JPH1078471A - Measuring method for partial discharge - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a measuring method, for a partial discharge, in which a skilled measuring operator is not required, in which the partial discharge can be judged simply and with high judgment accuracy and whose reliability is high. SOLUTION: The magnitude (q) of a discharged electric-charge amount in every cycle (t) obtained by a partial discharge detector attached to a power cable and information on a generation phase angle ϕ are arranged in a time- series manner, and three-dimensional ϕ-q-t data by which the magnitude of the electric charge amount of a measuring signal with reference to an electricity- application phase and the generation phase angle are indicated is created. Then, the electricity-application angle, the number of electricity-application cycles and the discharged electric-charge amount axis of the ϕ-q-t data are divided, and a ϕ-q-t pattern is created. Then, the ϕ-q-t pattern is input to a neural network Nu1 in which the ϕ-q-t pattern of a partial discharge signal is learned in advance, and the existence of the partial discharge signal is judged. In addition, ϕ-q-n data [where (n) represents the number of occurrences] is sorted into groups, a partial discharge is judged by using a sorted result, and the partial discharge can be measured with high accuracy in the same manner.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for measuring a partial discharge of a power cable, and more particularly to a method for measuring a partial discharge which can accurately determine the presence or absence of a partial discharge using a neural network.

[0002]

2. Description of the Related Art When measuring a partial discharge of a power cable, for example, an oscilloscope simultaneously observes a voltage phase applied to the cable and a measurement signal, and performs partial discharge based on a relationship between the observed pulse signal and the voltage phase. Is used. This is because the partial discharge pulse has a strong correlation with the voltage phase φ, and most of them have the first quadrant (0 ° to 90 °) and the third quadrant (180 °) as shown in FIG.
270 °).

Using the above relationship, a φ-qn pattern is created from the charge amount q of the measurement pulse, the potential application phase φ, and the unit frequency n, and this φ-qn pattern is previously stored in the partial discharge pulse. There is known a method of determining the presence / absence of a partial discharge by inputting φ-qn of the above to a learned neural network (for example, see Japanese Patent Application Laid-Open No. 5-256895).

FIG. 17 is a diagram showing an example of the φ-qn pattern. As shown in the figure, the phase angle and the charge amount axis are divided into a plurality of parts (in the figure, for example, the phase angle is divided into 20 and the charge amount axis is divided into 10).
Is represented by a plurality of cells (20 × 10 = 200 cells in the figure), and the number of occurrences n (or occurrence frequency) is assigned to the value of each cell. As a neural network, for example, FIG.
As shown in FIG. 3, a three-layer neural network having an input layer Ui, an intermediate layer Um, and an output layer Uo is used, and the above-described cells are replaced with the cells Ui1 to Ui of the input layer of the neural network.
n, the number of occurrences n (or occurrence frequency) is input to the cells Ui1 to Uin of each input layer. If the input layer and the output layer of the neural network are learned in advance by giving the φ-qn pattern and the presence / absence of partial discharge, the presence / absence of partial discharge can be obtained by giving the φ-qn pattern to the neural network. Can be determined.

[0005]

The technique of simultaneously observing the voltage phase imposed on the cable and the measurement signal with an oscilloscope or the like and judging partial discharge from the relationship between the observed pulse signal and the voltage phase is a measurement technique. In addition, there is a problem that skill is required for the determination and a skilled person is required for the partial discharge measurement. In addition, the above-mentioned φ-q-n
The method of determining the partial discharge by inputting the pattern to the neural network is to collect the information of the magnitude q of the partial discharge and the generation phase φ for a predetermined time and pattern it as the occurrence frequency n. Information is lost.

Therefore, in the case of the φ-qn pattern in which the information of the temporal change of the generated phase angle is lost, an erroneous determination may be made and there is a problem in reliability. Further, when the above-mentioned φ-qn pattern is input to a neural network to determine a partial discharge, it is necessary to learn a large amount of data in order to correspond to various measurement signals, and create an ambiguous neural network. Will be done. For this reason, the neural network may make a wrong decision, which may cause a problem in reliability.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problems of the prior art, and a first object of the present invention is to provide a simple and highly accurate judgment without requiring a skilled measurer. An object of the present invention is to provide a highly reliable partial discharge measurement method capable of performing discharge determination. A second object of the present invention is to make it possible to create measurement signal patterns without losing the characteristic of temporal change peculiar to partial discharge, and to reliably determine the occurrence of partial discharge from these measurement signal patterns. To provide a method for measuring partial discharge.
A third object of the present invention is to calculate a classification characteristic value from φ-qn data, divide the values into groups, and determine a partial discharge by using a partial discharge determination method corresponding to each group, thereby obtaining accuracy. An object of the present invention is to provide a partial discharge measurement method capable of performing a high partial discharge determination.

