Site: Mitsubishi Electric Corporation (MELCO)
Industrial Electronics and Systems Laboratory
8-1-1 Tsukaguchi Honmachi
Amagasaki-shi
Hyogo 661, Japan

Date Visited: March 26, 1992

JTEC Attendees:

Friedland
Feigenbaum
Nii
Chien

Hosts:

Dr. Toshiaki Sakaguchi

Manager, Advanced Systems Group,
Industrial Electronics & Systems Laboratory

Dr. Ikuyuki Hirata

Manager, Strategic Planning Department,
Central Research Laboratory

Dr. Toyo Fukuda

Manager, Advanced Systems Group,
Industrial Electronics & Systems Laboratory

Mr. Shinta Fukui

Asst. Mgr., Computer Control Systems Engineering Sect.,
Power Systems Dept.

Dr. Keinosuke Matsumoto

Senior Researcher, Advanced Systems Group,
Industrial Electronics & Systems Laboratory

Mr. Ichige

Senior Researcher,
Planning Division

ORGANIZATIONAL DATA

Mitsubishi Electric Corporation has four R&D laboratories in the Osaka area: Central Research; Manufacturing Development; Materials & Electronic Devices; and Industrial Electronics & Systems. There are approximately 1,000 scientific and technical personnel within those laboratories. The corporation as a whole has 13 R&D laboratories throughout the country. One of the highlighted accomplishments in its 1991 corporate profile is the development of the MELCOM PSI, a sequential inference machine for AI applications, and MELCOM PSI II, a large-scale integrated version.

There are about 10 AI researchers at the Central Research Lab, and another 20-30 in the Industrial Systems Laboratory.

DISCUSSION

We met with people in the Industrial Electronics and Systems Lab (IESL), a group especially focussed on power industry (electrical) problems. (Dr. Hirato of the Strategic Planning Department of the Central Research Lab was present but with minimal participation.) Therefore, we saw what is going on in a small section (one of a dozen research labs that Mitsubishi Electric runs) of a very large company.

Our hosts were most interested in ES for energy management systems (EMS) as applied to electric power distribution networks. MELCO sees uses for this technology in diagnosis, restorative operation, reliability assessment, dispatching control, and operations planning. There are currently three EMS systems (one from Mitsubishi) in practical use in Japan. The Mitsubishi one is in use at the Kansai power distribution center -- a system with about 200 rules. There are six more systems in field test.

Mitsubishi finds that Japanese power companies are very eager to use ES technology. They were led to believe that U.S. power companies were less interested in the technology even though EPRI appears interested.

IESL has a lot of experience in network diagnostic problems -- an industry survey shows about 50 percent of ESs in the diagnostic area -- thus does not have many failures in this area. In general, where an ES is built as an integral part of a larger system, the failure rate is very low. Mitsubishi considered integration from the very beginning, and thus did not experience problems of integrating stand-alone ESs after they were built.

MELCO has developed a domain-specific tool for network diagnosis, called DASH (Komai 1991 #594). It looks quite sophisticated, using both frames and rules. The tool is built to reuse knowledge (a library of network models), to be maintainable by the end-users, and to be verifiable (verification assumes correct models in the library). One application which was rewritten in DASH reduced the number of rules from 200 to 70-80 by using rules classes categorized by network types. In this area, the important needs for the future are an ability to reason fast with more than 1,000 network nodes and easier/faster knowledge acquisition.

Our hosts discussed the usual issues of system integration and realtime operation.

DETAILS

Our hosts stated that every ES building event has to start from scratch. No domain knowledge is saved in libraries; no task knowledge is saved. Their DASH tool is aimed at both problems. For the former, they are preparing domain specific libraries, e.g., relating to power networks. Structural components in the network are expressed as objects in an object-oriented programming system; functional specifications, however, are represented as state transitions in a transition network. They chose not to use message passing and methods attached to the objects (as is normal in object-oriented programming) because they found it difficult to express causal relations in this representation. We were given a demo of the system.

A system developer spoke on real ESs for energy management applications. There aren't any in practical use yet. The architecture is to put the ES on a workstation and connect it by a network to the computer that runs the EMS. That computer runs a huge program and is overloaded, so the ES needs its own platform.

This ES for diagnosis is a 200 rule system, but when implemented with DASH it is only 70-80 rules (some component knowledge is reusable). The system is fielded at Kansai Electric Power but is not in routine use yet. It speeds up the diagnosis from three to four minutes down to one minute.

MELCO's most successful ES is one that controls group elevator scheduling. This is an on-line real-time scheduling system that helps determine which elevator should stop where to pick up passengers. The typical application has four to five elevators in a 20 story building. The rules employ fuzzy logic. The system was first used two years ago in Kobe -- there are many installations now. The ES has been awarded many prizes, e.g., the Electric Industries Association award. MELCO has also developed a fuzzy logic control system for metal machining which became part of an electron beam cutting machine the company began selling three to four years ago.

Power industry ES applications world-wide in 1991:

Diagnosis               25%
Operations               25%
Monitoring and control each           15%
Planning               10%

Others: education/simulator; maintenance; design; system analysis

Dr. Sakaguchi's wishes for the "second generation:"

  1. Most of all, reusability of knowledge.
  2. Greater ease of maintenance. The power companies tell them that plant people must be able to maintain the system by themselves.
  3. Ease of verification of the system. At present it is tedious.

MELCO is also doing work on optical neuro-computers, but we didn't get into details.

MELCO's best AI systems are the elevator control application described above and the metal machining system.

The company did some work on an expert system for onboard use in control of the power system for the Japanese Experimental Module, to be attached to Space Station Freedom, but funding was cut before final development.

They also have an expert system for diagnosis of a "large rotating machine."

Several years ago a "knowledge media station" -- combination of AI and hypermedia -- was built to demonstrate the value of the PSI machine.

Current AI work is done on UNIX workstations using C, C++, and OPS83.

MELCO representatives expressed the view that ICOT was successful politically, but not technically. They see parallel computation important for simulation, not for AI. ICOT has had no influence on their operations yet. However they feel that it won a political victory in terms of showing the world that Japanese are spending money on basic research.

Note: The four MELCO laboratories at the Osaka location produced a brochure; in it are some interesting and diverse applications (including the fuzzy elevator control system), indicating that ES technology appears to be known in different parts of the company and is used routinely where appropriate.


Published: May 1993; WTEC Hyper-Librarian