http://pubs.acs.

org/journal/acsodf Article

Predictive Approach to Perform Fault Detection in Electrical

Submersible Pump Systems
Long Peng, Guoqing Han,* Xianfu Sui, Arnold Landjobo Pagou, Liying Zhu, and Jin Shu

Cite This: ACS Omega 2021, 6, 8104−8111 Read Online

ACCESS Metrics & More Article Recommendations

ABSTRACT: It has been a great challenge for the oil and gas
industry to timely identify any electrical submersible pump (ESP)
abnormal performance to avoid ESP failure. Given the high cost of
the ESP failure, more and more real-time surveillance systems are
applied to monitor ESP performance to generate alarms in the case
of failures. This paper presents a robust principal component
analysis (PCA) model to perform fault detection for ESP systems
continuously. A three-dimensional plot of scores of principal
components was used to observe different patterns during the
stable and failure periods. 47 cases of actual failure events and 40
cases of stable operating events were tested on the robust PCA
model to generate prediction results. The testing results
demonstrate that the robust PCA model has managed to identify
20 failure events before the actual failure time out of the 47 failure cases and has successfully distinguished all the 40 stable operating
wells. This study has concluded that PCA has the potential to be used as a monitoring platform to recognize dynamic change and
therefore to predict the developing failures in the ESP system.

1. INTRODUCTION real time. The oil and gas industry has been employing
The electrical submersible pumps (ESPs) are one of the most statistical modeling and machine learning techniques to detect
widely employed artificial lift pumping techniques, accounting patterns that allow for production optimization, failure
for over 60% of the total global oil production.1 More than prevention, and real-time event detection.6 Guo et al.7 used
90% of the offshore oil wells require ESPs for pumping oil and the support vector method to predict anomalous operation.
other fluids to the surface. However, it is often observed that Andrade Marin et al.8 analyzed random forest to obtain a high
the ESP system reaches the point of production disruptions value of accuracy and recall of ESP failure prediction in 165
that will result in an economic loss of hundreds of millions of cases. Bhardwaj et al.9 applied the principal component
oil barrels. analysis (PCA) and gradient boosting algorithm (XGBoost)
ESP failures are common occurrences in the oil industry, for anomaly detection and failure prediction in six cases and
and the ESP maintenance is an incredibly expensive issue for two samples. Developments in predicting the potential ESP
the oil operating company. A failure happens when key failures with machine learning technology, if successful and
parameters start deviating from safe operating thresholds and implemented, will improve the quality of forecasting and take
an ESP stops working forever.2 Ammeter charts have been the timely measures.
earliest and simplest ESP diagnostic tool in the past few years The ESP was previously monitored by traditional means,
and are still very common but require substantial human such as collecting ESP-related parameters, visiting the physical
resources for frequent wellsite visits to make necessary well site or evaluating the manual amperage charts (Amps),
adjustments.3 This monitoring method measures the motor hindering the ESP from reaching optimum performance.10 In
current and records it as a function of time on a circular chart.4 an effort to better understand ESP behavior, numerous
Ammeter charts are limited to detect compete ESP failures
because the change in power consumption is only linked to
voltage fluctuations, fluid density, and flow rate. In recent Received: November 30, 2020
years, nodal analysis has been applied to perform ESP system Accepted: February 22, 2021
modeling and determine some common ESP failures such as Published: March 16, 2021
broken shaft, blockage at pump intake, and so forth.5 However,
nodal analysis is also limited in the ability to analyze
performance and identify optimal operating conditions in
© 2021 The Authors. Published by
American Chemical Society https://dx.doi.org/10.1021/acsomega.0c05808
8104 ACS Omega 2021, 6, 8104−8111
ACS Omega http://pubs.acs.org/journal/acsodf Article

Figure 1. PCA model detection.

