Multi-modal information guided CAD model automatic assembly method based on boundary representation method

Abstract

The invention discloses a multi-modal information guided CAD model automatic assembly method based on a boundary representation method, and belongs to the technical field of computer aided design. The method comprises the following steps: constructing a B-Rep graph of a part to form a joint connection graph; extracting vertex features by using MLP and MPN, and extracting image features through ResNet; carrying out edge convolution analysis on the joint connection diagram after multi-modal features are fused, and predicting a potential joint axis position; and finally, optimizing joint attitude parameters by adopting a neural guidance search algorithm to realize automatic assembly. According to the method, by fusing geometric and image multi-modal information, the problem of single data modal in the prior art is solved, and the assembly precision is remarkably improved; b-Rep representation is combined with neural network guide search, so that the calculation complexity is reduced; the designed feature extraction and fusion mechanism enhances the adaptability of the method to different assembly scenes, is particularly suitable for automatic assembly of complex part models, and can be widely applied to the fields of machine manufacturing, aerospace and the like.

Description

Technical Field

[0001] The present invention belongs to the field of computer-aided design and intelligent manufacturing technology, and specifically relates to a multimodal information-guided CAD model automatic assembly method based on boundary representation (B-Rep), which is particularly suitable for the automated assembly scenarios of mechanical parts with complex geometric shapes. Background Art

[0002] Assembly of computer-aided design (CAD) models is a core component of modern manufacturing, widely used in the automotive, aerospace, and machinery industries. The assembly process involves precise coordination between multiple components, including geometric constraints, kinematic relationships, and functional matching. Automated assembly technology can significantly improve design efficiency, reduce human error, and shorten product development cycles. With the advancement of intelligent manufacturing, the demand for efficient and accurate automated assembly technology is growing.

[0003] Currently, two main approaches are used for CAD model assembly: traditional manual constraint methods and deep learning-based automated assembly methods. Traditional methods rely on manual geometric constraint setting between components (such as alignment, coaxiality, and fit). Designers spend a considerable amount of time and effort defining the precise position and assembly relationships of each part. While suitable for simple assembly scenarios, they are inefficient in complex assemblies and prone to assembly failures due to human negligence. In recent years, deep learning-based automated assembly technology has gradually emerged, using neural networks to train and identify assembly relationships between parts, reducing manual intervention. However, most existing methods rely on single-modal data (such as geometric or image information), resulting in insufficient recognition accuracy in complex assembly environments. For example, relying solely on geometric information makes it difficult to distinguish between parts with similar functions but different structures, while relying solely on image information is susceptible to interference from factors such as viewing angle and lighting. Furthermore, existing algorithms have high computational complexity and poor adaptability when dealing with multi-level constraints or non-standard parts, making them difficult to meet the needs of large-scale industrial applications.

[0004] The current CAD automatic assembly technology has the following main problems:

[0005] (1) Single data modality: Relying only on geometric or image information, it is difficult to fully understand the functions and assembly relationships between parts, resulting in a high recognition error rate.

[0006] (2) Low computational efficiency: In complex assembly scenarios, the algorithm needs to process a large number of constraints, which results in a heavy computational burden and makes it difficult to meet real-time assembly requirements.

[0007] (3) Insufficient adaptability: For non-standard parts, multi-degree-of-freedom assembly or dynamic assembly scenarios, the generalization ability of existing methods is limited and still requires manual adjustment.

[0008] (4) Dependence on manual labeling: Deep learning models require a large amount of labeled data for training. The cost of data acquisition is high, and the quality of labeling directly affects the model performance.

[0009] Therefore, there is an urgent need for an intelligent assembly method that integrates multimodal information (such as geometry, topology, functional semantics, etc.) to improve assembly accuracy, reduce computational complexity, and enhance adaptability to different assembly scenarios, thereby promoting the development of intelligent manufacturing. Summary of the Invention

[0010] In light of this, the present invention proposes a multimodal information-guided automatic assembly method for CAD models based on boundary representation. By integrating multiple modal information, such as geometry and images, this method improves assembly accuracy and efficiency, reduces reliance on manual intervention, and enables rapid and accurate automatic assembly of complex CAD models. This method addresses existing issues, such as incomplete assembly feature extraction and low accuracy in identifying complex assembly relationships, caused by reliance on a single data modality, as well as the inefficiency of traditional optimization algorithms in searching for assembly poses.

