A method and system for automatically generating a steel structure welding deepening model based on BIM

Abstract

The application belongs to the technical field of steel structure deepening design, and discloses a steel structure welding deepening model automatic generation method and system based on BIM, which comprises the following steps: performing part-level feature extraction and process semantic analysis based on a BIM design model, and constructing a part-level feature set and an associated feature set; wherein the part-level feature set comprises part geometric features, part attribute features and cross-section process semantic features; the associated feature set comprises connection types, part docking contact features and part pose features; features of the feature set are extracted by using a pre-trained encoder, and the features are taken as feature inputs to construct a primary weld prediction function and a secondary weld prediction function, so that the primary weld specification parameters and the secondary weld specification parameters are predicted; and the application provides an intelligent welding deepening model generation method which can meet welding requirements and is beneficial to improving the accuracy, consistency and automation level of welding design.

Description

Technical Field

[0001] This invention belongs to the technical field of steel structure detailed design, and more specifically, relates to a method and system for automatically generating detailed welding models of steel structures based on BIM. Background Technology

[0002] In the detailed design phase of steel structure engineering, the design model needs to be transformed into an executable fabrication and on-site installation model, with welding detailing being a critical step. Information such as the type, location, size, and sequence of primary welds (assembly welds between sub-plates within a component section) and secondary welds (connection welds between components or parts) directly affects the component fabrication accuracy, structural reliability, on-site construction efficiency, and the feasibility of subsequent welding robot operations. Common complex nodes in steel structures, such as beam-column joints, corbel joints, and column base joints, involve various component forms and complex spatial relationships, resulting in cumbersome weld layouts. The welding relationships of the components within these nodes typically rely on manual judgment based on the experience of engineers. Manual modeling is not only time-consuming but also prone to overlooking critical welds, leading to inconsistent design quality and consequently affecting construction efficiency and safety.

[0003] Currently, automated welding detailing often employs methods based on manual rules, component library weld diagrams, or fixed logic, such as setting default weld configurations according to component type or cross-sectional dimensions. While these methods are applicable to simple, standardized nodes, they lack the ability to characterize complex cross-sectional structures and actual geometric contact relationships. They struggle to distinguish between the specific welding requirements of primary and secondary welds, exhibiting insufficient adaptability and generalization capabilities, and failing to cover different node types, engineering conditions, and construction process requirements. As project scale and node types increase, the maintenance and updating of the rule base becomes increasingly cumbersome, and it is difficult to directly support the automatic generation of welding robot trajectories and process parameters, resulting in low overall automation efficiency.

[0004] In summary, existing technologies suffer from technical bottlenecks in the steel structure welding process, including high reliance on manual labor, low levels of intelligence, and insufficient ability to handle complex nodes. Summary of the Invention

[0005] In view of the above-mentioned defects or improvement needs of existing technologies, this invention provides a method and system for automatically generating detailed steel structure welding models based on BIM, which is used to solve the technical bottlenecks of existing technologies in the process of detailed steel structure welding, such as high dependence on manual labor, low level of intelligence, and insufficient ability to handle complex nodes.

[0006] To achieve the above objectives, according to one aspect of the present invention, a method for automatically generating a BIM-based steel structure welding detail model is provided, comprising:

[0007] S1, based on the BIM design model, perform part-level feature extraction and process semantic parsing, construct part-level feature sets for parts, and construct association-level feature sets for parts that are determined to be connected;

[0008] The part-level feature set includes part geometric features, part attribute features, and cross-sectional process semantic features. The cross-sectional process semantic features include cross-sectional type, assembly order of parts in the component to which the part belongs, and standard welding joint form of each weld in the component.

[0009] The associated feature set includes connection type, part pair contact features and part pair pose features. The part pair contact features include contact area overlap rate, connection angle and semantic identifier of contact part.

[0010] S2 utilizes the pre-trained encoder to extract feature representations of the part-level feature set and the association-level feature set, and uses the feature representations as feature inputs to construct primary weld prediction functions and secondary weld prediction functions respectively. By performing supervised learning on labeled sample sets, the prediction of primary weld specification parameters and secondary weld specification parameters is achieved, thereby generating a welding detail model.

[0011] According to the automatic generation method for detailed steel structure welding models based on BIM provided by the present invention, the geometric features of the parts include part position parameters, part orientation parameters, part dimensional parameters, and spatial enclosure parameters; the attribute features of the parts include part type, material designation, globally unique identifier of the part, component identifier field to which the part belongs, and component cross-section description string associated with the component to which the part belongs. The component cross-section description string is used to represent the construction type and cross-section type of the corresponding component; the construction type includes welding assembly type and steel section type, and the cross-section type is the specific shape type under the construction type;

[0012] The connected part pairs specifically include: display part pairs, which are obtained from the BIM design model and are displayed part pairs; and candidate part pairs, which are determined by the minimum spacing of the bounding boxes of the parts. When the minimum spacing of the bounding boxes is less than a preset spacing threshold, the corresponding part pairs are determined as candidate part pairs.

[0013] According to the BIM-based automatic generation method for steel structure welding detail model provided by the present invention, the construction of cross-sectional process semantic features is as follows:

[0014] When determining that a part belongs to a welded assembly type, the cross-section type is the cross-section type of the component. Based on the preset shape and specification parameters of each functional role within the component, the parts within the component are traversed to identify the functional role of each part in the cross-section composition as the basic role of the part. Thus, the assembly order of each part is determined based on the basic role of the part, and the standard welded joint form of the corresponding weld is determined based on the connection type between connected parts.

[0015] When determining that a component belongs to a steel section type, the cross-section type is the steel section type; the assembly sequence and standard welding joint form are both preset fixed quantities;

[0016] When determining that a part is an accessory, the cross-section type is obtained from the BIM design model; the assembly sequence and standard welded joint form are both preset fixed quantities.

[0017] According to the automatic generation method of steel structure welding detail model based on BIM provided by the present invention, the part-level feature set also includes part role semantic identifier; for parts whose components are of the welding assembly type, the part role semantic identifier is the basic part role; for parts whose components are of the section steel type and parts that are of the accessory type, the part role semantic identifier is a preset fixed quantity.

[0018] Correspondingly, for parts whose components are steel profiles, the equivalent functional roles of the steel profile geometric surfaces are identified based on the shape specifications of the preset equivalent functional roles. Then, using the equivalent functional roles as spatial reference benchmarks, the auxiliary functional roles of other accessories are identified based on the shape specifications of the preset auxiliary functional roles.

