TOWARD HYBRID VARIANT/GENERATIVE PROCESS PLANNING

Transcription

1 Proeedings of DETC 97: 1997 ASME Design Engineering Tehnial Conferenes September 14-17,1997, Saramento, California DETC97/DFM-4333 TOWARD HYBRID VARIANT/GENERATIVE PROCESS PLANNING Alexei Elinson Dept. of Computer Siene and Institute for Systems Researh University of Maryland College Park, MD Jeffrey W. Herrmann Dept. of Mehanial Engineering and Institute for Systems Researh University of Maryland College Park, MD Ioannis E. Minis Dept. of Mehanial Engineering and Institute for Systems Researh University of Maryland College Park, MD Dana S. Nau Dept. of Computer Siene and Institute for Systems Researh University of Maryland College Park, MD Gurdip Singh Dept. of Mehanial Engineering and Institute for Systems Researh University of Maryland College Park, MD ABSTRACT This paper desribes our ongoing work on a hybrid approah to proess planning, that attempts to ombine the best harateristis of both variant and generative proess planning while avoiding the worst limitations of eah. Our approah uses a database of designs and proess plans that are lassified using design signatures, graphial strutures based on detailed produt design attributes that are more meaningful and aurate than GT odes and an be omputed automatially from the designs stored in the database. We are developing ways to use design signatures to lassify and retrieve slies of designs and plans, so that when a proess plan is needed for a new design, we will be able to retrieve the plan slies that are most relevant, and ombine them and modify them to produe a plan for the new design. INTRODUCTION In proess planning pratie, variant tehniques are the tools of hoie: they urrently support almost all pratial implementations of omputer-aided proess planning. Several variant proess planning systems are ommerially available and have provided signifiant benefits but despite the relative popularity of this approah, variant proess planning has some well known drawbaks. A generative proess planner that provides realisti proess plans for a reasonably wide spetrum of produts would make a great impat on industrial pratie. Thus, a great deal of researh has been done on generative approahes, and a number of experimental systems have been developed for various aspets of proess planning. However, generative proess planning has proved quite diffiult. Most existing systems work only in restrited domains, and have not really ahieved signifiant industrial use. This paper desribes a hybrid proess planning approah that we are developing. This approah attempts to ombine the best harateristis of both variant and generative proess planning while avoiding the worst limitations of eah. As shown in Figure 1, our approah involves the following steps: Create a database of designs and proess plans similar to a variant database but instead of using GT odes to index and lassify the entries in the database, use detailed produt design attributes that are more meaningful and aurate than GT odes and an be omputed automatially from the designs stored in the database. Given a new design for whih a proess plan is needed, retrieve relevant proess planning information from the database but unlike traditional variant proess planning, this information is not a single plan but inludes instead portions or slies of several plans, eah of whih is relevant for a different portion of the design. Use generative plan-merging tehniques to ombine and modify the retrieved plan slies in order to synthesize a proess plan for the new design. This paper desribes the basi approah we are developing for eah of these steps. Some of the steps are implemented and working, and some of them are still work in progress. 1 Copyright 1997 by ASME

2 design analyze design attributes design signature further analysis signature slies database retrieval proess plan for the design shown above merge and modify proess plan slies database, indexed by a lassifiation tree of the signature slies Figure 1. Hybrid variant/generative proess planning. BACKGROUND Classifying Designs Group Tehnology. Group Tehnology (GT) is a manufaturing philosophy that was first introdued by Russian gun manufaturers during World War I and now widely used in industry [Mitr66]. Group Tehnology involves lassifying similar produts into groups in order to ahieve eonomies of sale normally assoiated with high-volume prodution. In order to implement GT, one must have a onise oding sheme for desribing produts and a method for grouping (or lassifying) similar produts. Many researhers have developed GT oding and lassifiation systems, inluding the popular Opitz, DCLASS [Bond88], and MICLASS [Orga86] shemes. In eah ase the basi idea is to apture ritial design and manufaturing attributes of a part in an alphanumeri string, or GT ode, that is assigned to that part. The typial GT ode [Hout75] onsists of two types of positions. In one ase, a position desribes some global property of the design suh as material, size, type, funtionality, et., and its meaning is ompletely independent of what values are stored elsewhere. In the other ase, a position represents some details that are relevant only for ertain types of designs, and thus its meaning depends on the values of other positions. GT lassifiation shemes are essentially tables and rules that help a designer determine the GT ode of a part from a drawing manually. One