No abstract

Ultrasonic and sonic logs are increasingly used to evaluate the quality of cement placement in the annulus behind the pipe and its potential to perform as a barrier. Wireline logs are carried out in widely varying conditions and attempt to evaluate a variety of cement formulations in the annulus. The annulus geometry is complex due to pipe standoff and often affects the behavior (properties) of the cement. The transformation of ultrasonic data to meaningful cement evaluation is also a complex task and requires expertise to ensure the processing is correctly carried out as well interpreted correctly. Cement formulations can vary from heavy weight cement to ultralight foamed cements. The ultrasonic log-based evaluation, using legacy practices, works well for cements that are well behaved and well bonded to casing. In such cases, a lightweight cement and heavyweight cement, when bonded, can be easily discriminated from gas or liquid (mud) through simple quantitative thresholds resulting in a Solid(S) - Liquid(L) - Gas(G) map. However, ultralight and foamed cements may overlap with mud in quantitative terms. Cements may debond from casing with a gap (that is either wet or dry), resulting in a very complex log response that may not be amenable to simple threshold-based discrimination of S-L-G. Cement sheath evaluation and the inference of the cement sheath to serve as a barrier is complex. It is therefore imperative that adequate processes mitigate errors in processing and interpretation and bring in reliability and consistency. Processing inconsistencies are caused when we are unable to correctly characterize the borehole properties either due to suboptimal measurements or assumptions of the borehole environment. Experts can and do recognize inconsistencies in processing and can advise appropriate resolution to ensure correct processing. The same decision-making criteria that experts follow can be implemented through autonomous workflows. The ability for software to autocorrect is not only possible but significantly enables the reliability of the product for wellsite decisions. In complex situations of debonded cements and ultralight cements, we may need to approach the interpretation from a data behavior-based approach, which can be explained by physics and modeling or through observations in the field by experts. This leads a novel seven-class annulus characterization [5S-L-G] which we expect will bring improved clarity on the annulus behavior. We explain the rationale for such an approach by providing a catalog of log response for the seven classes. In addition, we introduce the ability to carry out such analysis autonomously though machine learning. Such machine learning algorithms are best carried out after ensuring the data is correctly processed. We demonstrate the capability through a few field examples. The ability to emulate an "expert" through software can lead to an ability to autonomously correct processing inconsistencies prior to an autonomous interpretation, thereby significantly enhancing the reliability and consistency of cement evaluation, ruling out issues related to subjectivity, training, and competency.

No abstract

In recent years, the use of reinforcement learning and imitation learning to complete robot control tasks have become more popular. Demonstration and learning by experts have always been the goal of researchers. However, the lack of action data has been a significant limitation to learning by human demonstration. We propose an architecture based on a new 3D keypoint tracking model and generative adversarial imitation learning to learn from expert demonstrations. We used 3D keypoint tracking to make up for the lack of action data in simple images and then used image-to-image conversion to convert human hand demonstrations into robot images, which enabled subsequent generative adversarial imitation learning to learn smoothly. The estimation time of the 3D keypoint tracking model and the calculation time of the subsequent optimization algorithm was 30 ms. The coordinate errors of the model projected to the real 3D key point under correct detection were all within 1.8 cm. The tracking of key points did not require any sensors on the body; the operator did not need vision-related knowledge to correct the accuracy of the camera. By merely setting up a generic depth camera to track the mapping changes of key points after behavior clone training, the robot could learn human tasks by watching, including picking and placing an object and pouring water. We used pybullet to build an experimental environment to confirm our concept of the simplest behavioral cloning imitation to attest the success of the learning. The effectiveness of the proposed method was accomplished by a satisfactory performance requiring a sample efficiency of 20 sets for pick and place and 30 sets for pouring water.