Professional Documents
Culture Documents
Spe 214462 Ms Fallas
Spe 214462 Ms Fallas
This paper was prepared for presentation at the SPE Symposium: Leveraging Artificial Intelligence to Shape the Future of the Energy Industry held in Al Khobar, Saudi
Arabia, 17 - 18 January 2023. The official proceedings were published online on 19 January, 2023.
This paper was selected for presentation by an SPE program committee following review of information contained in an abstract submitted by the author(s). Contents
of the paper have not been reviewed by the Society of Petroleum Engineers and are subject to correction by the author(s). The material does not necessarily reflect
any position of the Society of Petroleum Engineers, its officers, or members. Electronic reproduction, distribution, or storage of any part of this paper without the written
consent of the Society of Petroleum Engineers is prohibited. Permission to reproduce in print is restricted to an abstract of not more than 300 words; illustrations may
not be copied. The abstract must contain conspicuous acknowledgment of SPE copyright.
Abstract
Despite being the most widely used artificial lift method for high-producing oil wells, ESPs still experience
unplanned failures that impact well productivity and overall field economics. Our advanced ESP Predictive
Failure Analytics (PFA) can help detect ESP events ahead of time and extend the overall ESP run life. PFA
enabled a major Latin American operator, experiencing frequent unplanned ESP failures, to identify critical
events while pumps were running and take remedial actions to extend ESPs run life.
Methods, Procedures, Process: PFA leverages artificial intelligence (AI), statistical and physics-based
models to reliably predict Remaining Useful Life (RUL) and possible failure cause. The models are trained
using historical sensor time-series from both running and failed ESPs. The trained models are deployed to
predict short-term events that may lead to immediate failure, such as a broken shaft, short-circuit, grounded
downhole failure; as well as long-term events which build up over time, such as pump low efficiency, sand,
scale deposition and gassy conditions.
Results, Observations, Conclusions: For this study, we used two ESPs. For ESP-1, PFA predicted broken
shaft/missed pump stages after a sudden decline in motor current and production rate. As the production
rate declined beyond the minimum recommended operating range, PFA identified downthrust condition and
estimated a significant RUL reduction. PFA enables the operator to quickly schedule a workover, reducing
downtime. For ESP-2, intake pressure and motor current started decreasing and motor temperature started
increasing. PFA predicted sand influx and estimated a significant RUL reduction. A chemical injection was
applied to reduce sand, and avoid an imminent failure leading to PFA estimating an increased RUL.
Novel/Additive Information: PFA is an innovative approach which combines AI, statistical and physics-
based methods to provide explainable predictions of ESP failure. Unlike commonly used threshold-based
approaches, PFA tends to generate fewer alarms which enables proactive optimization of ESP performance,
avoiding unplanned failures and extend ESP run life.
2 SPE-214462-MS
Introduction
Electrical Submersible Pumps (ESPs) are a widely used artificial lift method for high-producing oil wells.
To keep the production economical in such wells a long living ESP is required. Currently, despite being
very popular, ESPs experience many unplanned failures which impact the productivity and overall field
economics. A pump that fails prematurely leads to unplanned downtime which can result in deferred oil
production. If this happens frequently, it can have a significant financial impact on an operator, impacting
Figure 1— Run life distribution for 243 failed pumps from four assets in Latin America
The fact that most pumps do not make it to the target run life raises multiple questions including which
ESPs are more likely to fail? Will a certain ESP fail anytime soon? Was the failure preventable by using
preventive maintenance or optimizing operations? Can we be better prepared for a failure? Etc.
In this paper we will demonstrate how our advanced ESP Predictive Failure Analytics (PFA) can help
answer these questions by detecting ESP events ahead of time which enables the elongation of the overall
ESP run life.
Related Work
Monitoring and analyzing the performance of ESPs is of very high interest to both operators and ESP
manufacturers. Great progress has been made in recent years in the aim of developing monitoring systems,
but most operators still do not fully utilize a system that will assist in monitoring the health of an ESP.
Currently many operators and manufacturers use knowledge- and physics-based methods which generate
alarms when sensor trend changes are associated with potential ESP failures. These systems can detect
short- and long-term damage events i.e., motor overload/underload, imminent shutdown, pump wear tubing
leak etc. and a few also offer remedial recommendations for each possible damage event (Adesanwo et. al.