an use a database of the GT odes for design retrieval, variant proess planning, and other manufaturing appliations. Sine the 1980's several researhers [Shah89, Hend88, Srik94] have worked on automating this manual proess for lasses of mahined parts. One issue is the relevane of GT methods to speifi real world design retrieval problems. As it has been used during the last 35 years it works but it has an inherited drawbak: desribing designs as short strings reates a oarse lassifiation sheme. Moreover, from the beginning GT oding was intended to be human interpretable, hene the typial questions desribe somewhat subjetive human impressions of 2D drawings. This has aused diffiulty in automating the generation of GT odes. Geometri Approahes. Another possible basis for lassifying designs is to use geometri properties of solid and CAD models. Most of today's CAD/CAM systems use either onstrutive or boundary models to represent solids. The use of CSG trees as a way to lassify designs has two appealing harateristis: the analogy between volumetri CSG primitives and the volumes of material removed by mahining operations, and the ready availability of CSG trees as a basi representational sheme in several geometri modelers. However, the approah suffers from two drawbaks. First, the CSG representation for a design is not unique and a robust method for omputing a unique CSG representation for a design has not yet been found (many believe that suh a method simply annot be found [Lee87]). Seond, the CSG primitives that would be involved in suh a representation do not neessarily orrespond to the manufaturing operations that would be used to manufature the design and thus the lassifiation might not be very useful for manufaturing pratie. As far as we know, no methods to measure similarity on the basis of CSG trees were developed. Sun et al. [Sun95] have desribed a similarity measure for solids based on properties of their boundary representations. The approah involves representing a polyhedral approximation of a solid using a graph, in whih the verties orrespond to faes of a solid and have labels apturing the faes' orientation and area, and the edges orrespond to the adjaeny relation between solid faes, and are labeled by the orresponding solid angles. To ompare two solids they use a sophistiated algorithm to take the graphs of these solids and map them into eah other in suh way that the area and orientation of orresponding verties are as lose as possible. The results of suh mapping are expressed as a real number in a range from 0 to 1. As a new measure of relaxed geometrial similarity their work looks very interesting, but there are several diffiulties to be overome before it an be useful as a lassifiation sheme for manufaturing: As desribed in [Sun95], the method works only with polyhedral objets any non-planar faes of the designs must first be replaed with planar approximations. This may ause diffiulty in lassifying solids with a signifiant number of ylindrial or sulptured surfaes. The measure of similarity is not symmetrial (similarity between solids A and B is not equal to the similarity 2 Copyright 1997 by ASME

3 between B and A). This will ause diffiulties in using it as the basis for a traditional database indexing sheme, sine suh shemes assume a symmetrial measure. As desribed in [Sun95], the method does not inorporate (or reflet) manufaturing onsiderations, suh as approahability, fixturing, and operation interferene; and we do not see any obvious way to add them. Proess Planning The two primary approahes to omputer-aided proess planning are the variant and generative approahes, whih are desribed briefly below. For a more detailed survey of variant and generative approahes, see [Ham86, Ham88]. Variant Proess Planning. Variant proess planning (see Figure 2) is based on the use of the Group Tehnology oding shemes desribed earlier. Given a new design D for whih a proess plan is needed, the proess engineer first determines a GT ode for D, and then uses this ode as an index into a database to retrieve a proess plan P for a design D similar to D. One this is done, the proess engineer modifies the retrieved proess plan manually to produe a plan P for the design D. Some of our group s work on variant proess planning inludes [Cand95, Cand96, Iyer95]. In proess planning pratie, variant tehniques are the tools of hoie: they urrently support almost all pratial implementations of omputer-aided proess planning. However, variant proess planning also has some signifiant drawbaks. If the part mix varies over time, then for a new proposed design it may be diffiult to find existing designs in the database that satisfy similar design speifiations or require similar manufaturing proesses. Furthermore, if the proess plan retrieved by the variant system uses out-of-date proesses, then these will propagate to the proess plan for the new design unless the proess engineer makes a point of replaing them. Finally, a drasti redution in the bath size of an existing produt may require re-planning to derive an eonomially sensible proess plan. Generative Proess Planning. Given a new design D for whih a proess plan is needed, a generative proess planning system attempts to synthesize a proess plan diretly for D. For mahined parts, the typial approah is to do the planning on a feature-by-feature basis by retrieving andidate proesses from the manufaturing