JP2019200068A - Partial discharge diagnostic device - Google Patents

Abstract

An aspect of partial discharge is determined by using partial discharge data obtained by sampling a signal generated by partial discharge. A sensor that detects a signal generated by partial discharge of an electric device, a data acquisition unit that acquires partial discharge data by sampling a signal from the sensor, and processes partial discharge data acquired by the data acquisition unit. A data processing unit is provided, and the data processing unit performs machine learning using a learning data set consisting of partial discharge data acquired for each aspect of partial discharge, and changes the aspect from the partial discharge data acquired by the data acquisition unit. A machine learning unit that generates a machine learning model for determination, and a determination unit that determines the aspect of partial discharge using the machine learning model generated by the machine learning unit and the partial discharge data acquired by the data acquisition unit. Have. [Selection diagram] Figure 1

Description

  The present invention relates to a partial discharge diagnostic apparatus.

  For example, in electrical equipment such as a gas insulated switchgear, a power transformer, and a switchboard, partial discharge usually occurs before dielectric breakdown occurs. By detecting this partial discharge, it is possible to discover insulation diagnosis, that is, a sign of dielectric breakdown.

  Since electromagnetic waves, abnormal currents, ultrasonic waves, mechanical vibrations, etc. are generated when partial discharges occur, use electromagnetic wave sensors, current sensors, ultrasonic sensors, mechanical vibration sensors, etc. to diagnose the presence or absence of partial discharges. be able to.

  For example, the partial discharge diagnosis apparatus shown in Patent Document 1 or Patent Document 2 diagnoses the presence or absence of partial discharge from a detection signal detected by an antenna that detects electromagnetic waves. And this partial discharge diagnostic apparatus removes the noise component resulting from the noise different from partial discharge, in order to diagnose correctly the presence or absence of partial discharge.

JP 2010-276365 A JP, 2006-61733, A

  However, the conventional partial discharge diagnostic device detects whether or not partial discharge has occurred. For example, whether the partial discharge is a protrusion discharge, a void discharge, or a creeping discharge (occurrence) Aspect) cannot be determined.

  On the other hand, the inventor of the present application finds that the waveform of the detection signal of the sensor varies depending on the aspect of partial discharge, and by identifying this aspect, it is possible to identify the location of occurrence of partial discharge in an electrical device. It was.

  Accordingly, the main object of the present invention is to determine the state of partial discharge using partial discharge data obtained by sampling a signal generated by partial discharge.

  That is, the partial discharge diagnostic apparatus according to the present invention includes a sensor that detects a signal generated by partial discharge of an electrical device, a data acquisition unit that samples a signal from the sensor to acquire partial discharge data, and the data acquisition unit A data processing unit for processing the partial discharge data acquired by the machine, wherein the data processing unit performs machine learning using a learning data set including partial discharge data acquired for each aspect of partial discharge, and acquires the data A machine learning unit for generating a machine learning model for discriminating an aspect from partial discharge data acquired by the unit, a machine learning model generated by the machine learning unit, and partial discharge data acquired by the data acquisition unit And a discriminator for discriminating the aspect of the partial discharge.

  If this is the case, the state of partial discharge from the partial discharge data acquired by the data acquisition unit by the machine learning model generated using the learning data set consisting of partial discharge data acquired for each aspect of partial discharge. Can be determined. As a result, it is possible to identify the location (insulation degradation location) where the partial discharge of the electrical equipment occurs.

  As a learning algorithm of the machine learning unit, a support vector machine, a self-organizing map, an artificial neural network, a decision tree, a random forest, a k-average method, a k-nearest method, a genetic algorithm, a Bayesian network, or a deep learning method should be used. Can be considered.

  As a specific embodiment of the data acquisition unit, it is conceivable that partial discharge data is acquired by sampling the signal from the sensor for at least two cycles of the commercial power supply for each of a plurality of frequencies. Here, when the partial discharge diagnostic apparatus is portable, for example, it is difficult to capture the commercial frequency from the commercial power source, that is, it is difficult to synchronize with the commercial power source, and the phases of the sampling data acquired for each of the plurality of frequencies may be shifted from each other. is there. For this reason, it is desirable that the data processing unit further includes a phase correction unit that corrects the phase of the sampling data acquired for each of the plurality of frequencies.

