Electrical Engineering (2024) 106:2691–2703

https://doi.org/10.1007/s00202-023-02109-6

ORIGINAL PAPER

Machine learning for power generator condition assessment

Denis Ilić1 · Mileta Žarković2

Received: 7 September 2023 / Accepted: 25 October 2023 / Published online: 10 November 2023
© The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature 2023

Abstract
The digitalization of classic power systems enables the monitoring of diverse processes and occurrences in power electric
systems and big data collection. Manipulating data is a key role in decision-making and planning, and machine learning
(ML) algorithms can mimic human-like reasoning and automate decision processes. This manuscript centers on applying
ML algorithms to create maintenance schedules for power generators, primarily based on stator winding testing. The inves-
tigation showcases the implementation of multiple ML algorithms, encompassing fuzzy logic (FL), adaptive neuro-fuzzy
inference system (ANFIS), artificial neural network (ANN), autoencoder neural network (AENN), and decision tree (DT).
This research’s primary objective is to investigate and determine the most acceptable ML algorithm that fits the most with
the assessments made by human experts. Based on real data from power plants of Serbia best prediction correlation to the
human expert shows DT algorithms with 69.45%, then ANN 68.05%, ANFIS 59.56%, FL 47.81% and at the end AENN
6.83%. The presented ML algorithms results with risk maps prove to be a valuable instrument for condition and risk-based
maintenance. The presented methodology for decision making can be used to optimize power plants maintenance schedules,
minimize downtime, and maximize operational efficiency in one power system. This investigation contributes to the progress
of the electrical power industry by showcasing the capabilities of ML in facilitating informed decision-making processes and
enabling proactive maintenance strategies.

Keywords Asset management · Decision making · Diagnostic · Machine learning · Power generators

Abbreviations List of symbols

ML Machine learning RisoR Rotor winding insulation resistance

AI Artificial intelligence Te Age of the stator winding
FL Fuzzy logic N1 Polarization index
FLES Fuzzy logic expert system Riso Stator winding insulation resistance
ANFIS Adaptive neuro-fuzzy inference system PD Partial discharge
ANN Artificial neural network tgδ 0.2 Tgδ at 0.2Un test voltage
AENN Autoencoder neural network tgδ maxtip-up A maximum tgδ tip-up, per phase and per volt-
DT Decision tree age 0.2Un step
CBM Condition-based maintenance
RBM Risk-based maintenance
EIS Electrical insulation system

B Mileta Žarković
[email protected]
1 Introduction
Denis Ilić
[email protected]
Synchronous power generators are the base elements and
1 Faculty of Diplomacy and Security Belgrade, University of pillars of modern electrical power systems. Appropriate
Belgrade, Travnicka 2, 11000 Belgrade, Serbia maintenance of these machines is of great importance for the
2 School of Electrical Engineering, University of Belgrade, overall reliability of the power system and consequently for
Bulevar Kralja Aleksandra 73, 11020 Belgrade, Serbia the economic stability of an electrical power company [1–4].

