Abstract
This paper presents two methods for condition monitoring of high-voltage equipment based on the thermography approach. The overheating temperature of the hot spot has been obtained using the performed thermography procedure. MATLAB® technical computing software has been used to design the fuzzy controller and artificial neural network. The age of the element, the voltage level, the overheating temperature and the temperature of the previous overheating have been used as the reference inputs for the designed controller and artificial neural network. The developed software tool has been applied for the evaluation of the urgency of intervention in the function of the input data, designed rule base and the methods of defuzzification. Real measurements were used as input data in both methods so that the results were confirmed. The results might serve as a good orientation in the high-voltage equipment condition monitoring. The educational aspects of the application of this software tool are very important for both undergraduate and master's students studying Monitoring and Diagnostics of High Voltage Substations. During the past two academic years, the software application has received favorable comments from students.
Introduction
The reliable and adequate supply of electrical energy is the basic requirement which is placed on an electrical power system (EPS) by consumers of electrical energy. This requirement must be fulfilled by the EPS at all functional levels of production, transmission and distribution of electrical energy. The condition of the high-voltage equipment (HVE) has a significant influence on the reliability of EPS operation. The characteristics of the HVE irreversibly change during exploitation as a result of several factors, of which the most important are: aging due to the operating voltage, accelerated aging due to overheating, as well as the influence of atmospheric occurrences and chemical reactions.
The assessment of the HVE condition can generally be conducted on the basis of the methods which require interrupting operation (off-line methods) and methods which do not require operation to be interrupted (on-line methods). The use of off-line methods results in better insight into the condition of the HVE and significant expenses for disassembly, transport and laboratory testing, too. The on-line methods represent the most economically acceptable methods for the assessment of the HVE condition. Although less reliable, their use can give a preliminary picture of the HVE condition based on which further steps are taken, i.e. conducting field and/or laboratory testing.
One of the most commonly used on-line methods for the assessment of the HVE condition is thermography.1,2 Its use has increased dramatically with the commercial and industrial applications in the past 10 years. The problem that occurs after the thermography is establishing the time frame for the repair of the analyzed HVE. In the long term, warming can be used as an indicator of the HVE condition. Based on the overheating temperature and properties of the analyzed HVE, further testing will be made.
Artificial intelligence (AI) has been investigated and applied with success to solve some long-standing EPS problems where conventional methods experience difficulties. Fuzzy logic (FL) represents one of the most commonly used techniques of AI. 3 Generally speaking, a neuro-fuzzy system is an intelligent model which contains the learning capabilities of neural networks and knowledge illustration of fuzzy logic systems using linguistic expressions. 4 A review of the application of this powerful tool in meeting challenging different problems in the EPS has been given in the literature.5–9
In this paper, a fuzzy logic and ANN-based thermography approaches for condition monitoring of the HVE of the presented transmission system have been given. According to the performed thermography procedure, the overheating temperature of the hot spot for all analyzed elements has been obtained. MATLAB® technical computing software has been used to design the fuzzy controller. 10 The age of the element, the voltage level, the overheating temperature and the temperature of the previous overheating have been used as the reference input data for the designed controller. The corresponding rule base has been designed. The center of gravity (COG) has been used as the method for defuzzification. Each step in the design of the controller is gradually explained so that every student can understand and realize FL.
Based on the growing number of available different measurement data, it is necessary to decide the maintenance of HVE. Different scenarios of input data and faults show up every day in a plant. Those scenarios extend the base which already exists. Based on such data, an ANN is developed which gives decision on the possible maintenance of equipment. MATLAB® technical computing software is used to create ANN. All settings in this creation are explained in this paper.
Fuzzy logic
On one hand, the use of advanced technology makes it possible to significantly reduce the time, and on the other hand it increases the accuracy of the methods for monitoring and diagnostics. One such method is based on FL, and it allows the condition monitoring of the HVE and the appropriate intervention. FL is a mathematically formalized model which can show some uncertainties in linguistics. In classical, clear set theory any particular element (x) either belongs or does not belong to a defined set. In other words, belonging of elements is extremely distinctive. Fuzzy set (A) is, in that matter, a generalization of classical set (X), since the membership (i.e. membership level) of the element to fuzzy set can be characterized as a number from the interval [0, 1]. In other words, the membership function (μ
A
(x)) of the fuzzy set maps each element of the universal set of the mentioned interval of real numbers
Algorithm for implementation of FL and ANN
FL is simply aimed at overcoming communication problems related to the difference between rules that impose formal theories and ways of thinking, which describes the behavior of the human mind. Fuzzy controller provides a formal methodology for representing, managing and implementing human heuristic knowledge about how to make decisions.
