GB2431726A

GB2431726A - Identification of partial discharge using a neural network

Info

Publication number: GB2431726A
Application number: GB0600677A
Authority: GB
Inventors: Sun Geun Goo; Ki Jun Park; Jin Yul Yoon; Joo Sik Kwak
Original assignee: Korea Electric Power Corp
Current assignee: Korea Electric Power Corp
Priority date: 2005-10-27
Filing date: 2006-01-13
Publication date: 2007-05-02
Anticipated expiration: 2026-01-13
Also published as: DE102006008482B4; GB2431726B; KR20070045558A; JP2007124880A; KR100729107B1; DE102006008482A1; JP4244353B2; GB0600677D0

Abstract

Partial discharge signals from a partial discharge detector (fig 1, 130) coupled to high voltage power machinery (fig 1, 100) are automatically identified using a neural network. The partial discharge identifying method includes: plotting a graph of successive discharges, D n: D n-1, that is phase of a detected discharge against phase of the preceding detected discharge; transforming the graph; extracting a phase dependant summation and a phase independent summation; and inputting both of these to a neural network engine. The phase independent summation may be compared with phase dependant summation after phase shifting the phase dependant summation by 120 degrees or 240 degrees or other phases generated by training of the neural network, and may involve multiplication, and integration of the result. The result of such phase shifting may further be cross correlated with reference phase dependent summation and reference phase independent summation and then reasoning that where the correlation is a maximum is a phase of power applied to a point where a discharge signal is generated. Partial discharges can be identified without detecting phase information of the applied power. Partial discharges may be detected in three-phase machinery, such as transformers, gas insulated switchgear and power transmission lines, and power cables, and may be measured with a portable device.

Description

INPUT VECTOR FORMATION METHOD OF NEURAL NETWORKS FOR AUTO-

IDENTIFICATION OF PARTIAL DISCHARGE SOURCE

BACKGROUND OF THE INVENTION

Field of the Invention

The present invention relates to an electric power technology, and more particularly to an input vector formation method of neural networks for auto-identification of partial discharge source, which is capable of reducing time for measurement preparation since partial discharge is measured in a site in a state wherein it is difficult to get the phase information from a PT or a voltage divider, etc., as a neural network engine uses phase independent summation without phase information of applying voltage for high voltage machinery; of reducing measurement costs; of obtaining ratiocination result of relatively high reliability for discharge sources; of using an identical neural network engine in various types of partial discharge measurement apparatus, in which amplifier or signal attenuator characteristics are different from one partial discharge measurement apparatus to another, since input vector of a neural network engine is extracted from discharge types which do not include magnitude information of partial discharge signal; and of easy management of abnormal states of electric power machinery, since power phase at a portion in which partial discharge is generated can be detected using phase independent summation and phase dependent summation in electric power machinery having a plurality of phases, and partial discharge positions can be easily traced.

Description of the Related Art

A UI-IF, ultrasonic or other types of partial discharge as a precursor of break down could be generated in high voltage electric power machinery, such as a gas insulated switchgear (GIS), a gas insulated transformer, an oil transformer, rotating machinery, a gas insulated electric power transmission line, and power cables. When such a partial discharge occurs, a partial discharge measurement device measures the partial discharge, analyzes the measured signal type and patterns to determine whether they signify partial discharge, and diagnoses aging states of the electric power machinery, such that it can prevent a break down occurring in the electric power machinery.

With reference to Figs. 1 and 2, a partial discharge measurement device is described in detail below.

As shown in Fig. i, in a state wherein electric power machinery 100 has an abnormal state therein, when power 110 is applied to the electric power machinery 100, partial discharge as a precursor of break down is generated in the electric power machinery. Here, the electric power machinery 100 includes a GIS, a transformer, a motor, and a power cable.

Here, a partial discharge sensor 120 installed to the electric power machinery detects partial discharge signals, and carries the partial discharge signals to the partial discharge measurement device 130, thereby analyzing partial discharge signals.

