Hiroshi Narazaki

A new mill-setup system for hot strip rolling mill that integrates a process model and expertise

This paper describes a new mill-setup system for a hot strip rolling mill in the iron and steel making industry, that integrates mathematical models and operator expertise. Operator expertise is expressed in terms of roll gap differences between neighbouring rolling stands to avoid excessive rolling loads that may lead to shape defects in the rolled material. The problem is formulated as that of minimizing deviation between the know-how-based roll gap difference pattern and the actual one. Applying the system to an actual plant, it was found that operator intervention was reduced by 80%.

An alternative method for inducing a membership function of a category

Abstract We proposed an alternative learning method for category classification knowledge. Our method induces a membership function for a category from positive and negative examples. It can learn “topological knowledge” such as typicality of an example. Our method consists of two stages: example space configuration of a coordinate system in the first stage, induction of membership function that induces a membership function based on the distance in the newly configured example space in the second. Further, we investigate search strategies suitable for deriving the symbolic expression of a category by geometric analysis of the example space.

A case-based approach for modeling nonlinear systems

Abstract We propose a case-based approach for modeling nonlinear systems. Our method is a model-free approach which does not depend on any specific model structure. Given an input to a system, our method predicts the output by interpolating the outputs in the “similar” cases whose inputs are close to the given input. The similarity degree and coefficients used in interpolation are determined from the data in a context-dependent manner. We also propose a method to extract cases from a database to meet the specification on the prediction accuracy. Finally, applications to nonlinear dynamical systems are discussed.

An adaptive fuzzy modeling for continuous galvanizing line

This paper describes an adaptive fuzzy modeling method and its successful application to the feedforward control for a continuous galvanizing line (CGL) in an iron and steel making plant. In CGL, due to the time delay that exists in the process, the feedforward control is indispensable for the accurate control of the solidified zinc of a galvanized steel strip. For this feedforward control to be successful, a model having the following two properties is required: 1) the model should be able to describe the nonlinear characteristics of the process; and 2) the model should have an adaptive capability to maintain accuracy in the long run. In this paper, we present an adaptive fuzzy modeling method that satisfies the above two criteria. The usefulness of our method is demonstrated using experimental results.

A Fuzzy Satisficing Approach to Multiobjective Pass Scheduling for Hot Tandem Mills

This paper discusses the application of the Multiple Criteria Decision Making (MCDM) approach to pass scheduling for hot tandem mills. The multiobjective pass scheduling problem is formulated as a problem where criteria on thickness, temperature, shape, cost, and rolling time are to be optimized. The problem is treated as a Multiobjective Nonlinear Programming (MONLP) problem and Pareto optimal solutions are generated by the Generalized Reduced Gradient (GRG) method. Further, the fuzzy satisficing method is employed, and is proved to be a powerful methodology to reflect the Decision Maker (DM)’s preference in the optimization process.

Preference representation in a content-oriented information retrieval system for local assembly minutes

This paper presents a content-oriented information retrieval method for the local assembly minutes written in Japanese. In this work, the “information retrieval” means the sentence extraction based on our preference specification. In contrast to the conventional keyword-based approach, our discussion puts emphasis on the enrichment of the preference representation capability. We represent preference as a pattern on the syntactic and semantic features. Three preference representation methods are described: One is the rule-based method that requires the direct coding of the preference in the form of the constraints on the syntactic and semantic features. Though this approach enables us to specify the preference in a specification, more intuitive and simpler methods are desired from the viewpoint of application. In order to achieve this, we extend the rule-based method to the example-based and the unification-based methods. Both methods are simple in that preference is specified by example sentence presentation. In the unification-based method, part of the sentences can be left unspecified so that the search engine can fill the appropriate word in the course of search. Illustrative examples are also shown using the actual minutes texts.

Translation and Extraction Problems for Neural and Fuzzy Systems: Bridging Over Distributed Knowledge Representation in Multilayered Neural Networks and Local Knowledge Representation in Fuzzy Systems

Publisher Summary This chapter discusses the translation and extraction problems associated with the integration of qualitative and distributed knowledge representation techniques. The proposed methods bridge the gap between these two knowledge representation schemes as implemented in neural networks and in fuzzy systems. The translation mechanism helps in organizing knowledge into a neural network (NN) structure to exploit its learning capability. The extraction mechanism helps in understanding what the NN has learned from the data by extracting linguistic rules. The information contained in the NN is unreadable without extraction mechanism because of its distributed nature. The method described in this chapter makes possible a hybrid approach. Part of this approach is demonstrated in the incremental learning, where the NN is improved incrementally by adding a new piece of knowledge to the existing NN after it is translated into a NN structure. Another method is possible, whereby the NN based on the data is trained, the rule extraction algorithm to obtain information on the structure is applied, and the knowledge to reconstruct and train the NN is used. The interaction between the data-oriented learning NN and knowledge-oriented fuzzy systems promises to be a fertile topic of future research.

A Method to Use Uncertain Domain Knowledge in the Induction of Classification Knowledge Based on ID3

We propose a method to use uncertain and qualitative domain knowledge in inducing a classification tree based on ID3. We introduce a consistency degree between data and domain knowledge such as “The reduction rate in the latter stage is larger, the quality of a product is usually the better.” As criteria for inducing a decision tree, we use the consistency degree together with the traditional criterion based on the information-theoretic measure. In this work, the consistency degree is mainly used for pruning the hypotheses whose consistency degree with domain knowledge is below a pre-specified threshold. We demonstrate the effectiveness of our method using the data in a superconducting wire manufacturing domain.

Expert system for flatness control in aluminum foil rolling

The application of an expert system for an automatic flatness control (AFC) system is discussed. The expert system adjusts the target shape pattern in the AFC system according to the material characteristics and operating condition, and has shown the ability to improve all kinds of aluminum foil throughout a one-year trial in the operational rolling process. Moreover, the algorithm which maintains consistency among multiple control purposes and the adaptive inference method realized in this expert system proved to be useful for a class of adjustment problems in process control.<<ETX>>