Seats of Knowledge

Knowledge-based CAD (KBCAD), knowledge-based CAM (KBCAM) and knowledge-based CAD/CAM (KBCC) have introduced artificial intelligence (AI) to the design and manufacturing field. The goal of AI is to simulate intelligent behavior in computers. Some of the more generally useful paradigms of this technology include rule-based reasoning, constraint-based reasoning, and biologically modeled approaches such as genetic algorithms and neural nets.

Design
Considerable progress has been made in the use of AI in automating the product-design process. To optimize a design, genetic algorithms, modeled after the process of evolution, have become a popular choice. Sometimes, such as when the user’s knowledge is fragmentary, it helps to have a knowledge-based design system that can acquire knowledge on its own. In such cases another biologically modeled approach, neural net, is becoming popular. There are, however, many alternative techniques available for both optimization and for automated learning.

Indeed, the KBCAD field is broad, with a wide array of approaches. Here, we will concentrate on two KBCAD projects—one at Boeing Co. and one at Ford Motor Co.

Knowledge-based design at Boeing. The Generative Design Project at Boeing, part of DARPA’s Rapid Design Exploration and Optimization project, is a revolutionary approach to generating part designs and exploring design alternatives. This approach combines integrated design information with knowledge-based methods for evaluating and modifying designs. The project focuses on technologies that are useful on a very large scale and on complex designs with tens of thousands of parts. Boeing’s experimental generative design system, Genesis, is currently being applied to the design of aircraft hydraulic systems. It uses a grammatical framework for describing and automatically generating design alternatives.

The Generative Design Project addresses critical aspects of design technology that will lead to the following advances in capability:

• Design representations: Improving geometric, part classification, manufacturing-assembly-sequence, and functional representations of designs to allow better capture of design information and provide a better basis for design evaluation.
• Design rules and context: Controlling which subset of the design rules applies within any given subset of the design. This is critical when generating and evaluating large designs with large amounts of design knowledge.
• Design evaluation: Generic design-evaluation facilities that apply domain-specific evaluators to determine the cost and quality of design alternatives.
• Design exploration: Capabilities to generate designs and to systematically search and explore design alternatives.
• Design comparison and merge: Facilities for interactive and automated comparison and merging of design alternatives.

Knowledge-based design at Ford. IntelliGineering Corp. has developed a knowledge-based design system for Ford Motor Co. Called TransForm, this expert system automates the design of automobile transmissions. TransForm deals with transmissions at various levels of abstraction, including the kinematic, topological, stick-diagram, geometric, and gear-set levels, successively refining its designs from more abstract to more concrete levels. At each level, TransForm uses a similar constraint-satisfaction formulation, differing only in the particular variables, values, and constraints employed. The TransForm expert system has discovered thousands of patentable new transmission designs.

Manufacturing
Knowledge-based manufacturing is perhaps an even bigger field than knowledge-based design because of the variety of disciplines that manufacturing entails, including feature recognition, process planning, fixture planning, assembly planning, scheduling, process monitoring, process control, machine vision, robotics, CNC programming, inspection, diagnosis, and simulation.

In a fully integrated CAD/CAM system, the output of the CAD subsystem becomes the input to the CAM subsystem. Usually the CAD output is a solid 3-D CAD model of the product to be manufactured. The first step in a complete CAM system is the extraction of manufacturing features from such a CAD model. This often requires a knowledge-based approach, because the manufacturing features (the features by which the product is manufactured) are not necessarily the same as the geometric features (the features used to describe the product geometrically).

Led by Dana Nau, researchers at the University of Maryland, College Park, have developed knowledge-based approaches for manufacturing-feature extraction. The researchers have developed a feature algebra that, given one description of a design in terms of machinable features, generates several alternative descriptions. Having multiple alternative feature descriptions available allows more options in forming a process plan and, hence, allows the discovery of a better process plan.

CAPP
Having extracted the set of features to be used to machine a product, the next step is process planning (choosing the optimal order for the corresponding machining operations). Not all sequences are equally effective. Some of the leading researchers in applying knowledge-based approaches to computer-aided process planning (CAPP) include Dana Nau; Paul Wright, Integrated Manufacturing Lab, University of California at Berkeley; Tien-Chein Chang, Purdue University; and Carolyn Hayes, University of Illinois at Urbana-Champaign. (See For Further Reference for researchers’ Web sites or e-mail addresses.)

Paul Wright led the development of the MOSAIC platform, one of the first open-architecture CNC systems, and went on to develop MOSAIC-II. Wright later expanded the scope of the project to include CAD and process planning within the functionality of the controller. The idea was to create a rapid-prototyping system based on machining that was similar to the MOSIS system, currently used by integrated-circuit designers to fabricate chips rapidly. This expanded project was called Integrated Manufacturing and Design Environment (IMADE).

