A building assembly type decoration method based on BIM

Abstract

The application discloses a kind of based on BIM's building assembly type decoration method, comprising the following steps: S1, extract the decoration space structure information in building BIM model, and demarcate component arrangement unit and demarcate space limiting object;S2, check assembly constraint relationship and bind as component relationship diagram;S3, access improved Graphormer model;S4, edge relationship Token conversion layer truncation edge encoding is directly written into the path of attention bias, forms node Token and edge encoding fragment;S5, convert edge encoding fragment into edge relationship Token;S6, rewrite Transform encoder propagation path by node edge association matrix;S7, read out optimized component arrangement scheme and update BIM model.The application realizes component arrangement and construction data collaborative generation.

Description

Technical Field

[0001] This invention relates to the field of prefabricated building decoration technology, and in particular to a BIM-based prefabricated building decoration method. Background Technology

[0002] Existing prefabricated building decoration methods typically use BIM models as the design basis. By extracting building space dimensions, decoration area boundaries, and component specification information, the method completes the disassembly and arrangement of wall panels, floors, ceilings, keels, and connectors. It also combines collision detection to check pipelines, openings, equipment interfaces, and maintenance spaces, thereby outputting component lists, installation drawings, and construction data.

[0003] However, existing methods often use component relationships as external rules for verification. They typically generate the component layout first and then adjust conflicting positions, making it difficult to simultaneously handle the connections, avoidance, and installation relationships between components during the component layout calculation. In prefabricated decoration scenarios, the fact that component dimensions can be arranged into the decoration area does not mean that the connection relationship is valid, the space constraints can be avoided, or the installation sequence can be executed.

[0004] Furthermore, existing processing methods based on graph models or deep learning models often treat the edges between nodes as static relationships or attention biases in the calculation. The edge relationships themselves are difficult to propagate and be read out as independent objects within the model, resulting in a lack of deep coupling between the component layout results and the assembly constraint relationships, which affects the consistency of component layout optimization, construction data generation, and BIM model updates.

[0005] Therefore, how to provide a BIM-based prefabricated building decoration method is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a BIM-based prefabricated building decoration method. This invention combines BIM spatial objectification with an improved Grapher model. Through decoration area encapsulation, component layout unit segmentation, and spatial constraint object calibration, a component relationship graph capable of supporting component layout relationships is constructed. Furthermore, an edge relationship token conversion layer is embedded between the edge encoding and Transformer encoder in the improved Grapher model, transforming the assembly constraint relationships carried by the edges between nodes from static biased paths with attention bias into edge relationship tokens that can participate in model propagation. Based on this, the node-to-node direct propagation path within the Transformer encoder is rewritten into a node-to-edge-to-node propagation path through the node-edge association matrix, achieving collaborative optimization of component layout units and assembly constraint relationships. Finally, an optimized component layout scheme is formed, and component processing and construction data are extracted to update the building BIM model. This invention reduces the problem of disconnect between component layout and assembly constraints, improves the consistency of component splitting, spatial avoidance, installation acceptance, and construction data generation, and has the advantages of more adaptable layout results to prefabricated construction, stronger construction data connectivity, and more complete BIM model updates.

[0007] A BIM-based prefabricated building decoration method according to an embodiment of the present invention includes the following steps: S1. Extract the interior space construction information from the building BIM model, encapsulate the interior area and match it with the component parametric model library in terms of specifications, cut the component layout unit and mark the space constraint object. S2. Check the assembly constraint relationship along the component layout unit and spatial constraint object, organize the component layout unit and spatial constraint object into nodes, connect the assembly constraint relationship into the edge between nodes, and bind them into a component relationship diagram; S3. Connect the component relationship graph to the improved Graphormer model, wherein the improved Graphormer model embeds an edge relationship token conversion layer between edge encoding and Transformer encoder; S4. The edge relationship token conversion layer cuts off the direct write path from edge encoding to attention bias, organizes the nodes in the component relationship graph into node tokens, and organizes the assembly constraint relationship carried by the edges between nodes into edge encoding fragments. S5. The edge relationship token conversion layer converts edge encoded fragments into edge relationship tokens, and attaches the edge relationship tokens to the corresponding node tokens according to the endpoints of the edges between nodes in the component relationship graph. S6. Arrange the node-edge association matrix along the interlocking positions of the edge relationship token and the node token. The node-edge association matrix rewrites the direct propagation path from node to node in the Transformer encoder into a propagation path from node to edge and then to node, and updates the node token and edge relationship token along the rewritten propagation path. S7. Read out the updated node token and edge relationship token, merge them into the optimized component layout scheme, extract the component processing and construction data according to the optimized component layout scheme, and update the building BIM model.

[0008] Optionally, S1 specifically includes: S11. Extract interior space construction information from the building BIM model and delineate the boundaries of the interior surface area in the interior space construction information. S12. Close the disconnected decoration area boundary, remove the boundary area that does not participate in the component layout, and encapsulate it as a continuous decoration area. S13. Project the continuous decoration area onto the outer contour and edge boundary of the component in the component parametric model library, and cut the component layout unit along the projected component contour. S14. Check the occlusion of the space content that penetrates the continuous decoration area and occupies the component installation position. Mark the space content that is checked as a space restriction object. Merge the component layout unit and the space restriction object into the drawing object of the component relationship diagram.

[0009] Optionally, S2 specifically includes: S21. Assign aggregated component layout units and spatially restricted objects according to the decoration area to form candidate node groups on the same surface; S22. Compare the contact boundaries, fixed positions and installation directions between component arrangement units within the same candidate node group to screen out the assembly constraint relationships between component nodes. S23. Compare the occupancy positions and avoidance boundaries between component layout units and spatial constraint objects within the same candidate node group, and screen out the assembly constraint relationships between component nodes and constraint nodes. S24. Set the component layout units and spatial constraint objects to nodes, and snap the screened assembly constraint relationships into the edges between nodes to assemble a component relationship diagram.

[0010] Optionally, the improved Graphormer model specifically includes node embedding, centrality encoding, spatial encoding, edge encoding, edge relation token transformation layer, Transformer encoder, and readout layer.

[0011] Optionally, S4 specifically includes: S41. After the component relationship graph is connected to the improved Graphormer model, the node embedding receives the nodes in the component relationship graph, pushes the corresponding nodes of the component layout unit and the corresponding nodes of the space constraint object into the same node embedding sequence, and organizes them into the node initial token. S42, centrality encoding receives the number of node connections within the component relationship graph and forms a centrality bias; spatial encoding receives the graph distance between nodes within the component relationship graph and forms a spatial bias; centrality bias and spatial bias are retained within the attention bias write path. S43, edge encoding receives the edges between nodes in the component relationship diagram, performs edge encoding on the assembly constraint relationship carried by the edges between nodes, and organizes the endpoints and assembly constraint relationships corresponding to the edges between nodes into the same edge encoding sequence to form an edge encoding fragment; S44. The edge relationship token transformation layer is embedded between the edge encoding and the Transformer encoder, cutting off the path from the edge encoding to the attention bias to directly write the edge encoding fragment. The edge encoding fragment is separated from the attention bias writing path and transferred to the entry point of the edge relationship token transformation layer. S45. When the edge relationship token conversion layer receives the edge encoded fragment, it simultaneously carries the node endpoint attribution and assembly constraint relationship between the corresponding nodes of the edge encoded fragment. Edge encoded fragments whose endpoint attribution is not fully hit are retained in the attention bias write path. S46. The edge relationship token conversion layer organizes the initial token of the node into a node token and pushes the complete edge encoding fragment of the endpoint into the edge relationship token conversion layer entry.

