Recognition of Working Condition in Sucker-Rod Pumping Wells Based on Multi-View Kernel Learning and Hessian-Regularized Logistic Regression

Abstract

There are some research problems now in the existing working condition recognition methods of suckerrod pumping wells, such as false alarms easily caused by single information source, poor recognition effect and robustness arisen by traditional multiple information sources of feature connection, inferior engineering practicality induced by lots of labeled training samples. In order to overcome above problems and further improve the accuracy and practicality, in this paper, we propose a novel method for identifying the working condition of sucker-rod pumping wells based on multi-view kernel learning and Hessian-regularized logistic regression. The measured ground dynamometer cards, electrical power and wellhead temperature signal are combined together to recognize the working condition based on a model with the integration algorithm of multi-view kernel learning and Hessian-regularized logistic regression. The proposed method is applied to eleven kinds of working condition recognition of a certain block in Shengli oil field, and compared with traditional methods based on measured ground dynamometer cards, electrical power signal and multisource of feature connection respectively. Results show that the recognition rates increase by 2.4%, 11% and 13.8% respectively. The performance is even much better for cases of all marked and fewer marked training samples, especially in the case of 30% marked.