[0008]

FIGS. 1 and 2 are diagrams for explaining a partial discharge measuring method according to the present invention. In the present invention, the determination of partial discharge is performed as described in (a) and (b) below. . Next, the partial discharge measurement in the present invention will be described with reference to FIGS.

(A) Judgment of Partial Discharge Using φ-qt Data FIG. 1 is a diagram for explaining partial discharge judgment using φ-qt data in the present invention. The measurement signal detected by the partial discharge detector attached to the power cable is stored in cycle units. Then, the stored magnitude of the discharge charge and the generated phase information for each cycle are listed in chronological order, and a three-dimensional φ-q− indicating the magnitude of the charge and the generated phase angle of the measurement signal with respect to the applied potential phase. Create t data.

In general, φ-qt data includes an applied potential phase (360 °) on the X-axis, the number of applied cycles on the Y-axis,
The axis corresponds to the magnitude of the discharge charge amount, but may correspond to an arbitrary axis. FIG. 1A shows three-dimensional φ-q- of partial discharge.
An example of t data is shown. Although the magnitude of the discharge charge amount (Z axis) is shown by shading in the figure, the magnitude of the discharge charge amount can be more easily understood by indicating the magnitude of the change in color. From the three-dimensional φ-qt data obtained as described above, a measurement signal φ-qt pattern to be input to the neural network Nu1 is created. Therefore, the application potential phase angle (X-axis), the number of application cycles (Y-axis), and the discharge charge amount axis (Z-axis) of the three-dimensional φ-qt data are divided as shown in FIG.

Here, the main feature of the present invention is to perform data processing so as not to lose the feature of the temporal change of the generated phase angle. Therefore, in order to maintain the characteristic of the temporal change of the generated phase angle, a certain number of divisions is required for the X axis and the Y axis, but the number of divisions for the Z axis is not so important. That is, the applied potential phase on the X axis is represented by the following equation.
It is desirable to divide the number into about several hundreds, and it is desirable to divide the number of cycles on the Y-axis usually in units of cycles. However, in some cases, a plurality of cycles may be combined.

If an area surrounded by each of the X and Y axis division units divided as described above is a cell, each cell corresponds to each cell of the input layer Ui of the neural network Nu1, and The value input to each cell corresponds to the magnitude of the discharged charge amount. When the partial discharge charge amount is simply divided into two, the value input to the cell of the neural network is 0 or 1. 0 corresponds to a case where there is no signal in the cell (or a signal not exceeding a predetermined value), and 1 corresponds to a case where a signal exists in the cell (or a signal having a predetermined value or more). Further, a case may occur in which each cell includes a plurality of signals. However, as described above, since the X-axis and the Y-axis are set to the number of divisions so as to maintain the characteristic of the temporal change of the generated phase angle, the rate at which a plurality of signals are included is small. May exist, and they may be combined in any manner.

FIG. 1B shows an example in which the φ-qt data is patterned. In the figure, the X axis (application potential phase angle) is 18 °.
Is divided into 20 units, and the Y-axis (the number of power application cycles) is 25 units with 2 cycles as a division unit. Further, the Z-axis (the magnitude of the amount of discharged electric charge) is divided into two with a threshold of S / N = 2 based on the ratio of the measurement level S and the noise level N, and 1 if a signal is present (in FIG. Is painted out), and the others are set to 0. As is clear from FIG. 1B, the characteristic of the temporal change of the generated phase angle is maintained even after patterning.

There are many types of neural network, and any one may be used. Usually, a three-layer type (input layer, intermediate layer, output layer) shown in FIG. 1 using back propagation is used. It is better to use one.
In order to perform learning, a jigmoid function is used, and in order to improve determination accuracy even for unknown data, a method in which learning is performed by adding an error to learning data in advance is preferable.

It is best to use the data of the actually generated partial discharge signal as the learning data. However, in some cases, the data may be created so as to have the characteristics of the partial discharge signal. In any case, the partial discharge signal generated by the partial discharge signal or the φ-qt data having the characteristic of the partial discharge signal by the above-described method so as not to lose the characteristic of the generated phase angle and the temporal change. -Qt patterns are created, and the neural network Nu1 is trained with these patterns. In a trained neural network,
By inputting the φ-qt pattern obtained from the measurement signal as described above, a signal indicating the presence / absence of partial discharge can be obtained from the output layer of the neural network, and the presence / absence of partial discharge can be accurately determined. Can be.

