2d drawing to 3d drawing ai self-generation system

Abstract

The application relates to the field of engineering drawing self-generation, and specifically discloses an engineering 2D drawing to 3D drawing AI self-generation system. First, the input two-dimensional engineering drawing is subjected to deep format analysis and semantic feature extraction to be converted into structured engineering data. Then, relying on a knowledge graph reasoning engine, spatial logic reasoning and network flow direction reconstruction are performed on the structured data. Next, the system obtains parameterized assembly data through feature comparison and assembly rule mapping, and then drives the coordinate anchoring and automatic routing assembly of entity pipelines and equipment components in three-dimensional space to construct a high-fidelity three-dimensional engineering model. Finally, using a reverse geometric projection and attribute mapping algorithm, material features are directly extracted based on the generated three-dimensional model and view dimension reduction is performed, and engineering associated data including processing files and quantity and price lists are synchronously output, thereby eliminating the manual errors and data island phenomenon in the traditional method.

Description

Technical Field

[0001] This application relates to the field of self-generated engineering drawings, and more specifically, to an AI self-generated system for converting 2D engineering drawings into 3D drawings. Background Technology

[0002] In traditional engineering design and construction fields (such as water treatment, chemical engineering, and industrial pipelines), the operational flow of design and construction typically follows a core workflow from initial two-dimensional planar design to detailed three-dimensional modeling, and then to the export of engineering drawings and bills of materials. The three-dimensional engineering model plays a crucial role in this process; it is not only a carrier that intuitively displays the spatial layout and pipeline routing, but also the fundamental digital foundation for subsequent pipeline collision detection, quantity surveying, and material cost analysis. With the increasing complexity of engineering dimensions, the numerous professional symbols, implicit topological relationships, and system design intentions contained in two-dimensional drawings are difficult to extract quickly. How to efficiently, accurately, and automatically convert highly abstract, planar two-dimensional engineering drawings into three-dimensional digital models that can guide actual construction operations, and simultaneously correlate and derive associated engineering data to achieve seamless data flow throughout the entire design-quantity survey-cost estimation cycle, has become a pressing technical challenge in the intersection of engineering CAD and artificial intelligence. Therefore, constructing an AI-driven self-generating solution for converting 2D engineering drawings to 3D drawings is an inevitable requirement for promoting the end-to-end automation and intelligent upgrading of modern engineering design.

[0003] However, existing solutions for converting 2D engineering drawings to 3D models are mainly limited to manual modeling or semi-automated processes relying on shallow rules, resulting in several insurmountable technical flaws. First, the most common purely manual 3D modeling method requires designers to manually create pipe and equipment entities entirely based on 2D coordinates and flowcharts. This method is not only time-consuming but also highly dependent on the designer's individual experience, easily leading to spatial misalignment and loss of intent due to visual fatigue or subjective misinterpretation. Second, while some parametric and rule-driven modeling software incorporates standard libraries, it still requires manual intervention to extract drawing features and manually input core parameters such as pipe diameter and elevation. The underlying system cannot automatically recognize the complex pipe network flow patterns and spatial layout features in native 2D drawings, failing to achieve end-to-end direct connection from engineering drawings to 3D data. Furthermore, although existing AI technology is relatively common in general image recognition, general algorithms often struggle to deeply analyze industry-specific primitive dictionaries and simplified specifications in engineering drawings, and lack the ability to transform structured graphic features into knowledge network graphs for spatial logical reasoning. Meanwhile, under the existing system, once a 3D model is established, its material quantity statistics and associated cost budget still rely on a separate and cumbersome semi-manual accounting process, resulting in a serious disconnect between model construction and cost control.

[0004] Therefore, existing technologies urgently need an intelligent coordination system that can deeply analyze the semantics of drawings, autonomously reason about spatial topology rules, complete the parameterized assembly of entities, and generate quantity and price accounting lists in sync. Summary of the Invention

[0005] To address the difficulties in dynamic connection identification and the complexity of topology graph generation in existing technologies, this application is proposed. According to this application, an AI self-generating system for converting 2D engineering drawings to 3D drawings includes: an engineering drawing structuring module for parsing the format and extracting semantic features from 2D engineering drawings to obtain structured engineering data; an engineering data reasoning and reconstruction module for performing spatial logic reasoning and network flow reconstruction on the structured engineering data based on a knowledge graph reasoning engine to obtain spatial topology data; a data parameterization assembly module for performing feature comparison and assembly rule mapping on the structured engineering data to obtain parameterized assembly data; a 3D engineering model construction module for performing coordinate positioning and entity pipeline routing assembly on the spatial topology data and parameterized assembly data to obtain a 3D engineering model; and an engineering-related data generation module for performing view dimensionality reduction calculations and material feature extraction on the 3D engineering model based on reverse geometric projection and attribute mapping algorithms to generate engineering-related data.

[0006] Compared with existing technologies, this application provides an AI-driven self-generating system for converting 2D engineering drawings to 3D models. This system addresses technical issues such as over-reliance on manual drawing interpretation, difficulties in extracting complex topological relationships, and a severe disconnect between model construction and subsequent quantity surveying and cost estimation during the conversion from 2D engineering drawings to 3D models. First, it performs deep format parsing and semantic feature extraction on the input 2D engineering drawings, automatically converting high-dimensional graphic symbols into machine-readable structured engineering data, thus overcoming the efficiency and accuracy bottlenecks of manual drawing interpretation at the source. Then, relying on a knowledge graph inference engine, it performs spatial logic reasoning and network flow reconstruction on the structured data, accurately restoring the 3D spatial topological relationships of the pipeline network. Next, the system obtains parametric assembly data through feature comparison and assembly rule mapping, thereby driving the physical pipelines and equipment components to complete coordinate anchoring and automated routing assembly in 3D space, efficiently constructing a high-fidelity 3D engineering model. Finally, by using reverse geometric projection and attribute mapping algorithms, material features are directly extracted from the generated 3D model and the view is reduced in dimensionality. Simultaneously, engineering-related data containing processing documents and quantity and price lists are output, thereby realizing the integrated connection from drawing analysis, topology reasoning, 3D reconstruction to engineering cost, effectively eliminating human error and data silos in traditional methods. Attached Figure Description

[0007] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0008] Figure 1 This is a block diagram of an AI self-generating system for converting 2D engineering drawings to 3D drawings according to an embodiment of this application.

[0009] Figure 2 This is a schematic diagram of the data flow in an AI self-generating system for converting 2D engineering drawings to 3D drawings according to an embodiment of this application.

[0010] Figure 3 This is a block diagram of the engineering data reasoning and reconstruction module in the AI ​​self-generating system for converting 2D engineering drawings to 3D drawings according to an embodiment of this application.

[0011] Figure 4 This is a schematic diagram of the data flow in the AI ​​self-generating system for converting 2D engineering drawings to 3D drawings according to an embodiment of this application, specifically the module for generating engineering-related data. Detailed Implementation

[0012] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0013] It should be understood that the steps described in the method embodiments of this disclosure may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of this disclosure is not limited in this respect.

[0014] This application is made in response to the aforementioned deficiencies in the existing technology. Figure 1 and Figure 2 As shown, Figure 1 This is a block diagram of an AI self-generating system for converting 2D engineering drawings to 3D drawings according to an embodiment of this application. Figure 2This is a schematic diagram of the data flow in the AI ​​self-generating system for converting 2D engineering drawings to 3D drawings according to an embodiment of this application. In the overall system architecture, to ensure efficient collaboration among multiple modules, the system's front end also deploys a project library module and a main control panel module. The project library module, serving as a unified data entry point and management center, is responsible for creating design projects, centrally importing system floor plans and PID drawings in DWG or PDF format, and securely storing the final generated 3D model and all associated output files. The main control panel module, acting as the user interaction and control hub, provides an interface for initiating 2D-to-3D conversion tasks and dynamically configuring conversion parameters, and includes built-in automatic verification and compliance review tools for users to centrally review the AI's generated results. Under the scheduling of the aforementioned front-end interaction architecture...