knowledge repository, seleting the design GT oding sheme feasible proesses on the basis of geometri and manufaturingrelated onstraints, and ombining the hosen proesses in a proper sequene. A great deal of researh has been done on generative approahes, and a number of experimental systems have been developed for various aspets of proess planning [Mant89, Kamb93, Gupt94a, Yue94]. However, generative proess planning has unfortunately proved quite diffiult. Diffiulties arise from interation among various aspets of the problem, suh as workpiee fixturing, proess seletion, and proess sequening. As a result, most existing systems work only in restrited domains. Although one generative system, the PART system [Geel95]) is being marketed ommerially, generative systems have not really ahieved signifiant industrial use. Even in the absene of omplete and omprehensive solutions to the entire proess planning problem, generative proess planning tehniques an be useful in design for manufaturing [Boot94], in whih the designer tries to take manufaturability onsiderations into aount during the design stage. For example, by generating and evaluating operation plans for a part, it is possible to give feedbak to designers about possible manufaturability problems with the part, and/or to suggest hanges to the part that may improve its manufaturability [Mant89, Gupt94a, Gupt95, Das95, Lam95, Hebb96]. Hybrid Approahes. By a hybrid approah, we mean any approah that attempts to exploit knowledge in existing plans while generating a proess plan for a new design. Though some approahes have been proposed (two are desribed below), researhers have not yet developed omprehensive solutions: Park et al. [Park93] desribe an approah for aquiring knowledge useful for generating proess plans. Given a proess plan for a design, it uses inferene rules to find the explanations behind the plan (what part of the plan did what). Then it stores the knowledge as a shema, whih desribes how in general to make some olletion of features. Planning is done by seeking the relevant shema and inserting the neessary values to onstrut a valid plan. A relevant shema is one with the same olletion of features. This is a very simple design similarity measure: it uses no other manufaturing information (suh as preedenes or toleranes) to identify the relevant shema. Marefat and Britanik [Mare94] propose a hybrid approah that aptures plan knowledge that speifies the proesses GT ode database retrieval proess plan for the design shown above modifiations by the user proess plan for a similar design database of plans, indexed by GT odes Figure 2. Variant proess planning. 3 Copyright 1997 by ASME

4 Figure 3: A mahined part d 0, and a piee of stok from whih to make it. The part is an adaptation of a part design that Boeing Airraft ontributed to the NIST design repository. neessary to make a ertain feature (with a speifi size, hardness, surfae finish, and toleranes). Planning deomposes a design by generating subplans for eah feature and then searhing the old subplans for the most appropriate one. The most appropriate old plan is the one that makes a feature that is most similar to the new design's feature. Similarity here is hierarhial: the feature must be the same type, then the same dimensions, then the same toleranes. Corresponding to eah level of feature properties are proess apabilities. Beause the new and old features will be different at some level, the old plan is modified: the planner keeps the proess information that orresponds to the levels at whih the old and new features are idential, disards the remainder, and generates new information using proess apability rules. This approah assumes that eah feature an be made independently and thus does not group features in any way. As these examples show, the existing hybrid approahes have limited apabilities. A robust hybrid approah must onsider feature interations, preedenes, toleranes, and other ritial design information that impat proess planning. In addition it must onsider how to store, lassify, and retrieve useful design and proess planning information. OUR APPROACH Overview As desribed in the following setions, the overall approah inludes two major phases: Preproessing: given a database of existing designs and proess plans, build an indexing and lassifiation struture for searh and retrieval. Planning for new designs: given a new design, use the lassifiation struture to retrieve relevant planning information from the database, and use this information to synthesize a plan for the new design. Preproessing Design Signatures. Given a set of CAD designs and proess plans for those designs, we want to organize them into a database similar to a variant database. However, instead of using GT odes to index and lassify the entries in the database, we are developing a more detailed struture alled a design signature. A design signature is a graph struture that Figure 4: Volumetri mahining features for the part d 0 of Figure 3, found by the F-Rex feature extrator. 4 Copyright 1997 by ASME

