**4. Data Analysis from Acoustic Emission Signals**

Altogether, we collected 72 oscillograph records of partial discharges. Each of these records depicts a time window of one second. In general, many occurrences of partial discharge are registered in these time slots.

In addition to this phenomenon, the data acquisition system also recorded mechanical waves that were used to evaluate the gauging of acoustic emission sensors. These waves are the result of the break, near the surface where the sensor is installed, of graphite with speci‐ fications given by the manufacturer of acoustic emission sensors. The graphs resulting from this test are highlighted in Figure 6.

**Figure 6.** Acoustic emission signal resulting from the gauging process of sensors.

As shown in this figure, the signal is thus composed of two well-defined moments. The first of these relates to the instant when there was a mechanical disruption of graphite, while the second stage is the result of the impact of the pencil with the surface where the acoustic emission sensor is installed.

From Figure 8 we can see that each partial discharge results in a highly correlated mechani‐

Intelligent Systems for the Detection of Internal Faults in Power Transmission Transformers

http://dx.doi.org/10.5772/51417

13

Figure 10 illustrates the average frequency spectrum of an acoustic emission signal coming from a standard partial discharge. Through this frequency behavior, it can be seen that there is high signal energy at approximately 95 kHz and within the range between 160 kHz and 180 kHz. These values are of great importance in distinguishing partial discharge signals

cal wave. The graphs shown in Figure 9 highlight this relationship more clearly.

**Figure 9.** Detail of relationship between partial discharge current and acoustic emission.

**Figure 10.** Average frequency spectrum of acoustic emission signal coming from a partial discharge.

from other interferences.

Figure 7 shows in more detail the first moment of the mechanical wave in Figure 6, while Figure 8 illustrates how the mechanical waves are related to the currents resulting from par‐ tial discharges.

**Figure 7.** Details of the acoustic emission signal resulting from the gauging process of sensors.

**Figure 8.** Relationship between partial discharge current and acoustic emission waves.

From Figure 8 we can see that each partial discharge results in a highly correlated mechani‐ cal wave. The graphs shown in Figure 9 highlight this relationship more clearly.

**Figure 9.** Detail of relationship between partial discharge current and acoustic emission.

As shown in this figure, the signal is thus composed of two well-defined moments. The first of these relates to the instant when there was a mechanical disruption of graphite, while the second stage is the result of the impact of the pencil with the surface where the acoustic

Figure 7 shows in more detail the first moment of the mechanical wave in Figure 6, while Figure 8 illustrates how the mechanical waves are related to the currents resulting from par‐

**Figure 7.** Details of the acoustic emission signal resulting from the gauging process of sensors.

**Figure 8.** Relationship between partial discharge current and acoustic emission waves.

emission sensor is installed.

tial discharges.

12 Advances in Expert Systems

Figure 10 illustrates the average frequency spectrum of an acoustic emission signal coming from a standard partial discharge. Through this frequency behavior, it can be seen that there is high signal energy at approximately 95 kHz and within the range between 160 kHz and 180 kHz. These values are of great importance in distinguishing partial discharge signals from other interferences.

**Figure 10.** Average frequency spectrum of acoustic emission signal coming from a partial discharge.

In order to verify the behavior of the sensors for the tests, the voltage and current signals are processed in order to find the frequency response of these devices. In Figure 11 the ampli‐ tude versus frequency for the first calibration test has been recorded. The top of the graph highlights the energy and voltage signals sampled, and at the bottom there is the amplitude versus frequency. From the signal analysis it is then possible to observe a maximum re‐ sponse around 400 Hz and 100 kHz.