[0015] Specifically, the engineering drawing structuring module 110 is used to parse the format and extract semantic features of two-dimensional engineering drawings to obtain structured engineering data. It is understandable that early records of engineering design projects typically exist in the form of lines, symbols, and fragmented text annotations on various planes. While the original two-dimensional drawings visually show the pipe routes and preliminary equipment layouts, the underlying computing environment initially treats them as discrete coordinate points or geometric patterns without specific business attributes. The computing side itself cannot directly use two-dimensional projection to determine whether specific line segments in the drawing correspond to actual water supply pipe components, nor can it overcome the barriers of two-dimensional space to generate three-dimensional entity attributes containing parameters such as material. Given the underlying data barrier between the discrete state of these original primitives and the target 3D spatial computational reconstruction, this module is introduced at the very beginning of the overall process to perform in-depth format decryption and reshaping of the incoming 2D engineering drawings from the underlying data flow dimension. It transforms the simple visual line structure into related data features with physical and topological significance, breaks down and reassembles the original chaotic set of lines, and purifies it into structured engineering data that can be recognized by specific logic and used in subsequent calculations. This provides a high-value unified digital foundation for downstream execution of complex spatial logic reasoning and autonomous spatial mounting of 3D specimens.

[0016] In one exemplary embodiment, the engineering drawing structuring module 110 includes: an engineering drawing separation and mapping unit, used to perform layer separation and geometric vectorization mapping on two-dimensional engineering drawings based on a format parsing engine to obtain a basic set of graphic elements; a semantic feature recognition and association unit, used to perform semantic feature recognition and association on the basic set of graphic elements based on an AI model to obtain an independent set of graphic and textual features; and a feature decoding and encapsulation unit, used to perform entity category decoding and tree-structured protocol serialization encapsulation on the independent set of graphic and textual features through an engineering specification dictionary mapping and matching mechanism to obtain structured engineering data.

[0017] The relevant operational details are as follows: The engineering drawing separation and mapping unit receives externally imported two-dimensional engineering drawings as processing input. These two-dimensional engineering drawings contain key drawings of the project's piping system design, such as system layout plans and piping and instrumentation diagrams (PID diagrams). Their file format can be the DWG format commonly used by design institutes or the PDF format submitted for review. The two-dimensional engineering drawings are obtained as a file stream through methods such as user interface uploads or application programming interface calls. Upon receiving the file, the format parsing engine within the engineering drawing separation and mapping unit starts working. This format parsing engine is designed based on a plug-in architecture, integrating decoders for different file standards, such as a binary digital decoder for DWG files and a document object model parser for PDF files. The format parsing engine first reads and parses the underlying binary data stream of the two-dimensional engineering drawing. Taking a drawing named "Pump Station Plan.dwg" as an example, this drawing may contain multiple layers such as piping layers, equipment layers, annotation layers, and grid layers. The format parsing engine's layer separation function traverses the file's layer definition table, stripping away layers unrelated to core engineering entities, such as frame layers and annotation layers, retaining only layer data related to equipment, pipes, valves, and instruments, thus initially filtering out a large amount of background and non-critical information. Subsequently, the geometric vectorization mapping function is activated, parsing each graphic object in the retained layers. For vector formats like DWG, the engine directly extracts the mathematical parameters of defining primitives such as line segments, circles, arcs, and polylines. For example, a straight line representing a pipe will be extracted to show the coordinates of its two endpoints in paper space, such as P1(1500,2000) and P2(4500,2000), as well as its line type, color, and other attributes. Simultaneously, for text in the drawing, such as the annotation DN200, the engine extracts the text content DN200, its coordinate position, font size, and rotation angle, and calculates a rectangular area tightly enclosing the text, i.e., the text bounding box. After extracting all necessary graphic elements and text, to eliminate dimensional differences caused by different designers and projects using different drawing scales (e.g., 1:50 or 1:100), the engineering drawing separation and mapping unit performs a global coordinate system normalization mapping on all extracted coordinate data. This process is accomplished using the following linear transformation formula: ,in, Represents the new coordinates after normalization; These are the original coordinates extracted from the drawing file; This is a 2×2 scaling matrix, the values ​​of which are determined by the drawing scale; for example, for a 1:100 drawing, Set as The scale information can be automatically read from the drawing's metadata or determined according to preset configurations (such as processing all input drawings at 1:100 by default). This is a translation vector used to align the origin of the drawing's coordinate system to the origin of the normalized coordinate system. This vector value can be set to the coordinates of the lower left corner of the bounded area of ​​the drawing. This formula, through scaling and translation transformations, maps all primitives based on arbitrary scales and coordinate origins to a standard, dimensionless coordinate space. Taking the aforementioned endpoint of a straight line with coordinates P1(1500, 2000) as an example, at a 1:100 scale, if... If (0,0), then its normalized coordinates are... This will become (15, 20). After the above layer separation, vectorization mapping, and coordinate normalization processes, the engineering drawing separation mapping unit finally outputs a structured set of basic primitives. This set is a data list containing multiple primitive objects, where each object clearly records its type, such as line, arc, text, normalized geometric parameters, such as the start and end coordinates of the line segment, the position of the text, and its original attributes, such as layer name, color, etc.

[0018] The semantic feature recognition and association unit initiates its core processing flow. The core objective of this unit is to simulate the drawing interpretation logic of designers, intelligently and accurately associating graphic symbols, such as the icon of a pump, with nearby text describing its attributes, such as its tag number P-101 or model number. To achieve this goal, the semantic feature recognition and association unit incorporates a specially trained artificial intelligence recognition model. This model employs a transformer-based visual-text multimodal architecture. Regarding the deployment and implementation of the underlying core engine, this drawing understanding engine and subsequent computing engines can be directly integrated into the local terminal or deployed in the cloud to form a SaaS service model based on cloud computing power. Furthermore, in selecting the AI ​​model, in addition to using dedicated models trained in a closed loop using proprietary engineering data, the system can also adopt a technical approach of combining large models with professional domain data for fine-tuning, or jointly deploy multiple dedicated visual models, such as those specifically for recognizing pipes and equipment symbols, to achieve seamless workflow collaboration. This architecture includes a visual encoder branch and a text encoder branch. The model's training process utilizes a massive dataset of precisely annotated engineering drawings, where each graphic symbol is linked to its corresponding text annotation. Supervised learning is performed on this dataset, employing optimization methods such as contrastive learning loss functions. This allows for adjustments to the model's internal weights and biases until it accurately maps the visual features of the graphics to the semantic meaning of the text and determines the correlation between the two. When the basic set of primitives is fed into the AI ​​model, the graphic elements within the set (consisting of line segments, arcs, etc.) are fed into the model's visual encoder branch. The specific architecture of this visual encoder branch begins with a primitive embedding layer, which transforms the input sequence of discrete geometric primitives into a vector sequence that the model can process. Specifically, each geometric primitive, such as a line segment, is defined by its type (line segment, arc, etc.) and key geometric parameters (such as the normalized start and end coordinates of the line segment). This is then transformed into a fixed-dimensional initial feature vector through a learnable linear mapping layer. To enable the model to understand the relative spatial arrangement of these primitives, a two-dimensional spatial position encoding is calculated and injected into the feature vector of each primitive. This location encoding is generated based on the coordinates of the primitive's geometric center, ensuring that the model not only knows what a primitive is, but also where it is. This vector sequence carrying location information is then fed into a stacked multi-layer transformer encoder module. Each encoder module consists of a multi-head self-attention mechanism sublayer and a feedforward neural network sublayer. The multi-head self-attention mechanism can compute the association weights between primitives in parallel, allowing the model to simultaneously focus on the global and local geometric relationships constituting the symbol from different perspectives. For example, when recognizing the symbol of a pump, the attention mechanism can capture the specific spatial configuration between its main circular body and the outlet triangle.The feedforward neural network performs a nonlinear transformation on the features output by the self-attention layer, further deepening the feature representation. This branch, through the self-attention mechanism, captures the internal relationships between the various geometric parts that constitute the symbol, and ultimately outputs a high-dimensional graphical boundary feature vector that can represent the essential form of the graphic symbol. For example, a centrifugal pump symbol composed of circles and triangles is encoded as a 512-dimensional specific floating-point vector. Meanwhile, text elements from the basic primitive set, such as P-101, are fed into the model's text encoder branch. This text encoder branch follows the architecture of established natural language processing models. The input text string is first processed by a tokenizer, which segments the string into a series of tokens based on a pre-built vocabulary containing a large number of engineering domain terms. For example, P-101 might be segmented into two tokens: P- and 101. Each token is mapped to a high-dimensional word embedding vector through an embedding layer; this embedding vector is a numerical representation of the token's semantics learned by the model during training. Similar to the vision branch, to allow the model to understand the text's sequential information, a one-dimensional positional encoding vector is added to each token's embedding vector. The sequence of token vectors carrying positional information is then fed into a multi-layer transformer encoder stack similar in structure to the vision branch. The multi-head self-attention mechanism enables the model to capture contextual dependencies within the text. For example, when processing DN200, the model understands that DN and 200 together refer to the specific engineering concept of a nominal diameter of 200 millimeters. This branch leverages its language understanding capabilities, pre-trained on a large corpus, to encode text strings into a text semantic feature vector of the same dimension, containing its semantic information. .