[0011] The technical solution adopted in the present invention is as follows:

[0012] The present invention provides a multimodal information-guided CAD model automatic assembly method based on boundary representation, comprising:

[0013] S1. For each input part, construct its B-Rep graph, where the vertices represent the faces and edges of the part;

[0014] S2, connect the B-Rep graphs of each part to form a densely connected joint connection graph;

[0015] S3. Using a multilayer perceptron and a message passing network to extract features from the joint connection graph to obtain vertex features; wherein the multilayer perceptron is used to perform preliminary processing on the vertex features in the B-Rep graph to extract the local geometric features of each vertex; the message passing network further captures the mutual relationships between vertices through a message passing mechanism to enhance feature representation;

[0016] S4. Use the ResNet network to extract features from the single-view image formed after the parts are assembled, and fuse the image features into the vertex features to form a fused image;

[0017] S5. Use edge convolution to analyze the fused image to predict whether there is a potential joint on each edge and determine the possible joint axis position;

[0018] S6. Based on the predicted joint axis position, a neural-guided search algorithm is used to optimize the joint posture, adjust its offset, rotation angle, and flip parameters, find the optimal posture, and realize automatic assembly of parts.

[0019] Furthermore, a multilayer perceptron and a message passing network are used to extract features from the joint connection graph to obtain vertex features, including:

[0020] The vertex features x1 and x2 from the two parts are processed separately by two independent multilayer perceptrons to create the B-Rep graph vertex embedding; one multilayer perceptron is used for the vertex features of B-Rep faces, and the other is used for the vertex features of B-Rep edges;

[0021] Connect the generated vertex embeddings together to form joint connection graphs G1 and G2;

[0022] A two-layer graph attention network is used to perform message passing in each joint-connected graph to obtain new embeddings h1 and h2 for each vertex:

[0023] h1=f enc (x1,G1),h2=f enc (x v ,G2).

[0024] Furthermore, the ResNet network is used to extract features from the single-view image formed after the parts are assembled, and the image features are fused into the vertex features, including:

[0025] The ResNet network is used to extract features from the single-view image formed after the parts are assembled. The single-view image y is aligned with the size of the joint connection graph and then fused into the vertex features h1 and h2 to increase the probability that the corresponding position node is predicted to be a joint:

[0026] H(y)=F(y)+y.

[0027] Furthermore, edge convolution is used to analyze the fused image to predict whether there is a potential joint on each edge and determine the possible joint axis position, including:

[0028] Use a shared encoder network f enc Encode the node features x1 and x2 in graphs G1 and G2 to generate 384-dimensional vertex embeddings h1 and h2;

[0029] For a densely connected graph G j For each edge (u,v) in the graph representing the connection between G1 and G2, a logit value is predicted based on these embeddings to indicate whether a joint exists:

[0030]

[0031] Where φ: It is a 3-layer MLP. is the cascade operator, h u and hv According to G j The source and target vertices of each edge in are obtained from h1 and h2.

[0032] Furthermore, the loss function for training the network includes:

[0033] Item 1 is the edge prediction h uv and ground truth edge label j uv The cross entropy between ∈{0,1} is normalized to a probability distribution

[0034]

[0035] In the softmax operation, the subscript indicates that the operation is applied to the graph G j For all edges in , the loss function is defined as CE(p,q) = -∑ i p i logq i ;

[0036] Using symmetric cross entropy loss As the second term of the loss function:

[0037]

[0038] In the softmax operation, the subscript 2D indicates that it is performed on a single axis of the reshaped matrix, rather than being applied directly to each edge (u,v).

[0039] Furthermore, a neural-guided search is employed to enumerate the top-k joint axis predictions and directly consider the interaction between the two parts.

