CN-122019822-A - Three-dimensional part retrieval method based on self-supervision graph characterization learning

CN122019822ACN 122019822 ACN122019822 ACN 122019822ACN-122019822-A

Abstract

The application provides a three-dimensional part retrieval method based on self-supervision graph characterization learning, relates to the field of deep learning, and solves the technical problem of high retrieval cost caused by the fact that a part retrieval method in the prior art depends on manual labeling of parts. The method comprises the steps of obtaining and analyzing a CAD three-dimensional part data set in a boundary representation format, constructing an edge-face connection diagram according to a topological relation between a curved surface and a curved line in a three-dimensional part to obtain an undirected diagram, inputting the undirected diagram into an encoder-decoder structure, carrying out node-level self-supervision training on node information in the edge-face connection diagram, keeping a trained encoder as a node feature extractor, carrying out diagram-level self-supervision training on the edge-face connection diagram by using the node feature extractor to obtain a pre-trained part retrieval model, extracting feature vectors of all parts in the three-dimensional part data set by using the part retrieval model, and storing the feature vectors as a part feature library for part retrieval.

Inventors

Ma Zhaomeng
LIU YONGZHI
LIAO ZHOUYI
SONG YANZHI

Assignees

合肥人工智能与大数据研究院有限公司

Dates

Publication Date: 20260512
Application Date: 20260129

Claims (10)