[0019] Furthermore, based on the basic role, equivalent functional role, and auxiliary functional role of the part, a semantic identifier for the contact part is constructed from the associated feature set.

[0020] According to the automatic generation method for steel structure welding detail model based on BIM provided by the present invention, after S1 and before S2, it further includes: constructing a weld semantic-spatial-topology fusion diagram including nodes and connecting edges based on the part-level feature set and the association-level feature set;

[0021] In this context, nodes represent parts, and connecting edges are set between pairs of parts that are determined to be connected. For any graph node, the node is assigned an original feature vector based on the part-level feature set corresponding to the node. For any connecting edge, the connecting edge is assigned an original feature vector based on the corresponding association-level feature set.

[0022] Correspondingly, in S2, part-level feature representation is extracted based on graph mask autoencoder.

[0023] According to the BIM-based automatic generation method for detailed steel structure welding models provided by this invention, the training of the graph mask autoencoder specifically includes:

[0024] Multiple BIM design models were selected to construct weld semantic-spatial-topology fusion maps, and a pre-training dataset including multiple weld semantic-spatial-topology fusion maps was established.

[0025] For the original node features and original edge features of the samples in the pre-training dataset, random masking is performed on the features. The graph mask autoencoder is used to extract the features of the node mask features and edge mask features after random masking. A node feature reconstruction mapping function is constructed to reconstruct the node features based on the extracted node features and obtain the node feature reconstruction loss. An edge feature reconstruction mapping function is constructed to reconstruct the edge features based on the extracted edge features and obtain the edge feature reconstruction loss.

[0026] The combined loss for training a graph mask autoencoder includes node feature reconstruction loss and edge feature reconstruction loss.

[0027] According to the BIM-based automatic generation method for detailed steel structure welding models provided by this invention, the training of the graph mask autoencoder further includes:

[0028] A node-level process role classification head is constructed to identify the functional role of nodes based on extracted node features, and a loss for role prediction task is constructed based on the semantic identifiers of part roles in the part-level feature set. An edge-level connection pattern classification head is constructed to predict the connection type and contact semantic identifiers of edges based on extracted edge features, and a loss for edge-level connection pattern prediction task is constructed based on the connection type and contact semantic identifiers in the association-level feature set.

[0029] The overall loss for training a graph mask autoencoder also includes the loss for the role prediction task and the loss for the edge connection pattern prediction task.

[0030] According to the BIM-based automatic generation method for steel structure welding detail model provided by the present invention, S2 specifically includes:

[0031] Based on the component attribute characteristics, the components belonging to the welded assembly type are selected. The component-level feature set corresponding to the selected components is used as the input of the pre-trained encoder. The component feature representation output by the encoder is input into the weld prediction function to predict the weld specification parameters.

[0032] The parts pairs that are determined to be connected are filtered to find the parts pairs with secondary welds. The correlation-level feature set of the selected parts pairs is used as the input of the pre-trained encoder. The correlation feature representation output by the encoder is input into the secondary weld prediction function to predict the secondary weld specification parameters.

[0033] According to the automatic generation method of steel structure welding detail model based on BIM provided by the present invention, the primary weld specification parameters include: primary weld existence identifier, and primary weld process parameters including primary weld group intersection line coordinates, primary weld type and weld leg size parameters;

[0034] In the training phase of the primary weld prediction function, the primary weld existence prediction loss is constructed by combining the primary weld existence identifier and the actual label of the primary weld specification parameters. The primary weld process parameter prediction loss is constructed based on the predicted and labeled values ​​of the primary weld process parameters. The two losses are weighted and summed according to the preset weight coefficients to obtain the comprehensive loss for training the primary weld prediction function.

[0035] The secondary weld specifications include: the secondary weld existence identifier, and the secondary weld process parameters including the secondary weld connection interface boundary trajectory, secondary weld type and weld leg size parameters;

[0036] During the training phase of the secondary weld prediction function, the existence prediction loss of the secondary weld is constructed by combining the existence identifier and the actual label of the secondary weld specification parameters. The prediction loss of the secondary weld process parameters is constructed based on the predicted and labeled values ​​of the secondary weld process parameters. The two losses are weighted and summed according to the preset weight coefficients to obtain the comprehensive loss for training the secondary weld prediction function.

[0037] According to another aspect of the present invention, a BIM-based automatic generation system for detailed welding models of steel structures is provided. The system includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to perform the BIM-based automatic generation method for detailed welding models of steel structures as described above.

[0038] Overall, compared with the prior art, the BIM-based automatic generation method and system for steel structure welding detail models provided by this invention offer the following advantages:

[0039] 1. The extracted part-level feature set and association-level feature set can uniformly express the process semantics and micro-geometric contact relationships of component cross-sections. This enables the extracted feature representation to automatically understand and process complex component cross-sections and part contact relationships in BIM design models, and to distinguish and predict primary and secondary welds. It provides an intelligent welding detail model generation method that can meet welding requirements, which is beneficial for providing a high-quality weld data foundation for subsequent weld statistics, process planning, and offline programming of welding robots. This significantly improves the accuracy, consistency, and automation level of welding design, and promotes the development of intelligent manufacturing and digital construction technologies.

[0040] 2. This part-level feature not only covers microscopic geometric contact details (overlap rate, included angle, etc.), but also incorporates macroscopic process role semantics, which significantly improves the accuracy and generalization ability of the model in identifying primary / secondary welds when facing complex nodes and non-standard irregular cross-sections.

[0041] 3. A semantic-spatial-topological fusion graph model was constructed. Through a graph mask autoencoder pre-training mechanism, a unified representation learning of geometric attributes, spatial topological relationships, and process semantics at the component level of steel structures was achieved for the first time. Compared with traditional rule-based methods that rely solely on component type or simple geometric proximity relationships, this invention utilizes a self-supervised mask reconstruction task to force the model to learn and capture the implicit correlation between cross-sectional construction logic and actual contact relationships of components, generating a high-dimensional component-level fusion feature vector.

[0042] 4. A graph neural network framework for weld prediction with feature sharing and task decoupling is proposed. Based on a pre-trained graph encoder, node-level first-order weld prediction functions and edge-level second-order weld prediction functions are constructed respectively, enabling independent parallel reasoning for the internal assembly relationships of components and the connection relationships between components / attachments. During fine-tuning, freezing the graph encoder parameters effectively avoids negative interference between multiple tasks, ensuring the stability of the feature space. The model directly outputs a structured data package containing existence identifiers and process parameters (including pose trajectory, type, and size). This highly integrated and concrete parameter definition method eliminates the ambiguity of traditional location information descriptions, providing a high-quality data foundation with physical determinism for the automated construction of BIM detailing models and welding robot path planning. Attached Figure Description

[0043] Figure 1 This is a diagram illustrating the components of the automatic generation method for steel structure welding detail model based on BIM, as described in an embodiment of the present invention.

