Automated Drilling and Production Event Detection Using Advanced Time-Series Pattern Recognition Techniques

Abstract The identification of the well/rig state in time is a key component in the construction of accurate risk-assessment, event-detection, and efficiency-tracking tools in drilling and production operations. Traditionally, this state identification has relied on insufficient rule-based systems, which often results in inaccurate predictions and leads to non-reliable risk-assessment tools, imprecise event detection, and biased estimated efficiency. This paper compares three state-of-the-art, scalable methods for automatically identifying the well/rig state and presents three use cases in drilling and production stages. Identifying the well/rig state is a time-series multi-class classification problem, in which the data is collected at high frequency (typically 0.02–1 Hz) with sensors installed inside the well, on the wellhead, or on the equipment intervening the well, such as a rig or a coiled-tubing unit. This paper presents three solutions to this classification problem, namely a first-order logic inference system, a recurrent neural network (RNN) classifier, and a transformer-based classifier. We implement and compare these methods in three applications in drilling and production operations, including the detection of stuck pipe incidents and pressure trend abnormality. The models were evaluated on a withheld, pre-labeled test dataset consisting of 75 hours of 1-Hz drilling data and 1072 h of 0.17-Hz production data. This evaluation showed that the transformer-based classifier outperformed the other three methods in all three applications. Additionally, we observed that the deep learning-based classifiers were only slightly more computationally expensive than inference systems, making all models suitable for real-time prediction. In the test, we also investigated the value of accurate well/rig state identification in the assessment of drilling and well-integrity risks. The test revealed that the lack of accuracy in this identification task can severely bias the risk assessment, for instance, by overestimating the risk of pipe sticking and generating spurious predictions of abnormal pressure trends. This, in turn, can result in unnecessary, costly preventative measures. The applicability of the analyzed methods is not limited to the examples here provided; they are applicable to any task requiring the classification of time-series data and thus, can be employed in other well operations/stages and applications of event detection. The novelty of this paper is twofold. First, it is the first study to compare various methods for automatically identifying well/rig state, including one not evaluated before—a transformer model. Second, it is the first to propose an automated engine for well state identification during the production stage. This paper also analyzes the advantages and shortcomings of well/rig state identification models through three different applications.