drilling operations. Employing a data-driven approach for binary classification proves to be an effective method for differentiating between pre-kick and normal operational data. This research delves into the application of machine learning for well kick classification, critically addressing the challenge of model generalization. It proposes innovative solutions, including a novel PCA-based method, to overcome these hurdles. This study thoroughly examines the potential and obstacles in training models to achieve enhanced generalization capabilities. Notably, the Artificial Neural Network (ANN) model, with its remarkable nonlinear fitting abilities, demonstrates superior performance. This not only affirms the effectiveness of machine learning in detecting kicks but also signifies a substantial advancement in resolving issues related to data availability and model generalization in training processes. Consequently, this research represents a significant stride towards the practical and reliable application of machine learning in drilling operations, marking a crucial development in the field. 1. INTRODUCTION Well kick is a phenomenon where unexpected formation gas or fluid flows into the wellbore while drilling. The direct cause of a kick is that the hydraulic pressure provided by the drilling fluid cannot balance the formation pressure at the bottom of the hole. This imbalance may result from abnormal formation pressure, defective drilling fluid, or loss of circulation (Liu et al. 2023). A kick can compromise drilling efficiency by causing Non-productive Time (NPT) and, in severe cases, may lead to a blowout, resulting in substantial economic losses. In addition, the establishment and migration of kicks is a self-accelerating process where the lighter influx will further reduce hydraulic pressure at the bottom of the hole and drive more formation influx. Therefore, detecting, confirming, and treating kicks as early as possible is crucial for ensuring efficient and safe drilling operations. Traditional kick detection methods rely on monitoring data from one or more sensors, such as changes in pit volume, standpipe pressure or flow rate differences (Tarr et al. 2016). These methods offer significant improvements over relying solely on human experience. However, due to the transient nature of drilling and sensor disturbances, they often suffer from high rates of false alarms or slow detection times. Without adding new sensors and thus increasing drilling costs, traditional approaches encounter bottlenecks in further enhancing detection speed and improving kick detection accuracy. Data-driven methods that utilize the supervised machine learning paradigm have been proven to be powerful tools for solving many drilling engineering problems, both in terms of regression and classification (Kamyab et al. 2010, Singh et al. 2020, Olukoga et al. 2021, Wang et al. 2023, Zhang et al. 2024, Jing et al. 2024, Wang et al. 2024, Zheng et al. 2024). Adding binary labels to differentiate between pre-kick and normal operational states in real-time drilling data at each timestep transforms the task of kick detection into a supervised machine learning binary classification problem (Yang et al. 2019, Muojeke et al. 2020, Yin et al. 2020, Wang et al. 2022). At each moment, the multidimensional real-time drilling parameters serve as inputs to the machine learning model, while the target variable to be learned and predicted is the binary class (kick or normal), which represents the current tendency for a kick to occur. Multiple studies have attempted this approach, and the results are promising (Unrau et al. 2017, Alouhali et al. 2018, Obi et al. 2023). These research efforts share a common procedure: aggregating datasets from multiple wells, randomizing them, and then splitting them into training and testing sets. Models are trained on the training set and their accuracy is evaluated on the testing set. However, the generalization ability of models has seldom been studied or discussed. This refers to the accuracy of kick classification predictions for data originating from a new well (one that was not included in the training set) after the model has been trained. Subsequent sections of this paper presented a test specifically designed to analyze this issue. The examples demonstrate that the generalization capabilities of such models are indeed limited. This suggests that the development of a robust kick classification model necessitates training on multidimensional, real-time drilling data sourced from a significant number of wells, typically at the scale of an entire field.