[0019] After obtaining the feature vectors of all graphics and text in the image, the semantic feature recognition association unit employs a spatial-semantic joint attention mechanism to calculate the association score between any graphic and any text. This process is based on spatial proximity pruning, scoring only graphic-text pairs within a certain distance threshold, for example, set to 5% of the diagonal length of the drawing. The calculation follows the formula: In this formula, This represents the Euclidean distance between the geometric center of the graphic symbol and the center of the text bounding box. This value is calculated directly from the coordinate information of the underlying primitive set. For example, if the center coordinates of a pump symbol are (35.2, 58.4), and the center coordinates of a nearby text P-101 are (38.6, 59.1), then... The value is the straight-line distance between the two. It is a preset distance attenuation coefficient, for example, set to 0.1. It is used to adjust the influence weight of spatial distance in the correlation calculation. The farther the distance, the closer the value of the index term is to 0, and the smaller the contribution of this part of the correlation score. This is consistent with the intuitive understanding of marking the objects that are adjacent to the description in engineering drawings. The dot product operation is performed on the transpose of the graphic feature vector and the text feature vector, and the result reflects the semantic similarity between the two. The dot product of a feature vector that the model identifies as a graphic symbol of a pump and a text feature vector that is identified as an equipment tag number will be significantly higher than the dot product of it and a text representing the pipe diameter, such as DN200. This is a learnable bias term used to fine-tune the baseline of the graph score during training. Through calculation, a graph-text pair receives an association score; for example, the pump symbol might score 0.95 with P-101, while its score with another unrelated nearby text might be 0.12. The semantic feature recognition association unit selects the text with the highest score for each graph symbol as its associated text and outputs this graph-text pair as a whole, along with its fused feature information, to the next processing unit. After processing by this unit, the output data is transformed into a set of independent graph-text features, where each element is a successfully associated graph-text pair, signifying that the discrete primitive information has been assigned a preliminary logical relationship.

[0020] Finally, the feature decoding and encapsulation unit performs the final decoding and data construction tasks. This unit utilizes a pre-defined engineering specification dictionary to precisely translate the still somewhat ambiguous feature set identified by the artificial intelligence model into deterministic entity objects that conform to engineering standards. This engineering specification dictionary is a large database that stores all standard symbol information from design specifications of specific industries (such as chemical or water treatment) in a structured manner, for example, various valve, pump, and instrument symbols defined in the GB / T21458 standard. Each entry in the dictionary contains a template feature vector of a standard symbol, its corresponding engineering category name, such as centrifugal pump or gate valve, and the attribute fields that the category should possess, such as tag number, model, and material. The feature decoding and encapsulation unit traverses each graphic-text pair element in the received independent graphic-text feature set. For each element, it extracts the feature vector of its graphic portion. Then, a fast matching search is performed in the engineering specification dictionary. The matching process is completed by calculating the cosine similarity between the input vector and all template feature vectors in the dictionary. The entry with the highest similarity exceeding a certain preset threshold, such as 0.9, is considered the best match. For example, if the similarity calculated between the graphic feature vector of an input image-text pair and the template vector of centrifugal pump in the dictionary is 0.98, then this unit decodes this graphic entity into the centrifugal pump category. After decoding, the feature decoding and encapsulation unit extracts information from the associated text portion and fills it in according to the attribute list defined for this category in the dictionary. If the dictionary defines that centrifugal pump requires an associated tag number attribute, then the value can be extracted from the text P-101 in the image-text pair and assigned to that attribute. Similarly, if there is also a nearby associated Q=50m... 3 The text " / h, H=30m" will be used, and the corresponding flow rate and head attributes will also be assigned values. After the entity category decoding and attribute binding described above, all discrete graphic and textual feature elements are transformed into complete engineering data instances. However, these instances are still isolated. Therefore, the final step of the feature decoding and encapsulation unit is to serialize and encapsulate these data instances according to a unified, tree-like standard engineering protocol. This protocol can be a format based on JavaScript Object Notation (JSON) or Extensible Markup Language (XML). The feature decoding and encapsulation unit will create a top-level object representing the entire engineering project or drawing, and then create child nodes such as equipment lists and pipe segment lists under this object. The previously decoded P-101 centrifugal pump instance will be created as a JSON object and added to the equipment list array. This JSON object will contain the following information: Category: Centrifugal Pump, Tag Number: P-101, Coordinates: {x:35.2, y:58.4}, Flow Rate: 50m³ / h 3 Key-value pairs such as / h are used. Finally, this unit serializes the entire data tree into a string or file as structured engineering data.

[0021] Specifically, the engineering data reasoning and reconstruction module 120 is used to perform spatial logic reasoning and network flow reconstruction on structured engineering data based on a knowledge graph reasoning engine to obtain spatial topology data. Correspondingly, given that the preceding engineering drawing structuring module has successfully transformed the original, unstructured two-dimensional engineering drawings into structured engineering data composed of a series of independent engineering object instances carrying basic attributes, this data only addresses the question of what is on the drawings, i.e., identifying each discrete physical entity and its text annotations. It does not reveal how these entities are interconnected in the macroscopic process flow, how fluids flow in the pipeline network, or their actual spatial relationships in the three-dimensional physical world. In other words, these data instances lack system-level topological connectivity and spatial logic, and cannot be directly used to construct a functionally correct and spatially reasonable three-dimensional digital model. To bridge this cognitive gap from discrete entities to continuous systems, the engineering data reasoning and reconstruction module 120 is introduced to perform deep spatial logic reasoning and network flow reconstruction on this structured engineering data based on a knowledge graph reasoning engine, thereby generating spatial topology data containing complete three-dimensional connection relationships and flow information.

[0022] Figure 3 This is a block diagram of the engineering data reasoning and reconstruction module in the AI ​​self-generating system for converting 2D engineering drawings to 3D drawings according to an embodiment of this application. Figure 3 As shown, in an exemplary embodiment, the engineering data reasoning and reconstruction module 120 includes: an initial knowledge network graph construction unit 121, used to extract physical entities and map connected logical edges from structured engineering data using a graph parser to obtain an initial knowledge network graph; a connectivity measurement direction correction unit 122, used to perform semantic connectivity measurement and spatial alignment direction correction on the initial knowledge network graph to obtain a directed topological connection network; and a spatial topology generation unit 123, used to perform relative elevation calculation and positional relationship matrix construction on the directed topological connection network to obtain spatial topology data.

[0023] The relevant operational details are as follows: After receiving input data, such as a data file serialized in JSON format describing the entire pump station equipment and pipelines, the graph parser deployed in the initial knowledge network graph construction unit 121 begins to traverse each instantiated object in the data file. For each object, the parser extracts its core physical entity attributes and maps them to an entity node in the graph structure. For example, when the parser reads a JSON object {"Category":"Centrifugal Pump","Tag Number":"P-101","Coordinates":{"x":35.2,"y":58.4},"PID Identifier":"L-1001-OUT"}, it creates an entity node representing the centrifugal pump, which internally stores its key attributes such as category, tag number, and two-dimensional coordinates. Similarly, an object named {"Category":"Gate Valve","Tag Number":"V-101","Coordinates":{"x":45.8,"y":58.4},"PID Identifier":["L-1001-IN","L-1002-OUT"]} will also be instantiated as a corresponding gate valve entity node. After instantiating all entity nodes, the graph resolver will focus on processing the logical identifier data used to characterize pipeline connectivity. This data originates from the parsing results of the Piping and Instrumentation Diagram (PID diagram). In engineering practice, the PID diagram uniquely identifies a continuous process pipeline using a specific pipeline number. Based on the fundamental rule that identical pipeline numbers indicate physical connectivity, the resolver establishes connections between corresponding entity nodes. Specifically, the resolver will find that the PID identifier for the outlet of centrifugal pump P-101 is L-1001-OUT, while the PID identifier for a connection point of gate valve V-101 is L-1001-IN. Although the flow direction is not yet determined, the same core pipeline number L-1001 indicates that there is a physical pipeline between P-101 and V-101. Therefore, the parser will create a connection edge between the P-101 node and the V-101 node. At this stage, all created connection edges are undirected edges, because the correspondence of pipeline numbers alone is insufficient to determine the precise flow direction of the fluid. By aggregating and assembling all extracted entity nodes with the undirected connection edges mapped according to the PID logical identifier data, the initial knowledge network graph construction unit 121 finally constructs and outputs an initial knowledge network graph. This graph is structurally a standard graph data structure, containing the set of nodes of all physical entities and the set of undirected edges representing the potential connectivity between them.

