Partial-discharge analysis system

Abstract

Provided is a technology for expanding PRPD-pattern teaching data in partial-discharge measurement. This partial-discharge analysis system carries out analyses of partial discharge incidents in electrical equipment, and is provided with a control unit and a partial-discharge detection sensor. The control unit generates a PRPD pattern on the basis of input signals obtained by the partial-discharge detection sensor, and inputs the PRPD pattern as a two-dimensional image into an analysis model for analysis, thereby diagnosing the presence of partial-discharge incidents. Teaching data utilized in the analysis model contains, as a base image pixel-displayed as a two-dimensional image, a PRPD pattern obtained from past partial discharges, and contains as teaching data an image in which the base image has been shifted up / down, has been stretched up / down, or has been composited with a noise image.

Description

Partial Discharge Analysis System

[0001] The present invention relates to a partial discharge analysis system for analyzing the occurrence of partial discharge in electrical equipment.

[0002] As electrical equipment ages, the insulating performance of the insulators on the surface or inside the equipment deteriorates. When the insulating performance deteriorates, partial discharge occurs at the deteriorated area. If the deterioration of the insulating performance progresses further, the electrical equipment will suffer a breakdown. This breakdown can lead to serious accidents such as ground faults. For this reason, when maintaining electrical equipment, it is common to diagnose the deterioration state of the insulating performance of electrical equipment.

[0003] For example, Patent Document 1 discloses a degradation diagnosis system, a degradation diagnosis device, a degradation diagnosis method, and a computer program that can diagnose the degradation state of the insulation performance of electrical equipment. This technology determines the degradation state of electrical equipment based on degradation information from past cases of partial discharge occurrence.

[0004] Furthermore, Patent Document 2 discloses a partial discharge determination device and method that can reliably determine the progression of partial discharge occurring in an underground power transmission cable. This technology determines the progression of partial discharge occurring in the power transmission cable by analyzing and evaluating the distribution pattern of combinations of the charge amount and occurrence phase angle of partial discharge within a predetermined period.

[0005] Partial discharge measurement is also used as a method for diagnosing power equipment such as cables, distribution panels, and transformers in substations. There are various methods for measuring partial discharge, such as ultrasonic measurement and the residual charge method. Among these, there is a method that uses a transient earth voltage sensor to measure partial discharge without causing a power outage at the substation or opening the distribution panel. Because the measurement is safe and easy, this technique is used for partial discharge diagnosis in substations around the world.

[0006] Similarly, there is a method that uses a PRPD (phase resolved partial discharge) pattern to check for the occurrence of partial discharge. The PRPD pattern compares the power supply phase with the partial discharge pulse and visualizes the partial discharge signal by synchronizing it with the power supply sine waveform. This is a technology that can determine the occurrence of partial discharge by looking at the PRPD pattern.

[0007] JP 2020-085675 A JP 2021-117150 A

[0008] However, the conventional techniques have the following problems: When attempting to automatically diagnose insulation deterioration of power equipment due to partial discharge, as with the techniques disclosed in Patent Documents 1 and 2, past measurement data of specific equipment is required, and abnormalities can only be detected by comparing with that data.

[0009] Furthermore, the technology of determining whether partial discharges are occurring by observing PRPD patterns requires skilled techniques, knowledge, and know-how. Therefore, it is not an easy method that anyone can perform. Therefore, the most common method of partial discharge diagnosis worldwide is to measure partial discharges using a transient ground voltage sensor or a high-frequency voltage sensor and plot partial discharge waveforms using PRPD patterns. In recent years, partial discharge analysis technologies using AI (artificial intelligence) have begun to emerge, and active development of AI has been underway using measurement data from the past as training data. However, acquiring PRPD pattern data from power equipment where partial discharges are occurring is the same as performing measurements on power equipment where insulation breakdown is beginning to occur. Therefore, acquiring PRPD pattern data from power equipment where partial discharges are occurring is dangerous and difficult. To solve this problem, a method for expanding training data and building an analysis model using AI is needed.

[0010] The present invention has been made in consideration of the above circumstances, and aims to provide a technique for expanding training data related to PRPD patterns in partial discharge measurement.

