A novel method to detect plural kinds of shapes such as lines, circles, ellipses, and parabolas has been proposed. It is based on improved Genetic Algorithm (GA). We have developed the following techniques to detect those shapes: 1) each chromosome is interpreted differently: each chromosome represents plural kinds of shapes; 2) detected shapes are eliminated from the original image for efficient detection; and 3) initialization of the chromosomes is carried out during the process for fast detection. We confirmed validity and superior performance of the proposed method by computer simulations.

抄録全体を表示

A method for measuring the wetting angle of molten solder for the solderability test with a high accuracy and a high speed has been developed. By lighting the solder meniscus with a set of LED light sources, the surface angle of the meniscus is obtained from the light source of specular reflection. Brightness-ratio-coded method has been developed for the high-speed angle measurement, and a video camera having an expanded dynamic range is employed for the accurate angle measurement. A trial wetting angle measuring system has been developed using the method, and the ability and the effectiveness were confirmed by experiments.

抄録全体を表示

This paper describes BiCMOS technology with high potential to get the useful devices for analogue/digital mixed LSI such as signal processor for magnetic strage and graphics uses. We have developed simplified self-aligned emitter technology which is processed by reflowed PSG, photoetching for emitters, and emitter diffusion through polysilicon. At the same time, more lightly doped drain technology for nMOS transistors were employed. As a result, bipolar transistors having over 20 V BV_CEO and 2.8 GHz f_T and CMOS having over 20 V BV_SD were obtaind. The CMOS inverter delay time of 1.4ns was also achieved. For flexible design for analogue/digital mixed LSI, the backgate of nMOS transister is seperated from the p type substrate which has most negative voltage. So we get high performance nMOS transis tor not depending on any supply use, for examples, both positive and negative sources.

抄録全体を表示

1/f noise is found in various components and devices and the name 1/f is used because the power spectrum varies as 1/f^α with α close to unity. Many models for 1/f noise have been proposed but specific physical processes mostly remain to be identified. This paper describes a 1/f noise model based on the property of Brown motion, where the probability of first return to the origin is in proportion to the passage time to the minus three-second power. Therefore the paths of Brown motion cross the origin at intervals of 1/f spectrum. A practical example is given to illustrate how the model may be used to generate 1/f noise by electric circuits. An integrator and a comparator are connected in series. The comparator generates pulsed 1/f noise when white noise is applied to the input of the integrator. This circuit can be applied to generate 1/f noise in industrial applications for the low frequency range.

抄録全体を表示

In electric power companies and railway companies etc. with a lot of facilities, maintenance of facilities is a laborious task which should be computerized. In order to realize a computerized facility maintenance system, numerous paper-based facility drawings should be inserted to a database, whose process tends to cost much.
In this paper, we propose a construction method of facility database based on drawing image understanding. This method consists of the following three processes.
(1) Vectorizing process: to input facility drawings through an image scanner, and to analize the image data and to transform them into vector data.
(2) Organizing process: to decide relations among components of the drawings such as texts, lines and symbols, as results of drawing image understanding.
(3) Redaction process: to adjust the vector data and to insert them into a facility database.
In our method, a drawing image understanding is made by fitting figures into ‘relation patterns’ which are adjusted with contents of the drawings. For example, the patterns are relations between facilities and identification numbers, meanings of physical numerals attached to facilities, and so on.
We developed a facility database construction system based on this method. And the results of evaluation proved its effectiveness. The construction cost was reduced to one third compared with that of a former method which relies on human operation.

抄録全体を表示

This paper proposes a new approach to plan cogeneration systems known as one of distributed energy systems. The proposed approach uses structured genetic algorithms. The cogeneration systems planning provides optimal allocation of cogeneration systems, a layout of pipeline network structure for distributing heat energy between cogeneration systems and demand areas, and optimal heat and electric energy supply to meet the energy demands. The planning is formalized as a combinatorial optimization problem with minimizing cost of energy supplying for its objectives. The major method to solve it is based on mathematical programming methods as usual. But it is difficult to get an optimal solution as the number of areas increases because of property of combinatorial explosion and nonlinearity. This paper shows a new method to solve the cogeneration systems planning based on genetic algorithms. A solution of the cogeneration systems planning has a network structure. The proposed method in this paper applies structured genetic algorithms which genotype has a tree structure to represent a network structure. The characteristics of the proposed method are analized by applying the new method to empirical data of the area around K station.