[0024] Upon receiving the initial knowledge network graph, the connectivity measurement direction correction unit 122 is immediately activated. The core component of this unit is a knowledge graph inference engine. This engine first loads a built-in engineering graph rule base. This rule base is a collection of knowledge gathered and organized from a large number of standard design specifications, process manuals, and expert experience from industries such as chemical engineering and water treatment. It is stored as a series of if-then logical rules. For example, the library contains hundreds of deterministic process rules, such as the fluid flow direction of a centrifugal pump must be from its inlet to its outlet, fluid can only flow in one direction when passing through a check valve, and fluid always flows from the main pipe to the branch pipe. When the inference engine loads the initial knowledge network graph, it first attempts to apply these deterministic rules. For the undirected edge between node P-101 (centrifugal pump) and node V-101 (gate valve), the engine queries the rule base and finds a strong rule that the fluid from the centrifugal pump flows out from the outlet. Furthermore, the L-1001-OUT identifier of node P-101 explicitly indicates that this connection point is an outlet. Therefore, the engine can deterministically correct this undirected edge into a directed edge from P-101 to V-101. However, in complex pipe networks, such as at pipe intersections like tees and crosses, or in areas with design ambiguities, deterministic rules alone may not solve all problems. In such cases, the inference engine initiates its topology connection probability calculation logic to quantitatively evaluate the rationality of different connection schemes and selects the scheme with the highest probability to determine the connection relationship and flow direction. This calculation logic follows the formula below to calculate the topology connection probability from node i to node j. : In this formula, This represents the basic semantic connectivity between node i and node j, and its value is obtained by querying a pre-constructed graph rule association matrix. This matrix is ​​an N×N matrix, where N is the total number of all possible device / fitting interface types in the project, and each element in the matrix... The system stores semantic reasonableness scores (between 0 and 1) for connections from interface type a to interface type b. These scores are assigned by domain experts based on process reasonableness. For example, the score for connecting from a pump outlet to a valve inlet might be as high as 0.9, while the score for connecting from a pump outlet to a storage tank outlet might be close to 0. Represents spatial alignment, where It is the angle between the outflow direction vector of node i and the inflow direction vector of node j. In a two-dimensional plane, the direction vector of a pipeline can be directly extracted from the drawing. For example, a horizontal pipe segment pointing from (10,20) to (30,20) has a direction vector of (1,0). If the inflow direction of another pipe fitting is also horizontal, then... A value of 1 indicates perfect spatial alignment, which is the preferred connection method. If the entrance direction is vertical, with an included angle of 90 degrees, A value of 0 indicates that the connection is not smooth in space. and These are two weighting coefficients used to balance the importance of semantic connectivity and spatial alignment, respectively. Their sum is 1, and their specific values ​​are obtained through regression testing and parameter optimization on a large amount of historical project data to ensure that the calculated probabilities best match the actual design intent. For example, they can be set... =0.6, =0.4. The final one. The term is a pipe diameter matching attenuation index, which plays a crucial penalty role. and These are the nominal diameters of the ports connecting nodes i and j, respectively. If the pipe diameters of the two ports are exactly the same, for example, both DN150, then... =0, the exponent term has a value of 1, and does not cause any attenuation to the connection probability. However, if the pipe diameters do not match, for example, a DN150 port is attempted to connect to a DN100 port, then =50, the value of the exponent term will be significantly less than 1, thus greatly reducing the overall probability of the connection. This is a constraint that closely reflects engineering practice, because direct connections between pipes of different diameters require reducing fittings, and the probability of a direct connection is very low. This is a scaling factor, which can be set to 50 to control the penalty for pipe diameter mismatch. By calculating this probability value for all possible connection combinations, the inference engine ultimately selects the scheme with the highest probability for each uncertain connection and corrects undirected edges to their corresponding directed edges. After this series of processes, the final output is a directed topology network where all edges have a definite direction.

[0025] Finally, the spatial topology generation unit 123 receives the directed topology network as input. First, this unit projects the elevation of all nodes in the network into three-dimensional space based on an engineering drawing elevation specification dictionary. This elevation specification dictionary is another crucial knowledge base, storing the standard installation heights or burial depths of various equipment and pipelines under different operating conditions. These specifications are also derived from national standards, industry regulations, or internal design practices of enterprises. For example, the dictionary may contain rules such as the base installation elevation of a circulating water pump being ±0.00 meters, the center elevation of the main pipeline on the first floor of the process pipe gallery being +4.50 meters, and the soil cover depth of outdoor buried water supply pipes not less than 0.7 meters. The spatial topology generation unit 123 traverses each node in the directed network, querying the corresponding elevation rule in the dictionary based on its category and attributes. For example, for the centrifugal pump node with tag number P-101, its two-dimensional coordinates are (35.2, 58.4). A dictionary query reveals the standard base installation height is +0.5 meters, so the three-dimensional foundation anchorage coordinates of this node are determined to be (35.2, 58.4, 0.5). For the valve node V-101 on the pipeline, a query reveals the pipeline's design elevation is +0.5 meters, therefore its three-dimensional coordinates are also assigned a Z value of 0.5. After determining the precise three-dimensional coordinates of all nodes in the network, the spatial topology generation unit 123 can calculate the actual physical spatial length between any two nodes connected by a directed edge. This length is not the straight-line crossing distance on a two-dimensional plane. Since pipelines in real engineering environments must follow the principle of orthogonal laying along the X, Y, and Z axes, the system uses the three-dimensional orthogonal Manhattan distance formula. This is used to calculate the theoretically shortest orthogonal pipeline length. This estimated physical length, conforming to orthogonal wiring logic, is a key parameter for subsequent preliminary material statistics and initial stress pre-assessment of the pipeline network. Simultaneously, this unit also constructs a three-dimensional positional relationship matrix, an N×N matrix where N is the total number of nodes, and the elements of the matrix can be three-dimensional displacement vectors from node i to node j. Finally, the spatial topology generation unit 123 aggregates and serializes all the three-dimensional coordinates of all nodes, the directed topological relationships of all connecting edges, and the calculated precise physical length of each pipeline segment to form a complete and accurate spatial topology data.

[0026] Specifically, the data parameterization assembly module 130 is used to perform feature comparison and assembly rule mapping on structured engineering data to obtain parameterized assembly data. It should be understood that although the preceding modules have successfully transformed discrete two-dimensional drawing information into structured engineering data and further deduced the three-dimensional topological skeleton of the pipeline system, this only solves the problems of what exists, where it is, and how it is connected. To truly generate a three-dimensional model that can be used for refined design, construction simulation, and even digital twin delivery, it is necessary to solve the problem of the precise three-dimensional geometry of each piece of equipment, fitting, and valve, as well as its assembly constraints. Simply replacing the identified objects with basic geometric shapes such as cubes or cylinders obviously cannot meet the engineering accuracy requirements. Each standard component has a complex and precise shape, and their assembly process must also follow a series of strict industry standards and spatial constraints. To endow each node and edge in the topology network with a realistic geometric entity that conforms to engineering specifications, a data parameterization assembly module 130 is introduced to perform in-depth feature comparison and assembly rule mapping on the structured engineering data. This generates parameterized assembly data containing precise three-dimensional geometric parameters and assembly instructions for each engineering object, thereby transforming the abstract topology blueprint into a set of concrete entity instructions that can be directly rendered and assembled by the 3D engine.

