File generation method and device based on spatial algorithm and large model

Abstract

The application provides a file generation method and device based on a space algorithm and a large model. The method obtains a three-dimensional model and a bill of materials of a target. After format conversion, the three-dimensional model is data-aligned with the bill of materials to generate semantic material information of each part in the three-dimensional model. The semantic material information of each part is input into a preset large model for semantic sorting to generate an initial process sequence. Based on a space algorithm, each part in the initial process sequence is checked, and the sorting is adjusted according to the checking result to determine a part assembly sequence and an explosion direction. Based on the explosion direction, an explosion diagram of an optimal viewing angle is generated as an explosion view. The part assembly sequence is input into the preset large model for semantic reasoning to determine semantic process descriptions of each part. The semantic process descriptions of each part and the explosion view are inserted into a preset template to generate a process guide file. The application realizes accurate generation of an SOP file from a three-dimensional model to a picture-text combination.

Description

Technical Field

[0001] This application relates to the field of industrial document processing, and in particular to a document generation method and apparatus based on spatial algorithms and large models. Background Technology

[0002] In technology-intensive industries such as automotive, high-end equipment manufacturing, aerospace, and precision machining, Standard Operating Procedures (SOPs) are core guiding documents for production activities. They detail the processing / assembly steps, operating procedures, technical parameters, required equipment, tooling and fixtures, quality inspection standards, and safety precautions for product components. The accuracy, standardization, and enforceability of SOPs are crucial foundations for ensuring consistent product quality, production efficiency, and operational safety.

[0003] However, in current manufacturing process management practices, the compilation and management of process documents still face many severe challenges, which greatly restrict the efficiency improvement of process departments and the high-speed response capability of enterprises. Specifically, these challenges are as follows: First, they rely heavily on manual labor and experience, resulting in low compilation efficiency; second, the knowledge reuse rate is low and the degree of standardization is insufficient; third, it is difficult to synchronize changes and the maintenance cost is high; and fourth, they are disconnected from digital systems, resulting in poor data flow.

[0004] Therefore, the industry urgently needs a new solution for generating process operation documents (SOPs) that can fundamentally transform the traditional SOP compilation model, enabling efficient accumulation, intelligent reuse, and automatic generation of process knowledge. This would significantly reduce manpower consumption, shorten process preparation cycles, and improve document quality and consistency, thereby providing key support for enterprises' digital transformation and intelligent manufacturing upgrades. Summary of the Invention

[0005] This application provides a document generation method and apparatus based on spatial algorithms and large models to solve the problems in the existing process document generation schemes, which suffer from physical illusions and cannot combine three-dimensional models to give actual assembly routes, resulting in low generation efficiency, poor physical feasibility, and lack of intuitive exploded view guidance.

[0006] In a first aspect, this application provides a file generation method based on spatial algorithms and large models, comprising: acquiring a three-dimensional model and a bill of materials of a target; converting the format of the three-dimensional model and aligning it with the bill of materials to generate semantic material information for each part in the three-dimensional model; inputting the semantic material information of each part into a preset large model for semantic sorting to generate an initial process sequence; verifying each part in the initial process sequence based on a spatial algorithm, adjusting the sorting according to the verification results to determine the part assembly sequence and explosion direction, wherein the verification includes collision detection and movement direction prediction; generating an exploded view with an optimal perspective based on the explosion direction, wherein the optimal perspective for each part is calculated by the part assembly sequence; inputting the part assembly sequence into the preset large model for semantic reasoning to determine the semantic process description of each part; inserting the semantic process description of each part and the exploded view into a preset template to generate a process guidance document.

[0007] In some possible embodiments of the first aspect, after the 3D model is format-converted, it is aligned with the bill of materials to generate semantic material information for each part in the 3D model, including: converting the 3D model to obtain a model structure tree; aligning each node in the model structure tree with the bill of materials so that each node maps to at least one attribute information in the bill of materials, and each node is at least one part; classifying the unordered parts in the model structure tree according to the bill of materials type to obtain semantic material information for each part in the 3D model.

[0008] In some possible embodiments of the first aspect, the semantic material information of each part is input into a preset large model for semantic sorting to generate an initial process sequence, including: inputting the semantic material information of each part into a preset large model for semantic information extraction to determine semantic information; performing cosine similarity retrieval on the semantic information in a preset vector database to determine multiple similar process vectors, wherein the preset vector database includes a historical process vector database; using the preset large model to weight and sort the multiple similar process vectors according to semantic similarity, time popularity, and usage frequency, and arranging the vector sorting results from high to low as the initial process sequence.

[0009] In some possible embodiments of the first aspect, each part in the initial process sequence is inspected based on a spatial algorithm, and the sorting is adjusted according to the inspection results to determine the assembly order and explosion direction of the parts. This includes: performing spatial analysis on the potential features of each part based on the geometric topology information of the three-dimensional model of each part in the initial process sequence, and generating collision results based on a virtual collision detection algorithm; predicting the movement direction of each part in the initial process sequence to determine the movement direction information of each part; selecting the highest priority movement direction of each part in the movement direction information for path planning, and determining the part installation order by combining the initial direction and initial position in the movement direction information; and determining whether there is an interference relationship between each part and its adjacent parts based on the collision results, and determining the explosion direction according to the interference relationship corresponding to the part installation order.

[0010] In some possible embodiments of the first aspect, the movement direction prediction of each part in the initial process sequence is performed to determine the movement direction information of each part, including: determining the contact surface information of each part in the three-dimensional model; based on the contact surface information of each part, performing normal calculation according to principal component analysis and planar weighted clustering to determine all movable directions under all contact surfaces of each part; ranking all movable directions under all contact surfaces according to preset weights, and filtering based on the ranking results to determine the movement direction information of each part.

[0011] In some possible embodiments of the first aspect, the highest priority movement direction for each part in the movement direction information is selected for path planning, and the initial direction and initial position in the movement direction information are combined to determine the installation order of the parts, including: obtaining all movement direction information of each part in the three-dimensional model; filtering all movement direction information of each part to determine the highest priority movement direction; The highest priority movement direction is selected as the dominant direction. Based on the initial direction and initial position in the movement direction information, the assembly movement path of each part is determined. Based on the spatial constraint relationship between the assembly movement paths, the assembly sequence of the parts is determined.

[0012] In some possible embodiments of the first aspect, generating an exploded view with the optimal perspective based on the explosion direction includes: transforming the vertices of the three-dimensional model to a two-dimensional clipping space according to the three-view projection based on the set of surface normals of all parts in the three-dimensional model, obtaining the projection results of the parts in three orthogonal directions, extracting the outline of the parts based on the projection results from multiple perspectives, and determining the optimal projection perspective. Under the optimal perspective, the segmentation map of the initial image containing all parts is extracted by the panoramic segmentation network. The appearance features are extracted layer by layer and the channels are connected to form a conditional signal input into the preset diffusion model. The explosion distance is adaptively calculated by combining the translation transformation matrix to generate an unobstructed and clear explosion view.

[0013] In some possible embodiments of the first aspect, it further includes: simultaneously outputting a review report when outputting the process guidance document, and marking paragraphs with confidence levels below a preset threshold in the process guidance document; receiving instructions from process engineers to check and fine-tune the low-confidence paragraphs; and if a physical conflict is detected during the pre-simulation, generating a semantic feedback signal, feeding the semantic feedback signal back to the large model to trigger the correction and re-inference of the constraint vectors, and automatically feeding the modified high-quality data back to the vector database to achieve self-learning evolution.

[0014] In some possible embodiments of the first aspect, the method further includes: employing a multi-source knowledge fusion algorithm to perform weighted fusion of the semantic representation vectors of each source knowledge to reconstruct the optimal process description text; parsing a preset template, using regular expressions to traverse the text nodes in the preset template, identifying the positioning placeholders, and extracting strings containing type identifiers and name identifiers; establishing a mapping relationship between the process data to be filled and the positioning placeholders, injecting the process data in the optimal process description text into the corresponding position of the preset template based on the mapping relationship, and rendering it according to the format attributes to generate a process guidance document.