[0012] Optionally, S5 specifically includes: S51. The edge relationship token conversion layer reads the edge encoding fragment of the corresponding node pair bias table in the edge encoding, and extracts the assembly constraint edge sequence under the same node pair bias table item from the attention bias writing content. S52. The edge relationship Token conversion layer rearranges the edge encoding fragments based on the decoration region affiliation, the token positions of the two end nodes, and the assembly constraint direction associated with the assembly constraint edge sequence, forming an edge relationship to be encoded sequence with continuous arrangement of region affiliation, endpoint positions, and constraint directions. S53. The edge relationship Token transformation layer writes the edge relationship sequence to be encoded into the edge relationship Token construction bit of the Transformer encoder front end, and compresses it into an edge relationship Token that carries the decoration area affiliation, the Token positions of the two end nodes and the assembly constraint direction. S54. The edge relationship token transformation layer locates the two end node tokens corresponding to the edge relationship token in the node token sequence, embeds the edge relationship token into the propagation chain between the two end node tokens, and replaces the static bias connection of the original node pair bias table entry to the two end node tokens. S55. When there are multiple assembly constraint edge sequences under the same node pair bias table entry, the edge relation token conversion layer continuously places multiple edge relation tokens in the same propagation chain according to the edge sequence arrangement order in edge encoding. S56. The edge relationship token transformation layer outputs the edge relationship tokens embedded in the propagation chain, and arranges the edge relationship tokens embedded in the propagation chain and the corresponding node tokens together as the Transformer encoder front-end token sequence.

[0013] Optionally, S6 specifically includes: S61. Arrange the node tokens and edge relationship tokens into a Transformer encoder front-end token sequence, and establish the row and column positions of the node edge association matrix based on the token positions in the Transformer encoder front-end token sequence. S62. The position of the edge relationship token in the front-end token sequence of the Transformer encoder, respectively locating the start node token and the end node token corresponding to the edge relationship token; S63. In the node-edge association matrix, open the row and column intersection positions from the starting node Token to the edge relationship Token and from the edge relationship Token to the ending node Token, and hide the direct row and column intersection positions from the starting node Token to the ending node Token. S64. When the Transformer encoder calculates self-attention, it retains the attention scores corresponding to nodes to edges and edges to nodes according to the node-edge association matrix, and suppresses the attention scores directly corresponding to nodes. S65. During the value vector aggregation process, the Transformer encoder passes node tokens and edge relation tokens along the retained attention scores, and the assembly constraint relationships between node tokens are transferred through edge relation tokens. S66. The Transformer encoder updates the node token and edge relationship token along the value vector aggregation result, forming the node token and edge relationship token constrained by the node-edge association matrix.

[0014] Optionally, S7 specifically includes: S71. Read the updated node token and organize the component layout unit into component size, layout position and installation direction according to the node token's position in the decoration area. S72. Read out the updated edge relationship token, and according to the connection position of the edge relationship token in the node edge association matrix, rewrite the assembly constraint relationship into the constraint placement result; S73. Merge the component layout results and constraint placement results into a surface region to form an optimized component layout scheme. S74. Extract the component processing and construction data according to the optimized component layout scheme, and write the component layout results and component processing and construction data back to the building BIM model.

[0015] The beneficial effects of this invention are: This invention extracts interior space construction information from a building BIM model and organizes interior areas, component layout units, and spatial constraints into a component relationship diagram. This unifies the component layout objects and assembly constraints in prefabricated interior decoration into a calculable structure. Compared to the traditional method of arranging first and then verifying, this invention establishes the load-bearing structure between components and spatial constraints before component layout calculation, which helps reduce the disconnect between component size layout, spatial avoidance, and installation connection.

[0016] This invention introduces an improved Graphormer model, embedding an edge relation token transformation layer between the edge encoding and the Transformer encoder. This transforms the assembly constraints carried by the edges between nodes from the static bias path of attention bias into edge relation tokens. Furthermore, it rewrites the direct node-to-node propagation path within the Transformer encoder into a node-to-edge and then-to-node propagation path through the node-edge association matrix. Through this processing, component relationships are no longer merely secondary information affecting attention scores, but rather intermediate representations that can participate in internal model propagation and readout, thereby improving the coupling between component layout units and assembly constraints.

[0017] This invention can merge updated node tokens and edge relationship tokens to form an optimized component layout scheme, and further extract component processing and construction data to update the building BIM model. This method improves the consistency between component splitting, spatial avoidance, installation sequence, and construction data generation, reduces the workload of subsequent manual adjustments, and enhances the connection between prefabricated decoration design results and on-site installation processes. It has the advantages of more reasonable layout results, more coherent construction data generation, and more complete BIM model updates. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a BIM-based prefabricated building decoration method proposed in this invention; Figure 2 This is a schematic diagram of the improved Graphormer model for a BIM-based prefabricated building decoration method proposed in this invention. Figure 3 This is a schematic diagram of the structure of the edge relation Token conversion layer rewriting the Graphormer propagation path in the BIM-based prefabricated building decoration method proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-3 A BIM-based prefabricated building decoration method includes the following steps: S1. Extract the interior space construction information from the building BIM model, encapsulate the interior area and match it with the component parametric model library in terms of specifications, cut the component layout unit and mark the space constraint object. S2. Check the assembly constraint relationship along the component layout unit and spatial constraint object, organize the component layout unit and spatial constraint object into nodes, connect the assembly constraint relationship into the edge between nodes, and bind them into a component relationship diagram; S3. Integrate the component relationship graph into the improved Graphormer model, and embed an edge relationship token conversion layer between edge encoding and Transformer encoder in the improved Graphormer model; S4. The edge relationship token conversion layer cuts off the direct write path from edge encoding to attention bias, organizes the nodes in the component relationship graph into node tokens, and organizes the assembly constraint relationship carried by the edges between nodes into edge encoding fragments. S5. The edge relationship token conversion layer converts edge encoded fragments into edge relationship tokens, and attaches the edge relationship tokens to the corresponding node tokens according to the endpoints of the edges between nodes in the component relationship graph. S6. Arrange the node-edge association matrix along the interlocking positions of the edge relationship token and the node token. The node-edge association matrix rewrites the direct propagation path from node to node in the Transformer encoder into a propagation path from node to edge and then to node, and updates the node token and edge relationship token along the rewritten propagation path. S7. Read out the updated node token and edge relationship token, merge them into the optimized component layout scheme, extract the component processing and construction data according to the optimized component layout scheme, and update the building BIM model.

