MXPA06008705A - High impedance fault detection - Google Patents

Abstract

An apparatus, system, and method for detecting high impedance faults in electrical power lines using a composite high impedance fault detection system having a voter logic that samples the logical outputs from a plurality of independent high impedance detection systems and determines a high impedance fault if any two of the plurality of independent high impedance detection systems indicates a high impedance fault. Preferably, the plurality of high impedance detection systems include a wavelet based high impedance fault detection system having a first logical output, a higher order statistics based high impedance fault detection system having a second logical output, and a neural net based high impedance fault detection system having a third logical output. Preferably, each of the plurality of high impedance fault detection systems includes an independent high impedance fault detection application that independently detects a high impedance fault on the electrical power line.

Description

DETECTION OF HIGH IMPEDANCE FAULTS 1. Field of the Invention The present invention relates generally to failure detection in electric power systems, and particularly, to high impedance fault detection in power distribution lines. 2. Description of the Prior Art High impedance faults are characterized by a high, impedance at the point of failure. According to the above, a high impedance fault typically produces a small fault current level. High impedance faults can, therefore, generally be defined as those faults that do not carry sufficient fault current to be recognized and cleared by conventional overcurrent devices, such as protective circuit breakers. High impedance faults result when an energized primary conductor contacts an almost-insulating object, such as a tree, structure or equipment, a polar cross-member, or falls to the ground. Typically, a high impedance fault shows arc and scintillation production at the point of contact. The significance of these hard faults to detect is that they represent a serious public safety hazard as well as a risk of ignition of fire arc production. As such, the detection of high impedance faults has been a major protective disjunction interest for a long time. Protective circuit breakers are usually designed to protect equipment (line, transformer, etc.) from damage by isolating the equipment during high current conditions. High impedance faults, typically found in distribution circuits, result in very little, if any, current. High impedance faults do not pose a threat to the equipment and by their nature can not be detected with conventional overcurrent devices. However, the dangers of a downed driver are obvious at all. Possibility of fire, property damage, and someone coming into contact with the live driver are some of the main interests.

Also, legal matters for, promising a comprehensive solution can be very expensive for manufacturers (for example, liability issues). For example, disconnecting the breaker after the high impedance fault is not a clear cut choice. Although high impedance failure is a hazard, disconnection of the feeder unnecessarily will create new hazards by de-energizing homes, traffic signs, offices, etc. The effects of incorrectly determining a high impedance fault will have legal and economic ramifications. The service should always have the safety of the public as a top priority. However, the detection of high impedance faults has not been possible in the past and realistic detection algorithms are not anticipated to detect 100% of all downed conductors, while having 100% safety against bad operation. Services need an economical solution and a system that can reliably detect high impedance faults and are also safe in that they do not easily detect a HIF.