[0015] In a second aspect, this application provides a file generation device based on spatial algorithms and a large model, comprising: an acquisition module for acquiring a target 3D model and a bill of materials; a preprocessing module for converting the format of the 3D model and aligning it with the bill of materials to generate semantic material information for each part in the 3D model; a semantic sorting module for inputting the semantic material information of each part into a preset large model for semantic sorting to generate an initial process sequence; a sequence verification module for verifying each part in the initial process sequence based on a spatial algorithm, adjusting the sorting according to the verification results to determine the part assembly sequence and explosion direction, wherein the verification includes collision detection and movement direction prediction; a view generation module for generating an exploded view with an optimal perspective based on the explosion direction, wherein the optimal perspective for each part is calculated by the part assembly sequence; a process description module for inputting the part assembly sequence into the preset large model for semantic reasoning to determine the semantic process description for each part; and a file generation module for inserting the semantic process description of each part and the exploded view into a preset template to generate a process guidance document.

[0016] In a third aspect, this application also provides an electronic device, including a processor, a memory, and a communication bus; the communication bus is used to connect the processor and the memory; the processor is used to execute a computer program stored in the memory to implement the file generation method based on spatial algorithms and large models described above.

[0017] The beneficial effects of this application are as follows: A 3D model and bill of materials for the target are obtained; the 3D model is format-converted and aligned with the bill of materials to generate semantic material information for each part in the 3D model, establishing relationships between geometry, attributes, and functions; the semantic material information of each part is input into a preset large model for semantic sorting to generate an initial process sequence; each part in the initial process sequence is checked based on a spatial algorithm, and the sorting is adjusted according to the check results to determine the part assembly sequence and explosion direction, wherein the check includes collision detection and movement direction prediction; based on the explosion direction, an exploded view with the optimal perspective is generated as the exploded view, ensuring the integrity and clarity of information transmission and improving the readability and guidance of drawings; the part assembly sequence is input into the preset large model for semantic reasoning to determine the semantic process description of each part; the semantic process description of each part and the exploded view are inserted into a preset template to generate a process guidance document. The semantic process description, combined with the assembly path and spatial relationships, forms an executable and traceable process language, which is finally embedded in the preset template to form a standardized and reusable process guidance document. The large model provides a priori coarse sorting based on historical experience, which greatly reduces the complexity of spatial calculations; the spatial geometry algorithm solves the physical illusion problem of the large model, ensuring the absolute physical feasibility of the assembly sequence, and automatically realizes the accurate generation of SOP files with illustrations from the 3D model. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0019] In the attached diagram: Figure 1 A flowchart of a file generation method based on spatial algorithms and large models provided in an embodiment of this application; Figure 2 A flowchart illustrating the overall process of a file generation method based on spatial algorithms and large models provided in an embodiment of this application; Figure 3 A flowchart illustrating the application of a large model in a file generation method based on spatial algorithms and large models provided in an embodiment of this application; Figure 4 A complete flowchart of a file generation method based on spatial algorithms and large models provided in an embodiment of this application; Figure 5 A structural block diagram of a file generation apparatus based on spatial algorithms and large models provided in an embodiment of this application; Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0020] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0021] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. The drawings only show the components related to this application and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0022] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the present application. However, it will be apparent to those skilled in the art that embodiments of the present application may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the present application.

[0023] In related technologies, this application aims to address key technical bottlenecks in existing assembly process design, such as data disconnect between 3D models and bills of materials (BOMs), lack of logical basis for assembly sequence, semantic ambiguity in process description, unreasonable perspective of exploded views, and insufficient automatic generation capability of process guidance documents. Traditional assembly process design relies on manual experience and judgment, resulting in low efficiency, poor consistency, and susceptibility to errors. Especially in complex electromechanical products or high-end equipment, the rationality of the assembly path directly affects production efficiency and manufacturing quality. Therefore, there is an urgent need for an automated assembly process generation method that integrates 3D modeling, semantic understanding, and spatial reasoning to achieve an intelligent generation solution from product structure to process guidance.

[0024] To address the aforementioned problems, this application proposes a file generation method and apparatus based on spatial algorithms and large models. Please refer to [link / reference]. Figure 1The above is a flowchart illustrating a file generation method based on spatial algorithms and large models provided in an embodiment of this application, including: Step S110: Obtain the 3D model and bill of materials of the target; This application serves the entire industrial equipment manufacturing industry, and its broad applicability is reflected in its technical compatibility with mainstream 3D modeling software such as CATIA, SolidWorks, UG NX, Creo, and Inventor. It can read and convert models output by any 3D modeling software. The model conversion function supports the recognition of data formats from mainstream 3D modeling software, such as CATProduct, CATPart, JT, SLDASM, SLDPRT, ASM, and STP, and can recognize data in 3D feature representation formats such as BREP, CSG, and MESH.

[0025] Structure normalization is used to normalize the identified 3D modeling software data, forming a unified data structure throughout the system's operation. This eliminates data inconsistencies caused by different industrial software, enabling cross-platform compatibility. Topology optimization is used to optimize the data in the 3D modeling software, correcting abnormal topology issues such as broken topologies and non-manifolds.

[0026] For example, the 3D model (STEP, IGES, STL format) of the target product and the corresponding Bill of Materials (BOM) can be imported through a CAD system or a Product Data Management System (PDM). The BOM contains structured information such as the name, number, material, quantity, and hierarchy of each component.

[0027] Step S120: After converting the format of the 3D model, align the data with the bill of materials to generate semantic material information for each part in the 3D model; This can be understood as binding the input 3D model one-to-one with the Bill of Materials (BOM) list from the design and production side, obtaining information on all parts in the BOM list. Structure tree matching involves transforming the input 3D model to obtain a model structure tree. Each node in the structure tree is then matched by its node ID. For IDs with the same name, the nearest distance is checked between all models with the same name and their spatial locations provided in the BOM. All node data and BOM data are traversed, and the results are aligned one-to-one. This automated alignment solves the problems of inaccuracy and high labor costs associated with manual alignment of process documents, efficiently establishing a one-to-one mapping between the 3D model and structured data, providing accurate information input for subsequent process document generation.

[0028] Specifically, precise correspondence between BOM entries and model components is achieved through spatial topological relationships (such as parent-child relationships and adjacency relationships) and attribute matching rules. For example, "motor-bracket-left side" in the BOM corresponds to a component entity with specific positional and dimensional characteristics in the model. Exemplarily, the original 3D model is converted into a unified intermediate format (such as Parasolid or ACIS), and based on the hierarchical relationships in the BOM, the geometric entity of each component in the model is bound one-to-one with its attributes (such as name, material, and functional category), forming a triplet structure of part, geometry, and attributes, i.e., semantic material information.

[0029] Step S130: Input the semantic material information of each part into the preset large model for semantic sorting and generate the initial process sequence; It should be noted that by performing structured semantic extraction and vectorized sorting based on historical SOP documents, the problems of isolated historical experience, unusable repetitive semantic information, and missing core semantics in traditional documents are solved.

[0030] The process involves several steps: Corpus cleaning preprocesses historical documents, cleaning the text structure and parsing structured semantics. Semantic information extraction builds a pre-trained large model, fine-tuned with domain knowledge, to perform deep context-aware semantic parsing of historical SOP files, accurately extracting multi-level semantic elements such as operational entities, action instructions, environmental parameters, logical order, and constraints involved in each step. Vectorized sorting generates a set of historical vectors based on the semantic information extracted from the historical data. This vector extraction process is then repeated, extracting query vector information from the model's structure tree and BOM (Bill of Materials) to generate the SOP. The cosine similarity between query vectors and historical vectors is calculated to quantify their semantic association strength. A weighted sorting algorithm, primarily based on semantic similarity and secondarily on temporal popularity and usage frequency, is used to comprehensively sort all query vectors, generating an ordered list for structured output and computation by spatial algorithms.

[0031] By extracting semantic information, the SOP and other process documents with good data quality in the past are restructured and vectorized. Based on the input model information, the input is weighted and vectorized according to the historical order. The new model is first coarsely sorted according to historical experience, effectively utilizing historical experience to provide a verification basis for the correct output.