[0044] Figure 2 This is a flowchart of the multi-dimensional feature extraction and process semantic parsing of parts based on BIM, according to an embodiment of the present invention.

[0045] Figure 3 This is a flowchart illustrating the construction process of the weld semantic-spatial-topology fusion graph in an embodiment of the present invention.

[0046] Figure 4 This is a flowchart illustrating the part-level feature representation learning process based on a graph mask autoencoder according to an embodiment of the present invention.

[0047] Figure 5 This is a flowchart illustrating the prediction process for primary / secondary weld specifications in an embodiment of the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0049] Please see Figures 1-5 This embodiment provides an automatic generation method for steel structure welding detail model based on Building Information Modeling (BIM), the method including:

[0050] S1, based on the BIM design model, perform part-level feature extraction and process semantic parsing, construct a part-level feature set for any part, and construct an association-level feature set for any pair of parts that are determined to be connected;

[0051] The part-level feature set includes part geometric features, part attribute features, and cross-sectional process semantic features. The cross-sectional process semantic features include cross-sectional type, assembly order of parts in the component to which the part belongs, and standard welding joint form of each weld in the component.

[0052] The associated feature set includes connection type, part pair contact features and part pair pose features. The part pair contact features include contact area overlap rate, connection angle and semantic identifier of contact part.

[0053] S2 utilizes the pre-trained encoder to extract feature representations of the part-level feature set and the association-level feature set, and uses the feature representations as feature inputs to construct primary weld prediction functions and secondary weld prediction functions respectively. By performing supervised learning on labeled sample sets, the prediction of primary weld specification parameters and secondary weld specification parameters is achieved, thereby generating a welding detail model.

[0054] like Figure 2 As shown, the multi-dimensional feature extraction and process semantic parsing of parts based on the BIM design model specifically includes the following parts:

[0055] (1) Obtaining part geometric features and part attribute features based on BIM design model.

[0056] For example, by using the application programming interface (API) provided by BIM software, one can traverse the parts in the BIM design model to obtain the geometric features and attribute features of the parts. The BIM design model is the design model before welding detailing, and it does not contain explicit weld objects or welding parameter information. A part refers to the smallest independent unit or auxiliary connector that constitutes a structure, including but not limited to flanges, webs, stiffeners, connecting plates, end plates, etc.; a component is defined as the assembly parent object of a part, usually corresponding to a welded assembly in the factory prefabrication stage or the basic unit of engineering installation. A component contains at least one of the aforementioned parts.

[0057] The geometric features of a part include at least the part's position parameters, orientation parameters, dimensional parameters, and spatial bounding parameters, as detailed below:

[0058] Position parameters: Part Local coordinate system The origin is in the global coordinate system 3D coordinates of parts The geometric centroid of the solid in the global coordinate system The three-dimensional coordinates below;

[0059] Attitude parameters: parts Local coordinate system The three orthogonal basis vectors in the global coordinate system The unit direction vector below;

[0060] Dimensional parameters: parts In its local coordinate system The following are the external dimensions (length, width, thickness);

[0061] Spatial bounding parameters: and parameters derived from the solid geometry in the global coordinate system. The calculated parameters of the directed bounding box (OBB) are then used.

[0062] Global coordinate system Used to uniformly express the absolute spatial position of parts in the BIM design model; the part's local coordinate system Fixed to parts Used to express parts Its own orientation and external dimensions.

[0063] The attribute characteristics of a part include at least the following: part type, material number, part globally unique identifier (GUID), component identifier field to which the part belongs, and component cross-section description string associated with the component to which the part belongs. The component cross-section description string is used to represent the construction type and cross-section type of the corresponding component. The construction type includes welded assembly type and steel section type, and the cross-section type is the specific shape type under the construction type.

[0064] Furthermore, discrete text fields (such as part type and material designation) in part attribute features are mapped to numerical index codes. Component cross-section description strings are stored at the component level. When extracting part attributes, based on the dependency relationship between the part and the component, this description string is inherited or mapped to the attribute set of each sub-part. The component cross-section description string refers to the parametric text recorded in the BIM design model that defines the cross-sectional shape and specifications of the component to which the part belongs.

[0065] (2) Construction of component association features:

[0066] Component association characteristics include at least: component affiliation and connection topology. Component affiliation refers to the set of components to which a component belongs, obtained by grouping components based on the component identifier field in the component attribute characteristics. Connection topology refers to the relationship records used to characterize whether there is a connection / assembly adjacency between components. The relationship record at least contains component pair identifiers and their connection type identifiers. Connection types include mating connection, abutment connection, corner connection, and insert / penetration connection, which can be corresponding to coded identifiers.

[0067] Relationship records are primarily obtained by parsing explicit connection objects in the BIM design model. When the model does not provide explicit connection objects, as a preliminary screening method, spatial proximity is determined based on OBB parameters. The spatial intersection state of the OBBs of two parts is calculated, and part pairs that satisfy OBB intersection or a distance less than a preset threshold are selected as candidate connection topology records for precise geometric verification in subsequent steps. The distance threshold is a configurable parameter that takes into account modeling errors. Specifically, connected part pairs include: displayed part pairs (obtained from the BIM design model with explicit connections) and candidate part pairs (determined by the minimum spatial bounding box spacing; when the minimum spatial bounding box spacing is less than a preset spacing threshold, the corresponding part pair is selected as a candidate part pair).

[0068] (3) The construction of the semantic features of the cross-section process is as follows:

[0069] For each component belonging to a set of parts, the component cross-section description string carried by any part in that set is extracted (since parts belonging to the same component share this component-level description); and the component cross-section description string is parsed based on a preset cross-section rule library to identify the main cross-section construction type of the component. The cross-section rule library can define the mapping logic between the regular expression of the cross-section description string and the main cross-section construction type. The main cross-section construction type includes welded assembly type and steel section type. Welded assembly type refers to a cross-section form formed by welding multiple plates, such as I-beam, box-shaped, and cross-shaped. Steel section type refers to a standard cross-section form formed by hot rolling or other non-plate assembly, such as H-beam, I-beam, channel steel, and angle steel.

