System and method for failure prediction for rod pump artificial lift systems

• US 8,988,236 B2
• Filed: 05/27/2011
• Issued: 03/24/2015
• Est. Priority Date: 05/27/2010
• Status: Active Grant

• Current Assignee: University of Southern California
• Original Assignee: University of Southern California
• Inventors: Liu, Yintao, Olabinjo, Lanre, Yao, Ke-Thia, Liu, Shuping, Raghavendra, Cauligi Srinivasa, Seren, Fatma Burcu, Seddighrad, Sanaz, Gunasekaran, Dinesh Babu Chinnapparaja
• Primary Examiner(s): Phan, Hai
• Assistant Examiner(s): Wu, Zhen Y
• Application Number:
  US13/118,067
• Publication Number:
  US 20120025997A1
• Time in Patent Office: 1,397 Days
• Field of Search: 340/679, 166/244.1, 166/245, 166/254.2
• US Class Current: 340/679
• CPC Class Codes: E21B 47/009 Monitoring of walking-beam pump systems

Abstract

A computer-implemented reservoir prediction system, method, and software are provided for failure prediction for artificial lift well systems, such as sucker rod pump systems. The method includes providing well data from a production well. Attributes are extracted from the well data. Data mining is applied to the attributes to determine whether the production well is predicted to fail within a given time period. An alert is output indicative of impending production well failures.

Claims

1. A computer-implemented method for failure prediction for artificial lift well systems, the method comprising:

obtaining well data representing a plurality of measured attributes from a plurality of wells over a time period;

generating features from the well data based on the measured attributes;

applying a random peek semi-supervised machine learning technique to the features derived from the plurality of wells to generate a failure prediction model, the applying of the random peek semi-supervised machine learning technique including;

(a) splitting the well data into at least two data clusters such that a first cluster of well data is larger than a second cluster of well data, wherein the first cluster of well data is assumed to correspond to wells that do not contain one or more well failures and the second cluster of well data is assumed to correspond to wells that contain one or more well failures, and (b) adding well data from the first cluster to a training data set used to generate the failure prediction model;

determining whether one or more of the plurality of wells is predicted to fail within a subsequent time period based on the failure prediction model; and

outputting an alert indicative of impending failure of one or more of the plurality of wells.

2. The method of claim 1, wherein the each of the plurality of wells utilizes a sucker rod pump.

3. The method of claim 1, wherein data preparation techniques are applied to the well data prior to generating the features.

4. The method of claim 1, wherein generating the features further comprises generating a multivariate time series.

5. The method of claim 1, wherein generating the features further comprises applying a sliding window approach to extract multiple multivariate subsequences.

6. The method of claim 1, wherein the applying a machine learning technique to the features comprises applying an alternating decision tree method.

7. The method of claim 1, wherein the applying a machine learning technique to the features comprises applying a Bayesian Network method.

8. The method of claim 1, wherein the applying a machine learning technique to the features comprises applying a support vector machine method.

9. The method of claim 1, further comprising:

adjusting operating parameters of one or more of the plurality of wells to forestall a failure responsive to the alert.

10. The method of claim 1, further comprising:

scheduling maintenance for one or more of the plurality of wells responsive to the alert.

11. The method of claim 1, further comprising evaluating failure prediction accuracy using K-fold cross-validation method or a test-set method.

12. The method of claim 1, performing a cross-correlation by measuring similarity between two multivariates time sequences of the plurality of measured attributes to determine a dependency between the measured attributes across time.

13. The method of claim 1, further comprising performing a trend analysis and capturing trend information including determining a dependency between the measured attributes across time and determining a dependency between a current value of a measured attribute with a prior value of the measured attribute.

14. The method of claim 13, wherein determining the dependency between the measured attributes across time includes measuring a similarity of two measured attributes as a function of time-lag applied to one of the attributes.

15. The method of claim 13, wherein determining the dependency between the current value of the measured attribute with the prior value of the measured attribute includes applying an auto-correlation method.

16. A system for failure prediction for artificial lift well systems, the system comprising:

a database configured to store well data representing a plurality of measured attributes from a plurality of wells over a time period;

a computer processor configured to receive the stored data from the database, and to execute software responsive to the stored data; and

a software program executable on the computer processor to implement a method, the method comprising;

extracting features from the well data from the plurality of wells based on the attributes; and

generating a failure prediction model using a random peek semi-supervised machine learning technique, the random peek semi-supervised machine learning technique including;

(a) splitting the well data into at least two data clusters such that a first cluster of well data is larger than a second cluster of well data, wherein the first cluster of well data is assumed to correspond to wells that do not contain one or more well failures and the second cluster of well data is assumed to correspond to wells that contain one or more well failures, and (b) adding well data from the first cluster to a training data set used to generate the failure prediction model; and

determining whether one or more of the plurality of wells is predicted to fail within a subsequent time period based on the extracted features.

17. The system of claim 16, wherein the system further comprises a display that communicates with the computer processor executing the software program such that an alert indicative of impending production well failures is produced on the display.

18. The system of claim 16, wherein the software program is further executable on the computer processor to extract attributes from the well data in the form of a time series.

19. The system of claim 16, wherein the software program is further executable on the computer processor to reduce noise in the well data.

20. The system of claim 16, wherein the software program is further executable on the computer processor to generate a multivariate time series.

21. The system of claim 16, wherein the software program is further executable on the computer processor to apply a sliding window approach to extract multiple multivariate subsequences.

22. A computer program product, comprising a non-transitory computer usable medium having a computer readable program code embodied therein, the computer readable program code configured to be executed on a computer system to implement a method for failure prediction for artificial lift well systems, the method comprising:

extracting features from well data representing a plurality of measured attributes from a plurality of wells over a time period; and

applying a random peek semi-supervised machine learning technique to the features to create a failure prediction model based on the features from the plurality of wells, the random peek semi-supervised machine learning technique including;

(a) splitting the well data into at least two data clusters such that a first cluster of well data is larger than a second cluster of well data, wherein the first cluster of well data is assumed to correspond to wells that do not contain one or more well failures and the second cluster of well data is assumed to correspond to wells that contain one or more well failures, and (b) adding well data from the first cluster to a training data set used to generate the failure prediction model; and

determining whether one or more of the plurality of wells is predicted to fail within a given time period based on the failure prediction model.