[0027] In one exemplary embodiment, the data parameterized assembly module 130 includes: a spatial constraint unit, used to query hard spatial constraint conditions and bind attributes to structured engineering data to obtain a basic rule constraint set; a constraint set matching unit, used to perform feature cosine similarity matching and topology deviation penalty calculation on the basic rule constraint set to obtain hybrid assembly strategy data; and an injection unit, used to call standard part templates and inject dynamic deformation of engineering dimensions into the hybrid assembly strategy data to obtain parameterized assembly data.

[0028] The relevant operational details are as follows: The spatial constraint unit receives pre-parsed structured engineering data from the upstream module. This data includes equipment model parameters identified from the 2D drawings, such as centrifugal pump model ISG-80-160, pipe fitting specifications such as DN150, PN1.6, and connection type identifiers, such as flange connection. Once this data is received, the spatial constraint unit triggers its built-in, rule-based hard-matching mechanism. At the core of this mechanism is a rigorously formalized library of standards and specifications. This library is not simply a collection of documents, but a structured database that digitally includes a large number of national standards (such as GB), industry standards (such as HG / T, chemical industry standards), and internal design and process specifications of enterprises. These specifications are pre-parsed and transformed into a series of condition-result logical rules. For example, the specification regarding the maximum allowable spacing of horizontal pipes in the document will be transformed into a rule: if the nominal diameter of the pipe is 150 mm and the medium is water, then its maximum support spacing shall not exceed 4.5 meters. When equipment and fitting parameters from structured engineering data are input into this standard specification library, the specification parser within the spatial constraint unit begins executing a series of conditional mapping queries. This standard specification library belongs to the system's vast and comprehensive resource library module. As the knowledge foundation of the entire intelligent conversion system, the resource library module not only contains a standard specification library storing national and industry standards that constrain pipe spacing, connection methods, etc., but also fully integrates a strategy library covering design strategies, process steps, and component selection rules for direct AI access and simulation of expert decision-making. Furthermore, this module incorporates a user-owned 3D model library encompassing a massive amount of parametric standard parts and industry components, as well as a database containing brand information, supplier resources, historical prices, and historical contracts. With this rich resource library as a foundation, taking a horizontal process pipe with a nominal diameter of DN150 as an example, the parser will use "pipe," "DN150," and "horizontal" as query keywords. The query results may match multiple relevant specifications. For example, in addition to the support spacing rules mentioned above, they may also match regulations regarding the minimum installation slope for that pipe diameter, such as not less than 0.3%, bolt hole alignment requirements for flange connections, and minimum safe clearance from other pipes or structures. For a device identified as a safety valve, the parser will query hard installation orientation and location restrictions, such as the requirement to install vertically upwards and the outlet to be connected to a safe area. The spatial constraint unit structurally binds and combines all the hard spatial constraints retrieved through the query that must be unconditionally followed by the current device or fitting with the original device attributes.For example, for the DN150 pipe, in addition to the original pipe diameter, material, and other attributes, a new constraint set field will be added to its data object. This field contains a series of key-value pairs such as {"Constraint Type":"Maximum Support Spacing","Value":4.5} and {"Constraint Type":"Minimum Slope","Value":0.003}. Through this processing, the spatial constraint unit ultimately generates and outputs a basic rule constraint set. This constraint set assigns a clear and inviolable rule label to each standardized engineering object.

[0029] After obtaining the basic rule constraint set, the constraint set matching unit begins processing. This unit is established based on a common fact in engineering practice: not all engineering designs are completely standardized. In complex engineering projects, there are always some non-standard components, such as custom-made equipment composed of multiple standard components, irregular pipe connections, or customized parts that are shown on drawings but whose corresponding models cannot be found in any standard library. For these complex issues that cannot be directly mapped by the hard rule reasoning (RBR) of the previous unit, the constraint set matching unit adopts a more flexible and intelligent case-based reasoning (CBR) soft matching mechanism to seek solutions. This unit first traverses the input basic rule constraint set, comparing it with the standard specification library to isolate non-standard components that cannot be fully covered and explained. For each non-standard component, the constraint set matching unit constructs a target attribute feature vector from its key physical properties, such as known pipe diameter, material, and pressure rating. The system then constructs a target topology feature vector by analyzing the connection environment of the pipeline network topology, such as the number of nodes connected and the angles between various connection branches. Subsequently, this unit activates the CBR engine, which begins searching a vast historical project database in the system backend. This database stores complete data on all successfully completed 2D-to-3D conversion projects, including non-standard parts encountered in each project and the optimal solutions validated at the time. The CBR engine's mission is to find the most similar historical precedent for the current non-standard part problem within this treasure trove of experience. This similarity assessment is not subjective but is accomplished through a rigorous quantitative evaluation logic. The matching degree between the two cases... It is calculated using the following formula: This formula comprehensively measures the similarity between two cases through weighted combination. Part One It is the attribute feature vector of the current target case. The attribute feature vector of a certain historical case The cosine similarity between the two. This value measures the degree of similarity between the two at the level of basic physical properties. For example, consider a non-standard tee with three ports of DN150, DN150, and DN100, made of stainless steel. If a similar case with three ports, DN150, DN150, and DN80, also made of stainless steel is found in the historical database, then their cosine similarity in property characteristics will be very high. (Weighting coefficient) A value of 0.7 can be set to adjust the importance of attribute similarity. (The second part of the formula...) This is used to measure the differences between the two cases in terms of topological environment. It is a comprehensive metric for topological deviation distance, which can be the absolute value of the difference in the number of connected nodes between two cases, or the root mean square of the difference in branch angles, etc. When the topological environments of the two cases are exactly the same, =0, this item's value is 1; the greater the topological difference, the closer this item's value is to 0, exhibiting an exponential decay. This means that even if the basic attributes are similar, if the network environment differs greatly, the confidence in using the solution from that environment should be significantly reduced. Weighting coefficient and attenuation factor If set to 0.1, this is also obtained through machine learning training on historical data. Finally, the entire weighted sum is multiplied by a historical case reliability weight coefficient. This coefficient reflects the success rate of the historical case solution across all past applications. For example, a solution that has been successfully referenced multiple times and has proven effective has a high success rate. The value will be close to 1; while a solution that is rarely used or has caused subsequent problems will have a value close to 1. The value will be lower. The historical project library is traversed, and the matching score of all historical cases is calculated for each non-standard part. Finally, the solution of the historical case with the highest score is selected, such as a processing rule consisting of a standard tee and a reducer, as the recommended processing strategy for the current problem. This results in the output of a hybrid assembly strategy data set. This data not only includes hard rule constraints on standard parts but also provides the most reliable flexible assembly solutions for non-standard parts that have been historically tested.

[0030] After obtaining the mixed assembly strategy data containing all standard and non-standard processing schemes, the injection unit serves as the final stop for the data parameterization assembly module. This unit first parses the input mixed assembly strategy data, extracting the type identifier for each object to be processed, such as centrifugal pump-ISG-80-160, gate valve-Z41H-16C-DN150, and the corresponding assembly rule instructions. Based on the parsed type identifier, the injection unit accurately calls the corresponding model template from a vast 3D model resource library. This resource library is pre-built and contains a massive number of 3D parameterized models of engineering equipment, valves, and pipe fittings. Unlike traditional, fixed-size 3D models, the model template here is a skeleton or shell; its key geometric dimensions are not fixed values ​​but are constrained and driven by a series of variable parameters, such as flange outer diameter, valve body surface spacing, and pump body overall length. For example, when the injection unit needs to process a gate valve-Z41H-16C-DN150, it will call a general 3D model template for a Z41H type gate valve. Next, the injection unit executes its core injection operation. It dynamically injects the specific engineering dimensional parameters for the gate valve from the hybrid assembly strategy data, such as the flange outer diameter of 285 mm and the structural length of 400 mm extracted from the drawing annotations, into the feature tree constraints within the invoked gate valve model template as driving parameters. This injection process triggers the template's internal parametric engine, causing real-time and precise deformation of the template's geometry. The flange diameter automatically adjusts to 285 mm, and the valve body length extends to 400 mm, generating a unique 3D solid model that perfectly matches the requirements of the current actual design drawings. For non-standard parts for which solutions are found through CBR, the injection unit also follows strategy instructions. For example, it first invokes a standard tee template and a reducer template, performs parametric dimensional injection on them separately, and then combines them into a composite 3D solid according to the relative positional relationships defined in the strategy. After completing the dimensional-driven deformation of all objects, each object becomes a solid with precise 3D geometry. Finally, the injection unit will perform a final aggregation and serialization encapsulation of this 3D solid model object, along with its 3D spatial installation posture matrix (i.e., its position and rotation angle in the global coordinate system) which has been determined in the spatial topology data. The final output is a parametric assembly data containing the precise 3D geometry data of all engineering objects and their spatial assembly instructions.