[0021] In this embodiment, S1 specifically includes: After importing the building BIM model, the model units are standardized to millimeters, and the building surfaces, opening cut lines, main structure occlusion outlines, and MEP entity projections are registered to the same building coordinate system. The system extracts the interior decoration space construction information from the registered building BIM model. The interior decoration space construction information consists of the building surface outline enclosing the interior decoration space, opening cut lines, main structure occlusion outlines, and MEP entity projections. After extraction, the system delineates the boundary of the interior decoration area along the building surface outline. The boundary of the interior decoration area includes the outer boundary formed by the outer outline of the building surface and the inner boundary formed by the opening cut lines and the main structure occlusion outline. When a break occurs in the boundary of the interior decoration area, the break with an endpoint spacing of no more than 5 mm is closed by endpoint snapping; when the endpoint spacing is greater than 5 mm but not more than 30 mm, and the two boundary lines are located on the same building surface, they are closed by connecting the intersection of the extended boundary lines; the break with an endpoint spacing of more than 30 mm or no intersection after extension is written into the review list. The review list is not included in the current round of component layout unit segmentation. After BIM boundary correction, the interior decoration space construction information is re-extracted.

[0022] After the boundary of the decorative surface area is closed, the closed boundary forms a continuous decorative surface area. The system removes the outer surface pieces of the closed boundary. Fragment pieces with an area smaller than the minimum component installation area in the component parametric model library are merged into the adjacent edge finishing area. Fragment pieces that cannot be merged into the adjacent edge finishing area are written to the review list. After the continuous decorative surface area is formed, it enters the specification matching stage. Specification matching refers to the size adaptation between the continuous decorative surface area and the component specification boundaries in the component parametric model library. The component parametric model library saves the component outer contour, specification boundaries, minimum cutting width, allowable edge finishing width, and installation direction. The component outer contour is the projected contour of the component on the continuous decorative surface area. The specification boundaries are the long and short sides of the component outer contour that participate in the layout. The edge finishing boundaries are the position range where the edge of the continuous decorative surface area can be filled with edge finishing components. The system first aligns the long side of the component outer contour with the main laying direction of the continuous decorative surface area, and then calculates the edge allowance along the short side direction. When the edge allowance is less than the minimum cutting width, it is merged into the edge finishing boundary. When the edge allowance reaches the minimum cutting width, it is cut into a component layout unit. A component layout unit is a single processable and installable component area formed by matching and dividing a continuous decorative area according to specifications, and it is also involved in the spatial constraint object calibration process.

[0023] Spatial constraint object calibration is performed around component layout units. The system projects spatial content penetrating the continuous decoration area onto the plane where the component layout unit is located, and calculates the overlap width between the spatial content projection and the outer contour of the component layout unit. When the overlap width reaches 10 mm, it is determined that the spatial content occupies the component installation position. The system then pulls out the component installation channel along the installation direction given by the component parametric model library. When the spatial content enters the component installation channel and the clear distance is less than 20 mm, it is determined that the spatial content blocks the component installation. If either occupying the component installation position or blocking the component installation is true, the spatial content is calibrated as a spatial constraint object; if neither is true, the spatial content is only retained in the building BIM model and is not included in the current round of component relationship diagrams. Component layout units and spatial constraint objects are jointly merged into the drawing objects of the component relationship diagram. Objects that have not passed the boundary closure, specification matching, or occlusion check are written into the review list and are not included in the current round of drawing objects.

[0024] In this embodiment, S2 specifically includes: After the component layout units and spatial constraint objects are merged, the system first forms a candidate node group on the same surface. This candidate node group consists of component layout units and spatial constraint objects belonging to the same continuous decorative surface. When a component layout unit falls entirely within a continuous decorative surface, it is directly assigned to that surface. When a spatial constraint object spans multiple continuous decorative surfaces, the projected area of ​​the spatial constraint object within each surface is calculated, and the surface with the largest projected area is designated as the assigned surface. If the projected areas are the same, the surface where the spatial constraint object blocks the component insertion channel is designated as the assigned surface. After grouping, the component layout units become component nodes, and the spatial constraint objects become constraint nodes. Both component nodes and constraint nodes retain their respective continuous decorative surface, outline range, and installation direction, serving as the basis for filtering assembly constraint relationships.

[0025] Assembly constraints between component nodes are screened out from candidate node groups on the same surface. Screening means that candidate relationships are retained as sources of edges between nodes after being checked for distance, overlap, and installation direction. The system first calculates the shortest distance between the outline boundaries of two component layout units. When the shortest distance does not exceed 6 mm, a contact candidate is formed. When the outlines of two component layout units overlap, a contact candidate is still formed if the overlap width does not exceed the allowable joint width in the component parametric model library. If the overlap width exceeds the allowable joint width, the corresponding component layout unit returns to the specification matching and cutting process of the continuous decoration surface for reprocessing. After the contact candidate is formed, the system checks whether there are installable connection areas on both sides of the contact candidate and checks whether the installation channel of one component layout unit passes through the occupied area of ​​another component layout unit along the component installation direction. When installable connection areas exist and the installation channels do not block each other, the contact candidate is retained as an assembly constraint relationship between component nodes. When the installation channel is blocked, the blocking reason is written to the corresponding component node, and the contact candidate does not fall into the edge between nodes.

[0026] The assembly constraints between component nodes and constraint nodes are also filtered out from the candidate node group on the same surface. The system calculates the overlap width between the projection of the spatial constraint object and the outer contour of the component layout unit. When the overlap width reaches 10 mm, a constraint candidate is formed. After the constraint candidate is formed, the system checks whether there are any accommodating positions around the component layout unit. Accommodating positions are the boundary positions that still meet the minimum trimming width after the component contour is moved, opened, or trimmed. When accommodating positions exist, the constraint candidate is retained as the assembly constraint relationship between the component node and the constraint node; when no accommodating positions exist, the corresponding component layout unit returns to the specification matching and trimming process of the continuous decoration surface region for reprocessing. The retained assembly constraints are set as nodes-to-node edges, and the nodes-to-node edges are written with the node affiliation and assembly constraint relationship at both ends. Component nodes, constraint nodes, and nodes-to-node edges are assembled into a component relationship graph according to the continuous decoration surface region. The component relationship graph is then processed by the edge encoding of the improved Graphormer model.

[0027] In this embodiment, the improved Graphormer model specifically includes node embedding, centrality encoding, spatial encoding, edge encoding, edge relationship token conversion layer, Transformer encoder, and readout layer.

