BRepMAE: Self-Supervised Masked BRep Autoencoders for Machining Feature Recognition

TL;DR

A fine-tuned network that contains both the encoder and a task-specific classification network for machining feature recognition (MFR) achieves a significant improvement in recognition rate with the same amount of training data, especially when the number of training samples is limited.

Abstract

We propose a masked self-supervised learning framework, called BRepMAE, for automatically extracting a valuable representation of the input computer-aided design (CAD) model to recognize its machining features. Representation learning is conducted on a large-scale, unlabeled CAD model dataset using the geometric Attributed Adjacency Graph (gAAG) representation, derived from the boundary representation (BRep). The self-supervised network is a masked graph autoencoder (MAE) that focuses on reconstructing geometries and attributes of BRep facets, rather than graph structures. After pre-training, we fine-tune a network that contains both the encoder and a task-specific classification network for machining feature recognition (MFR). In the experiments, our fine-tuned network achieves high recognition rates with only a small amount of data (e.g., 0.1% of the training data), significantly enhancing its practicality in real-world (or private) scenarios where only limited data is available. Compared with other MFR methods, our fine-tuned network achieves a significant improvement in recognition rate with the same amount of training data, especially when the number of training samples is limited.

Figures (13)

• Figure 1: Given a BRep, we first convert it into a gAAG, where graph nodes represent BRep facets and graph edges represent BRep edges. Then, we embed the graph node and edge features into 256-dimensional spaces and pre-train a masked autoencoder to learn the BRep representation. Finally, the fine-tuning stage retains the encoder of our autoencoder and replaces the decoder with a classification head to train this fine-tuned model on various amounts of labeled data for MFR.
• Figure 2: For each face of a BRep model, we perform UV parameterization to discretize it into regular $10\times 10$ grid points. We extract the coordinate, surface normal, and trimming state of each grid point to form the face geometric information. Similarly, we uniformly sample 10 points in the curve parameter domain for each edge and extract the coordinates, tangent vectors, and normals of each point from both adjacent faces as the edge's geometric information.
• Figure 3: The workflow of our pre-training stage. After we convert the input BRep model into a gAAG, our model embeds the geometric information of faces and edges into a 256-dimensional space. Then we randomly mask some nodes of the gAAG and reconstruct their information using an autoencoder. A multi-branch decoder is adopted to recover the geometric information of BRep models. The reconstructed 256-dimensional features and geometric information are used in the loss to pre-train our model.
• Figure 4: Our BRep embedding. Various networks (such as 2D CNNs, MLPs, and 1D CNNs) are used to embed face and edge information. The face feature vectors are concatenated to form a 256-dimensional vector, which serves as the graph node feature. Similarly, the edge feature vectors are concatenated and projected to produce a 256-dimensional vector that serves as the graph edge feature.
• Figure 5: Given a graph with some nodes' features masked, our model first applies a GNN encoder (MPNNs) to extract the graph's latent representation, and then a GNN decoder (MPNNs) to reconstruct the masked node features. Later, a multi-branch decoder is implemented to decode the geometric and attribute information of the BRep face from the reconstructed node feature.