[0070] For the parsed main body cross-section structure type, perform semantic role recognition:

[0071] When determining that a part belongs to a welded assembly type, the cross-section type is the cross-section type of the component. Based on the preset shape and specification parameters of each functional role within the component, the system iterates through all parts within the component, identifying the functional role of each part in the cross-section composition as the basic role of the part. Specifically, if it is a welded assembly type, its standard cross-section parameters are analyzed (a welded assembly type component parameter library can be established, recording the functional roles of each component and the dimensions of each functional role, such as web height, flange width, and thickness values ​​of each plate). The standard cross-section parameters of the component can be used as search conditions to iterate through the set of parts to which the component belongs. Through dimensional value matching and spatial position verification, the functional role of each part in the cross-section composition (including the basic roles of flange plates, web plates, etc.) is identified. Thus, the assembly order of each part is determined based on its basic role, and the standard weld joint form of the corresponding weld is determined based on the connection type between connected parts. When determining that a part belongs to a steel section type, the cross-section type is the steel section type; the assembly order and standard weld joint form are both preset fixed quantities. When determining that a part is an accessory type, the cross-section type is obtained from the BIM design model; the assembly order and standard weld joint form are both preset fixed quantities.

[0072] For parts whose components are steel profiles, the equivalent functional roles of the steel profile geometric surfaces are identified based on the shape specifications of the preset equivalent functional roles. Then, using the equivalent functional roles as spatial reference benchmarks, the auxiliary functional roles of other accessories are identified based on the shape specifications of the preset auxiliary functional roles. Finally, based on the basic roles, equivalent functional roles, and auxiliary functional roles of the parts, semantic identifiers of contact parts in the associated feature set are constructed.

[0073] If the component is a steel section, based on a standard steel section geometry library (a steel section component geometry library can be established, recording the equivalent functional roles of each steel section component and the dimensions of each equivalent functional role), specific geometric surfaces of the steel section entity are mapped to equivalent functional roles, and the steel section itself is marked as the section matrix. Based on this, using the mapped equivalent functional roles as spatial reference benchmarks, the adjacency relationships between other auxiliary parts (such as stiffeners and connecting plates) and the equivalent flanges or equivalent webs are identified, thereby determining the auxiliary functional roles of the parts.

[0074] The standard steel geometry library is a pre-set parametric database that stores the cross-sectional dimension parameters of various hot-rolled steel sections (including H-beams, I-beams, channel steel, angle steel, etc.) as specified by national or industry standards, as well as the mapping rules between geometric surfaces and functional roles (for example, defining the upper and lower surfaces of H-beams as "equivalent flanges" and the middle part as "equivalent webs").

[0075] Finally, based on the identified functional roles of the parts (including basic roles and auxiliary roles) and the constructed connection topology, the assembly rules in the process knowledge base are retrieved, and the component cross-section type code, plate assembly order and standard welding joint form are mapped into numerical features to generate cross-section process semantic features that characterize the component cross-section assembly process features (for welded assembly components, characterizing their main weld features).

[0076] The process knowledge base pre-sets a set of rules for component assembly and welding, including at least: joint type mapping rules (defining the correspondence between connection types and standard welded joint types), assembly order logic (defining assembly priority based on the functional roles of parts), and feature encoding table (defining the numerical encoding index of the above information). Standard welded joint types include butt joints, T-joints, corner joints, lap joints, end joints, and cross joints.

[0077] (4) Generation of contact features of parts.

[0078] Based on the acquired geometric features of the parts, the boundary representation (B-Rep) data of each part inside the component is extracted, and the normal vectors of each geometric surface on the outer surface and the coordinates of the vertex of the surface boundary are obtained by parsing. For each pair of parts, the following feature extraction steps are performed:

[0079] Contact geometry analysis: For a pair of display parts, the intersection line or curve is calculated based on the vertex coordinates of the contact area geometry, and the area enclosed by the intersection line is calculated; the ratio of this area to the complete surface area of ​​the geometry of the smaller geometric entity in the part pair that participates in the contact is defined as the contact area overlap rate (used to distinguish between mating connection and perpendicular / angular contact); for a pair of candidate parts, the contact surface of the larger part is determined as the projection reference plane, the smaller part is projected onto the projection reference plane, and the area of ​​the overlapping region between the projected profile and the projection reference plane is determined. The ratio of the overlapping region area to the projection reference plane is the contact area overlap rate.

[0080] Relative position and attitude analysis: The cosine value of the included angle between the contact surfaces characterizes the connection angle between the two parts (such as perpendicular, parallel or inclined).

[0081] Semantic role association: Based on the established functional roles of parts (including the equivalent functional roles after the steel profile is mapped), the connection semantics of both sides of the contact pair are parsed to generate semantic identifiers for the contact parts (e.g., "stiffening rib-web" or "end plate-flange"). When there is steel profile in the part pair, the equivalent functional role corresponding to the contact geometry of the steel profile is selected to establish the semantic identifiers for the contact parts.

[0082] Finally, the overlap rate of the contact area, the cosine value of the contact surface angle, and the contact part identification are uniformly encoded to generate part pair contact features that characterize the weld features of the internal auxiliary component assembly welds (i.e., all welds).

[0083] In addition, the normalized centroid coordinates in the geometric features of the part can be used to calculate the relative Euclidean distance and relative rotation quaternion as the pose features of the part.

[0084] (5) Data standardization and normalization:

[0085] The acquired part geometric features and part attribute features, constructed connection type features, generated cross-sectional process semantic feature vectors, and generated part-to-part contact feature vectors and part-to-part pose feature vectors are collected into an original feature set, and then uniformly processed as follows:

[0086] Perform unit conversion on the physical quantity values ​​(such as coordinates, distance, area) in the geometric features of the part, the contact feature vector of the part, and the pose feature vector of the part, and unify them into standard units of measurement;

[0087] Normalization is performed on the continuous numerical features (including geometric dimensions, coordinate values, contact overlap rate, etc.) after unifying the units mentioned above, and they are mapped to a preset dimensionless numerical range to eliminate the order-of-magnitude differences between features with different dimensions.

[0088] Fill zero values ​​or mark anomalies for feature fields that fail to be calculated or are missing.

[0089] Finally, a standardized data sequence is output, which includes: a part-level feature set indexed by the part's globally unique identifier (containing part geometric features, part attribute features, and cross-sectional process semantic feature vectors). The part-level feature set may also include part role semantic identifiers. For parts whose component is a welded assembly type, the part role semantic identifier is the part's basic role. For parts whose component is a structural steel type and parts that are accessories, the part role semantic identifier is a preset fixed quantity. Also included is an association-level feature set indexed by part pair identifiers (constructed by combining the globally unique identifiers of two connected parts) (containing connection type features, part pair contact feature vectors, and part pair pose features).