In Figure 12, the signals were assigned in segments where the amplitude was more signifi‐ cant for detection purposes, which now represents the presence of different peak amplitudes

Intelligent Systems for the Detection of Internal Faults in Power Transmission Transformers

http://dx.doi.org/10.5772/51417

15

The energy signal shows an envelope having important information, making clear the dif‐ ferences between the acoustic emission signal and the reflections that are also registered. In order to better evaluate these peaks, segments of interest were amplified and the frequency

In the segment highlighted in Figure 12, there is clearly a large concentration of low frequen‐ cies, with maximum amplitude at 10 Hz. In contrast, Figure 13 presents a large concentra‐

It is worth noting that, in the light of the two analyses, the signal with higher energy, record‐ ed in the first segment, has an extremely low frequency wave. Thus, the propagation veloci‐ ty tends to be higher due to the proximity to the spectrum of mechanical waves. However, for higher frequencies, typically observed in electromagnetic waves, there is a decrease of the signal energy, because this wave will suffer large attenuation when propagating through the insulating oil. Thus, the signal perceived by the acoustic emission sensor has already suf‐ fered severe degradation before being detected. This attenuation phenomenon is of great importance for the location process of partial discharges when installing more sensors in the experimental tank. In fact, since the speed of wave propagation in the insulating oil is

The energy calculation is performed to obtain the full power of a signal. However, some sig‐ nals are negative and therefore a quadratic sum of the sampled points must be calculated, as

known, it is then possible to estimate the location of the source of discharge.

response was recalculated for this section, as reported in Figure 13.

**Figure 13.** Detail of frequency response of the acoustic emission signal (segment 2).

tion at 100 kHz and another at approximately 2.5 MHz.

shown in the following equation:

at various frequencies.

**Figure 11.** Frequency response of the acoustic emission signal.

**Figure 12.** Detail of frequency response of the acoustic emission signal (segment 1).

In Figure 12, the signals were assigned in segments where the amplitude was more signifi‐ cant for detection purposes, which now represents the presence of different peak amplitudes at various frequencies.

In order to verify the behavior of the sensors for the tests, the voltage and current signals are processed in order to find the frequency response of these devices. In Figure 11 the ampli‐ tude versus frequency for the first calibration test has been recorded. The top of the graph highlights the energy and voltage signals sampled, and at the bottom there is the amplitude versus frequency. From the signal analysis it is then possible to observe a maximum re‐

sponse around 400 Hz and 100 kHz.

14 Advances in Expert Systems

**Figure 11.** Frequency response of the acoustic emission signal.

**Figure 12.** Detail of frequency response of the acoustic emission signal (segment 1).

The energy signal shows an envelope having important information, making clear the dif‐ ferences between the acoustic emission signal and the reflections that are also registered. In order to better evaluate these peaks, segments of interest were amplified and the frequency response was recalculated for this section, as reported in Figure 13.

**Figure 13.** Detail of frequency response of the acoustic emission signal (segment 2).

In the segment highlighted in Figure 12, there is clearly a large concentration of low frequen‐ cies, with maximum amplitude at 10 Hz. In contrast, Figure 13 presents a large concentra‐ tion at 100 kHz and another at approximately 2.5 MHz.

It is worth noting that, in the light of the two analyses, the signal with higher energy, record‐ ed in the first segment, has an extremely low frequency wave. Thus, the propagation veloci‐ ty tends to be higher due to the proximity to the spectrum of mechanical waves. However, for higher frequencies, typically observed in electromagnetic waves, there is a decrease of the signal energy, because this wave will suffer large attenuation when propagating through the insulating oil. Thus, the signal perceived by the acoustic emission sensor has already suf‐ fered severe degradation before being detected. This attenuation phenomenon is of great importance for the location process of partial discharges when installing more sensors in the experimental tank. In fact, since the speed of wave propagation in the insulating oil is known, it is then possible to estimate the location of the source of discharge.

The energy calculation is performed to obtain the full power of a signal. However, some sig‐ nals are negative and therefore a quadratic sum of the sampled points must be calculated, as shown in the following equation:

$$E = \sum\_{l=1}^{N} \sum\_{j=1}^{M} \text{svar}\_{l,j}^{2} \tag{1}$$

**Figure 15.** Energy response calculated for sensor 2 (mounted on the front wall - top left) during experiments 1, 2 and 3.

