Publications

In Sucker Rod Pump (SRP) systems, analysis of dynamometer-card (dynacard) delivers valuable insight into the status of the pump and is used to indicate required action for the pump if issues are detected. All SRP wells in LK Field have installed Rod Pump Controller (RPC) where dynacard and other SRP data are sent automatically to the user’s platform using Global System Mobile Communication (GSM). Currently, engineers categorize the dynacards manually, taking up to 10 hours to categorize 20 cards whereas the system reads one dynacard per minute, summing up to 1440 dynacards for one well in one day. This inefficient method for categorization results in longer downtime and higher oil production deferment.

This study explores machine-learning techniques to improve surveillance of beam pumps from dynacard data by automatically categorizing the cards resulting in increased efficiency and reduced maintenance activities resulting from missed early diagnosis.

About 3,000 dynacards were manually labelled by experts into nine (9) classes that are categories usually occurring/found in the field of study/of interest. The dataset is split into 80% training, 10% testing datasets, and 10% validation datasets. Different machine learning algorithms were evaluated, and it is found that the top performing model, Convolutional Neural Network (CNN), achieves 99.54% accuracy.

After implementation of RPC and machine-learning, the categorization of SRP cards is automatic and early diagnosis is performed, decreasing the average downtime duration for wells from 4.8 hours per month per well to 1.7 hours per month per well (decreased 64%). Additionally, average loss potential oil was decreased from 57 barrels per month to 29 barrels per month (decreased 50%) or saving cost up to $25,000 per year.

This study found that the CNN model can aid engineers by automating dynacard diagnostics and categorization, which can reduce downtime and reduce loss potential oil.