The impact of distinct rolling window sizes in automated cement evaluation through well-logging and machine learning methods

Cementing operations in oil and gas (O&G) wells are critical procedures to guarantee zonal isolation, mitigate fluid migration, and maintain wellbore integrity. Traditionally, assessing cement sheath quality involves analyzing well logs, which requires specialized technical expertise. Considering recent developments in machine learning (ML) for automatically evaluating cement quality in offshore wells, this research examines the influence of different rolling window sizes on model performance. This work provides the first systematic evaluation of the impact of logging context interval size on machine learning prediction of cement bond quality and hydraulic isolation in offshore wells. The methodology consisted of developing multiple classification algorithms (random forest (RF), k-nearest neighbors (KNN), support vector machine (SVM), logistic regression (LR), and Naive Bayes (NB)) to automatically evaluate cement bond quality (BQ) and hydraulic isolation (HI), thus analyzing the effects associated with the application of distinct context intervals (CI), ranging from 1 to 13 meters. Feature engineering techniques were employed to extract relevant characteristics from raw data, minimizing biases associated with logging complexity and tool diversity. Classifier performances were evaluated using confusion matrices supported by four quantitative metrics, providing an overall score (OS) for each algorithm within each CI. Optimal performances were observed at smaller CI (3 to 8 meters). The RF model achieved the highest OS, with 75.5% for the BQ scenario (5- and 6-meter CI) and 85.3% for the HI scenario (5-meter CI).