(B) Judgment of Partial Discharge by Grouping Measurement Signals FIG. 2 is a diagram illustrating judgment of partial discharge by grouping of measurement signals. From the partial discharge signal obtained by the partial discharge detector attached to the power cable and the potential application phase, a discharge charge amount q, a generated phase angle φ, a generated number or a generated frequency n is obtained. Then, as shown in FIG. 2A, a classification characteristic value is calculated from these values. Since the classification characteristic value is an index for grouping, it is desirable that the classification characteristic value be a value obtained by summarizing the characteristics of the measurement signal data. Therefore, it is calculated by a statistical method using at least two or more values from q, φ, and n.

As the actual characteristic value, any characteristic value may be used as long as it is a value obtained by calculation. For example, the average fluctuation of the discharge electric charge, the dispersion fluctuation of the discharge electric charge, and the variation of the generated phase angle are used. It is desirable to use characteristic values such as average fluctuation, dispersion fluctuation of the generated phase angle, and fluctuation of the number of occurrences. Then, based on these classification characteristic values, the measurement signal data is classified into several groups Gr1 to Grn as shown in FIG. The classification group is created in advance for each purpose. If the main purpose is to determine the presence or absence of partial discharge, there are three types: a partial discharge group, a noise group, and a group in which partial discharge and noise are mixed.

If the main purpose is to determine the cause of the occurrence of the partial discharge, a group is created according to the type of defect assumed. If the main purpose is to determine the degree of partial discharge deterioration, a group is created according to the progress of deterioration. In any case, some groups having different partial discharge characteristics are created in advance. In creating a group, as many partial discharge data and noise data as possible for the purpose of measurement are collected, characteristic values for classification of each data are calculated, and grouping is performed based on the characteristic values. A cluster classification method is preferable as a classification method used when creating a group, but another classification method may be used.

On the other hand, any method may be used to classify the measured data into the groups created as described above, but a method using a neural network Nu2 as shown in FIG. is there. Therefore, representative data is selected from the data of each group divided as described above, and a classification characteristic value is obtained, which is used as learning data of the neural network Nu2 shown in FIG. 2B. Then, a classification characteristic value is given to the input layer of the neural network Nu2, and a group is given to the output layer, and learning is performed, thereby creating a classification determination neural network Nu2. The classification characteristic value calculated from the measurement data is input to the classification determination neural network Nu2, and the measurement data is classified into the groups Gr1 to Grn according to the output values.

Next, as shown in FIG. 2C, the measurement signals classified into each group are input to the partial discharge determination means D1 to Dn corresponding to each group, and the measurement signal is a partial discharge or a noise. Is determined. Partial discharge determination means D
For 1 to Dn, a method using a neural network is simple. As the learning data, it is desirable to use a characteristic value calculated from representative data of each group, but a value obtained by another calculation method may be used. Based on the learning data, a neural network is created for each group in which output values are made to correspond to, for example, partial discharge and noise.

The classified measurement data is input to the respective partial discharge determining neural networks, and it is determined whether the output is a partial discharge or a noise based on the output value. Further, as a determination method, any method of statistical calculation such as chaos, wavelet, fuzzy, and principal component analysis may be used, or a combination of these methods may be used. Further, the above method (a) is added to the above method (b), and the result obtained by the method (b) is applied to the result obtained by the method (a) using a neural network using the φ-qt pattern. By determining the partial discharge and the noise, the partial discharge can be determined with higher accuracy.

As described above, in the present invention, the above-mentioned problem is solved as follows. (1) In a method of measuring a partial discharge of a power cable, a generation phase angle φ of a measured signal and a magnitude q of a discharge charge amount are chronologically listed for each potential application phase cycle t. -T pattern is created and the partial discharge signal φ
-Qt pattern obtained from the measurement signal is input to the neural network that has learned the -qt pattern,
The presence or absence of a partial discharge signal is determined.

(2) In the method of measuring the partial discharge of the power cable, a phase angle φ at which a signal having a magnitude equal to or greater than a predetermined discharge charge amount is generated from the measured signals is determined every application potential phase cycle t. A φ-t pattern is created in a time series, and the measurement signal is input to a neural network in which the φ-t pattern of the partial discharge signal has been learned in advance to determine the presence or absence of the partial discharge.

(3) In the above (1) and (2), a plurality of potential application phase cycles are combined to form a pattern. (4) In the method of measuring the partial discharge of the power cable, a characteristic value for classification is calculated from the measured signal and the potential phase signal, and the measured signals are classified into at least two or more groups based on the characteristic value. , A partial discharge determination is performed using the classification result.

