United States Patent [19]
`Snow et al.
`
`illllllllllllllilllllilllllllllllilllllllllllllillllillllllllllllilllllllll
`5,537,327
`Jul. 16, 1996
`
`USOO5537327A
`Patent Number:
`Date of Patent:
`
`[11]
`
`[45]
`
`[54]
`
`METHOD AND APPARATUS FOR
`DETECTING HIGH-IMPEDANCE FAULTS IN
`ELECTRICAL POWER SYSTEMS
`
`[75] Inventors: Peter B. Snow, Colorado Springs,
`Colo.; Alexander P. Apostolov;
`Jefferson D. Bronfeld, both of
`Binghamton, NY.
`
`[73]
`
`Assignee:
`
`New York State Electric & Gas
`Corporation, Binghamton, NY.
`
`[21]
`[22]
`[51]
`[52]
`
`[58]
`
`[56]
`
`Appl. No.2 141,308
`Filed:
`Oct. 22, 1993
`
`Int. Cl.6 .......................... .. G06F 15/22; GOlR 31/08
`US. Cl. ........................ .. 364/483; 364/481; 364/482;
`395/50; 395/21; 395/22; 395/23; 395/907;
`361/91; 361/92; 361/93
`Field of Search ................................... .. 364/483, 482,
`364/481, 492; 361/93, 76, 30, 91-92; 324/508,
`509; 395/907, 915, 21-23, 50, 54, 75, 77
`
`References Cited
`
`U.S. PATENT DOCUIVIENTS
`
`4,617,636 10/1986 Johns et a1. ........................... .. 364/482
`5,390,106
`2/1995 Cederblad et a1. ..................... .. 361/90
`
`OTHER PUBLICATIONS
`Fernando; “High impedance fault detection using arti?cial
`neural network techniques”; IEEE; 6 vol. 2870 p. l, 1992
`(full text).
`Sarosh; “Toast: The Power System Operator’s Assistant”;
`IEEE; Jul. 1986, pp. 53-59.
`
`Abstract Title: “High Impedance fault detection using arti
`?cial nerual network techniques”, Society of Automotive
`Eng. 1992 vol. 2870 pp.
`S. Ebron et al. “A Nerual Network approach to the Detection
`of Incipient faults on power distribution feeders”, paper No.
`89 TD 377-3 PWRD.
`
`Primary Examiner-Emanuel T. Voeltz
`Assistant Examiner-Kamim' Shah
`Attorney, Agent, or Firm-Salzman & Levy
`
`[57]
`
`ABSTRACT
`
`The present invention features a method and apparatus for
`detecting and enabling the clearance of high impedance
`faults (HIFs) in an electrical transmission or distribution
`system. Current in at least one phase in a distribution system
`is monitored in real time by sensors. Analog current signa
`ture information is then digitized for processing by a digital
`computer. Zero crossings are identi?ed and current maxima
`and minima located. The ?rst derivatives of the maxima and
`rninima are computed and a modi?ed Fast Fourier Trans
`form (FPT) is then performed to convert time domain to
`frequency domain information. The transformed data is
`formatted and normalized and then applied to a trained
`neural network, which provides an output trigger signal
`when an HIF condition is probable. The trigger signal is
`made available to either a network administrator for manual
`intervention, or directly to switchgear to deactivate an
`aifected portion of the network. The inventive method may
`be practiced using either conventional computer hardware
`and software or dedicated custom hardware such as a VLSI
`chip.
`
`32 Claims, 6 Drawing Sheets
`
`i
`
`Accept Dcto
`
`5°
`
`1
`second Of (1010
`collected?
`
`Loccte zero crossing .__ .52
`points, ?nd maximum
`and minimum currents
`i
`take tirsi derivative
`of maximum and
`minimum currents
`
`i
`
`Periorrn FFT on
`iirst derivatives
`
`~56
`
`i
`
`Compuie power
`55
`spectrum for positive
`and negative currents /
`i
`Sum positive and
`50
`power /
`spectra
`1
`Average sums 0i
`positive and negative /
`power spectra
`
`64
`
`1
`
`NOITHOIIZG end lOi'mOi
`dOtO; OLtiDU‘l t0 ANN
`
`Terminate processi !