[0032] The large model performs pattern recognition and logical inference based on assembly rules accumulated in the training data (such as "install the base first, then the column" and "the motor must be in place before the drive shaft is installed"), sorting components according to dimensions such as "functional dependency—spatial pre-positioning—installation convenience." For example, the semantic material information of each part (including name, functional category, dependency relationship, installation location, etc.) is fed into a pre-trained large model (LLM, i.e., large language model) or a dedicated industrial semantic model. Through rules such as semantic similarity, functional dependency, and installation logic, it infers and outputs a preliminary assembly sequence. This generates an initial assembly path that conforms to engineering common sense, significantly reducing the trial-and-error costs in the early stages of manual design. This sorting is not simply based on position, but rather on a comprehensive judgment of functional logic and spatial constraints, reflecting semantic rationality.

[0033] Step S140 involves inspecting each part in the initial process sequence based on a spatial algorithm, and adjusting the sorting according to the inspection results to determine the part assembly sequence and explosion direction. The inspection includes collision detection and movement direction prediction.

[0034] Among these methods, spatial modeling based on bounding boxes, voxel meshes, or surface distance fields is used to evaluate the relative positional relationships between components in real time during assembly. Dynamic path planning and constraint solving ensure that the assembly path is physically feasible.

[0035] Among them, the collision detection algorithm is a key step in obtaining part assembly information. It is used to determine the intersecting and disjoint relationships between spatial topological structures, and then obtain the collision relationship between different structures.

[0036] For example, for each assembly operation in the initial process sequence, the following spatial checks are performed: collision detection, used to determine whether the part to be installed has geometric interference with other installed parts in the current assembly state; and movement direction prediction, based on the installation position and orientation of the part in three-dimensional space, to predict whether its installation path is feasible (e.g., whether rotation, translation, or tilting is required). If a collision or infeasible path exists, the process is backtracked to the previous steps, the assembly sequence is adjusted, or the installation orientation is modified, until all operations meet the spatial feasibility conditions. In this way, the assembly sequence and explosion direction of the parts are determined, ensuring that the assembly sequence is feasible in physical space and avoiding manufacturing infeasibility problems caused by design errors.

[0037] Step S150: Based on the explosion direction, generate an exploded view with the optimal perspective. The optimal perspective for each part is calculated based on the part assembly sequence.

[0038] Based on three-dimensional spatial geometry, a perspective optimization algorithm (such as maximum visible angle, minimum occlusion area, and shortest connecting line length) is used to calculate the viewing direction of each part, ensuring that the connection relationships between components in the exploded view are clearly distinguishable and minimizing overlap and occlusion. For example, according to the final determined assembly sequence, the optimal viewing angle for each part when it is "disassembled" during the assembly process is calculated. This viewing angle is defined as the angle from the part's final position before removal that most clearly shows its connection relationships, installation positions, and relative relationships with other components. In this way, the generated exploded view not only conforms to engineering standards but also possesses structural clarity and information completeness, intuitively reflecting the assembly relationships.

[0039] Step S160: Input the assembly sequence of the parts into the preset large model for semantic reasoning to determine the semantic process description of each part; The large model combines contextual information such as assembly path, spatial location, and connection method to generate conditions and perform language modeling, producing process descriptions that conform to engineering language specifications and embody the complete semantic chain of "action-object-parameter-constraint". For example, the determined assembly sequence is used as contextual input, combined with the semantic material information of the parts, and input into the large model for context-aware semantic reasoning, outputting a semantic process description for each part, such as "This bracket needs to be translated 5mm along the positive Z-axis and rotated 15° before fixing after the base is installed." In this way, the generated semantic process descriptions are readable, executable, and traceable, and can be directly used in production guidance documents, reducing manual writing and translation costs.

[0040] Step S170: Insert the semantic process description and exploded view of each part into a preset template to generate a process guidance document.

[0041] The template is designed based on industry standards (such as ISO13485 and GB / T50892) and includes modules such as title, table of contents, parts list, exploded view, process steps, and safety tips to ensure that the output document complies with the company's internal management specifications and regulatory requirements. For example, based on the assembly explosion direction, it determines multiple projection perspectives and calculates the optimal projection perspective, providing guidance for generating the exploded view.

[0042] The generated semantic process descriptions and corresponding exploded views of the parts are structurally assembled according to standard process document templates (such as assembly instructions, process cards, and work instructions) to form standardized process guidance documents. This achieves a fully automated closed loop from product design to process documentation, significantly improving document generation efficiency and consistency, reducing human error rates, supporting rapid output in multiple languages ​​and versions, and meeting document delivery needs in different application scenarios.

[0043] In some embodiments, after the 3D model is format-converted, it is aligned with the bill of materials to generate semantic material information for each part in the 3D model, including: Convert the format of the 3D model to obtain the model structure tree; Align each node in the model structure tree with the bill of materials so that each node maps to at least one attribute from the bill of materials and each node represents at least one part. The unordered parts in the model structure tree are classified according to the bill of materials type to obtain the semantic material information of each part in the 3D model.

[0044] For example, a 3D model (file) is acquired, its native format (such as CATPart, SLDPRT) is parsed, and it is converted into a preset intermediate geometric format (such as STEPAP242, JT, or GLTF). During the conversion process, the assembly hierarchy, part entity information, and spatial coordinates of the 3D model are extracted and reconstructed into a tree-like topology structure, i.e., the model structure tree. Each node in this tree corresponds to an assembly or part, and the root node represents the final assembly. All nodes in the model structure tree are traversed, and the node identifiers (such as name, ID, or geometric hash value) are extracted. At the same time, the BOM data is parsed, and the attribute information of the BOM entries (such as material code, specifications, and material) is extracted. Based on preset matching rules (such as string fuzzy matching, geometric feature comparison, or spatial location clustering), a one-to-one correspondence between nodes and BOM entries is established. After the mapping is completed, the attribute information in the BOM (such as "material: aluminum alloy") is attached to the corresponding model node. Based on the mapping relationship, the BOM type field (such as "standard part", "electronic component", "structural part") corresponding to each node is read. According to the preset classification system, the part nodes in the model structure tree are traversed and classified. Each part node not only contains geometric data and BOM attributes, but is also labeled with a clear semantic category, forming "semantic material information".

[0045] In this embodiment, the data fusion challenge between the 3D model and the bill of materials is systematically solved through a progressive process of format conversion, data alignment, and semantic classification. First, format conversion ensures the standardization of data sources; second, data alignment establishes the association between geometry and attributes, resolving the "who is who" problem; finally, semantic classification imbues the data with logical meaning, resolving the "what is it?" problem. These three processes work together to upgrade the traditional geometric model into a digital model rich in semantic information, providing a high-confidence data foundation for automated process generation, virtual simulation, and production control in intelligent manufacturing, effectively reducing manufacturing errors caused by data inconsistency.

[0046] In some embodiments, the semantic material information of each part is input into a preset large model for semantic sorting to generate an initial process sequence, including: Input the semantic material information of each part into the preset large model to extract semantic information and determine the semantic information; Semantic information is searched using cosine similarity in a pre-defined vector database to identify multiple similar process vectors. The pre-defined vector database includes a historical process vector database. Using a pre-defined large model, multiple similar process vectors are weighted and sorted according to semantic similarity, temporal popularity, and usage frequency. The vector sorting results are arranged from high to low as the initial process order.

[0047] For example, semantic material information for each part is obtained, including part name, functional category, material properties, assembly constraints, and module level. This structured information is input into a pre-trained Large Language Model (LLM). The model's internal attention mechanism captures key information features, outputting a high-dimensional semantic vector representing the core assembly attributes of the part and its corresponding keyword set—the semantic information. The cosine similarity between the semantic information vector and all historical process vectors in a pre-defined vector database is calculated. A similarity threshold is set, and candidate vectors exceeding the threshold are selected to form a set of similar process vectors. The pre-defined vector database is constructed from pre-processed and vectorized historical process documents (such as process specifications and work instructions), covering assembly process data for different product types. Multi-dimensional features are extracted from each similar process vector: semantic similarity score, time popularity (the timeliness weight of recent process data), and usage frequency (the number of times the process has been adopted in historical projects). These features are input into the pre-defined large language model, and a comprehensive score is calculated based on a built-in weighting strategy (such as weighted summation or attention weight allocation). Similar process vectors are sorted in descending order of their comprehensive scores, and the assembly step sequence corresponding to the top-ranked process vector is mapped to the initial assembly order of the current part, i.e., the initial process sequence.

