The apparatus for electron-beam excited plasma (EBEP) processes with an interface of polymer films was developed, which is composed with a reactor and an electron beam source vessel, which are separated by the polyester film of 1.5μm. The electron source is heated tungsten wire. The acceralated electron-beam was expected to be transmitted through the polymer film according to the space charge limited current law. The electron temperature in the plasma was measured to be 10_??_14 eV by a probe-measurement. When styrene vapor was filled at a pressure of 0.2 Torr, polymerized film was deposited on the substrate, which was confirmed to contain polymerized styrene. But the polymer interface was abrated at the large electron-beam current of 100μA, and the abrated polymer was observed to be deposited on the substrate. In CF₄ gas, 0.5μm lines on mask resist were transfered to Si wafer successuflly by this process.

View full abstract

It is well known that the neural network (NN) have capability to extract the characteristics from input data. The accuracy on the recognition must depend on the truth of characteristics that NN can extract from data by learning. The objective signal before feed into the input layer of NN is usually carried out many kinds of preprocessing. Since the variance of distribution of each class which behaves the same characteristics becomes narrow by means of these preprocessing, the recognition rate is expected to be risen. However, these preprocessing are not sure that to rise the recognition rate of NN, because NN have capability to absorb the some kinds of scattering of input signal. Thus, it is necessary to study how NN deal with the scattering of input signal in order to make the needless preprocessing clear.
In this paper, it is examined how NN learns and recognizes through the Japanese five vowels recognition. It becomes clearly that the width of distribution has on effect on the recognition and the form of distribution effects on recognition. The relation on location between the recognition data and the learning data influences to the recognition rate of NN. Thus, by using NN, the kinds of preprocessing which realize the variance of each class distribution getting narrow simply is not necessary to be perfomred.

View full abstract

A new audio media system utilizes a magnetic field modulation recording method (MFMR) on the optically excited magneto-optical (MO) media. MFMR is advantageous to be overwritable. The high laser power is apt to make the recorded marks expand and extend to the adjacent tracks and to erase prerecorded marks. A range of the write power with which the high quality readout signal is obtainable, is actually limited. We call it the write power margin (PwM).
The shapes and sizes of recorded marks using MFMR largely depend on the thermal characteris-tics of media as well as the ambient temperature. We therefore studied the design of MO media that can retain a specified range of PwM when the ambient temperature changes.
At first the recorded mark sizes were calculated by the heat conduction analysis with finite element method. Then the PwM were measured at various temperature for MO media prepared, and compared with simulated results. We could evaluate the necessitated range of PwM for various ambient temperature and for various media structures.

View full abstract

The Problem of local minima remains as a major subject in the research of artificial neural networks, although the studies have been developed remarkably. Gouhara, et al. have reported the structure of the learning surface where the learning is executed. The results of analyses are important for solving the problem of local minima of multilayered neural networks. One of the interesting results is a classification of minima in view of the intersection of the memory surfaces which are solution sets of simultaneous nonlinear equations.
Using simulations, we verified the results of analyses on minima considered of the intersection of memory surfaces.

View full abstract

In the process control field, the low order lag element plus dead time approximation of the process dynamics and PID control algorithms have been used extensively up to now.
In this paper, we consider a design method of digital PID, I-PD type controller based on the model matching method which minimizes the error integral of the step response between the control system and reference model.
The dynamic characteristic of the system is expressed by the step response of the system, then the optimal parameter of the controller can be easily obtained by the matrix computation. The derived algorithm can be implemented through a simple and systematically applied to the PID, I-PD, PI-PD type controller.
Illustrative numerical examples are presented, and it is found from experimental results that the method proposed here seems to be useful for the practical system.

View full abstract

In order to modify numerical values into linguistic values which fit human sense, an intelligent interface using fuzzy associative inference is proposed. The proposed interface uses associative memory networks and is constructed by a method which makes it easy to represent and to modify fuzzy rules. The associative memory networks use an association matrix M_e (which is a normalized matrix constructed from a correlation matrix M, a bias matrix B, and a scale parameter a) in order to easily carry out refinement and cut-and-paste operations on fuzzy rules, a difficult procedure for conventional systems. For the refinement of fuzzy membership functions, we propose “category based learning”, which carries out separated modifying on a fuzzy label category. Using the intelligent interface, we compose a command spelling corrector based on a multi-layer fuzzy rule set and uses an association matrix M_e to carry out associative inference. The spelling corrector application shows, (1) the usefulness of associative inference for a multi-layer fuzzy rule set using simultaneous bottom-up and top-down processing and, (2) the usefulness of category based learning for deriving fuzzy membership functions, which show human error feature.

View full abstract

This paper proposes the inverse estimation method of current distribution from magnetic fields by a combination method of genetic algorithm (GA) and neural-network. We regarded the estimation problem of current distribution as an optimum allocation problem of currents and used GA and neural-network to solve this optimum problem. Case stdies with computer simulation showed its effectiveness.

View full abstract