`
`i—
`
`AVS EXHIBIT 2004
`Toyota Inc. v. American Vehicular Sciences LLC
`IPR2013-00424
`
`

`
`US. Patent
`
`Jul. 16, 1996 I
`
`Sheet 1 of6
`
`5,537,327
`
`QA
`
`QB
`
`QC
`
`12a
`
`16
`
`20
`
`18
`
`Data
`
`Pre
`
`Processor
`
`ND
`
`Converter
`
`'\/14b
`
`Trained
`ANN
`
`1
`22
`
`/_\\_/ AV hv
`
`gig wza 7
`
`24
`Fault
`Indicating j
`Output
`
`7
`
`Visua| or
`Audi°
`Indicating
`H Device
`25a
`
`Other
`_ Indicating
`'
`or
`
`Recording
`r" Device
`25b
`
`Remotely
`Trippable
`Circuit
`f‘)
`Breaker
`250
`
`25d
`\ oth-er
`Electrically
`Actuatable
`Fault Clearing
`Device
`
`V
`
`

`
`U.S. Patent
`
`Jul. 16, 1996
`
`Sheet 2 0f 6
`
`5,537,327
`
`(
`
`Start Pre-process )
`
`4
`
`Accept Data
`
`~50
`
`1
`second of data
`collected?
`
`Locate zero crossing
`points, find maximum
`and minimum currents
`
`V
`
`Take first derivative
`of maximum and N54
`minimum currents
`
`1
`
`Perform FFT on N56
`first derivatives
`l
`Compute power
`58
`spectrum for positive
`and negative currents /
`
`7
`
`Sum positive and
`negative power
`spectra
`
`60
`
`62
`Average sums of /
`positive and negative
`power spectra
`
`64
`v /
`Normalize and format
`data; output to ANN
`
`

`
`US. Patent
`
`Jul. 16, 1996
`
`Sheet 3 0f 6
`
`5,537,327
`
`% HlZPROC
`% Developed by Dr. Peter B. Snow for NYSEG March 9, 1993
`% enter a file for processing - for example \ksc\hiz42_3.dat
`% now put the file in vector a - for example a = hiz42_3
`% now type in hizproc to activate this processing program
`% clear the old matrices
`clear b cl c2;
`% first get the size of the input vector
`[id, ja] = size(a);
`% now shift the signal to the origin y=0
`avg1 = mean(a);
`a = a - avg1;
`%now find the element numbers of the zero current crossings
`k = 1;
`for i=1zjd-1;
`if a(i) < 0.0;
`if a(i+1) > 0.0;
`b(k) = i;
`k = k + 1;
`end;
`end;
`end;
`[id, ja] = size(b);
`k = 1;
`% now find maximum values for each cycle
`for i = 1:ja-1;
`i1 = W):
`12 = b(i+1);
`
`k = k + 1
`end;
`% now take first derivative of max, min vectors
`e1 = diff(c1);
`e2 = diff(c2);
`% now take the FFT and convert to power spectrum
`91 = fl't(e1, 512);
`92 = fft(e2, 512);
`pg1 = Q1 .*c0nj(g1)/512;
`pg2 = Q2 .*conj(g2) / 512;
`sumfft1 ='sum(pg1);
`sumfit2 = sum(pg2);
`sumfft = ( sumfft1 + sumfft2 ) / 2;
`
`gig was 3
`
`

`
`US. Patent
`
`Jul. 16, 1996
`
`Sheet 4 of 6
`
`5,537,327
`
`32
`
`N0
`
`Micro
`
`Converter
`
`Processor
`
`Fault
`Indicating
`Output
`
`24
`
`Visual or
`Audio
`Indicating
`
`_ / Device
`25a
`
`Other
`_ Indicating
`'
`or
`
`Recording
`r.)