  In order to easily determine the state of partial discharge, the data processing unit further includes a feature amount extraction unit that extracts a feature amount of the partial discharge data, and the machine learning unit includes the learning data set as the learning data set. A machine learning model is generated using a feature amount of given partial discharge data, and the determination unit uses the machine learning model and the feature amount obtained by the feature amount extraction unit to perform the partial discharge. It is desirable to distinguish the aspect of

  Specifically, the feature quantity extraction unit performs singular value decomposition on the partial discharge data, and extracts the right singular vector (singular vector representing the characteristics of the time domain) of the maximum singular value obtained as a feature quantity. The machine learning unit may generate a machine learning model using a right singular vector of partial discharge data given as the learning data set.

Further, in order to speed up the arithmetic processing, the feature amount extraction unit performs singular value decomposition on the partial discharge data, and uses a feature vector obtained by performing Fourier transform on the right singular vector of the maximum singular value obtained as a feature amount. Preferably, the machine learning unit generates the machine learning model using a feature vector obtained by Fourier transforming a right singular vector of partial discharge data given as the learning data set.
In addition, the characteristic of periodicity can be clarified by Fourier transforming the right singular vector. Further, adjustment of time 0 at each measurement can be omitted. That is, the phase difference from the commercial frequency can be ignored.

  In order to determine a typical partial discharge aspect of an electric device, the machine learning unit generates a normal model, a protruding discharge model, a void discharge model, and a creeping discharge model as the machine learning model. Can be considered.

  According to the present invention configured as described above, it is possible to determine the aspect of partial discharge using partial discharge data obtained by sampling a signal generated by partial discharge.

It is a figure which shows typically the structure of the partial discharge diagnostic apparatus of this embodiment. It is a figure which shows the content of the partial discharge data of the embodiment. It is a three-dimensional graph of the partial discharge data of the same embodiment. It is a figure which shows the partial discharge data before and behind the phase correction of the embodiment. It is a graph which shows the right singular vector of the maximum singular value of the embodiment, and a feature vector. It is a schematic diagram which shows the machine learning by SVM of the embodiment. (A) three-dimensional graph of partial discharge data when normal, (b) partial discharge data after phase correction, (c) right singular vector data of maximum singular value, (d) feature vector data after Fourier transform FIG. (A) three-dimensional graph of partial discharge data in case of projecting discharge, (b) partial discharge data after phase correction, (c) right singular vector data of maximum singular value, (d) feature after Fourier transform It is a figure which shows vector data. (A) three-dimensional graph of partial discharge data in case of void discharge, (b) partial discharge data after phase correction, (c) right singular vector data of maximum singular value, (d) feature after Fourier transform It is a figure which shows vector data. (A) three-dimensional graph of partial discharge data for creeping discharge, (b) partial discharge data after phase correction, (c) right singular vector data of maximum singular value, (d) feature vector after Fourier transform It is a figure which shows data. It is a schematic diagram which shows the class discrimination | determination of the embodiment. It is a figure which shows the modification of the phase correction method between frequencies. It is a figure which shows the modification of the phase correction method between frequencies.

  Hereinafter, an embodiment of a partial discharge diagnostic apparatus according to the present invention will be described with reference to the drawings.

  A partial discharge diagnostic apparatus 100 according to the present embodiment shown in FIG. 1 diagnoses the presence and state (occurrence mode) of partial discharge X that can occur when an insulation abnormality occurs in an electrical device 200 such as a gas insulated switchgear, a power transformer, or a switchboard. It is a portable device.

  Specifically, the partial discharge diagnostic apparatus 100 detects an electromagnetic wave signal as a signal generated by the partial discharge X. As shown in FIG. 1, the electromagnetic discharge sensor 2 that detects the electromagnetic wave generated by the partial discharge X, and the electromagnetic wave A data acquisition unit 3 that acquires a partial discharge data by sampling a signal from the sensor 2 and a data processing unit 4 that processes the partial discharge data acquired by the data acquisition unit 3 are provided.

