Graph Neural Network Feature Extraction from 3D CAD Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing feature recognition systems in Computer Aided Manufacturing (CAM) struggle to accurately identify manufacturing features, especially with minor variations, and require expert knowledge for rule-based approaches, which are time-consuming and difficult to scale, often losing the Boundary Representation (B-Rep) relationship due to the use of voxel data structures.
Innovation Solution
A method and system that generate a graph from a 3D model using adjacency attribute matrices, assign scores to edges based on predefined criteria, extract sub-graphs, and utilize a Graph Neural Network (GNN) to determine manufacturing features, eliminating the need for rule-based systems and enabling flexible feature classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If rule-based approach is used for feature recognition, then recognition accuracy can be maintained for standard features, but the system becomes time-consuming and requires expert knowledge for rule development
Solution Approach 1:
The patent replaces the mechanical rule-based system with a neural network-based automated system. The GNN model learns feature recognition patterns automatically from training data, eliminating the need for manual rule development by experts while maintaining recognition accuracy for both standard and variant features.
Solution Approach 2:
The system enables self-service by allowing the neural network to automatically learn and adapt to different feature types and variations through training on diverse datasets. The model continuously improves its recognition capabilities without requiring external expert intervention for rule updates.
2Measurement precision
If template matching is used for feature recognition, then standard features can be identified, but the system becomes computationally expensive and incapable of handling minor variations
Solution Approach 1:
The patent employs a dynamic neural network model that can adapt to different feature variations rather than using static template matching. The GNN learns to generalize from training examples, enabling it to handle minor variations efficiently without requiring separate templates for each variation, thus reducing computational complexity.
Solution Approach 2:
The system changes the approach from fixed template parameters to learned parameters through neural network training. The model learns optimal feature representation parameters from data, allowing it to recognize features with minor variations without the computational burden of matching against multiple rigid templates.
3Ease of manufacture
If voxel data structure is used for feature recognition, then 3D model processing can be simplified, but the system loses B-rep relationship and becomes difficult to scale
Solution Approach 1:
The patent uses graph structures as an intermediary representation that preserves the topological relationships of B-rep models while being suitable for neural network processing. The graph structure maintains the connectivity and spatial relationships of the 3D model, enabling both simplified processing and preservation of geometric reasoning capabilities.
4Reliability
If existing feature recognition systems are used, then current features can be identified, but the system fails to recognize features with minor variations and is difficult to scale to new features
Solution Approach 1:
The patent performs preliminary action by training the neural network on a comprehensive dataset that includes various feature types and their variations before deployment. This pre-training enables the system to recognize both standard and variant features reliably without requiring reprogramming when new variations are encountered.
Solution Approach 2:
The neural network model serves as a universal feature recognition system that can identify multiple types of manufacturing features and their variations through a single unified model. The GNN architecture enables the system to generalize across different feature types, making it versatile and scalable to new features without requiring separate recognition systems.
Data Source
AI summary
The invention relates to method and system for extracting and classifying manufacturing features from a three-dimensional (3D) model of a product. The method includes generating graph corresponding to product based on 3D model of product. The graph includes nodes corresponding to faces of the product and links corresponding to edges of product. The graph generation includes determining adjacency attribute matrix from the 3D model. The method further includes assigning scores to each of links; determining a cumulative score for each of links; extracting sub-graphs from graph by discarding one or more links from links; extracting node parameters and edge parameters from 3D model of product; determining node feature vector based on node parameters and edge feature vector based on edge parameters; and determining a type of manufacturing feature based on corresponding node feature vector and edge feature vector using a Graph Neural Network (GNN) model.