In order to perform manufacturability analysis, a product design must be interpreted in terms of manufacturing features. Automated feature recognition has become the preferred technique for producing such feature-based representations, having been successfully employed for a variety of applications including process planning and part code generation for group technology. These feature technologies are rely heavily on the geometric and topological manipulation capabilities of solid modeling systems and deal predominantly with form or machining features.
Kyprianou [110] presented the first effort to use a
grammatical approach to parse solid models of parts for group coding.
Kramer [111] has presented a grammar-based method for
extracting non-intersecting features for a class of
2
-dimensional parts. Methods based on graph-grammars
have been used to both recognize features [112,113] and
translate between differing feature
representations [114]. Peters [115]
analyzes the combinatorial complexity of graph and grammatical
approaches to feature recognition and presents heuristics to reduce
these costs. In another effort to to address combinatorial problems
and handle realistic industrial designs, Gadh and
Prinz [116] describe techniques for abstracting an
approximation of the geometric and topological information in a solid
model and finding features in the approximation. More recently, Regli
et al [117] have outlined methods to
utilize multiple distributed processors. Their initial results show
that multi-processor techniques can be effectively employed to
significantly expand the class of mechanical designs that are feasible
and produce large improvements in system response times.
Woo [118], in an early effort on feature extraction, proposed a method for finding general depression and protrusion features on a part through decomposing the convex hull of the solid model. The approach had several limitations, including the existence of pathological geometric cases in which the procedure would not converge. The non-convergence of Woo's approach has been solved in recent work by Kim [119,120,121], whose system produces a decomposition of the convex hull of a part as general form features. Extension of this method from polyhedra to the more general surfaces required for realistic parts is currently under investigation [122].
Other volume decomposition approaches include the recent work by Sakurai [123]. Exhaustively, each combination of cells is then matched against user-defined feature templates. While the method is capable of generating all alternative feature interpretations composed of the primitive cells, it does so at a large combinatorial cost.
The seminal work of Henderson [124] employed rule-based systems on the feature recognition problem and has served as a foundation for more recent AI-based approaches. Henderson has also made extensive use of graph-based methodologies, first in [125] where graph-based algorithms are used to find protrusion and depression features. In Chuang and Henderson [126] use graph-based pattern matching to find feature patterns from part geometry and topology. Chuang and Henderson [127] were the first to explicitly address both computational complexity and decidability when defining the feature recognition problem. Their paper formalized the problem of recognition of features (including compound features) through parsing a graph-based representation of a part using a web grammar. Most recently, Gavankar and Henderson [128] adapted neural networks to recognize features from polyhedral objects. Also in this area, Peters [129] describes techniques for training neural networks to recognize feature classes that can be customized by the end user. In a recent paper, Henderson et al. [130] surveys a variety of feature recognition methodologies.
Other graph-based methodologies include the work of De
Floriani [131], who employed graph-based algorithms for
finding bi-connected and tri-connected components to partition a
polyhedral part into several varieties of protrusion and depression
features. Joshi's [132] approach used subgraph isomorphism
algorithms to match feature patterns to patterns in the topology of
polyhedral parts. Sakurai [133] developed a
graph-based system capable of handling limited types of user-defined
features, providing for a degree of application-specific
customizability. Corney and
Clark [134,135] have had success
extending the capabilities of graph-based algorithms to more general
2
-dimensional parts.
The work of Dong and Wozny [136,137,138] included formalization of a feature description language and was the first to employ a frame-based reasoning system to extract machining features for computer-aided process planning. Their approach included the ability to construct volumetric features from surface features and perform an analysis of tool accessibility.
Karinthi and Nau [139] presented the first systematic work on the generation of alternative interpretations of the same object as different collections of volumetric features. They present an algebra for computing alternate interpretations of parts resulting from algebraic operations on the features.
The ability to recognize interacting features has been a goal of a number of numerous research efforts, among them [116,132,136]. The approach of Marefat [140,141] built on the representation scheme of Joshi [132] and used a combination of expert system and hypothesis testing techniques to extract surface features from polyhedral objects and handle a variety of their geometric interactions. Marefat argues that his approach is complete over a class of polyhedral features, i.e., that it generates all features in his class that can be found from the geometry a part. Another recent approach [142] addresses completeness over a limited domain of iso-oriented polygonal parts. Regli et al [143,144] present a methodology for specifying the feature recognition problem and proving it is complete over a well-defined class of parts. Their features are based on a class of machining features that describe operations on three-axis machining centers and encompass a realistic class of parts bounded by analytic surfaces.
The most comprehensive approach to date for recognizing features and handling their interactions has been the OOFF system (Object-Oriented Feature Finder) of Vandenbrande [145]. Vandenbrande's work, using a knowledge-based approach like that of Dong and Wozny, provides a framework for recognizing machining features and building process plans via artificial intelligence techniques in combination with queries to a solid modeler.
Work of Laakko and Mäntylä [146] couples feature-based design and feature recognition to provide for incremental feature recognition. This type of approach identifies changes in the geometric model as new or modified features while preserving the existing feature information. They also provide for some form of customizability with use of a feature-definition language to add new features into the system.
Other related work includes feature recognition from 2D engineering drawings [147], feature recognition for sheet-metal components [148], and feature modeling by incremental recognition [146]. Many aspects of the feature recognition problem are still open and active areas of research. Among these are: recognizing and representing interacting features [145], incremental recognition of features [149], modeling alternative feature interpretations [140,143], reasoning about the manufacturability of features [25,150], and incorporation of user-customizable feature classes.