  The electromagnetic wave sensor 2 of the present embodiment is an electromagnetic wave antenna, and receives an electromagnetic wave and outputs a detection signal (for example, a voltage signal) proportional to the intensity to the data acquisition unit 3.

  The data acquisition unit 3 includes, for example, a spectrum analyzer, an A / D converter, a CPU, an internal memory, and the like. The data acquisition unit 3 acquires partial discharge data by sampling the detection signal of the electromagnetic wave sensor 2 for each of a plurality of frequencies at a time corresponding to at least two cycles of a commercial frequency (50 Hz or 60 Hz).

  Specifically, as shown in FIG. 2, the data acquisition unit 3 performs, for example, 100 sampling (t1) in a predetermined frequency band (for example, 1 to 100 MHz) in a time corresponding to two periods (commercial two cycles) of the commercial frequency every 1 MHz. To t100). When the commercial frequency is 50 Hz, the data acquisition unit 3 performs sampling every 0.4 [ms], and when the commercial frequency is 60 Hz, the data acquisition unit 3 performs sampling every 0.333 [ms]. Data acquired in each sampling is an amplitude value of the detection signal. In this embodiment, since it is the portable partial discharge diagnostic apparatus 100, without taking in a commercial frequency signal from a commercial power source, if the commercial frequency is 50 Hz, two commercial frequencies are set to 40 [ms], and the commercial frequency is If it is 60 Hz, 2 cycles of the commercial frequency are set to 33.3 [ms].

  The partial discharge data acquired by the data acquisition unit 3 is stored in the data storage unit 5 configured by an internal memory as waveform data for each frequency for a plurality of frequencies (f1 to f100). This partial discharge data is shown in FIG. 3 when it is three-dimensionally represented by three axes of a frequency axis, a time axis (that is, a sampling axis), and an intensity axis.

  A data processing unit 4 shown in FIG. 1 includes a CPU, an internal memory, and the like. The data processing unit 4 includes a phase correction unit 4a that corrects a phase shift between frequencies in the partial discharge data stored in the data storage unit 5, and a feature amount extraction unit that extracts a feature amount of the phase-corrected partial discharge data. 4b, a machine learning unit 4c that generates a machine learning model for determining the aspect of the partial discharge X, and a determination unit that determines the aspect of the partial discharge X using the machine learning model generated by the machine learning unit 4c. 4d.

<Phase correction between frequencies>
The phase correction unit 4a corrects the phase of the sampling data acquired for each of a plurality of frequencies, and corrects the phase shift by the following method.

  First, as a preparation for calculating a cross-correlation value, the respective intermediate values are subtracted for each of a plurality of frequencies (f1 to f100), and the intermediate values are made zero.

  A frequency component (f) having the strongest correlation with other frequencies for each frequency is selected from f1 to f100 and determined as a key frequency component (fkey).

  Then, with reference to fkey, a cross-correlation value between fkey and f1 (100 correlations of fkey and 100 correlations of f1) is obtained.

  Specifically, the element of f1 is shifted by 1 (t is shifted by 1), and the cross-correlation value between fkey and f1 is obtained. This is performed in all (100 times), and the number of shifts at which the cross-correlation value is maximized is obtained as the shift amount of sampling data.

  The same operation is performed for f2 to f100, and the shift amount at each frequency is obtained and set as the shift amount of the sampling data.

  Then, the phase is corrected by shifting each sampling data by the shift amount obtained at each frequency. The partial discharge data before and after such phase correction are shown in FIGS. 4 (a) and 4 (b). From the figure, it can be seen that phase correction between frequencies has been performed.

<Feature extraction>
The feature amount extraction unit 4b extracts feature amounts from the phase-corrected partial discharge data by the following method.

  First, singular value decomposition is performed on the partial discharge data after phase correction.

  Then, the right singular vector having the maximum singular value (singular vector representing the characteristics of the time domain, see FIG. 5A) is obtained. This right singular vector is 100-dimensional.

  This singular vector is Fourier-transformed and its absolute value is used as a feature vector (see FIG. 5B). This feature vector is 51 dimensional. This feature vector is used as the feature amount of the partial discharge data.