[0048] In this embodiment, a data-driven process sequence generation mechanism is constructed through cascaded processing of semantic extraction, vector retrieval, and multi-dimensional ranking. First, semantic extraction enables in-depth analysis of part assembly attributes; second, vector retrieval enables precise recall of historical process knowledge; and finally, multi-dimensional weighted ranking enables intelligent optimization of the recalled results. The synergistic effect of these three mechanisms transforms traditional "experience-driven" process design into "data-intelligence-driven" process generation, effectively improving the automation level and reliability of process design, and shortening the process planning cycle.

[0049] In some embodiments, each part in the initial process sequence is inspected based on a spatial algorithm, and the sorting is adjusted according to the inspection results to determine the part assembly sequence and explosion direction, including: Based on the geometric topology information of each part in the 3D model in the initial process sequence, the potential features of each part are spatially analyzed based on the virtual collision detection algorithm to generate collision results; Predict the movement direction of each part in the initial process sequence to determine the movement direction information of each part; Select the highest priority movement direction for each part in the movement direction information for path planning, and combine the initial direction and initial position in the movement direction information to determine the part installation sequence; Based on the collision results, it is determined whether there is an interference relationship between each part and its adjacent parts. The direction of the explosion is determined according to the interference relationship corresponding to the installation order of the parts.

[0050] For example, each part in the initial process sequence is traversed, and its geometric topology data in the 3D model is extracted, including vertex, edge, face, and voxel information. Based on the Bounding Volume Hierarchy (BVH) or Separating Axis Theorem (SAT), geometric proxy models (such as axis-aligned bounding boxes (AABB) or oriented bounding boxes (OBB)) of the part and its adjacent parts are constructed. The motion trajectory of the part along the predetermined assembly path is simulated, and the intersection volume between the part's geometric proxy model and the assembled environment is calculated in real time. If an intersection is detected, the spatial coordinates of the collision, the interference depth, and the identifier of the interfering part are recorded, and a collision result set is generated. For each part in the initial process sequence, its relative positional relationship with the assembly reference plane or connection feature (such as hole, groove, flange face) is analyzed. Based on the geometric symmetry of the part, the fit tolerance, and the principal axis direction of the connection feature, the principal geometric axes of the part are extracted using Principal Component Analysis (PCA). Based on assembly constraint types (e.g., insertion, fitting, threaded connection), the translation vector and rotation angle required for a part to move from an unconstrained state to a fully assembled state are predicted, generating movement direction information containing multiple candidate movement directions and their priority weights. The movement direction information is read, and the movement direction with the highest priority is selected as the preferred assembly path for the part. Based on this path, a displacement vector from the initial position (unassembled state) to the target position (assembled state) is constructed, combining the part's initial position (unassembled state) and target position (assembled state). All parts are traversed sequentially; if a part's preferred path is unobstructed in the current assembly environment (based on collision results), it is marked as installable. This process is iterated, dynamically adjusting the initial process sequence, placing parts with no collision paths first and parts with collision risks later, thus determining the final part installation sequence. Based on the part installation sequence, the disassembly process is analyzed in reverse. Based on collision results, the adjacent parts that first interfere with each part on the disassembly path are identified. The explosion direction is defined as the direction vector by which a part moves from its assembly position along its installation path in the opposite direction until it escapes all interference areas. If the part has multiple interference sources, the direction that requires overcoming the least interference resistance (such as the shortest moving distance or the smallest interference volume) is selected as the optimal explosion direction.

[0051] In this embodiment, a verification closed loop from logical process to physical process is constructed through deep coupling of geometric topology analysis, collision detection, movement direction prediction, and path planning. First, collision detection reveals the physical facts of spatial interference; second, movement direction prediction defines the boundaries of the degrees of freedom of motion; third, path planning and sequence adjustment ensure the feasibility of assembly actions; and finally, the determination of the explosion direction based on interference relationships ensures the authenticity of the visual representation. Each step supports the others, upgrading the traditional "two-dimensional drawings + experience-based imagination" process design mode to a digital design mode of "three-dimensional simulation + physical verification," effectively improving the accuracy and reliability of process design and reducing the trial-and-error costs in the manufacturing stage.

[0052] In some embodiments, the movement direction prediction is performed for each part in the initial process sequence to determine the movement direction information of each part, including: Determine the contact surface information of each part in the 3D model; Based on the contact surface information of each part, the normal is calculated according to principal component analysis and planar weighted clustering to determine all movable directions under all contact surfaces of each part; The movement direction information of each part is determined by ranking all movable directions under all contact surfaces according to preset weights and filtering based on the ranking results.

[0053] For example, all parts in the 3D model are traversed, and face-to-face, edge-to-face, or point-to-face contacts between parts are detected based on geometric topological relationships. Geometric faces that are in contact are extracted, and their vertex coordinates, normal vectors, areas, and corresponding paired part identifiers are recorded. All geometric faces on the same part that are in contact with different parts are aggregated to construct the contact surface set for that part, i.e., contact surface information. For the contact surface set of a single part, Principal Component Analysis (PCA) is performed on the normal vectors of each contact surface to extract the principal axis directions of the normal vector distribution. A planar weighted clustering algorithm is used, with the contact surface area as the weight, to perform K-means clustering or mean-shift clustering on the normal vectors, grouping normal vectors with similar directions into the same cluster. The weighted average normal vector of each cluster is calculated, and its opposite direction (i.e., the direction of detachment from contact) and the zero vector (representing movement along the tangential direction of the contact surface) are used together as the candidate set of movable directions for the part based on contact surface constraints. Preset weights are assigned to various movable directions. For example, the weight is highest along the direction opposite to the normal of the primary constraint, followed by the direction opposite to the normal of the secondary constraint, and lowest along the tangential direction. A comprehensive score is calculated for all candidate movable directions (e.g., the matching degree between the direction vector and the weighted cluster center multiplied by the preset weight). All candidate directions are ranked in descending order of their comprehensive scores. The top-ranked directions that are not collinear (with an angle greater than a preset threshold) are selected to constitute the movable direction information for the part.

[0054] In this example, intelligent solutions for the movement direction of a part are achieved through three progressive steps: contact surface extraction, normal statistical analysis, and weighted decision-making. First, the contact surface information clarifies the source of constraints. Second, PCA and weighted clustering reveal the primary and secondary relationships of constraints from a statistical perspective, determining physically feasible movement trends. Finally, ranking and selection based on preset weights transforms geometric possibilities into optimal engineering solutions. These steps work synergistically, transforming traditional empirical direction judgment into a quantitative calculation process based on geometric topology and statistical learning, significantly improving the accuracy and reliability of movement direction prediction.

[0055] In some embodiments, the highest priority movement direction for each part in the movement direction information is selected for path planning. Combining the initial direction and initial position in the movement direction information, the part installation sequence is determined, including: Obtain all movement directions information for each part in the 3D model; Filter all movement direction information for each part and determine the highest priority movement direction; The highest priority direction of movement is selected as the dominant direction. Based on the initial direction and initial position in the movement direction information, the assembly motion path of each part is determined. The assembly sequence of parts is determined based on the spatial constraints between assembly motion paths.

[0056] For example, 3D model data and preprocessed semantic information of parts are read, and the degree-of-freedom analysis results of each part in the assembly space are extracted. The movement direction information includes potential translation vectors, rotation axis information, and corresponding kinematic pair types (such as prismatic pairs and revolute pairs) derived from geometric topology analysis. The above information for each part is constructed into a structured dataset, where each entry contains a part identifier, direction vector, motion type, and constraint source. An evaluation index system for direction selection is established, including direction compliance (whether it conforms to assembly process specifications), motion smoothness (minimum change in path curvature), and constraint release degree (whether movement along this direction can release constraints the fastest). Each candidate movement direction is weighted and scored according to preset weight coefficients. The direction with the highest score is selected as the preferred movement direction for that part and is designated as the highest priority movement direction, while the remaining directions are stored as alternative redundant paths.

[0057] Using the highest priority direction of movement as the principal axis of motion, and combining the initial position (unassembled state) and target position (design reference position) of the part in the 3D model, a displacement vector is constructed. An interpolation algorithm (such as linear interpolation or spherical linear interpolation) is used to generate a continuous pose sequence from the initial position to the target position. During this process, key nodes on the path (such as starting point, obstacle avoidance point, and ending point) are recorded simultaneously to form a complete assembly motion path. The assembly motion paths of all parts are traversed, and pairwise cross-checks and spatial relationship analysis are performed. If the assembly path of part A traverses the space occupied by part B after assembly, it is determined that A and B have a spatial constraint relationship, and B must be installed before A. A directed acyclic graph (DAG) is constructed with parts as nodes and spatial constraints as directed edges. This graph is topologically sorted, and a linear installation sequence satisfying all spatial constraints is output.