[0011] Since electrical equipment that shows signs of deterioration is immediately repaired or replaced, it is difficult to collect partial discharge data from operating equipment, and the long time it takes to collect partial discharge data makes it difficult to accumulate training data. The inventors studied how to develop an AI (artificial intelligence) model built using supervised learning from a small amount of partial discharge data, and came up with the present invention.

[0012] The partial discharge analysis system of the present invention, which advantageously solves the above-mentioned problems, is a partial discharge analysis system for analyzing the occurrence of partial discharge in electrical equipment, and includes a control unit and a partial discharge detection sensor. The control unit generates a PRPD pattern based on an input signal obtained by the partial discharge detection sensor, and inputs the PRPD pattern as a two-dimensional image into an analysis model for analysis to diagnose the presence or absence of partial discharge. The training data used in the analysis model includes a base image in which a PRPD pattern obtained by a past partial discharge is displayed as a two-dimensional pixel image. The analysis is characterized in that the training data includes a combination of one or more of the following: (A) the training data includes a plurality of derived images obtained by shifting the PRPD pattern included in the base image by a predetermined pixel amount in the Y-axis direction; (B) the training data includes a plurality of derived images obtained by expanding or contracting the PRPD pattern included in the base image in the Y-axis direction; and (C) the training data includes a composite image obtained by combining an image in which the PRPD pattern included in the base image is made transparent at a predetermined density and an image in which a two-dimensional image of a PRPD pattern showing noise generated depending on the environment around the electrical equipment is made transparent at a predetermined density.

[0013] Furthermore, the partial discharge analysis system according to the present invention preferably has the following solutions: (1) further comprising a switch device provided between the control unit and a plurality of partial discharge detection sensors provided in each of a plurality of pieces of electrical equipment, and the switch device periodically or irregularly changes one of the plurality of partial discharge detection sensors to be connected to the control unit; (2) further comprising a display unit capable of displaying the diagnosis results of the control unit, and when the control unit diagnoses the occurrence of partial discharge, the display unit displays at least one partial discharge type that has the highest probability of being applicable out of a plurality of pre-set partial discharge types.

[0014] According to the present invention, it is now possible to sufficiently expand the training data for constructing an AI model that can accurately determine the presence and type of discharge. In addition, it is now possible to construct an AI model that takes into account offset voltage in order to remove noise whose magnitude varies depending on the substation.

[0015] Furthermore, if the partial discharge signal used as training data is weak, the training data waveform drawn on the PRPD pattern will be small, making it impossible to acquire training data with a large partial discharge signal. Conversely, if the partial discharge signal used as training data is large, the training data waveform drawn on the PRPD pattern will be large, making it impossible to acquire a weak partial discharge signal. By expanding or contracting the training data, training data for a weak partial discharge signal can be generated from a large discharge signal, and training data for a large partial discharge signal can be generated from a weak partial discharge signal, which has the effect of making it easy to expand the training data.

[0016] When measuring partial discharges, not only partial discharge data but also noise generated from surrounding equipment is acquired at the same time. If the training data consists only of partial discharge data, there is a risk that the AI ​​model will be affected by noise after it is built, resulting in a decrease in analysis accuracy. This not only generates new training data by combining noise with partial discharge data, but also has the effect of creating training data that is resistant to noise.

[0017] 1 is a schematic diagram of a partial discharge analysis system according to one embodiment of the present invention;

[0023] FIG. 1 is a functional configuration diagram of the partial discharge analysis system according to the embodiment;

[0024] FIG. 1 is an image diagram of an example of a raw waveform of a partial discharge signal converted into a PRPD pattern;

[0025] FIG. 1 is an image diagram of the image of FIG. 3(a) shifted in the positive direction of the Y-axis (vertical axis); and

[0026] FIG. 1 is an image diagram of the image of FIG. 3(a) stretched in the Y-axis direction.