(5) In the above (4), as the classification characteristic value, the classification characteristic value calculated by a statistical method from at least two or more values among the discharge charge amount q, the generation phase angle φ, and the generation number n. Is used. (6) In the above (4) and (5), the groups are at least one or more and two or more of the three types of partial discharge group, noise group, and group in which partial discharge and noise are mixed. (7) In the above (4), (5), and (6), the signal classified into the group is determined to be a partial discharge or a noise by using a predetermined partial discharge determination unit corresponding to the group.

(8) In the above (7), at least one or more determination methods from among statistical calculations such as neural networks, chaos, wavelets, fuzzy, and principal component analysis are used as the predetermined partial discharge determination method. . (9) In the method of measuring the partial discharge of a power cable, a classification characteristic value is calculated from a measurement signal and a potential application phase signal, and the measurement signals are classified into at least two or more groups based on the characteristic values. And performing a partial discharge determination using the classification result, and inputting a learning φ-qt pattern corresponding to the partial discharge determination result to a neural network prepared in accordance with the partial discharge determination result to perform learning. , A φ-qt pattern in which the measured phase angle φ of the measured signal and the magnitude q of the discharge charge amount are listed in time series for each potential application phase cycle t to generate the measured signal. A φ-qt pattern is created in which the phase angle φ and the magnitude q of the discharge charge amount are listed in time series for each potential application phase cycle t, and the partial discharge determination result is stored in each of the neural networks. The obtained φ-qt pattern is input, and the result of the determination of the partial discharge performed by the neural network using the classification result is evaluated.

In the invention of claims 1 to 3 of the present invention,
As described in the above (1) to (3), a φ-qt pattern of the measurement signal is created from the measurement signal and the potential application phase, and the measurement signal pattern which does not lose the characteristic of the temporal change of the generated phase angle. Is input to the neural network in which the learning of the partial discharge pattern has been completed in advance, and the partial discharge determination is performed. Therefore, the presence or absence of the partial discharge can be accurately determined. Therefore, the determination of the partial discharge can be easily performed without requiring a skilled measurer.

In the invention according to claims 4 to 8 of the present invention,
As described in (4) to (8) above, φ, q, and n are calculated from the measurement signal and the potential application phase signal to calculate a characteristic value for classification, and the characteristic value for classification is determined in advance according to the purpose of measurement based on the characteristic value. Since the group is classified into the created groups and the presence / absence of the partial discharge is determined by the partial discharge determination unit dedicated to each group corresponding to each group, the partial discharge determination with high determination accuracy can be performed. According to the ninth aspect of the present invention, since the configuration is made as in the above (9), a more accurate partial discharge determination can be performed.

[0029]

FIG. 3 is a diagram showing a configuration of a measuring circuit according to an embodiment of the present invention.
Using a CV cable (insulation thickness: 6 mm), a defect was simulated and a power application test was performed. In the figure, 1 is a CV cable, 2a and 2b are air terminals, 3 is a coupling capacitor,
4a and 4b are partial discharge detectors. Reference numeral 9 denotes a simulated defective portion provided on the power cable. In this embodiment, the simulated defect portion 9 facilitates the generation of partial discharge.
A part of the cable sheath and the shielding layer was peeled off to expose the outer conductive layer, and a treeing needle was inserted about 1 mm into the insulator from above the outer conductive layer.

5 is a balancer, 6 is a partial discharge measuring device, and the outputs of the partial discharge detectors 4a and 4b are input to the partial discharge measuring device 6 via the balancer 5, and the partial discharge measuring device 6
Is converted into a digital signal together with the application potential phase signal 8 and stored in the digital memory 7. Reference numeral 10 denotes a personal computer, and the personal computer 10 performs a partial discharge determination based on the partial discharge measurement signal and the potential application phase as described below. 11 is a display device, a φ-qn graph obtained by the personal computer 10,
A φ-qt graph, a partial discharge determination result, and the like are displayed. Reference numeral 12 denotes a power transformer.

Next, an embodiment of the present invention using the test circuit will be described. (1) Judgment of Partial Discharge Using φ-qt Data (i) Example 1 In the test circuit of FIG. 3, it was confirmed that partial discharge occurred at an applied voltage of 40 kV, and a measurement signal at this time was obtained. And the application potential phase signal into the digital memory 7. Next, the data in the digital memory was read out to the personal computer 10 to create 20 partial discharge signal patterns, which were used as neural network learning data realized by software of the personal computer 10. As shown in FIG. 4A, the number of phase axis divisions of the pattern was 20 (division unit: 18 °), and the time axis was 50 divisions of 25 divisions (division unit, 2 cycles). Then, the measured signal is divided into two with a threshold of S / N = 2 at a ratio of the signal level S to the signal level N, and the case where each cell contains more measurement signals is 1 FIG. 4B shows the case where the value is not included as 0.
The φ-qt pattern shown in FIG.