[0058] In this embodiment, a closed-loop processing mechanism of direction selection, path generation, and constraint sorting constructs an assembly logic chain from microscopic motion to macroscopic temporal sequence. First, direction selection establishes the optimal primitives for motion; second, path generation visualizes these primitives as spatial trajectories; and finally, topological sorting based on path constraints ensures the rationality of the global temporal sequence. Each step progresses progressively, transforming traditional assembly planning that relies on manual trial and error into a deterministic process based on automatic algorithm derivation. This significantly improves the automation level and reliability of assembly process design and reduces the risk of rework in the manufacturing process.

[0059] In some embodiments, generating an exploded view with an optimal perspective based on the explosion direction includes: Based on the set of surface normals of the 3D model containing all parts, the vertices of the 3D model are transformed to the 2D clipping space according to the three-view projection to obtain the projection results of the parts in three orthogonal directions. The part contour is extracted based on the multi-view search projection results to determine the optimal projection view. Under the optimal perspective, the segmentation map of the initial image containing all parts is extracted by the panoramic segmentation network, the appearance features are extracted layer by layer and the channels are connected to form a conditional signal input into the preset diffusion model. Combined with the translation transformation matrix, the explosion distance is adaptively calculated to generate an unobstructed and clear explosion view.

[0060] The multi-view search automatically determines multiple projection views and calculates the optimal projection view based on the assembly explosion direction, providing guidance for explosion view generation. The explosion view optimization selection automatically arranges the obtained optimal view parameters using heuristic layout optimization to generate the optimal exploded view.

[0061] For example, the surface mesh of all parts in the 3D model is traversed, the normal vector of each triangular facet is calculated, and the surface normal vector set is constructed. Based on perspective projection or orthogonal projection matrix, the coordinates of the model vertices are transformed from the world coordinate system to the clipping coordinate system, and after perspective division, they are mapped to the 2D screen space to generate projection results along the X, Y, and Z axes (i.e., top view, front view, and side view). Edge detection and contour extraction are performed on the projection results, and the number of independent closed contours, the total area of ​​the contours, and the contour overlap are counted under each viewpoint. A viewpoint evaluation function is constructed, and the viewpoint with the strongest contour independence and the lowest overlap is selected as the optimal projection viewpoint. The initial rasterized image of the 3D model is rendered under the optimal viewpoint. This image is input into a pre-trained panoptic segmentation network to identify the part instance and semantic category to which each pixel in the image belongs, and outputs a pixel-level segmentation mask image. The appearance features (such as texture, color, and edge intensity) of each part region in the segmentation image are extracted, and multi-channel features are stitched together according to spatial location to construct a tensor containing positional and semantic information as a conditional signal. The conditional signal is input into a preset denoising diffusion probabilistic model (DDIM). Simultaneously, the motion transformation matrix generated during the assembly path planning stage (recording the displacement vectors of each part relative to the assembly datum) is read, and pixel-level offsets are adaptively calculated based on part dimensions and gaps between adjacent parts. In the inverse denoising process of the diffusion model, the conditional signal guides the generation process, and affine transformations are performed on the segmented regions according to the calculated offsets, ultimately synthesizing an unobstructed exploded view.

[0062] In this embodiment, an end-to-end intelligent exploded view generation process is constructed by organically combining geometric perspective optimization and deep learning generation. First, a perspective selection mechanism based on normal and projection analysis solves the problem of "from which angle to view," ensuring the integrity of the information input. Second, a generation mechanism based on panoramic segmentation and diffusion models solves the problem of "how to generate clear and unobstructed images," achieving a leap from geometric models to semantic images. This breaks through the reliance on human experience in traditional exploded view generation, realizing automation, intelligence, and standardization of exploded view generation, significantly improving the quality and efficiency of digital product delivery.

[0063] In some embodiments, it also includes: When outputting process guidance documents, a review report is output simultaneously, and paragraphs with confidence levels below a preset threshold are marked in the process guidance documents; instructions from process engineers to check and fine-tune low-confidence paragraphs are received; if physical conflicts are detected during pre-simulation, a semantic feedback signal is generated and fed back to the large model to trigger the correction and re-inference of constraint vectors, and the modified high-quality data is automatically fed back to the vector database to achieve self-learning evolution.

[0064] For example, when generating each text paragraph in the process guidance document, the confidence estimation head of the large model is invoked simultaneously, or the confidence score of that paragraph is calculated based on the token entropy value. A preset threshold (e.g., 0.8) is set, and paragraphs with confidence scores below the threshold are marked as high uncertainty areas. The output review report lists the location, confidence value, and potential risk type (e.g., missing parameters, logical jumps) of each low-confidence paragraph. In the user interface of the process guidance document, the corresponding paragraphs are visually marked using highlighting, underlining, or sidebar annotations. In response to the process engineer's editing operations on the marked paragraphs, the modified text content and modification trajectory (e.g., added, deleted, or modified character sequences) are captured. The engineer's fine-tuning instructions are recorded, including corrected process parameters, adjusted word order logic, or supplementary constraints. The corrected paragraphs are compared with the original low-confidence paragraphs to extract correction vectors. During the virtual assembly pre-simulation phase, if the physics engine (e.g., collision detection, interference checking) detects physical conflicts arising from the process sequence generated in step one (e.g., part penetration, inaccessibility), a semantic feedback signal is automatically generated. This signal includes the conflict type, conflicting part ID, conflict spatiotemporal coordinates, and the corresponding process step index. This signal is encoded as a natural language prompt or structured constraint vector and injected into the context window of the large model, forcing the model to re-perform logical reasoning and process generation under the condition of including this negative feedback. The paragraphs confirmed through manual fine-tuning and the successful reasoning cases corrected by physical feedback, along with their corresponding input conditions (e.g., part semantic information, assembly constraints), are encapsulated into new high-quality data samples. These samples are vectorized, their feature embeddings are extracted, and metadata tags (e.g., "expert verification," "physical simulation passed") are updated. The new vectors are written to a preset vector database, replacing or supplementing existing low-quality vectors and optimizing the database's distribution structure.

[0065] In this embodiment, a closed-loop quality control system for intelligent process generation is constructed through a four-dimensional linkage of confidence review, human-computer interaction, physical feedback, and data feedback. First, the confidence mechanism provides risk warning capabilities; second, manual fine-tuning injects domain expertise; third, physical feedback establishes objective truth standards; and finally, data feedback enables experience accumulation. These steps complement each other, upgrading the isolated "generation-output" process into a continuous evolutionary cycle of "generation-verification-correction-accumulation," significantly improving the accuracy, reliability, and adaptability of process guidance documents and the system's self-adaptability, providing a reliable software infrastructure for the large-scale implementation of intelligent manufacturing.

[0066] In some embodiments, it also includes: A multi-source knowledge fusion algorithm is used to weight and fuse the semantic representation vectors of knowledge from each source to reconstruct the optimal process description text; Parse the preset template, use regular expressions to traverse the text nodes in the preset template, identify the placeholders, and extract the strings containing type identifiers and name identifiers; Establish a mapping relationship between the process data to be filled and the positioning placeholders. Based on the mapping relationship, inject the process data in the optimal process description text into the corresponding position of the preset template, and render it according to the format attributes to generate a process guidance document.

