Integrated Computer-Aided Engineering 30 (2023) 353–368 353

DOI 10.3233/ICA-230712
IOS Press

Connected system for monitoring electrical

power transformers using thermal imaging
F. Segoviaa,∗ , J. Ramíreza , D. Salas-Gonzaleza , I.A. Illána , F.J. Martinez-Murciaa ,
J. Rodriguez-Riveroa,b , F.J. Leivab , C. Gaitanb and J.M. Górriza
a
Department of Signal Theory, Networking and Communications and DASCI Research Institute, University of
Granada, Granada, Spain
b
Endesa Distribución, Madrid, Spain

Abstract. The stable supply of electricity is essential for the industrial activity and economic development as well as for human
welfare. For this reason, electrical system devices are equipped with monitoring systems that facilitate their management and
ensure an uninterrupted operation. bnbThis is the case of electrical power transformers, which usually have monitoring sys-
tems that allow early detection of anomalies in order to prevent potential malfunctions. These monitoring systems typically
make use of sensors that are in physical contact with the transformer devices and can therefore be affected by transformer
problems.
In this work we demonstrate a monitoring system for electrical power transformers based on temperature measurements
obtained by means of thermal cameras. Properly positioned, the cameras provide thermal data of the transformer, the incoming and
outgoing lines and their surroundings. Subsequently, by appropriate image processing, it is possible to obtain temperature series
to monitor the transformer operation. In addition, the system stores and processes thermal data in external equipment (placed
in locations other than the transformers) and is equipped with a communications module that allows secure data transmission
independent of the power grid. This aspect, along with the fact that there is no need to have physical contact with the transformer,
make this approach safer and more reliable than standard approaches based on sensors. The proposed system has been evaluated in
14 stations belonging to the Spanish power grid, obtaining accurate and reliable temperature time series.
Keywords: Connected system, mser segmentation, power transformers, thermal imaging analysis

1. Introduction transformation systems are equipped with a variety of

sensors and technologies that allow real-time monitor-
Electrical power transformers used in contemporary ing of the system’s behavior, so that a prompt action can
medium and low voltage distribution networks are criti- be taken in the event that the values of phase, intensity,
cal components of the electrical infrastructure in devel- energy, etc. provided by a power transformer deviate
oped nations. For this reason they are one of the main from those expected [1]. In addition, in recent years,
concerns of the industry and the sector’s regulators. some changes in society are threatening the resilience
These transformers are intended to be working without of the power system, which emphasizes the relevance
interruption and this involves to be able to handle peaks of these monitoring technologies [2].
demand and prevent that potential temporary changes On the one hand, the introduction of disruptive de-
in the transformer workload could result in unaccept- velopments such as electric cars, which demands large
able voltage levels or cause any damage to the hardware amounts of energy in a short time [3–6], or the increased
system. In order to accomplish this requirement, power use of renewable production sources (of an intermit-
tent character in the majority of cases [7]) are making
∗ Corresponding the power transformation management more compli-
author: F. Segovia, Department of Signal Theory,
Networking and Communications, University of Granada, Granada, cated and challenging [8], requiring more sophisticated
Spain. E-mail: [email protected]. monitoring systems.

ISSN 1069-2509 c 2023 – The authors. Published by IOS Press. This is an Open Access article distributed under the terms of the Creative
Commons Attribution License (CC BY 4.0) etgw45.
354 F. Segovia et al. / Connected system for monitoring electrical power transformers using thermal imaging

On the other hand, electric power systems are cur-

rently experiencing a digitization process, which in-
volves the integration of technologies associated with
computing, wireless communication, Internet of Things
(IoT) devices and a multitude of applications aimed at
facilitating real-time and seamless operations [9–11].
These advances have enabled the development of more
efficient semi-autonomous systems with reasonable op-
erating costs, but have also brought additional concerns,
including cyber security concerns, that should be ad-
dressed [12]. Fig. 1. Malfunction probability in different components of a power
The management of electricity networks must pay transformer.
special attention to potential overloads [13,14], as they
are one of the major problem sources affecting power insulation moisture analysis, and controlling the op-
transformers, whose failures represent approximately eration temperatures [24]. The measurement of both
12% of the total [15]. Transformer failures, in addi- the transformer temperature and the room temperature
tion, are particularly dangerous as they can cause fires is therefore part of standard protocols for monitoring
which, taking into account the urban location of trans- power transformers [25–27]. These are relevant vari-
formers, can result in serious economic and personal ables that should be measured accurately and uninter-
damages [16]. ruptedly and can also be used to forecast other variables
The malfunctions affecting power transformers de- such as reactive power or intensity [28].
pend to a large extent on the transformer design, with In conventional approaches, temperature measure-
important differences depending on the type of primary ment is usually carried out by sensors that are integrated
insulation [17]. In addition, the presence of certain com- into the transformer equipment and are, to some extent,
ponents that are not present in all transformers may part of the transformer system. In this regards, a vari-
cause failures to be concentrated in some elements or ety of sensors is available, and their characteristics and
others. Thus, for example, in transformers equipped principles of operation are highly dependent on the type
with On-Load Tap Changer (OLTC), the percentage of of transformer whose temperatures are monitored. In
failures associated with bushings and terminals is 14%, Liquid-Immersed power transformers, the temperature
while in transformers not equipped with OLTC the per- of the liquid (usually oil of different natures) is used
centage of failures of these components is more than as a estimation of the transformer temperature. In such
double (33.3%) [18]. cases, contact probes are used to measure the liquid
During the last years, several authors [19–21] as well temperature. The temperature of the transformer wind-
as different international organizations have published ing is also usually measured by contact sensors. This
a number of reports analyzing the probability of fail- approach is possibly the simplest and most straightfor-
ure of the components of a power transformer. One of ward, however, it has the disadvantage that it can be
these organizations, the International Council on Large affected by potential transformer problems, thus losing
Electric Systems (CIGRE) has divided the types of fail- its ability to predict transformer malfunctions [29].
ures into 6 categories based on the component causing In this manuscript we describe a system based on
the failure: winding, tap changer, core, tank, bushing, thermal imaging to monitor the room and transformer
auxiliary. The probability of failure for each is shown temperature of electrical power transformers. Thermal
in Fig. 1 [22]. imaging or thermography is a mature technology that
The reason for failures depends to a large extent on allows us to capture the invisible heat emitted by objects
the component in which they have occurred. Thus, for whose absolute temperature is above zero [30]. It has
example, tank failures can occur due to corrosion, which been successfully applied for monitoring purposes in
in turn is linked to environmental conditions (high hu- many fields, as civil structures [31,32], machinery in-
midity, solar radiation, etc.) and can cause fissures in the spection [33] or in the nuclear industry [34]. In addition,
tank walls, resulting in leaks. To avoid these failures, this technology has been recently proposed to correct
in addition to proper maintenance of the installations, temperature problems in electrical substations [35].
a number of measurements should be carried out on a The proposed approach has been implemented and
regular basis. This includes dissolved gas analysis [23], evaluated in 14 transformation centers belonging to the
 F. Segovia et al. / Connected system for monitoring electrical power transformers using thermal imaging 355

Fig. 2. PASTORA functional blocks and their relationship with the MONICA platform.