Since the prior art analysis method needs phase information of power 110, applied voltage phase information is provided to the partial discharge measurement device 130 using a power transformer PT, and a voltage divider 140 or etc. The partial discharge measurement device 130 obtains basic information related to discharge signals, such as time 220 when partial discharge signal 210 is generated, magnitude 230 of partial discharge signal, and phase 240 of applying voltage 200 of a corresponding machinery when partial discharge occurs, from partial discharge signals 210 as shown in Fig. 2.

A circuit for obtaining signals in the partial discharge measurement device 130 includes various types of amplifiers, a signal attenuator, a detector, and a neural network engine for classifying signals, such that the partial discharge signals 210 inputted from the partial discharge sensor 120 can be analyzed.

From analysis of partial discharge signals 210, generation of partial discharge, cause of partial discharge and the aging state of the electric high voltage power machinery can be diagnosed.

The conmion neural network engine of the partial discharge measurement device 130 uses partial discharge signals as an input vector, which are measured but not processed, or various parameters extracted in phase resolved pulse sequence (PRPS) or phase resolved partial discharge (PRPD) which are transformed from the partial discharge signals. Here, the various parameters include numbers of phase resolved discharged pulses, magnitude of discharge signals or statistical parameters such as skew or kurtosis. This kind of input vector is obtained by directly or indirectly using applied voltage phase information and magnitude of discharge signals for electric power machinery which generate discharge signals.

However, when discharge signals from high voltage electric power machinery 100 are measured using a portable type of partial discharge measurement device, phase information applied to the high voltage electric power machinery 100 cannot be easily provided. Namely, when phase information of an applied voltage 200 or phase of power at a position in which partial discharge occurs is unknown, the partial discharge measurement device 130 perfonns a measurement, assuming that its power phase is same as that of a phase of a voltage applied to the high electric power device 100. But such measurement causes a problem that the prior art neural network engine can ratiocinate wrong discharge source. As result of this error the user cannot properly manage such a problem.

Also, even in case that a partial discharge measurement system, which is installed on-line and is capable of normally obtaining phase information of applied voltage, regarding the high voltage electric power machinery 100 whose phases are plural, such as s three-phase bundle typed GIS or transformer, a three-phase motor, etc., unless a neural network engine is trained by assorted groups of partial discharge data according to each power phase applied to power machineiy, it is difficult to derive a precise -ratiocination result.

In addition, if magnitude of discharge signals is used as an input vector of the neural network engine, even in case of same partial discharge signals, the magnitudes of detected discharge signals could be different, according to the sensitivity of sensors, distance of a discharge source, characteristics of a signal acquisition device, the neural network circuit must be trained using various magnitudes of discharge signals. Especially, when gains of various amplifiers in a discharge signal acquisition device are changed, the magnitude of signals measured for the identical discharge signals may be changed. Therefore, the neural network engine must be retrained.

SUMMARY OF THE INVENTION

Therefore, the present invention has been made in view of the above problems, and it is an object of the present invention to provide an input vector formation method of neural network for auto-identification of partial discharge source, which is capable of automatically indicating discharge sources using a neural network engine from measured partial discharge signals, to prevent break down of high voltage electric power machinery, such as a GIS, a transformer, a motor, a power transmission cable.

It is another object of the present invention to provide an input vector formation method of neural network for auto-identification of partial discharge source, which is capable of ratiocinating partial discharge source in a state wherein phase information of power applied to a high voltage electric power machinery cannot be obtained, and analogizing a phase of power applied to a position in which discharge occurs when high voltage electric power machinery use several phases of power.

It is yet another object of the present invention to provide an input vector formation method of neural network for auto-identification of partial discharge source, which is capable of employing an identical neural network engine for analyzing signals measured by partial discharge measurement devices whose data acquisition circuit amplifier or signal attenuator characteristics are different from one partial discharge measurement device to another, which does not use magnitude information of discharge signal as an input vector of the neural network engine.