[0090] Furthermore, the process after S1 and before S2 includes: constructing a weld semantic-spatial-topology fusion graph, including nodes and connecting edges, based on part-level feature sets and association-level feature sets, where nodes represent parts, and connecting edges are set between pairs of parts that are determined to be connected; correspondingly, in S2, part-level feature representations are extracted based on a graph mask autoencoder. The specific construction of the weld semantic-spatial-topology fusion graph is as follows:

[0091] Using extracted standardized data sequences, a semantic-spatial-topological fusion graph capable of characterizing the complex modeling features of steel structures is constructed. This is achieved by structurally associating heterogeneous geometric data, process semantics, and spatial relationships, such as... Figure 3 As shown, the specific steps include:

[0092] Graph node construction:

[0093] Traverse the part-level feature set, using the globally unique identifier of each part as an index, and construct a graph node set. ,in This represents the total number of nodes.

[0094] For any graph node Based on the part-level feature set corresponding to the node, the node is assigned its original feature vector. ,in, This represents the number of original feature dimensions for the nodes; specifically, the following operations are performed:

[0095] (1) Convert the numerical index encoding of the part attribute features and the cross-sectional process semantic features into one-hot encoding vectors respectively, and generate graph node feature sub-vector 1 and graph node feature sub-vector 2; combine the normalized continuous numerical fields in the part geometric features into graph node feature sub-vector 3.

[0096] (2) Concatenate the graph node feature vector 1, graph node feature vector 2, and graph node feature vector 3 along the feature channel dimension to obtain the node. Node original feature vector .

[0097] Graph edge construction:

[0098] Traverse the associated feature set, using the identifier of each part pair as an index, and construct a graph edge set. , ,in, This indicates the connection to the node corresponding to the part. and The edges of the graph, Let be the total number of edges. For any graph, the number of edges is... Based on the corresponding association-level feature set, the graph edges are assigned original feature vectors. ,in, This represents the number of dimensions of the original edge features; specifically, the following operations are performed:

[0099] (1) Read the contact features of the part. If there are multiple discontinuous contact areas (i.e. multiple sets of feature records), select the record with the largest contact area and output the semantic identifier encoding and continuous numerical components of the unique contact part of the part.

[0100] (2) The connection type numerical index and the semantic identifier of the unique contact part in the part association features are respectively converted into one-hot encoded vectors to obtain graph-edge feature sub-vector 1 and graph-edge feature sub-vector 2; the unique continuous numerical components of the part pair are combined into graph-edge feature sub-vector 3; the nodes in the part geometric features are used to form graph-edge feature sub-vector 3. and The normalized centroid coordinates are used to calculate the relative Euclidean distance and relative rotation quaternion between the two, resulting in graph edge feature vector 4.

[0101] (3) Concatenate the graph edge feature vector 1, graph edge feature vector 2, graph edge feature vector 3 and graph edge feature vector 4 along the feature channel dimension to obtain the original edge feature vector. .

[0102] Matrixing of graph data:

[0103] Constructing a semantic-spatial-topology fusion graph model for welds And represent it as the original feature matrix of the nodes. Original feature matrix of edges .matrix The Behavior Nodes The original feature vector of the node Set of opposite edges China and its allies First, normalize the endpoint node index, let Press again The edges are sorted lexicographically and numbered to form an edge index; the original feature vectors of the corresponding edges are stacked according to the edge index rules to form a matrix. , among which the edge The corresponding row vector is the original feature vector of that edge. .

[0104] Part-level feature representation learning based on graph mask autoencoders, i.e., the training of graph mask autoencoders specifically includes:

[0105] A graph mask autoencoder is constructed and a self-supervised pre-training task is performed. The graph mask autoencoder includes a graph encoder and a graph decoder composed of node feature reconstruction mapping functions and edge feature reconstruction mapping functions. The graph decoder restores the features after masking, forcing the graph encoder to learn and represent the geometric attributes, spatial assembly relationships, and process semantics of the parts, thereby mapping the fused graph data into a part-level vector representation in a high-dimensional feature space. Figure 4 As shown, the specific steps include:

[0106] Construction of pre-training sample set:

[0107] Multiple BIM design models were selected to construct weld semantic-spatial-topology fusion graphs, thus establishing a pre-training dataset containing these graphs. For example, approximately 500 steel structure BIM design models with different building types, component forms, and node complexities were selected as the original pre-training samples. These 500 original samples were processed according to the steps described above to construct the corresponding set of weld semantic-spatial-topology fusion graphs, which served as the unlabeled pre-training dataset for the graph mask autoencoder.

[0108] Feature random masking processing:

[0109] During the input phase, the semantic-spatial-topological fusion graph model of the weld is... The original feature matrix of the nodes With the original feature matrix of the edges Random masking is performed to construct a prediction context for self-supervised learning. The specific process is as follows:

[0110] (1) The random sampling ratio is The node row indexes constitute the node mask index set. The random sampling ratio is The edge row indices constitute the edge mask index set. And according to the edge indexing rules in step S23, Mapped to the set of masked edges ;

[0111] (2) For those belonging to the set The node, its original feature vector Some or all of the feature dimensions are replaced with preset placeholder values;

[0112] (3) For those belonging to the set The edges, and their original feature vectors. Some or all of the feature dimensions are replaced with preset placeholder values.

[0113] The mask feature matrix generated by the masking process and This will serve as the sole input to the graph encoder during the pre-training phase, forcing the model to infer lost features through unmasked neighborhood information.

[0114] Construction of a graph encoder guided by edge features:

[0115] Constructing a graph encoder for semantic-spatial-topological fusion graph model of weld seams The graph encoder performs multi-layer message passing and feature aggregation. The layers are stacked to form a graph neural network with an attention mechanism, used to create a mask feature matrix output from S31. and The process for extracting high-dimensional representations is as follows:

[0116] (1) Input layer initialization:

[0117] No. Nodes in the layer graph encoder The node representation vector is denoted as Initially ;in, Represents a node The original features of the nodes after masking.

[0118] (2) Feature aggregation and state update:

[0119] Define nodes The set of adjacent nodes is In the layer to the first The state update formula for the node representation vector of a layer is defined as follows:

[0120]

[0121] in, , For the graph encoder The trainable weight matrix of the layer; It is a non-linear activation function; In the first Layer nodes For nodes Attention weight coefficients.