Spanish power grid [36]. The measures obtained with and its evolution. In this work, information exogenous
it were validated against the standard sensor-based ref- to the transformer (variables extracted from thermal
erence scheme. Compared to the latter, the proposed images) was included, demonstrating the relationship
approach offers some advantages. The most important between this information and other electrical variables.
of these is the fact that the system is completely inde- This work is part of the PASTORA project1 (ref.
pendent of the power transformer, and without physical EXP – 00111351/ITC-20181102), which is being de-
contact with it, so it is not affected by potential trans- veloped by ENEL and other international companies as
former failures (high temperature peaks or too frequent an extension of the MONICA project.2 Both projects
temperature fluctuations may damage the sensors). It is aim to develop an advanced monitoring system for the
also a more scalable system, capable of monitoring not distribution networks and the electrical transformation
only the temperature of the transformer, but also of dif- system that allows operators to perform real-time mon-
ferent transformer components independently, or of ad- itoring and preventive control of these assets. These
jacent equipment. This would be possible by only mod- objectives include the use of artificial intelligence to
ifying the thermal image processing algorithm, with- enable self-management between the downstream el-
out hardware changes. Finally, the proposed system ements of the network and the medium voltage sub-
based on thermal cameras can be used as a security sys- stations, through intelligent electronic devices and au-
tem capable of detecting the presence of living beings tomated control centers. Figure 2 shows a functional
(animals or people) near the transformer, based on the diagram of the PASTORA project, as well as the main
temperature changes detected. information flows between the different subsystems and
The contributions of this work are both technical and the MONICA platform. It includes the 5 parts in which
scientific. On the technical level, this is the first time the project objectives are divided and the companies of
(to our knowledge) that electrical variables and tem- the consortium in responsible for implementing each of
peratures obtained from sensors and temperature data them.
extracted from thermal images have been analyzed to-
gether. This study, in fact, is part of a pilot project to
evaluate the usefulness of this system in order to subse- 1 Official website: www.endesa.com/en/projects/all-projects/

quently extend it to the whole ENEL network in Europe. energy-transition/smart-grids/pastora-artificial-intelligence-

distribution-grid.
On the scientific level, this work includes an extension 2 Official website: www.endesa.com/en/press/press-room/news/
of the work presented in [37], in which Granger causal- energy-transition/smart-grids/monica-project-will-enable-endesa-
ity is used to characterize the state of the transformer to-obtain-real-time-data-about-low-and-medium-voltage-lines.
356 F. Segovia et al. / Connected system for monitoring electrical power transformers using thermal imaging

Fig. 3. Main devices required by the proposed monitoring system and connections among them.

The rest of the manuscript is organized as follows: Celsius. Nevertheless, this value highly depends
Section 2 contains the details about the system imple- on the environment in which the system is in-
mentation, including hardware description, communi- stalled, including both, the climatic conditions of
cation protocol and image processing. Section 3 shows the region and the size and ventilation of the room.
the comparison between the measures obtained by the
proposed system and those obtained by the standard 2.2. Thermal-imaging-based system description
approach. Then, Section 4 discusses the advantages and
limitations of each solution. Finally, Section 5 contains The monitoring system, designed and implemented
the conclusions and future work plans. to to meet the requirements specified in Section 2.1,
was composed by several devices placed in different
2. Material and methods locations. The acquisition equipment (AE) is located
next to the power transformer and composed by a ther-
2.1. System requirements mal camera, single-board computer (SBC) and com-
munication module. The storage and processing server
A temperature monitoring system for electrical trans- (SPS) is located in a data center, away from the power
formers based on thermal imaging must satisfy the fol- transformer (see scheme in Fig. 3). Further details about
lowing minimum requirements: the specification of these devices are given as follows:
– Uninterrupted operation. Since the electrical – Thermal camera. A commercial model from Op-
transformation system operates 24 hours a day, tris, the Xi 80, was selected. This model provides
365 days a year, the monitoring system must also thermal images of 80 × 80 pixels at a frame rate
be able to operate uninterruptedly. of 50 Hz. The field of view (FOV) is 80◦ at a fo-
– Acquisition frequency of at least one thermal im- cal length of 2.3, enough to cover the surface of
age per minute. In order to provide a promptly the transformer and part of the protection struc-
respond to temperature changes, the acquisition
ture, which is used to estimate the room tempera-
system must be able to take at least one thermal
ture. The purpose of this device is just to capture
image per minute.
the thermal data and send them through the USB
– Coverage of an area of at least 1.5 × 1.5 meters.
The acquisition instrument must cover an area of interface.
at least 1.5 × 1.5 meters when placed at a distance – Single-board computer. This device controls the
of 2 meters. This area permits to include most of thermal camera (which is connected via a USB
the surface of the transformer, the input/output port) and temporally stores the thermal images on
connections and the protective fence. a memory card, which works as a small buffer. In
– Secure external data storage. Thermal data and all addition, it periodically submits the images stored
information related to temperature control must on the memory card to the SPS using the com-
be securely stored in a location external to the munication module described bellow. This SBC is
transformation system. based on a Raspberry Pi 4 equipped with a 32 GB
– Secure communications system. Communication SD card.
for sending and storing data externally must be – Communication module. A Teltonika industrial
done in a secure and reliable manner. cellular router (RUT240 model) was used to en-
– Tolerate operating temperatures above 40 degrees able the SBC to access the storage server via In-
 F. Segovia et al. / Connected system for monitoring electrical power transformers using thermal imaging 357