[0031] Specifically, the 3D engineering model construction module 140 is used to perform coordinate positioning and entity pipeline routing assembly on spatial topology data and parametric assembly data to obtain a 3D engineering model. It is understandable that although the preceding modules have laid a solid data foundation for the entire automated modeling process, successively completing the conversion from unstructured drawings to structured data, then the reasoning and reconstruction of 3D topology logic, and finally the generation of parametric geometric data for each entity object, these results are essentially still a series of discrete, unintegrated sets of data instructions. They define what the entity is, where the entity is, and the precise 3D shape of the entity, but do not perform the final, global construction and assembly of this information in a unified 3D virtual space. A truly meaningful 3D engineering model requires all equipment to be precisely placed on their absolute coordinates, all pipelines to connect the various devices with optimal and collision-free paths, and the entire scene to conform to the real constraints of the physical world. To this end, a 3D engineering model construction module 140 is introduced to perform precise coordinate positioning and intelligent physical pipeline routing and assembly in a virtual 3D environment based on the spatial topology data and parametric assembly data transmitted from upstream, and finally construct and output a high-fidelity 3D engineering model that can be directly used for subsequent engineering applications.

[0032] In one exemplary embodiment, the three-dimensional engineering model construction module 140 includes: an initial positioning unit, used to perform global virtual coordinate system mounting and absolute position anchoring on spatial topology data and parameterized assembly data to obtain initial positioning scene skeleton data; a cost calculation connector assembly unit, used to perform pipeline routing cost calculation and connector entity interpolation assembly on the initial positioning scene skeleton data to obtain full initial assembly model data; and a model data detection and recalculation unit, used to perform spatial penetration detection and local path avoidance recalculation on the full initial assembly model data to obtain a three-dimensional engineering model.

[0033] The relevant operational details are as follows: The initial positioning unit first extracts the three-dimensional absolute coordinate information of all key equipment and pipe fitting nodes from the spatial topology data. For example, for a centrifugal pump with tag number P-101, it extracts the precise three-dimensional anchor point coordinates derived and determined by the preceding module, i.e., (35.2, 58.4, 0.5), corresponding to its position on the X, Y, and Z axes, respectively. Simultaneously, it extracts from the parametric assembly data an instantiated three-dimensional solid model corresponding to tag number P-101 that has undergone dimension-driven deformation. This model itself has an independent local coordinate system with its geometric center as the origin, and comes with a default installation attitude matrix, which is an identity matrix representing no rotation. Next, the core affine transformation operator inside the initial positioning unit begins to work. For each solid model that needs to be positioned, this operator performs a series of matrix operations to achieve a precise mapping from the local coordinate system to the global coordinate system. This process mainly involves translation and rotation transformations. First, a translation matrix is ​​created, with its translation vector being the global coordinates (35.2, 58.4, 0.5) of the device extracted from the spatial topology data. Then, based on possible installation orientation requirements in the parametric assembly data (e.g., a safety valve must be installed vertically upwards), a rotation matrix is ​​calculated. This rotation matrix rotates the model from its default orientation to the specified installation orientation. Finally, the original vertex coordinate matrix of the entity model is multiplied by this affine transformation matrix combining translation and rotation to calculate the new coordinates of each vertex in the global virtual 3D coordinate system. By performing this operation on all main devices, valves, and other core entities in the scene, the initial positioning unit completes the absolute position anchoring of the main engineering equipment. Finally, the unit combines and packages these spatially anchored device models, along with the precise spatial coordinates of the flanges or pipe openings pre-installed for connecting pipelines, to generate an initial positioning scene skeleton data that establishes the main framework of the project. Although the devices are still isolated from each other in this data, the framework of the entire scenario has been formed, and the locations of all interfaces to be connected have been clearly defined.

[0034] After the initial scene skeleton data is constructed, the cost calculation connector assembly unit begins its intelligent pipeline routing task. This unit first reads the input skeleton data and extracts the set of spatial coordinate points of all suspended flanges awaiting pipeline connection. These points appear in pairs, each pair originating from a directed logical connection edge defined in the spatial topology data, representing the need to lay a physical pipeline between them. For example, the cost calculation connector assembly unit identifies the actual outlet flange center point coordinates of centrifugal pump P-101 after parametric model invocation and dimension injection, based on the geometric stretching and offset of its local coordinate system. The coordinates (35.95, 58.4, 0.75) and the center point coordinates of the inlet flange of gate valve V-101 need to be carefully connected. Based on the hard data association established in the previous module, since the center anchoring coordinates of gate valve V-101 have been established as (45.8, 58.4, 0.5), combined with the 400 mm total structural length (0.2 m half-width) parametrically injected by the system from the 3D model library, the absolute coordinates of its inlet flange on the water-facing side are precisely reversed and induced back, generating the effective endpoint coordinates. (45.6, 58.4, 0.5). In this specific connection pair, the system detected a 0.25-meter three-dimensional spatial elevation misalignment between the actual outlet flange's 0.75-meter elevation and the downstream gate valve's 0.5-meter elevation. This represents a non-linear connection challenge that needs to be overcome. For each pair of starting and ending pipe ports requiring connection, the unit does not simply connect them with a straight line. In real engineering, pipelines need to be laid orthogonally (i.e., parallel along the X, Y, and Z axes) to facilitate construction and support, and also to avoid other equipment and obstacles. Therefore, the cost calculation connection assembly unit employs an intelligent pathfinding algorithm based on three-dimensional orthogonal mesh search, such as a variant of the A* (A-star) algorithm. This algorithm searches for an orthogonal path from the starting point to the ending point with the minimum total cost within a discretized three-dimensional spatial grid. Especially when dealing with elevation differences caused by varying equipment installation heights, the algorithm can automatically generate, evolve, and insert a short vertically descending riser at the feature node with the lowest overall cost, based on the three-dimensional Manhattan distance rule, thus naturally bridging the Z-axis spatial fault. Here, the cost is not a simple distance concept, but rather a comprehensive routing cost function. The definition and calculation formula are as follows: This cost function consists of three key parts. The first part represents the total cost of the pipeline's length. Among them, is the length of the i-th straight pipe segment in the path, and S is the total number of straight segments. The first part is an adjustable distance weighting coefficient, such as set to 1.0. It is directly related to the cost of pipeline materials; the longer the route, the greater this cost. The second part represents the complexity of the pipeline and construction costs. This refers to the number of bends in the entire path, i.e., the number of times the path direction changes by 90 degrees. Each additional bend not only means an additional cost for the pipe fitting (the bend itself), but also increases the pressure loss of the fluid flowing in the pipeline and increases the amount of welding or installation work on site. Therefore, a larger bend penalty weight... Setting it to 5.0 will be used to suppress the algorithm from generating too many unnecessary bends. The third part is a spatial safety distance penalty. In each step of the pathfinding algorithm, the minimum distance between the current path point and other existing obstacles in the scene, such as devices, walls, and other generated pipes, is calculated. .and This is the minimum safe distance that must be maintained for this pipe diameter, obtained from the standards and specifications database. For example, for a DN150 pipe, the minimum clearance from other objects is required to be 100 mm. If Greater than or equal to This means the current position is safe; at this point, the numerator of the exponent is negative or zero, and the value of the entire penalty term is very small, almost negligible. However, once the path exploration enters within the safe distance, i.e. When the numerator becomes positive, the penalty value increases exponentially as the distance decreases. Weighting coefficient. If set to 100.0 and attenuation factor Setting it to 10 controls the severity of the penalty. The ultimate goal of the optimization algorithm is to search and determine a route globally that satisfies the total routing cost function. The cost calculation unit finds the optimal three-dimensional orthogonal path with the minimum cost. After finding the optimal path, the unit generates a three-dimensional solid of the pipeline along that path and automatically inserts elbows of the appropriate specifications at each inflection point. By performing this process on all pipe pairs to be connected, the unit ultimately generates a complete initial assembly model data containing all equipment and the entire piping system.