The neural network is a three-layer type using back propagation (input layer, intermediate layer, output layer)
Was used. Then, as shown in FIG.
The number of i is 500, which is obtained by multiplying the number of phase axis divisions by the number of time axis divisions, and the number of intermediate layers Um is 50, and the number of output layers Uo
Was set to two. The learning uses a sigmoid function, and in order to improve the accuracy of determination even for unknown data, the learning data is trained 50,000 times by the back propagation method with an error of about 10% added in advance. In addition, a noise signal detected when power was not applied was similarly patterned and learned by the neural network without partial discharge.

Next, a power application test is performed on five 6 kV power cable samples, and the signal measured by the partial discharge measuring device is input to the neural network trained as described above to determine the presence or absence of partial discharge. Judged. The method of applying power and the method of measurement are the same as those in FIG. 3, but the simulated defect portion 9 is not provided. In the power application test, the voltage was raised in steps of 10 kV / 5 minutes, and 100 kV was set as the upper limit voltage. At each voltage step, partial discharge determination was performed by a neural network, and when partial discharge was determined, power application was stopped.
The test results are shown in Table 1 below.

[0034]

[Table 1]

As a result of the partial discharge determination by the neural network, 1 to No. 3 and No. 3 5 was determined to have no partial discharge at an upper limit voltage of 100 kV. No. 4
Determined that there was partial discharge at 80 kV, and stopped applying power. After that, No. When the sample of No. 4 was examined, it was confirmed that the insulator had a wound.

(Ii) Example 2 A partial discharge signal was measured during operation of a 275 kV class CV cable line laid at the site and during operation suspension (no power application), and the method according to the present invention was implemented. At the same time, a comparison was made with a conventional method of inputting a φ-q-n pattern to a neural network and determining partial discharge. Both the present invention and the conventional neural network using the φ-qn pattern learned typical partial discharge signals and noise signals generated in the first and third quadrants as shown in FIG. .

In the neural network of the present invention, the number of phase axis divisions of the pattern is 2 as in the case of FIG.
0 division (division unit: 18 °), and the time axis was divided into 50 divisions of 25 cycles (division unit, 2 cycles). Therefore, as shown in FIG. 5, the number of input layers is 500, the number of intermediate layers is 50, and the number of output layers is 2. Learning is 5000 by back propagation method
Trained 0 times. The measurement time was 4 hours during operation, and 4 hours during subsequent operation suspension.

In the measurement at the time of operation, an abnormal signal which was not seen at the start of the measurement was detected continuously after about 3 hours had passed. Based on the measurement signal at that time, a φ-qn pattern and a φ-qt pattern are created, and a conventional φ-qn pattern is created.
Each pattern was input to the neural network using the qn pattern and the neural network of the present invention, and the presence or absence of partial discharge was determined.

In the neural network of the present invention, it was determined that there was no partial discharge, but in the conventional neural network, it was determined that there was partial discharge. After that, the measurement was similarly performed in a state where the operation was stopped. An abnormal signal observed during the operation was measured, and it was confirmed that the signal was not a signal associated with power application. In the subsequent measurement, no abnormal signal was confirmed. From these measurement results, the effectiveness of the method of discriminating using the φ-qt pattern of the present invention is more effective than the conventional method using the φ-qn pattern in which the temporal change of the measured signal is lost. Verified.

(2) Partial Discharge Judgment by Grouping Measurement Signals (i) Example 3 Partial discharge measurement using group judgment was performed on unknown measurement data. A classification group was created in advance based on partial discharge data and noise data. As the partial discharge data, the test circuit shown in FIG. 3 was used, and a partial discharge signal generated when a voltage of 40 to 60 kV was applied to the 22 kVCV cable provided with the simulated defect 9 was used as data. In addition, as noise data, 275 kV25
Data measured on the 00sqCAZV cable line when no power was applied was used.

FIG. 6 shows an example of partial discharge data and noise data. In the figure, the horizontal axis represents the phase angle φ, the vertical axis represents the charge amount q, and the shading represents the number of occurrences n (occurrence frequency). Five types of characteristic values were obtained from the φ-q-n patterns of these data. The five characteristic values are the average variation of the discharge charge amount,
These are dispersion fluctuation of the discharge charge amount, average fluctuation of the generated phase angle, dispersion fluctuation of the generated phase angle, and fluctuation of the number of generation. The outline of each is shown below. 7 and 8 show examples of characteristic values obtained for partial discharge and noise. Average Fluctuation in Discharge Charge Amount The average charge in a certain phase φi is expressed by the following equation (1), and the average charge calculated for each phase is defined as the average charge fluctuation.

