Machine Learning Model for Identifying Drilling Dysfunctions from Surface Data and Bit Failure Forensics

Summary Drilling dysfunctions are those that prevent the drill bit from achieving the expected rate of penetration (ROP) performance despite increasing input energy such as weight on bit (WOB). As drilling depth and well structure complexity continue to increase, there is an urgent need to best determine how to identify and mitigate downhole drilling dysfunctions in a timely and accurate way. Traditional methods for identifying drilling dysfunctions on rigs have mainly relied on simple threshold alarms of drilling parameters, but most drilling operations lack the expensive downhole sensors needed, making it difficult to monitor the drilling dysfunctions from real-time downhole data. To address these challenges, in this work we integrated surface data and physical mechanisms with machine learning (ML) techniques to intelligently identify common drilling dysfunctions. The performance of four classical ML algorithms was evaluated, revealing that their precision for certain types of dysfunctions was unsatisfactory. To overcome this issue, we integrated autoencoders with these ML algorithms, which resulted in a more than 90% precision of the developed intelligent recognition model, with an improvement of up to 13.79% over previous versions. In addition, we used drill-bit failure forensics based on dull-bit photo as the validation source for the developed model, which compensated for the lack of downhole sensor data. The field trial showed that for the well using the model-recommended bit and optimized drilling parameters, the ROP and footage increased by 154% and 69%, respectively, compared with the offset well, demonstrating the effectiveness and reliability of the developed intelligent recognition model in practical drilling applications. The developed model had the potential to reliably identify drilling dysfunctions without downhole sensor data, thereby addressing the high cost of using downhole sensors.