Machining feature recognition based on deep neural networks to support tight integration with 3D CAD systems

Area of Science:

• Computer-Aided Design (CAD)
• Deep Learning
• Geometric Feature Recognition

Background:

• Current deep learning applications for 3D CAD models often convert boundary representation (B-rep) to voxel, mesh, or point cloud formats.
• This conversion process leads to reduced model resolution, feature loss, and difficulties in mapping converted data back to original B-rep faces.
• Existing methods struggle with preserving critical geometric information essential for accurate machining feature recognition.

Purpose of the Study:

• To propose a new method for recognizing machining features directly from 3D CAD B-rep models.
• To enable tight integration between 3D CAD systems and deep neural networks.
• To overcome limitations associated with traditional 3D model format conversions for deep learning applications.

Main Methods:

• Developed a method utilizing feature descriptors as direct inputs for deep neural networks.
• Feature descriptors explicitly represent key properties of a face within the B-rep model.
• Trained and evaluated a deep neural network using a dataset of 2236 3D CAD models (1430 training, 358 validation, 448 testing).

Main Results:

• The proposed method successfully recognized 17 types of machining features (16 types plus a non-feature) from B-rep models.
• Achieved accurate recognition for all 75 test cases.
• Demonstrated the effectiveness of feature descriptors in preserving crucial geometric information for deep learning.

Conclusions:

• The integration of feature descriptors with deep neural networks offers a robust solution for machining feature recognition in 3D CAD.
• This approach effectively addresses the resolution and feature loss problems inherent in traditional 3D model conversion methods.
• The method shows high potential for enhancing automated design and manufacturing processes.