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

2692 Electrical Engineering (2024) 106:2691–2703

The first model of electrical equipment maintenance devel- distribution grids can be used convolutional neural network
oped during the electrification period considered emergency [10] and FL [11]. ML techniques give very good results when
maintenance, dealing with equipment maintenance mainly if looking at one model of equipment or one device through his-
equipment failure has been detected. Over time, using a grow- torical tests and monitoring. The difficulties that this paper
ing experience, fixed time intervals have been determined for deals with arise precisely in the phase of interpreting the
specific equipment maintenance. This model is considered an results of performed electrical prophylactic tests. Namely, the
improvement, since, on some level, usual breakdowns were existing international standards largely define the procedures
detected prior to their occurrence. Lately, various optimized for performing certain measurements and define measuring
models of equipment maintenance evolved, supported by the instruments. However, the criteria for individual methods are
development of methods for continuous (online) and offline usually not presented. Due to the diversity of applied tech-
testing. These tests provide insight into machine condition, nologies of different manufacturers, as well as the progress
which has been considered the most valuable information. in the development of Electrical Insulation System (EIS),
This makes transition to CBM—condition-based mainte- the standards often do not define universal criteria. Even if
nance model possible. This model provides several technical criteria were present, it is always challenging to distinguish
and economic benefits, as limited maintenance assets can be unique conclusion from several test results. This paper ana-
organized only on machines that have had need for it. For lyzes the possibilities of modern artificial intelligence (AI)
the maintenance department to be able to implement CBM models for assessing the state of EIS synchronous generators
model, it is necessary to have input data which describes and data processing automation. Contributions of this paper
reliably, repeatedly, and objectively actual condition of the can be remarked as follows:
equipment.
Various approaches to electrical equipment maintenance
• Temperature, electrical, ambient, and mechanical aging
were developed through the development of the electric
factor of EIS is included with creation of monitoring
power industry. The first model of electrical equipment
database for generators in different operational conditions
maintenance developed during the electrification period con-
and lifetime.
sidered emergency maintenance, dealing with equipment
• Implementation and comparison of different ML algo-
maintenance mainly if equipment failure has been detected.
rithms (FL, ANFIS, ANN, DT, AENN) on real database
Over time, using a growing experience, fixed time intervals
from power plants of Serbia,
have been determined for specific equipment maintenance.
• Investigated correlation between the applied ML algo-
This model is considered as an improvement, since, on some
rithms and the assessments made by human experts,
level, usual breakdowns were detected prior to its occurrence.
• Incorporation of the thinking and experience of human
Lately, various optimized models of equipment maintenance
experts in decision making about EIS through the imple-
evolved, supported by the development of methods for con-
mentation of AI in the CBM methodology
tinuous (on-line) and offline testing. These tests provide
• ML algorithms results are used for a more reliable and
insight into machine condition, which has been considered
informative CBM and RBM in power plants to optimize
the most valuable information. This makes the transition
their maintenance schedules, minimizing downtime, and
to CBM—Condition Based Maintenance model possible.
maximizing operational efficiency.
This model provides several technical and economic ben-
• Investigation contributes to the progress of the electri-
efits, as limited maintenance assets can be organized only
cal power industry by showcasing the capabilities of ML
on machines that have had need for it. For the maintenance
in facilitating informed decision-making processes and
department to be able to implement CBM model, it is neces-
enabling proactive maintenance strategies.
sary to have input data which describes reliably, repeatedly,
and objectively actual condition of the equipment.
This paper proposes the implementation of machine learn- The paper is organized as follows. Section 2 presents
ing/ artificial intelligence to automate the processing of the facts for synchronous generator insulating system degrada-
raw test data and provide unique values which describe the tion and provides an overview usual monitoring tests in power
condition of the insulation system of the synchronous gener- plants. In Sect. 3, the motivation for the introduction of ML
ator. Optimized ML classification techniques is used for PD is elaborated. The subsections in Sect. 3 described how to
diagnosis in paper [5]. Papers [6, 7] use unsupervised ML implement different ML algorithms (FL, ANFIS, ANN, DT,
algorithms, K-mean and DT to classify the rows of mon- AENN). The results obtained using the previously described
itoring data of SF6 circuit breakers. To evaluate health of algorithms are compared and discussed in Sect. 4. Imple-
power transformer and do proper diagnostic, paper [8] uses mentation of presented ML results in maintenance strategy is
deep learning neural network and paper [7] analyzes differ- presented in Sect. 5. Conclusions of the research are derived
ent ways of using FL. To detect and find fault location in in Sect. 6.

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Electrical Engineering (2024) 106:2691–2703 2693

2 Synchronous generator insulating system with Hodrick–Prescott filter [19] and convolutional neural
degradation networks [20] are used. Aging and damage of equipment
in the power systems is caused by various effects, thermic,
During their working life, synchronous generators are mechanical, and electrical, that cannot be directly measured
exposed to the so-called. TEAM factor, the combined influ- together. For this reason, such a multidisciplinary critical
ence of Temperature, Electrical, Ambient, and Mechanical problem should be solved by applying AI/ML models that
aging factor of EIS [12, 13]. Isolation systems are a fairly can draw a conclusion based on a large number of measured
common problem and the reason for the failure of syn- different quantities.
chronous generators [13–16]. For example, according to a
study by the CIGRE International Committee [14], problems
related to EIS hydro generators account for 56% of the total 3 Creating ML models
number of outages, while according to another source [13],
this share of turbogenerators is lower, 23%. AI is widely used today in many home and industry appli-
Since the degradation of EIS in operating conditions is cations. It is a task of mimic human reasoning on machines.
practically inevitable, it is extremely important to recognize AI includes specific applications of expert systems, natural
timely accelerated degradation processes and prevent sudden language processing, speech recognition and machine vision.
outages. To diagnose these processes, electrical tests are per- ML is a specific subset of AI, which encapsulates techniques
formed. Also, monitoring the data of the installed devices for for training and creating ML models. It can be classified as
continuous monitoring is possible, but these devices are not supervised, unsupervised and reinforcement learning. Super-
so common in the powerplants in Serbia’s electrical power- vised ML is the most utilized machine learning. This type of
plants. machine learning feeds historical input and output data in ML
To detect these degradation processes and prevent out- algorithms, with processing in between each input/output
ages, different test methods are in routine practice by power pair that allows the algorithm to shift the model to create
companies in Serbia: outputs as closely aligned with the desired result as possible.
Common algorithms used during supervised learning include
1. Measuring the electrical insulation resistance, neural networks, decision trees, linear regression, and sup-
2. Measuring the dielectric loss and capacity of winding port vector machines. Unsupervised learning does not use
insulation, the same labeled training sets and data as supervised learn-
3. Measuring the intensity of partial discharges, and ing. Instead, the machine looks for less obvious patterns in
4. Testing with increased AC and/or DC voltage the data. This machine learning type is very helpful when
you need to identify patterns and use data to make decisions.
A common problem is that valid international standards, Common algorithms used in unsupervised learning include
such as IEEE, IEC and similar, usually do not provide a set hidden Markov models, k-means, hierarchical clustering,
of criteria for specific measurements. So, providing valu- and Gaussian mixture models. Reinforcement learning is
able insights from measurement values, especially for old the closest machine learning type to how humans learn. The
machines, already in operation, can be a very challenging algorithm or agent used learns by interacting with its envi-
task. This is an assignment that can be solved by the engaging ronment and getting a positive or negative reward. Common
possibility of modern AI/ML techniques. Also, the digi- algorithms include temporal difference, deep adversarial net-
talization of power plants and installing bigger number of works, and Q-learning.
monitoring sensors will make bigger database for diagnos- Algorithms for training machine learning models are
tic. The diagnostic problem to be solved is multidimensional widely implemented in several popular software libraries and
due to the influence of a larger number of monitoring tests tools, so they can be used easily without a lot of program-
and measured quantities. Monitoring many measured quan- ming skills. Programming languages such as Python, R, and
tities and their trends will not be possible for an engineer. Matlab provide a list of user-friendly libraries for data manip-
Experts cannot perceive such a problem through the dif- ulation and ML model creation. In this paper, several models
ferent operational conditions and lifetime of the equipment, based on a different ML technique will be analyzed as a sta-
and therefore, there is a clear need for the application of the tor condition assessment tool to present an upgrade to an
AI/ML techniques. Because of that, AI/ML models are used asset management toolbox. Matlab software will be used for
in the literature to improve the reliability of the power sys- model development.
tems. Wavelet-Seq2Seq-LSTM [17] and ensemble learning Before the training model, available data are analyzed,
methods [18] are used for time series forecasting of level of and a set of features are chosen. In the original database
dams in hydroelectric power plants. For prediction of insu- taken from power system of Serbia, data are populated for
lators faults in power system, ensemble learning methods years from 1980 and on, but until around 1997 there are a lot

