25-06-2012, 12:51 PM
A Knowledge Based Reverse Engineering for Mechanical Components
A Knowledge Based Reverse Engineering.pdf (Size: 810.77 KB / Downloads: 115)
ABSTRACT
This paper focuses on Reverse Engineering (RE) in mechanical design. RE is an activity
which consists in creating a full CAD model from a 3D point cloud. The aim of RE is to
enable an activity of redesign in order to improve, repair or update a given mechanical
part. Nowadays, CAD models obtained using modern software applications are
generally “frozen” because they are sets of triangles of free form surfaces. In such
models, there are not functional parameters but only geometric parameters. This
paper proposes the KBRE (Knowledge Based Reverse Engineering) methodology which
allows managing and fitting manufacturing and/or functional features.
INTRODUCTION
In mechanical engineering, Reverse Engineering (RE) is an activity which consists in creating a full CAD
model of a given mechanical part from a 3D point cloud. The 3D point cloud is often provided by 3D
scanners or as results of FEM approaches. The reasons people have to make RE operations are, most of
the time: (1) the original design is not supported by enough documentation and no plan is available or
correct; (2) the original provider of the considered mechanical part has disappeared and does not
manufacture the component anymore. In this paper, the following context is considered: little
information about the studied part and a 3D point cloud are available because the studied part has
been scanned.
THE STATE OF THE ART, THE REVERSE ENGINEERING WORKS
A full, accurate and automatic segmentation of a 3D point cloud is still a central problem in RE.
As said in the introduction, finding the design intents of a given part is the key to build a real CAD
model. There are not many related works that suggest using original design intents.
The Segmentation
The segmentation of a meshed 3D point cloud is a research field which consists in the division of the
3D point cloud of a given object into a set of n point clouds representing the n features that compose
this object. In RE, three segmentation techniques are commonly used. In a first place, region based
technique uses spatial coherence of the data to organize the mesh into meaningful groups. Least
squares approximation by plans is the simplest method. Besl and Jain's works [2] allow the
classification of the mesh in three under regions: plans, convex and concave faces are recognized. The
best techniques are based on the approximation by bi-polynomial surfaces [5] and allow the
recognition about simple forms such as plan, cylinder, spherical and conical surfaces. To summarize,
an adjacent region is absorbed if it satisfies the estimate by the polynomial surface of a given minimal
order. If this adjacent region is not absorbed, the smoothing by a polynomial of superior degree is
tried. The process stops when all regions are absorbed or when estimations with all polynomial
degrees have failed. Current computers improve execution speed of polynomial degrees but this
technique is sensitive to the noise of the cloud. Normal calculations or curvatures are often noised and
edges are not clearly defined. In a second place, the technique used is the edge-based method that
consists in intending to isolate discontinuities in the 3D point cloud.