Data processing: artificial intelligence – Neural network – Learning task
Reexamination Certificate
1999-03-01
2002-06-11
Davis, George B. (Department: 2122)
Data processing: artificial intelligence
Neural network
Learning task
C706S038000, C706S042000, C706S034000
Reexamination Certificate
active
06405184
ABSTRACT:
BACKGROUND INFORMATION
A method for obtaining fault classification signals characterizing faulty loops in a multiphase electric energy supply network is described in the Siemens device manual “Digitaler Abzweigschutz (Digital Feeder Protection) 7SA511 V3.0,” Order No. C53000-G1100-C98-1, 1995, page 36. In this conventional method, impedance starting occurs in the form of a loop-related starting process. In this process, following the execution of an initial process step for ground fault detection in the case of at least one detected ground fault, the conductor-ground loop, and in the case of no detected ground fault, the conductor-conductor loop, are monitored. A loop is considered started if the corresponding measured impedance vector is within the starting polygon for the particular loop. If several loops are started at the same time, an impedance comparison is made in which only loops whose impedance is no more than 1.5 times the smallest loop impedance are classified as started.
In a method for obtaining fault classification signals characterizing faulty loops described in a relatively old German Patent Application No. P 195 45 267.4, in order to eliminate with a great degree of certainty all of the loops which despite initial starting actually are not faulty, the actual faulty loops are determined through comparison of virtual impedances calculated with respect to the measured conductor-conductor loops, by absolute value and phase, with the impedances determined upon impedance starting if only loops without ground faults are found. If at least one loop with ground fault is found, faultless conductor-ground loops are recognized and eliminated through a comparison of the absolute values of the virtual impedance values formed from the impedance values of the conductor-ground loops detected as faulty with the smallest virtual impedance value formed from the impedance values of the fault-free conductor-ground loops. For further processing of the impedance values of the other, non-eliminated, perceived-as-faulty loops, variously configured test procedures are used in view of the number of simultaneously detected conductor-ground loops. Among such test procedures, the particular test procedure assigned to the particular number of conductor-ground loops detected is conducted.
The dissertation “Einsatz neuronaler Netze im Distanzschutz” (Use of Neural Networks in Impedance Protection], pages 71 through 76 by T. Dalstein, published in “Fortschritt-Berichte VDI” [VDI Progress Reports), Series 21: Elektrotechnik, No. 173, decribes the use of neural networks for the generation of fault classification signals. This network is trained by applying to it sampling values of current and voltage simulated for at least 50,000 malfunctions. The training must be performed individually for the particular installation site in an energy supply system, making manufacturing costs for a protective device equipped with such a neural network extremely high so that it cannot be considered for application in practice. A starting arrangement is assigned to the neural network.
SUMMARY OF THE INVENTION
The present invention relates to a method for generating fault classification signals which designate fault loops in a multiphase energy supply system which form in the case of a fault observed from a protective device with a starting arrangement in which a neural network is used which is trained with input values simulating faulty loops, and in which, in the event of faults, measured values derived from currents and voltages of the loops of the energy supply system are applied to the neural network at its inputs for the generation of fault classification signals in order to receive the fault classification signals at its outputs.
An object of the present invention is to further develop a method which can be carried out at a relatively low cost and therefore can be applied in practice.
To achieve this object according to the present invention, a neural network is used which is trained with input variables simulating faulty loops in the form of normalized resistance and reactance variables formed with consideration of the starting characteristic of the starting arrangement. For the purpose of generation of the fault classification signals in the event of a fault, resistance and reactance measured variables of the loops normalized taking into consideration the starting characteristic are applied to the neural network trained in such manner.
A method for obtaining a fault-identifying signal through a neural network arrangement is described in German Patent No. 43 33 257. In this method, normalized voltage values are supplied to the neural arrangement. A fault-identifying signal is generated with which it is possible to distinguish between a short circuit with arcing and a metallic short circuit. In addition, normalization of the voltage values apparently takes place in the conventional manner and not taking into consideration a starter characteristic of normalized resistance and reactance measured variables which in the method according to the invention is absolutely necessary for a suitable training process of the neural network arrangement.
The same also applies with respect to the use of normalized sampling signals with respect to an additional method described in German Patent No. 43 33 260. In this method, in contrast to the above described method and the method according to the present invention, a starting signal can be obtained in a selective protection arrangement.
An impedance protection device which contains a neural network arrangement as an essential component is described in German Patent No. 44 33 406. This arrangement has a separate neural network for each possible fault type on the monitored section of an energy supply network. Assigned to the neural network arrangement is a fault type identification device which is connected to the input of a device for the extraction of features. On the output side, this device is connected to contact devices which correspond in number to the possible fault types. The outputs of all contact devices are brought to a common output at which, in the event of a fault of a certain type in the monitored section, an output signal of the neural network occurs as a result of the corresponding contact device being triggered, such output signal being provided for detection of this specific fault.
A method for the generation of signals which identify the type of a fault with respect to single-pole faults to ground, bipolar faults with ground contact, bipolar faults without ground contact, and three-pole faults with or without ground contact is described in German Patent No. 43 33 258. Thus fault classification signals which identify faulty loops formed in the case of a fault are not generated. In this conventional method, a single neural network with several neurons in its output layer is used which is trained with current and voltage values normalized in a conventional manner so that in the event of a fault of a certain type, a particular neuron of the output layer transmits an output signal.
The book “Industrielle Anwendung Neuronaler Netze” (Industrial Application of Neural Networks) by E. Schoneburg, 1993, pages 51 and 324, decribes a system in which the measurement—in the framework of a gear unit diagnosis with neural networks—of all upshifts of the gear unit is investigated with a neural network in which the various response times and slip times are normalized.
German Patent No. 43 33 259 describes a method for generating a direction signal which indicates the direction of a short-circuit current. The direction signal indicates whether a short circuit occurred in a first or a second direction from the measuring point. A neural network is used for generating the direction signal. The neural network receives neurons on which subtraction values are used. The direction signal is formed using normalized sampling values of the currents. The neural network is trained accordingly. The normalized sampling values are formed in a normalizing component in the conventi
Böhme Klaus
Jurisch Andreas
LandOfFree
Process for producing fault classification signals does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Process for producing fault classification signals, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Process for producing fault classification signals will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2930460