Intelligent Systems for the Detection of Internal Faults in Power Transmission Transformers

http://dx.doi.org/10.5772/51417

17

**Figure 16.** Energy response calculated for sensor 3 (mounted on the side wall-lower right corner) for experiments 1,2 and 3

The sensor 4 showed an energy response similar to that already shown for sensors 1 and 2

**Figure 17.** Calculated energy response for sensor 4 (mounted on the side wall - upper left) during experiments 1, 2 and 3

(Figure 17).

where *N* is the *i*-th window, and *M* represents the *j*-th point of the data window (consisting of 1101 points per window).

Thus, it may be noted that each data window corresponds to an acoustic emission signal measured by a given sensor. In this case, 8 sensors are used and, therefore, for each partial discharge we have 8 data windows. In addition, 10 samples for each partial discharge are still considered, which were obtained at different moments. Thus, the energy calculation for each of the 8 acoustic emission sensors is shown form Figures 14 to 21. Moreover, three dif‐ ferent experiments were compared, where there was variation in the depth of the partial dis‐ charges in the oil tank used during the tests.

Experiment 1 represents a partial discharge located at 5 cm from the surface of the insulat‐ ing oil, while experiments 2 and 3 are respectively located at 21.5 and 40 cm from the surface of the insulating oil.

Experiment 3 also had a small variation in the distance of the partial discharge from the front of the experimental tank, where it was moved 1 cm with respect to the original posi‐ tion of tests 1 and 2.

It is important to mention that this displacement is made in such away that the partial dis‐ charge of experiment 3 could be detected by sensors closer to the front wall of the tank, where it was expected that sensors 1 and 2 allocated on the wall would be more sensitive in experiment 3 rather than in experiments 1 and 2.

From Figures 14 and 15 it is possible to observe the energy response supplied by sensors 1 and 2 (for each of 10 samples), which represents the greatest contribution of experiment 1 in sensitizing them, while sensor 3 shows an energy response which makes it difficult to define which experiment caused the highest sensitization (Figure 16).

**Figure 14.** Energy response calculated for sensor 1 (mounted on the front wall - bottom right) during experiments 1, 2 and 3.

Intelligent Systems for the Detection of Internal Faults in Power Transmission Transformers http://dx.doi.org/10.5772/51417 17

<sup>=</sup>åå <sup>2</sup>

=1 =1

*i j*

of 1101 points per window).

16 Advances in Expert Systems

of the insulating oil.

tion of tests 1 and 2.

and 3.

charges in the oil tank used during the tests.

experiment 3 rather than in experiments 1 and 2.

which experiment caused the highest sensitization (Figure 16).

*N M*

,

*E* (1)

*i j*

sinal

where *N* is the *i*-th window, and *M* represents the *j*-th point of the data window (consisting

Thus, it may be noted that each data window corresponds to an acoustic emission signal measured by a given sensor. In this case, 8 sensors are used and, therefore, for each partial discharge we have 8 data windows. In addition, 10 samples for each partial discharge are still considered, which were obtained at different moments. Thus, the energy calculation for each of the 8 acoustic emission sensors is shown form Figures 14 to 21. Moreover, three dif‐ ferent experiments were compared, where there was variation in the depth of the partial dis‐

Experiment 1 represents a partial discharge located at 5 cm from the surface of the insulat‐ ing oil, while experiments 2 and 3 are respectively located at 21.5 and 40 cm from the surface

Experiment 3 also had a small variation in the distance of the partial discharge from the front of the experimental tank, where it was moved 1 cm with respect to the original posi‐

It is important to mention that this displacement is made in such away that the partial dis‐ charge of experiment 3 could be detected by sensors closer to the front wall of the tank, where it was expected that sensors 1 and 2 allocated on the wall would be more sensitive in

From Figures 14 and 15 it is possible to observe the energy response supplied by sensors 1 and 2 (for each of 10 samples), which represents the greatest contribution of experiment 1 in sensitizing them, while sensor 3 shows an energy response which makes it difficult to define

**Figure 14.** Energy response calculated for sensor 1 (mounted on the front wall - bottom right) during experiments 1, 2

**Figure 15.** Energy response calculated for sensor 2 (mounted on the front wall - top left) during experiments 1, 2 and 3.

**Figure 16.** Energy response calculated for sensor 3 (mounted on the side wall-lower right corner) for experiments 1,2 and 3

The sensor 4 showed an energy response similar to that already shown for sensors 1 and 2 (Figure 17).