<Machine learning with SVM>
As shown in FIG. 6, the machine learning unit 4 c performs machine learning using a learning data set including partial discharge data (100 × 100 points of data) acquired for each aspect of the partial discharge X, and performs a machine learning model. Is generated. The machine learning unit 4c of the present embodiment uses a support vector machine (SVM) as a machine learning algorithm.

  The machine learning model generated by the machine learning unit 4c includes an individual discriminator (SVM # 1) for discriminating “normal” and an individual discriminator (SVM # 2) for discriminating “protrusive discharge”. , An individual discriminator (SVM # 3) for discriminating “void discharge” and an individual discriminator (SVM # 4) for discriminating “other discharge (creeping discharge)”.

  Therefore, the learning data set input to the machine learning unit 4c includes partial discharge data when “normal” and label data indicating “normal”, partial discharge data when “protrusive discharge”, and “ Label data indicating “protrusive discharge”, partial discharge data in the case of “void discharge”, label data indicating “void discharge”, partial discharge data in case of “creeping discharge”, and “creeping discharge” It is label data which shows.

  The partial discharge data (learning data set) is phase-corrected by the above-described phase correction unit 4a, and a feature vector is extracted by the feature amount extraction unit 4b. FIGS. 7 to 10 show (a) a three-dimensional graph of partial discharge data in each aspect, (b) partial discharge data after phase correction, (c) right singular vector data of the maximum singular value, and (d) Fourier transform. The latter feature vector data is shown.

  Then, the extracted feature vector is input to the machine learning unit 4c. Then, in the machine learning unit 4c, each feature vector and label data indicating its aspect are learned, and individual discriminators (SVM # 1 to SVM # 4) are generated.

<Difference of measurement data (partial discharge data)>
As illustrated in FIG. 11, the determination unit 4 d determines the aspect of the partial discharge X using the individual determination devices (SVM # 1 to SVM # 4) generated by the machine learning unit 4c.

  The determination unit 4d acquires the feature vector of the partial discharge data acquired by the data acquisition unit 3 from the feature amount extraction unit 4c. Then, the discriminating unit 4d inputs the acquired feature vector of the partial discharge data to each individual discriminator (SVM # 1 to SVM # 4) and obtains the individual score having the maximum score (score # 1 to score # 4). The label of the discriminator (SVM) is selected.

  Based on the selected labels (“normal”, “protrusive discharge”, “void discharge”, “creeping discharge”), the determination unit 4 d can determine the aspect of the partial discharge X.

<Effect of this embodiment>
According to the partial discharge diagnostic apparatus 100 of the present embodiment, the machine learning models (SVM # 1 to SVM # 4) generated using the learning data set including the partial discharge data acquired for each aspect of the partial discharge X. The aspect of the partial discharge X can be determined from the partial discharge data acquired by the data acquisition unit 3. As a result, the occurrence location (insulation degradation location) of the partial discharge X of the electric device 200 can be specified.

<Other modified embodiments>
The present invention is not limited to the above embodiment.

  In the embodiment, the electromagnetic wave is detected as a signal generated by the partial discharge. However, a current, ultrasonic wave, or mechanical vibration generated by the partial discharge may be detected. In this case, the partial discharge diagnostic apparatus includes a current sensor such as a Hall element or MR element, an ultrasonic sensor such as an electromagnetic microphone or a piezoelectric microphone, or a mechanical vibration sensor such as a piezoelectric ceramic element or a piezoelectric crystal element.

  Further, in the above-described embodiment, the aspect of partial discharge is determined to be normal, protruding discharge, void discharge, and creeping discharge, but other aspects are determined depending on the situation. May be.

In addition to the phase correction method between frequencies of the above-described embodiment, for example, the following method is conceivable as shown in FIG.
It is assumed that the phase gradient is a straight line, taking advantage of the fact that the phase shift of the frequencies (f1 to f100) is often constant (monotonic).
The time shift amount of f100 is changed by 1 with f1 being fixed, and the frequency between f1 and f100 is a time shift amount that is proportionally distributed. Then, the vertical standard deviation is taken with respect to the result after each time shift, and the time shift amount having the smallest standard deviation (small variation) is adopted. Then, phase correction is performed using this time shift amount.