[0027] FIG. 1 is a schematic diagram of the partial discharge analysis system according to one embodiment of the present invention;

[0028] FIG. 2 is an image diagram of the image of FIG. 3(a) stretched in the Y-axis direction;

[0029] FIG. 3(b) is an image diagram of the image of FIG. 3(a) stretched in the Y-axis direction;

[0029] FIG. 1 is an image diagram of the image of FIG. 3(a) shifted in the positive direction of the Y-axis (vertical axis);

[0029] FIG. 2 is an image diagram of the image of FIG. 3(a) shifted in the negative direction of the Y-axis;

[0029] FIG. 3 is an image diagram of the image of FIG. 3(a) stretched in the Y-axis direction;

[0029] FIG. 1 is an image diagram of the image of FIG. 3(a) shifted in the positive direction of the Y-axis (vertical axis);

[0029] FIG. 2 is an image diagram of the image of FIG. 3(a) shifted in the positive direction of the Y-axis (vertical axis);

[0029] FIG. 3 is an image diagram of the image of FIG. 3(a) shifted in the negative direction of the Y-axis;

[0029] FIG. 3 is an image diagram of the image of FIG. 3( FIG. 10 is a diagram showing an example of obtaining a PRPD pattern for multiple periods of the power supply frequency using a PRPD pattern image of 12 pixels by 12 pixels.

[0018] Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. The following embodiments are merely examples of devices and methods for embodying the technical concept of the present invention, and are not intended to limit the configuration to those described below. In other words, the technical concept of the present invention can be modified in various ways within the technical scope defined in the claims.

[0019] FIG. 1 is a schematic diagram of a partial discharge analysis system according to one embodiment of the present invention. A partial discharge 5 is a predictive signal that occurs in a power facility 100 and indicates deterioration of the insulation performance of the power facility. The power facility 100 includes a transformer 101, a power cable 102, a switchboard 103, and devices within the switchboard. A partial discharge signal 5A generated from the power facility 100 is acquired by a detection sensor 3, and the data is collected via a coaxial cable to a control unit 2 of a partial discharge analysis system 1 installed in a substation. The control unit 2 converts the partial discharge waveform into a PRPD pattern, and an analysis unit within the control unit 2 analyzes the type of partial discharge. The partial discharge signal 5A and the analysis results are transmitted to a data server 4. The data server 4 stores the partial discharge data and analysis results thus far.

[0020] 2 is a functional configuration diagram of the partial discharge analysis system 1 according to the above embodiment. The partial discharge data acquisition unit S1 refers to the detection sensor 3 that acquires the partial discharge signal 5A. In the control unit 2 of this system, for example, a maximum of 241 sensors can be attached to the partial discharge signal input expansion unit S21. In other words, if a controller and digitizer are installed in an 18-slot chassis, the remaining 16 slots can be equipped with switch modules. By stringing the switch modules together, a total of 15 x 16 + 1 = 241 sensors can be connected.

[0021] The partial discharge signal input expansion unit S21 is composed of a channel box and two switch modules in the PXI module. The channel box can connect, for example, 31 detection sensors 3 that can be connected with BNC plugs, i.e., coaxial connectors. These can then be grounded collectively. The two switch modules mounted on the PXI module can then be used to switch the detection sensor 3 to be measured at any time.

[0022] The partial discharge data calculation unit S22 analyzes the partial discharge signal 5A collected by the partial discharge data acquisition unit S1 in terms of time and voltage using an oscilloscope installed in the PXI module, and plots the PRPD pattern. The method for plotting the PRPD pattern is not particularly limited. Meanwhile, since the reference signal for one cycle is acquired from the commercial power supply for operating the PXI module, it is preferable to continuously overlay partial discharge signals acquired every cycle, for example, every 1 s / 50 or 1 s / 60, to represent them as a frequency distribution, i.e., a single histogram, and express them as a single partial discharge image. For example, FIG. 9 shows an example of plotting a PRPD pattern from a partial discharge signal. The vertical axis (Y axis) represents the discharge potential, and the horizontal axis (X axis) represents the time until one power cycle. In FIG. 9, the vertical and horizontal axes of each PRPD pattern are each 12 pixels by 12 pixels. FIGS. 9(a), (b), and (c) respectively represent the frequency of partial discharge signals in the first cycle, the next cycle, and the third cycle. The offset was set to 2 pixels on the vertical axis. Figure 9(d) shows the total frequency per second in each square.