**Figure 17.** Calculated energy response for sensor 4 (mounted on the side wall - upper left) during experiments 1, 2 and 3

By means of the energy response supplied by sensor 5 (Figure 18) it can be seen that there is a certain emphasis on the response of experiment 1, but its energy levels are very close to those of experiments 2 and 3.

**Figure 20.** Energy response calculated for sensor 7 (mounted on the side wall - bottom left) during experiments 1, 2

Intelligent Systems for the Detection of Internal Faults in Power Transmission Transformers

http://dx.doi.org/10.5772/51417

19

**Figure 21.** Energy response calculated for sensor 8 (mounted on the side wall - upper left) during experiments 1, 2

Finally, sensor 8 presented an energy response (Figure 21) similar to that already obtained

This section provides a theoretical foundation for fuzzy inference systems and artificial neu‐

Systems called fuzzy are built based on the theory of fuzzy sets and fuzzy logic, introduced by Zadeh in 1965, to represent knowledge from inaccurate and uncertain data. Fuzzy sys‐

by other sensors, whose higher sensitization was caused by experiment 1.

ral networks, as they are very prominent intelligent tools in the literature.

and 3

and 3.

**5. Intelligent Systems**

**5.1. Fuzzy inference systems**

**Figure 18.** Energy response calculated for sensor 5 (mounted on the rear wall - bottom right) during experiments 1, 2 and 3.

The energy response of sensor 6 (Figure 19) in almost all samples presented responses simi‐ lar to those obtained by sensors 1 and 2. However, in the first sample it can be seen that there are very similar levels of energy in the three experiments, although sensor 6 was a lit‐ tle more sensitive in experiment 3.

**Figure 19.** Energy response calculated for sensor 6 (mounted on the rear wall - top left) during experiments 1, 2 and 3

Sensor 7 presented the most complex energy response (Figure 20) because its response was unbiased for most samples. This is one factor that shows the complexity involved in the treatment of acoustic emission signals, making the application of intelligent systems very promising.

Intelligent Systems for the Detection of Internal Faults in Power Transmission Transformers http://dx.doi.org/10.5772/51417 19

**Figure 20.** Energy response calculated for sensor 7 (mounted on the side wall - bottom left) during experiments 1, 2 and 3

**Figure 21.** Energy response calculated for sensor 8 (mounted on the side wall - upper left) during experiments 1, 2 and 3.

Finally, sensor 8 presented an energy response (Figure 21) similar to that already obtained by other sensors, whose higher sensitization was caused by experiment 1.

#### **5. Intelligent Systems**

By means of the energy response supplied by sensor 5 (Figure 18) it can be seen that there is a certain emphasis on the response of experiment 1, but its energy levels are very close to

**Figure 18.** Energy response calculated for sensor 5 (mounted on the rear wall - bottom right) during experiments 1, 2

The energy response of sensor 6 (Figure 19) in almost all samples presented responses simi‐ lar to those obtained by sensors 1 and 2. However, in the first sample it can be seen that there are very similar levels of energy in the three experiments, although sensor 6 was a lit‐

**Figure 19.** Energy response calculated for sensor 6 (mounted on the rear wall - top left) during experiments 1, 2 and 3

Sensor 7 presented the most complex energy response (Figure 20) because its response was unbiased for most samples. This is one factor that shows the complexity involved in the treatment of acoustic emission signals, making the application of intelligent systems

those of experiments 2 and 3.

18 Advances in Expert Systems

tle more sensitive in experiment 3.

very promising.

and 3.

This section provides a theoretical foundation for fuzzy inference systems and artificial neu‐ ral networks, as they are very prominent intelligent tools in the literature.

#### **5.1. Fuzzy inference systems**

Systems called fuzzy are built based on the theory of fuzzy sets and fuzzy logic, introduced by Zadeh in 1965, to represent knowledge from inaccurate and uncertain data. Fuzzy sys‐ tems consist of a way to make a computational decision close to a human decision. Figure 22 shows a block diagram that expresses, in a simplified form, how a fuzzy system works.

