Abstract
Approximately 20% of all oilwells in the world use a beam pump to raise crude oil to the surface. The proper maintenance of these pumps is thus an important issue in oilfield operations. We wish to know, preferably before the failure, what is wrong with the pump. Maintenance issues on the downhole part of a beam pump can be reliably diagnosed from a plot of the displacement and load on the traveling valve; a diagram known as a dynamometer card. We demonstrate in this paper that this analysis can be fully automated using machine learning techniques that teach themselves to recognize various classes of damage in advance of the failure. We use a dataset of of 35292 sample cards drawn from 299 beam pumps in the Bahrain oilfield. We can detect 11 different damage classes from each other and from the normal class with an accuracy of 99.9%. This high accuracy makes it possible to automatically diagnose beam pumps in real-time and for the maintenance crew to focus on fixing pumps instead of monitoring them, which increases overall oil yield and decreases environmental impact.