improvement has been made on ESP sensors, supervisory X = TPT + E (1)

control, and data acquisition systems and surface remote
where X is the input matrix (n × p); T is the score matrix (n ×
terminal units over the past few decades.11 With the fast
k); P is the loading matrix (p × k); E is the residual matrix (n ×
development of the digital oil field, more and more machine
p); n is the number of time steps;15 p is the number of original
learning and data-driven models are applied to perform fault parameters;15 and k is the number of principal components.15
detection and predict impending failure for specific ESP The first principal component contains the highest variance
operation systems.12 The full realization of a data-driven model which means that the first principal component captures the
lies beyond the availability of continuous data acquisition. most information. The second principal component captures
Therefore, the key to perform fault detection on the ESP can the next highest variance. Each successive principal component
be better defined as a problem to build an accurate data-driven is orthogonal to its preceding principal components. Several
model that describes the ESP system dynamics. principal components are constructed to create new PCs to
This presented work is concerned with the application of the evaluate the original ESP system.
PCA in three dimensions to provide an intuitive way of Gupta et al.17 proposed that the PCA model contained two
monitoring the ESP system. The least squares method is used steps: model construction and model prediction. The ESP data
to determine the safety threshold of the ESP stable operation obtained from the stable operating zone (Xtraining) is used to
system. With this method, a predictive model is constructed to build the PCA model. The loading matrix Ptraining for the stable
perform fault detection in real time. The objective of this paper operating zone will be stored for prediction at a later stage.
is to evaluate this predictive approach to detect developing Once a robust PCA model is constructed, any new input
ESP failure. dataset can be selected as a testing dataset (Xtesting). The testing
dataset may be a stable operation zone or an unstable failure
2. APPLICATION OF PCA operation zone. The testing dataset is fed to the robust PCA
model as a new score matrix (Ttesting) is obtained. Patterns
PCA is widely used as a preprocessing method for data corresponding to stable periods or failure periods will be
visualization and dimensionality reduction.13 Jackson14 defined compared against patterns trained by the PCA model to draw
PCA as an unsupervised dimensionality reduction technique conclusions for making predictions.18 By this means, any
that converts a set of dependent variables into a set of linearly historical data in the ESP system can be detected as a stable
uncorrelated variables called “principal component”. As known operating zone or a failure zone. Figure 1 summarizes the
to all, ESP data are highly correlated. For example, increasing comments above.
the wellhead pressure will cause an increase in discharge
pressure and intake pressure, eventually increasing the motor 3. CONSTRUCTION OF A ROBUST PCA MODEL
pressure. A PCA model takes advantage of this interdepend- 3.1. Selection of ESP Variable Parameters. Much
ence to reduce the numbers of ESP original data and to create research has been done about the application of the PCA
a new principal component space (PCs). This reduction allows model in the XX oilfield XX block in China. The fast
the evaluation of the original ESP system only by several development of ESP sensors and data acquisition systems
principal components. Abdelaziz et al.11 built different PCA makes it possible for the ESP system to record dynamic data
models for each installation to conclude that PCA had the and historical data continuously. Dynamic production data
potential to be used as a tool to identify dynamic changes in were collected at a frequency of 20 min. Dynamic data
the ESP system. Gupta et al.15 proposed the PCA predictive recorded contain variables of casing pressure, intake pressure,
model, diagnostic model, and prescriptive model to detect discharge pressure, flowtubing pressure, intake temperature,
impending ESP failures. motor current, motor leakage current, motor power, motor
The PCA basic model was represented as16 temperature, motor vibration, motor voltage, tubing choke,
8105 https://dx.doi.org/10.1021/acsomega.0c05808
ACS Omega 2021, 6, 8104−8111
ACS Omega http://pubs.acs.org/journal/acsodf Article

Figure 2. Original ESP system.

