2021 IEEE Smart Information Systems and Technologies (SIST)

28-30 April, 2021, Nur-Sultan, Kazakhstan

Developing Predictive Oil Well Diagnostics

Based on Intelligent Algorithms
Zhanar Omirbekova 1, Aslan Amangeldiyev 2, Daur Aktaukenov 3 Abdelrahman Abdallah4
1
a) Al-Farabi Kazakh National University, 71 al-Farabi Ave., Almaty, e-mail: zhanomir@gmail.com,
https://www.kaznu.kz
2,3,4
Satbayev University, 22 Satbayev str., Almaty, email: reedgoody@gmail.com, dauraktaukenov@gmail.com,
www.satbayev.university

Abstract— The current competitive market condition in Various methods and tools are used for predictive fault
the oil industry is so concentrated on companies' budgets diagnosis without having to stop the process, thereby
that they require methods to extract oil from wells at the reducing the cost of repair work
lowest possible cost. All pumping units, new or old, require
regular preventive maintenance and constant inspection, III. DIAGNOSTIC METHODS FOR SUCKER-ROD PUMPS
and with 30 per cent of total oil production coming from
sucker rod pumps, the cost of diagnostics must be reduced A. The pump unit is connected to an electric or
accordingly combustion engine via a torque box, which converts
This article focuses on the intelligent diagnostics of the
the rotary motion of the engine into a reciprocating
rod and borehole pump for preventive maintenance and
monitoring during the life cycle of the well. The use of motion of the pump rods. The pump rods in turn
artificial intelligence methods for predictive diagnostics of transfer the mechanical energy received at the surface
equipment condition solves problems without production to the down-hole pump. Some energy is lost to
stoppages and without additional interventions from friction during this process. In figure 1 shows the
outside. main components of the SRP, namely the pump unit,
the pump rod, and the pump itself.
Keywords— Oil production, sucker rod pumps, artificial
intelligent system, dynamograph, preventive diagnosis,
machine learning.