Another factor that should be noted is the inference procedure, in which a variety of meth‐ ods can be used. Currently, the most commonly used methods are those of Takagi-Sugeno

Intelligent Systems for the Detection of Internal Faults in Power Transmission Transformers

http://dx.doi.org/10.5772/51417

21

Artificial neural networks are computational models inspired by the human brain, which can acquire and retain knowledge. Among the various neural network architectures, there is the architecture of multiple layers, called MLP (Multilayer Perceptron). This type of archi‐ tecture is usually used for pattern recognition, functional approximation, identification and control tasks [16]. The structure of a neural network can be developed according to Fig. 3.

As seen in Fig. 3, the neural network structure is basically composed of an input layer, hid‐ den neural layers and an output neural layer. Also, between the layers, there is a set of weights, which are represented by a matrix of synaptic weights that will be adjusted during the training phase. It is further worth commenting that, for each of the neurons (hidden neu‐ ral layers and output neural layer), it is necessary to implement activation functions in order to limit their output. In view of the basic configuration of the MLP neural network, other

During the training phase of MLP neural networks, some algorithms can be used. Currently, the backpropagation algorithm can be highlighted, which uses a descendent gradient calcu‐ lation to reach the best adjustment of the synaptic weight matrix. In addition to the backpro‐ pagation algorithm, the Levenberg-Marquardt algorithm has been widely used because of its ability to accelerate the convergence process, due to the use of an approximation of New‐

factors that should be explored are the training and validation stages.

and Mamdani.

**5.2. Artificial neural networks**

**Figure 23.** MLP neural network architecture.

ton's method for non-linear systems [16].

**Figure 22.** Diagram of a fuzzy inference system.

In the "Fuzzification" block, input values (in this case, information provided by the acoustic emission, gas concentration and electrical measurement sensors) are provided and condi‐ tioned, becoming fuzzy sets. Similarly, the "Defuzzification" block is responsible for trans‐ forming the outputs of the fuzzy system into non-fuzzy values (i.e., values which indicate the kind of internal fault and its location). The "Linguistic Rules Base" block has the func‐ tion of storing the linguistic sentences and is fundamental to guarantee good system per‐ formance. The linguistic rules base and membership functions related to the inputs and outputs can be provided by experts or by automated methods, such as the ANFIS system (Adaptive Neural Fuzzy Inference System). On the other hand, the "Inference Procedure" block maps a system by using the linguistic rules. Thus, if rules are combined with input fuzzy sets acquired by the fuzzification interface, the system is then able to determine the behavior of the output variables of the system so that they can be defuzzified, generating the corresponding output to a given input value.

When using a fuzzy inference system, fuzzy rules and sets are adjusted and tuned by expert information. However, in some cases, because of the complexity and nonlinearity of the problem, it is necessary to use hybrid systems, such as ANFIS, where adjustments are per‐ formed in an automated manner according to the data set that represents the process. How‐ ever, it is worth mentioning that, regardless of the setting, the whole fuzzy system has linguistic rules that can be represented as follows:

> 1 2 1 2 : 1is 2isx is y *i i R Input x Input Output* =× +× + *ax bx c* **If and Then**

Another factor that should be noted is the inference procedure, in which a variety of meth‐ ods can be used. Currently, the most commonly used methods are those of Takagi-Sugeno and Mamdani.

#### **5.2. Artificial neural networks**

tems consist of a way to make a computational decision close to a human decision. Figure 22 shows a block diagram that expresses, in a simplified form, how a fuzzy system works.