[0023] In the AI ​​model partial discharge analysis unit S23, the PRPD pattern acquired in the partial discharge data calculation unit S22 can be classified into a predetermined partial discharge type by applying it to an AI model that has been trained in advance to learn partial discharge types by machine learning. For example, partial discharge types can be exemplified as "no discharge," "corona discharge," "creeping discharge," "void discharge," etc. Furthermore, the analysis results by the AI ​​model can be output as a probability distribution as shown in FIG. 6.

[0024] The partial discharge analysis data transmission unit S24 transmits the PRPD pattern 7 acquired by the partial discharge data calculation unit S22 and the probability distribution 8 output by the partial discharge analysis unit using an AI model to the data server 4 serving as the partial discharge analysis data storage unit S3.

[0025] The data server 4, which is the partial discharge analysis data storage unit S3, stores the data received from the partial discharge analysis data transmission unit S24 as past data. In the example of Fig. 2, the control unit S2 includes an analysis unit consisting of the partial discharge analysis unit S23 using an AI model and the partial discharge analysis data transmission unit S24, but the analysis unit can also be configured independently or attached to the data server 4.

[0026] 3A and 3B are diagrams showing an example of extending training data in the partial discharge analysis system according to this embodiment. FIG. 3A shows training data obtained by directly converting the waveform of a partial discharge signal 5A into a PRPD pattern. The image is 64 pixels long and 64 pixels wide. As described above, the vertical axis (Y axis) represents the discharge potential, and the horizontal axis (X axis) represents the time required for one power supply cycle. In FIG. 3A, signals below a predetermined potential are recognized as noise, are offset and removed, and are not plotted.

[0027] Figure 3(b) shows the image of Figure 3(a) shifted upward by a predetermined amount to create training data. Figure 3(c) shows the image of Figure 3(a) shifted downward by a predetermined amount to create training data. In other words, shifting the image up and down on the vertical axis allows for different offset voltage settings to be used to obtain training data with noise removed.

[0028] Figure 4 shows a method for using an image obtained by stretching the image in Figure 3(a) in the vertical direction as training data. Figure 4(a) is the same original image as the image in Figure 3(a). Figure 4(b) shows the image stretched in the vertical direction by a predetermined number of degrees to be used as training data. By stretching the training data up and down in the vertical direction, weak partial discharge signals can be converted into strong partial discharge signals, and strong partial discharge signals can be converted into weak partial discharge signals, thereby obtaining new training data.

[0029] Figure 5 shows a method for synthesizing an example of the raw waveform of a partial discharge signal and an example of noise. Figures 5(a1) and 5(a2) show examples of converting the raw waveform of a partial discharge signal and noise into PRPD patterns, respectively. Figure 5(b1) shows an image obtained by converting the PRPD pattern of Figure 5(a1) with 20% transparency. Figure 5(b2) shows an image obtained by converting the PRPD pattern of Figure 5(a2) with 80% transparency. Figure 5(b3) shows an image obtained by superimposing the images of Figure 5(b1) and Figure 5(b2) as training data. Figures 5(c1) to 5(c3) and (d1) to 5(d3) also show methods for superimposing images with different transparency levels. By generating new training data by combining partial discharge data and noise data, an AI model capable of noise-resistant partial discharge identification can be constructed.

[0030] Figure 6 shows an example of the settings screen for capturing partial discharge signals and the image display of the control unit 2 outputting the acquired data in real time. Figure 6(a) shows the parameter settings for partial discharge measurement, and Figure 6(b) shows the actual measured data. The partial discharge waveform 6 acquired by the detection sensor 3 is converted into a PRPD pattern 7, and a probability distribution 8 for each type of discharge is calculated. The partial discharge waveform 6, PRPD pattern 7, probability distribution 8, and the discharge type 9 with the highest probability are displayed. In the example of Figure 6, the discharge type 9 is displayed as "normal," meaning there is no partial discharge.

[0031] Fig. 7 shows an example of a display output using the partial discharge analysis system according to this embodiment. PRPD pattern data is transmitted to the data server 4, and analysis can also be performed using the AI ​​model stored in the data server 4. Fig. 7(a) shows the latest analyzed PRPD pattern 7, and Fig. 7(b) shows the immediately preceding PRPD pattern 7. Fig. 7(d) is a bar graph showing the calculated probability of each partial discharge type by analyzing Fig. 7(a), expressed as a percentage, and Fig. 7(c) is a graph showing the time progression of the probability for each partial discharge type.