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

2694 Electrical Engineering (2024) 106:2691–2703

Fig. 1 Histogram for a N 1 , c (tgδ 0.2 )max and their dependence on the years of operation of the generator (b) and (d), retrospectively

of missing data since only a small scope of testing methods 6. Age, T e , of the stator winding (absolute value).
are used in the beginning of data collection. As a features for 7. Rotor winding insulation resistance, RisoR , (at 60 s).
stator condition assessment next features are derived from
test results:
These minimum and maximum values are used because
1. tgδ at 0.2Un test voltage (starting value of tg tip-up test). they represent the most critical values of parameters in the
The maximum value between phases is chosen. data set and reflect the worst possible generator state in that
2. A maximum tgδ tip-up, per phase and per voltage 0.2Un data set. In Fig. 1 are presented all measured data for input
step. parameters as histogram and as dependents of power gen-
3. Partial discharge, PD, intensity at 0.6Un, or phase volt- erator age. On the graphs, measurements indicating a good
age. The maximum value between phases is chosen. condition are marked in green, questionable measurements
4. Stator winding insulation resistance, Riso , (60 s value). A in blue, and dangerous generator conditions in red (the values
minimum value between phases is chosen. of these measurements are interpreted according to recom-
5. Polarization index N 1 as a ratio of stator winding insula- mendations and standards [14, 21–23]). It can be observed
tion resistance at 60 s and 15 s, a minimum value between that there is no clear dependence of the failure and aging of
phases is chosen. the generator insulation and that it is necessary to implement

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Electrical Engineering (2024) 106:2691–2703 2695

Fig. 2 Distribution of targets in available dataset

the knowledge of experts and apply ML techniques. By com-

mon practice, chief test engineers provide one score value to
assess the condition of the generator after test data have been
processed. These scores are given in linguistic form as good
Fig. 3 Created fuzzy logic expert system for synchronous generator
condition, suspicious, and bad condition. Dedicated corre- insulating system condition assessment
sponding numerical values are 3, 2 and 1, respectively. Those
values will be used as target values for model creation. An
example of dataset prepared for model training is presented features and targeted output values. The algorithm for appli-
in Table 1. After data cleaning, dataset used in this paper cation of FL for the purpose of diagnosing the state of energy
contains about 600 rows. equipment is explained in detail [3, 11, 13] and includes the
Data in each row is provided by one power generator elec- process of fuzzification, creating a rule base and defuzzifi-
trical testing. The available dataset is unbalanced, since it cation. Fuzzification is the coding of inputs via assignment
comprises most good condition targets. Targets distribution of the membership function (μA (x)) of fuzzy set which maps
is presented in Fig. 2. All designed models are tested on the each element of the input set A of the mentioned interval of
sample of 20 generator results which are excluded from the real numbers:
learning dataset.
Models analyzed in this paper are based on several differ- A  [x, μ A (x)|x X ] (1)
ent ML techniques [1, 9, 11, 24–26]:
Defining rule-base and the membership functions for each
1. Fuzzy logic expert system—FLES,
input and output is a time-consuming process. The created
2. Adaptive Neuro-Fuzzy Inference System—ANFIS,
model is represented in Fig. 3. Six input values are fuzzified,
3. Artificial Neural Network—ANN,
transforming these values into the fuzzy numbers, then fuzzy
4. Autoencoder Neural Network—AENN and.
rule base is created. Expert knowledge is modeled in FLES
5. Decision tree.
by 33 predefined rules. After that, center of gravity (CoG)
method for defuzzification is applied, which is the process of
obtaining a single number from the output (P) of the aggre-
3.1 Fuzzy logic
gated fuzzy set:
Fuzzy logic (FL) has been present in automation and control 
for at least half a century. By contrast to the Boolean logic μ(P) · P · dP
P  defuzzy(μ(P))   (2)
where the truth values of variables may only be the integer μ(P) · dP
values 0 or 1, FL is a form of many-valued logic in which
the truth value of variables may be any real number between The output is number from range 0–3 where value 0 means
0 and 1. Fuzzy set (A) present generalization of classical the worst state of insulating system and 3 the best state with
set (X), since the membership level of the element to fuzzy minimum of failure probability.
set can be characterized as a number from the interval [0,1]. These membership functions of output correspond to
This is very convenient for modeling of the human reasoning. scores of chief test engineers. The advantage of FLES is that
For creating FLES, data preparation is not needed, since FL it does not give discrete values like human but distributes
controller needs only raw expert knowledge incorporated in them and enables a better prioritization of the observed
set of if–then rule base which define relations between input generators.