5 D 4 : z D 4 : -z M 16 : z D 7 : -z D 7 : z M 2 : z M 11 : z M 1 : z M 10 : z M 0 : z M 9 : z D 0 : z D 0 : -z M 14 : x, -x M 15 : x, -x D 3 : -z D 3 : z D 5 : z M 5 : z M 4 : z D 5 : -z s, M 12 : z, -z D 8 : x D 8 : -x M 13 : z, -z s, D 9 : x D 9 : -x D 6 : -z D 6 : z M 8 : z M 7 : z M 3 : z M 17 : x, z M 6 : z D 1 : z D 1 : -z D 2 : -z D 2 : z Figure 5: The basi signature s 0 (d 0 ) for d 0. The verties represent the features in Figure 4, and the boxes and edges represent some of the relationships among them. For brevity, the other design attributes have been omitted. represents a number of design attributes that are relevant for manufaturing and an be omputed automatially from the designs. Below is a summary of the idea; for additional details we refer the reader to [Elin97]. A design an have multiple signatures, that represent the design at different levels of detail. The most detailed signature for a design is the design s basi signature. The basi signature represents all properties of the design that are relevant for the urrent appliation. Thus, if two designs have basi signatures that are isomorphi, then for the purposes of the urrent appliation, the two designs are idential. To reason about a design at a higher level of abstration, we may ompute other signatures for the design. These signatures will be simplifiations of the basi signature, produed by removing some of the less important details from the basi signature. The preise attributes one might want to represent in a design signature will depend on the partiular manufaturing appliation but for eah manufaturing appliation, the attributes should be things that are speified expliitly in the CAD model or an be dedued automatially from it. For mahined parts (our urrent fous), the attributes that we use inlude volumetri mahined features and various relationships among them, tolerane information, and many of the usual kinds of properties measured in GT oding shemes, suh as material, quantity, and so forth. As an example, onsider the mahined part d 0 shown in Figure 3. Using the F-Rex feature extrator developed by Regli et al [Regl95, Regl97], we get the volumetri mahining features shown in Figure 4. By augmenting this set of features to inlude the design attributes mentioned above, we get d 0 s basi signature s 0 (d 0 ), whih is shown in Figure 5. Design Similarity. Figure 6 shows a design d 1 that is different from d 0 but similar to it. The similarity between the two designs is refleted in similarities between their basi signatures s 0 (d 0 ) and s 0 (d 1 ), whih are shown in Figure 5 and Figure 7, respetively. If we simplify the basi signatures by removing some of the less important details, this will produe simplified design signatures for d 0 and d 1 that are isomorphi, as shown in Figure 8 and Figure 9. The more similar two designs are, the fewer the simplifiations needed in order to produe simplified design signatures that are isomorphi so we an judge how similar two designs by looking at how muh simplifiation is needed. We formalize this idea as follows. Figure 6: A mahined part d 1 similar to d 0. 5 Copyright 1997 by ASME

6 M 6 : z D 2 : z D 4 : z M 2 : z M 4 : x, -x M 5 : x, -x D 2 : -z D 3 : -z M 0 : x, -x M 1 : x, -x s, s, D 4 : -z D 5 : -z M 7 : z D 3 : z D 5 : z D 0 : x D 1 : x M 8 : z M 9 : z D 0 : -x M 3 : x, z D 1 :-x Figure 7: The basi design signature s 0 (d 1 ) for d 1. Just as in Figure 5, the verties represent the features found by the F-Rex feature extrator, the boxes and edges represent some of the relationships among these features, and other design attributes have been omitted for brevity. M 16 : z M 14 : x, -x M 15 : x, -x M 5 : z s, M 12 : z, -z D 8 : x M 13 : z, -z s, D 9 : x M 8 : z D 8 : -x D 9 : -x D 1 : z M 17 : x, z D 1 : -z D 2 : -z D 2 : z Figure 8: A simplified design signature for d 0. M 6 : z M 4 : x, -x M 2 : z M 5 : x, -x M 7 : z D 4 : z D 4 : -z M 0 : x, -x M 1 : x, -x s, s, D 5 : -z D 5 : z D 0 : x D 1 : x D 0 : -x M 3 : x, z D 1 :-x Figure 9: A simplified design signature for d 1. 6 Copyright 1997 by ASME

7 Let D be a set of designs. Let P be the set of all design properties that are relevant for the urrent appliation, and suppose we have partitioned P into subsets P 1, P 2,, P n ; where P 1 is the set of design properties that we onsider least important, P 2 is the set of design properties that we onsider seond-least important, and so forth. Let d be a design in D, and let s 0 (d) be its basi signature, whih represents the values that d has for all of the properties in P. Then we an produe a sequene of progressively simpler design signatures s 1 (d), s 2 (d),, s k (d), by taking s 1 (d) = the parts of s 0 (d) that represent properties in P 2,,P n ; s 2 (d) = the parts of s 1 (d) that represent properties in P 3,,P n ; ; s n 1 (d) = the parts of s n 2 (d) that represent properties in P n ; s n (d) =. Let d be another design in D, and let s 0 (d ), s 1 (d ),, s n (d ) be its design signatures, omputed in the same manner as the orresponding signatures for d. Then for k = 0, 1,, n, we define the relation R k (d,d ) to hold if and only if s k (d) = s k (d ). Thus, R k (d,d ) is an equivalene relation that has following basi properties: R 0 (d,d ) holds only if d and d are idential for the purposes of the urrent appliation. Thus the equivalene lasses of R 0 are singleton sets; Sine s n (d) = s n (d ) =, R n (d,d ) holds for all designs d,d in D. Thus R n has a single equivalene lass onsisting of D itself. For k=1,,n, the equivalene lasses of R k 1 are subsets of the equivalene lasses of R k. From these properties, it follows that the equivalene lasses of R 0, R 1,, R n an be arranged hierarhially in a tree struture suh as the one shown in Figure 10. We all this tree