[0028] In this embodiment, S4 specifically includes: The improved Graphormer model consists of node embedding, centrality encoding, spatial encoding, edge encoding, an edge-relation token transformation layer, a Transformer encoder, and a readout layer. In the existing Graphormer model, when processing graph data, node embedding converts graph nodes into node tokens, centrality encoding converts the number of node connections into a centrality bias, spatial encoding converts the graph distance between nodes into a spatial bias, and edge encoding converts the edges between nodes into edge codes. The centrality bias, spatial bias, and edge codes are all incorporated into the attention bias, which then enters the Transformer encoder for self-attention computation. However, in the prefabricated decoration component relationship graph, the edges between nodes carry assembly constraints. If the edge code fragments formed by edge encoding are directly incorporated into the attention bias, the assembly constraints can only participate in the computation as node-to-attention biases; they cannot enter the Transformer encoder as independent relation objects for propagation, nor can they independently handle component layout and assembly relationship information in the readout layer. This implementation embeds an edge relationship token conversion layer between edge encoding and Transformer encoder, so that the edge encoded fragments formed by edge encoding first enter the edge relationship token conversion layer, and then proceed to the edge relationship token generation and node edge association matrix arrangement process.

[0029] The training data for the improved Graphormer model comes from the historical prefabricated decoration refinement results corresponding to the BIM model. Training samples include component relationship diagrams, manually confirmed component layout schemes, and construction verification results. Node embedding, centrality encoding, spatial encoding, edge encoding, the edge relationship token conversion layer, and linear mapping, embedding tables, and attention parameters in the Transformer encoder are used in training. Node affiliation, edge affiliation between nodes, and geometric coordinates of the decoration surface in the component relationship diagram are used as input items but not as trainable parameters. During training, the model parameter updates are jointly constrained by the consistency of component layout scheme positions, the consistency of assembly constraint relationships, and the penalty for non-assembly relationships. Position consistency is calculated by the coordinate deviation between the predicted component positions and the manually confirmed component positions; the consistency of assembly constraint relationships is calculated by the matching results between the predicted edge relationship positions and the manually confirmed relationship positions; and the penalty for non-assembly relationships is calculated by the number of erroneous relationships corresponding to occlusion, intrusion, and installation path blocking. Training stops when the loss value changes below a set convergence threshold for three consecutive training epochs. The convergence threshold is one-thousandth of the loss value in the previous epoch. During inference, the improved Graphormer model calls the trained parameters, and the edge relation token transformation layer performs edge encoding fragment takeover and endpoint attribution verification along the trained linear mapping and embedding parameters.

[0030] After the component relationship graph is integrated into the improved Graphormer model, the node embedding receives nodes from within the graph. These nodes originate from component layout units and spatially constrained objects. The node input vector consists of the node source, the decoration region affiliation, the local 2D coordinate range, and the installation direction encoding. The node source distinguishes between nodes corresponding to component layout units and nodes corresponding to spatially constrained objects. The decoration region affiliation defines the continuous decoration region where the node resides. The local 2D coordinate range is obtained from the minimum bounding rectangle of the node's outer contour in the continuous decoration region coordinate system. The installation direction encoding is obtained from the insertion direction given by the component parametric model library or the occlusion direction of the spatially constrained object. The node embedding uses a trainable linear mapping to uniformly map the node input vector to the model's hidden dimension, which is set to 256. After the nodes corresponding to component layout units and spatially constrained objects are pushed into the same node embedding sequence, the node embedding outputs the initial node token. This initial token carries both the node corresponding to the component layout unit and the node corresponding to the spatially constrained object.

[0031] Centrality encoding receives the number of node connections within the component graph and forms a centrality bias. The number of node connections is calculated from the number of edges between nodes connected to the same node. When the number of node connections exceeds the maximum index of the encoding table, it is written to the maximum index position; when the number of node connections is zero, it is written to the isolated index position. Centrality encoding maps the number of node connections to the centrality bias through a trainable embedding table. Spatial encoding receives the graph distance between nodes within the component graph and forms a spatial bias. The graph distance between nodes is calculated from the number of the shortest edges between two nodes in the component graph; when there is no connected path between two nodes, it is written to the unreachable index, which forms a weak connection bias within the attention bias. The attention bias is a bias matrix between node tokens, where the row and column positions correspond to the token order in the node token sequence. The centrality bias is written to the row and column positions corresponding to the same node token, and the spatial bias is written to the row and column intersection positions corresponding to two node tokens. Both the centrality bias and the spatial bias are retained within the attention bias writing path and continue to serve as the source of structural bias when the Transformer encoder calculates self-attention.

[0032] Edge encoding receives edges between nodes in the component relationship graph and performs edge encoding on the assembly constraints carried by these edges. The endpoint affiliation of an edge is determined by the indices of the two ends connected by the edge, and the assembly constraints are determined by the content of the edges in the component relationship graph. Edge encoding first reads the indices of the two ends of the edge and then maps the assembly constraints to edge encoding vectors. The dimension of the edge encoding vector is consistent with the hidden dimension of the model, which is 256. When there are multiple edges between the same node pair, edge encoding arranges multiple edge encoding vectors according to the order in which the edges are established in the component relationship graph, forming an edge encoding sequence at the same node pair position. When there is only one edge between the same node pair, the edge encoding sequence contains only one edge encoding vector. After the endpoint affiliation and assembly constraints corresponding to the edges are organized into the same edge encoding sequence, an edge encoding fragment is formed. This edge encoding fragment serves as the receiving object of the edge relationship token conversion layer entry.

[0033] The edge relationship token transformation layer is embedded between edge encoding and the Transformer encoder, specifically located after the edge encoding fragment output of edge encoding and before the token input of the Transformer encoder. In existing Grapher, the edge encoding fragment output by edge encoding is directly written to the corresponding node pair position in the attention bias. In this implementation, a truncation position is set on the path from edge encoding to attention bias where the edge encoding directly writes the edge encoding fragment. The truncation position is located between the edge encoding output and the attention bias merging entry point. After truncation, the edge encoding fragment is separated from the attention bias writing path and transferred to the edge relationship token transformation layer entry point; the attention bias still receives the centrality bias formed by centrality encoding and the spatial bias formed by spatial encoding. Thus, edge encoding is still responsible for generating the edge encoding fragments corresponding to the edges between nodes, and the edge encoding fragments are no longer directly merged into the attention bias as node pair biases.

[0034] When the edge relationship token conversion layer receives edge-coded fragments, it simultaneously carries the endpoint attribution and assembly constraints between the corresponding nodes. Complete endpoint attribution means that both node indices corresponding to the edge-coded fragment can be matched in the initial node token sequence; incomplete endpoint attribution means that one node index is missing, both node indices are missing, or the node index cannot match the token position in the initial node token sequence. Edge-coded fragments with complete endpoint attribution are pushed into the edge relationship token conversion layer entry point and establish a convertible relationship with the corresponding initial node token; edge-coded fragments with incomplete endpoint attribution do not enter the edge relationship token conversion layer entry point and remain in the attention bias write path. The edge relationship token conversion layer simultaneously organizes the initial node tokens into node tokens, and the node tokens and edge-coded fragments with complete endpoint attribution together serve as the preceding objects in the edge relationship token generation process.