In the "Fuzzification" block, input values (in this case, information provided by the acoustic emission, gas concentration and electrical measurement sensors) are provided and condi‐ tioned, becoming fuzzy sets. Similarly, the "Defuzzification" block is responsible for trans‐ forming the outputs of the fuzzy system into non-fuzzy values (i.e., values which indicate the kind of internal fault and its location). The "Linguistic Rules Base" block has the func‐ tion of storing the linguistic sentences and is fundamental to guarantee good system per‐ formance. The linguistic rules base and membership functions related to the inputs and outputs can be provided by experts or by automated methods, such as the ANFIS system (Adaptive Neural Fuzzy Inference System). On the other hand, the "Inference Procedure" block maps a system by using the linguistic rules. Thus, if rules are combined with input fuzzy sets acquired by the fuzzification interface, the system is then able to determine the behavior of the output variables of the system so that they can be defuzzified, generating the

When using a fuzzy inference system, fuzzy rules and sets are adjusted and tuned by expert information. However, in some cases, because of the complexity and nonlinearity of the problem, it is necessary to use hybrid systems, such as ANFIS, where adjustments are per‐ formed in an automated manner according to the data set that represents the process. How‐ ever, it is worth mentioning that, regardless of the setting, the whole fuzzy system has

: 1is 2isx

1 2

*Output* =× +× + *ax bx c*

is

*R Input x Input*

**If and**

1 2

**Figure 22.** Diagram of a fuzzy inference system.

20 Advances in Expert Systems

corresponding output to a given input value.

linguistic rules that can be represented as follows:

y

*i*

*i*

**Then**

Artificial neural networks are computational models inspired by the human brain, which can acquire and retain knowledge. Among the various neural network architectures, there is the architecture of multiple layers, called MLP (Multilayer Perceptron). This type of archi‐ tecture is usually used for pattern recognition, functional approximation, identification and control tasks [16]. The structure of a neural network can be developed according to Fig. 3.

As seen in Fig. 3, the neural network structure is basically composed of an input layer, hid‐ den neural layers and an output neural layer. Also, between the layers, there is a set of weights, which are represented by a matrix of synaptic weights that will be adjusted during the training phase. It is further worth commenting that, for each of the neurons (hidden neu‐ ral layers and output neural layer), it is necessary to implement activation functions in order to limit their output. In view of the basic configuration of the MLP neural network, other factors that should be explored are the training and validation stages.

During the training phase of MLP neural networks, some algorithms can be used. Currently, the backpropagation algorithm can be highlighted, which uses a descendent gradient calcu‐ lation to reach the best adjustment of the synaptic weight matrix. In addition to the backpro‐ pagation algorithm, the Levenberg-Marquardt algorithm has been widely used because of its ability to accelerate the convergence process, due to the use of an approximation of New‐ ton's method for non-linear systems [16].

On the other hand, the validation stage has the purpose of verifying the integrity of previ‐ ously conducted training, so that the learning ability (generalization) of neural networks can be analyzed.

**•** Data validation (use of other data than those used in training/tuning stage);

the hardware used in the acquisition of the signals, we used the experimental tank.

tions to be performed by the software is set out below (Figure 24).

tinguish these signals and to locate the source of partial discharges.

previously, these are provided as input to the expert system.

**Figure 24.** Overview of the embedded software.

sal Serial Bus).

literature.

**•** Performance analysis of the methodology in relation to other methodologies found in the

Intelligent Systems for the Detection of Internal Faults in Power Transmission Transformers

http://dx.doi.org/10.5772/51417

23

It is worth mentioning that, out of the 6 steps mentioned above, most attention should be given to the allocation and acquisition of data, because bad data acquisition can affect the whole process of identifying and locating faults. It is also important to emphasize that the calculations made during the preprocessing of the signals was devised in order to extract the characteristics that best represent the positioning of the partial discharge in relation to the acoustic emission sensor. However, for this first stage of testing the expert system and

In order to better represent the embedded software, a block diagram detailing the calcula‐

As can be seen in Figure 24, it may be noted that the embedded software, after obtaining the acoustic signal, applies some computations in order to extract the characteristics that may represent the signal appropriately. Through these features, the expert system is able to dis‐

In this context, during the preprocessing step of the signs, the following calculations are performed: RMS, Energy, Length, Amplitude, Rise Time and Threshold. Finally, after ob‐ taining the signal characteristics, they are sent to the computer through a USB (Univer‐

Upon receipt of these data, the expert system is then responsible for providing information regarding the location of any partial discharge in the transformer. In order to better repre‐ sent the overview of expert system, a block diagram is shown in Figure 25. In this figure, it may be noted that, after the received data concerning the characteristics commented upon
