Feature Recognition¶
Home / Engineering / System Design / CAM Pipeline / Feature Recognition
Status: Skeleton
This document is a starting point. Sections marked with TBD need to be filled in.
Purpose¶
Identify machinable features from part geometry and produce a structured feature map that AutoCAM uses to plan manufacturing operations. This is the bridge between raw geometry and manufacturing intent.
Responsibilities¶
- Analyze standardized B-Rep input and identify machinable features: holes, pockets, faces, fillets, threads, slots, chamfers, etc.
- Classify features with attributes relevant to CAM planning (depth, diameter, tolerance requirements, surface finish)
- Output a structured feature map tied to the part geometry
Model Candidates¶
The feature recognition system will run one or more ML models trained on annotated machining data:
| Model | Notes |
|---|---|
| AAGNet | Attribute Adjacency Graph Network for machining feature recognition |
| BRepMFR | B-Rep based manufacturing feature recognition |
| BRepFormer | Transformer-based approach for B-Rep feature understanding |
TBD: Model selection, training data requirements, and accuracy targets.
Interfaces¶
| Direction | System | Data |
|---|---|---|
| Input | Data Ingestion | Standardized B-Rep model with PMI annotations |
| Output | AutoCAM Pipeline | Structured feature map (feature types, locations, attributes, tolerances) |
Training Data¶
- Primary source: distributed annotation workflow using offshore machinists and CAM experts (Mechanical Turk-style) who label features on thousands of CAD files
- Secondary source: real customer parts providing authentic edge cases
- Annotation tool (support tool) is used to create and manage training data
Open Questions¶
- Which model architecture performs best on our target geometry class (prismatic aluminum parts)?
- What is the minimum training dataset size to reach useful accuracy?
- How do we handle features that span multiple setups?