[0035] After initial automated assembly, the model data detection and recalculation unit serves as the final quality control checkpoint. Although previous routing algorithms considered collision avoidance, in high-density, complex scenes, especially when multiple pipelines are navigating in parallel, localized, minor collisions or failures to meet safety clearance requirements can still occur. To address this, the model data detection and recalculation unit employs an efficient two-stage collision detection strategy. The first stage is a coarse spatial exclusion based on the Axis-Aligned Bounding Box Tree (AABB Tree) algorithm. This algorithm calculates a minimum bounding box (the smallest cube whose faces are parallel to the coordinate axes) that perfectly encloses each 3D entity (equipment, pipe, fitting) in the scene. Then, an efficient tree-like data structure quickly determines which object bounding boxes overlap in space. Only entity pairs with overlapping bounding boxes are considered potential collision risk objects and are selected for the next stage of precise detection. This method significantly reduces the number of entity pairs requiring precise calculation, greatly improving detection efficiency. For all potential collision entity pairs selected in the first stage, the model data detection and recalculation unit performs a second stage of precise interference detection. This stage employs a Boolean intersection algorithm based on polygonal meshes. This algorithm directly performs geometric intersection calculations on the 3D surface meshes of the two objects (composed of numerous tiny triangular facets). If the calculation results show any form of penetration or volumetric intersection between the meshes of the two objects, then a physical collision can be accurately determined. Furthermore, the model data detection and recalculation unit checks whether the entities meet the reserved operating or maintenance space requirements; for example, sufficient space must be left around a valve handwheel for personnel operation. Once any rigid body penetration or space encroachment is detected, the model data detection and recalculation unit accurately records the coordinates of the interference point and sends this feedback signal containing obstacle information back to the cost calculation connector assembly unit from the previous stage. Upon receiving the feedback, the cost calculation connector assembly unit uses the marked interference point as a high-cost region and triggers a local path avoidance recalculation for the relevant pipeline. The new path will be forced to bypass this confirmed collision point. The model data inspection and recalculation unit will iteratively execute the process of inspection-feedback-recalculation-re-inspection until no more unexpected mesh interferences are detected in the entire 3D space. Finally, after confirming that all geometric models are in a spatially reasonable and collision-free state, the unit will perform a final sealing and baking of all the finally verified and qualified 3D geometric mesh data, along with the Bill of Materials (BOM) attribute tags (such as materials, specifications, supplier information, etc.) inherited from various stages, into a single file with a complete data structure that conforms to the Building Information Modeling (BIM) standard format (such as IFC or Revit format).In terms of 3D visualization output, the system not only delivers a basic 3D system model but also simultaneously generates intuitive 3D rendered displays and exploded assembly views. To improve the collaborative efficiency of modern engineering management, the system supports the generation of files based on this model that can be used for distributed BIM collaboration, allowing project team members to perform lightweight multi-device browsing and annotation online. Furthermore, for extended applications in digital twins and virtual reality, this 3D model can be directly output in a low-level format suitable for mainstream game engines (such as Unity and Unreal Engine) or lightweight web rendering frameworks (such as Three.js), enabling seamless integration into immersive simulation training or online 3D interactive display scenarios. This file is the final product of the entire automated process: the 3D engineering model.

[0036] It is understandable that in the aforementioned three-dimensional spatial interference verification and position correction logic, interference detection relies entirely on the geometric bounding box algorithm for the first stage of spatial exclusion, and combines it with Boolean intersection of polygonal meshes for the second stage of precise verification. This single rigid body physical penetration and static operational space encroachment detection mode exposes serious lag and limitations when facing real complex engineering environments. Specifically, this mechanism fails to consider the multi-physics field coupling radiation correlation between engineering entities across physical boundaries. In actual industrial scenarios, there are strong non-contact field-space interactions between pipelines and equipment. For example, even if a carbon steel main pipeline transporting high-temperature and high-pressure steam does not have any physical mesh interference with the adjacent PVC dosing pipeline on the three-dimensional geometric mesh, the high-intensity thermal radiation field emitted by it is enough to cross the spatial gap and cause the PVC pipeline to soften, deform, or even rupture and fail; similarly, if the high-frequency vibration pipe section of a high-power centrifugal pump is close to a high-precision sensitive component, even if there is a geometric gap of several centimeters between them, the mechanical vibration coupling field transmitted through the air or shared support will cause severe signal distortion. If the aforementioned non-contact field interference phenomena are ignored, the generated 3D model may appear collision-free visually, but it could easily lead to serious safety accidents during actual engineering implementation and production operation. To overcome the limitations of pure geometric detection, a non-contact implicit physical field interference detection mechanism is introduced when processing the full set of initial assembly model data. This mechanism includes three core transformation steps from surface geometric assembly to deep physical mechanism simulation.

[0037] Based on this, in the second exemplary embodiment, the model data detection and recalculation unit is used for:

[0038] Physical field boundary expansion and attribute decoupling are performed on the full initial assembly model data to obtain multidimensional field entity skeleton data. It is understandable that the aforementioned fixed safety distance cannot truly reflect the dynamic threat posed by different media to the surrounding environment. This fixed geometric envelope pattern is prone to missed risk assessment when facing invisible high-temperature heat waves or mechanical resonance sources. To construct a field enclosure that adaptively extends with physical characteristics, physical field boundary expansion and attribute decoupling are performed on the full initial assembly model data to obtain multidimensional field entity skeleton data. During execution, using the full initial assembly model data as input, while extracting the pipeline geometric mesh, the physical characteristics of the contained medium, such as fluid temperature and pipeline amplitude, are forcibly decoupled and extracted. Based on the above decoupling attributes, a dynamic field expansion radius exceeding the basic geometric boundary is calculated for each entity. The semi-normal calculation formula is as follows: ;In the formula, The dynamic field expansion radius characterizes the i-th entity; The basic geometric outer diameter representing the i-th entity is directly extracted from the design specifications of the 3D pipe model. This is the thermal radiation diffusion coefficient, the value of which is determined based on the thermal conductivity characteristics of the air at the site. This represents the temperature difference between the fluid temperature inside the i-th entity and the ambient temperature. The ambient air heat transfer threshold; The vibration radiation coefficient; Refers to the working amplitude of the i-th entity; To characterize the vibration transmission gain factor of mechanical wave amplification, and Based on the dynamic testing experience of mechanical equipment, the system is pre-defined and fixed at the bottom layer. Specifically, this applies to the centrifugal pump outlet pipe section, which has been located in three-dimensional space using coordinates. Referring to the previous configuration data, its corresponding pipe diameter and outer diameter can be determined. The value is 0.075. During the decoupling process, the temperature of the medium transported in this pipe section was obtained as 120 degrees Celsius, while the set ambient temperature was 20 degrees Celsius. The absolute temperature difference was calculated. The temperature is 100 degrees Celsius. The system adjusts the heat transfer threshold based on the workshop environment. The preset temperature is 10 degrees Celsius, and the thermal radiation diffusion coefficient is... The preset value is 0.05. Simultaneously, the operating amplitude of the centrifugal pump is acquired. The vibration radiation coefficient is 0.01 meters. The default value is 0.2, which is the gain factor. The preset value is 40. Substituting the above real data into the formula, the logarithmic term was used to smooth the calculation and found that thermal radiation expanded the danger radius by an additional 0.12 meters. At the same time, the exponential term was used to accurately calculate that mechanical vibration increased the danger airspace by about 0.3 meters. The combination of these three factors resulted in the total dynamic field expansion radius of the entity, which originally had an outer diameter of only 0.075 meters, being expanded to nearly half a meter.

