Diagnosis of Operating Conditions and Sensor Failures in Wells Operating by Mechanical Pumping Using Machime Learning
Sucker-rod pumping, Machine Learning and Dynamometric cards
In oil fields with many wells operating by Sucker-rod pumping, due to the lack of early diagnosis of operating conditions or sensor faults, several problems can go unnoticed. These problems can generate large losses, such as increased downtime, increased OPEX (Operational Expenditure), decreased efficiency and lost production. In practice, the identification and diagnosis of operating conditions are carried out from surface and downhole dynamometric cards, using pre-established standards, with human visual effort in the operation centers. This task requires a lot of time and work, in addition to requiring experience, as it can be influenced by subjective factors. However, in recent years, with the facilities inherent to Machine Learning (ML) algorithms, several researches on the subject have achieved good results in the diagnosis of operating conditions, showing that ML can be used for this purpose. However, it is still common to have doubts about the difficulty level of the dynamometer card classification task, the best algorithm, the best shape descriptor, the best metrics and what is the impact of the imbalanced datasets. In the search for answers to these questions, this work used real data from 38 wells operating by sucker-rod pumping in the region of Mossoró, RN, Brazil. More than 50,000 cards have been classified by specialists and distributed in eight modes of operation and two common sensor faults in this field. Sixty tests were performed and divided into seven groups. Three algorithms were tested (with and without hyperparameter tuning): Decision Tree, Random Forest and XgBoost, in addition to pipelines provided by Automated Machine Learning (AutoML). The descriptors used were: Fourier descriptors and Wavelet descriptors, in addition to the load values of the downhole dynamometric card. Balanced and imbalanced datasets were also tested. The results confirm the feasibility of applying ML for diagnosis of operating conditions and sensor faults in sucker-rod pumping systems, since 75% of the tests reached accuracy above 92% and a maximum accuracy was 99.84%.