2016, Awaid et. al. 2014, Bermudez 2014, El Gindy 2015, Gagner et. al. 2010, Grassian et. al. 2017). The
main disadvantage of such systems is using fixed thresholds for trend detection and overlapping expert rules
for event detection, both of which tend to generate an overwhelming large number of alarms, in addition
to the fact that they are difficult to generalize.
SPE-214462-MS 3
Other approaches currently being used include an enhanced Cox Proportional Hazard Method (CPH)
which is used to predict ESP failures based on historical ESP pulls (Bailey et. al. 2018), as well as the use of
different signal processing tools such as Fourier- and Wavelet- transformation, on motor current to enable
predicting ESP failures which are caused by scale build up (Noui-Mehidi et. al. 2019).
Other attempts to predict the probability of an ESP failure utilize different Machine Learning (ML)
models. One approach used semi-supervised ML models to classify timeseries data from sensors which
originated in historical failed pumps. Three classes were defined, normal, pre-failure and failure which were
Data Description
ESP specification information was collected from ESP tracking spreadsheets, this information includes
well name, run number, installation date, start date pull/failure date etc. For the failed ESPs the Dismantle,
Inspection, and Failure Analysis (DIFA) reports were used to collect data regarding pull reason and failure
cause which was used to create ground truth labels.
Pump failures can be classified into different categories including mechanical failures and electrical
failures. Failures in these different classes are then further diagnosed to identify a specific failure cause.
In this case study we will focus on over 200 ESPs being operated in Latin America. In this region, failed
pumps are caused due to different issues where sand is the leading cause for pump failure. This specific
4 SPE-214462-MS
failure cause can be caused by either a mechanical or an electrical issue. See Figure 2 for a detailed failure
cause distribution. In this case study we will focus on mechanical and electrical failures only.
The failed pumps at this region can be described using the Weibull distribution which is widely used
in survival analysis as well as failure analysis (see Figure 3a). Using the best fitting Weibull distribution,
we can analyze the survival rate as well as the hazard rate and life expectancy. Looking at the survival
function for this collection of pumps, a pump has more than 80% chances of not failing if it survived 500
days (see Figure 3b).
Figure 3— Failed ESP analysis a. Weibull distribution function that fits failed ESP distribution best. b. Survival function
for the current pump collection. Probability of not failing (y axis) given that a pump has survived to this age (x axis).
The start date and pull/failure date from the ESP specification information was used to collect historical
data from the operator's historian / SCADA system for both running and failed pumps. This data included
historical timeseries data for surface and downhole sensors, which was collected at a 1-minute sampling
rate for all ESPs. For failed ESPs, the pulled data was between the first day the ESP was started and up
SPE-214462-MS 5
to the reported failure date. For running ESPs, the data was collected up to the current date. The ESP's
timeseries data was evaluated and curated prior to model training to remove any low-quality data. The
data preprocessing steps include dealing with missing data, as well as evaluating and scoring available data
including dealing with outliers and physically meaningless values. The effect of removing such outliers and
meaningless values is shown in Figure 4. Furthermore, advanced interpolation techniques were applied to
deal with missing data points. Data scores consider data gaps, sensor error as well as constant values. The
quality scores range between 0 (low quality) and 1 (high quality). Data quality is further used to weight
Figure 4— Intake pressure value distribution. a. Distribution before removing any values which are not within
the valid sensor range (red vertical lines) b. Distribution after removing values outside of valid sensor range.
Methodology
An overview of the methodology implemented in PFA to predict pump failure and provide remedial
recommendation has been previously described (Silvia et al., 2022). In this section, we will summarize the
general structure of PFA architecture, as well as introduce a subset of recently added enhancements.
An overview of the methodology implemented in PFA to predict pump failure and provide remedial
recommendation has been previously described (Silvia et al., 2022). In this section, we will summarize the
general structure of PFA architecture, as well as introduce a subset of recently added enhancements.