[0042]

(Equation 1)

Dispersion Variation of Discharge Amount The charge amount variance is represented by the following equation (2), and the value obtained by calculating the charge amount variance for each phase is defined as the dispersion change value of the average charge amount.

[0044]

(Equation 2)

Average Variation of Generated Phase Angle The average generated phase angle is expressed by the following equation (3), and the average generated phase angle obtained for each phase is defined as the average generated phase angle fluctuation value.

[0046]

(Equation 3)

Variance Variation of Generated Phase Angle The phase angle variance is represented by the following equation (4), and the phase angle variance obtained for the angular phase is defined as the variance variation value of the average generated phase angle.

[0048]

(Equation 4)

Variation of Generation Phase Angle Assuming that the sum of ni in a certain phase φi is N, N obtained for each phase is a variation value of the number of occurrences. Based on these characteristic values, grouping by cluster classification was performed. Cluster classification is a general term for classification algorithms, and a commonly used aggregation method (tree clustering) was used. This is a method of sequentially connecting clusters under a certain aggregation rule (joining rule) using similarity between clusters and distance (distance between patterns in a multidimensional space). The Ward method, which has a characteristic that small groups can be easily classified, was adopted as the aggregation rule.

FIGS. 9 and 10 show tree diagrams created by calculating five characteristic values from the classification data and performing cluster classification based on the characteristic values. The data group is divided into two large groups (Gr
1, Gr2), and the data in each group forms a subgroup at a position with a coupling distance of 30 or less. Assuming that the coupling distance is about 8 as a threshold, the data group has five subgroups (Gr
1-1, Gr1-2, Gr1-3, Gr2-1, Gr2
-2). Here, Table 2 shows a breakdown and main features of the data classified into each group.

[0051]

[Table 2]

Most of Gr1 is occupied by partial discharge, and most is classified into Gr1-1 and Gr1-2.
Gr1-3 is classified into cases in which the first and third quadrants are partially included or the single quadrant occurs.
Many of them are dominated by noise. In the case of Gr2, all are noise except for 5% of the partial discharge. Based on the results of the above-described cluster classification, a group determination neural network for performing grouping as shown in FIG. 2 and a partial discharge determination neural network corresponding to each group were created.

The neural network used was of three layers (input layer, intermediate layer, and output layer) utilizing back propagation. The input layer of the neural network for group determination was 92, the intermediate layer was 46, and the output layer was 5. The neural network for partial discharge determination corresponding to each group has 92 input layers and 46 intermediate layers.
And two output layers. For the training data, select representative data from each group, use a jigmoid function for training, and give random noise of about 10% until it converges to improve the decision accuracy even for unknown data. I learned.

FIG. 11 is an outline of a partial discharge judgment flow of the present embodiment utilizing group judgment. As shown in the figure, five classification characteristic values are calculated using 400 pieces of unknown data, and the five classification characteristic values are input to a neural network for group determination, and the group Gr1 is calculated based on the characteristic values. -1 to Gr2-2. Further, the classification characteristic value is input to the partial discharge determination neural network for each group, and the partial discharge and noise (PD in FIG.
NOISE represents noise). On the other hand, as a comparison, a neural network for inputting data of the φ-qn pattern was created using learning data of all the groups, and evaluation was performed with the same data. Table 3 below shows the evaluation results.

[0055]

[Table 3]

From the above Table 3, the correct answer rate of 86.3% was obtained when there was no group judgment. However, when the group judgment was used, the correct answer rate varied among the groups. It can be seen that the judgment accuracy is improved to 8%.

(Ii) Fourth Embodiment When the group determination method in the third embodiment is used, erroneously determined data is confirmed.
For 1, Gr1-2, erroneous noise occurring in the first and third quadrants of the φ-q-n pattern was relatively large. Therefore, in order to further improve the correct answer rate, in order to identify the noise generated in the first and third quadrants, the partial discharge determination using the neural network for group determination of the third embodiment has been described in the first and second embodiments. Neural network judgment by φ-qt pattern was additionally performed.

That is, when looking at a narrow time range, the partial discharge signal is randomly generated within a certain phase range, while the noise is continuously generated in the same phase (hereinafter, referred to as synchronous noise). ) Or the generation phase tends to shift with time. These features are input to the φ-qt pattern determination neural network described in the first and second embodiments to make a determination.

FIG. 12 shows an outline of the flow of the partial discharge determination of this embodiment. As shown in the drawing, groups Gr1-1 to Gr1-1 are formed by a neural network for group determination.
2-2, and partial discharge and noise were determined by a partial discharge determination neural network for each group. Then, the partial discharge and the noise were discriminated by the φ-qt pattern determination neural network corresponding to the partial discharge and the noise, respectively. φ-qt
The input data of the pattern judgment neural network is
The phase angle φ is divided into 20, and the time axis t is divided into 20 cycles of 10
The division was performed, and the partial discharge charge amount q was input to the value of each cell.
In the case where a plurality of signals are included in the same cell, the signal is represented by the largest charge amount.