`Device
`
`Remotely
`Trippable
`Circuit
`Breaker
`
`Other
`Electrically
`Actuatable
`Fault Clearing
`Device
`
`gig LL18 4a
`
`

`
`US. Patent
`
`Jul. 16, 1996
`
`Sheet 5 0f 6
`
`5,537,327
`
`40
`
`Current -—'—" VLSl
`
`
`
`Input Signals
`
`,
`
`Chip
`
`Fault
`24
`Indicating
`Output /
`
`Visual or
`Audio
`
`~
`
`Indicating
`r4 Device
`25a
`
`Other
`‘ Indicating
`
`,
`
`or
`
`Recording
`f" Device
`25b
`
`Remotely
`, Trippable
`Circuit
`r‘) Breaker
`25c
`
`25d
`\ Other
`Electrically
`
`_ Actuatable
`Fault Clearing
`Device
`
`was 41;
`
`

`
`US. Patent
`
`Jul. 16, 1996
`
`Sheet 6 of 6
`
`5,537,327
`
`0.12 l——
`
`0.10 ----~
`
`0.08 ——
`
`0.00-—
`
`0.04 ——
`
`0.02 ——
`
`O
`
`1
`
`2
`
`3
`
`4
`
`5
`
`6
`
`7
`
`8
`
`9
`
`10
`
`U Series1
`[E Series 2
`
`gl?ww 5a
`
`0.60 ——
`
`0.50 ——
`
`0.40 ——
`
`0.30 ——
`
`0.20 ——
`
`0.10 —
`
`E Series 1
`Series 2
`
`I
`
`“15 5LZ
`
`Series 1 = Phase B of 13 kV Line - Dropped on Ice
`Series 2 = Phase B of 13 kV Line
`
`

`
`1
`METHOD AND APPARATUS FOR
`DETECTING HIGH-IMPEDANCE FAULTS IN
`ELECTRICAL POWER SYSTEMS
`
`5,537,327
`
`FIELD OF THE INVENTION
`
`The invention pertains to the detection of faults in an
`electrical power system and more particularly, to the detec
`tion of high-impedance faults in power transmission and
`distribution systems caused by downed conductors, tree
`limbs across conductors and the like.
`
`BACKGROUND OF THE INVENTION
`
`In the electrical power generation and distribution indus
`try, monitoring transmission and distribution networks for
`fault conditions is very important. The term distribution
`network is used herein to refer to any electrical power
`transmission or distribution facility. Moreover, a fault con
`dition is any abnormal (unexpected) current-conducting path
`in a distribution network. Fault conditions often present
`danger to people and property. They also waste electrical
`energy.
`One type of fault is a bolted (short-circuit) fault of one or
`more legs of a distribution network to another leg thereof or
`to ground. This type of low-impedance (low-Z) fault con
`dition is easily detected by conventional circuit overcurrent
`protective devices such as fuses or circuit breakers. In other
`words, a complete short circuit (a low-impedance path)
`usually trips a circuit breaker or blows a fuse. A circuit of the
`distribution network experiencing such a fault condition is
`quickly removed from service until such time as repairs can
`be effected (i.e., until the fault is cleared).
`Another type of fault condition occurs when an unin
`tended high impedance (high-Z) conductive path occurs
`between transmission line legs or between one leg and
`ground. Such high-Z paths may occur when a tree limb or
`the like falls across a transmission line or when a single leg‘
`of a transmission line breaks (due to ice or wind, for
`example) and touches the ground. Generally, a single con
`ductor of the distribution network dropping to the ground
`will not create a short circuit, but will continue to allow
`current ?ow at a relatively low rate. Such current ?ow often
`causes arcing. This condition presents a great danger of
`electrocution to people or animals happening across the
`downed conductor. The arcing can also result in ?res.
`A problem constantly plaguing the electrical power indus
`try is ?nding an effective way to differentiate between a
`high-Z fault (HIF) condition and similar effects caused by
`changes in the loads attached to the distribution network. In
`addition to load switching events, power factor correcting
`capacitor banks are frequently switched on and off the
`network and transformer taps are automatically changed to
`keep the network voltage constant. Both of these events also
`create conditions on the network which may appear similar
`to an HIF condition. Any effective system for detecting HIFs
`must be able to distinguish fault conditions from normal
`load switching events. A system which ignores legitimate
`HIF conditions risks the aforementioned dangers while a
`system which falsely trips in response to normal load
`switching events can wrack havoc with consumers relying
`on uninterrupted electrical service. Interruption of electrical
`service to certain manufactming processes, for example,
`may destroy work-in-process and result in large expense to
`the manufacturer. An interruption of medical apparatus can
`also be inconvenient at best, and disastrous at worst.