a lassifiation tree. equivalene lasses of R 4 d 0 d 1 d 2 d 3 equivalene lasses of R 3 equivalene lasses of R 2 equivalene lasses of R 1 individual designs Figure 10. A simple example of a lassifiation tree. Eah vertex represents an equivalene lass. Finding similar designs. Given a new design d, the lassifiation tree provides an easy way to find existing designs similar to d. For eah vertex v of the tree, we say that v mathes d if d is a member of the equivalene lass represented by v. It immediately follows that if v mathes d, then at most one of v s hildren mathes d. Furthermore, the deepest vertex mathing d represents the set of designs that are most similar to d, and we an find this vertex using the following searh algorithm: proedure find-best-math(d) v = the root vertex of the lassifiation tree loop if d does not math any of v s hildren then exit, returning v. else v := whihever hild of v mathes d Figure 11 gives an example of the operation of this algorithm. verties that math d d 0 d 1 the designs most similar to d verties that don t math d d 2 d 3 Figure 11. Using the searh algorithm to searh the lassifiation tree of Figure 10. Computational Issues. To determine whether a vertex v of depth k in the lassifiation tree mathes a design d, the above algorithm ompares the design signatures s n k ( d) and s n k ( d ), where d is some design in the equivalene lass represented by v. This is a speial ase of the subgraph isomorphism problem, whih is well known to be omputationally diffiult. It is not known whether or not the subgraph isomorphism problem is NP-omplete, but no known algorithm for it runs in less than exponential time. In our ase, there are ways to speed up the omputation onsiderably. The details of how we do this appear in [Elin97], but here are two examples: We make use of speial properties of our problem that do not our in the general subgraph isomorphism problem. For example, by keeping trak of whether orresponding edges have the same edge labels, we an often very quikly detet ases in whih d and d do not math. The algorithm will not even attempt to hek whether d mathes a vertex v at depth k of the tree unless d mathes v s parent at depth k-1. Sine the design signatures used at depth k 1 are simpler than those used at depth k, heking whether d mathes v s parent an take exponentially less time than heking whether d mathes v. Validation. To validate our lassifiation system, we wanted to show that for a reasonable olletion of realisti objets, the tehniques orretly and effiiently (1) find idential objets and (2) return reasonably intuitive estimations of similarity between different objets. One major obstale is the lak of a generally available large and varied data set of CAD models for mehanial designs. A further ompliation is the lak of an agreed-upon standard for what it means to be similar from the manufaturing point of 7 Copyright 1997 by ASME

8 view -answers vary depending on the individual to whom the question is posed. Hene, large-sale validation was not possible and we hose to perform ontrolled experiments to determine how well the prototype system operates on reasonable examples. To perform our study, we used a olletion of solids taken from the NIST Design, Proess Planning, and Assembly Repository and employed the experimental F-Rex [Regl95, Regl97] feature reognizer previously developed at the University of Maryland at College Park. The prototype system used few feature parameters but was able to identify idential designs, and its similarity estimations orresponded losely to the design families. The experiments, one pre-proessing was performed, ran in moderate user-time, and we did not observe that the isomorphism heks reated any omputational bottleneks. Interestingly, we notied that the approah worked even on those parts where the F-Rex feature reognizer had diffiulty or produed spurious results. We believe that we an enhane the implementation by using the design features in the CAD model and integrating a ommerially tested features tool. Planning for New Designs If one wants to onstrut a proess plan for a new design using the traditional approah to variant proess planning, one begins by retrieving (from the variant database) a proess plan for an existing design that is similar to the new design. One drawbak to this approah is that although some portions of the existing design may losely math the orresponding portions of the new design, other portions may not math so well. Our goal is to find lose mathes to as muh of the new design as possible, using the following approah: deompose the design signature for the new design into slies that are meaningful from the point of view of proess planning; retrieve proess plan slies that orrespond to similar design slies; ombine and modify the plan slies to produe a proess plan for the new design. Below, we desribe the tehniques that we are urrently developing for these tasks. h 2 p 1 Figure 12. In this design, the hole h 1 and the poket p 1 are dependent on eah other for manufaturing purposes, but both are independent of the hole h 2. Sliing Designs. Given a design, some features of a design may be independent of eah other for manufaturing purposes, and some features may depend on other features. For example, in the simple design shown in Figure 12, the hole h 1 h 1 and the poket p 1 depend on eah other, beause the proess details for eah will depend on