[0040] Furthermore, a neural-guided search algorithm is used to optimize the joint posture, including: using the cost function To evaluate candidate joint configurations, the function is based on two key criteria: the overlap volume and the contact area between parts, as given by:

[0041]

[0042] Let V1 and V2 be the volumes of the two parts, V 1∩2 is their overlapping volume; A1 and A2 are the surface areas of the two parts, A 1∩2 is their contact area;

[0043] When the overlap ratio When , set λ = -10 to reduce the penalty for overlap; otherwise, set λ = 0 to increase the penalty for overlap;

[0044] Given the cost function, the Nelder-Mead algorithm is used for derivative-free optimization to search for the optimal joint pose.

[0045] Furthermore, when predicting the joint axes of each part, the joint axes are aligned by three parameters: the offset distance along the joint axis, the rotation around the joint axis, and the flip parameter that reverses the direction of the joint axis.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] This paper proposes a multimodal information-guided automatic assembly method for CAD models based on boundary representation. By fusing geometric features represented by B-Rep with image visual features, a multimodal information fusion framework is constructed. A graph neural network is used to model the topological connection relationships between parts, enabling intelligent understanding of complex assembly scenarios. A neural-guided optimization algorithm is then employed to efficiently search for assembly poses, significantly improving the accuracy and efficiency of automatic assembly of CAD models. This method overcomes the limitations of traditional assembly techniques that rely on single-modal data, achieving full automation of the entire process, from part feature extraction and assembly relationship prediction to final assembly pose optimization. This method effectively improves the assembly accuracy and efficiency of complex CAD models. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0049] Figure 1 This is a flowchart of a multimodal information-guided B-Rep-based CAD model automatic assembly method in an embodiment of the present invention;

[0050] Figure 2 is a network structure diagram according to an embodiment of the present invention;

[0051] Figure 3 1. is a schematic diagram of a single-view image and joint axis prediction results in an embodiment of the present invention;

[0052] Figure 4 It is an assembly result diagram in an embodiment of the present invention. DETAILED DESCRIPTION

[0053] This paper provides a multimodal information-guided automatic assembly method for CAD models based on B-Rep representation to improve assembly accuracy and efficiency. By integrating geometric and image information, this method enables fast and accurate automatic assembly of complex CAD models, reducing reliance on manual intervention.

[0054] In order to make the person skilled in the art better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should belong to the scope of protection of the present application.

[0055] It should be noted that the terms "first", "second" and the like in the specification and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or a chronological sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0056] As shown in Figure 1 , a multi-modal information guided B-Rep representation based CAD model automatic assembly method in the embodiment is as follows:

[0057] S1, for each input part, construct its B-Rep graph.

[0058] Wherein, the B-Rep graph represents the geometry of the part by accurately describing the boundary of the part, including face, edge and other elements; the vertex in the B-Rep graph represents the face and edge of the part;

[0059] S2, connect the B-Rep graph of each part to form a densely connected joint connection graph.

[0060] By converting the geometric information of the part into a graph structure, a foundation is provided for subsequent feature extraction and assembly relationship prediction.

[0061] S3, using multi-layer perceptron (MLP) and message passing network (MPN) to extract features of the formed joint connection graph, to obtain vertex features.

[0062] Wherein, the MLP is used to preliminarily process the vertex features in the B-Rep graph, and extract the local geometric features of each vertex; the MPN further captures the mutual relationship between the vertices through the message passing mechanism, and enhances the feature representation.

[0063] S4. Use the ResNet network to extract features from the single-view image formed after the parts are assembled, and fuse the image features into the vertex features to form a fused image;

[0064] Among them, the single-view image formed after the parts are assembled is used as the verification image. The ResNet network can effectively extract high-level semantic features from the image. These features, combined with geometric features, can more comprehensively describe the characteristics of the parts and improve the accuracy of assembly relationship recognition.

[0065] S5. Use edge convolution to analyze the fused image to predict whether there is a potential joint on each edge and determine the possible joint axis position.

[0066] Edge convolution aggregates the edge information connecting two parts to generate a feature vector for prediction. This process uses a neural network to automatically identify potential assembly relationships between parts by learning from a large number of assembly examples, reducing manual intervention.

[0067] S6. Based on the predicted joint axis position, a neural-guided search algorithm is used to optimize the joint posture, adjust its offset, rotation angle, and flip parameters, find the optimal posture, and realize automatic assembly of parts.

[0068] By optimizing the joint posture, the assembly results are ensured to meet the physical constraints and functional requirements, thus improving the quality and reliability of the assembly.