`
`20
`
`35
`
`45
`
`50
`
`55
`
`65
`
`2
`I-HF detection solutions as simple as lowering the trip
`points of conventional circuit protective devices have been
`tried. Because HIF-drawn current is often a very small
`percentage of the total network current, this solution has
`done nothing more than cause excessive service interrup
`tions. Most H[F research has focused upon the problem of
`?nding detectable differences in measurable parameters in a
`distribution network under HE and normal load switching
`conditions. Some of the parameters measured and compared
`have included phase current, ground current, ratio of ground
`current to positive sequence current, and frequency spec
`tra—both near line frequency (typically 60 Hz.) and at
`higher harmonic relationships to line frequency.
`One system with potential for detecting I-lIFs is disclosed
`in U.S. Pat. No. 5,223,795 issued to Frederick K. Blades,
`titled “Method and Apparatus for Detecting Arcing in Elec
`trical Connections by Monitoring High Frequency Noise”.
`Blades discloses a system wherein high-frequency noise
`caused by arcing is detected and, when measured above a
`pre-programmed threshold level, trips a circuit protective
`device. While the Blades system is intended for residential
`branch circuit uses, it is representative of a class of HIF
`detection strategies relying on high-frequency noise for use
`in power distribution networks. These techniques have not
`proven effective in the detection of HIFs, since noise gen
`erated by HIFs varies widely in both spectrum and intensity.
`In addition, capacitor banks, automatically switched on and
`off the network for power factor correction, tend to short the
`high-frequency noise signals to ground, adding additional
`uncertainty to the detection and analysis process.
`Another approach to HIF detection is taught in U.S. Pat.
`No. 5,216,621 issued to Richard T. Dickens, titled “Line
`Disturbance-Monitor and Recorder System”. Dickens dis
`closes a system comprising analog current and/or voltage
`sensors placed at selected positions in a distribution net
`work. Analog signals from the detectors are digitized by
`analog-to-digital (A/D) converters and presented to a high
`speed digital signal processor (DSP) as digital sample
`words. The DSP computes both the real and imaginary
`phasor components of the operating parameter or param
`eters. These phasor components are then used to calculate
`various measures of power transmission performance
`according to known phasor equations. Trigger means imple
`mented within the DSP provides an output signal when
`pre-programmed limits of a measured or calculated quantity
`are exceeded. Memory in cooperation with the DSP captures
`and stores digital sample words associated with abnormal
`events for future analysis. The Dickens apparatus appears to
`be expensive to build and, even with a state-of-the-art DSP,
`the system may only detect HIFs with well-known charac
`teristics. Each installation on a network may have to be
`individually calibrated to the characteristics of that network
`and, as the loads changed on the network (e.g., by adding or
`removing power customers), the system would have to be
`re-calibrated.
`A third approach to HIF detection is disclosed in U.S. Pat.
`No. 4,878,142 issued to Sten Bergman, et al., titled “High
`Resistance Ground Fault Protection”. Bergman discloses a
`system for analyzing the non-harmoniccomponents of phase
`currents. A, estimate of Fourier coe?icients is computed,
`thereby transforming the time-domain information into the
`frequency domain. Both the original digitized signals and
`the transformed frequency domain signals are applied to
`detection circuitry. Logical decisions based on comparison
`to known fault parameters are made and a fault-indicating
`trip signal is provided when appropriate. The Bergman
`system suffers from many of the same shortcomings as the
`
`

`
`3
`aforementioned Dickens system. The Bergman system must
`be calibrated to each network and re-calibrated when the
`load pro?le on the network changes.