In accordance with the present invention, the above and other objects can be accomplished by the provision of an input vector formation method of neural network for auto-identification of partial discharge source, in which the input vector is used iii a multi-layer perceptron structure, self organized map or other types of neural network which automatically ratiocinate source of partial discharge signals which are generated in high voltage power machinery, such as a GIS, a transformer, a power cable, an electricity distribution device, etc. The presented invention comprise the steps of: forming a cI)n: tn-1:N graph using discharge signals measured from a partial discharge measurement apparatus; transforming the c11n: Dn-1:N graph from the bottom right of cI)n: n-1:N graph to its upper left; extracting phase dependent summation to be used as the input vector of the neural network engine; extracting phase independent summation to be used as the input vector of the neural network engine; and inputting the phase dependent summation and the phase independent summation as an input vector of the neural network engine thereto.

Preferably, the method may further comprise the step of comparing phase dependent summation obtained from discharge signals and its phase shifted form by 1200 and 240 or other with shape of reference phase independent summation generated from training data of partial discharge source in neural network, and then obtaining a phase of power applied to a point wherein a discharge signal is generated.

Preferably, the method may further comprise the step of multiplying phase dependent summation obtained from discharge signals and its phase shifted form by 1200 and 240 or other by reference phase independent summation generated from training data of partial discharge source in neural network; integrating the multiplied value; and selecting a phase dependent summation having the largest integral area as a phase independent summation which indicate exact power phase of the a point wherein a discharge signal is generated.

Preferably, the method may further comprise the step of obtaining cross correlation between phase dependent summation obtained from discharge signals and its phase shifted form by 120 and 240 or other with shape of reference phase dependent summation generated from training data of partial discharge source in neural network, and then obtaining a phase of power applied to a point wherein a discharge signal is generated.

BRIEF DESCRIPTION OF THE DRAWINGS

The above and other objects, features and other advantages of the present invention will be more clearly understood from the following detailed description taken in conjunction with the accompanying drawings, in which: Fig. 1 is a block diagram of an apparatus for measuring partial discharge of the prior art; Fig. 2 is a graph of partial discharge signals, which describes information of the partial discharge signals measured by an apparatus for measuring partial discharge of the prior art; Fig. 3 is a graph of an initialized 1n: Dn-l:N; Fig. 4 is a view illustrating a discharge pattern by a cIln: 1-1:N visualization method when there are two successive partial discharge signals; Fig. 5 is a view illustrating a discharge pattern by a 1n: tn-1:N visualization method when there are three successive partial discharge signals; Fig. 6 is a graph of partial discharge signals illustrated by a t'n: c1n-1:N visualization method for describing a method of input vector formation for auto-identification of partial discharge source using neural networks according to the present invention; Fig. 7 is a view for describing concept of a method for extracting phase dependent summation and phase independent summation; Fig. 8 is a view for describing a method for extracting phase dependent summation and phase independent summation; Fig. 9 is a view illustrating a multi-layer neural network circuit; and Fig. 10 is a view illustrating a method for discriminating phase of power of discharge generation position using phase dependent summation according to the present invention.

DESCRIPTION OF THE PREFERRED EMBODIMENTS

Reference will now be made in detail to the embodiments of the present invention, examples of which are illustrated in the accompanying drawings, wherein like reference numerals refer to the like elements throughout The embodiments are described below to explain the present invention by referring to the figures.

An input vector formation method of neural network for auto-identification of partial discharge source according to the present invention is described in detail below.

Input vector is extracted, in which the input vector does not include magnitude of discharge signals from partial discharge signals and phase information of applied voltage.

Through a neural network engine using the input vector, in a state wherein phase information of a voltage applied to a high voltage power machinery does not exist, source of partial discharge can be automatically ratiocinated. Also, since the two input vectors do not use magnitude information of discharge signals, the neural network circuit must not be trained considering the magnitude of signals. In addition, when various phases of power are applied to the high voltage power machinery, input vectors including ratiocinated partial discharge sources and phase information are generated such that a phase of applied power at the point where partial discharge generates can be analyzed.