Figure 3. New PCs.

and VFD frequency. Thirteen variable parameters were production time of 47 wells were used to construct a robust
selected as training datasets or testing datasets. Historical PCA model. Three principal components captured more than
data recorded include historical well events and well failure 70% information of the original training dataset to create a new
reports. The information when a trip or failure occurred in PCs. Taking well #X1 as an example, Figure 2 showed the
each ESP well was recorded in the historical data. original stable operating ESP data recorded containing 13
3.2. Comparison of the Original System and PCs. A variable parameters shown. Three principal components were
data-driven model using the PCA method was constructed to extracted to form the new PCs to evaluate the original system
reduce the dimensionality of the historical stable training shown in Figure 3. It could be seen from the comparison of
dataset containing 13 key variable parameters. 174,538 Figures 2 and 3 that the trends of the original 13 variable
historical stable operating data according to different parameters were retained in the form of three principal
8106 https://dx.doi.org/10.1021/acsomega.0c05808
ACS Omega 2021, 6, 8104−8111
ACS Omega http://pubs.acs.org/journal/acsodf Article

components according to the robust PCA model. Each or failure, the cluster started deviating away from the training
principal component was determined by calculating the dataset and was finally located outside the sphere.
singular value decomposition of the selected datasets. More-
over, a special eigenvalue and eigenvector corresponding to a 4. MODEL TESTING AND RESULTS
unique principal component would be determined. Each 4.1. Testing of Potential Failure Wells. The historical
principal component explained the different percentage of stable operating data of 47 wells were used to construct the
the total variance, and the top-ranked principal components robust PCA model, and the unstable operating period of these
had larger eigenvalues. Only three principal components were wells served as a training dataset.
used to evaluate the original ESP system, making the ESP According to the ESP failure nomenclature standard and
system fault detection much easier. Three principal compo- failure definitions by Alhanati et al.,19 this paper represented
nents on a three-dimensional plot provided an intuitive way of the different types of ESP failures from the collected datasets
monitoring the ESP system. shown in Figure 5. There were mainly 6 different types of
3.3. Scores of Three Principal Components. A three- failures, and the causes for each type of failure were also
dimensional plot of scores of principal component 1, principal different.
component 2, and principal component 3 was used to observe
different patterns during stable and failure periods. The three
principal components were extracted from the original training
dataset. Figure 4 represented the scores plot of three principal

Figure 5. Different types of ESP failures in the XX oilfield.

When the ESP wells were put into production, the ESP wells
would start from the stable region to the failure region as the
time step increased. The unstable operating region contained
stable operating data before a trip/failure and unstable
behavior ultimately leading to a trip/failure. 47 potential
failure wells were tested by the robust PCA model. For each
well of the testing dataset, the robust PCA model generated
one prediction result. Taking four wells as examples, Figure 6
was an intuitive visual of ESP well monitoring for a historical
failure event. It was observed that the stable operating zone of
the testing dataset formed clusters inside the training sphere.
Figure 4. Scores of three principal components. In the meanwhile, a failure zone started deviating away from
the sphere while a failure or trip took place.
4.2. Testing of Stable Operating Wells. The robust
components; it was viewed that the historical stable operating PCA model was tested against a total of 40 stable operating
zone forms a large cluster. The X-axis represented the first wells. These 40 wells were also from the XX oilfield XX block
principal component, the Y-axis represented the second in China. However, the difference was that these 40 wells were
principal component, and the Z-axis represented the third the daily report data rather than real-time data. Similarly, these
principal component. The new PCs also contained 174,537 40 wells were tested by the robust PCA model to generate the
stable operating data according to different production times in prediction results. Taking 4 wells as examples, Figure 7
three principal components rather than 13 original variable represented three-dimensional plots of scores for ESP fault
parameters. The three-dimensional sphere was built up to detection on the stable operating wells. The analysis of Figure
enclose all the stable operating data. The center of the sphere 7 showed that the stable operating daily report data formed
of these 174,537 data was determined based on the least clusters inside the sphere, which meant that these four wells
square method, and the radius was the maximum distance from did not have any abnormal behavior during the operating
all stable operating data to the center of the sphere. There was periods.
only one sphere for all 47 wells. The radius of the sphere was 4.3. Testing Results and Discussion. A confusion matrix
the failure threshold of all the stable operating data of 47 wells. was used to summarize the performance of the robust PCA
By this means, when the scores for the testing dataset model, showing what the prediction results were getting right
corresponded to a stable operation period, the cluster formed and what types of errors the prediction results were making.
inside the sphere, similar to the training dataset. However, There were four possible binary classification outcomes in a
when the scores for the testing dataset corresponded to a trip confusion matrix.8
8107 https://dx.doi.org/10.1021/acsomega.0c05808
ACS Omega 2021, 6, 8104−8111
ACS Omega http://pubs.acs.org/journal/acsodf Article

