A Nonlinear Finite Element-Based Supervised Machine Learning Approach for Efficiently Predicting Collapse Resistance of Wireline Tool Housings Subjected to Combined Loads

Abstract

Abstract Wireline tools are designed in such a way that they can be lowered into a wellbore on the end of the wireline cable. They can be used for helping evaluate the reservoir and rock properties and formation pressures, locate casing collars, identify the properties of liquid present in the reservoir, and so on. Wireline tools can be subjected to harsh conditions such as those found in many modern oil, gas, and geothermal wells. For instance, pressures in gas wells can exceed 15,000 psi, while temperatures can reach 400 deg Fahrenheit in some geothermal wells. The tools can also experience bending loads when measurement is performed in the curved section of a well, and tension when the tool is pulled out. Housings are critical tubular-like structures that are intentionally designed to withstand such combined loads acting on the tool and protect the measurement components in the tool from damage during operation. Collapse of the housings has been one of the major failure modes and concerns for structural integrity of the wireline tools. Moreover, with continuous operation over time, these housings can get worn out, which could result in reduced collapse resistance. Collapse failures, particularly for those nuclear tools with radiation sources and nuclear detectors, can cause not only loss of the expensive assets, but also harm to the environment. Therefore, there is a strong demand to develop a fast-loop model that can be used for efficiently predicting the collapse resistance of the housings under combined service loads. Placement of the tool in a pressure vessel for physical testing could take several weeks’ effort, and is only affordable on a few limited sample configurations and pressure-only conditions. Conventionally, pressure rating of the housing was evaluated with empirical equations and an imposed safety factor, and its accuracy was limited to a ballpark estimation. When a housing needs to be qualified for high pressure applications, a high fidelity FEA (finite element analysis) model needs to be used. However, nonlinear FEA of a pressure housing could take hours due to its high-fidelity nature, which prevents it from being used as a computing engine for real-time applications. It is demonstrated in this paper that such challenges can be overcome with a newly developed computational framework by combining state-of-the art machine learning and nonlinear FEA for efficient and accurate collapse pressure predictions. First, a nonlinear FEA model that captures dimensions of both new and worn housings, elasto-plastic material properties, and environmental loads such as pressure, bending, tension, and temperature was developed, parametrized, and automated for parametric studies. Second, a carefully selected synthesized dataset containing simulation-based solutions, which had been pre-produced with automated FEA of the housings, was used to train a fast-loop predictive model. Third, feature engineering was performed to discover the compressed feature space representation of the data. Finally, modern supervised machine learning (ML) algorithms with hyperparameter tuning techniques were then evaluated to select a best performing model that can map the underlying pattern between the features and the collapse pressure of the housings. The resulting collapse pressure predictions for the housings were found to agree favorably well with the holdout dataset, which contains the experimental measurements from pressure vessel testing. This newly developed surrogate model has been deployed to a cloud-based platform for enabling prompt model-based job planning and asset utilization optimization without consuming expensive commercial FEA software licenses. The implemented model takes 1–2 minutes to calculate the pressure rating of a housing with satisfactory calculation accuracy.