[0035] Compared to existing Graphormer models, the improvement of this invention lies in the addition of an edge relationship token transformation layer between edge encoding and Transformer encoder, and the change in the entry path of the edge encoded fragments output by edge encoding. In existing Graphormer, the edge encoded fragments formed by edge encoding are directly incorporated into the attention bias, and edge information only participates in the calculation as a node to the attention bias. This invention separates the edge encoded fragments with complete endpoint affiliation from the attention bias writing path and transfers them to the edge relationship token transformation layer entry point, while retaining the centrality bias formed by centrality encoding and the spatial bias formed by spatial encoding to continue entering the attention bias. The assembly constraint relationship is transformed from the static bias content in the attention bias into a pre-level object generated by the edge relationship token, and is carried over into the node-to-edge and then-to-node propagation path in the Transformer encoder, providing an internal relationship representation for the readout of the component layout scheme.

[0036] In this embodiment, S5 specifically includes: After receiving the edge-encoded fragments from edge encoding, the edge-relationship token conversion layer first establishes a correspondence between the edge-encoded fragments and the node-pair bias table. The node-pair bias table is a two-dimensional bias position table corresponding to the node token sequence in the attention bias. The row and column positions of the node-pair bias table correspond to the token order in the node token sequence; the row position serves as the query-side node token position, and the column position serves as the key-side node token position. The original edge encoding process involved writing the edge-encoded fragments at the intersection of the row and column positions in the node-pair bias table. After the edge-relationship token conversion layer intervenes, it first reads the truncated edge-encoded fragments in the writing path and then checks the row and column positions of the node-pair bias table along the endpoint affiliation carried by the edge-encoded fragments. When both the row and column positions match the node token sequence, the edge-encoded fragment enters the edge-relationship encoding sequence preparation process; when either the row or column position does not match the node token sequence, the edge-encoded fragment does not enter the edge-relationship token construction position and remains in the attention bias writing content.

[0037] The edge relationship encoding sequence is formed by organizing edge-encoded fragments under the same node's bias table entry. The edge relationship token transformation layer first extracts the decoration area affiliation from the edge-encoded fragments, which comes from the continuous decoration areas to which the node token belongs; then it extracts the positions of the two node tokens, which come from the row and column positions of the nodes in the bias table; finally, it extracts the assembly constraint direction, which comes from the loading and receiving direction recorded when the edge is placed between nodes. The edge relationship token transformation layer rearranges the edge-encoded fragments in the order of decoration area affiliation, two node token positions, and assembly constraint direction to form the edge relationship encoding sequence. The edge relationship encoding sequence only originates from the edge-encoded fragments output by edge encoding and does not introduce external business fields.

[0038] The edge relation token construction site is located within the front-end token sequence of the Transformer encoder, and is allocated between the two end node tokens by the edge relation token transformation layer. After the edge relation to be encoded sequence enters the edge relation token construction site, the decoration surface domain affiliation, the positions of the two end node tokens, and the assembly constraint direction are mapped to 256-dimensional vectors through a trainable embedding table. These three 256-dimensional vectors are concatenated sequentially to form a 768-dimensional concatenated vector. The 768-dimensional concatenated vector is then compressed into a 256-dimensional candidate relation vector after entering the linear mapping layer. The candidate relation vector undergoes layer normalization and nonlinear activation processing to form a 256-dimensional edge relation token. The nonlinear activation uses GELU activation. When there are multiple assembly constraint edge sequences under the same node pair bias table entry, the edge relation token transformation layer performs embedding, concatenation, linear mapping, normalization, and activation processing on each edge in the order in which edge encoding forms the edge encoded fragments. Each assembly constraint edge sequence forms an edge relation token.

[0039] After the edge relation token is formed, the edge relation token transformation layer locates the corresponding two node tokens in the node token sequence along the positions of the two end node tokens carried by the edge relation token, and embeds the edge relation token into the propagation chain between the two node tokens. The propagation chain is a local arrangement structure in the front-end token sequence of the Transformer encoder, with one node token at the front, the edge relation token in the middle, and the other node token at the end. After the edge relation token is embedded into the propagation chain, the static bias connection between the node pair bias entries and the two node tokens no longer carries the responsibility of transmitting assembly constraint relationships. When multiple edge relation tokens are formed under the same node pair bias entry, the multiple edge relation tokens are placed consecutively in the same propagation chain according to the edge encoding order in edge encoding. The edge relation tokens and node tokens after embedding are arranged together to form the front-end token sequence of the Transformer encoder, and are then used in the process of establishing the node edge association matrix.

[0040] Compared to existing Graphormer models, the improvement of the edge relation token transformation layer lies in changing the existence form and entry position of the edge encoded fragments output by edge encoding. In existing Graphormer models, the edge encoded fragments formed by edge encoding are directly incorporated into the node pair bias positions in the attention bias, and the assembly constraint relationship can only participate in attention calculation as a static bias between node pairs. In this invention, within the edge relation token transformation layer, the edge encoded fragments are extracted from the content written to the attention bias, and rearranged into an edge relation to be encoded sequence according to the decoration region affiliation, the positions of the two end node tokens, and the assembly constraint direction, and then compressed into edge relation tokens. Through the above processing, the assembly constraint relationship is transformed from static bias content into an independent propagation object in the front-end token sequence of the Transformer encoder. The edge encoded fragment no longer stays at the node pair bias position, but is embedded in the propagation chain between the two end node tokens, providing a clear token object and connection position for the arrangement of the node edge association matrix.

[0041] In this embodiment, S6 specifically includes: The node-edge association matrix is ​​established after the Transformer encoder front-end token sequence is formed. The Transformer encoder front-end token sequence includes node tokens and edge relationship tokens. Each token position in the token sequence serves as both the row and column position of the node-edge association matrix. The row position represents the query source token, and the column position represents the key source token. If the Transformer encoder front-end token sequence contains N tokens, the node-edge association matrix is ​​N rows and N columns. A pass value or a block value is written at each row and column intersection. A pass value indicates that the corresponding query-key match is retained and participates in the self-attention calculation, while a block value indicates that the corresponding query-key match is superimposed with a negative blocking amount in the self-attention calculation. During matrix initialization, a pass value is written between each token and itself, and a block value is written between different node tokens that have not passed through an edge relationship token.

[0042] After an edge relation token enters the node edge association matrix, the matrix rewrites the row and column intersections along the positions of the two end node tokens carried by the edge relation token. When the edge relation token carries the positions of the starting and ending node tokens, the matrix writes a pass value at the intersection of the row corresponding to the starting node token and the column corresponding to the edge relation token, and also at the intersection of the row corresponding to the edge relation token and the column corresponding to the ending node token. When the edge between nodes carries a bidirectional assembly constraint, the pass values ​​are written simultaneously at the intersections of the row corresponding to the edge relation token and the column corresponding to the starting node token, as well as the intersections of the row corresponding to the ending node token and the column corresponding to the edge relation token. Direct row and column intersections between the starting and ending node tokens without passing through the edge relation token retain blocking values. When an edge relation token lacks either end node token position, the matrix only retains the pass value for the corresponding position of the edge relation token itself, and does not open intermediate positions between the edge relation token and the node token.