Figure 6. ESP fault detection on potential failure wells.

The robust PCA model was tested on 47 cases of actual that the cause for each type of failure was different. For
failure events and 40 cases of stable operating events to instance, there were no obvious abnormal behaviors before the
generate prediction results. Figure 8 summarized the matrix pump broken shaft, so the robust PCA model had failed to
results in a confusion matrix. detect the failure events before the actual failure data but had
It could be drawn from Figure 8 that out of the 47 ESP identified when these failure events exactly happened. In the
failure events, the robust PCA model managed to detect 20 meanwhile, the robust PCA model had managed to detect the
failure wells before the actual failure date but failed to identify motor failures before the actual failure date because the motor
the remaining 27 impending failure events. For the stable current, motor voltage, or motor temperature performed
operating well testing results, the robust PCA model had abnormally before the actual failure time captured by the
successfully predicted all the 40 nonfailure cases. It was robust PCA model. Secondly, the sampling of 20 min was too
observed that the robust PCA model in three-dimension long and much dynamic changes of different variable
presented 42.5% accuracy for predicting developing ESP failure parameters leading to a failure had been missed. Lastly, the
events before the actual failure date. The appearance of these robust PCA model in three principal components only
prediction results came for some reasons. The main reason was extracted 70% information of the original ESP operation
8108 https://dx.doi.org/10.1021/acsomega.0c05808
ACS Omega 2021, 6, 8104−8111
ACS Omega http://pubs.acs.org/journal/acsodf Article

Figure 7. ESP fault detection on stable operating wells.

system, and 30% of the variance was discharged, resulting in minimize downtime if the clusters are far away from the
part of fault behavior missed. Future work must be focused on sphere.
these ways to improve the prediction results in detecting
developing failure events.
5. RESULTS AND DISCUSSION
Overall, the PCA method serves as a real-time platform to This paper develops a robust PCA model to perform fault
automatically perform fault detection of any ESP system. This detection for the ESP systems. 47 cases of actual ESP failure
events and 40 cases of stable operating events are tested by the
detection solution can be of significant importance in PCA model. It can be concluded that the PCA technique has
predicting the impending ESP failure in real-time. The three- the potential to monitor dynamic change and predict the
dimensional plot of scores of three principal components impending failure in the ESP system. The following
creates a visual medium to continuously monitor the ESP conclusions has been developed from this work:
operation for keeping users apprised of what is happening with (1) A three-dimensional plot of scores of principal
the monitored system. Moreover, oilfield engineers will be components is used to observe different patterns of
reminded to take corrective action or timely solve the issues to stable and failure period in real time. The dataset
8109 https://dx.doi.org/10.1021/acsomega.0c05808
ACS Omega 2021, 6, 8104−8111
ACS Omega http://pubs.acs.org/journal/acsodf Article

Notes
The authors declare no competing financial interest.

■ ACKNOWLEDGMENTS
The authors of this article would like to express their gratitude
to the National Natural Science Foundation of China (grant
no. 51574256) and Open Foundation of Shaanxi Key
Laboratory of Carbon Dioxide Sequestration and Enhanced
Oil Recovery (grant no.YJSYZX20SKF0011). The authors are
grateful to the editor and the reviewers for their critical
comments, which were helpful in the preparation of this article.