Some of the previous HIF detection schemes included arc production failure detection techniques that used low frequency current components and are described in B. Russell, K. Mehta, R. Chinchali, "An arcing fault detection technique using low frequency current components performance evaluation using recorded data ", IEEE Transactions on Power Delivery, Vol. 3, No. 4, pp. 1493-1500, October 1988. This technique of conventional arc production failure detection examined low frequency currents, mainly 180 and 210 Hz, to determine if a failure of arc production has occurred. A hierarchical detection scheme was developed, which is based on the signature energy and uses a dynamic threshold. Although the algorithm was tested in the field, its performance limited its practical application. U.S. Pat. No. 4,466,071 entitled "High Impedance Fault Detection Apparatus and Metered" describes a conventional high impedance fault detection technique that monitors and evaluates the high frequency components in each cycle of an alternating current. The occurrence of a high impedance fault is determined based on a significant increase in magnitude of the high frequency components for a prescribed period of time and a prescribed pattern. However, this patent only considers a very specific feature of high impedance fault currents and does not consider high impedance fault currents in a multi-resolution structure. U.S. Pat. No. 5,512,832 entitled "Energy Analysis Fault Detection System" describes a conventional detection technique that compares load current energy values to preset thresholds to detect an arc production failure in an energy line. However, this patent is very basic in nature and does not consider or use other load current characteristics to detect a fault. Another HIF detection scheme spotted a neural network detection system that is trained to identify HIF faults and is described in S. Ebron, D. Lubkeman, and M. White, "A neural net ordk approach to the detection of incipient faults on power distribution feeders ", IEEE Transactions on Power Delivery, Vol. 5, No. 2, pp. 905-914, April 1990. Although this study is conducted, in the previous days- of neural network development does not highlight the potential of adaptive learning systems in HIF detection. This pilot study does not include any experimentally generated data but was developed and tested using stimulated HIF data generated using electromagnetic transient program software (EMTP). In addition, the neural network used in that study had a relatively simple architecture. The 'US Patent No. 5,537,327 entitled "Meted and Apparatus for Detecting High-impedance Faults in Electrical Power Systems" describes a conventional high-impedance fault detection technique that uses a neural network to detect high-impedance faults. However, the features used by U.S. Pat. No. 5,537,327, is based on derivatives of the maximum and minimum load current. Also, zero crosses are used to detect a fault. The results of long-term HIF detection tests conducted with the collaboration of five services and including failures in stages, faults that occur naturally, and normal system operation, were published by B. Russell and C. Benner, in "Arcing fault detection for distribution feeders: security assessment in long term field triais ", IEEE Transactions on Power Delivery, Vol. 10, No. 2, pp. 676-683, April 1995. This study describes in more detail some of the practical aspects of HIF detection, but shows that 75% HIF detection was the best performance achievable using its algorithm based on randomness. British Columbia Hydro and Powertech Labs Inc. tested three HIF detection systems including that of Russell et al. , described directly above (see, V. Bucholz, M. Nagpol, J. Nelson, B. Parsi, and W. Zarecki, "High impedance fault detection device tester", IEEE transactions on Power Delivery, Vol. 11, No. 1 , pp. 184-190, January 1996). The most significant result was that the higher frequencies of HIF signatures played an important role in HIF detection and in distinguishing HIF from other types of faults or normal arc production operations. In another HIF screening study, the results of years of experience with HIF detection and testing are summarized and the formal evaluation of the performance of a HIF detection algorithm based on randomness is described. (see, C. Benner 'and B. Russell, "Practical high impedance fault detection on distribution feeders," IEEE Transactions on Power Delivery, Vol. 33, No. 3, pp. 635-640, May / June 1997). Despite this simplicity, and its ad hoc nature, this algorithm provides relatively reliable HIF detection. This technique is implemented in a circuit breaker sold by General Electric known as the Digital Feeder Monitor (DFM) and uses nine algorithms based mainly on energy, randomness, interarmonics, etc. in layers as shown in Figure 4 on page 8 by R. Patterson, W. Tyska, B. Don Russell and B. Michael Aucoin, "A Microprocessor-Based Digital Feeder Monitor with High-Impedance Fault Detection", presented at the 47th Annual Conference for Protective Circuit Breaker Engineers, Texas A &M University, College Station, TX, March 21-23, 1994. In contrast, the technique of the present invention as described in the preferred embodiment uses only three algorithms based on neural networks, statistics and wave trains and has a scheme of choice that determines - the detection of the HIF fault based on each of the algorithms independently detecting a HIF condition. Other publications that describe HIF failure detection studies and techniques are: a. Snider, L.A .; Yuen, Y. S., International Conference on Power Systems Transients 1999, PP. 235-40, a publication of a document given at the conference held on June 24, 1999 in Budapest, Hungary, which presents a 'circuit-breaker based on artificial neural network (ANN) algorithm that uses Fourier analysis to determine harmonic vectors of under order they are then fed to a perceptron or neural network of feed. Networks are trained by feeding them with input vectors consisting of low-order harmonic phase angles or magnitudes. The present invention uses a neural network algorithm that is different since it is a multilayer perceptron that uses the waveform samples as an input, not the discrete Fourier transform (DFT) or fast Fourier transform (FFT) of the low order harmonicas. b. Don Russell, B.; Benner, Carl L., Electric Power Systems Research v 31 n 2 Nov 1994, p. 71-77, which presents an intelligent analysis system that processes the outputs of several algorithms to determine the confidence that a fault exists. This is what is basically implemented in the GE Digital Feeder Monitor (DFM). This system processes the odd, non-harmonic pairs, and uses the Energy and Randomness algorithms for arc detection analysis. The system uses the Arc Confidence Level Generator to generate the level of confidence and the Analyzer of Expert Load Pattern for persistence of arc production. c. Benner, C. L.; Russell, D., Rurar Electric Power Conference 1996 PP. B2 / 38-43 sponsored by Rural Electr. Power Committee of IEEE Ind. Soc. Of Requests. IEEE, a publication of a document given at the Rural Electric Power Conference held April 28-30, 1996 in Forth Worth, TX that focuses on having multiple detection algorithms to successfully balance fault detection with discrimination failure. The document describes the Energy and Randomness algorithms of DFM, as well as the Generated Arc Confidence Level and - Expert Load Pattern Analyzer, but does not go into great detail considering this implementation. The document basically states that multiple algorithms (such as with DFM) improve the detection of these types of failures. The document is to say such commercial development has occurred with GE. Even though this document supports multiple algorithm detection it does not challenge any of the specific algorithms used in the present invention. d. Lazkano, A.; Ruiz, J .; Aramendi, E .; Leturiondo,. L. A., Proceedings of Ninth International Conference on Harmonies and Quality of Poser, Vol. 3, pp. 1005-1010, 2000, "A New Approach to High Impedance Fault Detection Using Wavelet Packet Analysis," a publication of the document given at the conference held in Orlando, FL that presents an HIF detection technique related to train packet analysis cool. It appears to be similar to the wavelength algorithm implemented in the present invention as it uses the wave train, mother Daubechies_4 with 4 level decomposition of the arc current signal. However, it only analyzes the 2nd harmonic value of the current.

