Diagnostics of Oil Well Pumping Equipment by Using Machine Learning

Abstract

If speaking of timely detection of deviations in operation of pumping equipment, there is a problem of the current coverage of the oil well stock with telemetry sensors. Some data analytics, for example, analysis of dynamograms, is still performed manually. The present work attempts to create an automation solution for diagnostics of the condition of well pumping equipment. For sucker-rod pumps, a dynamogram classification model based on a convolutional neural network has been developed, which makes it possible to identify working conditions of a pumping unit. For electric centrifugal pumps (ECPs), a virtual sensor model has been developed based on modern machine learning technologies, which enables prediction of temperature and pressure gradients at the pump intake in the absence of submersible sensors. In the work, we tested a set of classical machine learning algorithms based on linear models and ensembles of decision trees, as well as advanced deep learning methods, e.g., transformers. The virtual sensor models developed are embedded directly into the automated process control system (APCS), and thus technologists and operators can be warned timely, almost in real time, of a possible shortening of the planned time between failures of ECP units and their possible mailfunctioning for various reasons.