whih of them is mahined first. However, no suh dependeny exists between these two features and the hole h 2. The idea of a design slie is that it is a olletion of d s features that depend on eah other for manufaturing purposes. Manufaturing dependenies an arise in a number of ways; below are a few examples: if two features interset (like h 1 and p 1 in Figure 12), then they will usually depend on eah other; this dependene may form a preedene onstraint if one feature establishes or denies the approahability, aessibility, or emergene onditions of another feature. One feature will depend on another if the first feature has a tolerane attribute that is defined relative to a datum that is reated by the seond feature. One feature will depend upon another if there is a thin setion between them. One feature will depend upon another if the same tool is used to produe both. One feature will depend upon another if both have the same approah diretion. In pratie, it is not appropriate to try to apture all manufaturing dependenies in a design slie, beause we want our haraterization of a design slie to depend only on the design and the design signature, and some dependenies would be quite diffiult to detet without knowing the proess plan as well. Thus, for pratial purposes, we will onsider a feature f to depend on another feature g only under the following onditions: f and g interset; a fae reated by f has a tolerane that is defined relative to a datum that is reated by g; f and g have the same orner radii; f and g have the same approah diretion. Given this set of onditions, there are several possible things that we might mean by a design slie. The first possible meaning is that given a design d and a feature f of d, the slie of d by f is the set of all features in d that f depends on. This an be omputed using the algorithm slie1 shown below; and the set of all suh slies of d is thuss = {slie1(d,f) : f is a feature of d}. proedure slie1(d,f) s := t := {f} while t is not a subset of s do s := s t t := {all features of d that feature g depends on} return s U g t Not every set of features found by slie1 will atually be of interest to us. For example, if one feature f depends on another feature g, then we will not are about the set slie1(d,g), 8 Copyright 1997 by ASME

9 beause it will be a subset of the set slie1(d,f). This leads to the seond possible meaning of a design slie: that it is any set of features found by slie1 that is not a subset of some other set of features found by slie1. The set of all suh slies of d an be found using the algorithm slie2 below. proedure slie2(d) S := for eah feature f of d do s := slie(d,f) if s is not a subset of any of the sets in S then do remove from S every set that is a subset of s S := S {s} end return S For the design shown in Figure 12, slie2 will return a set S onsisting of two slies: {h 1,p 1 } and {h 2 }. In this ase the slies are disjoint, but in other ases they may overlap. This leads to a third possible meaning of a design slie: that the slies of d onsist of the sets found by slie2, but modified by taking the unions of all overlapping sets. These sets an be omputed using the algorithm slie3 shown below. proedure slie3(d) S := slie2(d) let the members of S be alled s 1,s 2,,s k for i = 1 to k do if s i is still in S then for j = i+1 to k do if s j is still in S and s i s j then remove s j from S and set s i := s i s j return S We have not yet determined whether it is preferable to use the sets omputed byslie2 or slie3. Deiding this will be an important topi for future work, as it has onsequenes for the details of the plan retrieval and ombining tehniques desribed in the following setions. Retrieving Plans and Extrating Plan Slies. One we have omputed the design slies for a new design d as desribed above, the next step for eah slie s of d is to math s to a orresponding slie s of some design d in the database. We an do that by using the tehniques desribed earlier in the Preproessing setion to reate a lassifiation tree of design slies and searh for a design slie similar to s. One we have found s and d, we need to take the proess plan p for d and extrat the plan slie q of p that orresponds to the design slie s. This plan slie onsists of the set of all operations in p that reate the features in d, as well as any other operations in p that are needed to establish the neessary preonditions. The onditions under whih we onsider an operation m to establish a preondition of another operation o are similar to the onditions for feature dependeny above: m is the setup needed for operation o. m reates a feature f that must preede the feature g that operation o reates. We an use this set of onditions to extrat the plan slie q using the following algorithm: proedure slie-plan(p,s ) q := r := {the operations of p that reate the features in s } while r is not a subset of q do q := q r for eah operation m of r do r := {all operations of p on whih m depends} return q U m r' Combining Plan Fragments. Suppose that for a design d, we have used the tehniques of the previous setions to ompute design slies s 1, s 2,, s k ; and suppose that using these slies we have retrieved plan slies q 1, q 2,, q k. Eah plan slie q i onsists of a linear sequene of operations, and we wish to interleave these sequenes to form a single plan q; and hange the parameters of the operations in q so that they will fit the requirements