This paper provides an overview of our research into development of enhanced computer support environments for mechatronic design conducted as part of the Concept Database project at the Berkeley Expert System Technology Laboratory. Due to rapid advances in information technology and increased emphasis on life-cycle considerations in product design, design support tools require intelligent filtering, archiving and information retrieval capabilities. The Concept Database is an engineering information system developed to expand the information available to designers and builds upon an architecture integrating hypermedia, a relational database, engineering analysis software and electronic product-data access . Heuristic, deterministic, decision-analytic and case-based methods are developed to provide diverse information navigation strategies to the designer seeking to incorporate life-cycle considerations in his/her design decision making.

INTRODUCTION

The Concept Database aims to provide design teams with improved access to information during the product development phase. Continued funding was made available under the 1994-95 MICRO Program to support development of an improved case representation scheme for mechatronic products leading to an integrated prototype. The Concept Database has been implemented around a product model for mechatronic products - our test domain is electric motor selection - that extends traditional geometry based CAD models to include functional and behavioral specifications, designer annotation, links to theory and analysis as well as considerations of life-cycle issues. Motor selection is a representative example of complex design decision making in the mechatronics domain because of the rapidly changing technology of motors and their applications and the importance of motors as components within embedded computing/mechatronics applications.

SUMMARY OF PROGRESS

The goal of extending abstractions of mechatronic design descriptions to support design reuse as well as integration with design tools from our industrial partners has been substantially accomplished. Since communication among designers is typically not restricted to a single medium, the system acts as a repository of media-rich design information that includes analytic models, text based design documents, CAD drawings, electronic catalog components and design related video and pictures. Such variety in the types of information available has led to the development of both structured and unstructured interfaces for navigation of design information. Various aspects of the catalog component selection problem with uncertain parameters have been previously addressed by Bradley and Agogino[1] and Bradley et. al.[2], using an Expected Value of Perfect Information(EVPI) metric that trades off the value of the designer's time in improving his/her state of information against the potential improvement such information may contribute to the designer's objectives. A multi-objective approach to prototype selection has been developed by Dong and Agogino [3]. These analytic methods are complemented by a case-based-reasoning approach that recognizes that archived design cases can play an important role in placing a new problem in context and provide guidance towards a solution. Consequently, the system framework of the Concept Database provides for storage and retrieval of multimedia design elements at multiple levels of granularity. Remote and concurrent access to stored design information is provided through hypertext interfaces, utilizing standard internet protocols and tools like World Wide Web, gopher and Wide Area Information Services (WAIS). Design elements are indexed and linked through structured database queries as well as relatively unstructured free text queries over verbose description and designer annotation associated with the elements. Machine learning techniques have been applied to the problem of condensing underlying archived information into meaningful concepts for the designer. Enabling seamless exchange of design information between the different modules that constitute the Concept Database architecture has been an essential step in integrating with analysis tools made available by our industrial partners.

SYSTEM ARCHITECTURE

The Concept Database seeks to provide intelligent design support and encourage reuse of applicable archived design information. The user interacts with the system via a hypertext interface that supports both unstructured browsing as well as systematic, iterative refinement of the problem. Broad information-seeking queries like "retrieve information about brushless motors and their torque speed curves" are parsed and mapped onto keywords and text annotations linked to various media elements in the Concept Database.

Information retrieved may include pages of scanned textbooks relating to motor design, images of torque speed curves for the requested motor class as well as links to previous design cases associated with brushless motors. Beyond this point, navigation of available information and judgment of relevance is left to the user. A relational database is used for storing establishing and representing information about the relationships of entities - in this case, information used to select and model engineering components. Access to commercially available motors is provided through an electronic catalog. Engineering models comprising of equation sets have been stored as subsets of a comprehensive list of equations relating to electric motors. Hypertext design documents called templates provide a platform for representing useful problem solving processes or "concepts" in the system. Templates provide integrated access to models, catalog components, preference models, objective functions, and text descriptions that are focused towards a well defined and potentially reusable problem solving methodology. System defined templates encode basic relationships that occur in motor selection applications like wire gauge based motor customization and it is expected that each user may appropriate instantiate these templates to suit his or her application.

DEVELOPMENT OF ANALYSIS MODULES AND INTEGRATION WITH DESIGN TOOLS

Our development effort has focused on integrating mechatronic analysis models with software provided to us by our industrial partners. Type selection is an important design decision made early in the mechatronic design process and a multimedia motor tutorial and prototype expert system have been developed to aid the user in this task [4]. The electronic component catalog and product models associated with the Concept Database have been linked to Design Sheet - a general purpose symbolic and numerical solver developed by Rockwell and capable of operating over very large design models [5].

Extending abstraction levels for encoding life-cycle design information, design rationale and supporting design reuse for mechatronic system design has been a critical goal of this project. We have developed mechatronic design templates for capturing design information at various levels of design decision making. Fig. 2 shows an example of a template that captures the design considerations relating to selection of a motor for a solar powered racing car developed at UC Berkeley for international competition. Access to functional objectives, design models, electronic product data and analysis capabilities in a hypertext environment provide a powerful medium to the designer in which to explore design tradeoffs while being mindful of the life cycle considerations affecting the design.