[0043] When the Transformer encoder performs self-attention computation, node tokens and edge relation tokens are first processed through trainable linear mappings to form query vectors, key vectors, and value vectors, respectively. The dimensions of the query vectors, key vectors, and value vectors are all consistent with the model's hidden dimensions, taking 256 dimensions. The query vector and key vector are multiplied by a dot product, and the result is divided by the square root of 256 to obtain the attention score. The attention score is then superimposed with the centrality bias and spatial bias retained in the attention bias, and then constrained by the pass value or block value of the node edge association matrix. For positions in the node edge association matrix where pass values ​​are written, the attention score retains the superimposed value; for positions where block values ​​are written, the attention score is further superimposed with a negative masking amount, which is set to -10000, so that the weight of the block position after softmax normalization is close to zero. After softmax normalization, value vector aggregation only forms effective weights along the positions corresponding to pass values; the positions corresponding to block values ​​do not participate in the effective value vector aggregation.

[0044] In the value vector aggregation phase, the starting node Token transmits the layout position, region affiliation, and installation direction encoding to the edge relation Token along the corresponding position of the pass value. The edge relation Token then transmits the assembly constraint relationship to the ending node Token along the corresponding position of the pass value. When reverse verification is enabled, the ending node Token transmits the assembly constraint relationship back to the starting node Token through the edge relation Token. The value vector aggregation result enters the residual connection and feedforward network of the current layer of the Transformer encoder. The residual connection adds the current layer input Token to the value vector aggregation result. The feedforward network sequentially performs linear mapping, GELU activation, and linear mapping, outputting the updated node Token and edge relation Token. The updated node Token continues to carry the layout state of component layout units and spatially constrained objects, and the updated edge relation Token continues to carry the intermediate expression of the assembly constraint relationship in the propagation chain, and together they enter the readout layer.

[0045] Compared to existing Graphermer models, the improvement of the node-edge association matrix lies in changing the propagation scope of self-attention computation and value vector aggregation within the Transformer encoder. In existing Graphermer models, the Transformer encoder mainly performs self-attention propagation around direct query-key matching between node tokens, and edge encoding only affects the attention score of node pairs through attention bias. This invention opens up the row and column intersection positions of node tokens to edge relation tokens and edge relation tokens to node tokens through the node-edge association matrix, and suppresses the direct node-to-node matching positions that do not pass through edge relation tokens, so that value vector aggregation is performed along the path from node to edge and back to node. Through the above processing, the assembly constraint relationship is no longer just a bias term within the Transformer encoder, but participates in the node token update in the edge relation token. The readout layer can extract the component layout content and assembly constraint content from the updated node token and edge relation token, respectively.

[0046] In this embodiment, S7 specifically includes: After receiving the updated node tokens and edge relationship tokens, the readout layer first distinguishes between the component layout unit tokens and the spatial constraint object tokens based on the node source in the node tokens. The component layout unit tokens are entered into the component layout result readout branch, while the spatial constraint object tokens only participate in constraint placement verification and do not directly output component dimensions. The component layout result readout branch reads the decoration surface region affiliation, local coordinate range, and installation direction code from the node tokens. The decoration surface region affiliation determines whether the component is located on the wall, floor, or ceiling; the local coordinate range determines the outermost range of the component in the two-dimensional coordinate system of the decoration surface region; and the installation direction code determines the component's installation direction. The readout layer inputs the component layout unit tokens into the linear mapping layer, which outputs the component width correction, component height correction, planar position offset, and direction response vector. The component width and height corrections are added to the original cut width and height of the component layout unit, respectively, to form the component dimensions. The planar position offset is added to the original lower left corner coordinate of the component layout unit to form the layout position. After the direction response vector is normalized by softmax, the direction with the largest response value is taken as the installation direction. When the dimension correction exceeds the allowable cutting range of the component parametric model library, the readout layer limits the component dimensions to the nearest processable specification and writes the limited component dimensions into the component layout result.

[0047] The edge relationship token enters the constraint placement result readout branch. The readout layer determines the two end node tokens corresponding to the edge relationship token based on the pass value position of the edge relationship token in the node edge association matrix. When both end node tokens exist, the readout layer reads the assembly constraint relationship vector in the edge relationship token and outputs the constraint placement vector through the linear mapping layer. The constraint placement vector includes the connection point offset, the avoidance boundary offset, the edge closing point offset, and the installation acceptance sequence response. The connection point offset is superimposed on the contact boundary of the component contours corresponding to the two end node tokens to form the component connection position; the avoidance boundary offset is superimposed on the projection boundary of the spatial constraint object to form the avoidance boundary; the edge closing point offset is superimposed on the edge margin position of the decoration area to form the edge closing position; the installation acceptance sequence response is normalized by softmax, and the order item with the largest response value is taken as the installation sequence. When the edge relationship token only connects to one end node token or does not connect to any node token, it does not enter the constraint placement result. The position corresponding to the edge relationship token is written into the abnormal position list, which is output along with the optimized component layout scheme for review.

[0048] The region merging process revolves around the component layout and constraint placement results. The readout layer first groups component dimensions, layout positions, and installation directions into the same decorative region based on their region affiliation. Then, it groups component connection positions, avoidance boundaries, edge finishing positions, and installation sequence relationships into their corresponding decorative regions. During merging, it first checks whether the outer contours of components within the same decorative region extend beyond the region's boundary. If the excess width does not exceed the allowable edge finishing width, the excess portion is merged into the edge finishing position. If the excess width exceeds the allowable edge finishing width, the corresponding component layout result is returned to the size correction process for re-limiting. Next, it checks the clearance between the avoidance boundary and the component's outer contour. If the clearance is less than the minimum installation clearance required by the component's parametric model library, the component layout position is first moved away from the projected boundary of the spatial constraint object, with the movement distance limited to reaching the minimum installation clearance. If, after moving, the component extends beyond the decorative region boundary or still cannot meet the minimum installation clearance, the corresponding edge relationship token is added to the abnormal position list. After region merging, the component layout results and constraint placement results are combined into an optimized component layout scheme.