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

2696 Electrical Engineering (2024) 106:2691–2703

3.2 ANN works by training the network for several different combina-
tions of hyperparameters on training data, while performance
ANN represents supervised ML model based on the actual is measured on validation data. The combination of hyperpa-
functioning of biological neurons. ANN represents a mathe- rameters that gives the best results on the validation data is
matical model of a human neural network which consists declared optimal and adopted. The created ANN processes
of a large number of neurons. Each neuron in ANN has 6 input sizes and 3 output sizes, one column for each cate-
inputs (x j represents the jth input, j  1,…n) that are con- gory: good, suspicious, and bad EIS generator. The network
nected to it through synapses. The strength of synapses is is designed as acyclic with 50 neurons in the hidden layer and
defined through synaptic weight ωij , where the positive value 3 output neurons. The number of hidden neurons is adjusted
of ωij corresponds to excitatory synapses, while negative val- to achieve the lowest system error with the optimal number
ues model inhibitory synapses. All inputs multiplied by the of neurons. The minimum error was achieved in 54 iterations
appropriate weight coefficients are integrated into the trans- (epochs), with a gradient of 0.0294.
fer function, and this value is compared to the value of the
threshold. In the case that the transfer function is greater than 3.3 ANFIS
the threshold, activation value equals 1; otherwise, it equals
0. This statement can formally be written in the following An adaptive neuro-fuzzy inference system or adaptive
form: network-based fuzzy inference system (ANFIS) is a kind
⎛ ⎞ of artificial neural network that is based on Takagi–Sugeno
 n
fuzzy inference system, similar as in [26].
φ  φ⎝ x j · wi, j − θ j )⎠ (3) It encapsulates benefits from both—ANN and fuzzy logic
j1
systems. Based on the database created with experts’ ratings,
the ANFIS methodology automatically performs an inde-
The most used activation function of a neuron is sigmoid
pendent fuzzification of the input and creates a rule base in
function, defined by:
order to extract the knowledge of the expert who, through
1 the scores, indicated the state of generator insulating system.
oj  (4) Final training of the system after a number of available set-
1 + e−α·φ
tings and iterations was completed after 4 iterations with a
For each set of raw input data, the parameters are adjusted calculated system error of 0.0293. After training, the algo-
so that they best reflect the neural network’s patterns. Cre- rithm defined a total of 2187 fuzzy system rules that define
ating the model demands data preparation and forming a relations between 6 input and output values. The ANN and
feature dataset and target dataset. After initial training, a ANFIS application algorithm is shown in Fig. 4.
group of hyperparameters can be tuned to achieve the best
possible model precision. The training and testing process of 3.4 AENN
ANN is carried out through three phases. The training data
must be selected from the whole set of available data. Train- AENN or autoencoder is a supervised ML model in which
ing uses 70% of the database, testing 15% of the database, and we leverage neural networks for the task of representation
15% are used for validation. The data are normalized to avoid learning. An autoencoder is a neural network that is trained
saturation. These events in the database were randomly cho- to approximately copy its input to the output layer. The net-
sen and formed. This is very important because a uniform part work is divided into two parts: the encoder and the decoder.
of the base can lead to wrong conclusions. If the base is not The autoencoder has a hidden layer whose output is used
mixed, it can lead to the entire training set of variables being for dimensional reduction of the input. The output of that
only good conditions, with a rating of 3, while, for example, hidden layer is called a latent representation. The encoder
the testing set contains all bad conditions with a rating of 1. function, denoted by α maps the original data x, to a latent
Such an ANN would not give adequate results, and during its space l, which is present at the bottleneck. The decoder func-
testing, it would made a huge mistake. Such a mistake and tion, denoted by β, maps the latent space l at the bottleneck
big mean square error of ANN would lead to wrong hyper- to the output:
parameters of the ANN, which will lead to problems in the
application of the methodology. Training the ANN has been α:x →l (5)
done using the back propagation algorithm. The best ANN
hyperparameters are learned from validation data. A test set β :l →x (6)
cannot be used to learn hyperparameters and would not get
a reliable estimation of network performance. The hyperpa-  2
rameter learning procedure is called cross-validation, and it α, β  arg minx − (α ◦ β)x  (7)
α, β