[0067] For example, process-related knowledge from multiple data sources is acquired, including but not limited to design drawing parameters, historical process cards, simulation analysis reports, and material attribute databases. Pre-trained language models (such as BERT or RoBERTa) are used to transform unstructured text from each source into high-dimensional semantic representation vectors. A multi-source knowledge fusion algorithm is constructed, assigning dynamic weights to each vector based on the authority level, timeliness factor, and semantic relevance to the current task of each data source. The weighted vectors are aggregated using weighted summation or an attention mechanism to generate a fused feature vector. This fused vector is input into a decoder, and a bundle search strategy is used to generate grammatically correct, logically rigorous, and informationally complete process description text. A pre-defined process guidance document template (such as XML, HTML, or a Word document in a specific format) is loaded. All text nodes in the template document object model (DOM) tree are traversed. Syntax rules for locating placeholders (such as "${Type_Name}") are defined, and the corresponding regular expression patterns are compiled. A regular expression engine is used to match the content of each text node and identify strings that conform to the rules. The successfully matched strings are segmented and parsed to extract type identifiers (e.g., "Material", "Torque") and name identifiers (e.g., "Bolt_M8", "Step_3"), constructing a placeholder index table. Based on the extracted type and name identifiers, a search and match is performed in the generated optimal process description text and associated process data packages. A key-value mapping relationship between "placeholders" and "data content" is established. The corresponding rendering engine is called according to the type identifier (formatting "Date" type as "YYYY-MM-DD" and adjusting "Image" type to the specified resolution). The data content replaces the corresponding placeholders in the preset template, and predefined formatting attributes such as font, font size, color, and paragraph spacing are applied. Pagination, header and footer updates, and table of contents generation operations are performed, outputting the final process guidance document.

[0068] In this embodiment, a fully automated chain from knowledge integration to document delivery is constructed through the synergistic effect of multi-source knowledge fusion and intelligent template filling. First, multi-source fusion solves the "content quality" problem, ensuring the depth and breadth of process descriptions; second, template parsing and mapping filling solve the "form standardization" problem, ensuring the standardization and compliance of output documents. This breaks through the bottleneck of traditional process compilation, where content relies on manual editing and formatting relies on manual adjustment, achieving high quality, high efficiency, and high consistency in the generation of process guidance documents.

[0069] Please see Figure 2This document presents an embodiment of a file generation method based on spatial algorithms and large models, providing an overall flowchart. The file generation device based on spatial algorithms and large models generates entirely new process guidance documents, eliminating repetitive and time-consuming tasks for process engineers, such as manually aligning, matching, and inputting process descriptions based on the model. The entire process achieves intelligent, automated, and accurate output of process documents. This application consists of two modules: a historical data pre-training module for the large model, which uses high-quality historical process documents for pre-training before system startup to continuously increase the training volume and improve model matching accuracy; and a system-wide working module that automatically aligns the input 3D model with the existing BOM data, automating the entire process of outputting process guidance and process guidance documents.

[0070] Please see Figure 3 The following is a flowchart illustrating the application of a large model in a file generation method based on spatial algorithms and large models, provided in an embodiment of this application. High-quality existing process guidance documents are used for training to provide semantic generation guidance for subsequent semantic generation of large models.

[0071] Step 1: Historical Data Organization and Structured Input All provided historical data fields are organized and arranged, and key fields such as tools used, part numbers, tooling and accessories, and process descriptions are automatically extracted from the file information. This key information is then extracted into structured input, and each part is integrated into a complete structured input.

[0072] Step 2: Text Segmentation Because the volume of incoming historical data is enormous, long documents are cut into more precise text blocks based on contextual relationships and structured key information for accurate retrieval.

[0073] Step 3: Text Cleaning and Standardization The structured input information undergoes text cleaning to remove irrelevant information and formatting, retaining only the key information corresponding to each data point. This provides standardized input for the subsequent creation of the vector library. Noise and inconsistent errors are eliminated to provide high-quality input for the model.

[0074] Step 4: Vectorized Embedding and Vector Database Establishment The text blocks are transformed from their discrete symbolic form into continuous, semantically rich, high-dimensional vector representations. The large model provides each text block with a fixed-dimensional floating-point vector; the position of this vector in space represents the semantics of the text block. After vectorizing all text blocks, a vector database is built to efficiently store all related data. An indexing process is quickly constructed using approximate nearest neighbors, providing data for subsequent semantic output.

[0075] Step 5: Semantic Output of Different Functions Different inputs will generate different semantic outputs for the large model. If the input is unordered parts, the unordered parts will be reordered based on vector similarity and then output semantically. If the input is ordered parts with assembly relationships, the large model's language reasoning ability will be invoked based on the prompt word "engineering" to automatically generate descriptions of processes and steps.

[0076] Please see Figure 4 The following is a complete flowchart of a file generation method based on spatial algorithms and large models provided in an embodiment of this application. It processes newly input 3D models from users and automatically generates process operation instructions for the input models. The overall process and steps are as follows: Step 1: Data Input Following the system instructions, users input the address of the assembled model from the cloud or locally to obtain the corresponding 3D digital model information and connect to the BOM management system to retrieve the BOM information matching the digital model. If access to the BOM management system is not possible, a BOM table matching the current digital model is created.

[0077] Step 2: Model Transformation and Data Alignment To adapt to current mainstream industrial software, this system needs to convert all input digital model files into a unified data structure for calculation and rendering within the system. The converted digital model files must be aligned with the BOM management system, associating the digital model in each node with each row of data in the BOM to achieve a precise one-to-one match and obtain all attribute information from the BOM.

[0078] Step 3: Model Node Classification and Matching The converted model nodes are classified according to the BOM information and the part type. The classified models are then matched with each row of data in the BOM to obtain all the BOM information.

[0079] Step 4: Semantic Arrangement Order of the Large Model The semantics of the categorized unordered parts are used as input. A pre-trained large model is used to provide an initial order for the unordered parts based on an approximate search of the model vector library. The initial order depends on the order of the process documents that exist in history.

[0080] Step 5: Spatial geometry algorithm reordering The large model is input in the initialization order, and a recursive spatial geometry algorithm is performed on all parts. The purpose of the initial input is to reduce the computational load of the geometry algorithm. The spatial geometry algorithm performs virtualized movement and collision detection based on the part order to determine the rationality of the parts' installation in a specific order, and calculates the explosion direction of the parts, providing directional guidance for the generation of the exploded diagram. Based on interference detection and feasibility analysis, all parts are reordered.

[0081] For example, based on the large model sorting results as prior input for spatial collision checking, movement collision detection and reachability analysis are performed on each 3D model in the sorting. The order is then rearranged through a dynamic adjustment algorithm to achieve a precise output of the model installation order. Compared to traditional methods that rely solely on large model generation, this precise geometric algorithm provides the most accurate results.

[0082] Collision Detection Algorithm: Combining 3D geometric topology information, spatial analysis is performed on the feature region. An improved collision detection algorithm is used to perform spatial interaction analysis on potential mating features. By detecting the containment relationship, gap distance, and contact area between part features, it identifies whether the part's position in the sequence will cause interference.

[0083] Movement direction prediction algorithm: The algorithm obtains contact surface information through collision detection, calculates the normals based on principal component analysis and planar weighted clustering, calculates all possible movement directions under all contact surfaces, and ranks and filters them according to weights, filtering out obviously abnormal directions and retaining the movement direction information.

[0084] Continuous movement collision detection algorithm: Continuous movement collision detection is a key step in virtual simulation analysis. It is used to determine whether there is a collision interference relationship with other parts when the movement direction and path are determined, so as to ensure the rationality of the installation sequence.

[0085] For high-priority movement directions, route planning is performed. Based on the initial position and direction, a movable path is calculated. It is then determined whether collision information is generated under the specified movement direction and path. If no collision information is generated, it means that the target parts in the current installation sequence can be automatically assembled. If collision information is generated, the installation sequence of the target parts needs to be rearranged.

[0086] Rational sorting algorithm: The optimal sorting algorithm optimizes the rational sorting of parts based on the results of continuous collision detection and interference detection, as well as the constraints of tool usage and workstation conditions. It performs virtual movement collision detection on the movement path and direction of the target part, that is, it constructs a spatial movement physical model, judges the interference relationship between the target part and the preceding part, and determines the optimal sorting position of the target part without interference during the virtual movement based on the interference judgment and some additional input reachability constraints.

[0087] The spatial geometry algorithm module is the core component of this application. It is responsible for accurately calculating and generating a reasonable part assembly sequence based on the part's 3D geometric model and spatial constraints, providing crucial structured input for subsequent process instruction generation. This module deeply integrates typical assembly sequences summarized from historical assembly experience and dynamically corrects sequence errors caused by physical deviations, changes in working conditions, or assembly complexity in historical experience through geometric analysis and interference verification of the actual 3D model. This achieves precise optimization and logical rationalization of the assembly sequence, significantly improving the feasibility and correctness of the generated process.

