Geometric/Knowledge Modeling Challenges

By John M. Switlik, Boeing

The modeling of data, such as geometry or knowledge for CAD/CAM/CAE, RE or KBE, involves trades related to representation, methods, and quality. This article provides a brief look at the background, considers the trades, looks at some practical experience, and lists a few open issues that deserve our attention.

The article uses three major categories important to the subject (CoT, Exp, and Cyb). These are provided mainly to support discussion. CoT includes the sum total of CAD/CAM/CAE/etc. capability plus the human side. The Expertise that drives a system ("Top Gun") is an integral part of the CoT equation. The KBE realm has two components, Exp and Cyb.

• Exp denotes the Expertise as encapsulated in a domain Expert that is more than that of the "Top Gun" Expertise. Exp includes an incredibly large set of talents and knowledge.
• Cyb is the computational part of model (hardware, software, control) that can demonstrate Expertise bordering on intelligent behavior. Cyb may turn out to be subsumed under CoT but that has not been the case to date.

Background
CAD/CAM/CAE (CoT) has data modeling concerns that are not easy to manage. If we look at the progress in computational capability over the years, CoT has kept up by making use of the new ability and in many cases paving the way. The efforts to bring CoT under control have resulted in an every-growing and powerful base.

Those who deal with the practice of CoT can tell us about their successes and about their view of the realities, the challenges and the problems. Recently, a SIAM workshop looked at issues of CoT and reported concerns about the underlying framework. The discussion focus was on the elusiveness of meeting the "tight integration" goals, such as between CAD and CAE. The workshop report stated that systems attained a certain degree of maturity; however, the report also stated that at the heart of the problem there are "some deep mathematical issues, concerned with the computation, representation and manipulation of complex geometries."

One recent addition to the CoT mix is KBE. In its short life, this addition has attracted attention through its potential and power as witnessed by documented savings in time and in real costs. A lot of the benefit of KBE resulted from closing the gap between the Expert human (Exp) and the computer such as to allow Expertise to be exhibited by a system (Cyb). The Exp interacts with computer in a form that is close to software development. In many cases, an Exp was solely responsible for KBE content and required only minimal assistance from computer Experts. There are many cases where the Cyb left unattended essentially acted as a collection of many Exp seats. Undoubtedly, there is much to this phenomenon that we still need to study and understand.

The application of knowledge methods does bring forward a lot of potential benefit. But, it adds to the control requirements through sensitivity to issues of representation and methods.

Representation
Representation deals with technical concerns, such as structure and parameter definitions. Structure would include details of data and function, given the object paradigm. The choices related to representation have a broad influence that includes who can use data, how they can use data, and for what purposes the data can be used.

In terms of "who can use the data," we now need to consider the user as a system entity (Cyb) as well as the usual set of Exp talent. Representation, in the KBE context, requires a broad mixture of structures in order to support the more-smart processing requirement. We have seen an emergence of this the past decade, however CoT has barely rounded the first corner.

"How used" and "for what purpose" relate to the fact that any representation scheme must deal with the need for transformations to other representations. These types of manipulations of data are known by experience to be problematic. An example is movement between vendor systems. Another example is the mapping to presentation elements necessary for visualization. When knowledge is in the mix, issues of explanation complicate interactions with the Exp.

There are other necessary transforms. Just making adjustments in memory to meet specifics of an algorithm/heuristic may not be trivial. This problem relates to the "deep" issues mentioned by SIAM. Of note is that many have the opinion that we have not seen a theory that resolves the problem. On the other hand, practice of the proper sort will help found an improved basis; practice as exemplified by COE members (and other such groups) will be essential.

Methods
Methods deal both the Exp (human) and Cyb (system) elements. In terms of the former, the CoT industry has been putting more power into the hands of the user. One major talent that an Exp will exhibit is judgment which when trained can help resolve the seemingly insurmountable issues of complexity (Exp in the loop). Though lights-out operation has been attained in constrained cases, fully autonomous operations are still mostly a goal (witness autonomous vehicle research). Given the human element in the equation, strategies such as continual improvement are, and will continue to be, essential. How an Exp will interact with the system may become more complicated since the introduction of the knowledge element may cause a problem to exhibit a multi-disciplinary flavor.

Methods on the Cyb side will evolve through time into a collection of very capable agents. How well this works will depend upon how well we do our homework. But, talking of the current state of the art, methods as encapsulated in systems are going to be getting more and more difficult to understand. Two issues of method concern trust and ability.