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Electrical Engineering (2024) 106:2691–2703 2697

Fig. 4 The ANN and ANFIS

application algorithm

Fig. 5 The AENN application

algorithm

In this paper, AENN is trained to learn a representation of 3.5 Decision tree

the good condition of the generator insulating system. Later,
it may be possible to detect any disturbance in the data, which The decision tree represents one of the popular ML models
corresponds to the problem in the. The prediction of this which can be used for regression and classification problems.
model will not be class of the insulating system condition. The decision trees can handle both numerical and categor-
It will mean squared error of the prediction from the trained ical data of insulating system conditions. The decision tree
good values. For clarification, if error is high, that implies algorithm starts with a single node that represents the entire
bad insulating system condition, and similarly, if error is low, dataset and then on the most informative feature within the
then insulating system is treated as good. The final version of data splits the dataset into three subsets of insulating system
the autoencoder was created with 75 neurons in the hidden condition. The “splitting” process involves finding the feature
layer, with 5000 training iterations and a mean square error and its corresponding threshold that separates the data into
of 0.0551. The algorithm for AENN is shown in Fig. 5. three groups with distinct insulating system conditions. The

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

2698 Electrical Engineering (2024) 106:2691–2703

Table 1 Example of data

prepared for model training Gen. Input data Output
No.
tgδ0.2 tgδ PD [nC] Riso [M] N1 RisoR [M] Age, Te
[%] max [year]
tip-up
[%]

1 0.95 4 4 8000 5.6 4500 6 3

2 2.4 8 6 1200 1.5 3000 26 2
3 3.7 7 18 800 1.2 1680 56 1

Table 2 Results and comparison

for all five created models Human expert FLES ANFIS ANN DT AENN (mse)

1 1.37 0.64 2 1 9.48E+02

1 1.46 1.82 2 1 3.26E+00
1 2.52 1.65 2 1 4.29E+00
2 1.32 3.81 2 2 1.64E+06
2 1.36 2.09 3 3 4.23E+04
2 2.96 2.11 3 3 3.00E−02
2 1.79 2.96 3 3 2.46E+05
2 1.50 2.21 3 3 9.27E+03
2 1.33 2.86 3 3 1.28E+05
2 1.81 2.35 3 3 6.68E+04
2 2.04 2.54 3 3 4.42E+03
2 2.98 2.24 3 3 1.00E−02
2 1.32 2.98 3 3 1.34E+06
2 2.73 3.00 3 3 4.81E+03
2 1.98 2.56 3 3 6.12E+05
3 2.96 3.00 3 3 9.50E−02
3 3.03 2.92 3 3 1.57E+04
3 2.60 2.45 3 2 1.32E+00
3 1.88 2.59 3 3 2.73E+04
3 2.99 2.88 3 3 3.52E−01

algorithm uses these splits to create branches in the tree, and normalization, and in this solution the min–max normaliza-
each branch leads to another node representing a subset of tion is used:
the data. This branching continues until the algorithm deter-
mines the final classification for each data point based on the
majority class within the corresponding leaf node. In order x − min(x) x − min(x)
to carry out cluster data, similarity between two data sets is x∗   (9)
range(x) max(x) − min(x)
quantified based on their characteristics using the Euclidean
distance, given by the following relationship:

So, the model creates a classification tree that will be


n
d E (x, y)  x − y  (xi − yi )  d E (y, x)
2
(8) trained so that the input data set classifies, i.e., score 1, 2
i1 or 3, like the set on which the tree was trained. The fitctree
function was used to train the tree, which trains a binary clas-
where x  (x 1 ,…,x n ) and y  (y1 ,…,yn ) are attribute values sification tree on the prepared data set. All models training
of two data sets of length n. Cluster algorithms require data has been done in MATLAB® , software.