The wavelength algorithm of the present invention analyzes current components in the 320-400 Hz region. In this way, the theory and implementations are different and the document only describes theoretical results. and. Al-Dabbagh, M.; Al-Dabbagh, L., IJCNN'99. International Joint Conference on Neural Networks. Proceedings, Vol. 5, pp. 3386-90, 2000, "Neural Networks Based Algorithm for Detecting High Impedance Faults on Power Distribution Lines", a publication of the document given at the conference held July 10-16, 1999 in Washington, DC, which describes an algorithm Neural network detection which is very similar to the neural network detection algorithm of the present invention. However, there is a significant difference between the two algorithms. The algorithm in this document is also a multilayer network with backup propagation, but this algorithm uses DFT of the voltage and current signals as inputs, instead of the individual samples themselves. This document does not indicate the number of inputs, outputs, and hidden layers associated with the network, and tests the algorithm using EMTP / ATP simulations. As can be appreciated, conventional means for detecting high impedance faults in electric power lines are typically not always conclusive and / or reliable and can be expensive. Therefore, a need exists for a new reliable and economical solution to detect high impedance faults in power lines that directs the legal and engineering ramifications to detect and determine what to do once a high impedance fault is detected.

BRIEF DESCRIPTION OF THE INVENTION A method for detecting high impedance faults in electric power lines comprising: providing a plurality of high impedance fault detecting means, each having an output; independently detecting a high impedance fault condition in the electric power lines using the plurality of high impedance fault detection means; and determining a presence of a high impedance fault using a decision means, wherein the decision means determines a high impedance fault if either of two or more independent outputs are indicative of one associated with the plurality of high means. Fault detection has detected a high impedance fault condition. A system for detecting high impedance faults in an electrical power system having an alternating current flowing therethrough, comprising: a supply of electrical energy; one or more interconnected electric power conductors; and a composite high impedance fault detection system connected to one or more electric power conductors for detecting a high impedance fault when at least two of a plurality of individual high impedance fault detection systems, each independently detect the occurrence of a high impedance fault in the electrical power conductors. An apparatus for detecting a high impedance fault in electric power lines comprising: a wave-based system having a first logic output for detecting a high impedance fault condition in the electric power lines; a system based on higher order statistics having a second logical output to detect a high impedance fault condition in the electric power lines; and a system based on neural network having one. third logical output to detect a high impedance fault condition in the electric power lines, the wave-based system, a system based on higher order statistics and a neural network based system detecting each independently the same High impedance fault condition in electric power lines. An apparatus for detecting a high impedance fault in electric power lines comprising: a plurality of high impedance fault detecting means each having an output, each of the plurality of. High impedance fault detection means independently detecting a high impedance fault condition in power lines; and a decision means for determining a high impedance fault if any two or more of the independent outputs are indicative that an associated one of the plurality of high fault detecting means has detected a high impedance fault condition. A protective circuit breaker for electric power distribution lines, comprising: one or more computational devices, only one of the computational devices used to detect both non-high impedance faults and high impedance faults in power distribution lines.

BRIEF DESCRIPTION OF THE DRAWINGS The brief description above, as well as the following detailed description of the preferred embodiments, is best understood when read together with the accompanying drawings. For the purpose of illustrating the invention, it is shown in the embodiments of the drawings that are currently preferred, it being understood, however, that the invention is not limited to the specific methods and instrumentation described. In the drawings: Figure 1 is a schematic diagram of an exemplary electrical power distribution system having a composite high-impedance fault detection system in accordance with the present invention; Figure 2 is a block diagram showing a composite high impedance fault detection system, exemplary according to the present invention; Figure 2a is a block diagram showing a typical embodiment for the composite high-impedance fault detection system shown in Figure 2; Figure 3 is a block diagram showing an exemplary laboratory model developed to experimentally classify high impedance faults and collect data for testing and evaluation; Figure 4 shows a flow chart showing a high-impedance fault detection system based on wavelength; Figure 5 shows a flow diagram showing a high impedance fault detection system based on higher order statistics; Figure 6 shows a flow diagram showing a high impedance fault detection system based on neural network.