[0088] Step Six: Description of the Semantic Generation Process for Large Models The precise part assembly sequence calculated by spatial geometry algorithms and the corresponding part attributes in the BOM information are used as input prompts. The semantic reasoning capabilities of the large model and the semantic block information formed by the multi-element fusion algorithm are used to automatically generate semantics that are contextually coherent, accurate and consistent with the process document description.

[0089] For example, vectorized feature extraction is performed on the description of process documents in history, and the corresponding words are segmented and clustered. Then, the structured semantic prompts are input into the large model based on the precise order of the sorted parts and the process parameters (auxiliary material requirements, assembly environment, torque requirements, etc.) provided by the context, and the corresponding natural language description of the process is automatically generated.

[0090] Structured semantic input: Spatial geometry algorithms calculate the precise assembly sequence of parts, extract the corresponding process parameters from the BOM data, and combine them into structured semantic prompts, which serve as the input to the large model.

[0091] Multi-source knowledge fusion and logical reasoning: In the process of semantic generation, large models need to integrate multi-dimensional information, including semantic similarity, vector usage frequency, vector generation time, etc., and calculate the optimal vectorized semantics to generate semantic information with the highest description accuracy.

[0092] Dynamic adaptation and context-coherent generation: When generating text semantics, the large model based on the Transformer structure dynamically adjusts the description style and the level of detail of the generated text according to the context of the current task and the generated structure, and automatically optimizes the generation of fluent and unambiguous process paragraphs.

[0093] The large-scale model semantic generation module employs multi-dimensional semantic reasoning and dynamic context adaptation mechanisms. It can deeply integrate high-quality descriptions from historical process data, automatically extract high-frequency and highly relevant semantic fragments, and generate them coherently based on the current task context. Through joint optimization of semantic similarity and process specification compliance, this module ensures that the output process text not only conforms to industry standards but also possesses good readability and executability, significantly improving the accuracy and applicability of process document generation.

[0094] Step 7: Optimal viewpoint generation and exploded view generation Based on the assembly orientation of the parts calculated using spatial geometry algorithms, the optimal viewing angle for each part is automatically calculated. Based on the generated optimal viewing angle and the size of the parts, the explosion distance for each part is adaptively generated. Using an optimization algorithm, an exploded view with minimal occlusion and optimal viewing angle for each part is captured.

[0095] For example, based on automatically generated process step sequences and movement trajectory predictions derived from spatial algorithms, this application can automatically generate exploded views required for SOP (Standard Operating Procedure) work instructions. This module dynamically selects the optimal visualization perspective based on the dimensional attributes and spatial relationships of each component, thereby generating exploded views that are clear, logically laid out, and conform to assembly logic, effectively improving the intuitiveness and operability of the work instructions.

[0096] Dynamic viewpoint selection algorithm: Based on the input part movement direction, the algorithm automatically calculates the distance relationship between the viewpoint and the part, generates virtual cameras at different angles around the viewpoint, and obtains the optimal viewpoint selection direction based on the projection characteristics and collision interference relationship at different camera points. Then, based on the size of the part, it determines the reasonable distance relationship between the viewpoint and the part position.

[0097] Exploded view generation algorithm: Based on the optimal viewpoint and spatial calculation results, automatically calculate the movement transformation matrix of the part, match the movement action frames of the part, and automatically generate the explosion visual effect of the part.

[0098] The optimal viewing angle is determined based on dynamic perspective selection. This angle displays the most complete installation state of the part, minimizes occlusion variables, and provides the optimal viewing distance. An exploded view conforming to assembly logic is generated from the exploded diagram, providing the most intuitive action guidance for operations and instructions during the process document generation.

[0099] Step 8: Generation of process instruction documents Based on different operational requirements, templates for Standard Operating Procedures (SOPs) are generated. Utilizing model information, semantically generated process descriptions and exploded views of each part are inserted into the model, completing the insertion of each data point in the template. Finally, a complete SOP is generated and saved in different output formats.

[0100] Step Nine: Manual Verification Stage The process instruction document generation process also automatically generates a corresponding review report, which describes the accuracy of the semantic generation of the large model. Users can check the documents with lower accuracy in the work instructions based on the review report to see if they need to be modified, and then output the modified documents.

[0101] This application employs a file generation method based on spatial algorithms and large models, which has the following technical advantages: First, it achieves data normalization processing for mainstream 3D design software, converting different input data into a self-developed data structure type and performing corresponding structural optimization and data compression. At runtime, it can run independently as a plugin for 3D design software, without relying on other additional platforms, significantly reducing operating costs.

[0102] Secondly, instead of relying on single historical or model inputs, it uses historical experience as a priori guidance for 3D model calculations. This improves the accuracy of spatial geometric calculations of the 3D model and provides accurate input for preserving and generating historical experience. It ensures accurate results from empirical data and accurate process guidance for generating new input data, providing precise assembly sequence guidance and assembly direction generation.

[0103] Third, the 3D model is aligned with the existing BOM in the system to ensure the absolute accuracy of the installation materials and other information in each piece of data. It also ensures that each piece of data is aligned with the data in the system management version, thus ensuring data consistency throughout the entire process.

[0104] Fourth, a computational logic is employed that combines large-scale model reasoning with spatial geometric algorithms as the primary method. The initial 3D model assembly sequence is derived from the large-scale model, and then recursively calculated sequentially using spatial geometric algorithms. Interference detection analyzes the rationality of the assembly sequence, and a rationalized sorting is used to output the optimal assembly result. This approach ensures that the large-scale model can extract historical experience, reducing the complexity of spatial calculations, while also obtaining the most accurate part assembly sequence through precise calculations. This optimizes both computational speed and result accuracy.

[0105] Fifth, relying on the results of spatial geometric algorithms, the most accurate parts installation sequence was obtained. By adopting a multi-source knowledge fusion algorithm, the optimal vectorized semantics were calculated by weighting semantic similarity, vector usage frequency, vector generation time, etc., to generate semantic information with the highest description accuracy.

[0106] Sixth, 3D models were incorporated for calculation and simulation. During spatial geometry algorithms for calculation and interference checks, the correct installation sequence and explosion direction for each part were obtained. Dynamic optimal viewpoint generation was employed, adaptively generating the explosion distance and viewpoint position based on the part's size. Through occlusion detection and ray tracing, the optimal viewpoint for each part was selected, minimizing occlusion and displaying the most complete view for operational guidance.

[0107] Please refer to Figure 5 This embodiment of the application also provides a structural block diagram of a file generation device based on spatial algorithms and large models, including: The acquisition module 510 is used to acquire the three-dimensional model and bill of materials of the target. The preprocessing module 520 is used to convert the format of the 3D model and align it with the bill of materials to generate semantic material information for each part in the 3D model. The semantic sorting module 530 is used to input the semantic material information of each part into a preset large model for semantic sorting and to generate an initial process sequence. The sequence verification module 540 is used to inspect each part in the initial process sequence based on a spatial algorithm, and adjust the sorting according to the verification results to determine the assembly sequence and explosion direction of the parts. The inspection includes collision detection and movement direction prediction. The view generation module 550 generates an exploded view with the optimal perspective based on the explosion direction. The optimal perspective for each part is calculated based on the assembly sequence of the parts. The process description module 560 is used to input the assembly sequence of parts into a preset large model for semantic reasoning to determine the semantic process description of each part. The document generation module 570 is used to insert the semantic process description and exploded view of each part into a preset template to generate a process guidance document.

[0108] It should be noted that the file generation apparatus based on spatial algorithms and large models provided in the above embodiments belongs to the same concept as the file generation method based on spatial algorithms and large models. Specifically, the acquisition module 510 executes step S110, the preprocessing module 520 executes step S120, the semantic sorting module 530 executes step S130, the sequence verification module 540 executes step S140, the view generation module 550 executes step S150, the process description module 560 executes step S160, and the file generation module 570 executes step S170. The specific methods of execution of each module have been described in detail in the method embodiments and will not be repeated here. In practical applications, the file generation method based on spatial algorithms and large models provided in the above embodiments can allocate the above functions to different functional modules as needed, that is, divide the internal structure of the system into different functional modules to complete all or part of the functions described above. This is not a limitation here.