Fig. 4. Device scheme showing the integration of several AEs and communication between AEs and SPS by means of an intermediate server.

ternet. This model can access to Internet using collected. After analysis them, the Optris Xi 80 model
commercial mobile networks, by means of a SIM was selected due to its cost, which was substantially
card of any carrier. This communication module lower than other proposals, while maintaining the main
is connected to the single-board computer through requirements in terms of resolution, frame rate and field
its Ethernet interface. of view.
– Storage and processing server. This equipment
is used to store and analyze the thermal data re- 2.3. Logical control and security aspects
ceived from the AE located in power transformers
The logical control of the proposed monitoring sys-
(it can serve to several AEs located in different
tem is carried out by three processes that are run in
power transformers). Its main task is to convert
the SBC and the SPS using a structure focused on re-
thermal images into a series of values representing liability and security (note that thermal data is trans-
the room and transformer temperature, and storing mitted via the Internet and special attention should be
them. Due to security concerns, the SPS is logi- paid to the communication security). On the one hand,
cally located behind an additional device that op- two processes running on the SBC are responsible for
erates as intermediary, therefore in practice two thermal data acquisition and temporal storage. The first
servers are needed for this task. More details are one captures thermal images at a rate of one image per
given in Section 2.3. minute and stores them on the temporal storage of the
It is worth noting that, if the temperature of several SBC. The second one submits the thermal data acquired
power transformer should be monitored, several AEs in last minutes to the SPS for processing and perma-
(one per power transformer) are required, but only one nent storage. The destination of this submission is in-
SPS. deed an intermediate server that work as firewall for
This selection of components was carried out on the the SPS. Both, data acquisition and sending rates are
basis of both functional and economic criteria. First, parameterized, so they can be easily adapted to future
a selection of components was made with those that requirements. On the other hand, a process running on
met the required characteristics. Subsequently, the low- the SPS periodically collects the thermal data that has
est cost component was selected from among those se- been placed on the intermediate server. The connec-
tion scheme is shown in Fig. 4. This structure, with
lected. Special attention was paid to the economic crite-
an intermediate server and two processes to move data
ria since the system is intended to be an alternative the
between the temporary storage in the SBC and the SPS,
cheaper sensor-based monitoring system and, to that
allows the SBC to not have access to the SPS, which
end, the price increase compared to the latter must not
improves the security of the entire system by preventing
be excessive. In this regard, the most critical element the SPS from being accessed if the security of any AE
of the system in terms of price was the thermal camera, is compromised.
which represented the majority of the hardware cost.
Thus, an exhaustive search was carried out for this com- 2.4. Image processing
ponent and several quotations from the main manufac-
turers (including Infratec, Fluke, Flir and Optris) were After storage in the SPS, the thermal images were
358 F. Segovia et al. / Connected system for monitoring electrical power transformers using thermal imaging

Fig. 5. Thermal image (left) and visible image (right) of a power transformer currently serving a residential area in southern Spain. Note that both
images were acquired with different cameras as thermal camera used in this work only provide thermal images.

processed to extract temperature information. Both, Gaussian Mixture Model to model the image. Pixels
room and transformer temperatures, were estimated as belonging to each component (Gaussian) were treated
the average temperature of pixels at specific locations as different regions and, finally, spatially separated re-
in thermal images. Transformer temperature was ob- gions of equal intensity are considered independently
tained from pixels containing the surface of the power by applying the MSER algorithm. All these algorithms
transformer while room temperature was obtained from are briefly described below.
the pixels of the fence that is part of the metallic struc-
ture that protect the power transformer. Figure 5 shows 2.4.1. SLIC superpixel method
a thermal image (left) and its visible-range equivalent Simple linear iterative clustering (SLIC) is a clus-
(right) of a power transformer currently serving a resi- tering method proposed by [38]. It is a variant of k-
dential area in southern Spain. The blue T-shape region means [41] that makes more efficient use of memory
in the lower half correspond to the protection fence and achieves higher speed. The main differences from
and was used to estimate the room temperature. The k-means are as follows [38]:
transformer temperature was estimated from the pixels – SLIC significantly reduces the number of calcula-
containing the transformer surface, placed on lower left tions and complexity by limiting the search region
and right corners. to the size of a superpixel.
In order to efficiently process the time series of ther- – It uses an adaptive distance measurement that com-
mal data, the calculation of room and transformer tem- bines spatial proximity and color, allowing both
peratures was carried out into two phases. First, a ref- size and compactness of superpixels to be con-
erence made-up thermal image (representing tempera- trolled.
tures independent of time) was segmented in order to The algorithm begins by establishing the center of k
obtain two masks, one for selecting pixels of the fence clusters, uniformly distributed in the image. The param-
and other one for selecting pixels of the power trans- eter k represents the number of superpixels and is the
former surface. Then, both masks were applied to all only parameter of the model. Each pixel is then associ-
the images in the time series to extract the pixels of the ated to the closest cluster center within its search area,
fence and the pixels of the transformer surface sepa- defined asp a square centered on the pixel and of side
rately and, from them, obtain the two target temperature equal to 2 N/k, being N the total number of pixels.
series. Next, cluster centers are recalculated so that they are
For the first phase, two segmentation approaches centered on the region they cover. The assignment and
were evaluated. The first one carried out the process in update steps are repeated iteratively until the error (dif-
3 steps: i) downsampling by means of the SLIC super- ference between cluster centers in successive iterations)
pixel method proposed by Achanta et al. [38], ii) re- converges.
quantization using several thresholds computed accord-
ing to [39] and iii) separation of the regions of equal 2.4.2. The Otsu’s method
intensity by applying the Maximally Stable Extremal The Otsu’s method [39] is a non-parametric algo-
Regions algorithm [40]. The second approach used a rithm frequently used for threshold selection in image
 F. Segovia et al. / Connected system for monitoring electrical power transformers using thermal imaging 359