[0122] The attention weight coefficient Used to characterize different connected edges The relative importance of nodes in the feature aggregation process is determined by the node representation vector. and And the original feature vector of the corresponding edge after masking. Jointly determined, and at the node Normalization is performed within the neighborhood of .

[0123] As an optional implementation, the attention weight coefficients can be obtained using an attention calculation mechanism based on node representations and edge features.

[0124] (3) Output of hidden layer node embedding matrix and edge embedding vector.

[0125] go through Layered message passing and feature aggregation yield the node embedding matrix. :

[0126]

[0127] in, The Behavior Nodes The final node embedding vector; The dimension of the vector embedded in the final node.

[0128] Simultaneously, the edge is obtained by combining the embedding vectors of the two endpoints with the original feature vectors of the corresponding edges after masking. edge embedding vector :

[0129]

[0130] in, It is a multilayer perceptron.

[0131] Mask feature reconstruction task:

[0132] Using the final node embedding vectors generated by the graph encoder, numerical restoration is performed on the masked node and edge features. Self-supervised pre-training is achieved by minimizing the reconstruction residual. The specific process is as follows:

[0133] (1) Node reconstruction:

[0134] Node embedding matrix As input, a node feature reconstruction mapping function composed of a multilayer perceptron structure is used to perform a nonlinear mapping on each node embedding vector, restoring it to the original node feature dimension, thus obtaining the node reconstructed feature vector. Arrange the reconstructed feature vectors of all nodes according to their node indices to obtain the node feature reconstruction matrix. .

[0135] Node feature reconstruction loss Defined as:

[0136]

[0137] (2) Side reconstruction:

[0138] Embedded vectors by edges As input, the edge feature reconstruction mapping function, composed of a multilayer perceptron structure, performs a nonlinear mapping on the embedding vector of each node, restoring it to the original edge feature dimension, thus obtaining the edge reconstruction feature vector. Arrange the reconstructed feature vectors of all edges according to their edge indices to obtain the edge feature reconstruction matrix. .

[0139] Edge feature reconstruction loss Defined as:

[0140]

[0141] (4) Calculation of reconstruction error:

[0142] Feature Reconstruction Loss Defined as:

[0143]

[0144] in, This is the balance coefficient. Specifically, for the original node features and edge features of samples in the pre-training dataset, random feature masking is performed separately. A graph mask autoencoder is then used to extract features from the randomly masked node and edge features. A node feature reconstruction mapping function is constructed to reconstruct node features based on the extracted node features and obtain the node feature reconstruction loss. Similarly, an edge feature reconstruction mapping function is constructed to reconstruct edge features based on the extracted edge features and obtain the edge feature reconstruction loss. The overall loss during graph mask autoencoder training includes both node feature reconstruction loss and edge feature reconstruction loss.

[0145] Furthermore, the training of the graph mask autoencoder also includes:

[0146] A node-level process role classification head is constructed to identify the functional role of nodes in parts based on extracted node features, and a loss for role prediction is built based on the semantic identifiers of part roles in the part-level feature set. An edge-level connection pattern classification head is constructed to predict the connection type and contact semantic identifiers of edges based on extracted edge features, and a loss for edge-level connection pattern prediction is built based on the connection type and contact semantic identifiers in the association-level feature set. The comprehensive loss of the graph mask autoencoder training also includes the loss for role prediction and the loss for edge-level connection pattern prediction. (Process semantic auxiliary classification task)

[0147] To enhance the graph encoder's ability to discriminate component manufacturing process patterns and inter-part connection logic, extracted features are used as weak supervision signals to construct node-level process role prediction tasks and edge-level connection pattern prediction tasks. The specific process is as follows:

[0148] (1) Node-level process role prediction task:

[0149] The final node embedding vector output by the graph encoder As input, the functional roles of parts at nodes are predicted through a node-level process role classification head composed of a multilayer perceptron structure.

[0150] Extract the one-hot encoded components corresponding to the functional roles of parts from graph node feature sub-vectors 2 or part role semantic identifiers as true labels; the loss of the node-level process role prediction task. Cross-entropy loss is used.

[0151] (2) Edge-level connectivity pattern prediction task:

[0152] Embedded vectors by edges As input, the connection type and contact point identifier of the edge are predicted by an edge-level connection pattern classification head composed of a multilayer perceptron structure.

[0153] The generated graph edge feature vector 1 and graph edge feature vector 2 are used as the true labels for connection type and contact location, respectively; the loss of the edge-level connection pattern prediction task. Defined as connection type prediction loss Predicted loss at contact points sum:

[0154]

[0155] in, and Both use cross-entropy loss; This is the task balance coefficient.

[0156] Joint optimization of overall loss function and parameters:

[0157] By combining the various losses generated in the preceding steps, an overall pre-training loss function is constructed. Parameters are optimized through multi-task joint optimization.

[0158]

[0159] in, , , , This is the weighting balance coefficient; The neighborhood consistency loss is used to constrain the node representation vectors of connected nodes. and In the feature space, they tend to be similar to each other to improve the smoothness of the feature representation.

[0160] The graph encoder is iteratively trained using the gradient descent algorithm until the overall pre-training loss function is reached. Convergence yields a graph encoder capable of extracting part-level fused feature vectors.

[0161] The specific predictions for the primary / secondary weld specifications in S2 are as follows:

[0162] A pre-trained graph encoder is used to extract part-level feature representations, which are then used as a unified feature input to construct node-level primary weld prediction functions and edge-level secondary weld prediction functions, respectively. Supervised parallel fine-tuning is performed on manually labeled sample sets to achieve prediction of primary and secondary weld specification parameters, such as... Figure 5 As shown, the specific steps include: filtering parts whose components belong to the welded assembly type based on their attribute features; using the part-level feature set corresponding to the filtered parts as input to a pre-trained encoder; and inputting the part feature representation output by the encoder into a primary weld prediction function to predict primary weld specification parameters. For part pairs that are determined to have connections, filtering out part pairs with secondary welds; using the association-level feature set of the filtered part pairs as input to a pre-trained encoder; and inputting the association feature representation output by the encoder into a secondary weld prediction function to predict secondary weld specification parameters. Specifically, it includes the following parts:

[0163] (1) Data preparation and feature extraction:

[0164] One hundred steel structure BIM design models representing different building types (such as industrial plants and high-rise buildings) were selected as original samples. These samples covered various composite steel sections and complex component connection nodes. Further, the 100 original samples were processed according to the steps described above to construct corresponding weld semantic-spatial-topology fusion diagrams.

