document preview

Abstract

Artificial Neural Networks (ANN) are computational algorithms well suited to identify non-evident regularities and correlations in data. ANN are especially useful when there is missing or noisy process data, and when a mathematical model of the process is not readily available, as is the case with some Artificial Lift methods. Some ANN models exhibit excellent dimensionality reduction and visualization capabilities, mapping non-linear statistical relationships between high-dimensional data into simple geometric relationships, preserving the most important topological relationships of the data set. Thus, useful insight on the behavior of the process under study can be gained and used in its analysis and diagnosis. All this makes ANN a very useful tool in the exploratory phase of data mining. Results of several case studies applying ANN in pattern recognition of some Artificial Lift systemas are presented.

Introduction

Artificial Neural Networks are parallel information processing systems, formed by simple processing units, often called neurons, arranged and interconnected in diverse topologies, according to the function they are intended to perform. The units operate only on their local data and on the inputs they receive from their connections. The robustness of ANN, due to their massive parallelism approach, and their capability to learn, adjusting elements of their topology based on their input data, makes them adequate to solve problems in noisy or highly uncertain environments, and when there is not an explicit process model. ANN can be divided in supervised (trained by example) and unsupervised (learning without knowing the correct answer).

We have been using unsupervised ANN capabilities for data mining, in order to identify and adequately classify operation and failure patterns of Artificial Lift methods such as Progressive Cavity Pumping (PCP), which has no such information available to date, due to its particular characteristics. We are also working with unsupervised ANN to enhance and refine the existing knowledge on methods like Electric Submersible Pumping (ESP) and Sucker Rod Pumping (SRP). Results of classification for ESP and PCP are presented in this work.

Pattern Recognition and Data Mining

Pattern Recognition

This term encompasses a wide range of information processing problems, which humans can solve in a nearly effortless manner, but are very difficult when computational means are used, such as voice and handwriting character recognition. Bishop¹ divided pattern recognition problems in two categories: classification, or assignation of input values between a discrete number of classes, and regression, where output values represent continous variables. In this work, we will focus on classification of Artificial Lift operating conditions.

Data Mining^2,3,4

Is the automated analysis of large databases, searching for relationships, patterns and non-evidents trends, which otherwise would remain unnoticed. The process of knowledge discovery via data mining can be divided into four basic activities: selection (creating the target data set), pre-processing (preparing the data set for analysis), data mining (submitting the cleaned data to analysis by the data mining software) and interpretation of the results.