PFA is an ensemble of models used to train a final predictive life model. This ensemble includes four
main subcomponents described as follows:
PFA is an ensemble of models used to train a final predictive life model. This ensemble includes four
main subcomponents described as follows:
1. Weibull-based survival model trained on asset-level survival data and clustered based on relevant
well, pump, and environment related factors.
2. Machine-learned models trained on pump-level timeseries data considering various subsets of
engineered sensor data and trained on various failure events.
3. Expert knowledge-based methods that use engineering rules and ESP manufacturing information to
identify critical events from ESP sensor data.
4. Physics-based nodal analysis model based on thermodynamic, multiphase and PVT correlations, heat
transfer, hydraulic, electromagnetic and reservoir inflow performance.
6 SPE-214462-MS
All these four components together provide a better holistic view of the overall health of the pump and
result in a better final failure prediction model. As seen in Figure 5, a Bayesian optimizer determines the
relative weight for each model. This Bayesian optimizer is fine tuned for different fields separately and is
re-tuned regularly with updated weights when new data becomes available, and models are retrained.
All these four components together provide a better holistic view of the overall health of the pump and
result in a better final failure prediction model. As seen in Figure 5, a Bayesian optimizer determines the
relative weight for each model. This Bayesian optimizer is fine tuned for different fields separately and is
re-tuned regularly with updated weights when new data becomes available, and models are retrained.
Focusing on expert knowledge-based systems, the models can be classified into high-resolution and low-
resolution data models. The low-resolution class of models are not suited for short-lived event signatures
(events occurring and clearing within 6-12 hours, e.g., voltage spikes). Their advantage is in detection
of long-term trends, many of which can be obscured due to factors such as subtleness of the trends and
data display and scale. Furthermore, the PFA processing pipeline refines low resolution data to filter out
transients and other disruptions and ensures the highest quality low-resolution data feed into the rule-based
engine as summarized in Figure 6:
For most rule-based routines, the entire data is not of interest; instead, the engine should only be activated
following certain events such as a pump cycle or a setpoint change, or only after a period of steady pump
operation.
SPE-214462-MS 7
An example of a model that can be activated after an event is one that compares the current power to
the power that a healthy pump is expected to consume given its past operating conditions. The ability to
quantify this expectation provides valuable insight into the health of pump. The actual power consumed
being less or more than the expected value can be indicative of certain conditions.
The model uses numerical techniques to establish a baseline for each pump, prior to which no estimate on
power expectation is made. The functions f and g above are physics-based; however, they have additional
degrees of freedom built-in to be trained on asset-level data, with clustering when warranted. This makes
The advantage of being an ESP Original Equipment Manufacturer (OEM) is that specific and proprietary
diagnostic parameters are available to be used as models. One great example of such capability in PFA
is the expected range of parameters within which the pump operates optimally without damaging thrust
conditions. The areas outside of this operating window are where damaging vibrations, upthrust, or severe
downthrust conditions may exist and degrade pump's useful life. For pumps manufactured by our company
as the OEM, the operating window is calculated, using API calls to internal proprietary models, as operating
conditions change.
An example of such an operating window is presented in Figure 7. This example shows a case in which
the operating window for liquid production was calculated and displayed. Operations outside of this window
lead to damaging thrust conditions. In addition to this window the optimal operating region is available. In
the case of this specific pump displayed in Figure 7, in early life it was experiencing severe downthrust. Such
impacts are not only valuable as a part of the input to the ensemble model, but they provide a valuable visual
to engineers and operators looking to obtain insight in pump's operating conditions and ways to improve
the pump run life.
Figure 7— Display of operating window for liquid production (green area). The green dashed line
represents the optimal point. The blue solid line is the liquid production values from well test. The red line
represents the quasi-steady version of liquid production data filtering out pump cycles and transients.
Example Cases
In this section, we will focus in more detail on the PFA output for two different ESPs and how it assisted
the production engineers in making operational decisions. These cases are fully anonymized (including
references to ESP numbers, dates, and other identifying information).