[0032] Figure 8 shows the analysis accuracy of the partial discharge analysis system in the form of a confusion matrix. The vertical columns indicate the correct classes, and the horizontal rows indicate the classes predicted by the AI ​​model. The stronger the diagonal elements, the higher the analysis accuracy of the AI ​​model. Table 1 shows an example of the analysis results of the partial discharge analysis system.

[0033]

[0034] The overall accuracy of partial discharge analysis was 98.7%. Precision in Table 1 is an index showing "how many of the predicted values ​​were correct." It is effective in verifying the accuracy of the predicted values. Precision does not take into account the incorrect judgments that were actually correct. Recall is an index that complements Precision. Recall indicates "how many of the actual correct values ​​were predicted to be correct." Precision and Recall have a mutually beneficial relationship. In this example, we believe that Precision should be given the highest priority. The reason for this is that we believe it is absolutely necessary to avoid judging that there is no abnormality even when partial discharge is occurring. This corresponds to the Precision for "No Discharge" in Table 1. With the AI ​​model used in this example, there were zero incorrect judgments for the Precision for "No Discharge." From this, it can be said that the AI ​​model used in this example has sufficient performance to confirm the presence or absence of partial discharge. F-measure is the harmonic mean of Precision and Recall. This is expressed by the following equation 1. F-measure is an index that represents the reliability of the AI ​​model. Since F-measure is calculated using the values ​​of Precision and Recall, if either value is extremely low, F-measure will also show a low value. If F-measure is low, even if either Precision or Recall is high, the AI ​​model will have low reliability. Since the F-measure of the AI ​​model used in this example was all 97% or higher, it can be said to be sufficiently reliable. The analysis accuracy was calculated by randomly selecting data that served as training data and performing inference on the constructed AI model.

[0035]

[0036] According to the present invention, it is possible to expand partial discharge data, which can contribute to the development of partial discharge analysis AI models and the like, and is of great industrial value.

[0037] 100 Power equipment 101 Transformer 102 Power cable 103 Distribution board 1 (Partial discharge) analysis system 2 Control unit 3 (Partial discharge) detection sensor 4 Data server 5 Partial discharge 5A (Partial discharge) signal 6 Partial discharge waveform 7 PRPD pattern 8 Probability distribution 9 (Most probable) discharge type

Claims

1. A partial discharge analysis system for analyzing the occurrence of partial discharge in electrical equipment, comprising a control unit and a partial discharge detection sensor, wherein the control unit generates a PRPD pattern based on an input signal obtained by the partial discharge detection sensor, and inputs the PRPD pattern as a two-dimensional image into an analysis model for analysis to diagnose the presence or absence of partial discharge, and training data used in the analysis model includes a base image in which a PRPD pattern obtained from a past partial discharge is displayed in pixels as a two-dimensional image, and the system performs analysis by combining one or more of the following: (A) the training data includes a plurality of derived images obtained by shifting the PRPD pattern included in the base image by a predetermined pixel amount in the Y-axis direction, (B) the training data includes a plurality of derived images obtained by expanding or contracting the PRPD pattern included in the base image in the Y-axis direction, and (C) the training data includes a composite image obtained by combining an image in which the PRPD pattern included in the base image is made transparent at a predetermined density and an image in which a two-dimensional image of a PRPD pattern showing noise generated depending on the environment around the electrical equipment is made transparent at a predetermined density.

2. The partial discharge analysis system according to claim 1, further comprising a switch device provided between the control unit and a plurality of partial discharge detection sensors provided in each of a plurality of pieces of electrical equipment, wherein the switch device periodically or irregularly changes one of the plurality of partial discharge detection sensors to be connected to the control unit.

3. The partial discharge analysis system according to claim 1 or 2, further comprising a display unit capable of displaying the diagnosis results from said control unit, and when said control unit diagnoses the occurrence of a partial discharge, at least one partial discharge type that has the highest probability of being applicable out of a plurality of pre-set partial discharge types is displayed on said display unit.