II. INTRODUCTION
The current state of the oil and gas industry in
Kazakhstan is characterized by the fact that the vast
majority of operations are at a late stage of development,
characterized by increased water cut, reduced average
flow rates from oil production wells, and higher costs per
unit of oil produced. In addition, the oil and gas industry
are just recovering from a severe downturn that led to
significant operating losses for oil producers. Figure 1 Mechanics of the sucker-rod pump
According to a 2019 report by the Oil and Gas
Information and Analysis Centre of Kazakhstan, 90.4 The depth pump, which is shown in figure 2, transfers the
million tons of oil were produced, of which 30% were resulting mechanical energy into the multiphase fluid (oil,
produced using sucker-rod pumps [1, 2]. gas, sediment, and water).
As is known, a sucker-rod pump (SRP) is a single-acting
pump that converts the rotary mechanism of the motor
into a vertical reciprocating motion to drive the pump
shaft and is used to mechanically lift oil-bearing fluid
from wells [].
According to Ambyint, it costs about $250,000 to install
the SRP together with all the starting equipment, and up
to $100,000 to repair or replace it [3]. This fact, with
declining oil prices and decreasing profits, is forcing oil
producers to use smart maintenance technologies to
reduce costs, improve productivity and make their assets
more profitable. Figure 2. Depth pump working principle
The main components of a submersible pump are the friction of the rod against the liquid reduces the stroke
plunger, casing, pressure valve and suction valve. length of the plunger and the suction valve closes at A1
Together they form a positive displacement pump system. (interval f↓) instead of point A.
When the plunger moves downwards, the pressure valve Stage 2 - the polished rod changes direction and moves
opens, and the suction valve closes. In this way, the upwards. When the direction of movement of the plunger
volume of the oily liquid column is supported by the is changed, the process is captured by the straight-line
suction valve assembly and the liquid flows into the segment A-A2. Starting from point A2 the pump rod
housing. The inside of the plunger is filled with oily takes the load from the weight of the liquid column Pf. At
liquid. Plunging the pump rods into the liquid causes a point B1 the load is equal to the sum of the weight of the
small discharge due to the volume that has been rod with the liquid and the friction forces. That is, at A2-
displaced. B1 the pump rod stretches and the load increases, but the
As the plunger moves upwards, the pressure valve closes, polished rod does not move and is still in the bottom end
and the suction valve opens. The oily liquid moved by the position of the plunger. At this point the pump suction
plunger comes to the surface, while the body is filled valve opens and the oily liquid flows into the pump
through the suction valve. In this way, the volume of cylinder.
liquid in the pipe is transferred to the pump rod. Stage 3 - polished rod and pump plunger move upwards
During continuous operation of the deep well pump, due which is described by the line B1-C1. The line is also
to the reciprocating motion of the rods, the pumped oily parallel to the s-axis of the graph due to constant rod
liquid begins to fill the pump rod and rise to the surface. pressure and friction force.
There are three methods to diagnose the condition of the Stage 4 - The polished rod changes direction and moves
sucker-rod pump without stopping it and raising the downwards, the friction force also changes its direction.
plunger to the surface: the dynamometer method, the The process is recorded by the straight-line C-C2. The
wattmeter method and using artificial intelligence based load change corresponds to C2-D1 when the rod string is
on the analysis of sucker-rod diagrams. unloaded, and the pipes are loaded. Point D is the
opening of the pump discharge valve and the start of the
B. Dynamometry downward movement of the plunger.
Dynamometry of the rod pumping unit is the most
important source of information about the performance of
the rod pump, the rod string, the condition of the
