Feature Recognition and Process Planning

I wrote a Ph.D. thesis on Feature Recognition from Solid Models. What that means is that I look for machinable regions in a solid model description of a mechanical part (really, in the difference between the stock and the final part, a.k.a the Delta volume). The feature recognizer then classifies these regions (i.e., is it some kind of hole, slot, or pocket), and extracts the necessary info (basically the solid to be removed with regions that overlap with other features identified) so that a downstream process planner has sufficient information to plan the actual machining of the part.

The main problem my work tried to address was to deal with interacting features. When features interact their topology in the part (i.e., face adjencies) get messed up. For example, a generic slot is toplogically characterized by a floor connected to two walls = "|_|"). However, if another feature is now embedded in, say the floor of the slot, this rule won't hold anymore, and so you end up with either a bunch of rewrite rules, parsers, languages, etc, chasing after face/edge adjacencies.

To circumvent these problems, I used an approach based on hints, where hints can be generated from a variety of sources such as geometry, topology, design features and tolerances. I concentrated on generating hints from geometric information because it tends to be less perturbed by feature interferences. For example, a geometric hint for a slot are two parallel and opposing walls. This configuration is unperturbed by any other feature interfering with either the floor or the side faces, as long as some minimal amount remains of the two parallel and opposing faces (see the Minimal Existence Rule). Even though I concentrated on generating hints via geometric means, it does not mean I dismiss the value of other searches. In fact, this method encourages to start with the easiest and quickest feature search method. For example, one should first start the search with design features and then revert to other methods to find what other methods have not picked up. If someone used a Hole in design, chances are that it will be drilled. But it also means that you cannot blindly trust that information and the feature still has to pass all the "Machining Validity" rules such as Minimal Existence, Accessibility, Non-Interference, and applicable Technological domain (a 20" "hole" is typically not drilled).

The second contribution my work made is to analyze the information we thought a process planner needs to do its job. The feature recognizer extends the volumetric features in the natural sweep directions of the machining feature. This will derive the maximum sized feature of a specific type in that cavity without interfering with the part. The consequence of this extension process is that it collects all the portions of a machining feature that were disconnected by interferences with others. This is useful for machining because you typically want to machine the largest possible region with a cutter.

After the extension process, the system then looks for accessibility regions, and regions where it overlaps with other features or thin air. This has often been ignored though important because it provides the process planner with information on how much, for example, a cutter can overshoot into neighboring regions without gouging the part. (See Optional/Required Volumes).

I implemented these ideas in Knowledge Craft (tm), which is/was an Object Oriented AI Shell built on Lucid Common Lisp, hooked up with the PADL-2 modeler developed at the University of Rochester. The feature recognizer is called OOFF, for Object Oriented Feature Finder, and oddly, it also resembles my exclamation after my defence.

For more information on this work, two of our papers are now available on line:

• Feature Recognition. An introductory paper describing the general feature recognition process.
• Feature Completion. This paper is the follow up of the first paper (though there is some overlap) that concentrates on the feature completion aspect of my work.
• Index to other research by Prof. Requicha and his students.

• JungHyun Han, who worked on incremental feature recognition. He's now Director of   Computer Graphics Laboratoty at SKKU in Korea.
• Bill Regli, who investigated whether feature recognition makes any sense (some cool examples). He's now at Drexel.
• Dusan Sormaz (now at Ohio U), who investigated the Feature Recognition to Process Planning hook up, and developed a prototype system. This is the collection of all of his papers on the net:
• NAMRC
• KREP2
• ECP
• Future
• Steve Brooks at (the former) Allied Signal plant in Kansas City, created a hint based Feature Machining Husk called FBMach that he is/was trying to market. Other FBMach refs: STEPTools.

Here are some general links to other fine academic institutions:

• University of Rochester
• University of Southern Califonia
• UCLA
• Vrije Universiteit van Brussel

Work History

Here are a couple of fancy pictures made by UG. In general, UG is a CAD/CAM package that allows you to design, analyze, and machine mechanical products like car body dies,aerospace parts, molds for shoes soles and computer enclosures. Chances are you are either using, wearing, driving or flying with products designed and manufactured by UG. Pretty cool! Of course, you all probably figured out by now that we generally don't play with fancy shaded immages and texture mappings all day in the R&D groups (unless that happens to be your project of course).

In 1997, I joined Boeing Applied Research and Technology where I'll be working on Automated NC (what else is new) and Generative Design.

Disclaimer: This home page does not reflect the view of any academic institution I attended or any place I ever worked. Any link provided here are informational only and we cannot be held liable for any consequences of reading, browsing, using or copying any of the information held herewith.
Copyright © 1994-2094. Vandenbrande. All rights reserved.