[0049] The component fabrication and construction data are extracted from the optimized component layout scheme. The system first generates component fabrication data based on component size and layout position, recording the component's outer contour dimensions, cutting boundaries, opening positions, and edge finishing positions. Then, it generates construction assembly data based on installation direction and sequence, recording component number, installation area, installation direction, and assembly sequence. Component numbers are formed by combining the decoration area affiliation, component layout sequence, and component type codes from the component parametric model library. Within the same decoration area, components are arranged from bottom to top and from left to right; in the ceiling area, they are arranged according to the main keel direction. When updating the building BIM model, the optimized component layout scheme is written to the corresponding decoration area, component fabrication data is linked to the corresponding component family instance, construction assembly data is written to the assembly attributes of the component family instance, and the avoidance boundaries and abnormal location list corresponding to spatially constrained objects are written to the BIM model's verification attributes. After writing, the component geometry, fabrication attributes, and construction assembly attributes in the building BIM model maintain the same component number association, forming an updated model capable of supporting detailed design, component fabrication, and on-site installation.

[0050] Example 1: To verify the feasibility of this invention in practice, it was applied to a residential prefabricated decoration detailed design project. This project involved the integrated assembly layout of walls, floors, and ceilings. The building BIM model contained spatial constraints such as opening boundaries, electromechanical pipelines, equipment interfaces, and maintenance spaces. Traditional methods typically involve first arranging the components according to the dimensions of the decoration area and standard component specifications, then using collision checks to identify issues such as pipeline obstructions, equipment interface coverage, occupied maintenance spaces, or discontinuous installation sequences. Designers then repeatedly adjust component dimensions, joint positions, opening locations, and installation directions. This method is prone to inconsistencies between component layout drawings, fabrication lists, and installation sequence tables, especially in areas with numerous intersections between ceilings and electromechanical pipelines, dense wall interfaces, or irregularly distributed maintenance openings, resulting in substantial post-installation modifications.

[0051] During implementation, the structural information of the interior space is first extracted from the building BIM model. Walls, floors, and ceilings are encapsulated into interior surfaces, which are then matched with the component parametric model library to determine component layout units. Contents penetrating the interior surfaces or occupying installation space are designated as spatial constraint objects. Subsequently, assembly constraints are checked against the component layout units and spatial constraint objects, which are then organized into nodes. Assembly constraints are linked into edges between nodes, forming a component relationship diagram. After the component relationship diagram is integrated into the improved Graphormer model, the edge relationship token transformation layer truncates the direct write path from edge encoding to attention bias, organizes the assembly constraints carried by the edges between nodes into edge-encoded fragments, and then converts them into edge relationship tokens. The node-edge association matrix rewrites the direct node-to-node propagation path within the Transformer encoder into a node-to-edge-to-node propagation path, enabling node tokens and edge relationship tokens to be updated together within the model. After reading the updated node tokens and edge relationship tokens, the system merges them to form an optimized component layout scheme, extracts the component processing and construction data, and writes it back into the building BIM model.

[0052] To verify the feasibility of this invention in practice, 18 standard apartment types from the same prefabricated decoration detailing project were selected as verification objects, covering 96 decoration areas including walls, floors, and ceilings, involving 8620 decoration components. Manual verification confirmed 418 instances of component relationships in this batch of samples, including pipe obstruction, equipment interface occupation, maintenance space intrusion, and installation sequence conflicts. The traditional BIM manual layout method, BIM rule-based layout combined with the conventional Graphormer method, and the method of this invention were compared. The number of component layout conflicts, the proportion of components requiring manual adjustment, the number of installation sequence conflicts, the consistency rate of component processing and construction data, the completeness rate of BIM model write-back, and the average detailing adjustment time per apartment type were statistically analyzed. The results are shown in Table 1 below.

[0053] Table 1. Comparison of Consistency between Prefabricated Decoration Component Layout and Construction Data

[0054] As can be seen from the data in Table 1 above, the method of the present invention is superior to the traditional BIM manual layout method and the conventional Graphormer method in terms of component layout conflict control, manual adjustment workload, installation sequence continuity, construction data consistency and BIM model write-back integrity. Under the same verification scale of 8620 components, the traditional BIM manual layout method generated 418 3D conflicts, requiring manual adjustment of 18.6% of the components, and 126 installation sequence conflicts. This indicates that when relying on manual experience for component layout, there is a tendency for disjointedness between component size, spatial constraints, and installation connection relationships. BIM rule layout combined with the conventional Graphormer method reduced the 3D conflicts to 236 and the proportion of components requiring manual adjustment to 11.3%. This shows that the introduction of the graph model can improve the expression of component relationships, but it still mainly relies on direct propagation between nodes and static edge offsets, which is insufficient for the internal bearing capacity of assembly constraint relationships. The method of this invention further reduced the 3D conflicts to 89, the proportion of components requiring manual adjustment to 4.7%, and the installation sequence conflicts to 21. The consistency rate of component processing and construction data increased to 98.6%, the completeness rate of BIM model write-back increased to 97.9%, and the average time for detailed adjustment per unit type was shortened from 6.4 hours to 2.3 hours. This shows that the invention can achieve a more stable comprehensive effect in reducing conflicts, reducing manual corrections, and improving the closure of model data.

[0055] This embodiment organizes the structural information of the decoration space into decoration areas, component layout units, and spatial constraint objects. Then, it assembles the component layout units and spatial constraint objects into a component relationship graph, unifying the previously scattered component layout relationships, spatial avoidance relationships, and installation connection relationships in prefabricated decoration into a unified graph structure. The improved edge relationship token transformation layer in the Graphermer model separates the assembly constraint relationships carried by the edges between nodes from the edge encoding to the attention-biased direct write path, converting them into edge relationship tokens that can participate in the Transformer encoder propagation. The node-edge association matrix further rewrites the direct propagation path from node to node into a node-edge-to-node propagation path, allowing node tokens and edge relationship tokens to be updated together within the model. Therefore, the component layout result no longer relies solely on region size and post-collision correction, but simultaneously integrates spatial constraints and assembly constraints during model propagation, ultimately forming a component layout scheme more suitable for processing, installation, and model rewriting, demonstrating significant engineering application value.

[0056] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A BIM-based prefabricated building decoration method, characterized in that, Includes the following steps: S1. Extract the interior space construction information from the building BIM model, encapsulate the interior area and match it with the component parametric model library in terms of specifications, cut the component layout unit and mark the space constraint object. S2. Check the assembly constraint relationship along the component layout unit and spatial constraint object, organize the component layout unit and spatial constraint object into nodes, connect the assembly constraint relationship into the edge between nodes, and bind them into a component relationship diagram; S3. Connect the component relationship graph to the improved Graphormer model, wherein the improved Graphormer model embeds an edge relationship token conversion layer between edge encoding and Transformer encoder; S4. The edge relationship token conversion layer cuts off the direct write path from edge encoding to attention bias, organizes the nodes in the component relationship graph into node tokens, and organizes the assembly constraint relationship carried by the edges between nodes into edge encoding fragments. S5. The edge relationship token conversion layer converts edge encoded fragments into edge relationship tokens, and attaches the edge relationship tokens to the corresponding node tokens according to the endpoints of the edges between nodes in the component relationship graph. S6. Arrange the node-edge association matrix along the interlocking positions of the edge relationship token and the node token. The node-edge association matrix rewrites the direct propagation path from node to node in the Transformer encoder into a propagation path from node to edge and then to node, and updates the node token and edge relationship token along the rewritten propagation path. S7. Read out the updated node token and edge relationship token, merge them into the optimized component layout scheme, extract the component processing and construction data according to the optimized component layout scheme, and update the building BIM model.