1. A three-dimensional part retrieval method based on self-supervision graph characterization learning is characterized by comprising the following steps: Acquiring and analyzing a CAD three-dimensional part data set in a boundary representation format, and constructing an edge-face connection diagram according to the topological relation between curved surfaces in the three-dimensional part to obtain an undirected diagram; inputting the undirected graph into a coder-decoder structure, performing node-level self-supervision training on node information in the boundary connection graph, and reserving the trained coder as a node characteristic extractor; Performing graph-level self-supervision training on the boundary connection graph by using a node feature extractor to obtain a pre-trained part retrieval model; And extracting feature vectors of all parts in the three-dimensional part data set by using the part retrieval model, and storing the feature vectors as a part feature library, wherein the part feature library is used for carrying out cosine similarity calculation on the feature vectors of the three-dimensional parts to be retrieved and returning information of the parts according to cosine similarity sequencing.
2. The three-dimensional part retrieval method based on self-supervision graph characterization learning according to claim 1, wherein the constructing an edge-face connection graph according to the topological relation between curved surfaces in the three-dimensional part comprises: Analyzing the CAD file in the boundary representation format by adopting an open source geometric kernel to obtain the geometric structure information of the three-dimensional part in the file; extracting topological association relation between the curved surface and the curve from the geometric structure information, defining the curved surface as a node of the undirected graph, and defining an intersecting curve of the intersecting curved surface as an edge of a corresponding node; And integrating all the nodes and the edges to obtain an edge-face connection diagram.
3. The three-dimensional part retrieval method based on self-supervised graph representation learning of claim 1, wherein the node-level self-supervised training of the node information in the edge-to-edge connection graph comprises: calculating original characteristics of nodes and edges in the edge-face connection graph; Randomly selecting a plurality of node marks as invalid in the edge-face connection diagram, and recording subscripts of the nodes marked as invalid to obtain a second edge-face connection diagram; inputting the second edge connection diagram into an encoder for encoding to obtain a first output characteristic; Marking the corresponding node in the first output characteristic as invalid according to the subscript of the node marked as invalid, so as to obtain a second output characteristic; Inputting the second output characteristic as a node characteristic of the second edge-face connection diagram into a decoder, and reconstructing the node characteristic marked as invalid in the second output characteristic by using the decoder; and carrying out normalization processing on the original characteristic and the output characteristic of the decoder, and iteratively training the encoder-decoder structure with the aim of minimizing the mean square error of the original characteristic and the output characteristic of the decoder until the loss converges.
4. A three-dimensional part retrieval method based on self-supervised graph representation learning as recited in claim 3, wherein the computing of the original features of nodes and edges in the edge-to-face connection graph comprises: Uniformly sampling the UV parameter space of the curved surface to form curved surface grid points, and representing the curved surface grid points as two-dimensional vectors to obtain the original characteristics of the nodes; And uniformly sampling the parameter domain of the curve to form curve grid points, and representing the curve grid points as one-dimensional vectors to obtain the original characteristics of the edges.
5. The method for three-dimensional part retrieval based on self-supervised graph representation learning of claim 1, wherein the encoder in the encoder-decoder architecture comprises a two-dimensional convolutional neural network, a one-dimensional convolutional neural network, and a graph-annotation force network, The two-dimensional convolutional neural network is used for processing the original characteristics of the nodes and outputting initial geometric embedded vectors of the nodes; the one-dimensional convolutional neural network is used for processing the original characteristics of the edges and outputting initial geometric embedded vectors of the edges; The graph attention network is used for fusing the initial geometric embedded vectors of the nodes and the edges and the topological structure of the graph, calculating the message transmission weight among the nodes through an attention mechanism, and outputting the node characteristics containing global context information.
6. The method for three-dimensional part retrieval based on self-supervised graph representation learning of claim 5, wherein the graph annotation force network employs a dynamic graph attention mechanism in which edge weights between nodes i and j The calculation formula of (2) is as follows: Wherein, the method comprises the steps of, 、 Node characteristics of the node i and the node j respectively, W is a parameter matrix capable of learning, d is a characteristic dimension, sigma is an activation function, lambda is a balance coefficient, The similarity of the curved surfaces of the node i and the plane where the node j is located is calculated based on Gaussian curvature.
7. The three-dimensional part retrieval method based on self-supervised graph representation learning of claim 1, wherein the graph-level self-supervised training of the boundary surface connection graph with the node feature extractor comprises: Inputting the edge-face connection graph into a trained node feature extractor, and outputting node features containing aggregate information and topology information; Carrying out random data transformation on the boundary connecting graphs, taking the transformed graphs as positive samples of the current boundary connecting graphs, and selecting other boundary connecting graphs in the same training batch as negative samples of the current boundary connecting graphs, wherein each training batch comprises N boundary connecting graphs and N corresponding positive samples; Inputting the node characteristics into a graph neural network, and training the graph neural network to learn a global trapping representation of the edge-face connection graph by taking a comparison learning loss function as an optimization target, wherein the comparison learning loss function is used for quantifying similarity differences between the current edge-face connection graph and positive and negative samples, and training the graph neural network by minimizing the similarity differences between the current edge-face connection graph and the positive samples; and stopping training after the contrast learning loss function converges to obtain a pre-trained part retrieval model.
8. The three-dimensional part retrieval method based on self-supervised graph representation learning of claim 7, wherein the graph-level self-supervised training loss function is: wherein L represents a loss value, s ij represents the similarity between the current edge face connection graph i and a positive sample j, s ik represents the cosine similarity between the current edge face connection graph i and a negative sample k, N is the training batch size, τ is a temperature super-parameter, the positive sample represents a graph generated by carrying out random data transformation on the current edge face connection graph, and the negative sample represents other edge face connection graphs except the current edge face connection graph in the same training batch.
9. The three-dimensional part retrieval method based on self-supervised graph representation learning of claim 2, wherein the constructing the boundary surface connection graph further comprises: constructing a three-layer hierarchical graph structure comprising a curved surface level graph, a feature level graph and a part level graph; the curved surface level graph is the edge-face connection graph; the feature level diagram is a feature subgraph which characterizes a specific geometric function through curvature clustering of curved surface nodes, wherein the special subgraph comprises holes, chamfers and bosses; The part level diagram is a global diagram for aggregating all curved surfaces and features; and the three layers of layered graph structures realize bidirectional message transmission through cross-layer connection edges.
10. The three-dimensional part retrieval method based on self-supervision graph characterization learning according to claim 9, wherein the part feature library is a multi-granularity vector database for storing global graph embedded vectors of the part and embedded vectors of all feature subgraphs, the information of the part is returned according to cosine similarity ordering, and the three-dimensional part retrieval method comprises the following steps: The first level search comprises the steps of calculating cosine similarity of the query part and the embedded vector of the part global diagram in the part feature library, screening out a result with the similarity higher than a first threshold value, and obtaining a first screening result; in the first screening result, calculating cosine similarity of the embedded vector of the feature subgraph in the query part and the first screening result, and screening out a result with the cosine similarity higher than a second threshold value to obtain a second screening result; And thirdly, screening the second-stage screening result, performing parameterized geometric equation comparison and tolerance fit surface analysis, generating assembly compatibility, and sorting according to compatibility scores to obtain a final retrieval part list.