抄録全体を表示

This paper proposes a practical handwriting model to produce character patterns which resemble those written by a human. The practicability of the model has been examined by handwriting simulation and handwritten character recognition by a neural network built with the model. As a successful application of the model, this paper also proposes a new efficient learning of a neural network for handwritten character recognition. Like human learning, the proposed learning acquires excellent recognition ability for unknown character patterns only from a small number of typical character patterns. The recognition rates exceed those by a conventional statistical method. This application not only provides an effective means for handwritten character recognition but also proves the validity of the proposed handwriting model.

抄録全体を表示

This paper proposes a novel neural network which realizes practical recognition ability for handwritten characters by modeling human active learning and recognition. The neural network incorporates a human handwriting model for active learning and a self-organizing feedback mechanism for active recognition. Emphasis is placed on the feedback mechanism which imitates human selective attention to a particular portion of a character. The result of the first rough classification by a structured neural network is fed back to the early processing of an input character. This feedback mechanism is self-organized by using neural network technology. The recognition ability of the proposed neural network has been evaluated for actual field data. The evaluation results have shown that the self-organizing feedback mechanism considerably improves the recognition ability and that the combination of the human handwriting model and the feedback mechanism leads to a practical level of recognition ability.

抄録全体を表示

In the previous paper, we proposed the Adjusting Neural Network (AJNN) for accurate model tuning, which consisted of two multi-layered neural networks, The AJNN had the error calculation neural network (ECNN), which calculated the output error of conventional neural network (CNN) and was added in parallel to the CNN. By subtracting the output of ECNN from that of CNN, the accurate tuning value can be obtained. A remaining problem, however, is that the tuning value is not always accurate. Especially when the output of CNN is large, it is often inaccurate because of improper output of ECNN.
In this paper, we discuss how to obtain more accurate tuning value regardless of the outputs of CNN and propose the extended structure of AJNN, which limits the output of ECNN properly by using a non-linear function unit at the output side of ECNN. The mechanism for improving tuning capability of the extended AJNN is also clarified by comparing the behavior of tuning value with that of the conventional neural network and the AJNN proposed in the previous paper. Finally the suitable non-linear function is investigated by several simulation results with a reheating furnace plant and is found that a simple piecewise linear function, even though it is not optimal, can work effectively to decrease tuning epochs for convergence.
For the extended AJNN, an unique training algorithm is not necessary, which was proposed in the previous paper, for example. It can be created by conventional back propagation algorithm with small execution time.

抄録全体を表示

A 3-D pose determining method for a flexible object which is overlapped complicatedly by itself is presented. By transforming the stereo matching problem of images into the matching of two curvature representations, the 3-D pose of a flexible object is able to be determined. In order to get two curvature representations of an object, the skeletonized images must be coded as chaincodes. Three basic types of self overlappings are introduced and analyses of various complex situations, based on them are discussed. Two dividing principles to code a skeletonized image as short chaincodes and the dividing method in cases that the dividing principles do not work are described. Three of image features and the statistical data of curvature representations are used in the matching process of two sets of chaincodes. Experimental results with real flexible objects show the effectiveness of the method presented.

抄録全体を表示

Recently the need for automation of higher proficiency adjustment is increasing. However, it has taken enormous time to make the adjustment algorithm and timely development of the automation equipment has been missed. The purpose of this paper is to reduce the term for the development of the adjustment algorithm. This paper presents a new construction method of fuzzy neural networks. We propose a method to select the input parameters that can recognize vague waveforms. To operate the learning method easily we propose an automatic learning rate set-up method and also propose a method to avoid the local minimum automatically. We also propose a virtual experiment system using fuzzy neural networks to evaluate the ability of learned algorithm. We have developed a system which integrates the proposed construction method of fuzzy neural networks and the virtual experiment system. We proved the efficiency of the new system by the application of VTR tape running mechanism adjustment.

抄録全体を表示

Higher order derivatives of Universal Learning Network (U.L.N.) has been derived by forward and backward propagation computing methods, which can model and control large scale complicated systems such as industrial plants, economic, social and life phenomena. In this paper, a new concept of nth order asymptotic orbital stability for U.L.N. is defined by using higher order derivatives of U.L.N. and sufficient condition of asymptotic orbital stability for U.L.N. is derived. It is also shown that if 3rd order asymptotic orbital stability for recurrent neural network is proved, higher order asymptotic orbital stability than 3rd order is guaranteed.

抄録全体を表示