The present invention is related to generation of input vector of neural network circuit for ratiocinating precise discharge sources. As shown in Fig. 2, it is difficult to normalize discharge time 220 of information related to basic discharge signals, since values of the discharge time 220 vary from relatively very small to relatively very large. Also, since it has a problem that a neural network engine must be trained for the various magnitudes 230 of discharge signals. So the input vector of the neural network circuit is generated using only a phase 240 of applied voltage 200.

In order to generate the input vector, as shown in Fig. 3 to Fig. 6, a pattern of discharge is formed by a "tn: 1n-1:N visualization method" using a phase (b) 240 of applied voltage when partial discharge occurs.

The c1n: n-1:N visualization method describes relation between phases of voltages applied to a high voltage power device at time points when discharges are successively generated.

Namely, when discharge signals are successively generated in a high voltage power device, phase of the voltage is divided into K columns between 0 to 360 of the voltage phase, as show in Fig. 3. After that, X-axis is defined by phase of a voltage which is applied thereto when a preceding discharge signal occurs, and Y-axis is defined by phase of a voltage which is applied thereto when a following discharge signal occurs. For example, as shown in Fig. 2, let's assume that there are three successive discharge signals 210. In first two discharge signals, when phase n-1 of applied voltage of the preceding discharge signal corresponds to i-th (1 = i = K) phase, and phase c1n of applied voltage of the following discharge signal corresponds toj-th (1 = j = K) phase, a value of(i,j) in a two-dimensional graph whose initial value is set to zero, Pj, is added by one, as shown in Fig. 4. Therefore, as shown in Fig. 5, phases of successive discharge signals can be described by the tn: (I)n-1:N visualization method.

Similarly, of the three successive discharge signals 210 in Fig. 2, when phase cbn of applied voltage of the preceding discharge signal of the second successive discharge signal corresponds to j-th (1 = j = K) phase, and phase n+l of applied voltage of the following discharge signal corresponds to j'-th (1 = j' = K) phase, a value of (jj') in a two-dimensional graph whose initial value is set to zero, P', is added by one. Therefore, as shown in Fig. 5, phases of successive discharge signals can be described by the c1n: n-1:N visualization method.

n: n-1:N pattern of the discharge signals which are measured for a given time (T) by the above-mentioned method can be expressed as Fig. 6. Each coordinate is expressed by color value 600 indicating the number of successively generated partial discharge signals, in Fig. 6. Here, X-and Y-axes indicate phase axes and are normalized within a range of 0 to 360 automatically.

Here, the number 600 of points in Fig. 6 can be expressed by density of points which are expressed for a predetermined time.

On the other hand, since shapes of such a discharge pattern by the 1n: 1n-l:N visualization method are different according to types of discharge sources, it is adequate to generate input vectors based on the discharge sources such that the input vectors can be discernable according to type of partial discharge sources. In order to obtain input vector excluding phase information, Fig. 5 is modified. Namely, when bottom right 500 of Fig. 5 is vertically moved to locate at the upper side of Dn: (1n-1:N of the prior art, as shown in Fig. 7, new discharge pattern 710 can be obtained, which is referred to as a" modified)n: (I)n-1:N visualization method". hi such a graph of the modified n: n-1:N, when the discharge number of Y' axis is summed with respect to X' axis, the number of discharges generated in the given 11n-l can be obtained. Since such a summation includes phase information, the present invention defines that the summation is called a phase dependent summation (PDS) 720. A term PDS of the i-th PDS 720 can be expressed by the following equation:

K

PDS= P1 1= I When summation is performed with respect to X' axis direction while a point is moved along the Y' axis in Fig. 7, since each summed point is not a summation for specific phases on the C1n axis and the c1n-1 axis but a summation for the number of discharges corresponding to phases which are different from each other, a new graph 730 without phase information can be obtained, which is called a phase independent summation (PIS). A term PIS of thej-th PDS 730 can be expressed by the following equation: PJS = P1(/(11)) +1K(f 2) K-j_2 1-K-1 = 0, when j=1) Fig. 8 is a view for describing a method for extracting phase dependent summation 810 and phase independent summation 820, in which n: Dn-I:N pattern for substantial discharge signals of Fig. 6 is expressed by a modified cIn: 1n-1:N visualization method 800.