DESCRIPTION OF THE PREFERRED MODE (S) (S) The present invention considers high impedance faults in a last-resolution structure. The present invention relates to a new approach for high impedance fault detection that includes a multi-scheme high impedance fault detection scheme by employing a plurality of individual fault detection systems each having its own algorithm application that uses Various characteristics of ground currents and / or phase to individually detect a high impedance fault. The proper characteristics of currents include their wavelength signatures, their fourth order momentum, their sample values as observed by a neural network, and the like. Figure 1 shows a schematic diagram of an electric power distribution system having an electrical power distribution line 10 and a composite high impedance detection system 12. The solid vertical bars in figure 1 are bus bars 16 and represent the interconnection of multiple distribution lines. The composite high impedance detection system 12 includes a plurality of individual high impedance fault detection systems that are not shown in Figure 1 but are shown in Figure 2. Also shown in Figure 1 is the potential transformer PT and The CT current transformer that provides the typical analog inputs for a protective circuit breaker. These individual high impedance fault detection systems have individual algorithms to individually detect high impedance faults. These algorithms can use, for example, wave train, higher order statistics, neural network, and the like to identify the presence of high impedance fault independently of each of the other algorithms of the system. The individual high impedance fault detection algorithms can each have a different confidence level. A fault is identified as a high impedance fault once it is independently detected by the algorithms and processed through a decision logic. Figure 2 shows a composite high impedance fault detection system 12 including a high impedance fault detection system based on higher order statistics 20 identified in Figure 2 as a 2nd order statistical system., a high impedance detection system based on wave train 24, and a high impedance detection system based on neutral network 26. As shown in figure 2, an input connection 28 marked "Acquisition" in the Figure 2, and an output connection 30 marked "Detection Decision" in Figure 2 are provided to communicate an electrical signal between the electric power distribution system and the high impedance fault detection systems 22, 24, 26. For example, the input connection 28 receives an electrical signal from a detector device coupled to the electric power distribution line. The detector device may include any suitable detector device, such as the current transformer shown in Figure 1. The acquisition output 28 is processed through data filtering 29 which provides limited signals per band to each fault detection system. individual high-impedance 22, 24, 26. As shown in Figure 2, each individual fault detection system 22, 24, 26 includes a logic output that communicates with the composite high-impedance fault detection system shown in FIG. figure 2 as "Decision Logic" 32 that determines whether a high impedance fault has occurred. The composite high-impedance fault detection system detects and identifies a fault as a high-impedance fault once it determines that at least two individual high-impedance fault detection systems 22, 24, 26 have independently detected a high failure. impedance. This composite feature provides increased security against false identification while improving the probability of detecting all high impedance faults. Each high impedance fault detection system 22, 24 ,. 26 and its associated algorithm, as well as the composite algorithm are discussed in detail below. The output connection, ie "Decision Detection" 30 of Decision Logic 32 provides the logical output of each of the high impedance detection systems based on individual wavelengths, i.e. the high detection system. impedance based on higher order statistics 22, the high impedance detection system based on wave train 24 and the high impedance detection system based on neural network 26, to the high impedance detection system composed. The high impedance detection system based on higher order statistics 22, the high impedance detection system based on wave train 24 and the high impedance detection system based on neural network 26 and the logic of decision 32 are implemented in a microprocessor that is also used to implement non-HIF detection algorithms such as protection, other than HIF detection, and control algorithms and if desired, measurement and / or monitoring algorithms. In this way, in the present invention, a microprocessor is used to implement both algorithms, HIF detection and non-HIF detection. Figure 2a shows in block diagram form a typical embodiment for acquisition 28, filtration 30, high impedance fault detection system 20 and decision logic 30 of Figure 2 as well as other elements typically associated with a protective circuit breaker. As shown in Figure 2a, acquisition 28 is the combination of the potential transformer PT and current transformer CT shown in Figure 1 whose outputs are filtered by an associated filter 21 and provide a multiplexer 23. The output of multiplexer 23 it is connected by an analog-to-digital converter 25 to the input of the digital signal processor 27. The modality shown in FIG. 2a also includes a memory 32 and a CPU 34 board including a microprocessor 34a, a random access memory 34b and a read-only memory 34c. As described above in connection with Figure 2, each of the individual high impedance fault detection systems 22, 24, 26 shown in that figure are implemented in the microprocessor 34a. Also as described above in connection with Figure 2, the microprocessor 34a is also used to implement non-HIF detection algorithms such as protection, other than HIF detection, and control algorithms and, if desired, measurement and / or monitoring algorithms. The CPU 34 board output, which is an indication that a high impedance fault or a non-high impedance fault condition is determined, is connected to -the alarm 36. Test Dates Figure 3 shows a simplified form of an exemplary laboratory model to be developed to experimentally classify high impedance faults and collect data for testing and evaluation. The exemplified facility includes two 120/4500 V, 1 kVA 42 transformers connected in parallel and energized from a power source 120 V, 15 A, 60 Hz 40. As shown in Fig. 3, an exposed conductor 46 is connected to a terminal of the secondary transformers to simulate a downed transmission line. The other secondary terminal is connected to a copper plate 44 buried in the ground 48-, thus simulating the terrestrial electrode and the earth. The discovered conduit is dropped on a variety of soil surfaces to investigate differences in the resulting currents. The HIF current signatures are collected using a data acquisition system based on the use of signal conditioning and data acquisition boards National Instruments with Lab-VIEW software operating in Windows NT. The data is sampled at 20 kHz, quantized at 14 bits and stored in a binary format. Each HIF test case is conducted for a duration of 50 seconds. Fifteen cases are run for seven different wet surface conditions (wet and frozen turf, soil, asphalt, gravel, sand and concrete) for a total of 105 HIF cases. This data acquisition scheme is also used to collect signatures for non-HIF currents for non-linear single-phase loads (eg, TV, fluorescent lamp, PC, bridge rectifier, phase-controlled motor drive, and arc welder) . A total of 22 non-HIF files is created.