[0069] In the specific implementation, for graph vertex features, B-Rep face and edge information is used to determine that B-Rep faces use one-hot encoding (one-hot) vectors to represent surface types (such as planes and cylinders) and include a flag to indicate whether the surface is reversed. B-Rep edges use one-hot vectors to represent curve types (such as lines and arcs), combined with the edge length and a Boolean flag to indicate whether the curve is reversed relative to the edge direction. This example evaluates the performance of these input features combined with other features to optimize model performance.

[0070] For B-Rep faces, the one-hot vector of the surface type (e.g., plane, cylinder), the area of ​​the face, and a flag indicating whether the surface is inverted are evaluated. For B-Rep edges, the one-hot vector of the curve type (e.g., line, arc), the edge length, a Boolean flag indicating whether the curve is inverted relative to the edge, and the one-hot vector of the edge's dihedral angle and its concavity (e.g., convex, concave) are evaluated. Combinations of these features are used to analyze their impact on model performance. The model is trained by sequentially removing input features and evaluating the top-1 accuracy on the validation set to determine the contribution of each feature, including comparisons using all features. The study found that the length feature was the most critical to model performance, followed by the entity type. This result is consistent with the common use of cylinders at part interfaces, as the network is able to infer the B-Rep surface type and the length of adjacent B-Rep edges. Specifically, because the length of the circular edge at the end of a bolt or the outer edge of a hole is proportional to the hole radius, edge length is an effective means of distinguishing bolts and holes of similar sizes. In contrast, the area feature did not significantly improve performance and may even have a negative impact. Therefore, the model performance can be improved by removing the worst performing features: area, dihedral angle, and convexity.

[0071] like Figure 2 As shown, the upper half of the encoder neural network f enc First, vertex features x1 and x2 from the two parts are processed separately through two independent multi-layer perceptrons (MLPs) to create B-Rep graph vertex embeddings. One MLP is used for vertex features of B-Rep faces, and the other is used for vertex features of B-Rep edges; the generated vertex embeddings are then concatenated together to form joint connection graphs g1 and g2. Next, two layers of GATv2 (graph attention network) are used to perform message passing in each joint connection graph to obtain new embeddings h1 and h2 for each vertex:

[0072] h1=f enc (x1,G1),h2=f enc (x2,G2);

[0073] The lower part of the encoder uses a ResNet network to extract features from the assembled image, aligns these image features y with the size of the joint connection graph, and fuses them into the vertex features h1 and h2 to increase the probability that the corresponding position node is predicted to be a joint:

[0074] H(y)=F(y)+y;

[0075] Among them, F(y) represents the image features after feature extraction,

[0076] When creating joints, defining the joint axis is a crucial design consideration, allowing the two parts to align and constrain each other. This joint axis not only defines the degrees of freedom between the parts but also provides the basis for downstream tasks such as assembly, part mobility, and animation. Predicting the joint axis is formulated as a link prediction problem, where the core goal is to identify connections between graphs G1 and G2 that correctly align the two parts along the true joint axis. This is accomplished by aggregating information along the edges connecting the two parts using edge convolutions.

[0077] Specifically, we first use a shared encoder network f enc Encode the node features x1 and x2 in the graphs G1 and G2 to generate 384-dimensional vertex embeddings h1 and h2. Then, for the densely connected graph G j For each edge (u,v) in the graph representing the connection between G1 and G2, a logit value is predicted based on these embeddings to indicate whether a joint exists:

[0078]

[0079] Where φ: It is a 3-layer MLP. is the cascade operator, h u and h v According to G j The source and target vertices of each edge in are obtained from h1 and h2.