`Yet a fourth HIF detection system is described in “A
`Neural Network Approach to the Detection of Incipient
`Faults on Power Distribution Feeders”, paper No. 89 TD
`377-3 PWRD, S. Ebron, D. Lubkemen and M. White (1989).
`That HIF detection system relies on the monitoring, digiti
`zation, and comparison of voltage and current conditions in
`the distribution network. The mechanism for deciding
`whether an event is an HIF or a normal load switching event
`is a partially or fully-trained neural network. The 200-node
`neural network described by Ebron et a1. is trained using
`data obtained from a computer-simulated distribution net
`work using the Electromagnetic Transients Program
`(EMTP) by Systems Control, Inc. Data is collected for ten
`cycles (two subsequent zero-crossings going from negative
`to positive) of simulated operation. Digitized data is pre
`processed to extract features such as peak transient current
`over three phases, and phase currents immediately before
`and after the occurrence of a detected transient. The
`extracted feature vector representing ten cycles of simulated
`network operation is then applied to a neural network. The
`network, once at least partially trained, then identi?es pat~
`terns in the data as either HIF or normal load switch events.
`A trigger signal can be generated when an HIF is detected.
`It is therefore an object of the present invention is to
`provide a method and apparatus for monitoring an electrical
`power distribution network and for distinguishing HIF con
`ditions from normal load, capacitor bank or transformer tap
`switching conditions with extremely high accuracy.
`It is a further object of the invention to produce a system
`self-adaptable to a variety of networks and one that need not
`be calibrated for changes in load on the network.
`It is still a further object of the invention to provide a
`self~contained, low-cost, single-chip hardware implementa
`tion of the inventive method for use either as a standalone
`HIF monitor, or as an integral part of a circuit-interrupting
`device for completely self-contained fault clearing.
`SUMMARY OF THE INVENTION
`
`In accordance with the present invention, there is pro
`vided a method and apparatus for detecting and clearing
`high impedance faults (HIFs) in an electrical transmission or
`distribution system. Current in at least one phase in a
`distribution system is monitored in real time by sensors.
`Analog current/voltage signature information is then digi
`tized for processing by a digital computer. Zero crossings are
`identi?ed and current maxima and rninima for each cycle
`located. The ?rst derivatives of the maxima and minima are
`computed and a modi?ed Fast Fourier Transform (FFT) is
`then performed to convert these time~domain derivatives to
`frequency domain information. The transformed data is
`formatted, summed and normalized and then applied to a
`trained neural network, which provides an output trigger
`signal when an HIF condition is probable. The trigger signal
`is made available to either a network administrator for
`manual intervention, or directly to switchgear to deactivate
`an aifected portion of the network. The inventive method
`may be practiced using either conventional computer hard
`ware and software or dedicated custom hardware such as a
`VLSI chip.
`BRIEF DESCRIPTION OF THE DRAWINGS
`
`A complete understanding of the present invention may be
`obtained by reference to the accompanying drawings, when
`
`45
`
`55
`
`60
`
`65
`
`5,537,327
`
`4
`taken in conjunction with the detailed description thereof
`and in which:
`FIG. 1 is a functional block diagram in accordance with
`the present invention;
`FIG. 2 is a ?ow chart showing the pre-processing steps of
`the method of the present invention;
`FIG. 3 is a program listing of the pre-processing com
`mands;
`FIGS. 4a and 4b show block diagrams of alternate hard
`ware embodiments of the present invention; and
`FIGS. 5a and 512, respectively, show fault detection prob
`ability without and with ?rst derivative calculation.
`
`DESCRIPTION OF THE PREFERRED
`EMBODlMENT
`
`10
`
`15
`
`25
`
`35
`
`40
`
`Generally speaking, the invention relates to the ?eld of
`fault detection in electrical power transmission and distri
`bution systems and networks. It should be understood that
`the terms “power distribution network” or “network” are
`hereinafter used to refer to any alternating current electrical
`transmission or distribution system or facility. More speci?
`cally, there is disclosed a system for the identi?cation of
`high-impedance faults (HIFs) which is capable of differen
`tiating with high accuracy fault conditions drawing as small
`a current as 200 mA in a power distribution network. HIF
`conditions may, however, draw currents as high as 700
`amps. The system is applicable to both single-phase and
`poly-phase networks.