segmentation problems. Its main idea is to select the modeled by a sum of k Gaussians:
threshold that maximizes the separability of the result- k
X
ing classes in gray levels [42]. p(x) = ωi fi (x|µi , Σi ) (2)
The procedure uses the zeroth- and the first-order i=1
cumulative moments of the gray-level histogram and where ωi , µi and Σi are respectively the weight (within
reduce the problem to an optimization problem that the mixture), mean and standard deviation of each of k
maximizes the class separability. Formally, in a problem Gaussians. Therefore, by finding these parameters we
with two classes the threshold, k ∗ , is selected as: can model each of the underlying groups of our data.
GMM parameters can be accurately estimated using
2
σB (k ∗ ) = max σB
2
(k) (1)
1<k<L an iterative Expectation-Maximization (EM) approach.
where σB 2
represents the class separability and can be This algorithm starts by initializing the ωi , µi and Σi
computed as ω0 ω1 (µ1 −µ0 )2 , being ωi and µi the prob- with the values provided by a k-means algorithm and
ability and mean of class i. proceeds by iteratively applying two steps:
– For each of the samples in our data set, the poste-
2.4.3. Maximally stable extremal regions algorithm rior probability of belonging to each component
SLIC and Otsu’s methods allow us to group the pix- (Gaussian) is calculated.
els of an image according to their intensity level but – Using the maximum likelihood, component means,
do not take into account the distance between pixels. covariance matrices, and mixing ratios are com-
The Maximally Stable Extremal Regions Algorithm puted. Posterior probabilities of component mem-
(MSER) algorithm [40] includes this information in the bership calculated in the previous step are used as
segmentation procedure thus, areas of the image whose weights.
pixels have similar intensity levels but are not connected
(are separated by pixels with different intensity levels) 2.6. Background on Wiener-Granger causality
are considered different regions.
Wiener-Granger causality (a.k.a. G-causality) is a sta-
The algorithm was designed to address the stereo
tistical test intended to analyze the relationship between
problem in image analysis, i.e. the problem of estab-
two time series. Originally introduced by Wiener [46]
lishing the correspondence between several images of
and subsequently adapted to the analysis of economic
the same scene taken from different viewpoints, but it
studies by Granger [47], G-causality allows us to de-
can be used for other image analyses as tracking color
termine weather a time series can be used to forecast
objects [43] or detecting regions based on color [44]. another time series or, more precisely, weather a time
The general idea of MSER can be explained as fol- series X, conditional on its own past, also depend on
lows: Let us imagine a grayscale image as a 3D struc- the past of another time series Y . In such case, it is said
ture in which each point has a height corresponding Y G-cause X [37,48].
to its intensity level. If we drop a fluid on a point, it The conditional dependence of one time series on
will spread out, occupying a wider and wider area with another can also be quantified by a statistic, F, which
the points (pixels) that have less height (intensity). If measures how much better the adjustment (lower error
this process is done simultaneously at different points, variance) of the conditional model is compared to the
regions will appear and as the fluid level rises, they non-conditional model. Let’s suppose two time series,
will be connected to each other. The process ends when X and Y , can be described as follows:
the stability criterion is reached, which is continuously X p
checked as the fluid falls. X(t) = A11,k X(t − k) +
k=1
2.5. Gaussian mixture models p
X
A12,k Y (t − k) + X (t) (3)
Gaussian mixture models (GMMs) are probabilistic k=1
models that allow the identification of subpopulations p
X
within the same heterogeneous population [45]. They Y (t) = A21,k X(t − k) +
use the density estimation of the data set to define sub- k=1
sets and thus delimit several groups. GMM algorithms p
X
suppose that the data, x = x1 , x2 , . . . , xN are sam- A22,k Y (t − k) + Y (t) (4)
ples drawn from a probability distribution p(x) that is k=1
360 F. Segovia et al. / Connected system for monitoring electrical power transformers using thermal imaging

Fig. 6. Top row: Reference images obtained by computing respectively the minimum, mean and maximum (pixel wise) of 1440 thermal images
captured along 24 hours at a rate of 1 per minute. Bottom row: Histogram of images in the top row.

where Aij,k are the regression coefficients, p is the the thermal cameras, which was carried out by setting
model order (maximum number of previous observa- the factory calibration through the library provided by
tions, which implies p < T ) and X and Y are the the manufacturer to configure and control the device.
prediction errors. The dependence of X on Y can be In our case, this factory calibration was adequate since
expressed as: we did not focus on absolute temperature values, but
ΣX on temperature changes. Beside calibration, some envi-
FY →X = ln (5) ronmental factors, as camera positioning, can affect the
ΣY →X
temperature measurements. This influence is small (less
where ΣY →X is the residual covariance matrix defined than 1% at distances between 0.5 and 2.5 meters) for the
as: camera model used, however, in order to avoid differ-
ΣY →X = Cov(X ) (6) ences between the operating temperatures assumed to
be normal from one transformer substation to another,
and ΣX is the residual covariance matrix for the as- all the cameras were installed at a distance of approxi-
sumption that X does not depend on the past of Y . It mately 2 meters from the surface of the transformer.
can be derived from Eq. (3) omitting the A12,k coeffi- As mentioned in Section 2.4 the image processing
cients. carried out to extract two temperatures from each ther-
mal image required a previous stage of segmentation of
a reference image representing the temperature profile
3. Experiments and results of each transformer. In order to build this image, 1440
thermal images taken over 24 consecutive hours were
The proposed thermal imaging-based monitoring sys- used, and three aggregation measures (mean, minimum
tem was installed at 14 transformer stations of the Span- and maximum) were considered. This way, the image is
ish power grid. They were configured to capture ther- independent of the time of day and the whole method is
mal data at a rate of one image per minute and data was able to deal with images captured at any time. Figure 6
stored and processed on a remote server as described compares the result of these three options. Note that all
in Sections 2.2 and 2.3. All equipment was configured the three have a similar temperature profile and can be
in our laboratory prior to installation in the transformer used indifferently for this purpose. For simplicity we
centers. This configuration included the calibration of selected the last one, where each pixel was calculated
 F. Segovia et al. / Connected system for monitoring electrical power transformers using thermal imaging 361

Fig. 7. Variation of the average temperature measured from thermal images throughout the day.

Fig. 8. Segmentation procedure based on SLIC superpixel, Otsu’s and MSER methods

as the maximum number of pixels in the same position – First, the image was downsampled using the SLIC
in all images. superpixel method proposed by Achanta et al. [38].
The parameter k was set to 500.
Two segmentation approaches were evaluated. The
– Then, the resulted image was requantized using
first one, based on Otsu’s method for thresholds selec- several thresholds (multiclass version) computed
tion, carried out the segmentation into three step: according to [39].
362 F. Segovia et al. / Connected system for monitoring electrical power transformers using thermal imaging

Fig. 9. Intensity profile of the reference image computed as the maxi- Fig. 10. Result of segmenting the reference image computed as the
mum. maximum using a GMM model with 3 Gaussians.