[0165] Furthermore, the generated fused graph is input into a pre-trained graph encoder, which maps the original attributes of each node in the graph into a final node embedding vector containing global context information, and simultaneously outputs edge embedding vectors representing the connection relationships.

[0166] (2) Prediction of primary weld specifications at the node level:

[0167] Select nodes that are classified as welded composite steel sections. and its corresponding final node embedding vector Eliminate individual parts that do not require internal assembly.

[0168] Furthermore, the selected final node embedding vectors are then... The input is fed into the node-level primary weld prediction function, and the output is the primary weld specification parameters; the node-level primary weld prediction function is composed of a multilayer perceptron structure.

[0169] The primary weld specification parameters include: primary weld existence identifier. And primary weld process parameters, including the coordinates of the intersection line of the primary weld group, the primary weld type, and the weld leg size parameters. .

[0170] During the training phase of the primary weld prediction function, the primary weld existence prediction loss is constructed by combining the primary weld existence identifier with the manually labeled true labels of primary weld specification parameters. Defined as cross-entropy loss; the prediction loss of primary weld process parameters is constructed based on the predicted and labeled values ​​of the primary weld process parameters. Mean squared error loss is used, according to preset weighting coefficients. By performing a weighted sum of the two losses, the total loss for a single welding task can be obtained. :

[0171]

[0172] With the overall loss as the objective, the gradient descent algorithm is used to iteratively update only the trainable parameters of the node-level first-order weld prediction function, freezing the graph encoder parameters during the fine-tuning process until the overall loss of the first-order weld task is achieved. convergence.

[0173] (3) Prediction of secondary weld specifications at the edge level:

[0174] Based on the generated cross-sectional process semantic features and part pair contact features, edges that meet the preset welding conditions are selected, and their corresponding edge embedding vectors are obtained. The process involves filtering component pairs that are determined to have connections, specifically filtering for component pairs with secondary welds. This includes removing component pairs with primary welds and removing component pairs with contact area overlap rates less than a preset overlap rate threshold, thereby identifying component pairs with secondary welds. The details are as follows:

[0175] Based on the overlap rate of the contact area, edges with substantial physical contact are retained;

[0176] By using the identified functional roles of parts and structural types of components, the connection edges of the internal assembly welds (primary welds) of welded composite steel sections are eliminated, and only the parts pairs corresponding to secondary welds, such as steel section accessory connections and inter-component connections, are retained.

[0177] Furthermore, the selected edges are embedded into the vector. The input is fed into the edge-level secondary weld prediction function, which outputs secondary weld specification parameters. The edge-level secondary weld prediction function is constructed using a multilayer perceptron structure. The secondary weld specification parameters include: a secondary weld existence identifier. And secondary weld process parameters, including the boundary trajectory of the secondary weld connection interface, the type of secondary weld, and the weld leg size parameters. .

[0178] During the training phase of the secondary weld prediction function, the existence prediction loss for secondary welds is constructed by combining the existence identifier of secondary welds with the actual labels of the secondary weld specification parameters manually annotated. Defined as cross-entropy loss; the secondary weld process parameter prediction loss is constructed based on the predicted and labeled values ​​of the secondary weld process parameters. Mean squared error loss is used, according to preset weighting coefficients. The combined loss of the two losses is obtained by weighted summation. :

[0179]

[0180] With the overall loss as the objective, a gradient descent algorithm is used to iteratively update only the trainable parameters of the node-level secondary weld prediction function, freezing the graph encoder parameters during fine-tuning until the overall loss of the secondary weld task is achieved. convergence.

[0181] The detailed generation of the BIM welding model for the steel structure is as follows:

[0182] For the input steel structure BIM design model to be further developed, the corresponding part-level fusion feature vectors are first extracted according to the above steps; then, the feature vectors are input into the trained primary / secondary weld prediction function for forward inference to obtain the primary / secondary weld specification parameters.

[0183] Furthermore, a confidence threshold is set for the output primary weld existence identifier and the output secondary weld existence identifier. All prediction results are iterated through, and only weld instances with an existence identifier greater than the threshold are retained, while false results determined by the model inference as requiring no welding are discarded. The existence identifier can be 1 to indicate substantial existence or 0 to indicate non-substantial existence.

[0184] Further, for the selected valid instances, their corresponding primary or secondary weld process parameters are analyzed to drive the generation of 3D geometric entities:

[0185] (1) Generation of primary weld: The coordinates of the intersection line of the primary weld group are parsed from the primary weld process parameters as the path, and the primary weld type and weld leg size parameters are parsed as the cross-sectional profile; sweep modeling is performed on the cross-sectional profile along the path to generate the primary weld geometric entity inside the component.

[0186] (2) Secondary weld generation: The secondary weld connection interface boundary trajectory is parsed from the secondary weld process parameters as a path, and the secondary weld type and weld leg size parameters are parsed as a cross-sectional profile; sweep modeling is performed on the cross-sectional profile along the trajectory to generate the secondary weld geometric entity between components or accessories.

[0187] Furthermore, the numerical information from the primary and secondary weld process parameters used to generate the geometric entities is directly loaded into the property set of the corresponding weld geometry. Simultaneously, the association between the weld entity and its corresponding part is established, and the output steel structure BIM welding detail model is used to guide subsequent production.

[0188] Furthermore, a BIM-based automatic generation system for detailed welding models of steel structures is also provided. The system includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to perform the BIM-based automatic generation method for detailed welding models of steel structures described above.

[0189] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for automatically generating detailed welding models of steel structures based on BIM, characterized in that, include: S1, based on the BIM design model, perform part-level feature extraction and process semantic parsing, construct part-level feature sets for parts, and construct association-level feature sets for parts that are determined to be connected; The part-level feature set includes part geometric features, part attribute features, and cross-sectional process semantic features. The cross-sectional process semantic features include cross-sectional type, assembly order of parts in the component to which the part belongs, and standard welding joint form of each weld in the component. The associated feature set includes connection type, part pair contact features and part pair pose features. The part pair contact features include contact area overlap rate, connection angle and semantic identifier of contact part. S2: The pre-trained encoder is used to extract feature representations of the part-level feature set and the association-level feature set. The feature representations are used as feature inputs to construct the primary weld prediction function and the secondary weld prediction function respectively. Supervised learning is performed on the labeled sample set to predict the primary weld specification parameters and the secondary weld specification parameters, thereby generating a welding detail model. The part after S1 and the part before S2 includes: constructing a weld semantic-spatial-topology fusion graph including nodes and connecting edges based on the part-level feature set and the association-level feature set; In this context, nodes represent parts, and connecting edges are set between pairs of parts that are determined to be connected. For any graph node, the node is assigned an original feature vector based on the part-level feature set corresponding to the node. For any connecting edge, the connecting edge is assigned an original feature vector based on the corresponding association-level feature set. Correspondingly, in S2, part-level feature representation is extracted based on graph mask autoencoder.