Here, the phase dependent summation 810 and the phase independent summation 820 are parameters related to discharge number and are always positive. I0

Fig. 9 is a view illustrating a multi-layer neural network engine. As shown in Fig. 9, the multi-layer neural network engine includes a first layer 910, N-th layer 920 composed with perceptrons, neurons (operations) 950 and input vectors 900, which are extracted from measured discharge signals. Also, the circuit includes an output layer 930 for operating output vectors 940 of the neural network circuit, which indicates partial discharge sources of electric power machinery, and synapses 950 connecting to respective layers.

When the phase independent summation 820 is used as an input vector of the multi-layer neural network circuit, since phase information of an applied voltage is not employed, the partial discharge signal measurement device can ratiocinate relatively precise partial discharge sources regardless of acquisition of phase information of applied voltage.

Also, when the phase independent summation 820 and phase dependent summation 810 are used as input vectors of the neural network circuit, since information for a magnitude of a discharge signal is not used, a neural network engine, which is trained by a discharge signal measured in a specific signal acquisition device, can be employed in another partial discharge signal acquisition device, whose signal amplification characteristic is different from the specific signal acquisition device, without retraining.

The neural network circuit employing the phase dependent summation 810 and the phase independent summation 820 uses much more discharge information than that of a neural network circuit which uses only a phase independent summation 820, when phase information can be obtained. Therefore, the neural network circuit employing two summations can little bit increase recognition ratio, compared with a neural network circuit using only phase independent summation 820. But in case that phase information can be obtained from high voltage power machinery, neural network with only phase independent summation as input vector shows good classification ability because phase independent summation doesn't use phase information.

Also, when using the phase dependent summation 810 and the phase independent summation 820, a phase of a voltage applied to a position in which discharge occurs can be analogous without specific information for phase of a position at which discharge is generated, wherein the discharge occurs in a high voltage electric power machinery to which power having a plurality of phases is applied.

For example, in a three-phase bundle type GIS in which phases R, G and B simultaneously exist, measurement of discharge signals using a partial discharge measurement device which is synchronous to phase R is described in detail below, in which partial discharge occurs at phase G. Firstly, as shown in Fig. 10, after phase dependent summation 1000 and phase independent summation are extracted from discharge signals, respectively, discharge sources are ratiocinated through a neural network engine which employs the phase independent summation of the discharge signals.

Afterwards, shape of reference phase independent summation 1030, which is used in training a neural network circuit for ratiocinating discharge sources, is compared with those of phase dependent summation 1000 obtained from discharge signals and phase dependent summations 1010 and 1020, which are generated as phase of the phase dependent summation 1000 is shifted by 120 and 240 , respectively.

Here, the most analogous shape with the reference phase independent summation 1030 is a phase dependent summation whose phase is shifted by 120 . Therefore, partial discharge is generated at phase G leading by phase 120 , in which the phase G is lagged by phase R which is synchronous to the partial discharge measurement device. From such a procedure, a phase of applied voltage at a position in which partial discharge occurs can be reasoned.

In this case, there are several methods for numerically expressing analogy between a reference phase independent summation 1030 and a phase dependent summation 1000 of measurement signals or its phase shifted phase dependent summations 1010 and 1020 which are generated as phase of the phase dependent summation 1000 is shifted by 1200 and 240 , respectively. The simplest one of the methods is to multiply two phase dependent summations with each other. According to such a method, the more shapes of the two phase dependent summations are similar to each other, the larger value of integrating the multiplying value (which is an area for a curve of multiplying phase dependent summations).

