Machine learning-based cement integrity evaluation with a through-tubing logging experimental setup

Abstract

Assessing the integrity of the cement layer and the quality of its bond to the casing and formation is paramount to ensure that the wellbore is hydraulically isolated from the surrounding environment before permanently sealing the well. Such inspection is part of Plugging & Abandonment (P&A) operations, and it is usually achieved through well logging tools, which provide a vast amount of data that a skilled specialist interprets. The process is human-dependent, error-prone, and time-consuming. There has been an increasing interest in solutions that allow an automatic interpretation of the logging data since the recent panorama of the oil and gas industry indicates a growing P&A demand. Such solutions aim at reducing the dependence on human knowledge and consequently increasing the reliability and accuracy of the cement integrity evaluation. Therefore, this work presents an experimental setup capable of emulating defective cement layer configurations of real-world oil wells in single or multi-string arrangements. Such flexibility enables acquiring logging data for multiple well conditions and building a rich logging database to develop a supervised learning framework and define the most suitable model for performing the cement integrity evaluation. Several logging runs were performed, producing a database with 130 samples, including varied tubing eccentricity levels and cement layer conditions. A complete analysis of the data both in the time domain as well as in the frequency–wavenumber domain was performed, highlighting the complexity of the interpretation task. A resampling-based workflow was employed to evaluate machine learning models of different families. The models were tested under three scenarios, and accuracy and computational complexity metrics were computed to compare their performance. The results showed that shallow learning models can perform satisfactorily well even with less data available for training. The support vector machine stood out, achieving a mean accuracy score higher than 0.99 while being able to predict the cement sheath’s condition in less than 1 ms. This paper contributes to the research on the cement integrity evaluation by presenting a study that combines an experimental setup mimicking several oil well conditions and the employment of machine learning as a diagnostic tool, which has no precedents in recent literature regarding the acoustic logging knowledge field.