Advanced Data Recognition Technique for Real-Time Sand Monitoring Systems

Abstract

Sand production in oil and gas wells is a serious issue for the petroleum industry around the world. The commonly used non-intrusive sand monitoring systems are based on - acoustic emission measurement techniques. This research presents advanced data recognition techniques that can significantly improve the accuracy of sand monitoring. At the first step, factor analysis was used to identify key acoustic features of sand particles. Then, the following machine learning techniques have been applied: support vector machines, logistic regression, random forest method and gradient boosting. For training and testing the recognition system we used the acoustic database obtained in the laboratory of the oilfield service company SONOGRAM LLC (Kazan, Russia). The database consisted of acoustics signals from sand particles impacting on the inside and outside of a pipe wall in various scenarios (dry and wet gas, different flow rates, etc.). It was shown that the use of support vector machines with the Gaussian kernel reduces false positives compared with the algorithm that is based on ultrasound power peaks detection.