FIG. 13 shows an example of partial discharge and synchronous noise. The φ-qt pattern determination neural network uses a three-layer type in the same manner as the group determination, and has an input layer of 2 layers.
00 (20 × 10), 50 intermediate layers, and 3 output layers. The three output layers corresponded to "partial discharge", "1.3 quadrant noise", and "other noise", and data having respective characteristics were selected as learning data. The learning method is the same as described above.

Using 400 unknown data, the group judgment neural network and the partial discharge judgment neural network of each group have φ-q as described above.
Five characteristic values analyzed from the −n pattern were input, and in any case of the partial discharge and the noise, the evaluation of the partial discharge discrimination was finally performed by the φ-qt pattern. As a comparison, φ-q using learning data of all groups
A neural network for inputting -n pattern data was created, and the same data was input and evaluated. Table 4 shows the evaluation results.

[0062]

[Table 4]

From the above Table 4, when there is no group judgment,
When the correct answer rate of 86.3% is used, when the group judgment to which the neural network by the φ-qt pattern is added as in the present embodiment is used, the total is 96.8.
%, The judgment accuracy is greatly improved. Example 3
The correct answer rate is higher than that of.

Incidentally, as described above, the voltage phase, the partial discharge signal and the synchronous noise appear as shown in FIG. 14, for example, and the partial discharge signal varies in the generated phase among the voltage phases φ1 and φ3. There is a characteristic that the generated phase has almost no variation. Therefore, when the partial discharge signal and the synchronous noise are displayed in a φ-qt pattern as shown in FIG. 13, the phase position of the synchronous noise is stable with respect to the voltage phase, but the partial discharge signal is stable with respect to the voltage phase. The positions of occurrence are randomized, and the difference between the two can be clearly recognized and distinguished.

On the other hand, since the characteristic of temporal change is lost in the φ-qn pattern, when the partial discharge signal is displayed in the φ-qn pattern, the result becomes as shown in FIG. Is difficult to understand, but the φ-qn pattern provides information necessary for discriminating the defect type and the like.
Therefore, by displaying both the φ-qn pattern and the φ-qt pattern on the display device 11, the features of the respective display methods can be utilized.

FIG. 15 shows an example of the display result. In the figure, the upper part is a φ-qn pattern, the middle part is a temporal change of the partial discharge charge amount q, the lower part is a φ-qt pattern, CH
1, CH2 and CH3 are measurement channels. Although the number of generations n and the amount of charge q are shown by shading in the figure, the number of generations n, the amount of charge q and the like can be easily grasped by changing the shading to color. Note that the type of graph displayed on the display device 11 may be changed depending on the measurement target. For example, when noise determination is performed, the charge amount-generation frequency may be displayed at the top.

As described above, by displaying at least two or more types of patterns on one screen, it is possible to determine the partial discharge with high accuracy by utilizing the features of the respective display methods. Therefore, the accuracy of the determination of the partial discharge can be further improved by using the above display method alone or in combination with the partial discharge determination method shown in the first to fourth embodiments.

[0068]

As described above, the following effects can be obtained in the present invention. (1) By using the φ-qt pattern, it is possible to create measurement signal patterns without losing the characteristic of the temporal change of the generated phase angle peculiar to partial discharge, and to use these measurement signal patterns to generate neural signals. Since the partial discharge determination is performed in the network, it is possible to easily determine the partial discharge without requiring a skilled measurer.

(2) From the measured signal and the applied potential phase signal, φ,
q and n are calculated to calculate the characteristic values for classification, and the characteristic values are classified into groups created in advance according to the purpose of the measurement, and the partial discharge determination means dedicated to each group corresponding to each group is used. Since the presence / absence of the partial discharge is determined, the partial discharge determination with high determination accuracy can be performed. Also,
The φ-q-
By adding the neural network determination based on the t pattern, the determination accuracy can be further improved,
Partial discharge determination with high accuracy can be performed.

[Brief description of the drawings]

FIG. 1 is a diagram illustrating a partial discharge determination using a φ-qt pattern of the present invention.

FIG. 2 is a diagram illustrating a partial discharge determination based on grouping of measurement signals according to the present invention.

FIG. 3 is a diagram showing a configuration of a test circuit of the present invention.

FIG. 4 is a diagram showing an example of a φ-qt pattern.

FIG. 5 is a diagram showing a configuration of a neural network according to the first embodiment of the present invention.

