Smart Sucker Rod Pump Failure Analysis with Machine Learning

Publikationen: Thesis / Studienabschlussarbeiten und HabilitationsschriftenMasterarbeit

Standard

Smart Sucker Rod Pump Failure Analysis with Machine Learning. / Ben Smida, Ameni.
2019.

Publikationen: Thesis / Studienabschlussarbeiten und HabilitationsschriftenMasterarbeit

Harvard

Ben Smida, A 2019, 'Smart Sucker Rod Pump Failure Analysis with Machine Learning', Dipl.-Ing., Montanuniversität Leoben (000).

APA

Ben Smida, A. (2019). Smart Sucker Rod Pump Failure Analysis with Machine Learning. [Masterarbeit, Montanuniversität Leoben (000)].

Bibtex - Download

@mastersthesis{6ba22c59534c45d7be24bf44f9438df2,
title = "Smart Sucker Rod Pump Failure Analysis with Machine Learning",
abstract = "Sucker rod pumping is the most frequently used artificial lift method for boosting oil production. To insure a good operation of this system, a continuous monitoring of its working conditions is essential to maintain acceptable productivity levels. The most valuable tool for analyzing the rod pumping system performance is the dynamometer card. However, the interpretation of such cards is time consuming and requires the knowledge of an experienced person. A new trend came along and solved the problem of time and human expertise dependency. This trend is in instance artificial neural networks (ANN). In this work two types of ANN are used, the first one is the back propagation neural network (BPNN) which is considered traditional as it requires feature extraction from the data before using it for data learning and the second one is the convolutional neural network (CNN) which is able to use image data directly without processing it prior to training. Both networks use 6132 dynamometer cards which for BPNN requires processing as follows; each dynamometer card, which is represented by an image file stored in a PNG format, was prepared as a set of (x,y) values which are then converted into a set of Elliptic Fourier Descriptors which fully describe the whole card. After performing the data preparation, the two machine learning classification models were created, evaluated using precision and recall as well as confusion matrix and F1-score and tested by use of cross validation. The proposed models are trained and tested by using real field dynamometer cards data. About 30% of these cards represent normal sucker rod pumping condition and 70% represent malfunctions. The data contain in total five different pump states, pump off, gas interference, travelling valve leak, pump hitting on top and normal pump condition. For training the dataset was separated into different sub-sets, 80% of the data were used for training and 20% for testing. The CNN as well as the BPNN produced very good results. This study is an original contribution to the automatically investigations of the dynamometer cards and the accurate and quick recognition of the rod pumping systems failures.",
keywords = "Tiefpumpenausfall, Tiefpumpen, Fehleranalyse, maschinelles LernenK{\"u}nstliche Neuronale Netze, Dynamometerkarten, Convolutional Neural Network, Back Propagation Neural Network, K{\"u}nstliche Neuronale Netze, Bilddaten, sucker rod pump, machine learning, dynamometer cards, rod pump failure, artificial neural networks, back propagation neural network, convolutional neural network, image data",
author = "{Ben Smida}, Ameni",
note = "no embargo",
year = "2019",
language = "English",
school = "Montanuniversitaet Leoben (000)",

}

RIS (suitable for import to EndNote) - Download

TY - THES

T1 - Smart Sucker Rod Pump Failure Analysis with Machine Learning

AU - Ben Smida, Ameni

N1 - no embargo

PY - 2019

Y1 - 2019

N2 - Sucker rod pumping is the most frequently used artificial lift method for boosting oil production. To insure a good operation of this system, a continuous monitoring of its working conditions is essential to maintain acceptable productivity levels. The most valuable tool for analyzing the rod pumping system performance is the dynamometer card. However, the interpretation of such cards is time consuming and requires the knowledge of an experienced person. A new trend came along and solved the problem of time and human expertise dependency. This trend is in instance artificial neural networks (ANN). In this work two types of ANN are used, the first one is the back propagation neural network (BPNN) which is considered traditional as it requires feature extraction from the data before using it for data learning and the second one is the convolutional neural network (CNN) which is able to use image data directly without processing it prior to training. Both networks use 6132 dynamometer cards which for BPNN requires processing as follows; each dynamometer card, which is represented by an image file stored in a PNG format, was prepared as a set of (x,y) values which are then converted into a set of Elliptic Fourier Descriptors which fully describe the whole card. After performing the data preparation, the two machine learning classification models were created, evaluated using precision and recall as well as confusion matrix and F1-score and tested by use of cross validation. The proposed models are trained and tested by using real field dynamometer cards data. About 30% of these cards represent normal sucker rod pumping condition and 70% represent malfunctions. The data contain in total five different pump states, pump off, gas interference, travelling valve leak, pump hitting on top and normal pump condition. For training the dataset was separated into different sub-sets, 80% of the data were used for training and 20% for testing. The CNN as well as the BPNN produced very good results. This study is an original contribution to the automatically investigations of the dynamometer cards and the accurate and quick recognition of the rod pumping systems failures.

AB - Sucker rod pumping is the most frequently used artificial lift method for boosting oil production. To insure a good operation of this system, a continuous monitoring of its working conditions is essential to maintain acceptable productivity levels. The most valuable tool for analyzing the rod pumping system performance is the dynamometer card. However, the interpretation of such cards is time consuming and requires the knowledge of an experienced person. A new trend came along and solved the problem of time and human expertise dependency. This trend is in instance artificial neural networks (ANN). In this work two types of ANN are used, the first one is the back propagation neural network (BPNN) which is considered traditional as it requires feature extraction from the data before using it for data learning and the second one is the convolutional neural network (CNN) which is able to use image data directly without processing it prior to training. Both networks use 6132 dynamometer cards which for BPNN requires processing as follows; each dynamometer card, which is represented by an image file stored in a PNG format, was prepared as a set of (x,y) values which are then converted into a set of Elliptic Fourier Descriptors which fully describe the whole card. After performing the data preparation, the two machine learning classification models were created, evaluated using precision and recall as well as confusion matrix and F1-score and tested by use of cross validation. The proposed models are trained and tested by using real field dynamometer cards data. About 30% of these cards represent normal sucker rod pumping condition and 70% represent malfunctions. The data contain in total five different pump states, pump off, gas interference, travelling valve leak, pump hitting on top and normal pump condition. For training the dataset was separated into different sub-sets, 80% of the data were used for training and 20% for testing. The CNN as well as the BPNN produced very good results. This study is an original contribution to the automatically investigations of the dynamometer cards and the accurate and quick recognition of the rod pumping systems failures.

KW - Tiefpumpenausfall

KW - Tiefpumpen

KW - Fehleranalyse

KW - maschinelles LernenKünstliche Neuronale Netze

KW - Dynamometerkarten

KW - Convolutional Neural Network

KW - Back Propagation Neural Network

KW - Künstliche Neuronale Netze

KW - Bilddaten

KW - sucker rod pump

KW - machine learning

KW - dynamometer cards

KW - rod pump failure

KW - artificial neural networks

KW - back propagation neural network

KW - convolutional neural network

KW - image data

M3 - Master's Thesis

ER -