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Electrical Engineering (2024) 106:2691–2703 2699

Fig. 6 Prediction correlation to the human expert scores

Fig. 8 Comparison of all results obtained by created models in Table 2

DT algorithms make a mistake in only 2 cases out of 20 new

Fig. 7 Comparison of all results obtained by created models in Table 2 test case, so 10% of decisions are not like human expert.
The presented algorithms have certain weaknesses and
limitations in their implementation. The application of FL is
4 Model performance analysis based on the knowledge of orientation limit values of inputs
and the assignment of a rule base based on the experience of
The dataset for model performance testing on 20 generators experts. The application of AENN is based on the existence
is like the dataset presented in Table 1. For all created mod- of a database with good values that indicates the absence of
els, predictions are made, and results are compared to the failure. The disadvantage of ANFIS is that it gives discrete
actual results of the expert assessment. Results of predic- outputs and not fuzzy membership function. The application
tions for all created models are presented in Table 2. As can of ANN is that it is necessary to provide a database with
be seen from Fig. 6, the top 3 performing models correspond exact outputs. An unbalanced database can lead to wrong
to the decision tree, ANN, and ANFIS, respectively. These conclusions by any ML algorithms.
three models have the largest correlation factor with human The confusion matrix for this specific decision tree model
expert scores. A comparison of all results is presented on a is presented in Fig. 6. Based on the display of the confusion
bar chart in Figs. 7 and 8. All models had the most difficul- matrix and TPR/FNR (true positive rate/false negative rate)
ties with the “Suspicious” class of the stator condition, score coefficients, it can be concluded that the accuracy of the tree
2. This is completely corresponding with a usual practice created in this way in classifying the input values into the
where even in human expertise, this second class is always set of “good” values is as high as 97.7%, the set of “suspi-
the most challenging. In the power system of Serbia, this cious”, i.e., the value “2” is 65%, while the accuracy of the
class corresponds to the condition which has some deteriora- classification into the set of “bad” EIS is equal to 77.5%.
tion, but it is not getting worse over time. This finding can be From Fig. 9, as expected, model has made almost per-
significantly improved if a significant amount of data related fect predictions for good generators. It can be demonstrated
to this assessment is collected in the future. Also, there is that model has not misclassified a significant number of bad
doubt about the correctness of the evaluations of the expert generators as good. Errors are mainly within the neighbor-
who makes the decision, because the ML was trained on 600 ing classes. Figure 10 presents created decision tree graph
previous historical data. Does the expert have sufficient expe- because of model that has best prediction correlation to the
rience and did he really examine, monitoring the observed human expert scores. According to Table 2, a decision tree is
generator. The advantage of such ML algorithms is to avoid the model which can identify “Bad” conditions, so to stay on
human error and subjectivity when making a decision. The the safety side this model will be chosen as a representative
DT algorithm with great accuracy observed the results that to demonstrate possibilities for implementation of this “tool”
were evaluated as bad, which represents a great advantage. in CBM and RBM asset maintenance models.

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

2700 Electrical Engineering (2024) 106:2691–2703

Fig. 9 Confusion matrix for the

created decision tree model

Fig. 10 Created decision tree graph

5 Implementation in maintenance strategy cost of the undelivered electricity can exceed the financial
resources that would be invested in maintenance by many
The electricity market and the development of electrification times.
and test methods also conditioned the evolutionary develop- At the same time, the implications for the stability of
ment of different approaches to the maintenance strategies the entire power system can be great when looking at large
of power equipment. Further development of the electricity turbogenerators, which most often form the base power of
market showed that the costs of sudden stoppages and the the power system. The higher the power of the machine,

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Electrical Engineering (2024) 106:2691–2703 2701

Table 3 RBM maintenance

schedule Gen. No. Nominal power Health Index RBM health index Maintenance
(MVA) (decision tree) schedule by RBM

1 40 3 120 16
2 250 3 750 3
3 350 3 1050 2
4 4 2 8 19
5 8 2 16 18
6 15 2 30 17
7 60 2 120 15
8 100 2 200 14
9 120 2 240 13
10 150 2 300 12
11 210 2 420 8
12 290 2 580 7
13 300 2 600 6
14 360 2 720 4
15 650 2 1300 1
16 4 1 4 20
17 360 1 360 10
18 360 1 360 11
19 400 1 400 9
20 650 1 650 5

the greater the negative financial consequences. Namely, In Table 3, a case study has been shown for the set of the
synchronous machines for smaller hydropower plants are 20 generator datasets (used for model evaluation). Appar-
made with nominal powers of up to a few megawatts, ent from generator order number and corresponding nominal
while machines in powerful thermal and thermonuclear pro- power, in third column HI calculated by a decision tree model,
duction plants are made with nominal powers of up to 2 numbers are sorted descending, by HI and Sn. In this exam-
gigawatts. The implications of sudden outages of these pow- ple, HI is derived from the model prediction, so number 3
erful machines on the reliability and stability of power system corresponds to the worst condition, and number 1 the best.
is extremely large compared to small hydro generators. So, CBM strategy would simply use the fact that generators 1, 2
investment in testing and maintenance strategy can provide and 3 should be maintained first, based on their bad scores.
many technical and financial benefits. Upgrading maintenance strategy to the RBM level as pro-
CBM strategy is based on the actual condition of the posed in this paper, schedule is reordered, as shown in the last
machine to schedule proper maintenance. In this purpose, column of Table 3. This way, for example, generator which
simply score generated as a model prediction can be used as is scheduled first for maintenance is scored as “Suspicious”,
it represents so called „Health Index “ of a machine. If a more but it is large and powerful, so in case of failure, impact on
advanced maintenance model is to be implemented, such as the system stability is a lot bigger than outage of the 40MVA
RBM—model based on the risk of failure, then authors of generator No. 1. Considering a powerful power generators,
this paper suggest derived health index by relation 1. RBM, for example, suggests that attention should be directed
to the generator No. 3 with nominal power of 350MVA, prior
the generator No. 20 with nominal power of 650MVA, even
HIRBM  HI × Sn (10) though this one is as much as twice as power It is logical that
RBM get focus maintenance on high power generators. But
in power systems with a big number of power generators and
where HI is health index derived from the prediction of the
with limited resources for maintenance and repairs, it is very
model and S n represents nominal power of the machine. This
important to use risk maps to make a decision. Also, it is
way, it is taken care that the most powerful machines get
even more relevant to use the hourly produced energy of the
treated with the most attention, since their outage can produce
generator multiplied by the hourly price of energy from the
a big influence on power system stability an performance.