Trust deals with veracity, such as demonstrated by the supporting basis of the associated knowledge. There is no doubt that systems are difficult to build and to maintain, as the domains of CoT software are not trivial, by any means. One of the criteria of trust would be the demonstration of continued ability through time. Ability deals with method effectiveness and with how the method (knowledge) evolves (or devolves) through its lifecycle. In the computational era, methods imply management of external and internal entities. One ongoing debate concerns the openness of a system to review. The types and kinds of openness of an external system will be clarified and defined further. There is no guarantee that internal systems are open, in the sense that the understanding of a system is not limited how it is expressed or controlled. That is, domain Expertise embedded in CoT is both difficult (deep understanding) and complicated (breadth and depth); this characteristic cannot be managed away.

Capability involves an effective mix of procedures, algorithms, and heuristics. One benefit of KBE has been to provide better means to affect heuristics than was allowed with prior efforts. Increases in computational features lead to increases in interrelationship and complexity properties, thereby necessitating 'truth management.'

Quality
Quality varies by context, is driven by data usage and measurement capabilities, and affects representation and methods. There are many examples of classic trades that exist between some quality criteria. One good example would be trades for a curve or surface (fit, continuity, and footprint). Being 'faithful' to real-world data many times means a discontinuous representation that can have an influence on methods.

Quality criteria can differ by whether the use is interactive or automated. In the case of the former, human judgment can be trained to be part of the process. Examples abound that show that the talents of the Exp (for instance, the human eye) afford a broad evaluative set that can be difficult to duplicate via automation. In the case of the latter, a major concern is how to juggle conflicting constraints dealing with the interplay of Cyb and Exp spaces. That is, explanation must be provided to effect oversight and governance. This problem is related to the trust and ability issues of methods.

But, too, if the Cyb element will be making decisions, those need to be well founded (and worthy of trust). One issue will be the paradigm to support the effort; control engineering's feedback loop is an appropriate metaphor. But, in the KBE case, some elements are of a virtual nature only. How well we know and can control their behaviors is an open issue of importance.

KBE has allowed a better grasp on the subtleties of quality management and will continue to play a key role. The related trades have a dynamic flavor that can only be given proper attention with the appropriate mathematical and computational methods.

An example project has balanced many disciplines in one area of model quality.

Example Practice
Data for modeling can originate different ways. The origination may be from a particular representation that needs to be transformed or improved. The source may be measurement data obtained from real entities, such as a real-world part. Data may result from experiments where the data represents physical phenomena. KBE may deal with the first example. Any method, whether of the Cyb or Exp, will be directly dependent upon the representation. RE is an example domain that has the problem of dealing with data originating from physical measurements. In terms of the last, CAE is very much dependent upon how well this modeling is accomplished.

An example KBE project worked with several types of data requirements which included: design/analysis models at several stages in the life-cycle; sampling/measuring models/techniques related to representing physical parts; Experimental models associated with calculating physical properties (materials). The project integrated internal spline methods with rules expressed using ICAD in the CATIA v4 environment. Work has started to apply CATIA v5 Knowledgeware in the KBE process.

The project implemented knowledge related to geometry (spline) creation and evaluation. The spline methods generated entities that demonstrate optimal smoothness (manifold) properties. The project had a goal of establishing quantitative (computational) frameworks that would ensure the success of applying adaptive control to problem solving. There was specific concern in the areas of representation, methods, and quality.

Strategies for control were several. An Exp in the evaluation loop was helpful. In terms of Cyb control, the project used both mathematical and adaptive approaches. The project, for example, could apply optimization to surface and curve fitting thereby balancing fit, continuity, and footprint. This approach demonstrated the ability to handle problematic data (minimal, sparse, noisy, or originating from multiple/conflicting sources) and has been used to accomplish physical modeling from a difficult set of data.

Project quality for software followed the regime set out by the SEI/CMM. Model (product) quality is more problematic, and efforts continue in this area. Some quality measurements for models are easily established using an Exp in the loop. However, there is a growing set of quantitative measures whose use is Expected to grow.

Conclusion
Three areas of concern to CoT and KBE will be representation, methods, and quality. Some progress in these areas has been reported. Of interest would be joint effort within the industry to establish standard methods and means that can be applied generally, both for the more foundational issues as well as the operational framework needed for KBE success.