Signal Model The HIF detection algorithms developed in this exemplary study are based on the use of HIF current signatures in all 3 phases and / or terrestrial that are considered non-stationary, varying temporarily, and of varying discharge durations. Even though HIF signals do not look like the third harmonic component of the current, other harmonic components as well as non-harmonic components can play a vital role in HIF detection. One challenge is to develop a data model that recognizes that high impedance faults could occur at any time within the current observation window and could be randomly delayed and substantially attenuated depending on the fault location away from the measurement station. The exemplary model is motivated by previous HIF search, current experimental observations in the lab, and traditionally represents an exact representation of a non-stationary signal with time-dependent spectrum. The HIF detection problem addressed in this exemplary study is formulated as such: Hypothesis HO: r (t) = s (t) + n (t) (1) Hypothesis Hl: r (t) = s (t) + n (t) + f (t) (2) where r (t) represents the terrestrial currents and / or monitored phase. It is assumed that all current records are corrupted with additive Gaussian noise n (t). The signature HIF is denoted by f (t) and represents the instantaneous value of the HIF current. The normal charge signals are denoted by s (t) and thus Hypothesis HO represents a non-HIF situation and Hypothesis Hl represents a HIF situation. High Impedance Fault Detection (HIF) using Wave Rail Based System Figure 4 is a flow chart showing an exemplary waveform based HIF detection application. After the data is acquired at 50, it is filtered at 52, using a bandpass filter of 320-400 Hz, and then, as described in detail below, it is decomposed into low wavelet train decomposition filters. and high step, separated by 54. The energy is then calculated at 56 and the calculated energy is compared with a threshold at 58 to determine if HIF has occurred. The following is an exemplary application of high impedance fault detection using a high-wave-based fault impedance detection system. The continuous wavelet transformation of r (t) is where, the wave train is f (t), p is the position and s is the scale. The position argument keeps track of the temporal change in current harmonics that is essential for HIF detection and the scale change maintains record of frequency bands of the current load. Both position and scale are continuous, therefore the previous transformation is not suitable for computing. A discrete version of the transformation is necessary, which is given by, = S) k (* - (4 where,, m and n are all integers. The above transformation is implemented by multiresolution analysis where the signal is decomposed into a high pass component and low pass through two high pass filters and low pass - known apart - as wavelet decomposition filters. After filtering, both high-pass and low-pass signals are displayed by a factor of 2. The high-pass signal component corresponds to the first detail look of the signal. The second detail look can be obtained by additional decomposition of the low current pass signal in two new high pass and low pass components. The detail signals, third, fourth, etc., may be obtained by further decomposition of subsequent low-pass components. The original signal can be reconstructed with minimal error of its high pass and low pass components in a reverse pyramid fashion. It is in these high-pass components that different HIF characteristics can be located and distinguished from other firms' signatures. non-linear loads of discharge and transient nature. Decomposition filters are associated with the type of mother-wave train used. Most of the exemplary tests of this technique are conducted using the Daubechies-4 wave train which is not a very smooth wave train but requires less computing time. The use of other wave trains or other Daubechies wave trains does not show any observable change in performance or threshold parameters used. The exemplary HIF detection algorithm developed for the wavelength-based system examines advantages of coating the current at different scales and details through a wavelet transformation. Although the appropriate HIF detection can be done using more than a single scale, the experimental test indicated that the energy component of the seventh detail signal carries the most significant HIF information that is more distinguishable from other normal arc loads or normal nonlinear loads. The additional preprocessing required is an FFT to turn the current with all its random delay components to the unresponsive position. In this way, the preferred algorithm is based on evaluating the energy of the seventh detail signal of the magnitude of FFT of a current. That energy is compared with a fixed threshold and with the energy of the previous data segment. The combined decision results in a failure / failure determination. This detection scheme provides approximately 80% detection with approximately a 0.5% phase alarm rate in the absence of arc welding loads. If the HIF attenuation parameters were lower limited to 0.1 (ie, typically high impedance fault detection systems are not interested in detecting very weak currents), the detection rate increases to approximately 95% with