of the design d. In doing this, we need to be areful about the way in whih we interleave the operations, for not every possible interleaving will work orretly. As an example, onsider the design shown in Figure 12. Regardless of whih design-sliing algorithm we use, we will get the slies s 1 = {h 1,p 1 } and s 2 = {h 2 }. Suppose that from a database of plans for mahining prismati parts on a vertial mahining enter, we retrieve the following plan slies (for brevity, many details have been omitted from these plans): plan slie q 1 : Step 1. setup the part so that h 1 and p 1 are vertial Step 2. end-millp 1 Step 3. drill h 1 plan slie q 2 : Step 1. setup the part so that h 2 is vertial Step 2. drill h 2 In the plan slie q 2, setting up the part so that h 2 is vertial is one of the preonditions needed in order to drillh 2 on a vertial mahining enter. If we interleave the two plan slies in suh a way that the first step of q 1 omes between the first and seond steps of q 2, then the first operation of q 1 will deny this preondition, and thus it will not be possible to drill h 2. If the design slies are independent and eah plan slie fully reates the orresponding design slie, then one simple planmerging approah is to do eah plan slie in turn. That is, the new proess plan would equal the sequene q 1, q 2,, q k. However, it may be possible to reate a better plan by reduing the number of setups and tool hanges, as desribed below. Note that in the mahining domain, eah plan slie onsists of one or more subslies (whih must be done in the speified order). Eah subslie begins with a setup followed by one or more mahining operations performed in that setup. The next subslie begins with a setup in a different diretion. In the new plan formed by merging the plan slies, it is possible for a subslie to preede one or more subslies with the same setup diretion. In this ase we an often merge the subslies into a single subslie, by removing the following subslies setups (sine the first subslie s setup would be suffiient). Having 9 Copyright 1997 by ASME

10 done this, we may now also be able to remove tool hanges if onseutive mahining operations within the subslie use the same utting tool. More generally, we are interested in hoosing an order for interleaving the plan slies that is onsistent with the relevant manufaturing onstraints and maximizes the amount of subslie merging that an be done. Finding an optimal solution to this problem is NP-hard but we been able to solve a losely related problem quite effiiently using using branh-and-bound tehniques [Yang92], and we plan to adapt that approah to the urrent problem. SUMMARY AND CONCLUSIONS Current Status At this point we have made signifiant progress towards implementing our hybrid approah. We have reated routines that allow the user to define the design and the relevant feature relationships. The feature extrator is omplete, and we are developing the program to reate the design graph of these features. We need to onstrut routines that divide a design graph into design slies and searh for similar design slies. We will devise a design lassifiation sheme that reflets the need to find useful proess plans. We have a routine for reating the proess plan slies. Future work inludes defining algorithms and developing programs that ombine the plan slies and modify them to form a new proess plan. Antiipated Benefits Our approah, when ompleted, will have the following primary benefits: Aelerating the produt development proess. Like traditional variant proess planning, our hybrid approah will onstrut proess plans that the proess engineer may need to improve. However, by automatially adapting the retrieved plans to the new design requirements, our hybrid approah will minimize the need for suh improvements. Our approah will utilize, in an innovative way, the strengths of both variant and generative proess planning. This approah also inludes sophistiated feature reognition and plan-based design evaluation in an integrated methodology. Our approah will provide the designer feedbak about the ahievable produt quality, the ost, and time needed to manufature the produt. By identifying those design elements that are espeially diffiult to produe in a osteffetive manner, our approah will help the designer develop produts that are easy to manufature. This will redue the need for redesign during the prodution run, resulting in redued lead time and produt ost. In summary, adaptive proess planning and plan-based design evaluation will support agile manufaturing by supporting a quik response to ever-hanging market opportunities. By using our hybrid variant-generative approah, a firm will be able to develop a new produt quikly and manufature a small prodution run eonomially. ACKNOWLEDGMENTS This work is supported in part by ONR grant DABT C-0037, NSF grants NSF EEC , IRI , and DDM , ARPA grant DABT63-95-C-0037, and in-kind ontributions from Spatial Tehnologies and Bentley Systems. Any opinions, findings, and onlusions or reommendations expressed in this material are those of the authors and do not neessarily reflet the views of the funders. REFERENCES [Bond88] A. Bond and R. Jain, The Formal Definition and Automati Extration of Group Tehnology Codes, Proeedings of the ASME Computers in Engineering Conferene, pp , [Boot94] G. Boothroyd. Produt design for manufature and assembly, Computer-Aided Design 26:7, 1994, pp [Bour95] D. Bourne