SPE-214462-MS 9
ESP-1 is an example of a failure prediction which can enable proactive workover scheduling. Figure 8
represents an overview of PFA output. In this figure we see that PFA predicted a broken shaft (damage
indicator event, near the top), which ultimately was confirmed as a few missed stages out of the 122 total
stages in mid-Q1 of 2022. Concurrently, a strong alarm indicating the power consumed by the pump is lower
than expected was raised (purple line in second graph) and the pump's remaining useful life has decreased
to just under two months (orange line in first graph). This low RUL alerts the operator to prioritize this
pump for a work-over as its failure is inevitable due to the irreversible nature of this damage. This pump
Figure 8—PFA output for ESP-1 presenting an event of irreversible damage (broken shaft)
As showed in Figure 9, ESP-2 is an example of a reversible damage event in which the pump recovered
after remedial action. The well in which this pump operates tends to produce sand, causing the ESP to
wear out quickly or end up with a broken shaft. PFA output for ESP-2 is presented in Figure 9. In the case
of ESP-2, around June 2021, intake pressure and motor current started decreasing and motor temperature
started increasing. PFA predicted sand influx (blue line in second graph) and estimated a significant RUL
reduction (orange curve in first graph). Shortly thereafter, the operator shut the well in to perform remedial
intervention including injection of chemicals (gray vertical areas in first graph indicate the pump was shut
10 SPE-214462-MS
off). After opening the well back up and starting the ESP, the sand influx was remediated and the PFA
predicted RUL increased. The recovery in the remaining useful life removed this pump from the "high
priority" bucket of ESPs within PFA's asset-level dashboard (i.e., ESPs that need immediate attention and
either need to be pulled or to have remedial action performed on). It is noteworthy that PFA algorithm is
robust enough to detect critical events even when important data such as flow rate is missing.
Conclusion
In this case study, Predictive Failure Analytics (PFA) was trained and deployed for a major operator in Latin
America, who was experiencing short ESP run life in four of its high producing assets. PFA performed
reasonably well on the operator's set of pumps with a balanced accuracy of 73%. PFA performance met the
client's objectives in terms of key performance indicators and the solution was deployed.
Based on these outcomes, the production engineers could rely on the failure alarms, and the RUL
estimation generated by PFA, to take preventive actions. The operator was experiencing frequent broken/
fractured shaft and pump worn out failures due to sand. Using the PFA, the operator was able to identify
wells which had ESPs at high risk of failure. Specifically, the operator was able to see that PFA raised
an early broken-shaft/ missed pump stages alarm for ESP-1. This alarm enables the operator to schedule
procurement of services in a timely manner, reducing the downtime by two weeks, on average. This helps
SPE-214462-MS 11
the operator save $450K from reduced deferred production, in addition to saving other costs associated
with an emergency workover. For ESP-2, after PFA detected sand influx, the operator applied a chemical
injection to resolve the sand issue, this action extended the predicted ESP run life. To this date, the ESP is
still running without any further issue.
In summary, operators try to manage production and workover logistics, while also holding down costs
- both OPEX and CAPEX. Production and facility engineers have many roles to play in a developing field.
It is very difficult to glance briefly at a SCADA display and derive insights in a timely manner, while
Acknowledgments
The authors would like to thank Baker Hughes for the permission to publish this work.
Nomenclatures
ESP Electrical Submersible Pump
PFA Predictive Failure Analytics
DIFA Dismantle, Inspection, and Failure Analysis
RUL Remaining useful life
RULpred Predicted RUL
RULbase Baseline Rule
RULactual Actual RUL
AI Artificial Intelligence
ML Machine Learning
CPH Cox Proportional Hazard Method
PCA Principal Component Analysis
NLP Natural Language Processing
SVM Support Vector Machine
XGBoost eXtreme Gradient Boosting
KPI key performance indicator
BPD Barrel Per Day
PSI Pound Per Square Inch
SCADA Supervisory Control and Data Acquisition
PDP Pump Discharge Pressure
PIP Pump Intake Pressure
Ti Intake Temperature
Tm Motor Temperature
V Volts
A Motor Current
Hz Drive Frequency
Ct Current Leakage
Q Total Fluid
API American Petroleum Institute
MLE Motor Lead Extension
GDH Grounded Downhole
TPR True Positive Rate
TP True Positive
12 SPE-214462-MS
References
Abdelaziz, M., Lastra, R., & Xiao, J. J. (2017). ESP data analytics: Predicting failures for improved production