11
ANN successfully predicted the behavior and state of HVE. The problem that we are facing is how to make the output signal and connect it with the actual measurements. At the same time all input and output data must be consolidated into a database. The whole process is organized by the algorithm presented in Figure 1.
Algorithm for implementation of FL and ANN for monitoring of HVE.
The output information denotes the urgency of intervention of the analyzed HVE. The results explicitly indicate the effects of denoted input data and the applied methods upon the urgency of intervention, which might serve as a good orientation in the HVE condition monitoring. In this way, students are taught how to use real measurements from monitoring and thus perform automation of diagnostics using software tools. Use of the software tool makes it possible to confirm the theoretical bases in an environment which is very similar to the actual conditions in engineering practice. In this kind of modern learning environment, students acquire practical knowledge while simultaneously mastering the theoretical.
Application of FL
MATLAB® technical computing software has been used to design the fuzzy controller. The layout of the performed fuzzy controller is displayed in Figure 2(a). From the layout presented in Figure 2(a) it can be seen that the reference inputs to the controller are: age of the element, voltage level, the overheating temperature, the temperature of the hot spot and ambient temperature. The output signal is a number from the interval [0, 1] which refers to the condition of the element and the urgency of the intervention on it. If the output number is higher, closer to 1, the intervention is more urgent and the tested element is potentially more defective.
(a) Description of fuzzy controller and fuzzy membership functions for set of (b) lifetime, (c) previous overheating, (d) voltage level, (e) temperature of overheating of element and (f) output.
One of the operations in the fuzzy controller is fuzzification which simply modifies the input signals, so that they can be properly interpreted and compared with the rules in the rule base. The reference signal is converted into an appropriate fuzzy shape. This is provided by the membership functions, which actually map the degree of the truth claims. Membership functions are a continuous measure of safety if the variable is classified as the linguistic value. To be concrete, the age of element is introduced to continuous membership function (μ
IN1
(x)) which determines the degree of lifetime of the element. In a similar way, membership functions (μIN2, 3, 4(x)) are assigning to the other three inputs. The only membership function to be fulfilled is to be scaled and to have values from 0 to 1. In Figure 2(b), the continuous fuzzy membership function for set of age of element is presented. Trapezoidal membership functions are illustrated in Figure 2(c). Figure 2(d) shows Dirac membership functions of fuzzy set voltage level. According to the literature,
12
overheating is divided into three groups, used to assess defective equipment: up to 10℃, between 10℃ and 30℃ and greater than 30℃. Three bands with some overlap are also included (Figure 2e). The output signal shown in Figure 2(f) has five membership functions that overlap and are related to:
do nothing (element is correct), pay attention (element is uncertain), required intervention within 60 days, required intervention within 30 days and urgent (intervention is needed as soon as possible).
Rule base contains knowledge on how to control the system, in the form of a set of logical (if–then) rules.
13
The interface is a mechanism for evaluating fuzzy controller which controls rules that are relevant to the current state of the system and decides, with the logic circuit, what will be the control signal, i.e. output. It is assumed that this is the best way to manage the system described through sentences in a certain language. The task is to find a way to enter this “linguistic” knowledge of the process of the fuzzy controller. The aim of fuzzy controller is to use FL to represent the mapping of the inputs into outputs of the controller. Primary mechanism for that is if−then list of statements, called rules. All the rules are carried out parallel and the order does not matter. This list of rules is called a rule base (rule–base). The rules refer to the linguistic variables and their properties. If all the terms and all the features that define those terms, i.e. variables, are previously defined, the system design that interprets the rules can be accessed.
With an aim to express the result produced by the current values of input variables, a set of rules has to be formed. These rules have the form: if <condition> then <consequence>, and it is possible to have multiple parallel if-rules that are connected by the connectors “and” “or” and “not”, so that using them complex statements can be built.