Description

Three-dimensional part retrieval method based on self-supervision graph characterization learning Technical Field The invention belongs to the field of deep learning, and particularly relates to a three-dimensional part retrieval method based on self-supervision graph representation learning. Background With the development of industrial digitization, the number of CAD models for computer-aided design has grown dramatically, and how to efficiently retrieve target parts has become a key issue in the engineering field. Traditional CAD part retrieval relies primarily on manually designed geometric features, but such methods have limited generalization capability for complex structures. With the rise of deep learning, the method based on 3D deep learning, multi-view rendering and 2D CNN significantly improves the retrieval precision. However, parameterized geometric features are lost when CAD is converted into point cloud/grid, and the main stream 3D retrieval method is based on sensor data and cannot utilize the complete design semantics of the CAD original format. And the labeling cost of the 3D parts is higher, so that not only is a great deal of time consumed by professional engineering personnel to accurately and manually label the information such as the category, the functional attribute, the topological characteristic and the like of the parts, but also a large number of CAD models are difficult to cover to form a large-scale high-quality labeling data set, and the method is limited to be widely applied to the improvement of the retrieval performance of the large number of CAD models in industrial scenes. Disclosure of Invention The application provides a three-dimensional part retrieval method based on self-supervision graph characterization learning, which solves the technical problem of high retrieval cost caused by the fact that the part retrieval method relies on manual labeling of parts in the prior art. In order to achieve the above purpose, the application adopts the following technical scheme: In a first aspect, a three-dimensional part retrieval method based on self-supervised graph representation learning is provided, including: Acquiring and analyzing a CAD three-dimensional part data set in a boundary representation format, and constructing an edge-face connection diagram according to the topological relation between curved surfaces in the three-dimensional part to obtain an undirected diagram; inputting the undirected graph into a coder-decoder structure, performing node-level self-supervision training on node information in the boundary connection graph, and reserving the trained coder as a node characteristic extractor; Performing graph-level self-supervision training on the boundary connection graph by using a node feature extractor to obtain a pre-trained part retrieval model; And extracting feature vectors of all parts in the three-dimensional part data set by using the part retrieval model, and storing the feature vectors as a part feature library, wherein the part feature library is used for carrying out cosine similarity calculation on the feature vectors of the three-dimensional parts to be retrieved and returning information of the parts according to cosine similarity sequencing. According to the technical scheme, in the three-dimensional part retrieval method based on self-supervision graph representation learning, through a two-stage self-supervision learning mode, the characteristic representation of different three-dimensional parts can be automatically learned without manually marking the category or attribute information of the three-dimensional parts, meanwhile, the topological relation between a curved surface and a curve is directly extracted from the Brep format CAD data through constructing an edge-face connection graph, the essential geometric structure and the topological information of the three-dimensional parts are accurately reserved, in addition, the node-level self-supervision training is used for focusing on the reconstruction of local curved surface characteristics, the graph-level self-supervision training is used for focusing on the distinguishing property of a global structure, the two-stage training synergistic effect enables the model to learn the local geometric details and the global topological structure of the parts at the same time, the extracted characteristic vector is more representative, the quick matching of the parts to be retrieved and the characteristic library can be realized, the retrieval precision and the retrieval efficiency are suitable for the actual industrial part retrieval requirement. Further, the constructing an edge-face connection graph according to the topological relation between the curved surface and the curve in the three-dimensional part includes: Analyzing the CAD file in the boundary representation format by adopting an open source geometric kernel to obtain the geometric structure information of the three-dimensional p