The IMADE system consists of three modules: a feature-based CAD system, a process-planning system, and MOSAIC-II. The process planning in IMADE is done in two stages—a high-level stage called macroplanning and a low-level stage called microplanning. This division reflects the traditional hierarchical division of planning in AI. The macroplanner is responsible for operation and setup sequencing, while the microplanner is responsible for details of each individual operation, from cutting parameters to tool paths.

Although inspired by MACHINIST (an earlier CAPP system that Wright co-developed with Carolyn Hayes), the macroplanning tasks in IMADE are performed in a procedural rather than a declarative way. Using MACHINIST-like rules to generate precedences, the planning task is reduced to a graph searching problem. In fact, much of the code in the macroplanning system in IMADE is devoted to graph algorithms that minimize the number of setups and tool changes in the process plan. The rules in the macroplanner are organized into a frame-based knowledge base and linked to a solid modeler. Much effort was devoted to ensuring that rules had some element of geometric intelligence attached, which earlier systems like MACHINIST lacked. In these systems, the lack of powerful geometric engines forced the systems to represent geometric conditions in terms of awkward logic predicates.

The low-level microplanner in IMADE is similarly knowledge-based. The recommended speeds, feeds, cutting tools, and cutting strategies for various cutting situations are organized into a frame-based decision system. An interesting aspect of the IMADE system is that the macroplanner and microplanner can communicate with each other and iterate to find a valid process plan. This may be necessary, for example, when the setups suggested by the macroplanner require longer cutting tools than the microplanner is willing to accept. Under these circumstances, the microplanner can tell the macroplanner to generate an alternative setup plan, which, while perhaps not optimal, might facilitate better cutting performance according to the rules encoded in the microplanner.

At the University of Illinois, researchers are developing MEDIATOR, a constraint-based, resource-adaptive CAD/CAPP integration system. With a resource-adaptive system, users can modify MEDIATOR easily so that it produces features and manufacturing operations that are appropriate for the resources of their shops. MEDIATOR performs both feature recognition and manufacturing-operation generation for 3-axis machining. The resource-adaptive feature is important, because most process planners produce a single plan to suit all shops. Therefore, some shops may not be able to use the plan because of their particular tools and fixtures.

MEDIATOR intertwines feature extraction and manufacturing decisions; it is part of a larger architecture for CAD/CAPP integration called MAPP. MAPP uses constraint-based knowledge representations and generates planning options through an iterative process of constraint identification, constraint-based search control, and knowledge refinement. Given a CAD model of a part and a description of the available manufacturing resources, MEDIATOR identifies the set of tools and fixture combinations that can be used to create a section of the part and combines surfaces that share the same tools and fixtures into features. The result is a set of features and a set of manufacturing methods that describe how to use the available equipment to create each feature. Since MEDIATOR uses manufacturing information to help define features, the features it extracts are more likely to be manufacturable. Machined areas of the part that cannot be parsed into features are flagged as being unmanufacturable with the processes available.

To adapt MEDIATOR to a different set of manufacturing resources, users need only change the tool and fixture descriptions. MEDIATOR has a Tool Editor that allows users to graphically input their own profiles, cutting surfaces, and cutting motions. This allows users to enter nonstandard-shaped, custom-designed rotating tools. MEDIATOR then can automatically calculate how those tools can be used in a plan. Attempts are being made to adapt MEDIATOR to other material-removal processes, including multiaxis milling, turning, mill/turning, and die-sinker EDM. Future plans include selection and sequencing of the operations found by MEDIATOR into a plan, integration with detailed fixture design and plan-optimization algorithms, and human/ computer interfaces. Work with MEDIATOR has led to the development of a new planning technique called Effecter-Based Goal and Operator Construction. The result is a planner that is easier for users to maintain and to adapt to their own specialized needs.

Tien-Chein Chang of Purdue University has developed a CAPP system capable of even higher level planning than IMADE. This distributed-process planning system is intended for companies with multiple plants scattered around the world, multiple manufacturing cells in each plant, multiple machines in each cell, and hierarchical process planning with local autonomy. The corporate-level process planner checks plant capability and selects the most suitable plants available for the manufacture of a given part. The plant-level process planner checks manufacturing-cell capability, compares cell-level cost, and selects machines for parts by comparing machine-level costs. The machine-level process planner checks tool capability, selects tools for features, and optimizes machining parameters. Chang is developing a computer-aided distributed-process planning system that is suitable for collaborative manufacturing by enlarging the scope of his earlier Quick-Turnaround Cell (QTC) system to the multiple-plants and multiple-machines environment.

Machine Monitoring
Machine vision plays an important role in areas such as process monitoring and control and quality-control inspection. Neural nets, among other knowledge-based techniques, have proven effective in the development of intelligent visual and other types of sensors for machine monitoring and control.

Even during the manufacturing stage itself, there is much room for knowledge-based assistance, especially for in-process monitoring and control. David Dornfeld, head of the Laboratory for Manufacturing Automation at the University of California at Berkeley, has led the development of neural-net, genetic-algorithm, and fuzzy-logic techniques for the detection of tool wear and breakage. One project, being conducted in conjunction with Paul Wright, concentrates on the use of an AI-based supervisory controller for in-process cut optimization. Methods have been developed to use online force-measuring instruments to adjust for tool deflection during cutting and to use data from touch-trigger probes to adjust cutting conditions for better accuracy.

