电机定子故障保护

摘 要：该文研究一种由模糊逻辑和人工神经网络（ANN）组成的，对发电机定子线圈进行故障保护的综合方法。该方法只采用发电机机端电压、电流信号，并对这些信号进行特征提取，然后用到FNN的故障诊断和定位上。该技术由两个阶段组成：基于人工神经网络（ANN）的故障类型分类和基于一个包括模糊逻辑以及人工神经网络的综合网络的进行定子绕组故障精确定位。
关键词：发电机定子保护；人工神经网络；模糊化神经网络；小波分析

1 INTRODUCTION
When internal faults occurs in generator stator windings, there is a possibility of damaging the generator. Some protective relays are used to minimise this damage. However, it is very difficult to locate the internal faults when they happen very close to winding neutral, especially within the region 0% to 5% from neutral. Based on the simulation of generator internal faults and their data analysis, it is apparent that some new techniques are necessary to define the location of faults across the full 100% of the windings. Artificial neural network (ANN)–based techniques [1-3,4,9] have the potential advantage over conventional techniques in significantly improving the accuracy. However, there are still a number of contingencies under which an ANN-based fault location technique’s performance can be adversely affected. The technique presented herein thus proposes the use of fuzzy logic to further improve the accuracy of an ANN-based fault location technique.
This paper mainly discusses and investigates an integrated approach comprising fuzzy logic and ANN.
The technique is based on utilizing voltages and currents at the terminal and comprises of two stages: fault type classification based solely on an ANN and precise location of a fault based on an integrated network comprising fuzzy logic and an ANN. The approach presented here is designed to provide further improvement in the accuracy of the fault detection/ location technique over that presented in [5].
Firstly, data based on feature extraction from voltage and current waveforms is used to train an ANN for fault classification. This ANN can identify the type of faults, such as single-phase ground, two-phase ground, three-phase ground. Secondly, the data is input into a Fuzzified Neural Network (FNN), whose outputs are fuzzified as membership functions to locate where the fault is in the winding. The output of the ANN is fuzzified to increase the accuracy and the dynamic range of the ANN’s output and also to minimise the effect of noise-ridden measurements.
2 SELECTION OF ANN AND FNN
Based on data from simulation studies [6] which consist of three-phase voltage and three-phase current in the time domain, the inputs of ANN must be chosen firstly. If this data is directly input into the ANN, there will be a huge volume of data and a large portion of the data does not contribute significantly to information about faults; thus special features must be extracted [7]; This is commonly known as data preprocessing. The technique adopted here for feature extraction is based on time-domain frequency decom-position of voltage and current waveforms using the Discrete Fourier Transform (DFT). The Power Spectral Density (PSD) of the three-phase voltage and current is calculated based on different harmonics. The overall principle is shown in Figure.1 [8].



When the faults are very close to neutral end, the difference between normal and fault condition is small and is not easily discernible. Nevertheless, when faults occur, the harmonics are still produced, albeit of small magnitudes. The technique presented here is based on utilizing these harmonics in an effective manner.
In order to design a neural network, it is vitally important to train it correctly and then test it. The ANN involved in the first stage is trained with the data obtained from the fault simulation.
The inputs to the ANN comprise of a set of features based on three-phase voltages and three-phase current. With regard to the procedure for feature selection, an acceptable simple criterion used here is that a variable as a feature for the ANN input should provide more information for fault type classification than those not selected. In this respect, a series of studies have revealed that the following frequency components (attained through the previously mentioned time-domain frequency decomposition of the fault waveforms) are representative of the vast majority of different system and fault conditions encountered in practice:
（1）DC component.
（2）Fundamental frequency component.
（3）Second order harmonic component.
（4）Third order harmonic component.
（5）Components over 200-1000 Hz range.
These are then converted into five features for each measured signal: from (1) to (4), the accurate components under which frequencies are used; those associated with (5) above comprise of the summated signal energy over 200-1000 Hz range. With this approach, it becomes possible to confine the number of inputs into the ANN to 30 elements for 6 signals.
In the fault type classification stage, the outputs of the ANN comprise of four variables A, B, C and G (shown in Figure 2). Of these, a value close to unity for any of the first three variables corresponds to the appropriate a, b, or c phases being faulty and a near unity value of G signifies that ground is involved in a fault. This ANN logic is depicted in the following example:




The most critical problem in constructing the ANN is the choice of the number of hidden layers and the number of neurons in each layer. Using too few neurons in the hidden layer may prevent the training process to converge, while using too many neurons would produce long training time and may cause overfitting, thereby causing a large error when the net is confronted with test data. Fig 3 depicts the problem. It is apparent that the choice of hidden neurons should be within the flat region. One hidden layer has been found sufficient to implement most of the practical problems.



Based on the aforementioned, the adopted three layers feed-forward 30 inputs and 4 outputs network yields 11 hidden neurons and the ANN architecture is as shown in Figure 2:
Not withstanding the foregoing, a problem that persists during neural network training is that of ‘overfitting’. The error associated with the training set is driven to a very small value, but when new data is presented to the network, the error is large; this network has memorized the training examples, but it has not learned to generalize to new situations. One method for improving network generalization is to use a network which is just large enough to provide an adequate fit. The larger a network one uses, the more complex the functions that the network can create. If we use a small enough network, it will not have enough power to overfit the data. The problem is that it is difficult to know beforehand how large a network should be for a specific application. There are two other methods for improving generalization: regularization and early stopping. In this paper, automated regularization is used to avoid overfitting.
After the detector has identified the internal faults in the generator, a hybrid fuzzy neural network is fired. The neural network used in this study is chosen as the multi-layer feed-forward network (BP). The output of the neural network is fuzzified to increase the accura-cy and the dynamic range of the neural network’s output and also to minimize the effect of noisy measurements. The network output is defuzzified to indicate the precise fault location.
As in the first stage, the training data is acquired by the data acquisition system. Extracted features from the data are used as inputs to train the fuzzified neural network. The features would be acquired by the PSD of voltages and currents in stage 1. Namely, the inputs of this fuzzified neural network are the same as those in stage 1: the number of inputs is 30; however, the number of outputs are different.
The location of the short circuit is coded into a number of fuzzy membership functions determined by the desired resolution of the fault location. The number of output neurons of the neural network is the same as the number of the fuzzy membership functions. The location is partitioned into six groups. Each output neuron corresponds to the value of the corresponding membership function. For a fault at location 25%, the membership (neural network output) is [0 0.75 0.25 0 0 0]T.
During testing, the output of the neural network is defuzzified where each membership function is weighted by the state of the corresponding output neuron. The weighted membership functions are then added and the centre of mass (first moment) of the sum is the fault location. From Figure 4, the number of output neurons is 6 and the number of input neurons is 30.



There are two methods to locate the position of faults. There is only a little difference between two methods, namely, the different numbers of FNN’s outputs. One is that outputs are just the membership of one single phase. For example, considering phase A, there are six outputs of FNN to locate where the fault is on phase A. The six outputs will locate the position of fault in this phase. The location has been partitioned into six groups. Each output neuron corresponds to the value of the corresponding membership function. The details and results are presented in [5]. The same FNN is also suitable for other phases: phase B and C. So the same FNN is used several times to locate faults in different phases. Of course, the premise is that the three phases are symmetric. The input number is the same as that of Fault Type Classification ANN.
Thus in the first method, the FNN architecture is based on 30 inputs, 6 outputs and 14 neurons in the hidden layer ( shown in Figure 4).Like previous stage, the algorithm used to modify weights and bases is also automated regularization.
3 TRAINING AND TESTING OF ANN AND FNN
The training data includes different internal fault types and different location. The fault types include one-phase-earth fault, two-phase-earth fault, and 3-phase-earth fault. The performance of

[1][2]下一页

——此文章转载于互联网，文中观点与本网站无关，如有侵权请联系删除