– Finally, spatially separated regions of equal in- approaches for image segmentation produced almost
tensity are considered independently by apply- identical time series of temperatures.
ing the Maximally Stable Extremal Regions algo- The evaluation of these systems was carried out by
rithm [40]. comparing the temperatures obtained by them with
those obtained by the more standard approach, based
The intermediate results of these steps are shown in
on sensors. Figure 11 shows a comparison between the
Fig. 8. Note how, after the last step (MSER applica-
temperatures registered by the proposed system and
tion), four clearly differentiated regions appear. The
those obtained by sensors. Note that, despite the offset
two with the highest average intensity correspond to
probably due to differences between the calibration of
the transformer and were used to select the transformer the sensors and the thermal cameras, there is an almost
pixels while the one with the lowest average intensity perfect correlation between the temperatures recorded
corresponds to the protective fence used to estimate the by the two systems. This correlation can be easily seen
room temperature. if we remove the offset, as is done in Fig. 12, where
The second segmentation approach proposed in this the temperature obtained from the thermal camera was
work modeled the reference image with a mixture of shifted so that it has the same mean as the temperature
Gaussians and then identified each region with a Gaus- obtained from sensors.
sian. After obtaining the Gaussian mixture parameters In order to further investigate the relationship be-
(see Eq. (2)) using a Expectation-Maximization algo- tween the temperatures obtained by the standard and the
rithm, the MSER algorithm was also used to separate proposed monitoring systems, and even the relationship
into regions areas of equal intensity located at distant between temperatures and other operating variables, we
positions. Given the shape of the intensity profile of the performed an additional analysis based on G-causality.
reference image (see Fig. 9), a model with 3 Gaussians The results are shown in Figs 13 and 14. The former
was used to fit the image. This would ideally associate figure represents the F statistic for some variables of
one component with pixels of higher intensity (located interest, including temperatures computed using the
on the surface of the power transformer), another with standard and proposed monitoring systems. It was com-
pixels of lower intensity (representing the protection puted using the expression (5). Figure 14 shows a cir-
fence) and a third component with the rest of the pixels cular graph with stronger connections among the same
having an intermediate temperature. In practice, some variables.
pixels with a temperature similar to that of the protec-
tion fence were associated with the same component as
the fence, as shown in Fig. 10. In order to isolate the 4. Discussion
pixels of the fence, we applied the MSER algorithm
as in the alternative approach based on SLIC super- Temperature control plays a fundamental role in the
pixels and Otsu’s methods. In our experiments, both ability of electrical power transformers to operate fault-
 F. Segovia et al. / Connected system for monitoring electrical power transformers using thermal imaging 363

Fig. 11. Comparison of transformer and room temperatures obtained by the proposed system and those obtained by sensor based approach.

Fig. 12. Transformer temperature obtained by the proposed system based on thermal cameras (red) and by sensors (blue). The former temperature
has been shifted so that both have the same mean and in order to the correlation between them can be easily appreciated.

lessly [49]. Room and transformer temperatures are a function of past values of outputs and exogenous in-
variables of great relevance in the analysis of trans- puts. This scenario is also supported by the result of
former operation and has a clear relationship with other the analysis based on G-causality carried out in this
electrical variables. In [37], a methodology based on work (figures 13 and 14). This analysis, similar to the
Granger causality is used to study the ability of some one carried out in [37], but including the temperatures
electrical variables to predict others. The results show obtained by the thermal imaging-based system, meets
that there is a clear relationship between the transformer two objectives: i) to confirm the dependence of some
temperature and other relevant variables such as cur- electrical variables on temperature and ii) to corrobo-
rent, active power, reactive power and voltage. The last rate the relationship between the temperatures obtained
experiment carried out in this work has shown that by the proposed system and the sensor-based system.
this relationship between the temperatures obtained by The experiments carried out demonstrate that it is
the proposed system and other electrical variables is possible to obtain reliable information on the operating
also present. The relationship between transformer tem- temperature of an electrical power transformer using a
perature and active power has also been corroborated system based on thermal imaging. It could improve the
in [50], which shows that it is possible to predict the temperature monitoring of electrical power transform-
transformer temperature in the near future from the ac- ers and, that way, reduce the probability of errors that
tive power of the three transformer lines together with affect the quality of the power supply and shorten the
the ambient temperature using long short-term memory transformer lifetime.
(LSTM) networks. A similar approach was also demon- After data acquisition, thermal images were ana-
strated in [28]. In this case, a non-linear autoregressive lyzed in order to extract room and transformer tempera-
neural network with exogenous inputs (NARX) was tures. This analysis included the image segmentation by
used to predict the temperature of the transformer as means for classical algorithms, however the use of these
364 F. Segovia et al. / Connected system for monitoring electrical power transformers using thermal imaging

Fig. 13. F matrix based on conditional G-causality (see Eq. (5)) and containing a measure of the dependence of some electrical variables on others.

and time-consuming). In addition, it has all the advan-

tages of eliminating the human factor, as subjectivity
and error-proneness [51,52]. In this work two proposals
based on classical methods for automatic segmentation
of the thermal images have been presented. Both pro-
vided a correct separation of the regions that allow us to
calculate the transformer and room temperatures. Thus,
there was no reason to opt for a manual segmentation.
Figures 11 and 12 show that the thermal camera-
based system provides accurate temperatures, similar
to the those obtained by the standard sensor-based sys-
tem. A close examination of the figures shows, how-
ever, that there are some small differences, beyond the
offset (possibly due to calibration differences). On the
one hand, the data from the sensors have lower tem-
poral resolution, which is closely related to sampling
frequency of each system. Moreover, temperature peaks
Fig. 14. Circular graph with the stronger relations between electrical are slightly higher in sensor data probably due to sensor
variables. inertia. Larger differences at specific instants may be
due to inhomogeneous heat distribution that causes heat
algorithms is not strictly necessarily. Given the fixed to be detected earlier by one system than by another.
position of the thermal cameras and the limited number On the other hand, differences in room temperature are
of locations where they have been installed, a manual more noticeable. These are due to the malfunctioning
segmentation would have been feasible. However, the of the room temperature sensors, which were blocked
automation of this task has the advantage of facilitating for hours at a time. In any case, the proposed thermal
the extension of the proposed monitoring system to the imaging-based approach provides some interesting ad-
entire power grid (manual segmentation can be tedious vantages:
 F. Segovia et al. / Connected system for monitoring electrical power transformers using thermal imaging 365