Figure 8. Confusion matrix of well failure prediction results.

■ REFERENCES
(1) Dunham, C. 27th ESP Workshop: Summary of Presentations;
Oilfield Automation Consulting, 2013; p 1. Retrieved from https://
corresponding to a stable operation period forms the www.spegcs.org/media/files/files/cebfcc3a/2013-ESP-Workshop-
clusters inside the sphere and the dataset corresponding Summary-of-Presentations.pdf.
to a trip or failure starts deviating away from the sphere. (2) Alamu, O. A.; Pandya, D. A.; Warner, O.; Debacker, I. ESP Data
(2) Out of the 47 ESP failure events, the PCA model has Analytics: Use of Deep Autoencoders for Intelligent Surveillance of
Electric Submersible Pumps. Offshore Technology Conference, Houston,
managed to identify 20 impending failure events and
Texas, USA, May 2020.
mispredict 27 failure before the actual failure time. (3) Bates, R.; Cosad, C.; Fielder, L.; Kosmala, A.; Hudson, S.;
(3) The PCA model has successfully detected all the 40 Romero, G.; Shanmugam, V. Taking the pulse of production wells
stable operating wells out of the 40 nonfailure cases. ESP surveillance. Oilfield Rev. 2004, 16, 16−25.
(4) The PCA technique can serve as a real-time platform to (4) Gabor, T. Electrical Submersible Pumps Manual: Design,
Operations, and Maintenance; Elsevier: Burlington-Massachusetts,
perform fault detection for the ESP system and as an 2009.
unsupervised machine learning technique to predict the (5) Awaid, A.; Al-Muqbali, H.; Al-Bimani, A.; Al-Yazeedi, Z.; Al-
impending ESP failures. Sukaity, H.; Al-Harthy, K.; Baillie, A. ESP Well Surveillance using