If the element is old, at the end of its life span, if the voltage level is high, 400 kV, and if the overheating is above 30℃, and there was previous overheating, it is clear that the urgent intervention is needed. This rule has the form: If (life span is old age) and (previous-overheating is YES) and (temperature is high) and (voltage-level is 400 kV) then (output1 is urgent). Display of rules-base for a particular combination of inputs.
Defuzzification, which is the final step in the fuzzy-controller, transforms the interface conclusion in a signal that can be a signal representing the output.
Defuzzification process is essentially the opposite process of fuzzification and it is called decoding. This is in fact a process that needs to transform the result of aggregation, which is basically a surface cross-section, into the signal that the process can recognize. The output controller must have a unique value, usually represented by a real number. The most commonly used methods for defuzzification is the center of the gravity (COG). The following expression has been used for denoted method
Urgency of intervention in the function of the overheating temperature and the age of element (a) and the overheating temperature and the temperature of the previous overheating (b).
Application of ANN
ANNs are used for making a decision on how to maintain HVE. Based on the input data and experience from EPS the base data is formed of samples upon which the trained ANN. The multilayer feedforward neural networks applied in this paper consist of many interconnected signal processing elements called neurons. These neurons form a layered configuration of network through only feedforward interlayer weighted connection. 14
The structure of the ANN proposed in this paper is shown in Figure 5(a). The optimum ANN can be achieved by adjusting suitable number of hidden layers, number of neurons in each hidden layer, learning rate, momentum and parameters of the activation function. These are the parameters which have strong effects on the operation of an ANN. Once the ANN architecture has been proposed and its parameters have been determined it should be examined through three phases.
3
Displayed network that gives the best results is network with a single hidden layer of 25 neurons. ANN is trained on the formed database size 1000 × 5, where the first four columns represent the input and the last column is the desired output. Output can get a value of 1, 2, 3, 4 and 5, which correspond to output membership functions in fuzzy controller. The training data have to be selected among the whole set of available data. For this training 70% of the database was used, 15% for testing and 15% for validation. The data are then normalized to avoid saturation. Training the ANN has been done using the back propagation algorithm. The most commonly used sigmoid function (arc tang) is applied as activation function. The results of testing the network (confusion matrix) are displayed in Figure 5(b).
The proposed ANN structure and confusion matrix of testing ANN.
Confusion matrix is 6 × 6 dimensions and it is given a number and percentage of the hits and errors. There are five target classes (horizontal) and five output classes (horizontal). Element on position (4, 3) in matrix denotes that the number of samples of class 3 was found to belong to the class of fourth in fact, all the many values that belong to the main diagonal are classified correctly samples, and all other values were misclassified samples. The sixth species is the percentage value of hits and errors for each of the five inputs (target) class. The sixth column shows the percentage value of hits and errors of output classes. As the most important value is shown at the position (6, 6), it tells how the classification was done for all classification of all samples of each class.
Results
FL and ANN based tool were realized using MATLAB® technical computing software as the development platform for creating the user interface (Graphical User Interfaces (GUIs) Toolbox).
15
By providing an interface between the user and the application's underlying code, GUIs enable the user to operate the application without knowing the commands. For this reason, applications that provide GUIs are easier to learn and use for students. Students can easily test the formed fuzzy controller and ANN. Command window for the calculation of the coefficient of emergency intervention is given in Figure 6(c) and (f). The following information should be entered for the analyzed element: age in the range of 0 to 30 years, previous overheating temperature in the range from 0 to 30℃ (if the previous overheating temperature is higher than 30℃ the element will be already replaced), overheating temperature in the range from 0 to 50℃ and voltage level. By pressing the RUN button the result will be determined. FL enables the coefficient of emergency intervention as a result. By applying ANN a message about the required procedure in the maintenance of the tested element is obtained.
Terminal bushing of power transformer.
An illustration of this process is provided in the following examples. The thermovision recording of the power transformer put into operation 25 years ago was done. Characteristic of this power transformer is the observed repeated warm place on the outer terminal bushing. The temperature of the previous overheating was 20℃ and the current temperature is 29℃. The voltage level of the element is 220 kV. The terminal bushing of power transformer and the corresponding thermal image are presented in Figures 6 and 7, respectively. Fuzzy controller gives the coefficient of emergency intervention for the given inputs. The value is 0.89754, which means that the intervention is urgent (Figure 8). ANN gives the same result.