– It has been set up as an external system that is not being monitored. A communication system is therefore
coupled to the transformer, i.e. there is no physical required between the acquisition equipment and the
contact between the devices of both systems and servers external to the transformer. Possibly, the most
no data transfer through the electrical network, direct application is to use the power grid itself to trans-
which guarantees that errors in the transformer mit the monitoring data, but this makes the monitoring
will not affect the monitoring system. system vulnerable to failures in the distribution net-
– Furthermore, the proposed approach is easily scal- work and its implementation entails substantial costs.
able to obtain a higher temperature granularity, In order for the monitoring system to be completely
i.e. to obtain not only the room and transformer autonomous from the power grid, it is necessary that
temperatures but also those of other elements such the data transmission is carried out independently from
as the connection terminals or the input and output the grid. However, implementing an independent data
lines. This would only require changing the image network from scratch (whether wired or wireless), in-
processing and eventually the camera orientation, volves high costs. In the proposed approach, a commer-
but no hardware changes (additional cameras, etc.) cial mobile phone network was used for this purpose.
would be necessary. This allows significantly lower implementation costs by
– Finally, the proposed system based on thermal using already deployed infrastructure, while allowing
imaging can also serve as a security system since data transmission to be carried out independently of
it allows monitoring the transformer environment. the power grid. A similar approach was previously de-
Thus, for example, it would be possible to detect scribed in [53]. Unlike the approach proposed by Jalil-
temperature changes in devices close to the trans- ian et al. the communication system proposed in this
former but external to it, which could end up af- paper is unidirectional. The transmission is made only
fecting the transformer’s operation. Furthermore, from the equipment installed next to the transformer to
it would be possible to detect the presence of per- the external database. This facilitates the system to have
sons attempting to tamper with the transformer. a certain resilience since synchronous communication
The comparison between the temperatures obtained is not required and in case of connection error the data
by sensors and those obtained by the proposed system are temporarily stored in a buffer memory to be sent
(Figs 11 and 12) also shows the differences between the later in a new connection attempt.
sampling frequencies of each system, which is much The main limitations of this work can be divided
higher in the case of the thermal imaging-based system. into two aspects. On the one hand, the proposed system
This difference may lead to lower reaction times of one suffers from the limitations implicit in the collection
system versus the other, however, it is important to note equipment used, i.e., thermal cameras. Therefore, it is
that it is not so much due to the technology itself, but not possible to directly obtain other variables of inter-
rather, to the configuration of the system. The reaction est, such as input or output line currents or voltages. Al-
time of the system to a failure resulting in overheating though, as mentioned above, it is theoretically possible
also depends on the element in which the failure occurs. to estimate these variables from the temperatures, this
Thus, an increase in temperature of an element that is estimation would not be free of error and would pos-
not visible from the position of the thermal camera will sibly not be as accurate as the direct measurement by
not be detected by the camera-based system until the means of specific devices. In addition, thermal camera
heat has been transferred to the air or other elements that are substantially more expensive than the sensors used
can be captured by the camera. Similarly, a temperature in the standard approach. However, despite the high
sensor immersed in the oil will not be able to detect a value of the thermal cameras, their price is insignificant
temperature rise in the windings until that temperature compared to that of the monitored equipment. More-
rise has caused a temperature rise in the oil as well. over, a malfunction of the equipment can result in in-
For this reason, the safest and most reliable monitoring calculable damage (from hundreds of citizens deprived
strategy would be one that includes sensors of different of a basic good such as electricity to a fire in an urban
types, including thermal cameras. area with unpredictable consequences), which justifies
Regardless of the system used for data acquisition, any investment in the prevention of these events.
data must be permanently stored in a database that al- On the other hand, the use of commercial mobile net-
lows later analyses [24]. Ideally this permanent storage works for data transmission between the AE equipment
is carried out externally to the transformer, i.e. in lo- and the SPS can be a concern when installing the moni-
cations other than the transformer whose operation is toring system in power transformers placed in locations
366 F. Segovia et al. / Connected system for monitoring electrical power transformers using thermal imaging

