Abstract

Abstract Over the lifetime of multiple wells, in different fields, data produced from integrity assessment of the casing and mechanical parts of oil and gas wells accumulates to huge amounts and diversity. Knowledge derived from these test records can help in integrity assessment of other wells and explain contributing factors. This paper presents application of a unique Deep-Learning algorithm, to automate well-integrity assessment by extracting entire knowledge from an existing database of millions of integrity-tests, using a complex Deep Neural Network (DNN), and transfer this knowledge into another simple DNN model to provide an explainable integrity assessment and contributing factors for end user. Herein, we present a two-phase algorithm-development process. It uses values of annular pressure, maximum allowable pressure, production annulus shut-in pressure, surface wellhead emission rate, corrosion etc. from 105301 oil and gas wells. Firstly, a complex Deep Neural Network (DNN) highly regularized with drop-outs and equivalent to summation of exponential number of models, extracts knowledge-representation, i.e., mapping between quantification of mechanical properties, their evolution and factors contributing to these properties, from well-integrity tests records of all the fields. In the second phase, the knowledge-representation learned by the complex DNN is passed on to the simple DNN. It extracts well-specific information from the knowledge-representation, with its two objective functions and provides an explainable integrity assessment for end users to make better decisions. First DNN with thousands of trainable parameters is cumbersome, unexplainable and has very slow execution speed. Second DNN with only a few hundred parameters outputs one-hot encoded target vector of values for Sustained Casing Pressure (SCP), Casing Vent Flow (CVF) and corrosion, to quantify integrity of a well. These vectors and soft probability from knowledge representation of the first DNN, combines in the first objective, to ensure transfer of entire knowledge from the first DNN to the second DNN. Second objective function performs optimization between calculated probability of SCP, CVF and corrosion, and corresponding truth values in a very small training set. Second DNN fails to perform if it does not use knowledge-transfer from the first DNN. With the second objective, the second DNN achieves an accuracy of 93%. Development-database consists of records of well-integrity assessments performed in Raton, San-Juan, Denver-Julesburg, Appalachian, Permian and Piceance basin of Colorado, New Mexico and Pennsylvania, between 1991 and December, 2020. Proposed experiments were performed on Nvidia RTX 2060 SUPER 2x8GB GPU using deep learning framework. Novelty of this paper lies in demonstration of one of the initial applications of knowledge-distillation, a deep-learning algorithm, to automate well-integrity assessment. It is a unique method of transferring knowledge-representation learned from a huge database by a complex DNN, to a simpler DNN, for explainable and fast assessment.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call