Thermal image of the terminal bushing of power transformer. Command window for calculation of the coefficient of emergency intervention for the analyzed power transformer.

Another example is thermovision of the current transformer whose voltage level is 110 kV. Element is in operation for 15 years. There was no previous overheating and the current temperature of overheating is 29℃. Photography and thermal image of the tested current transformer are presented in Figures 9 and 10, respectively. ANN gives an output message ‘pay attention’, and fuzzy controller gives the output value 0.343 (Figure 11). This value corresponds to the second membership function of the fuzzy controller which also means to pay attention.
Image of current transformer. Thermal image of current transformer. Command window for the calculation of the output message using ANN for analyzed current transformer.


The output information for disconnectors as a function of age, temperature of (previous) overheating and voltage level.
Due to the present uncertainty in the output membership functions overlap, ANN results are always on the side of safety. From the results presented in Table 1, it can be concluded that the instructions for intervention explicitly depend on the age of elements, temperature of (previous) overheating and voltage level.
Educational aspects and students' feedback
The course, ‘Monitoring and Diagnostics of High Voltage Substations’ is taken during master's degree studies at the Faculty of Electrical Engineering, the University of Belgrade. This course is selected by students who wish to acquire knowledge from the field of condition monitoring of HVE in power substations. Aside from the lecture component, which consists of 3 h per week of theoretical instruction presented as an obligatory form of teaching, the course also includes computer exercises in which students master the material envisaged by the curriculum. Students are provided with the material for conducting the exercises in electronic and paper form. The instructions contain step-by-step explanations for each stage in the realization of the exercises. During the exercises, students are required to record and analyze the obtained results, as well as compose a report on the exercise which they hand in to be checked. The use of the software tool makes it possible to confirm the theoretical bases about AI in an environment which is very similar to the actual conditions in engineering practice. In this kind of modern learning environment, students acquire practical knowledge while simultaneously mastering the theoretical bases. The use of software such as MATLAB® enables the computer to become an integral component of theoretical and practical engineering education. Students are required to form a fuzzy controller which is explained step-by-step. Students analyze real thermal images of various elements of high-voltage substations. On the basis of these thermal images and details of the analyzed elements they test their fuzzy controller. At the end, they have to confirm their results with the results of ANN that is already formed. Comparing the results of the two methods is very easy with the GUI. By verifying their results, students gain confidence. This mode of work gives the students a sense of pleasure, because they are able to apply their knowledge and create part of applications that can be used in practice.
Evaluation form for students and their responses.
Conclusions
The output data of thermography is the temperature of the warm place. It is necessary to load input data used in FL and ANN from an existing database, which are associated with the tested element. The input data are discussed according to certain criteria and decisions are made. Based on this information it is possible to:
estimate the trend, make a prediction about behavior of each component of the element, timely alert responsible staff and block the element in order to prevent greater harm or damage.
It is clear that such a decision should be made quickly and accurately. FL provides just that. FL allows use of expert knowledge and experience of experts in the field of thermography. A database that is growing with new measurements is even better for use of ANN. ANN confirms the results of the FL and allows taking into account the experiences in the maintenance of the equipment. The obtained and validated results can be used in decision-making for maintenance of equipment in power systems. The presented methods and the results confirm the validity of applying FL and ANN in making appropriate decisions about the HVE condition monitoring.
Finally, using FL and ANN as an educational tool gives users the opportunity to solve real problems in a very sophisticated way. In this manner, the users achieve a better understanding regarding the design of automation monitoring and diagnostics of HVE. Additionally, the users acquire excellent active learning skills.
Footnotes
Acknowledgements
The authors would like to thank the Ministry of Science and Technological Development of the Republic of Serbia which, within the framework of Project III 45003 “Optoelectronic nanodimensional systems – road towards applications”, Subproject: “Nanostructural optoelectronic sensing systems” and Project III 42009 “Smartgrid”, made this work possible.
Conflict of interest
None declared.
Funding
The second author would like to thank the Alexander von Humboldt Foundation, Bonn, FR Germany, for its support for his scientific research work.