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

2702 Electrical Engineering (2024) 106:2691–2703

Fig. 11 Exploitation risk map

present priority map for
maintenance

a)

b)

market as an economic variable. In this way, in addition to already in operation. A set of measurement methods produces
the installed power, the produced energy in the entire power a lot of different values, which are hard to synthesize into
system would be considered. one, easy-to-use value, as a representative of the generator’s
Furthermore, these results can be graphically presented, condition. Temperature, electrical, ambient, and mechanical
so management of the fleet of the generators can be even aging factor of EIS must be a part of monitoring database for
more intuitive. An example of “heatmap” based on the RBM generators in different operational conditions and lifetime.
values has been shown in Fig. 11a. As the generator score Paper present implementation and comparison of mod-
point is more distant from the origin point, generator should ern ML algorithms (FL, ANFIS, ANN, DT, AENN) to extract
be treated with more concern. Figure 11b presents 3D map valuable information from a set of real data from power
and directly indicates the generator with highest RBM index plants of Serbia and calculate the condition, based on real
and the first priority for maintenance. measurements. The accent of the paper points on investiga-
tion correlation between the applied ML algorithms and the
assessments made by human experts. The presented method-
6 Conclusions ology enables incorporation of the thinking and experience of
human experts in decision making about EIS of generators.
To properly maintain synchronous generators in large power The presented algorithms can automate raw data pro-
production companies, it is a matter of essence to pos- cessing and produce numeric values which correspond to
sess an objective, precise mathematical tool to describe its the condition of the generator on the predefined scale. This
condition repeatedly and objectively, based on measured val- enables us to sort generators by their condition. Using this
ues. Known international standards usually do not provide approach can easily introduce the RBM model, as an advan-
criteria for all machine types, and especially for machines tage of the existing CBM model in asset management, and

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Electrical Engineering (2024) 106:2691–2703 2703