2. The BIM-based prefabricated building decoration method according to claim 1, characterized in that, S1 specifically includes: S11. Extract interior space construction information from the building BIM model and delineate the boundaries of the interior surface area in the interior space construction information. S12. Close the disconnected decoration area boundary, remove the boundary area that does not participate in the component layout, and encapsulate it as a continuous decoration area. S13. Project the continuous decoration area onto the outer contour and edge boundary of the component in the component parametric model library, and cut the component layout unit along the projected component contour. S14. Check the occlusion of the space content that penetrates the continuous decoration area and occupies the component installation position. Mark the space content that is checked as a space restriction object. Merge the component layout unit and the space restriction object into the drawing object of the component relationship diagram.

3. The BIM-based prefabricated building decoration method according to claim 1, characterized in that, S2 specifically includes: S21. Assign aggregated component layout units and spatially restricted objects according to the decoration area to form candidate node groups on the same surface; S22. Compare the contact boundaries, fixed positions and installation directions between component arrangement units within the same candidate node group to screen out the assembly constraint relationships between component nodes. S23. Compare the occupancy positions and avoidance boundaries between component layout units and spatial constraint objects within the same candidate node group, and screen out the assembly constraint relationships between component nodes and constraint nodes. S24. Set the component layout units and spatial constraint objects to nodes, and snap the screened assembly constraint relationships into the edges between nodes to assemble a component relationship diagram.

4. The BIM-based prefabricated building decoration method according to claim 1, characterized in that, The improved Graphormer model specifically includes node embedding, centrality encoding, spatial encoding, edge encoding, edge relation token transformation layer, Transformer encoder, and readout layer.

5. A BIM-based prefabricated building decoration method according to claim 1, characterized in that, S4 specifically includes: S41. After the component relationship graph is connected to the improved Graphormer model, the node embedding receives the nodes in the component relationship graph, pushes the corresponding nodes of the component layout unit and the corresponding nodes of the space constraint object into the same node embedding sequence, and organizes them into the node initial token. S42, centrality encoding receives the number of node connections within the component relationship graph and forms a centrality bias; spatial encoding receives the graph distance between nodes within the component relationship graph and forms a spatial bias; centrality bias and spatial bias are retained within the attention bias write path. S43, edge encoding receives the edges between nodes in the component relationship diagram, performs edge encoding on the assembly constraint relationship carried by the edges between nodes, and organizes the endpoints and assembly constraint relationships corresponding to the edges between nodes into the same edge encoding sequence to form an edge encoding fragment; S44. The edge relationship token transformation layer is embedded between the edge encoding and the Transformer encoder, cutting off the path from edge encoding to the attention bias to directly write the edge encoding fragment. The edge encoding fragment is separated from the attention bias writing path and transferred to the entry point of the edge relationship token transformation layer. S45. When the edge relationship token conversion layer receives the edge encoded fragment, it simultaneously carries the node endpoint attribution and assembly constraint relationship between the corresponding nodes of the edge encoded fragment. Edge encoded fragments whose endpoint attribution is not fully hit are retained in the attention bias write path. S46. The edge relationship token conversion layer organizes the initial token of the node into a node token and pushes the complete edge encoding fragment of the endpoint into the edge relationship token conversion layer entry.

6. A BIM-based prefabricated building decoration method according to claim 1, characterized in that, S5 specifically includes: S51. The edge relationship token conversion layer reads the edge encoding fragment of the corresponding node pair bias table in the edge encoding, and extracts the assembly constraint edge sequence under the same node pair bias table item from the attention bias writing content. S52. The edge relationship Token conversion layer rearranges the edge encoding fragments based on the decoration region affiliation, the token positions of the two end nodes, and the assembly constraint direction associated with the assembly constraint edge sequence, forming an edge relationship to be encoded sequence with continuous arrangement of region affiliation, endpoint positions, and constraint directions. S53. The edge relationship Token transformation layer writes the edge relationship sequence to be encoded into the edge relationship Token construction bit of the Transformer encoder front end, and compresses it into an edge relationship Token that carries the decoration area affiliation, the Token positions of the two end nodes and the assembly constraint direction. S54. The edge relationship token transformation layer locates the two end node tokens corresponding to the edge relationship token in the node token sequence, embeds the edge relationship token into the propagation chain between the two end node tokens, and replaces the static bias connection of the original node pair bias table entry to the two end node tokens. S55. When there are multiple assembly constraint edge sequences under the same node pair bias table entry, the edge relation token conversion layer continuously places multiple edge relation tokens in the same propagation chain according to the edge sequence arrangement order in edge encoding. S56. The edge relationship token transformation layer outputs the edge relationship tokens embedded in the propagation chain, and arranges the edge relationship tokens embedded in the propagation chain and the corresponding node tokens together as the Transformer encoder front-end token sequence.

7. A BIM-based prefabricated building decoration method according to claim 1, characterized in that, S6 specifically includes: S61. Arrange the node tokens and edge relationship tokens into a Transformer encoder front-end token sequence, and establish the row and column positions of the node edge association matrix based on the token positions in the Transformer encoder front-end token sequence. S62. The position of the edge relationship token in the front-end token sequence of the Transformer encoder, respectively locating the start node token and the end node token corresponding to the edge relationship token; S63. In the node-edge association matrix, open the row and column intersection positions from the starting node Token to the edge relationship Token and from the edge relationship Token to the ending node Token, and hide the direct row and column intersection positions from the starting node Token to the ending node Token. S64. When the Transformer encoder calculates self-attention, it retains the attention scores corresponding to nodes to edges and edges to nodes according to the node-edge association matrix, and suppresses the attention scores directly corresponding to nodes. S65. During the value vector aggregation process, the Transformer encoder passes node tokens and edge relation tokens along the retained attention scores, and the assembly constraint relationships between node tokens are transferred through edge relation tokens. S66. The Transformer encoder updates the node token and edge relationship token along the value vector aggregation result, forming the node token and edge relationship token constrained by the node-edge association matrix.

8. A BIM-based prefabricated building decoration method according to claim 1, characterized in that, Specifically, S7 includes: S71. Read the updated node token and organize the component layout unit into component size, layout position and installation direction according to the node token's position in the decoration area. S72. Read out the updated edge relationship token, and according to the connection position of the edge relationship token in the node edge association matrix, rewrite the assembly constraint relationship into the constraint placement result; S73. Merge the component layout results and constraint placement results into a surface region to form an optimized component layout scheme. S74. Extract the component processing and construction data according to the optimized component layout scheme, and write the component layout results and component processing and construction data back to the building BIM model.

Images