Further, as shown in FIG. 13, the following method is also conceivable.
A peak position is found for each graph in the time axis direction of data of frequency (f1 to f100). Here, as a method of finding the peak position, a method of differentiating twice in the time axis direction to obtain the difference between the minimum value and the adjacent maximum value and certifying the minimum value having the large difference as the peak position can be considered. . Data of frequencies for which no peak exists at this point is excluded from the following calculation.
An approximate curve is drawn using, for example, the top 30 peaks. Using that line as a correction amount, phase correction is performed.
In the above, 100 × 100 pieces of data may be subjected to a smoothing process for smoothing irregularities by, for example, an envelope creation process or a moving average process.

  Furthermore, in the above-described embodiment, the aspect is determined using the feature vector that is an absolute value after Fourier transform as the feature amount, but the aspect is determined using the right singular vector having the maximum singular value without performing Fourier transform. You may comprise as follows.

  In addition, although the machine learning algorithm of the embodiment is SVM, other than that, a self-organizing map, an artificial neural network, a decision tree, a random forest, a k-average method, a k-nearest method, a genetic algorithm, a Bayesian network, or It may be a deep learning method. For example, when a self-organizing map is used, it is conceivable to use a self-organizing map data in which a label of an aspect is displayed in color.

  In addition, it goes without saying that the present invention is not limited to the above-described embodiment, and various modifications can be made without departing from the spirit of the present invention.

DESCRIPTION OF SYMBOLS 100 ... Partial discharge diagnostic apparatus 200 ... Electrical equipment X ... Partial discharge 2 ... Sensor 3 ... Data acquisition part 4 ... Data processing part 4a ... Phase correction part 4b ... Feature amount extraction unit 4c... Machine learning unit 4d.

Claims (7)

A sensor for detecting a signal generated by partial discharge of an electrical device;
A data acquisition unit that samples a signal from the sensor to acquire partial discharge data;
A data processing unit for processing the partial discharge data acquired by the data acquisition unit,
The data processing unit
Machine that performs machine learning using a learning data set including partial discharge data acquired for each aspect of partial discharge, and generates a machine learning model for discriminating an aspect from partial discharge data acquired by the data acquisition unit The learning department,
The partial discharge diagnostic apparatus which has a discrimination | determination part which discriminate | determines the aspect of the said partial discharge using the machine learning model produced | generated by the said machine learning part, and the partial discharge data which the said data acquisition part acquired.   The machine learning unit is configured using a support vector machine, a self-organizing map, an artificial neural network, a decision tree, a random forest, a k-average method, a k-nearest neighbor method, a genetic algorithm, a Bayesian network, or a deep learning method. The partial discharge diagnostic apparatus according to claim 1. The data acquisition unit acquires partial discharge data by sampling a signal from the sensor for at least two cycles of a commercial power supply for each of a plurality of frequencies.
The partial discharge diagnosis apparatus according to claim 1, wherein the data processing unit further includes a phase correction unit that corrects a phase of sampling data acquired for each of the plurality of frequencies. The data processing unit further includes a feature amount extraction unit that extracts a feature amount of the partial discharge data,
The machine learning unit generates a machine learning model using a feature amount of partial discharge data given as the learning data set,
The partial discharge diagnosis according to any one of claims 1 to 3, wherein the determination unit determines the aspect of the partial discharge using the machine learning model and the feature amount obtained by the feature amount extraction unit. apparatus. The feature amount extraction unit performs singular value decomposition on the partial discharge data, and extracts the right singular vector of the maximum singular value obtained thereby as a feature amount,
The partial discharge diagnosis apparatus according to claim 4, wherein the machine learning unit generates a machine learning model using a right singular vector of partial discharge data given as the learning data set. The feature amount extraction unit performs singular value decomposition on the partial discharge data, and extracts a feature vector obtained by Fourier-transforming the right singular vector of the maximum singular value obtained as a feature amount,
The partial discharge according to claim 4 or 5, wherein the machine learning unit generates the machine learning model using a feature vector obtained by Fourier-transforming a right singular vector of partial discharge data given as the learning data set. Diagnostic device. The partial discharge diagnosis apparatus according to claim 1, wherein the machine learning unit generates a normal model, a protruding discharge model, a void discharge model, and a creeping discharge model as the machine learning model. .