approximately a false alarm rate at 0.1%. The detection performance falls to approximately 65% in the presence of arc welding signals and without considering any lower limit in attenuation. The false alarm rate remains below approximately 1%. The seventh detail signal obtained through wave train decomposition corresponds to the frequency range between the second and fifth harmonics (approximately 156-312 Hz). The importance of the proposed detection scheme is to 'consider the third, fourth, and fifth harmonics as well as the frequencies between harmonics as a block of characteristics for HIF detection. In addition, the temporary change in HIF currents is considered in the temporal change of the seventh detail signal. High Impedance Failure Detection (HIF) using Higher Order Statistics Based System Figure 5 is a flow chart showing an HIF detection system based on higher order exemplifying statistics. The data is acquired in 50, filtered in 52 using a bandpass filter of 320-400 Hz. The acquisition and filtering of data in this application are both the same as the acquisition and filtering of data described for the detection system HIF based wave train of Figure 4 and thus have in Figure 5 the same reference numbers as used in Figure 4 for those functions. The energy is then calculated at 60 and the calculated energy is compared to a threshold at 62 to determine if a HIF has occurred. An exemplary detection system and algorithm based on examining the higher-order statistical characteristics of normal currents has been developed and tested, as discussed below. Higher-order spectra, mainly the bispectrum and the tribespectrum, are traditionally recognized as important feature extraction mechanisms that are associated with third- and fourth-order accumulators of random signals. The bispectrum and trispectrum are by definition the. two-dimensional and three-dimensional Fourier transformation of third and fourth order accumulators defined as, C 2 (m, n) = E. { r (t) r (t + m) (t + n)} (5) C 3 (m, n, k) = E. { (t) r (t + m) r (t + n) r (t + k)} (6) where E remains with the expected value. The exemplified algorithm implemented in this study is partly due to Tugnait (see, J. Tugnait, "Detection of Random Signals by Integrated Polyspectral Analysis", IEEE Transactions on Signal processing, Vol. 44, No. 8, pp. 2102-2108 , August 1996) and uses the integrated polyspectros of single phase current loads. This reference is incorporated herein by reference in its entirety. The detector is developed so that a detection decision is made either by using second order statistics in a preliminary stage or by using third and fourth order statistics in an additional stage. The basic concept is as follows: what is the detection decision achievable assuming accessibility to second statistics, third and fourth order for a given set of data and a fixed false alarm rate. First, it is determined if there is a fault using only second order statistics. If the detection can not be done, an alternative test based on accumulators of third and fourth order is activated. Both combined tests are designed so that the false alarm probability is set and predetermined by the system operator. Clearly, this detector uses additional information beyond the power signatures. Preferably, this detector is based on all current spectra including the harmonics between them as generated by the pre-processing filter described above. The HIF detector is detailed as follows: I. Declares a fault, if thirst >; T s Where the signature sed denotes the second order statistics of the data r (t) and Tas is the threshold. thirst is defined as, N s ed =? r n i = 1 (, (7) where, 2 s n is the evidence of r (t) given Hypothesis H (0) which is a situation of no failure. The Tas threshold is chosen so that, p (zA2 o) > ta? ) = ah (8) where denotes the non-centered qui-square distribution of N degrees of freedom. The parameters as and ah are set by the designer so that, as + (l - s) ah = a (9) where, a is the determined false alarm probability. II. If a detection can not be done with the previous test, then the next stage is used. Declare a fault, yes, sh > Tah where, the signature sh denotes the third order statistics of the data r (t) and Ta ^ is the threshold. The Tah threshold is chosen so that, where, *? * (<) denotes the non-centered qui-square distribution of N degrees of freedom. The parameters as and ah are set by the designer so that as + (l-as) ah = a (11) where, a is the determined probability of false alarm. The data of length N is divided into segments L each of length NB. The third order statistics sh is scalar and is defined as, The vector vh is defined as Vh (? M) = Re 2rr m)} \ m ^ C2f r (»)} Re { sC3rr (m)} ^^, (mf (13) The transposed symbol used is. T and lian with m = 1,2,, In this way,? m represents all the spectral components of the registered current. The real and imaginary components are denoted by Re and Im respectively. The inverted matrix D? m0 used in the example above is defined as a diagonal matrix with elements representing the integrated polyspectros of the signal of no failure. lSc3ncin. { < »M)} (14) The integrated bispectral bispectral components are defined as, i K 1 sc, r -j? Cyr. { ? m) R ^% m) NR Go (15) where, Cl2 FFT. { Cir) and the accumulators are defined as, cir-'fr). 2jl)? R2 (,) 16).

