Risk Assessment of Drilling Tool Failure—A Data Driven Technique Assisted by Expert Knowledge

Abstract

Abstract Risk assessment of drilling tool failure prior to field deployment plays a critical role in well construction. Drilling tool failure can be very expensive, due to rig downtime, missed reservoir targets and pre and post job costs. In recent years, data-driven techniques, such as machine learning (ML) methodologies, have gained popularity in drilling operations, including predicting drilling tool failure. This is due to the abundance of operational drilling data, the advancement of ML algorithm capabilities, and computational infrastructure. However, ML models built upon pure data knowledge have several pitfalls: lack of generalization and explainability, data imbalance, and difficulty in optimization. This paper proposes a data-driven technique assisted by expert knowledge during the model building process, comprised of descriptive and predictive analytic frameworks. The descriptive analytic framework considers multiple drilling parameters, such as temperature, vibrations, bit rotary speed, stick slip, torque, weight, drilling fluid properties, circulating hours, and drilling hours. Additionally, data labeling covers information on the failure, run parameters, and active state of the tool. This step evaluates all plausible contributing factors to the failure, which experts then assess, feeding back into the analysis of the set of stress factors to consider. In the following predictive analytics framework, an ensemble of predictive models estimates the parameter importance supporting the feedback from the experts. Partial dependence analysis and individual conditional expectation evaluates interaction between parameters and estimates the contribution of each parameter to the failure event. This technique helps with model explainability, which adds confidence in the results. Boosting technique generalizes the model, and a developed random under-sampling routine handles the imbalance of the data label (failure and survivor). Finally, the Bayesian optimization technique handles the difficulty found during the optimization process. An application demonstrates the novelty of the method proposed in this paper. The application is to evaluate the deployment failure risk of a hydraulic motor in a Measurement-While-Drilling directional drilling unit, and stator-alternator, in the communication-power module of a drill string. The paper also demonstrates an affirmation of the insights from data-driven technique with the laboratory results during failure investigation. Moreover, it is shown that infusing expert knowledge in the analysis framework can result in an improved model performance for failure risk assessment, which in turn contributes significantly to reducing the failure rate. ML models built upon pure data knowledge have several pitfalls, including lack of generalization and explainability, data imbalance, and difficulty in optimization. The method described here tackles these pitfalls by incorporating subject-matter-expert (SME) knowledge into analytics, machine learning context. This results in a powerful tool, illustrated here in evaluating the failure risk of a component of a downhole tool.