■
Pattern Recognition Analysis, Oil Wells, Petroleum Development
AUTHOR INFORMATION Oman. International Petroleum Technology Conference, Doha, Qatar,
January 2014.
Corresponding Author (6) Stone, P. Introducing predictive analytics: Opportunities. Digital
Guoqing Han − State Key Laboratory of Petroleum Resources Energy Conference and Exhibition, Houston, Texas, U.S.A., April 2007.
and Prospecting, China University of Petroleum (Beijing), (7) Guo, D.; Raghavendra, C. S.; Yao, K.-T.; Harding, M.; Anvar, A.;
Beijing 102249, China; Shaanxi Key Laboratory of Carbon Patel, A. Data driven approach to failure prediction for electrical
Dioxide Sequestration and Enhanced Oil Recovery, Xi’an submersible pump systems. SPE Western Regional Meeting, Garden
710075, China; Email: CUP_706ac@163.com Grove, California, USA, April 2015.
(8) Andrade Marin, A.; Busaidy, S.; Murad, M.; Balushi, I.; Riyami,
Authors A.; Jahwari, S.; Ghadani, A.; Ferdiansyah, E.; Shukaili, G.; Amri, F.
Long Peng − State Key Laboratory of Petroleum Resources ESP Well and Component Failure Prediction in Advance using
and Prospecting, China University of Petroleum (Beijing), Engineered Analytics-A Breakthrough in Minimizing Unscheduled
Subsurface Deferments. Abu Dhabi International Petroleum Exhibition
Beijing 102249, China; Shaanxi Key Laboratory of Carbon
& Conference, Abu Dhabi, UAE, November 2019.
Dioxide Sequestration and Enhanced Oil Recovery, Xi’an (9) Bhardwaj, A. S.; Saraf, R.; Nair, G. G.; Vallabhaneni, S. Real-
710075, China; orcid.org/0000-0001-9654-953X Time Monitoring and Predictive Failure Identification for Electrical
Xianfu Sui − China National Offshore Oil Corporation Submersible Pumps. Abu Dhabi International Petroleum Exhibition &
(CNOOC) Research Institute, Beijing 100027, China Conference, Abu Dhabi, UAE, November 2019.
Arnold Landjobo Pagou − State Key Laboratory of Petroleum (10) Shuwaikhat, H. A.; Ramos, M.; Aifan, A.-R.; Al-Sadah, A. A.
Resources and Prospecting, China University of Petroleum Innovative Approach to Prolong ESP Run Life Using Algorithmic
(Beijing), Beijing 102249, China; Shaanxi Key Laboratory of Models. Abu Dhabi International Petroleum Exhibition & Conference,
Carbon Dioxide Sequestration and Enhanced Oil Recovery, Abu Dhabi, UAE, November 2017.
Xi’an 710075, China (11) Abdelaziz, M.; Lastra, R.; Xiao, J. ESP data analytics: Predicting
failures for improved production performance. Abu Dhabi Interna-
Liying Zhu − State Key Laboratory of Petroleum Resources
tional Petroleum Exhibition & Conference, Abu Dhabi, UAE, November
and Prospecting, China University of Petroleum (Beijing), 2017.
Beijing 102249, China; Shaanxi Key Laboratory of Carbon (12) Jahnke, P. Machine Learning Approaches for Failure Type
Dioxide Sequestration and Enhanced Oil Recovery, Xi’an Detection and Predictive Maintenance; Technische Universität Darm-
710075, China stadt, 2015; p 19.
Jin Shu − State Key Laboratory of Petroleum Resources and (13) Raschka, S., Predictive Modeling, Supervised Machine Learning,
Prospecting, China University of Petroleum (Beijing), Beijing and Pattern ClassificationThe Big Picture; Sebastian Raschka, 2014.
102249, China; Shaanxi Key Laboratory of Carbon Dioxide (14) Jackson, J. E. A User’s Guide to Principal Components; John
Sequestration and Enhanced Oil Recovery, Xi’an 710075, Wiley & Sons, 2005; Vol. 587.
China (15) Gupta, S.; Saputelli, L.; Nikolaou, M. Applying big data
analytics to detect, diagnose, and prevent impending failures in
Complete contact information is available at: electric submersible pumps. SPE Annual Technical Conference and
https://pubs.acs.org/10.1021/acsomega.0c05808 Exhibition, Dubai, UAE, September 2016.

8110 https://dx.doi.org/10.1021/acsomega.0c05808
ACS Omega 2021, 6, 8104−8111
ACS Omega http://pubs.acs.org/journal/acsodf Article

(16) Ratcliff, D.; Gomez, C.; Cetkovic, I.; Madogwe, O. Maximizing

Oil Production and Increasing ESP Run Life in a Brownfield Using
Real-Time ESP Monitoring and Optimization Software: Rockies Field
Case Study. SPE Annual Technical Conference and Exhibition, New
Orleans, Louisiana, USA, September 2013.
(17) Gupta, S.; Nikolaou, M.; Saputelli, L.; Bravo, C. ESP Health
Monitoring KPI: A Real-Time Predictive Analytics Application. SPE
Intelligent Energy International Conference and Exhibition, Aberdeen,
Scotland, UK, September 2016.
(18) Bimani, A.; Kulkarni, R.; Lee, C. Y.; Giuliani, V.; Musallami,
M.; Awaid, A.; Kamyani, N.; Kumar, N.; Gala, R. Case Study Toward
Digital Oil Field: How the ESP Operation is Changing by Using
Automatic Well Models in PDO’s ESP Fields. SPE Gulf Coast Section
Electric Submersible Pumps Symposium; The Woodlands: Texas, USA,
May 2019.
(19) Alhanati, F.; Solanki, S.; Zahacy, T. ESP failures: can we talk
the same language? SPE Gulf Coast Section Electric Submersible Pump
Workshop, Houston, Texas, April 2001.

8111 https://dx.doi.org/10.1021/acsomega.0c05808
ACS Omega 2021, 6, 8104−8111