downhole hole, etc. It is carried out by means of special
technical means [4]. It is carried out by means of special
technical means [4]. Data from the dynamometer are
transferred to a portable data acquisition module and then
to a centralized upper-level system, or to a programmable
controller installed at the well site and allowing to
process, analyze and control operating modes of pumping Figure 3. Dynamogram profile
equipment [5]. The dynamogram is a graph of load
dependence, at the suspension point of the rod, on the The reading of the dynamograms allows quantitative and
movement of the polished rod. The theoretical qualitative determination of the pump's performance
dynamogram for normal pumping operation is based on loads and stresses in the polished rod, plunger and rod
gravity, elasticity, friction, and Archimedes' principle. stroke lengths, degree of pump filling, tightness of the
Insufficient consideration of other influencing factors, pump suction and pressure valves, gas impact, correct
such as the force of inertia and the properties of the plunger fit, leaky tubing, rod or rod coupling turnovers
pumped liquid, limits the capability of this method. and breakages, plunger jamming. The dynamometer can
The dynamogram is shown in figure 3, which is a describe 30 different state parameters of the sucker-rod
diagram of the rod load - p as a function of the polished pumping equipment without lifting it. The transformation
rod stroke - s. The process can be divided into four steps. curve is a multi-resolution method that has been widely
Point A in the dynamogram indicates the bottom-most used to solve feature extraction problems [6,7,8].
position of the plunger and point C corresponds to the The shape of the dynamometer depends on the type of
top-most position. failure of the sucker-rod pump and can be used to identify
Stage 1 - polished rod and pump plunger move the following types of failures in the sucker-rod pumping
downwards. Line D1-A1 corresponds to the load equipment: fluid leakage from the tubing and valves
difference from rod pressure and friction force P, the line (suction and pressure); mechanical failures such as high
is parallel to the zero line (s axis) of the dynamometer and low plunger fit, rod breakage, plunger jamming, etc;
due to the constant rod pressure and friction force. Line other problems related to pumping liquids gas, sand,
D-A corresponds to the static pressure of the rod in the paraffin.
liquid Prod without friction force. Consequently, the
C. Wattmetrics process when lifting of formation fluid, high dynamics and
Modern SRPs are driven by variable frequency drives, exceeding of normative values of loads in elements of
which provide control, speed adjustment and parameter sucker-rod pump.
control of the SRP. The method of watt metering is based
on the power consumption of these variable frequency
drives. Wattmetrics is the process of obtaining a watt-
hour diagram which shows the relation between the
power N(φ) consumed by the pump motor and the angle
of the crankshaft, or the dependency of the power
consumed by the installation on time [4]. The analysis of
the wattmeter spectrum enables the detection of vibration Figure 5. Wattmetrograms of unbalanced SRP (low load)
and shock loads, which makes it possible to diagnose
gearbox and bearing defects [9]. The difference between The analysis of wattmetrograms makes it possible to
the Wattmeter and the dynamometer is that the predict the most common faults of deep well pumps. [4]
dynamometer is mainly used for the submersible part of The classification of common faults can be solved by
the shaft, while the Wattmeter is more informative for the linear partitioning of the phase plane into parts, the
above-ground part. appearance of each of which is determined by the
During the stroke, the pump rods are subjected to static presence or absence of certain faults. The classification
and dynamic loads. The downward movement of the features in the wavelet transform wattmetergram are
pump rods at the very end of the stroke causes a higher determined by the equation [10]:
load, while the upward movement of the pump rods at the
very end of the stroke causes a lower load. Since the i 