[0109] The above methods are used to obtain the target's 3D model and bill of materials. After format conversion, the 3D model is aligned with the bill of materials to generate semantic material information for each part in the 3D model, establishing relationships between geometry, attributes, and functions. The semantic material information of each part is input into a pre-defined large model for semantic sorting, generating an initial process sequence. Each part in the initial process sequence is checked using spatial algorithms, and the sorting is adjusted based on the check results to determine the part assembly sequence and explosion direction. The check includes collision detection and movement direction prediction. Based on the explosion direction, an exploded view with the optimal perspective is generated to ensure the integrity and clarity of information transmission, improving the readability and guidance of the drawings. The part assembly sequence is input into the pre-defined large model for semantic reasoning to determine the semantic process description for each part. The semantic process description and exploded view of each part are inserted into a pre-defined template to generate a process guidance document. The semantic process description, combined with assembly paths and spatial relationships, forms an executable and traceable process language, which is finally embedded into the pre-defined template to form a standardized and reusable process guidance document. The large model provides a priori coarse sorting based on historical experience, which greatly reduces the complexity of spatial calculations; the spatial geometry algorithm solves the physical illusion problem of the large model, ensuring the absolute physical feasibility of the assembly sequence, and automatically realizes the accurate generation of SOP files with illustrations from the 3D model.

[0110] In some embodiments, an electronic device is also provided, which may be a server, and its internal structure diagram is shown below. Figure 6As shown, the electronic device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. The computer program is executed by the processor to implement the functions or steps of the server-side method described above.

[0111] In some embodiments, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the methods described above.

[0112] In some embodiments, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0113] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or electronic device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0114] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation that may be implemented in the methods and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0115] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A file generation method based on spatial algorithms and large models, characterized in that, include: Obtain the target's 3D model and bill of materials; After converting the format of the 3D model, it is aligned with the bill of materials to generate semantic material information for each part in the 3D model. The semantic material information of each part is input into a preset large model for semantic sorting to generate an initial process sequence; Each part in the initial process sequence is inspected based on a spatial algorithm, and the sorting is adjusted according to the inspection results to determine the assembly sequence and explosion direction of the parts. The inspection includes collision detection and movement direction prediction. Based on the explosion direction, an exploded view with an optimal perspective is generated as the exploded view, wherein the optimal perspective for each part is calculated based on the assembly sequence of the parts; The assembly sequence of the parts is input into the preset large model for semantic reasoning to determine the semantic process description of each part; The semantic process description and the exploded view of each part are inserted into a preset template to generate a process guidance document.

2. The file generation method based on spatial algorithms and large models according to claim 1, characterized in that, After converting the format of the 3D model, it is aligned with the bill of materials to generate semantic material information for each part in the 3D model, including: The 3D model is converted to a new format to obtain a model structure tree. Each node in the model structure tree is aligned with the bill of materials so that each node maps to at least one attribute in the bill of materials, and each node represents at least one part; The unordered parts in the model structure tree are classified according to the bill of materials type to obtain the semantic material information of each part in the 3D model.

3. The file generation method based on spatial algorithms and large models according to claim 1, characterized in that, The semantic material information of each part is input into a preset large model for semantic sorting to generate an initial process sequence, including: The semantic material information of each part is input into a preset large model for semantic information extraction to determine the semantic information; The semantic information is searched for cosine similarity in a preset vector database to determine multiple similar process vectors, wherein the preset vector database includes a historical process vector database. The multiple similar process vectors are weighted and sorted according to semantic similarity, time popularity and usage frequency using the preset large model, and the vector sorting results are arranged from high to low as the initial process order.

4. The file generation method based on spatial algorithms and large models according to claim 1, characterized in that, Each part in the initial process sequence is inspected based on a spatial algorithm, and the sorting is adjusted according to the inspection results to determine the assembly sequence and explosion direction of the parts, including: Based on the geometric topology information of each part in the three-dimensional model in the initial process sequence, the potential features of each part are spatially analyzed based on the virtual collision detection algorithm to generate collision results; For each part in the initial process sequence, the movement direction is predicted to determine the movement direction information of each part; The highest priority movement direction for each part in the movement direction information is selected for path planning. The initial direction and initial position in the movement direction information are combined to determine the installation sequence of the parts. Based on the collision results, it is determined whether there is an interference relationship between each part and its adjacent parts, and the explosion direction is determined according to the interference relationship corresponding to the installation order of the parts.

5. The file generation method based on spatial algorithms and large models according to claim 4, characterized in that, For each part in the initial process sequence, the movement direction is predicted to determine the movement direction information of each part, including: Determine the contact surface information of each part in the three-dimensional model; Based on the contact surface information of each part, the normal is calculated according to principal component analysis and planar weighted clustering to determine all movable directions under all contact surfaces of each part; The movement direction information of each part is determined by ranking all movable directions under all contact surfaces according to preset weights and filtering based on the ranking results.

6. The file generation method based on spatial algorithms and large models according to claim 4, characterized in that, Path planning is performed by selecting the highest priority movement direction for each part from the movement direction information. Combining this with the initial direction and initial position from the movement direction information, the part installation sequence is determined, including: Obtain all movement directions information for each part in the 3D model; Filter all movement direction information for each part and determine the highest priority movement direction; The highest priority direction of movement is selected as the dominant direction. Based on the initial direction and initial position in the movement direction information, the assembly motion path of each part is determined. The assembly sequence of parts is determined based on the spatial constraints between the assembly motion paths.

7. The file generation method based on spatial algorithms and large models according to claim 1, characterized in that, Based on the explosion direction, an exploded view with an optimal perspective is generated as the exploded view, including: Based on the set of surface normals of the 3D model containing all parts, the vertices of the 3D model are transformed to the 2D clipping space according to the three-view projection to obtain the projection results of the parts in three orthogonal directions. The part contour is extracted based on the projection results from multiple perspectives, and the optimal projection perspective is determined. Under the optimal perspective, the segmentation map of the initial image containing all parts is extracted by the panoramic segmentation network. The appearance features are extracted layer by layer and the channels are connected to form a conditional signal input into the preset diffusion model. The explosion distance is adaptively calculated by combining the translation transformation matrix to generate an unobstructed and clear explosion view.

8. The file generation method based on spatial algorithms and large models according to any one of claims 1 to 7, characterized in that, Also includes: When outputting process guidance documents, a review report is output simultaneously, and paragraphs with confidence levels below a preset threshold are marked in the process guidance documents. Receive instructions from process engineers to check and fine-tune low-confidence segments; Furthermore, if a physical conflict is detected during the pre-simulation, a semantic feedback signal is generated and fed back to the large model to trigger the correction and re-inference of the constraint vectors. The modified high-quality data is automatically fed back to the vector database to achieve self-learning evolution.

9. The file generation method based on spatial algorithms and large models according to any one of claims 1 to 7, characterized in that, Also includes: A multi-source knowledge fusion algorithm is used to weight and fuse the semantic representation vectors of knowledge from each source to reconstruct the optimal process description text; Parse the preset template, use regular expressions to traverse the text nodes in the preset template, identify the location placeholders, and extract the strings containing type identifiers and name identifiers; Establish a mapping relationship between the process data to be filled and the positioning placeholders. Based on the mapping relationship, inject the process data in the optimal process description text into the corresponding position of the preset template, and render it according to the format attributes to generate a process guidance document.

10. A file generation device based on spatial algorithms and large models, characterized in that, include: The acquisition module is used to acquire the target's 3D model and bill of materials. The preprocessing module is used to convert the format of the 3D model and align it with the bill of materials to generate semantic material information for each part in the 3D model. The semantic sorting module is used to input the semantic material information of each part into a preset large model for semantic sorting and to generate an initial process sequence. The sequence verification module is used to verify each part in the initial process sequence based on a spatial algorithm, and adjust the sorting according to the verification results to determine the assembly sequence and explosion direction of the parts. The verification includes collision detection and movement direction prediction. The view generation module generates an exploded view with the optimal perspective based on the explosion direction, wherein the optimal perspective for each part is calculated based on the assembly order of the parts; The process description module is used to input the assembly sequence of the parts into the preset large model for semantic reasoning to determine the semantic process description of each part. The document generation module is used to insert the semantic process description and the exploded view of each part into a preset template to generate a process guidance document.

Images