Predicting Subsea Wellhead Fatigue Damage with a Machine Learning Based Tool

Abstract Subsea wells are susceptible to wellhead fatigue damage during operation. Accurate and timely prediction of this fatigue damage is crucial for informed rig decisions, ensuring operational safety while minimizing nonproductive time (NPT). However, conventional methods for assessing wellhead fatigue overestimate damage, leading to conservative, time-consuming, and costly design, and operational decisions. To address these limitations, a web-based tool hosting machine learning (ML) models was recently developed to assess the wellhead fatigue damage in a less conservative, real-time manner and comes at no incremental cost to end users. The machine learning models based on existing industry algorithms and the one specifically developed to address high fatigue damage scenarios and data gaps are presented. Validation against field measurement results from Gulf of Mexico (GOM) demonstrates a high success rate for the tool. The ML models utilize measured fatigue damage rate as the target variable and incorporate various environmental data as features, including sea current profiles, wave, and sea water temperature. Due to the high cost, only a small number of wells in the GOM have been equipped with wellhead fatigue monitoring systems during offshore operations, resulting in a scarcity of measurement data. In addition, most of the collected data corresponds to low fatigue damage conditions. To address this challenge and enhance the generality of the machine learning models, an extrapolation workflow has been established. The decision of whether to use interpolation or extrapolation models was made using a Gaussian Mixture Model, followed by employing an XGBoost regressor for interpolation and a linear regression model for extrapolation. This workflow aims to enhance model performance when the input data significantly deviates from the training data. Incorporating extrapolation techniques allows for more robust predictions even in cases where the input data deviates significantly from the previously observed patterns. This groundbreaking tool, the first of its kind in industry, achieved a 100% success rate in validating wells previously monitored for fatigue damage in the GOM. Validation involves comparing the predicted fatigue damage from the tool with the measured fatigue damage. Currently, the tool is being piloted for fatigue risk assessment in subsea well operations in the GOM and has demonstrated superior performance. It may replace conventional wellhead fatigue assessment methods, including physics-based simulation and analysis, as well as fatigue monitoring using measurement equipment like sensors. The novelty of the tool and the developed ML models lies in their ability to assess wellhead fatigue in a real-time manner, incurring no additional cost. The predicted fatigue damage is less conservative, resulting in cost savings for system design and rig operations.