M. Sone

Harmonic detection by using neural network

A methodology of harmonic detection in active power filter from the power line by using neural network is presented. Harmonic detection by using Fourier transformation is an advantageous method to compensate a specific harmonic component. However, it needs input data for each cycle of the current waveform and needs time for the analysis in each cycle. Therefore, the harmonic compensation will be delayed for more than 2 cycles. This paper proposes a new idea to detect harmonic from a given distorted wave by using the partial association of a multilayer neural network. By providing sequentially the amplitude values of distorted wave to neural network, the context of each harmonic component will be detected to each corresponding unit in output layer. In this paper, the detection of 3rd and 5th harmonic components from the distorted waves has been verified by means of the simulation. As neural network can correctly detect the context of each harmonic, we conclude that the proposed algorithm is available to be used for active power filter.

ANN accelerator by parallel processor based on DSP

Artificial neural networks (ANN) become popular in various fields, especially in the pattern recognition. In the recognition stage of ANN the time that is required is very short but in the learning stage it becomes long according to the number of learning data and ANNs scale. This is a serious problem in the case of simulating ANN by an emulation software or a small-size emulation system. So a super computer or general purpose computer is generally needed for pursuing these simulations. A high speed parallel processor was realized by using digital signal processors (DSP) which is effective to high speed data processing. This parallel processor realized a ANN that has three layers architecture with the backpropagation learning algorithm (BP method) at high speed.

Determination of conduction current based on water cluster structures investigated by FT-IR

Because dissolved water in dielectric liquids influences the conduction phenomena, the relation between the water and electrical properties has been investigated. However, the amount of water measured was not sufficient to explain the difference in the change of the dielectric property in different kinds of dielectric liquids. This is because the difference in the state of water as well as its association with the dielectric liquids is not taken into consideration. In this paper, we examined the amount of water as well as the strength of the hydrogen bonding energy of the water combined with the dielectric liquids.

Effect of water in insulating oil on conduction phenomena under high electrical field

The conductivity of dielectric liquids including and not including the water dimer were measured under DC electrical fields up to 100 [kV/cm], and the properties of the conductivity of the two kinds of dielectric liquids compared. The results show that the conductivity of liquid paraffin which includes only free water depends on Ohms law, but the conductivity of di-octyl-phthalate (DOP) which includes a bound dimer water does not depend on Ohms law. The conductivity decreases and increases again according to electrical field. Since the strength of the electrical field at which the relationship changes decreases according to the amount of water in DOP, it seems that the bound water dimer does not effect ion conduction phenomena like free water, but effects the transport of electrons in hopping theory.

A study of a divided learning method

A new learning method is proposed called the divided learning method. It is able to learn a large number of data. The method is based on human learning. Selecting the number of neurons in the hidden layer makes the learning time almost equal to or shorter than that of the usual method. Using this method, learning for character recognition was carried out with NEURO TURBO, a neural network accelerator

Chinese character recognition by the auto recognition system

It is well known that neural networks (NN) have a capability to extract characteristics from input data. So the research into application of NNs for pattern recognition has been very active. However, lots of researches look at recognition for character patterns for under one hundred cases, few researches look at recognition of thousands of character patterns. However, in the case of dealing with general recognition problems, the number of learning data increase. Thus is it necessary to adapt NNs for this problem. In this paper, the authors propose the auto recognition system which can recognize 6709 Chinese characters by plural combined NNs. The system has four parts, rough classification part, fine classification part, recognition part, and auto judgement part. The system classifies input data by character density at the rough and fine classification parts, and recognizes by the multi-layered NN(BPNET) at the recognition part. The auto judgement part judges and sends out Chinese character from four different BPNETs output values. As a result, the authors get 100% recognition rate for 6709 Chinese characters.

Application of recurrent neural network for active filter

Active filters remove harmonic current by pouring in a compensation current which is equal to the quantity of harmonic current with opposite sign. They require a high performance harmonic analyzer. Recurrent neural networks (RNN) have the ability of conversion without affecting phase change. They also learn how to convert load current to fundamental current by themselves. These abilities enable RNN to be applied to the harmonic current analyzer of active filters. We suggest such a use for RNN and investigate their ability to eliminate harmonics. We show that RNN can eliminate harmonic current without being influenced by the composition rate and phase of the harmonic current and that they can work as a high performance harmonic analyzer.

Curvature of the cathode and electron emission

An experiment was carried out to investigate the relationship between the initial electron emission and the effective field strength under a nonuniform field, as in the case of the cathode-anode needle-plane and hemisphere-plane geometries. The curvature of the cathode was altered to change the characteristics of the surface field strength, and the initial electron current was measured. The curvatures were 1.2 mu m to infinity (plane). The range of the current was 1*10/sup -19/ A to 1*10/sup -14/ A in a high vacuum (2*10/sup -7/ torr). The effect of the effective field strength on the initial electron emission under nonuniform and uniform fields was evaluated. The effective field strength is mostly constant at 10/sup 8/ V/cm, irrespective of the cathode shape. In the case of gas adsorption, the effective field strength of the inception of the electron emission is mostly constant.<<ETX>>

Effect of bound water cluster in liquid dielectrics on conductivity

The relation between the conductivity and the binding force is considered by calculating the hydrogen bonding energy among the molecules. As a result, it is shown that the conductivity of hydrous dielectrics is determined by the strength of the hydrogen bonding energy and the amount of water, and an experimental equation for conductivity is derived from this relation.

Effect of water dimer in insulating oil on conduction phenomena

An experimental study of the effect of water dimer in insulating oil on conduction phenomena is presented, and results are given. It is shown that saturated hydrocarbons do not include water dimer which combines with molecules of liquids by hydrogen bonds. The conductivity of saturated hydrocarbons is lower than that of alphatic hydrocarbons. The effect of water dimer on conductivity is larger than the conductivity of saturated hydrocarbons. The effect of water dimer on conductivity depends on the size of the molecules of the dielectric liquids.<<ETX>>