`Referring now to FIG. 1, there is shown a functional block
`diagram of the present invention. A portion of a typical
`three-phase power distribution network is shown generally
`at reference numeral 10. Phase conductors 12a, 12b and 12c
`of distribution network 10 each are monitored by current
`transformers 14a, 14b and 14c, respectively. Since current
`transformers are well known in the art, it should be noted
`that any current transformer or transducer capable of gen
`erating a low-noise analog signal representative of the
`current may be employed. Typically a Model No. MF1254O
`MF slip-over bushing current transformer manufactured by
`Associated Engineering Corp. could be used. The HIF
`detection system of the present invention could be used
`e?fectively to monitor a single phase network or a single
`phase of a poly-phase distribution network.
`Analog signals representing phase currents from phase
`conductors 12a, 12b, and 12c monitored by current trans
`- formers 14a, 14b and 14c are provided to inputs of an
`analog-to-digital (A/D) converter 16. A/D converter 16 is a
`50
`three or more channel device with 12 bit resolution, typically
`a Model No. DAS-50 manufactured by Keithley Metrobyte.
`Sampling rate is approximately 5 kHz. Digitized represen
`tations of analog phase currents are applied from A/D
`converter 16 to a pre-processor 20 via data bus 18. Pre
`processor 20 performs several mathematical and data for
`matting operations described in more detail hereinbelow.
`Pre-processed data is then applied to the input of a trained
`arti?cial neural network (ANN) 22. An ANN is a computer
`model of the parallelism and interconnectedness of the
`human brain. Connectionist models exist which have the
`ability to derive “rules” by analyzing patterns. The ANN
`differs from more conventional expert systems or general
`arti?cial intelligence approaches in that the latter require the
`existence of well-de?ned rules as a prerequisite for effective
`operation. The training of an ANN is the process whereby
`the ANN learns to associate input states to output states by
`
`

`
`5,537,327
`
`25
`
`5
`adjusting weights and biases. In the training mode, the
`expected or desired outcome based upon the applied data is
`used. If training data covering a broad enough range of
`conditions is provided, the ANN eventually self-develops a
`pattern relating inputs to outputs allowing it to analyze any
`applied unknown data. A trained ANN is therefore highly
`effective at pattern recognition tasks. That is, pattern recog
`nition is accomplished at extremely high speed and with a
`high probability of a correct outcome, based upon the
`applied data.
`'10
`Training the ANN 22 is necessary to establish the “rules”
`or weights to be used to classify events occurring on the
`network as HIFs or normal switching events. Training is
`accomplished by providing multiple sets of known data
`representing both HIF and normal network conditions to the
`ANN 22 along with the correct “answers” corresponding to
`each data set. Data sets my be compiled using computer
`simulation techniques or may consist of actual ?eld-col
`lected data corresponding to_both fault and no-fault condi
`tions of distribution networks.
`Actual ?eld-collected data and superposition combina
`tions of ?eld data were used to train the ANN 22 of the
`present invention. Over 300 sets of ?eld-collected and super
`position data were used for the training. A typical training
`data set consisted of 2860 data points representing 10
`seconds of sampling time on a real or simulated network.
`Digitization frequency was 5000 samples/second. The train
`ing data sets were pre-processed essentially as will be
`described in detail hereinbelow. The backprapogation of
`errors method was the learning technique chosen to train
`ANN 22 of the present invention. A learning rate of 0.05 and
`a momentum (a method of changing weights based on a
`previous weight) of 0.1 were chosen. Once trained (i.e.,
`appropriate weights established), an ANN my be duplicated
`and distributed and applied in its intended application as a
`?xed-weight (non-trainable) ANN.
`In the preferred embodiment, ANN 22 comprises 386
`MATLAB software version 3.5M from Mathworks, Inc.
`operating on an IBM® compatible personal computer under
`the Microsoft® DOS operating system version 5.0 or
`greater.
