NEURAL-NETWORK APPROACH TO THE RECONSTRUCTION OF FREEFORM SURFACES FOR REVERSE ENGINEERING

Reconstruction and manufacturing of existing freeform surfaces are of paramount importance for reverse engineering. These are particularly useful for the situations where the surfaces are partially damaged or the surface models are not available. In order to generate models based on the existing freeform surfaces, the surfaces are first digitized by either a laser scanner or a coordinate measuring machine. The digitized points are then used to construct the models. The paper presents a neural network approach to reconstruction of computer models for existing freeform surfaces, and manufacturing of surfaces. To evaluate the effectiveness of the approach, a mathematically known surface, a nonuniform B-spline surface, was used for generating a number of samples for training the networks. Three 4-layered neural networks were designed and trained using a modified back propagation algorithm. The trained networks then generated a number of new points which were compared with the calculated points using the known surface equations. These points were also used to generate toolpaths for machining the surface. The results show that the approach can be used to reconstruct computer representations of the existing surfaces, and to manufacture these surfaces. An example is included to illustrate the approach.