To support the objective of maximizing product variety in mechatronic design with standard components, we have developed analysis models for customization of DC motors from catalog components. The models developed relate base-level motor variables like magnet-type, wire material, number of wire turns, etc. to catalog parameters like the torque constant, thermal resistance, etc. Relationships between the catalog parameters and performance requirements like torque, efficiency, etc. have also been identified. Design Sheet has been used to obtain a bipartite graph of the D.C. brushless motor model. The graph is viewed in the context of available manufacturing processes to determine which of the base-level motor variables must be fixed and which can be varied. Cost sensitivity information is being used to guide this process.Monotonicity is used to identify variables, the variation of which would not influence the performance requirements favorably. After the problem has been thus simplified, a cost function is developed for the remaining free base-level variables and genetic algorithms have been used as the preferred optimization technique. By viewing the graph in the context of available manufacturing processes and cost sensitivity information, multiple levels of decision making have been accounted for.

CASE BASED INDEXING AND REASONING

Numeric valued design data is only a subset of the range of design information that impacts design decisions and this fact precludes the use of only statistical learning methods to characterize available information. Using direct case experience as well as design concepts "compiled" from design case knowledge has been shown to be a practical and cognitively plausible model for incorporating background knowledge into the design decision making process. Making effective use of case experience also stresses the need for efficient indexing mechanisms. The case based reasoning component of the Concept Database utilizes hyperlinked storage of process, function and artifact centered documents that encode partial solutions, design rationale, models and objectives. These models are WAIS indexed by content [6]. Part of the structure for indexing case experience is provided by the design of our relational database for design entities as well as the implementation of free text queries over design documents. Our research into concept formation from raw design data has shown that machine learning approaches can play a significant role in construction of useful abstractions. Such abstractions may be obtained from structured descriptions of designs as represented in the relational database as well as from the text annotations associated with various design entities. Adaptive Resonance Theory based neural networks have been applied to the task of clustering attribute - value type case data into meaningful design segments at graded abstraction levels[7]. Bayesian belief networks have been applied to the task of creating automated, condensed representations of concepts underlying design text [8]. By merging the processes of design documentation and design data management through the concept of "smart drawings"[9], a framework for managing CAD-based design information information among interdisciplinary design teams has been described. Together, these approaches augment conventional querying capabilities in the Concept Database to provide a powerful suite of techniques by which the designer may utilize available information resources for maximum impact.

Documents related to this project are available on WWW and may be viewed at the URL :

http://pawn.berkeley.edu

REFERENCES

[1] Bradley, S. and Agogino, A.M., "Computer-Assisted Catalog Selection with Multiple Objectives", Design Theory and Methodology - DTM `93, ASME DE-Vol. 53, pp. 139-147., 1993

[2] Bradley, S., A.M. Agogino and W.H. Wood, "Intelligent Engineering Component Catalogs", AI in Design `94, pp. 641-658., 1994.

[3] Dong, Andy, and Agogino, Alice, " A Spectral Optimization Algorithm for Multi-Objective Prototype Selection," in Proceedings of the 1995 ASME Design Theory and Methodology Conference, Boston, MA., 1995.

[4] Huang, Zhijie "An Electronic Product Catalog", unpublished Masters of Engineering Project Report., 1995.

[5] Buckley, M., K. Fertig and D. Smith, "Design Sheet: An Environment Facilitating Flexible Trade Studies During Conceptual Design," AIAA 92-1191 (1992 Aerospace Design Conference), American Institute for Aeronautics and Astronautics, Washington D.C., 1992.

[6] Wood, W.H., "Supplying Concurrent Engineering Information to the Designer: The Conceptual Design Information Server", Ph.D Dissertation, University of California at Berkeley., 1996.

[7] Varma, Anil, Wood,W.H. and Agogino, Alice M.: "A Machine Learning Approach to Automated Design Classification, Association and Retrieval",To appear in the Proceedings of the Fourth International Conference on Artificial Intelligence in Design, June 24-27, 1996, Stanford,CA.

[8] Dong, Andy and Agogino, Alice M.: "Text Analysis for Learning Design Representations", To appear in the Proceedings of the Fourth International Conference on Artificial Intelligence in Design, June 24-27, 1996, Stanford, CA.

[9] Dong, A., F. Moore, C. Woods, and A.M. Agogino: "Managing Design Knowledge in Enterprise-Wide CAD"

, Advances in Formal Design Methods for CAD,(eds, J.S. Gero and F. Sudweeks), Preprints of the IFIP WG 5.2 Workshop on Formal Design Methods for CAD, Key Centre of Design Computing, University of Sydney, pp. 329-347, 1995.

ACKNOWLEDGEMENTS

The MICRO funds have been leveraged by NSF grant # DDM-9300025. We wish to acknowledge the other members of the Concept Database team : Andy Dong and Jorge E. Barreto.