For further discussion of the content of this article, please e-mail the author. The content of this paper will be covered in more detail at the COE Fall 2003 Conference & TechniFair.

About the author
John M. Switlik is an Associate Technical Fellow (Boeing, Commercial Airplane, Wichita Define Systems, KBS, john.m.switlik@boeing.com) with experience in the entire software process life cycle and with successful experience in several computational domains. His primary interests are system/integration/quality for life-cycle control, software and methodology (SEI/CMM, applied/computational mathematics), knowledge modeling and management, and advanced techniques (adaptive approaches). Software practice success has been shown with processes that received the SEI/CMM Level 4+ rating.

Abbreviations:
RE - Reverse Engineering
KBE - Knowledge Based Engineering, http://www.ktiworld.com/pdf/understanding-tis-2.pdf
CAD - Computer Aided Design
CAM - Computer Aided Manufacturing
CAE - Computer Aided Engineering
SEI - Software Engineering Institute
CMM - Capability Maturity Model, http://www.sei.cmu.edu/cmm/cmm.html
CoT - CAD/CAM/CAE/etc. includes the sum total of capability plus the human side. How the system is driven ("Top Gun") is an integral part of the CoT equation.
Exp - Expertise as encapsulated in a domain Expert includes an incredibly large set of talents the nature of which we do not fully understand but upon which the sustained success of Cyb may very well depend.
Cyb - The whole shebang of the computational model (hardware, software, control, KBE) that can demonstrate Expertise bordering on intelligent behavior. Cyb with Exp governance provided the early KBE successes. Autonomous Cyb can be a worthy goal, with the appropriate caveats.

Getting Dynamic Results

By Jeffrey Rowe

DaimlerChrysler engineers ditched their old homegrown kinematics and dynamics apps for ultra-modern dynamic simulation and saved time and aggravation.

Three-dimensional, dynamic multibody simulation recently helped DaimlerChrysler (DCX) make significant improvements in its diesel engine design practices. The automaker used dynamic multibody simulation using software from LMS International for improving the valve trains of its new engine designs. Valve trains are vital subsystems of both diesel and gas engines because they control the opening and closing of valves and can greatly impact engine performance. Simply put, valve trains are comprised of a camshaft, rocker arms, valve lifters, and valve springs.

Multibody simulation refers to dynamic motion simulation of a mechanical system. Dynamic motion simulation extends beyond a kinematics simulation (often used to evaluate range of motion and detect interferences), and includes the full dynamic behavior of a system under actual operating conditions, including the flexibility and deformability of discrete bodies.

For more information please visit www.deskeng.com/articles/03/may/index.htm

Variation Behavioral Modeling and Analysis Optimized for Virtual Product Design Through Manufacturing Process Development
By Timothy V. Bogard, President & CEO, Sigmetrix, LLC

Variation behavioral modeling is fast becoming a key component for virtual prototyping and design robustness validation as a part of the overall product lifecycle management (PLM) vision. Other initiatives like 6 Sigma, Design for Manufacturing, Collaborative Design, all require the ability to understand the real world effects of variation sources on the performance and cost of a product. Virtual teams are working on virtual designs with virtual manufacturing process definitions in order to minimize the risk of errors by reducing the need for physical prototypes, while transitioning from design to manufacturing. There are many issues - some of them not readily seen - that must be considered to make this virtual vision a reality.

The magnitude of the cost of tolerancing problems due to variation is not usually fully understood, even in financial terms. A study conducted on over 1000 mechanical drawings demonstrated that over half of all the changes requested to be made after the design was approved for production were the direct result of dimensioning and tolerancing errors. In each case, these errors could have been avoided if variation issues had been better assessed earlier in the design process. Most companies realize that these changes can be expensive, but few seem to have quantified this cost. Here is an example. A typical part is changed at least four times as a direct result of dimension-based manufacturing and design errors. A typical assembly has more than parts. The administrative costs alone of a typical design change to a released part are between $1,500 (all dollar figures are in U.S.) and $10,000. And this does not take into account the costs of scrap or re-working, etc., let alone the potential market costs involved in production delays. In other words, the STARTING COST of the changes made for tolerancing problems on an assembly is between $15,000 and $100,000.