Finally, R (Í »¿) is the Fourier Transformation r (t).

The exemplary algorithm is tested using the data collected and the results indicate a detection probability of approximately 97.14% with a zero false alarm rate for a total of 630 cases including two charge welding charges. The thresholds are set so that the false alarm rates c s = 0. 0 5 and ah = 0. 05 correspond to a false total alarm rate of approximately 0.09. Higher order statistics are invoked approximately 4% of the time. These results indicate that higher order signatures are distinguishable from welding and other non-linear loads. High Impedance Failure Detection (HIF) using Artificial Neural Network (ANN) Figure 6 is a flow chart showing application of HIF detection based on neural network. The data is acquired at 50, it is filtered at 52 using a bandpass filter of 320-400 Hz. The acquisition and filtering of data in this application are both the same as the acquisition and filtering of data described for the detection system HIF based wave train of Figure 4 and thus have in Figure 6 the same numerical references as used in Figure 4 for those functions. The samples are transformed into 64 using a fast Fourier transformation (FFT) that is used only in. the second neural network modality described below, and then plotted in the HIF plane at 66 using the neural network algorithm and compared to a threshold at 68 to determine if a HIF has occurred. The following is an exemplary application of high impedance fault detection using a high impedance fault detection system based on neural network. Artificial Neural Networks (AANs) have been used successfully in many applications to solve complex classification problems due to their ability to create non-linear decision limits. The most common and flexible neural network is the multilayer perceptron (MLP) that is constructed from a series of neurons. The first neural networks investigated used 1000 input nodes to the network (100 samples during 3 cycles at 20 Hz). No attempt is made to synchronize the zero crossing of the monitored current to the first input node of the neural network with the hopes of reducing the complexity of implementation. The best results occurred when 200 nodes were used in the hidden layer. The network trains with 600 input / target cases (300 HIF and 300 non-HIF) and had a square-sum error of 1.4 after learning completion (1 missing HIF detection and 0 false alarms). The generalization performance is determined by testing the network in 3600 new 3-cycle windows (1800 HIF and 1800 non-HIF). Considering all network output greater than 0.5 to indicate the presence of an HIF event, the network achieved a detection speed of 70.83% with a false alarm rate at 22.06%. One modality of a neural network design used the spectrum of the window of 3 data cycles. The magnitude of FFT of the 1000 samples is truncated on the harmonic 13. This resulted in a reduction to only 40 input nodes for the neural network. This network had fewer weights and deviations and could be trained almost an order of magnitude faster. The best results occurred when 30 nodes are used in the hidden layer. The network trains with 600 cases and had a square-sum error of 11.8 - (8 missing detections and 4 false alarms). The generalization test on 3600 new entries resulted in approximately a detection rate at 86.06% with approximately a false alarm rate at 17.06%. .. The increased performance of this network over the previous network is probably due to the invariance of the frequency spectrum to phase changes. These performance figures again are based on using approximately 0.5 as the exit threshold to indicate a detected HIF. An attempt is made to reduce the false alarm rate by increasing the exit threshold to approximately 0.75. This resulted in an approximately 83.7% detection rate with approximately a false alarm rate of 14.8%. The threshold increase to approximately 0.95 resulted in approximately a detection rate of 77.7% and approximately a false alarm rate of 11.8%. Another exemplary network architecture was a combination of the two previous networks operating in parallel. If the output of both networks is greater than 0.5, then a positive HIF decision is indicated. A decision that no HIF is present is made if the output of both networks was less than 0.5. For the cases in which the two neural networks do not agree as to the presence of a HIF current, the output of the two networks are summed and a variable threshold is used to make the decision. A threshold of 0.1 corresponded to making the final decision based on which the network was more confident in its own decision. For example, if network output 1 was 0.9867 and network output 2 was 0.0175, then the sum would be less than 1.0 and a non-HIF decision would be made because network output 2 is closer to the ideal value of 0 of which the output of network 1 is of the ideal value of 1. On the other hand, if a more conservative approach is desired in which it is chosen to reduce the false alarm rate, a larger threshold approaching 1.5 would be selected. In essence, a larger threshold gives more weight to the network that indicates a non-HIF situation. Table 1 summarizes the performance as this threshold is varied.

Table 1 . Change in Detection and False Alarm Rates with Threshold The results indicate that the network using the spectrum (FFT) of the monitored current is more capable of detecting HIF than the network using the current current samples. Using the random sample stream network with the spectrum-based network can reduce the false alarm rates, however, it does not seem to increase the detection rate significantly. The lack of synchronization of the zero crossing of the current during training and generalization can prohibit this neural network from detecting any of the patterns or characteristics attributed to HIFs, such as asymmetry of average cycles and variations from cycle to cycle. The results are encouraging given that the detection is carried out in only one fast trip of data of 3 cycles. High Pedal Ignition Detection (HIF) using a Composite System Referring again to Figure 2, an exemplary composite HIF detection system 12 is shown that includes all three different techniques described above. As can be seen from the test results of the three different HIF techniques 22, 24, 26 described above, none of them can detect all HIF faults while ensuring non-false alarms. The present invention evaluates the presence of HIF failure with all the techniques above and uses a multi-resolution structure having a decision logic 32 to detect the presence of high impedance failure. A fault is identified as a high impedance fault once it is independently detected by either of two of a plurality of individual high impedance fault detection systems. An exemplary decision logic is described below: yes (Technique 1 = true); and if ((Technique 2 = true) or (Technique 3 = true)), then HIF = TRUE purpose; also, yes (Technique 1 = false); and if ((Technique 2 = true) Y (Technique 3 = true)) then HIF = TRUE end purpose. where, Technique 1 is the logical output (true or false) of the wave-based algorithm; Technique 2 is the logical output of the algorithm based on higher order statistics; and Technique 3 is the logical output of the ANN-based technique. For the previous example, the logical output of any individual technique is true if this technique detects an HIF, otherwise it is false. It should be understood, however, that although numerous features and advantages of the present invention have been set forth in the foregoing description, together with details of the structure and function of the invention, the description is illustrative only, and changes can be made in detail, especially in matters of form, size and installation of parts within the principles of the invention to the full extent indicated by the broad general meaning of the terms in which the appended claims are expressed.