In the above example, when integrating multiplication of phase independent summations 1000, 1010, 1020 and a reference independent summation 1030, to produce integral areas (1040, 1050, 1060), the phase dependent summation 1050 of which phase is shifted by 1 20 which makes the integral area be largest is the most analogous to the reference phase independent summation 1030.

Also, methods, such as cross correlation, etc., can derive phrase.

As mentioned above, the input vector formation method of neural network for auto-identification of partial discharge source according to the present invention has advantages in that fact that measurement preparation efforts and costs can be reduced since there is no need to get the phase information of electric power machinery, in a state wherein a phase signal can't be easily inputted from a PT or a voltage divider, etc., when a neural network circuit using phase independent summation is employed. In addition the method has advantages in that ratiocination result of relatively high reliability of auto-classification for discharge sources can be obtained, and an identical neural network circuit can be used in other partial discharge measurement apparatus, in which characteristics of an amplifier or a signal attenuator, which are in the partial discharge measurement device, are different each other, since input vector of a neural network engine is extracted from discharge types which do not include magnitude information of partial discharge signal.

Furthermore, the method can give more opportunity to recover against abnormal states of electric power machinery having a several power phases, since phase of applied voltage at a portion in which partial discharge is generated can be detected using phase independent summation and phase dependent summation and so positions where partial discharges are generated can be easily traced.

Although the preferred embodiments of the present invention have been disclosed for illustrative purposes, those skilled in the art will appreciate that various modifications, additions and substitutions are possible, without departing from the scope and spirit of the invention as disclosed in the accompanying claims.

Claims

WHAT IS CLAIMED IS: 1. An input vector formation method of neural

network for auto-identification of partial discharge source which automatically ratiocinates source of partial discharge signals which are generated in high voltage power machinery, such as a GIS, a transformer, a power cable, a electricity distribution device, and whose from can be of various neural network engines, such as multi-layer perceptron, a selforganization map, etc., the method comprising the steps of: forming a n: q)n-1:N graph using discharge signals measured from a partial discharge measurement apparatus; transforming from the bottom right ofn: cI)n-1:N graph to its upper left; extracting phase dependent summation to be used as the input vector of the neural network engine; extracting phase independent summation to be used as the input vector of the neural network engine; and inputting the phase dependent summation and the phase independent summation as an input vector or the neural network engine thereto.

2. The method as set forth in claim 1, further comprising the step of: comparing phase dependent summation obtained from discharge signals and phase dependent summation obtained after the phase dependent summation is shifted by 1200 and 240 or other phases with shape of reference phase independent summation which is generated from training data of neural network, and then reasoning phase of best similar phase shifted phase dependent summation to reference phase dependent summation as a phase of power applied to a point wherein a discharge signal is generated.

F

3. The method as set forth in claim I or 2, further comprising the step of: multiplying phase dependent summation obtained from discharge signals and phase dependent summation obtained after the phase dependent summation is shifted by 1200 and 240 or other phases by reference phase independent summation which is generated from training data of neural network; integrating the multiplied value; and selecting a phase dependent summation having the largest integral area as a phase independent summation which is the most analogous to the reference phase independent summation; reasoning the phase of best similar phase shifted phase dependent summation to reference phase dependent summation as a phase of power applied to a point wherein a discharge signal is generated 4. The method as set forth in any preceding claim, further comprising the step of: obtaining cross correlation between phase dependent summation obtained from discharge signals and phase dependent summation obtained after the phase dependent summation is shifted by 120 and 240 or other phases with shape of reference phase dependent summation, and reference phase independent summation, and; reasoning the phase where magnitude of cross correlation is maximum as a phase of power applied to a point wherein a discharge signal is generated.

5. An input vector formation method substantially as hereinbefore described with reference to Figures 3 to 10 of the accompanying drawings.