Abstract
Abstract
Each year, coiled tubing (CT) well intervention fleets produce terabytes of multimodal data; these are recorded from surface and downhole sensors on each job. Among these data are job type(s) and technologies used on each job; traditionally, a field crew manually supplies this information. Given the diversity of data, many acquisition labels are often missing or inaccurate. A multimodal framework is presented that automatically identifies the job type and technologies used during an acquisition.
The proposed framework leverages different types of data depending on the job and technology to identify. Most job types (e.g., CT milling, CT fishing), and general technologies (e.g., downhole telemetry, jetting nozzles), are identified through a natural language processing (NLP) algorithm applied to operational reports. The presence of downhole technologies with greater granularity and certainty (e.g., downhole 2 1/8-in. pressure and temperature sonde) lies in the detection of meaningful information or noise on specific channels and follows a logic mimicking human interpretation. Finally, CT cementing and electronic firing heads are identified through statistical metrics and pattern recognition.
The framework leverages several tested methods. Primary job types and general technologies are classified using NLP. Of 366 acquisitions in the cloud archival system, the algorithm labels 97% with job type and 26% with one or more technologies. The second method extends the resolution and number of detected technologies to cover 12 unique real-time telemetry modules such as pressure, temperature, casing collar locator, and gamma ray sondes. For this method, over 50 acquisitions are analyzed with an accuracy of 94%. Electronic firing head signatures for three unique types of firing heads are identified successfully 97% of the time on 44 different acquisitions. This identification is sped up by automatically identifying three major stages of a CT well intervention (i.e., initial run in hole, service delivery, and final pull out of hole) and restricting the search space to the relevant stage. CT cementing job type identification is tested on 52 different jobs, with an accuracy of 92% and less than 30% of false negatives.
The innovative framework automatically classifies jobs and technologies. The inherent methods combine domain knowledge with the power of machine learning to enable efficient mining of data that would otherwise remain out of reach. By automating labeling, human error is largely eliminated and the reliability of the contextual metadata is significantly improved to provide crucial insights regarding operations.