Claims (13)

1. CLAIMS 1. Method for detecting high impedance faults in electric power lines comprising: providing a plurality of high impedance fault detecting means, each having an output; independently detecting a high impedance fault condition in said electric power lines using said plurality of high impedance fault detecting means; and determining a presence of a high impedance fault using a decision means, wherein said decision means determines a high impedance fault if any two or more of said independent outputs are indicative of an associated one of said plurality of means of High fault detection has detected a high impedance fault condition. Method according to claim 1, characterized in that said plurality of high impedance fault detection means are at least three. Method according to claim 1, characterized in that said plurality of high impedance fault detection means are at least three and each provide a logic output having a state indicative that said associated one of said three detection means has detected a high impedance fault and said decision means is a decision logic and said method further comprises: said decision means determining a high impedance fault if any two of said three logic outputs are in an indicative state to detect a high failure. impedance. Method according to claim 1, characterized in that said decision means is a decision logic and said method further comprises: providing at least three means of detecting high impedance faults having a logic output which in a state is indicative of said associated with said three detection means has detected a high impedance fault; and said decision means determining a high impedance fault if at least two of said at least three logic outputs are in said state indicative of detecting a high impedance fault. 5. System for detecting high impedance faults in an electrical power system having an alternating current flowing through a system comprising: a supply of electrical energy; one or more interconnected electric power conductors; and a composite high impedance fault detection system connected to said one - or more electrical power conductors for detecting a high impedance fault when at least two of a plurality of detection systems. of individual high impedance faults, each independently detects the occurrence of a high impedance fault in said electric power conductors. System according to claim 5, characterized in that said plurality of independent single high impedance fault detection systems further comprise: a wave-based system having a first logic output for detecting a high impedance fault condition in said electric power line; a system based on higher order statistics having a second logical output to detect a fault condition of high impedance in said power line; ... and a system based on neural network having a third logical output to detect a high impedance fault condition in said electric power line, said wave-based system, said system based on higher order statistics and said system based on neural network each independently detecting the same failure condition of high impedance in said electric power lines. System according to claim 6, characterized in that said composite high-impedance fault detection system further comprises a decision logic for determining an occurrence of a high impedance fault, wherein said decision logic determines said presence of high failure. impedance if either of two of said first logic output, said second logic output, and / or said third solid output are in a state indicative of the detection of a high impedance fault condition. The system of claim 5, further comprising a sensing device coupled to one or more of said one or more electrical power conductors for detecting current flow in said conductors and a bandpass filter positioned between said one or more said detector device and said composite high impedance fault detection system. The system of claim 5, further comprising one or more processors that receive and process data indicative of current flow in said one or more electrical power conductors of said detector device and logic outputs of each of said fault detection systems of high individual impedance, and which determines a high impedance fault in said one or more electrical power conductors when any two of said individual high impedance fault detection systems independently detect a high impedance fault. 10. Apparatus for detecting a high impedance fault in electric power lines comprising: a wave-based system having a first logical output to detect a high impedance fault condition in said electric power lines; a system based on higher order statistics having a second logical output to detect a high impedance fault condition in said electric power lines; and a neural network based system having a third logical output to detect a high impedance fault condition in said electric power lines, said wave-based system, said system based on higher order statistics and said system based on neural network, each one independently detecting the same high impedance fault condition in said electric power lines. 11. Apparatus for detecting a high impedance fault in electric power lines comprising: a plurality of high impedance fault detection means each having an output, each of said plurality of high impedance fault detection means. independently detecting a high impedance fault condition in said electric power lines; and a decision means for determining a high impedance fault if any two or more of said independent outputs are indicative that an associated one of said plurality of high fault detecting means has detected a high impedance fault condition. 12. Protective circuit breaker for electric power distribution lines, comprising: one or more computational devices, only one of said computational devices used to detect both non-high impedance faults and high impedance faults in said electric power distribution lines. 13. Protective circuit breaker according to claim 12, characterized in that said only one of said computational devices detects a high impedance fault in said electric power distribution lines by independently detecting a high impedance fault condition in said electric power lines using a plurality of high impedance fault detection means and determines a presence of a high impedance fault using a decision means, wherein said decision means determines a high impedance fault if any of two or more of said plurality of detection means of high-impedance independent faults have detected a high-impedance fault condition.