[0080] The network is trained using a loss function with two terms. The first term is the edge prediction h uv and ground truth edge label j uv The cross entropy between ∈{0,1} is normalized to a probability distribution

[0081]

[0082] In the softmax operation, the subscript indicates that the operation is applied to the graph G j For all edges in , the loss function is defined as CE(p,q) = -∑ i p i logq i This cross entropy loss encourages true joints to have higher prediction values ​​while suppressing the prediction values ​​of non-joints. However, since positive labels (i.e. true joints) are relatively sparse, this approach is suboptimal because it involves summing over a large number of negative samples. In order to better focus on the key loss terms and enhance the contrast between true joints and possible non-joints, a symmetric cross entropy loss is introduced in this embodiment. As the second term of the loss function:

[0083]

[0084] In the softmax operation, the subscript 2D indicates that it is performed on a single axis of the reshaped matrix, rather than being applied directly to each edge (u,v). Specifically, the graph G j The predictions and ground truth labels for all edges in are reshaped into an n × m matrix. This reshaping allows softmax to be applied across the columns (or rows) of the matrix, ensuring that each joint prediction is normalized in the correct context.

[0085] The B-Rep entities predicted by the network enable us to query the ground truth B-Rep data to obtain the joint axis predictions for each part. These axes are aligned by three parameters: the offset distance along the joint axis, the rotation around the joint axis, and the flip parameter that reverses the direction of the joint axis. A neural-guided search is used to enumerate the top-k joint axis predictions and directly consider the interaction between the two parts. In this way, not only the accuracy of the joint axis prediction is ensured, but also the overall quality of the joint configuration is optimized, so that the final joint is both physically constrained and functionally reasonable. In order to evaluate the candidate joint configurations, a cost function is proposed in this embodiment. The function is based on two key criteria: the overlap volume and the contact area between parts, and is formulated as:

[0086]

[0087] Let V1 and V2 be the volumes of the two parts, V 1∩2 is their overlapping volume; A1 and A2 are the surface areas of the two parts, A 1∩2 is their contact area. Intuitively, minimizing the cost function should encourage larger contact areas while penalizing overlapping volumes to prevent penetration. When λ = -10, the penalty for overlap is reduced; otherwise, the penalty for overlap is increased by setting λ = 0. Given this cost function, the Nelder-Mead algorithm is used for derivative-free optimization to search for the optimal joint pose. This algorithm efficiently finds the joint configuration that minimizes the cost function in complex space, ensuring close and physically plausible alignment between the two parts.

[0088] like Figure 3 、 Figure 4 As shown, Figure 3 The one on the left is a single-view image after the parts are assembled. Figure 3 The middle right side shows the result of joint axis prediction.

[0089] In the above embodiment, feature extraction based on B-Rep representation uses B-Rep (boundary representation) to represent the part's geometry, enabling a more accurate description of geometric features such as faces, edges, and vertices. By constructing a B-Rep graph and joint connection graph, efficient extraction and representation of part features are achieved, providing a solid foundation for subsequent assembly relationship prediction.

[0090] Multimodal Information Fusion: This technology combines part geometry with image information, leveraging the complementary strengths of multimodal data to improve the accuracy and robustness of assembly relationship recognition. Geometric information provides the precise shape and size of a part, while image information captures its appearance and texture details, helping to more accurately match parts in complex scenarios.

[0091] Neural-guided search algorithm: A neural network-guided search algorithm is introduced to optimize the joint pose determination process. By learning from a large number of assembly examples, the neural network can intelligently guide the search direction, quickly find the optimal assembly pose, and improve assembly efficiency and quality.

[0092] Efficient feature extraction and fusion mechanism: Design an efficient feature extraction and fusion module that can extract the most discriminative features from multimodal data and effectively fuse these features together to improve the model's expressive power and prediction accuracy.

[0093] In summary, the present invention overcomes the limitations of traditional assembly methods through the automatic assembly method of CAD models guided by multimodal information, and provides a more efficient and reliable automatic assembly solution for the design and manufacture of complex products. It has important application value and broad market prospects.

[0094] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multimodal information guided CAD model automatic assembly method based on boundary representation, characterized in that: include: S1. For each input part, construct its B-Rep graph, where the vertices represent the faces and edges of the part; S2, connect the B-Rep graphs of each part to form a densely connected joint connection graph; S3. Using a multilayer perceptron and a message passing network to extract features from the joint connection graph to obtain vertex features; wherein the multilayer perceptron is used to perform preliminary processing on the vertex features in the B-Rep graph to extract the local geometric features of each vertex; the message passing network further captures the mutual relationships between vertices through a message passing mechanism to enhance feature representation; S4. Use the ResNet network to extract features from the single-view image formed after the parts are assembled, and fuse the image features into the vertex features to form a fused image; S5. Use edge convolution to analyze the fused image to predict whether there is a potential joint on each edge and determine the possible joint axis position; S6. Based on the predicted joint axis position, a neural-guided search algorithm is used to optimize the joint posture, adjust its offset, rotation angle, and flip parameters, find the optimal posture, and realize automatic assembly of parts.