and C. Wang. Design and manufaturing of sheet metal parts: using features to resolve manufaturability problems, in A. Busnaina, Ed., Computers in Engineering 1995, ASME, 1995, pp [Cand95] A. Candadai, J. W. Herrmann, and I. Minis. A Group Tehnology-based variant approah for agile manufaturing, in Conurrent Produt and Proess Engineering, edited by A.R. Thangaraj, R. Gadh, and S. Billatos (ASME, New York, 1995), [Cand96] A. Candadai, J.W. Herrmann, and I. Minis, Appliations of group tehnology in distributed manufaturing, Journal of Intelligent Manufaturing, Volume 7, pages , [Chan90] Chang, T. Expert Proess Planning for Manufaturing, Addison-Wesley, Reading, MA, [Das95] D. Das, S. Gupta, and D. Nau. Generating redesign suggestions to redue setup ost: a step towards automated redesign. Computer Aided Design, to appear. [Elin97] A. Elinson, D. S. Nau, and W. C. Regli. Featurebased similarity assessment of solid models. Pro. ACM Solid Modeling Conferene, May [Geel95] R. Geelink, O. Salomons, F. van Slooten, F. van Houten, and H. Kals. Unified feature definition for feature based design and feature based manufaturing, in A. Busnaina, Ed., Computers in Engineering 1995, ASME, 1995, pp [Gupt94a] S. K. Gupta, D. S. Nau, W. C. Regli, and G. Zhang. A methodology for systemati generation and evaluation of alternative operation plans. In Jami Shah, Martti Mantyla, and Dana Nau, editors, Advanes in Feature Based Manufaturing, pages Elsevier/North Holland, [Gupt95] S. K. Gupta and D. S. Nau. A systemati approah for analyzing the manufaturability of mahined 10 Copyright 1997 by ASME

11 parts. Computer Aided Design, 27(5): , [Ham86] I. Ham., D. Marion., and J. Rubinovih, Developing a Group Tehnology Coding and Classifiation Sheme, Industrial Engineering, Vol. 18, No. 7, pp , [Ham88] I. Ham and S. Lu (1988), Computer Aided Proess Planning, the Present and Future, Annals of CIRP, 37(2). [Hebb96] K. Hebbar, S. J. Smith, I. Minis, and D. Nau. Plan-based evaluation of designs for mirowave modules. Pro. ASME Design Tehnial Conferene, August [Hend88] M. Henderson and S. Musti, Automated Group Tehnology Part Coding from a Three-Dimensional CAD database, Journal of Engineering for Industry, Vol. 110, No. 3, pp , [Hout75] A. Houtzeel. MICLASS, a lassifiation system based on group tehnology. Tehnial Report Working Paper MS #75-721, Soiety of Manufaturing Engineers, [Iyer95] S. Iyer and R. Nagi, Identifiation of Similar Parts in Agile Manufaturing," in Conurrent Produt Design, edited by R. Gadh (ASME, New York, 1994), [Kamb93] S. Kambhampati, M. Cutkosky, J. Tenenbaum, and S. Lee. Integrating general purpose planners and speialized reasoners: ase study of a hybrid planning arhiteture. IEEE Transations on Systems, Man, and Cybernetis 23:6,1993. [Lam95] G. Lam, High Level Proess Planning and Manufaturability Analysis in Agile Manufaturing, M.S. Thesis, Department of Mehanial Engineering, University of Maryland, College Park, Maryland, [Lee87] Y. C. Lee and K. S. Fu. Mahine understanding of CSG: extration and unifiation of manufaturing features. IEEE Computer Graphis & Appliations, Vol. 7, No. 1, pp , [Mant89] M. Mäntylä, J. Opas, and J. Puhakka. Generative proess planning of prismati parts by feature relaxation, in B. Ravani, Ed., Pro. 15th ASME Design Automation Conf., ASME, 1989 pp [Mare94] M. Marefat and J. Britanik, Case-based proess planning with hierarhial plan merging, IEEE, [Mitr66] S. P. Mitrofanov, Sientifi Priniples of Group Tehnology, English Translation, National Library for Siene and Tehnology, Washington, D.C., [Orga86] Organization for Industrial Researh, OIR Multi-M Code Book and Conventions, Waltham, Massahusetts, [Park93] S.C. Park, M.T. Gervasio, M.J. Shaw, and G.F. DeJong, Explanation-based learning for intelligent proess planning, IEEE Transations on Systems, Man, and Cybernetis, Volume 23, Number 6, pages , [Regl95] W. C. Regli, Satyandra K. Gupta, and D. S. Nau. Extrating alternative mahining features: An algorithmi approah. Researh in Engineering Design 7:3,1995, pp [Regl97] W. C. Regli, S. K. Gupta, and D. S. Nau. Toward multiproessor feature reognition. Computer Aided Design, 1997, to appear. [Srik94] A. Srikantappa and R. Crawford. Automati part oding based on interfeature relationships, in J. Shah, M. Mäntylä, and D. Nau, Eds., Advanes in Feature Based Manufaturing, Elsevier, Amsterdam, 1994, pp [Shah89] J. J. Shah and A. Bhatnagar, Group Tehnology Classifiation from Feature-Based Geometri Models, Manufaturing Review, Vol. 2, No. 3, pp , [Sun95] T.-L. Sun, C.-J. Su, R.J. Mayer, and R.A. Wysk. Shape similarity assessment of mehanial parts based on solid models. ASME Design Engineering Tehnial Conferenes 83:2, 1995, pp [Wang91] H.-P. Wang and J.-K. Li, Computer-Aided Proess Planning. Advanes in Industrial Engineering 13, Elsevier Siene Publishers, [Wozn94] M., Wozny, M., Pratt, and C. Poli, Topis in feature-based design and manufaturing, in J. Shah, M. Mäntylä, and D. Nau, Eds., Advanes in Feature Based Manufaturing, Elsevier, Amsterdam, 1994, pp [Yang92] Q. Yang, D. S. Nau, and J. Hendler, Merging Separately Generated Plans with Restrited Interations, Computational Intelligene 8:2, 1992, pp [Yue94] Y., Yue and J. Murray, Workpiee seletion and lamping of omplex 2.5D omponents, in J. Shah, M. Mäntylä, and D. Nau, Eds., Advanes in Feature Based Manufaturing, Elsevier, Amsterdam, 1994, pp Copyright 1997 by ASME