`ANN 22 continuously analyzes data from pre-processor
`20 and provides an output signal 24 which indicates either
`a normal condition or an HIF. Signal 24 may be used to alert
`a network administrator (not shown) of a potential HIF by
`means of indicating/recording devices 25a and/or 25b.
`Optionally signal 24 may be utilized directly to clear the
`assumed fault by sending a trip signal to a protective circuit
`interrupter 25c or 25d as is well known in the art.
`Refer now also to the ?ow chart of FIG. 2 and also to the
`program listing of FIG. 3, which represents the code applied
`to ANN 22. In learning mode operation, pre-processor 20
`?rst accumulates approximately 5000 digitized current data
`points for each line phase 12a, 12b and 120 being monitored.
`This represents approximately one second of network opera
`tion, step 50. Cycles are identi?ed by their zero-crossing
`points. Zero-crossing points are identi?ed in the data for
`each phase conductor 12a, 12b and 120. This is accom
`plished by calculating when the current crosses from nega
`tive to positive sign. The maximum and minimum current
`for each cycle is determined, step 52. The ?rst derivative
`(i.e, the rate of change with respect to time) of the maximum
`and minimum current values is then taken, step 54.
`It has been found through experimentation with both
`simulated and ?eld recorded data that calculating the ?rst
`derivative is essential to the process of accurately detecting
`
`50
`
`55
`
`65
`
`20
`
`35
`
`40
`
`45
`
`6
`I-IIFs. For purposes of illustrating the importance of the ?rst
`derivative in detecting a high-impedance fault, refer now
`also to FIGS. 5a and 5b. There are shown bar graphs of ANN
`output (relative value or relative HJF probability) vs. time
`for a 10 second period for an actual H]F condition occurring
`on a single phase of a 13 kV three-phase transmission line
`(not shown). For this fault, the phase B leg of the transmis
`sion line was dropped onto ice. Series 1 data shows the
`faulted condition while Series 2 data shows the line under no
`fault conditions. In FIG. 5a the ?rst derivative has not been
`calculated and, as may be seen, there is little difference in the
`relative value of Series 1 or Series 2 data. In FIG. 5b,
`identical data is plotted. However, the ?rst derivative has
`been calculated, and consequently there is pronounced dif
`ference in the output of the ANN 22 in response to the HIF
`(relative value of the Series 1 data).
`A Fast Fourier Transform (FFT) is then performed on
`each of the two ?rst derivatives, step 56 (FIG. 2). Next, a
`power spectrum is computed for both positive and negative
`currents, step 58. This is accomplished by taking the product
`of the FFT with its complex conjugate. Total power is
`calculated by summing the power spectrum for the one
`second data window, step 60. The positive and negative
`summed powers are then averaged, step 62. The averaged
`power thus calculated is then formatted and scaled for
`compatibility with the non-linear transfer function of the
`trained ANN 22, step 64. This is accomplished by repre
`senting each one second averaged power as the number of
`standard deviations from the average Z-scaling. The steps
`are repeated continuously, if additional data is processed,
`step 66, starting again with step 50. Processing terminates,
`step 68 when all data has been exhausted.
`Refening to FIGS. 1, 4a and 4b, a fault indicating output
`signal 24 is generated by trained ANN 22 and provided to
`signaling/recording or fault clearing devices 25a, 25b, 25c
`and 25d. These signaling/recording and/or fault clearing
`devices are well known in the art and may be con?gured in
`any suitable combination. A visual indicator or audible
`annunciator 25a are used to alert operating personnel (not
`shown) of the presence of a fault. This fault indication may
`prompt action to clear the fault by manually deenergizing the
`faulted line. A recording device 25b such as a digital fault
`recorder, well known in the art, may also be attached to
`monitor output signal 24. In other installations, a remotely
`trippable circuit breaker 250 or other electrically actuatable
`fault clearing device 25d may operate automatically upon
`receipt of the fault indicating output'signal 24 to clear the
`faulted line.