2. The method for automatically generating detailed welding models of steel structures based on BIM as described in claim 1, characterized in that, The geometric features of a part include part position parameters, part orientation parameters, part dimensional parameters, and spatial enclosure parameters; the attribute features of a part include part type, material designation, globally unique identifier of the part, identifier field of the component to which the part belongs, and component cross-section description string associated with the component to which the part belongs. The component cross-section description string is used to represent the construction type and cross-section type of the corresponding component. The structural types include welded assembly types and steel section types, and the cross-section type is the specific shape type under the structural type; The specific part pairs that are connected include: display part pairs, which are obtained from the BIM design model and are displayed part pairs; And candidate part pairs are determined by the minimum spacing of the bounding boxes of the parts. When the minimum spacing of the bounding boxes is less than a preset spacing threshold, the corresponding part pair is determined to be a candidate part pair.

3. The method for automatically generating detailed welding models of steel structures based on BIM as described in claim 1, characterized in that, The construction of the cross-sectional process semantic features is as follows: When determining that a part belongs to a welded assembly type, the cross-section type is the cross-section type of the component. Based on the preset shape and specification parameters of each functional role within the component, the parts within the component are traversed to identify the functional role of each part in the cross-section composition as the basic role of the part. Thus, the assembly order of each part is determined based on the basic role of the part, and the standard welded joint form of the corresponding weld is determined based on the connection type between connected parts. When determining that a component belongs to a steel section type, the cross-section type is the steel section type; the assembly sequence and standard welding joint form are both preset fixed quantities; When determining that a part is an accessory, the cross-section type is obtained from the BIM design model; the assembly sequence and standard welded joint form are both preset fixed quantities.

4. The method for automatically generating detailed welding models of steel structures based on BIM as described in claim 3, characterized in that, The part-level feature set also includes part role semantic identifiers; for parts whose components are welded assembly type, the part role semantic identifier is the basic part role; for parts whose components are structural steel type and parts that are accessory type, the part role semantic identifier is a preset fixed quantity. Correspondingly, for parts whose components are steel profiles, the equivalent functional roles of the steel profile geometric surfaces are identified based on the shape specifications of the preset equivalent functional roles. Then, using the equivalent functional roles as spatial reference benchmarks, the auxiliary functional roles of other accessories are identified based on the shape specifications of the preset auxiliary functional roles. Furthermore, based on the basic role, equivalent functional role, and auxiliary functional role of the part, a semantic identifier for the contact part is constructed from the associated feature set.

5. The method for automatically generating detailed welding models of steel structures based on BIM as described in claim 4, characterized in that, The training of a graph mask autoencoder specifically includes: Multiple BIM design models were selected to construct weld semantic-spatial-topology fusion maps, and a pre-training dataset including multiple weld semantic-spatial-topology fusion maps was established. For the original node features and original edge features of the samples in the pre-training dataset, random masking is performed on the features. The graph mask autoencoder is used to extract the features of the node mask features and edge mask features after random masking. A node feature reconstruction mapping function is constructed to reconstruct the node features based on the extracted node features and obtain the node feature reconstruction loss. An edge feature reconstruction mapping function is constructed to reconstruct the edge features based on the extracted edge features and obtain the edge feature reconstruction loss. The combined loss for training a graph mask autoencoder includes node feature reconstruction loss and edge feature reconstruction loss.

6. The method for automatically generating detailed welding models of steel structures based on BIM as described in claim 5, characterized in that, Training the graph mask autoencoder also includes: A node-level process role classification head is constructed to identify the functional role of nodes based on extracted node features, and a loss for role prediction task is constructed based on the semantic identifiers of part roles in the part-level feature set. An edge-level connection pattern classification head is constructed to predict the connection type and contact semantic identifiers of edges based on extracted edge features, and a loss for edge-level connection pattern prediction task is constructed based on the connection type and contact semantic identifiers in the association-level feature set. The overall loss for training a graph mask autoencoder also includes the loss for the role prediction task and the loss for the edge connection pattern prediction task.

7. The method for automatically generating detailed welding models of steel structures based on BIM as described in claim 1, characterized in that, S2 specifically includes: Based on the component attribute characteristics, the components belonging to the welded assembly type are selected. The component-level feature set corresponding to the selected components is used as the input of the pre-trained encoder. The component feature representation output by the encoder is input into the weld prediction function to predict the weld specification parameters. The parts pairs that are determined to be connected are filtered to find the parts pairs with secondary welds. The correlation-level feature set of the selected parts pairs is used as the input of the pre-trained encoder. The correlation feature representation output by the encoder is input into the secondary weld prediction function to predict the secondary weld specification parameters.

8. The method for automatically generating detailed welding models of steel structures based on BIM as described in claim 1, characterized in that, The specifications of a primary weld include: the existence identifier of the primary weld, and the process parameters of the primary weld, which include the coordinates of the intersection line of the primary weld group, the type of primary weld, and the weld leg size parameters. In the training phase of the primary weld prediction function, the primary weld existence prediction loss is constructed by combining the primary weld existence identifier and the actual label of the primary weld specification parameters. The primary weld process parameter prediction loss is constructed based on the predicted and labeled values ​​of the primary weld process parameters. The two losses are weighted and summed according to the preset weight coefficients to obtain the comprehensive loss for training the primary weld prediction function. The secondary weld specifications include: the secondary weld existence identifier, and the secondary weld process parameters including the secondary weld connection interface boundary trajectory, secondary weld type and weld leg size parameters; During the training phase of the secondary weld prediction function, the existence prediction loss of the secondary weld is constructed by combining the existence identifier and the actual label of the secondary weld specification parameters. The prediction loss of the secondary weld process parameters is constructed based on the predicted and labeled values ​​of the secondary weld process parameters. The two losses are weighted and summed according to the preset weight coefficients to obtain the comprehensive loss for training the secondary weld prediction function.

9. A BIM-based automatic generation system for detailed welding models of steel structures, characterized in that, The system includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it performs the automatic generation method for BIM-based steel structure welding detail model according to any one of claims 1-8.

Images