However, using such information to control machine tools has been difficult, because most CNC systems today have proprietary interfacing standards and are therefore closed to external integration. The early efforts of Paul Wright to develop MOSAIC and MOSAIC-II were motivated by the need to integrate his tool-wear monitoring system into the functionality of numerical control.

Sanjay Sarma of Massachusetts Institute of Technology has worked to redefine the G-codes used to program CNC machines. Sarma is developing an open-architecture update of G-codes called OpenG, which is intended to enhance the functionality of G-codes into a full-fledged, event-based programming language with modular plug-and-play interfacing for external devices. OpenG will permit users to program intelligent, self-correcting NC functions, in which the controller can communicate with sensory agents, factory-automation systems, and knowledge bases. OpenG will permit the user to write “daemons” that wait until an exceptional event is reported and then determine and implement corrective action. Sarma hopes to integrate OpenG with a real-time planning system.

When would such a real-time planning system come in handy? Consider this scenario. A common cause of scrap in machining is tool breakage. Visual or acoustic sensor systems can detect imminent tool breakage during a cut. Today, such sensory information cannot be used efficiently; apart from aborting the entire operation, there is no facility for any automated readjustment of the motion commands. In contrast, an intelligent daemon written in OpenG can correct the motion commands in-process. While the cut is in progress, the controller first consults a knowledge base to determine the tools available in the toolchanger. It then communicates with an external replanner to determine the recommended speeds and feeds. Finally, it looks ahead at the upcoming G-codes and inserts a tool-change operation in the middle of the next rapid-move operation. This way, the cutting operation can continue without a stoppage.

	Figure 1: High-level diagram of the proposed NIST METK.

Integrated Manufacturing Systems
Hundreds of manufacturing-engineering software applications have become available over the past decade. These applications can help manufacturing engineers perform the various tasks necessary to transform a product design into a physical reality. Most of these applications, however, have focused on specific engineering functions of the overall product life cycle. An issue facing industry today is that these applications are not designed to work together and cannot be easily integrated. Solving this problem would save significant time in a product’s manufacturing cycle.

Working toward this solution, researchers have begun integrating CAM and CAD into integrated manufacturing-engineering systems—MAPP, IMADE, OpenG, QTC, MOSAIC, and MOSAIC-II are examples of these efforts.

Another approach to the integration problem is the computer-aided Manufacturing Engineering Toolkit (METK), currently under development at the National Institute of Standards and Technology (NIST). The project is a collaborative effort by users, vendors, academic researchers, and representatives of other government agencies. The toolkit is being used for the following purposes:

• To demonstrate that tools are commercially available to perform CAM system engineering;
• To develop a better understanding for individual engineering tools and the overall environment; and
• To identify integration standards and issues that must be addressed to implement plug-compatible environments in the future.

The toolkit consists of commercial off-the-shelf manufacturing-software applications housed together on a high-speed computer workstation. METK is envisioned to be an integration of these applications to support sharing of data between the applications. Figure 1 shows a high-level diagram of the proposed system.

METK is intended to provide product-data and workflow management, process planning, and engineering-data validation using simulation. METK generates and validates the information needed to perform the manufacturing operations required to produce a part. This information contains various elements, including CAD models, process plans, tool paths, NC programs, operations sheets, routing data, setups, tool lists, fixture lists, and machine lists. The functionality of METK will be based on extensions to the capabilities of the commercial off-the-shelf applications included in the toolkit.

The METK prototype currently consists of the following software: 1) Adra Systems’ Matrix™, a product-data-management application; 2) Parametric Technology Corp.’s Pro-Engineer™, a CAD application; 3) Control Data’s ICEM™ Part, a generative process-planning application; and 4) Deneb Robotics’ Quest™ and VNC™, manufacturing-simulation applications used for data validation. These applications reside together and execute on a single UNIX-based Silicon Graphics workstation, located in the Advanced Manufacturing Systems and Networking Testbed facility at NIST. The workstation is connected to the Internet and therefore is capable of file-transfer protocol to accommodate the transfer of data files from other sites participating in the project.

Computers have largely turned the practice of engineering design and manufacturing into CAD and CAM, respectively. AI is now turning CAD/ CAM into KBCC, in which the computer is used not only as a convenient way to represent designs and the processes for manufacturing them, but also to automatically generate and select designs and the processes for manufacturing them. KBCC has made dramatic progress in the 1990s. Important research systems have been developed and some have even been deployed in large companies such as Ford, Xerox, and Boeing. However, very few commercial KBCC products are currently available for small and medium-size companies. The good news is, it appears that commercial and Internet-based KBCC systems will soon be available for the rest of us and will lead to a quantum jump in manufacturing productivity and product quality.

About the Author
Bernard Nadel is president of IntelliGineering Corp., Southfield, MI.