where there is little or no mobile network coverage. References

This can be stated at least for the climatic conditions
of the locations in which our experiments were carried
out, i.e., average temperatures between 9 (night) and [1] Mullen S, Onsongo G. Decentralized agent-based underfre-
quency load shedding. Integrated Computer-Aided Engineer-
30 (day) degrees Celsius, as is typical in autumn on the ing. 2010; 17(4): 321-329.
southern coast of Spain. Although additional experi- [2] Santofimia MJ, del Toro Garcia X, Villanueva FJ, Villa D,
ments in other regions with different climates would Escolar S, Lopez JC. A comprehensive common-sense-based
be desirable, the results shown can be, in our opinion, architecture for understanding voltage-sag events in electrical
grids. Integrated Computer-Aided Engineering. 2018; 25(4):
reliably extrapolated to other regions with similar or 397-416.
colder climates. [3] Zhang Z, Guo M, Fu D, Mo L, Zhang S. Traffic signal opti-
mization for partially observable traffic system and low pene-
tration rate of connected vehicles. Computer-Aided Civil and
Infrastructure Engineering. 2022; 37(15): 2070-2092.
5. Conclusions [4] Shi H, Zhou Y, Wang X, Fu S, Gong S, Ran B. A deep re-
inforcement learning-based distributed connected automated
In this work we have described a system for monitor- vehicle control under communication failure. Computer-Aided
ing electrical power transformers using thermal cam- Civil and Infrastructure Engineering. 2022; 37(15): 2033-2051.
[5] Ding C, Dai R, Fan Y, Zhang Z, Wu X. Collaborative control
eras. The proposed system allows the monitoring of of traffic signal and variable guiding lane for isolated inter-
transformer temperatures, which are critical variables section under connected and automated vehicle environment.
to anticipate potential failures in the system operation. Computer-Aided Civil and Infrastructure Engineering. 2022;
Using thermal cameras, we collected information on 37(15): 2052-2069.
[6] Chen S, Dong J, Ha PYJ, Li Y, Labi S. Graph neural network
the temperature of the transformer and its surroundings
and reinforcement learning for multi-agent cooperative control
at a rate of 1 image per minute. Subsequently, by ap- of connected autonomous vehicles. Computer-Aided Civil and
propriate thermal image processing, we obtained two Infrastructure Engineering. 2021; 36(7): 838-857.
temperature values from each image, representing the [7] Sovetkin E, Steland A. Automatic processing and solar cell de-
tection in photovoltaic electroluminescence images. Integrated
room and transformer temperature, and resulting in two
Computer-Aided Engineering. 2019; 26(2): 123-137.
temperature time series (after processing of a series of [8] Foros J, Istad M. Health index, risk and remaining lifetime
images). estimation of power transformers. IEEE Transactions on Power
This approach has been implemented and evaluated Delivery. 2020; 35(6): 2612-2620.
[9] Martinez-Monseco FJ. An Approach to a Practical Optimiza-
in 14 transformer stations belonging to the Spanish
tion of Reliability Centered Maintenance. Case Study: Power
electricity system. The experiments carried out showed Transformer in Hydro Power Plant. Journal of Applied Re-
that the temperatures obtained by the proposed system search in Technology & Engineering. 2020; 1(1): 37-47.
are as good as those obtained by widely used sensor- [10] Li J, Jiao J, Tang Y. Analysis of the impact of policies inter-
vention on electric vehicles adoption considering information
based systems. In addition, the system based on thermal
transmission – based on consumer network model. Energy
imaging provides some advantages such as greater reli- Policy. 2020; 144: 111560.
ability as it is a system totally independent of the trans- [11] Crozier C, Morstyn T, McCulloch M. The opportunity for
formation system and therefore not affected by the pos- smart charging to mitigate the impact of electric vehicles on
transmission and distribution systems. Applied Energy. 2020;
sible errors of the latter, as well as certain surveillance
268: 114973.
capabilities. [12] Elsisi M, Tran MQ, Mahmoud K, Mansour DEA, Lehtonen
As future work, we plan to develop time series analy- M, Darwish MMF. Effective IoT-based deep learning platform
sis methods to predict possible anomalies in transformer for online fault diagnosis of power transformers against cy-
operation from the registered temperature series. berattacks and data uncertainties. Measurement. 2022; 190:
110686.
[13] Müllerová E, Hrůza J, Velek J, Ullman I, Stˇríska F. Life cycle
management of power transformers: results and discussion of
Acknowledgments case studies. IEEE Transactions on Dielectrics and Electrical
Insulation. 2015; 22(4): 2379-2389.
[14] Vītolin, a S. Development of lifetime data management algo-
This work was supported by the Consejería de rithm for power transformers. 2014 5th International Confer-
Economía, Innovación, Ciencia y Empleo (Junta de ence on Intelligent Systems, Modelling and Simulation. 2014.
Andalucía) and FEDER under B-TIC-586-UGR20 and [15] Yazdani-Asrami M, Taghipour-Gorjikolaie M, Moham-
P20-00525 projects and by the University of Granada mad Razavi S, Asghar Gholamian S. A novel intelligent protec-
tion system for power transformers considering possible elec-
and Endesa Distribución under the PASTORA (ref. trical faults, inrush current, CT saturation and over-excitation.
EXP – 00111351/ITC-20181102) and RESISTO (ref. International Journal of Electrical Power & Energy Systems.
2021/C005/00144188) contracts. 2015; 64: 1129-1140.
 F. Segovia et al. / Connected system for monitoring electrical power transformers using thermal imaging 367