FIG. 6 is a diagram showing examples of partial discharge data and noise data.

FIG. 7 is a diagram showing an example of characteristic values obtained for a partial discharge.

FIG. 8 is a diagram showing an example of a characteristic value obtained for noise.

FIG. 9 is a diagram showing a tree diagram created by performing cluster classification.

FIG. 10 is a diagram showing a tree diagram (continued) created by performing cluster classification.

FIG. 11 is an outline of a partial discharge determination flow using a group determination.

FIG. 12 is an outline of a partial discharge determination flow using a group determination and a φ-qt pattern.

FIG. 13 is a diagram illustrating an example of partial discharge and synchronization noise.

FIG. 14 is a diagram illustrating a voltage phase, a signal, and noise.

FIG. 15 is a diagram illustrating a display example of two or more types of patterns.

FIG. 16 is a diagram showing a partial discharge occurrence angle with respect to a potential application phase.

FIG. 17 is a diagram illustrating an example of a φ-qn pattern.

FIG. 18 is a diagram showing an example of a partial discharge determination neural network.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 CV cable 2a, 2b Air terminal part 3 Coupling capacitor 4a, 4b Partial discharge detector 9 Simulated defect part 5 Balancer 6 Partial discharge measuring instrument 8 Potential application phase signal 7 Digital memory 10 Personal computer 11 Display device 12 Power supply transformer

 ──────────────────────────────────────────────────続 き Continuing on the front page (72) Inventor Hiroaki Tsuchiya 2-6-1 Marunouchi, Chiyoda-ku, Tokyo Inside Furukawa Electric Co., Ltd. (72) Keiichi Shibata 2-6-1 Marunouchi, Chiyoda-ku, Tokyo Inside Furukawa Electric Co., Ltd. (72) Inventor Yoko Narui 2-6-1 Marunouchi, Chiyoda-ku, Tokyo Inside Furukawa Electric Co., Ltd.

Claims (9)

[Claims] 1. A method for measuring a partial discharge of a power cable, comprising: generating a phase angle φ of a measured signal and a magnitude q of a discharge charge amount in a time-series manner for each potential application phase cycle t;
-Qt pattern is created, and the neural network, which has learned the φ-qt pattern of the partial discharge signal in advance, obtains the φ-q-
A partial discharge measuring method, comprising inputting a t-pattern and determining the presence or absence of a partial discharge signal. 2. A method for measuring a partial discharge of a power cable, comprising: determining a phase angle φ at which a signal having a magnitude equal to or greater than a predetermined discharge charge amount among measured signals is generated every application potential phase cycle t. Partial discharge measurement characterized in that a φ-t pattern is created by enumerating the series, and the above measurement signal is input to a neural network in which the φ-t pattern of the partial discharge signal has been learned in advance to determine the presence or absence of a partial discharge. Method. 3. The partial discharge measurement method according to claim 1, wherein a plurality of potential application phase cycles are combined to form a pattern. 4. A method for measuring a partial discharge of a power cable, comprising: calculating a classification characteristic value from a measurement signal and a potential application phase signal; and dividing the measurement signal into at least two or more groups based on the characteristic value. A method for measuring partial discharge, comprising classifying and performing partial discharge determination using the classification result. 5. A classification characteristic value calculated by a statistical method from at least two values out of a discharge charge amount q, a generation phase angle φ, and a generation number n as a classification characteristic value. The partial discharge measurement method according to claim 4. 6. The method according to claim 4, wherein the group is at least one type and two or more groups among three types of a partial discharge group, a noise group, and a group in which partial discharge and noise are mixed. The partial discharge measurement method according to claim 5. 7. The method according to claim 4, wherein the signal classified into the group is determined to be a partial discharge or a noise using a predetermined partial discharge determination method corresponding to the group. The partial discharge measurement method according to claim 6. 8. A predetermined partial discharge determination method includes neural network, chaos, wavelet, fuzzy,
The partial discharge measurement method according to claim 7, wherein at least one or more determination methods are used from among statistical calculations such as principal component analysis. 9. A method for measuring a partial discharge of a power cable, comprising: calculating a classification characteristic value from a measurement signal and a potential application phase signal; and dividing the measurement signal into at least two or more groups based on the characteristic value. Classify and perform a partial discharge determination using the classification result, and input a learning φ-qt pattern corresponding to the partial discharge determination result to a neural network prepared according to the partial discharge determination result. The generated phase angle φ of the measured signal and the magnitude q of the discharge charge amount are listed in time series for each potential application phase cycle t.
-Qt pattern is created, the φ-qt pattern obtained the partial discharge determination result is input to each of the neural networks, and the partial discharge determination result is performed by the neural network using the classification result. And measuring the partial discharge.