The machining features in a three dimensional (3D) computer-aided design
(CAD) model are shapes with engineering meaning created through
machining processes such as cutting or turning. With recent advancements
of process automation technology in the m...
The machining features in a three dimensional (3D) computer-aided design
(CAD) model are shapes with engineering meaning created through
machining processes such as cutting or turning. With recent advancements
of process automation technology in the manufacturing industry, researches
on machining feature recognition from a 3D CAD model are being
conducted to automatically generate machining commands and tool paths
based on recognized machining features.
The previous researches have focused on face-level recognition, which
means classifying each face of a 3D CAD model as a machining feature
class. However, face-level recognition cannot recognize a set of faces that
constitute machining features. As machining information such as machining
command is mainly composed of machining feature instances, it is not
possible to extract machining information from the result of face-level
recognition.
To solve this problem, this study proposes a method for instance-level
machining feature recognition using graph neural networks (GNNs).
Instance-level machining feature recognition means identifying a set of faces
that belong to the same machining feature instance. This enables it to
recognize a set of faces that constitute a machining feature and extract
meaningful machining information.
To achieve this, encoders are defined to embed the surfaces and curves of
the 3D CAD model, extracting feature vectors of surfaces and curves.
These feature vectors are then used as input feature vectors for the GNNs.
In addition, a two stage GNN architecture is defined. The first stage GNN
uses the feature vectors of surfaces and curves to classify faces into
machining feature classes. The second stage GNN classifies faces on
instance-level using the face-level recognition results and the feature
vectors of surfaces and curves.
In order to train the proposed model, a dataset for it is required. However,
as publicly available CAD model datasets contain only face-level labeling
information, they cannot be used for this study. Therefore, in this study, the
automatic generation code of MFCAD++ dataset, a publicly available CAD
model dataset, is modified to automatically generate a CAD model dataset
containing instance-level labeling information. The generated dataset
contains 50,900 CAD models. The dataset is split into 30,540 training
datasets, 10,180 validation datasets, and 10,180 test datasets.
Finally, the proposed model is trained with the training dataset. And the
hyper-parameters are tuned based on the random search method with the
validation dataset. Subsequently, the performance of the model is evaluated
on the test dataset. As a results, the proposed method exhibits an accuracy
of 97.82% and pIoU of 95.93% for instance-level recognition.