Manufacturing-based variations are not normally considered as a part of the design in the CAD systems, but are a critical part of the overall PLM process. Companies are now realizing that a great design has to comprehend both functional design features and the associated manufacturing process capabilities in order to produce a robust product. Current CAE and CAD geometry systems base all analysis on idealized manufactured and assembled conditions. For example, interference checking is conducted on a CAD model that is assumed to be "perfect." Structural analysis assumes a perfectly accurate form and correct location of parts in an assembly. Variations in products are introduced as functions of manufacturing processes used to produce parts that deviate from nominal on the surfaces of parts. Manufacturing variation then propagates or moves through assemblies as a function of the contacting surfaces in the assembly. This can result in significant cost and quality effects due to product failures or higher than necessary product costs.

CAD assembly and part constraint systems have not been optimized for variation behavior modeling, yet are still assumed to represent the true physical behavior of the assembled parts in real world. CAD systems are optimized to create intelligent nominal designs that can be analyzed in many ways, but not with respect to the effects of mechanical variation. The challenge is to allow the use of CAD data and constraints that are valid for real world variation behavior modeling, while allowing for alternative associative methods for rapidly including real world assembly process and fabrication process effects on the design.

One way to understand this issue is to address the three key fundamentals that define variation behavior modeling (VBM) for analysis:

• VBM is a CAE process that allows the user to analyze the effects of variation in the design and manufacturing process parameters on certain important product performance measurements.
• Variation analysis is based on the concept of generating a population of realized designs and determining the population of measurement results.
• The analysis is obviously limited by the ability of the input parameters to represent the full range of variation that can be realized in a population of manufactured designs.

With the continuing pressure to develop designs more quickly, the opportunities for introducing variations in those design increases. The remedy lies in implementing more intelligent integration of the CAD and CAE tools to support quicker transition to stable production. Understanding the design and geometry capture practices within the CAD model and the information required for manufacturing process modeling with regards to a critical design requirement is fundamental to a great variation modeling and analysis application.

To help understand these concepts, we need to define three more key terms:
Variable -- Variables are the size, location and orientation dimensions of the part features.
Variation -- Variation describes how much the feature variables (dimensions) deviate from their nominal (design) condition. Errors in the manufacturing and assembly processes cause variation in the variables.
Tolerance -- Tolerances convey design requirements and design intent. They do not contain true variation information. Rather, they control one or more variables by defining how much variation is acceptable.

Variation information is collected from a number of sources and is used to relate the variability of the manufacturing processes to the limits of the design tolerances. Tolerance analysis assumes that tolerances can be achieved by manufacturing. Variation behavior based analysis uses actual manufacturing process data as the source of variation values.

In addition, assembly variation models provide an optimized assembly constraint definition based on real-world physical contact without dependency on design intent. When performing a variation analysis of a design, the user's ultimate objective should be to analyze the behavior the assembly will exhibit after production, rather than to analyze the behavior of the constraints used to manage geometry. Some additional concerns include environmental loads, such as a shaft in a bearing resting on one side of the bearing. Fasteners allow features to "float" between them, ensuring that they will not always be located in the same place. Fixtures tend to "adjust" the assembly during manufacturing and can introduce their own error in the assembled conditions.

Variation analysis technologies have been evolving to address these fundamental variation behavior modeling issues. CETOL 6 Sigma is a CATIA CAA V5 Based application optimized to address the ability to understand the effects of variation on design intent. Essentially, it unites the "ideal" world of the model with the "real" world of product design:

• Fundamental to the approach is the use of existing CAD surface information to re-create the physically correct assembly conditions that can automatically understand variation in the model and be used to conduct a variety of critical analysis.
• The first by-product of a robust variation model is the ability to conduct sensitivity analysis with respect to a more precise assembly constraint system.
• Being able to "bias" assemblies from their nominally precise conditions provides unique insight into the critical functional surfaces that must be more closely studied.
• Contribution analysis considers the effect of sensitivities in the design and accounting for the size of the variation as a way to separate the critical few features from the rest.
• Finally, the ability to directly introduce manufacturing process quality data as the source of the variation allows for a very precise understanding of the design through manufacturing process and driving robust design decisions.

In summary, tolerance analysis is a critical part of the PLM process because it reflects the fact that variation behavior modeling and analysis techniques are critical to truly matching design requirements while addressing manufacturing realities.

Sigmetrix Contact Information:

Tim Bogard
President & CEO, Sigmetrix, LLC
105 W. Virginia Street
McKinney, TX 75069
Phone: 972-542-7517 x11
Fax: 972-542-7520
Email: tbogard@sigmetrix.com
Website demo: http://www.sigmetrix.com