`Alternate embodiments of the present invention may be
`implemented wherein the entire A/D converter 16, data
`pre-processor 20 and ANN 22 are implemented as a dedi
`cated self-contained device. FIG. 4a shows generally, at
`reference numeral 30, the functions of AID converter 16
`(FIG. 1), data pre-processor 20 andANN 22 all implemented
`using a commercial microprocessor 32, AID chip(s) 34 and
`a commercial EPROM 36 customized with all necessary
`program instructions. FIG. 4b shows an embodiment where
`A/D converter 16, data pre-processor 20 andANN 22 are all
`implemented as a single VLSI chip or equivalent chip shown
`as reference numeral 40.
`In yet another embodiment of the present invention, the
`VLSI or equivalent chip may be packaged as part of a
`fault-clearing or circuit protecting device. Output from the
`I-HF detecting circuit chip would then trip the circuit pro
`tecting device directly.
`In still another embodiment of the present invention, input
`data leading to an HIF determination by the ANN 22 would
`
`

`
`5,537,327
`
`7
`be captured and stored. Such data would then be used either
`to further train other ANN s (not shown) or, in cooperation
`with other analysis software not part of the present inven
`tion, to create a set of heuristic “rules” in a form usable by
`more general expert systems. An expert system (e.g., a
`“fuzzy logic” inference engine) using rules so derived could
`replace the ANN 22 as the HIF decision-making component
`of the present invention.
`Since other modi?cations and changes varied to ?t par
`ticular operating requirements and environments will be
`apparent to those skilled in the art, the invention is not
`considered limited to the example chosen for purposes of
`disclosure, and covers all changes and modi?cations which
`do not constitute departures from the true spirit and scope of
`this invention.
`Having thus described the invention, what is desired to be
`protected by Letters Patent is presented in the subsequent
`appended claims.
`What is claimed is:
`1. A method for detecting high impedance faults on a
`power line, the steps comprising:
`a) sensing a plurality of cycles of currents attributable to
`a power line and generating a signal representative
`thereof;
`b) computing a ?rst derivative of cycle-to-cycle maxima
`or minima of said current;
`0) transforming said derivatives into a frequency domain
`representation thereof, the power spectrum of which
`frequency domain representation being usable by
`means for processing information;
`d) applying said power spectrum to means for processing
`information using arti?cial intelligence techniques to
`detect a high-impedance fault and to generate a fault
`signal indicative thereof; and
`e) actuating a switch in response to said fault signal.
`2. The method for detecting high impedance faults on a
`power line in accordance with claim 1, wherein said means
`for processing information comprises a neural network.
`3. The method for detecting high impedance faults on a
`power line in accordance with claim 2, wherein said neural
`network is trained by applying signals representative of line
`currents indicative of a high impedance fault.
`4. The method for detecting high impedance faults on a
`power line in accordance with claim 3, wherein said neural
`network training is accomplished by the use of simulated
`data.
`5. The method for detecting high impedance faults on a
`power line in accordance with claim 4, wherein said simu~
`lation is based on superposition of actual data.
`6. The method for detecting high impedance faults on a
`power line in accordance with claim 3, wherein said neural
`network training is accomplished by applying actual data.
`7. The method for detecting high impedance faults on a
`power line in accordance with claim 1, wherein said trans
`forming step (c) further comprises determining an envelope
`based on minimum and maximum signals and processing
`said envelope.
`8. The method for detecting high impedance faults on a
`power line in accordance with claim 1, wherein said actu
`ating step (e) results in operating a circuit clearing device.
`9. The method for detecting high impedance faults on a
`power line in accordance with claim 8, wherein said clearing
`device is a clearing device disposed along said power line.
`10. The method for detecting high impedance faults on a
`power line in accordance with claim 1, wherein said actu
`ating step (e) results in operating an indicating device.
`
`10
`
`15
`
`25
`
`35
`
`45
`
`50
`
`55
`
`65
`
`8
`11. The method for detecting high impedance faults on a
`power line in accordance with claim 1, the steps further
`comprising:
`f) prior to applying said power spectrum to said means for
`processing information, mathematically conditioning
`said power spectrum.
`12. The method for detecting high impedance faults on a
`power line in accordance with claim 11, wherein step (f)
`comprises the steps of:
`i) summing positive and negative power spectra; and
`ii) averaging said sums.
`13. A