Drilling with AI-Based Systems: To Trust, or Not to Trust Model Forecast with Low Quality Real-Time Data?

Abstract Data-driven systems typically require input data of the same quality as that used to train their underlying algorithms in order to function effectively. However, drilling support data often falls short of these standards. The performance of AI-based systems designed to make drilling more efficient is directly affected by the consistency, frequency, completeness, and accuracy (e.g., presence of anomalies) of the incoming real-time drilling data stream. The paper discusses data quality control approaches for real-time drilling data, data pre-processing steps, and the influence of these approaches on the final AI-based model decisions. The impact of low quality real-time data on the decision-making process is considered using the example of an information system designed for predicting drilling accidents on real wells using data from the WITSML protocol stream. The methodology describes different aspects of consistency analysis of drilling parameters among each other and data quality calculations. Additionally, the impact of missing data fragments and missing channels on the prediction accuracy is analyzed. The quality of the algorithm operation as part of the information system is evaluated as the percentage of correct predictions and the number of false alarms per well per day. Lack of connectivity, data loss and anomalies are almost ubiquitous in the drilling support process. Single data losses of less than 10 minutes do not significantly change the accident prediction curve if they do not contain precursors of emergency situations. At the same time, small losses in individual channels can be recovered considering normal drilling conditions. Data quality is a critical factor for any AI algorithm. Automatically detecting and handling corrupted or inconsistent data enhances the decision-making process of AI-based systems.