eventually, produce better exploitation of the machines which 9. Žarković M, Stojković Z (2017) Analysis of artificial intelligence
will imply financial benefits to the electric power system. Fur- expert systems for power transformer condition monitoring and
diagnostics. Electr Power Syst Res 149:125–136
ther work is the creation of an application that will be used in 10. Mirshekali H, Keshavarz A, Dashti R, Hafezi S, Shaker HR (2023)
the energy sector of Serbia. Future research will be focused Deep learning-based fault location framework in power distribution
on the use of online monitoring data as input. grids employing convolutional neural network based on capsule
network. Electr Power Syst Res 223:109529
Acknowledgements The authors would like to thank the Ministry of 11. Žarković M, Stojković Z (2015) Artificial intelligence based ther-
Science and Technological Development of the Republic of Serbia mographic approach for high voltage substations risk assessment.
which, within the framework of Project III 42009 “Smartgrid“, made IET Gener Transm Distrib 9(14):1935–1945
this work possible. 12. Stone G, Culbert I, Boulter E, Dhirani H (2014) Electrical insula-
tion for rotating machines: design, evaluation, aging, testing, and
Author contributions DI: investigation, writing/original draft, method- repair. Wiley, Hoboken
ology, software; MŽ: supervision, conceptualization, writing—review 13. Ilić D, Žarković M, Stojković Z (2022) Artificial intelligence sys-
& editing. tem for stator condition diagnostic. Electr Eng 104:1503–1513
14. Cigre WG A1.10 (2009) Survey of hydrogenerator failures. Cigre
Funding The author(s) received no financial support for the research, Report 392, CIGRE Paris, France
authorship, and/or publication of this article. 15. Babin A, Polyakov R, Savin L, Tyurin V (2020) Statistical anal-
ysis of turbo generator sets failure causes. Vibroeng PROCEDIA
31:129–133
Availability of data and materials The data that support the findings
16. Seinsch HO (2001) Monitoring und Diagnose elektrischer maschi-
of this study are available from the corresponding author, M.Z., upon
nen und Antriebe. Stand der Forschung, Entwicklungstendenzen,
reasonable request.
Frankfurt am Main, VDE-Haus
17. Stefenon SF, Seman LO, Aquino LS, Coelho LS (2023) Wavelet-
Declarations Seq2Seq-LSTM with attention for time series forecasting of level
of dams in hydroelectric power plants. Energy 274:127350
Conflict of interest No author associated with this paper has disclosed 18. Stefenon SF, Ribeiro MHDM, Nied A, Yow K, Mariani VC, Coelho
any potential or pertinent conflicts which may be perceived to have LS, Seman LO (2022) Time series forecasting using ensemble
impending conflict with this work. learning methods for emergency prevention in hydroelectric power
plants with dam. Electr Power Syst Res 202:107584
Consent to participate Not applicable. 19. Seman LO, Stefenon SF, Mariani VC, Coelho LS (2023) Ensemble
learning methods using the Hodrick–Prescott filter for fault fore-
Consent to publish Not applicable. casting in insulators of the electrical power grids. Int J Electr Power
Energy Syst 152:109269
20. Corso MP, Stefenon SF, Singh G et al (2023) Evaluation of visible
contamination on power grid insulators using convolutional neural
References networks. Electr Eng 105:3881–3894. https://doi.org/10.1007/s0
0202-023-01915-2
1. Ilić D (2022) Synchronous generators electrical insulation system 21. IEC 60034-1:2022 Rotating electrical machines: rating and perfor-
condition diagnostic based on artificial intelligence Ph.D. disser- mance
tation, Dept., Power systems. Serbia., Belgrade Univ., Belgrade. 22. IEEE Std 115-1995 and 2019 Guide for test procedures for syn-
Serbia chronous machines including acceptance and performance testing
2. Nian L, Fusheng T (2000) The new intelligent monitoring method and parameter determination for dynamic analysis
for operating states of large generator units. In: IEEE Power Engi- 23. EPRI (1989) Hydropower plant modernization guide. EPRI
neering Society Winter Meeting. Conference Proceedings—Singa- 24. Russell S, Norvig P (2010) Artificial intelligence: a modern
pore 23–27, 223–227. https://doi.org/10.1109/pesw.2000.849959 approach, 3rd edn. Pearson Education, London
3. Tavner P, Ran L, Penman J, Sedding H (2008) Condition monitoring 25. Klement EP, Slany W (1997) Fuzzy logic in artificial intelligence.
of rotating electrical machines. IET, eISBN: 978-0-86341-991-1 In: Encyclopedia of computer science and technology, vol 34, suppl
4. Zenios S, Ziemba W (2007) Handbook of asset and liability 19. Marcel Dekker, pp 179–190
management. Applications and case studies, vol 2. Elsevier, North- 26. Jang J-SR (1993) ANFIS: adaptive-network-based fuzzy inference
Holland. ISBN: 9780444528025 system. IEEE Trans Syst Man Cybern 23(3):665–685
5. Mansour D-EA, Taha IBM, Farade RA, Wahab NIBA (2022) Partial 27. Eisenecker U (1995) AI: the tumultuous history of the search for
discharge diagnosis in GIS based on pulse sequence features and artificial intelligence. AI Commun 8:45–47
optimized machine learning classification techniques. Electr Power
Syst Res 211:108162
6. Bezerra FE, Di Santo SG, Junior GM (2023) An approach based Publisher’s Note Springer Nature remains neutral with regard to juris-
on wavelets and machine learning to build a prediction model for dictional claims in published maps and institutional affiliations.
SF6 filling pressure of high-voltage circuit breakers. Electr Power
Syst Res 216:108974 Springer Nature or its licensor (e.g. a society or other partner) holds
7. Žarković M, Stojković Z (2019) Artificial intelligence SF6 circuit exclusive rights to this article under a publishing agreement with the
breaker health assessment. Electr Power Syst Res 175:105912 author(s) or other rightsholder(s); author self-archiving of the accepted
8. Xing Z, He Y, Chen J, Wang X, Du B (2023) Health evaluation manuscript version of this article is solely governed by the terms of such
of power transformer using deep learning neural network. Electr publishing agreement and applicable law.
Power Syst Res 215(Part B):109016

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”), for small-
scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are maintained. By
accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use (“Terms”). For these
purposes, Springer Nature considers academic use (by researchers and students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or a personal
subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription
(to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the Creative Commons license used will
apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data internally within
ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking, analysis and reporting. We will not
otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of companies unless we have your permission as
detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that Users may
not:

1. use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to circumvent access
control;
2. use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil liability, or is
otherwise unlawful;
3. falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by Springer Nature in
writing;
4. use bots or other automated methods to access the content or redirect messages
5. override any security feature or exclusionary protocol; or
6. share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer Nature journal
content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates revenue,
royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain. Springer Nature journal
content cannot be used for inter-library loans and librarians may not upload Springer Nature journal content on a large scale into their, or any
other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any information or
content on this website and may remove it or features or functionality at our sole discretion, at any time with or without notice. Springer Nature
may revoke this licence to you at any time and remove access to any copies of the Springer Nature journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express or implied
with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or warranties imposed by law,
including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be licensed
from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other manner not
expressly permitted by these Terms, please contact Springer Nature at

[email protected]