 f (t ) ( a
i
pump rods are directly connected to the pump itself, these w(i , k ) n  a02 0 t  kb0 )dt (1)
loads are transferred to the variable speed drive. Thus, 
information about the condition of the underground
equipment is contained in the wattmeter spectrum. Where, φ(i,k) is the basis wavelet function;
Experimentally, each state-specific wattmetergram can be f(t) is the wavelength function.
linked to one of the various dynamometer classes, which i,k ϵZ.
are widely used in rod pump monitoring and diagnostics. Diagnostic signs of sucker-rod failure are an increase
This can enable rod pumps to be diagnosed directly from in peak active power consumption by the motor when the
the wattmetrogram. Furthermore, rotor defects, static and sucker-rod head is lowered (the weight of the sucker-rod
dynamic eccentricities and bearing defects can be string does not compensate for the counteracting torque
identified by the spectral power density. created by the crank). At the same time, the half-period of
A wattmeter diagram of a fully balanced and serviceable power increase is absent in the watt diagram during rod
SRP is shown in Figure 4. For each full stroke of the lowering, which is caused by absence of external load
boom the SRP has two half-periods with significant peaks acting during lifting of formation fluid column of the
corresponding to the horizontal position of the crank. borehole. Similarly, defects in the pressure valve of the
According to the regulations, the difference in maximum submersible pump are shown. They consist in a significant
power consumption by the drive motor of the pump unit decrease of power, consumed by the engine during the
during the lowering of the depth pump rod must not boom lift, because of leakage of most of formation fluid
exceed 10% [15]. through the defective pressure piston valve back to the
pump cylinder, and as a result, a significant drop in the
point of suspension of the pump rods load. Uneven load
growth during lowering of sucker-rod string, increase of
dynamic force components at sucker-rod drive are the
evidence of faulty sucker-rod pump valve.
The spectrum analysis of active power signal, shown
on a figure 6, allows to reveal frequency components from
Figure 4. Wattmeter diagram of a fully balanced SRP
0 to 30 Hz, characterizing oscillations of kinematic chain
"downhole pump - pump unit - reducer - driving motor".
Insufficient amount of counter torque, created by [4] Defects in working units and mechanical transmissions
cranked loads during lowering of sucker-rod pump of sucker-rod drives determine occurrence of variable
column at unbalanced sucker-rod pump (fig. 5), leads to loads, that cause appearance of new spectral components.
transition of drive motor to generator mode, and working Periodic measurement of values in a power spectrum,
stroke of pump rod is accompanied by increased loads on characterizing concrete defects in the drive engine and
gearbox and pumping unit motor. The consequence of mechanical transmission, allows to carry out estimation of
these processes is increase of specific energy consumption technical condition of SRP in the simplest mode and, if
necessary, to carry out repair actions for prevention of large: first, the computational complexity would become
failures. too great; second, increasing the size would eventually
lead to a decrease in performance [13]. To reduce the
dimensionality of the feature space, there are two
different approaches. One is to discard certain elements
of the feature vector and leave the most representative
ones. This type of reduction is feature selection [14]. The
first harmonic of the
reversible motor
other is called feature extraction, in which the original
feature vector is converted into a new feature vector using
a special transformation and the new features have much
smaller dimensions
Figure 6. Wattmeter spectra in the presence of a pump-motor unit
defect. IV. SIMULATION
Spectral analysis of energy consumption records in The simulation in the first stage is run-to-fail (RTF),
asynchronous motors (Figure 6) can be used to detect which means that wells are maintained only after an
defective electrical rotor parts, including rotor winding accident occurs, and the goal of RTF simulation is to
breaks and short circuits between layers; stator electrical generate a large volume of telemetry and repair data.
parts, including power winding breaks and electrical Machine learning (ML) models are then trained on the
asymmetry, short circuits between layers; static and telemetry and repair logs to predict which production
rotating eccentricities; bearing defects resulting in wells are likely to fail soon. Once the machine learning
fluctuating air gap shapes. Gearwheel faults, gear fit on models are built, the simulation is repeated in predictive
the shaft, misalignment of driven shafts and their rotation maintenance mode, which uses these machine learning
supports, and defects in the V-belt transmission can be models to mark those wells that are likely to fail soon and
diagnosed by the nature of the change in the power signal then assigns available service personnel to perform
spectrum. preventive maintenance on the marked wells.
D. Artificial intelligence methods A. Initial Data
Dynamometry provides an exactly accurate indication of
the condition of the SRP, but also requires intervention The input data for the task is generated in the form of an
by workers, because a dynamograph will need to be excel spreadsheet with the following well parameters:
installed and the instrument read. In Kazakhstan, number of SRPs, number of time steps, time interval each
dynamometer readings are taken once a month in oil sensor has to wait before sending any telemetry,
wells, which means there is no continuous diagnostics of maintenance strategy used in the simulation, number of
the SRP's condition, and it is impossible to predict its technicians available to service oil wells, and time needed
condition. for technicians to repair the well where an emergency
The disadvantages of this method of evaluating the failure has occurred.
technical condition of the SRPs with respect to their
energy consumption include the difficulty in detecting B. Simulation results and discussion
several defects at the initial stage of development. Firstly, Well flow rates decrease over time due to oil
these are roller bearing failures, crankshaft rotation, wear, accumulation, and production is also interrupted when
and destruction of connecting rod pins. The occurrence emergencies occur and they are repaired, as shown in the
and development of these defects is accompanied by a graph of well No. 123's production efficiency over time.
change in the diagnostic signals of the spectra in the
higher frequency range [16]. In real plant conditions, in
the presence of liquid leaks in the pump, gas influences
and other irregularities in the normal operation of the
pump, the deciphering of the diagnostic readings
becomes even more complicated and, as a rule, the
influence of these irregularities is mixed and it is difficult
to distinguish, in an explicit way, the influence of a single
indicator on the pump malfunction and on its delivery
[11]. The use of artificial intelligence techniques for
predictive diagnosis of the condition of equipment can
solve these problems without stopping production and Figure 7. Processing capacity graph over time for machine 123
without additional intervention from outside. One of the
problems associated with pattern recognition is the so- Different problems by design are more or less likely in
called curse of dimensionality [12]. There are two reasons different parts of the temperature, pressure and load
why the dimensionality of a feature vector cannot be too parameter space, and the following scatter diagram shows
that plough jamming is most likely to occur when the estimation, also used as a loss function for regression
load on the pump rods becomes large. tasks in training an algorithm. It usually expresses
accuracy as a percentage, and is defined by the equation:

1 n Yi  Yˆi
MAPE =
n

1 Yi
(2)

Figure 8 Scatter diagram of the "Plunger jamming" incident

The scatters of failures and deviations had to be Figure 11 quality of the Simulation algorithms
compiled over the simulation period in terms of classify
failures rapidly and accurately in the future The simulation result shows that we are able to reduce the
MAPE error to 0.01 % (fig 11). At the same time, the
number of training cycles can be reduced, as already after
10 training cycles, the model losses are no longer
significant.
Based on the results of the ML training experiments,
model number 3 can be selected for further testing on
online data in the SCADA system. For this purpose, the
model was saved on TensorFlowServing. New data are
fed to the model by means of API requests in JSON
format and the result from the trained model was fed back
to any SCADA, where graphs were already built based on
previous readings and predicted values.
Figure 9 Load simulation results to process parameters
ACKNOWLEDGMENT
The result of the simulation is shown above, depending
This research was conducted during the master's and PhD
on the current process parameters during loading and the
thesis of Satbayev University and the focus in simulation
identification of the optimum parameter in the process,
which ensures optimal operation of the pump in order to thanks to open data access and the support of al-Farabi
Kazakh National University
avoid failures.
REFERENCES
[1] Oil and Gas Information and Analysis Centre. Infographics // Nur-
Sultan, 2020.
[2] NC Kazmunaigas JSC. Annual Report 2019 // Nur-Sultan, 2020.
[3] David Wethe 'Nodding donkeys' challenged to pump oil from
horizontal wells // Bloomberg Dec. 15, 2017
[4] Shinyakov Y., Sukhorukov M., Torgaeva D., Soldatov A. Methods
Figure 10. Production efficiency and the use of technicians with a and means of control of the rod pump operation // International
predicate Journal of Engineering and Technology, 2018
[5] V.D. Kovshov, M.E. Sidorov, S.V. Svetlakova, Dynamometry,
modelling and diagnostics of the deep rod pumping unit condition
As can be seen from the graph, the average use of //
technicians for repairs due to the failure of the SRP has [6] D.O. Bruno, M.Z.D. Nascimento, R.P. Ramos, V.R. Batista, L.A.
decreased, and the average efficiency of production has Neves and A.S. Martins, LBP operators on curvelet coefficients as
also increased. an algorithm to describe texture in breast cancer tissues, Expert
Systems with Applications. 55 (2016) 329–340.
C. Testing and validation of simulation results [7] F. Li, F. Fang and G. Zhang, Unsupervised change detection in
SAR images using curvelet and L1-norm based soft segmentation,
The quality of algorithm performance is verified by International Journal of Remote Sensing. 37 (14) (2016) 3232–
methods known in machine learning: MAPE, MSE, 3254.
RMSE The mean absolute percentage error (MAPE), also [8] Y. Li, Q. Yang and R. Jiao, Image compression scheme based on
curvelet transform and support vector machine, Expert Systems
known as the mean absolute percentage deviation with Applications. 37 (4) (2010) 3063–3069.
(MAR), is a measure of prediction accuracy of a
prediction method in statistics, such as in trend
[9] Khakimyanov M.I., Pachin M.G., Monitoring of sucker-rod and SVM with application to indicator diagram diagnosis, Expert
pumping units condition by the results of wattmetrogram analysis Systems with Applications. 38 (10) (2011) 12721-12729.
// Electronic scientific journal "Oil and Gas Business" № 5, 2011 [13] van der Heijden, F., Duin, R., De Ridder, D., & Tax, D.. (2004).
[10] Dunayev I.V., Diagnostics and monitoring of downhole sucker- Classification parameter estimation and state estimation: An
rod pumping unit condition on the basis of dynamometry and engineering approach using MATLAB. John Wiley & Sons Inc.
neural network technologies // Department of Technical [14] Theodoridis, S., & Koutroumbas, K. (2003). Pattern recognition.
Cybernetics at Ufa State Aviation Technical University, 2007[in New York: Academic Press.
Russian] [15] Bubnov M., Zuzev A. Means of diagnosing equipment of rod
[11] A.G.Rzaev, S.R.Rasulov, A.M.Abdurakhmanova // Research of deep-well pump installations // Proceedings of the First Scientific
deep pump condition diagnostics, Institute of Control Systems, and Technical Conference of Young Scientists of the Ural Power
Azerbaijan State University of Oil and Industry, 2016 Engineering Institute. Yekaterinburg, 2016. [in Russian]
[12] K. Feng, Z. Jiang, W. He and B. Ma, A recognition and novelty [16] Barkov A., Barkova N., Borisov A. Methods for diagnostics of
detection approach based on Curvelet transform nonlinear PCA mechanisms by electric current consumption // North-West
Educational Center. Saint Petersburg, 2012.)[in Russian]