[16] Dolata B, Coenen S. Online condition monitoring becomes [32] Lu L, Dai F, Zaniewski JP. Automatic roller path tracking and
standard configuration of transformers – practical application mapping for pavement compaction using infrared thermog-
for optimized operation, maintenance and to avoid failures. In: raphy. Computer-Aided Civil and Infrastructure Engineering.
E-ARWtr2016 Transformers, Advanced Research Workshop 2021; 36(11): 1416-1434.
on Transformers. vol. 2. La Toja Island, Spain; 2016. [33] Bagavathiappan S, Saravanan T, George NP, Philip J, Jayaku-
[17] Barkas DA, Chronis I, Psomopoulos C. Failure mapping and mar T, Raj B. Condition monitoring of exhaust system blowers
critical measurements for the operating condition assessment using infrared thermography. Insight: Non-Destructive Testing
of power transformers. Energy Reports. 2022; 8: 527-547. and Condition Monitoring. 2008; 50(9): 512-515.
[18] Setayeshmehr A, Akbari A, Borsi H, Gockenbach E. A proce- [34] Itami K, Sugie T, Vayakis G, Walker C. Multiplexing thermog-
dure for diagnosis and condition based maintenance for power raphy for international thermonuclear experimental reactor di-
transformers. In: Conference Record of the 2004 IEEE Interna- vertor targets. Review of Scientific Instruments. 2004; 75(10
tional Symposium on Electrical Insulation; 2004. p. 504-507. II): 4124-4128.
[19] AJ C, Salam MA, Rahman QM, Wen F, Ang SP, Voon W. [35] Zarco-Periñán PJ, Martínez-Ramos JL, Zarco-Soto FJ. A novel
Causes of transformer failures and diagnostic methods – a method to correct temperature problems revealed by infrared
review. Renewable and Sustainable Energy Reviews. 2018; 82: thermography in electrical substations. Infrared Physics &
1442-1456. Technology. 2021; 113: 103623.
[20] Moravej Z, Abdoos AA. An improved fault detection scheme [36] Segovia F, Ramírez J, Salas-Gonzalez D, Illán IA, Martinez-
for power transformer protection. Electric Power Components Murcia FJ, Rodriguez-Rivero J, et al. Temperature control
and Systems. 2012; 40(10): 1183-1207. and monitoring system for electrical power transformers using
[21] Haghjoo F, Mostafaei M. Flux-based method to diagnose and thermal imaging. In: Ferrández Vicente JM, Álvarez-Sánchez
identify the location of turn-to-turn faults in transformers. IET JR, de la Paz López F, Adeli H, editors. Bio-Inspired Systems
Generation, Transmission and Distribution. 2016; 10(4): 1083- and Applications: From Robotics to Ambient Intelligence. Lec-
1091. ture Notes in Computer Science. Cham: Springer International
[22] Abbasi AR. Fault detection and diagnosis in power transform- Publishing; 2022. p. 573-582.
ers: A comprehensive review and classification of publications [37] Rodriguez-Rivero J, Ramirez J, Martínez-Murcia FJ, Segovia
and methods. Electric Power Systems Research. 2022; 209: F, Ortiz A, Salas D, et al. Granger causality-based informa-
107990. tion fusion applied to electrical measurements from power
[23] Bakar NA, Abu-Siada A, Islam S. A review of dissolved gas transformers. Information Fusion. 2020; 57: 59-70.
analysis measurement and interpretation techniques. IEEE [38] Achanta R, Shaji A, Smith K, Lucchi A, Fua P, Süsstrunk
Electrical Insulation Magazine. 2014; 30(3): 39-49. S. SLIC superpixels compared to state-of-the-art superpixel
[24] Kunicki M, Borucki S, Zmarzły D, Frymus J. Data acquisition methods. IEEE Transactions on Pattern Analysis and Machine
system for on-line temperature monitoring in power transform- Intelligence. 2012; 34(11): 2274-2282.
ers. Measurement. 2020; 161: 107909. [39] Otsu N. A threshold selection method from gray-level his-
[25] Peimankar A, Weddell SJ, Jalal T, Lapthorn AC. Evolutionary tograms. IEEE Transactions on Systems, Man, and Cybernet-
multi-objective fault diagnosis of power transformers. Swarm ics. 1979; 9(1): 62-66.
and Evolutionary Computation. 2017; 36: 62-75. [40] Matas J, Chum O, Urban M, Pajdla T. Robust wide-baseline
[26] Velasquez-Contreras JL, Sanz-Bobi MA, Galceran Arellano stereo from maximally stable extremal regions. Image and
S. General asset management model in the context of an elec- Vision Computing. 2004; 22(10): 761-767.
tric utility: Application to power transformers. Electric Power [41] Hartigan JA, Wong MA. A K-means clustering algorithm.
Systems Research. 2011; 81(11): 2015-2037. Journal of the Royal Statistical Society Series C (Applied
[27] Górriz JM, Ramírez J, Ortíz A, Martínez-Murcia FJ, Segovia Statistics). 1979; 28(1): 100-108.
F, Suckling J, et al. Artificial intelligence within the interplay [42] Ćurković M, Ćurković A, Vučina D. Image binarization
between natural and artificial computation: Advances in data method for markers tracking in extreme light conditions. Inte-
science, trends and applications. Neurocomputing. 2020; 410: grated Computer-Aided Engineering. 2022; 29(2): 175-188.
237-270. [43] Donoser M, Bischof H. Efficient maximally stable extremal
[28] Ramírez J, Martínez-Murcia FJ, Segovia F, Ortiz A, Salas- region (MSER) tracking. In: 2006 IEEE Computer Society
González D, Carrillo S, et al. Power transformer forecasting Conference on Computer Vision and Pattern Recognition
in smart grids using NARX neural networks. In: Valenzuela (CVPR’06). vol. 1; 2006. p. 553-560.
O, Rojas F, Herrera LJ, Pomares H, Rojas I, editors. Theory [44] Chavez A, Gustafson D. Color-based extensions to MSERs. In:
and Applications of Time Series Analysis. Contributions to Bebis G, Boyle R, Parvin B, Koracin D, Wang S, Kyungnam K,
Statistics. Granada, Spain: Springer International Publishing; et al., editors. Advances in Visual Computing. Lecture Notes
2020. p. 401-414. in Computer Science. Berlin, Heidelberg: Springer; 2011. p.
[29] de Melo AS, Calil WV, Paz Salazar PD, Liboni LHB, Costa 358-366.
ECM, Flauzino RA. Applied methodology for temperature nu- [45] McLachlan GJ, Peel D. Finite mixture models. New York;
merical evaluation on high current leads in power transformers. 2000.
International Journal of Electrical Power & Energy Systems. [46] Wiener N. The theory of prediction. In: Modern Mathematics
2021; 131: 107014. for Engineers. New York: McGraw-Hill; 1956. p. 165-190.
[30] Mariprasath T, Kirubakaran V. A real time study on condition [47] Granger CWJ. Investigating causal relations by economet-
monitoring of distribution transformer using thermal imager. ric models and cross-spectral methods. Econometrica. 1969;
Infrared Physics & Technology. 2018; 90: 78-86. 37(3): 424-438.
[31] Sirca GF, Adeli H. Infrared thermography for detecting de- [48] Seth AK. A MATLAB Toolbox for granger causal connectivity
fects in concrete structures. Journal of Civil Engineering and analysis. Journal of Neuroscience Methods. 2010; 186(2): 262-
Management. 2018; 24(7): 508-515. 273.
368 F. Segovia et al. / Connected system for monitoring electrical power transformers using thermal imaging

[49] Karakoulidis K, Fantidis J, Kontakos V. The temperature mea- [52] García-Aguilar I, García-González J, Luque-Baena RM,
surement in a three-phase power transformer under different López-Rubio E, Domínguez E. Optimized instance segmenta-
conditions. Journal of Engineering Science and Technology tion by super-resolution and maximal clique generation. Inte-
Review. 2015; 8: 19-23. grated Computer-Aided Engineering. 2023; Preprint(Preprint):
[50] Martinez-Murcia FJ, Ramirez J, Segovia Román F, Ortiz A, 1-14.
Carrillo S, Leiva FJ, et al. Prediction of transformer tempera- [53] Jalilian M, Sariri H, Parandin F, Karkhanehchi MM, Hookari
ture for energy distribution smart grids using recursive neural M, Jirdehi MA, et al. Design and implementation of the moni-
networks. In: International Conference on Time Series and toring and control systems for distribution transformer by us-
Forecasting. Granada, Spain; 2019. ing GSM network. International Journal of Electrical Power &
[51] Lin J, Ma L, Yao Y. A spectrum-domain instance segmen- Energy Systems. 2016; 74: 36-41.
tation model for casting defects. Integrated Computer-Aided
Engineering. 2022; 29(1): 63-82.