2. The method for automatic assembly of CAD models guided by multimodal information and based on boundary representation according to claim 1, characterized in that: The joint connection graph is subjected to feature extraction using a multilayer perceptron and a message passing network to obtain vertex features, including: The vertex features x1 and x2 from the two parts are processed separately by two independent multilayer perceptrons to create the B-Rep graph vertex embedding; one multilayer perceptron is used for the vertex features of B-Rep faces, and the other is used for the vertex features of B-Rep edges; Connect the generated vertex embeddings together to form joint connection graphs G1 and G2; A two-layer graph attention network is used to perform message passing in each joint-connected graph to obtain new embeddings h1 and h2 for each vertex: h1=f enc (x1,G1),h2=f enc (x2,G2)。 3. The multimodal information-guided CAD model automatic assembly method based on boundary representation according to claim 2, characterized in that: Use the ResNet network to extract features from the single-view image formed after the parts are assembled, and fuse the image features into the vertex features, including: The ResNet network is used to extract features from the single-view image formed after the parts are assembled. The single-view image y is aligned with the size of the joint connection graph and then fused into the vertex features h1 and h2 to increase the probability that the corresponding position node is predicted to be a joint: H(y)=F(y)+y.

4. The method for automatic assembly of CAD models guided by multimodal information and based on boundary representation according to claim 1, characterized in that: Edge convolution is used to analyze the fused image to predict whether there is a potential joint on each edge and determine the possible joint axis position, including: Use a shared encoder network f enc Encode the node features x1 and x2 in graphs G1 and G2 to generate 384-dimensional vertex embeddings h1 and h2; For a densely connected graph G j For each edge (u,v) in the graph representing the connection between G1 and G2, a logit value is predicted based on these embeddings to indicate whether a joint exists: h uv =φ(h u ⊕h v ); in is a 3-layer MLP, ⊕ is a cascade operator, h u and h v According to G j The source and target vertices of each edge in are obtained from h1 and h2.

5. The method for automatic assembly of CAD models guided by multimodal information and based on boundary representation according to claim 1, characterized in that: The loss function for training the network includes: Item 1 is the edge prediction h uv and ground truth edge label j uv The cross entropy between ∈{0,1} is normalized to a probability distribution In the softmax operation, the subscript indicates that the operation is applied to the graph G j For all edges in , the loss function is defined as CE(p,q) = -∑ i p i logq i ; Using symmetric cross entropy loss As the second term of the loss function: In the softmax operation, the subscript 2D indicates that it is performed on a single axis of the reshaped matrix, rather than being applied directly to each edge (u,v).

6. The method for automatic assembly of CAD models guided by multimodal information and based on boundary representation according to claim 1, characterized in that: A neural-guided search is employed to enumerate the top-k joint axis predictions and directly consider the interaction between the two parts.

7. The multimodal information-guided CAD model automatic assembly method based on boundary representation according to claim 6, characterized in that: A neural-guided search algorithm is used to optimize the joint posture, including: using a cost function To evaluate candidate joint configurations, the function is based on two key criteria: the overlap volume and the contact area between parts, as given by: Let V1 and V2 be the volumes of the two parts, V 1∩2 is their overlapping volume; A1 and A2 are the surface areas of the two parts, A 1∩2 is their contact area; When the overlap ratio When , set λ = -10 to reduce the penalty for overlap; otherwise, set λ = 0 to increase the penalty for overlap; Given the cost function, the Nelder-Mead algorithm is used for derivative-free optimization to search for the optimal joint pose.

8. The multimodal information-guided CAD model automatic assembly method based on boundary representation according to claim 6, characterized in that: When the joint axes of each part are predicted, the individual joint axes are aligned by three parameters: the offset distance along the joint axis, the rotation around the joint axis, and the flip parameter that reverses the direction of the joint axis.

Images