[0039] Multiphysics coupling interference intensity estimation is performed on the multidimensional field entity skeleton data to obtain the field interference heatmap matrix. That is, when the dynamic field enclosures of different entities spatially overlap, it is extremely difficult to distinguish the devastating consequences of the overlapping area on surrounding devices based solely on the boundary state. Treating all overlaps with equal weight would lead to extremely uneven distribution of subsequent avoidance costs and deviation from true physical laws. When faced with full initial assembly model data containing massive pipeline components, forcibly requiring the underlying logic to perform full physical field traversal calculations on any two components would inevitably lead to an extremely large accumulation of computational load and even cause the computing power architecture to collapse. Therefore, the dynamic field expansion radius carried in the multidimensional field entity skeleton data output from the preceding steps serves as an efficient pre-trigger threshold for coarse spatial resistance elimination operations. Specifically, the processing mechanism first performs a preliminary spatial intersection test at the underlying level, only considering the actual spatial Euclidean distance between entity i and entity j. Strictly smaller than the dynamic field expansion radius of both and Only when the sum of their expansion radii is reached does the system determine that the two entities have entered a high-risk radiation zone, and then accurately extract the core attributes of the pair, such as temperature and material, to activate subsequent precise physical microscopic characterization. Conversely, if the actual distance is greater than the sum of their expansion radii, the system directly determines that the environment is absolutely safe and the interference intensity is zero, and quickly skips the cumbersome calculation process for the entity pair. Once it is confirmed that two industrial objects are deeply embedded within each other's dynamic field influence range, the intensity of interference damage they experience will completely eliminate boundary assumptions and strictly follow the actual distance-power decay effect and the absolute energy of the source in the laws of physics for rigorous calculation. At this time, the macroscopic boundary expansion radius only serves as a broad outer contour, perfectly fulfilling its defensive mission of massive data screening and collision warning, and therefore is no longer substituted as an explicit algebraic variable into the subsequent core equation. To accurately quantify the conflict intensity between these invisible field boundaries and consider the resistance differences of the target objects, multi-physics field coupling interference intensity estimation is performed on the multi-dimensional field entity skeleton data to obtain the field interference heat map matrix. During execution, for any two entities whose field enclosing volumes overlap, the resistance sensitivity parameter of the target material is introduced to calculate the bidirectional field coupling interference intensity. This logic not only considers the absolute energy of the radiation source but also rigorously evaluates the tolerance of the target entity. The formula for calculating the interference intensity is as follows: ;In the formula, Characterizes the coupling interference strength of entity i to entity j; and The system uses multi-physics field weight normalization parameters to balance the dimensional differences between the thermal and vibrational fields. These parameters are preset using an engineering expert scoring method. It is 0.6. It is 0.4; The thermal sensitivity index representing the material of entity j is obtained by querying the thermal resistance physical and chemical property library of the target pipeline material. Let be the shortest spatial Euclidean distance between the surfaces of entity i and entity j; To prevent the denominator from overflowing due to tiny positive numbers, the system defaults to 0.0001. The resonance penalty function, derived from the mapping of the natural frequency difference between the two entities, is dynamically calculated by the system using a nonlinear curve based on the difference between the operating frequency of the radiation source and the natural frequency of the receiver. This function is applied when the operating frequency of radiation source i... The intrinsic frequency of receptor j At this point, the function value exhibits a non-linear surge. Referring to the previous example, let's denote the high-temperature pipe section at the centrifugal pump outlet as entity i, and a nearby dosing pipeline made of polyvinyl chloride as entity j. To match the energy calculation dimensions of the underlying normalization, the system presets the thermal sensitivity index of polyvinyl chloride. The quantization value is 0.0075, but the significant difference between its natural frequency and the pump's vibration frequency results in a resonance penalty function score of only 1.2. The shortest spatial Euclidean distance between the surfaces of the two pipelines was calculated using a three-dimensional coordinate system. The distance is 0.3 meters. Substituting this into the formula, due to the extreme heat resistance of PVC and the close proximity, the reciprocal of the square of the distance greatly amplifies the destructive effect of the original 100-degree Celsius temperature difference. Combined with the normalized sensitivity operator, the thermal interference intensity component is accurately calculated to be approximately 5.0. In contrast, the vibration interference, based on the inverse cube law of pure physics, is drastically reduced in its denominator, resulting in a vibration interference component of approximately 0.18. After weighted calculation using the formula, the obtained comprehensive interference intensity value is determined to be approximately 5.18, significantly exceeding the baseline warning line. This step establishes a non-contact, implicit interference intensity theoretical model for the engineering entity, encapsulating the conflict scores of all similar entities into a field interference thermogram matrix. This achieves the effect of exposing deep-seated engineering accident hazards in advance during the digital model stage and presenting them intuitively with quantitative parameters.

[0040] Based on the field interferometric heatmap matrix, adaptive gradient repulsion and route correction are applied to the full initial assembly model data to obtain a 3D engineering model. Ultimately, after obtaining the interferometric heatmap distribution, traditional random avoidance or brute-force coordinate recalculation easily leads to deadlock in the global pipeline routing, and trial-and-error optimization without directional guidance consumes extremely high levels of underlying computational resources. To form a controlled and clearly directional spatial obstacle removal mechanism, adaptive gradient repulsion and route correction are applied to the full initial assembly model data based on the field interferometric heatmap matrix to obtain a 3D engineering model. In this execution phase, which transitions from quantitative evaluation to geometric correction, vulnerable pipeline nodes are forced to perform escape displacements. The absolute direction and feedback force of their escape depend entirely on the microscopic physical burning and oscillation experienced by the target in its current coordinate domain, as well as the direction of the 3D spatial absolute tensor of the core centroids of both the exertor and receiver. Since the physical meaning of the previously defined dynamic field expansion radius is limited only to the extreme value of the farthest dangerous boundary reached by the field energy wave, when sensitive equipment or pipelines have been detected by the array as being deeply trapped inside the radiation field and suffering from the damage of multi-physics field coupling interference, the computational focus of the routing escape mechanism is automatically extended and completely locked on the steep distribution difference of the internal interference gradient. The design intention of the entire spatial obstacle removal mechanism is to guide the interference nodes to quickly escape the physical field penalty coverage area with minimal geometric reshaping and recalculation costs. Therefore, the macroscopic dynamic field expansion radius label data is formally stripped from the underlying displacement calculation logic and completely handed over to the pre-generated absolute interference intensity, environmental spatial intensity gradient, and centroid geometric offset vector to take over and dominate the spatial force superposition driving calculation. During execution, the spatial routing processing logic synchronously reads the field interference heat map matrix and skeleton model data, uses the artificial potential field method to transform high-risk areas into potential energy peaks, applies a repulsive force with a clear spatial orientation to the affected pipeline nodes that need to be avoided, and forces them to undergo three-dimensional spatial displacement in the safe direction where the repulsive force gradient decreases the fastest. The specific calculation logic for the repulsive force vector used in displacement correction is as follows: ;In the formula, The spatial repulsion correction vector that needs to be executed by the affected node j; This represents the set of all adjacent hazardous entities that exert field interference on node j; It refers to the gradient value of the coupling interference intensity in three-dimensional space; The system presets the interference tolerance safety threshold as specified by the engineering standard as a constant value according to the national industrial pipeline specification. In order to ensure the dimensional consistency of the underlying mathematical calculation and to prevent overflow, the system maps it to a normalized constant value of 1.0. and These represent the three-dimensional coordinate vectors of the centroids of entities j and i, respectively. The fractional part calculates the unit direction vector between the two points pointing to the victim entity, where the denominator is the physical Euclidean distance represented by the vector modulo operation. Continuing with the previous scenario example, the centroid coordinates of the PVC pipe segment... Under intense radiation threat, the calculated interference intensity is as high as 5.18, far exceeding the set normalized safety threshold of 1.0. At this point, the exponential term exp(5.18-1.0) generates a nonlinear penalty multiplier of approximately 65.3. This magnitude not only completely avoids the risk of underlying computing power overflow but also effectively amplifies the gradient value with tremendous power. By extracting and normalizing the coordinate differences between the high-temperature pipe segment vector set and the PVC vector set, a three-dimensional unit direction vector is calculated, pointing directly from the high-temperature pipe source to the vulnerable dosing pipe. Combining this with the massive scalar value obtained through summation, the system automatically derives a large correction vector that moves away from the pump body along the spatial plane, such as an outward offset of 0.45 meters along the ordinate axis. When the sensitive pipeline mistakenly enters a high-temperature or high-vibration radiation zone, the algorithm, relying on its high physical intuition, directly calculates the optimal route to escape the danger zone with minimal geometric deflection. With the iterative redrawing operation guided by repulsion until the entire heat map data is cleared, this step completely eliminates the material damage traps hidden in the complex mesh of traditional models. This enables the 3D engineering model containing information about many components to not only achieve a seamless mesh without penetration, but also to achieve extremely high reliability for digital twin production in the harsh environment of intertwined thermal and dynamic forces.

[0041] Specifically, the engineering-related data generation module 150 is used to perform view dimensionality reduction calculations and material feature extraction on the 3D engineering model based on reverse geometric projection and attribute mapping algorithms to generate engineering-related data. That is, although a high-fidelity, collision-free 3D engineering model is the core product of the entire automated generation process, in actual engineering project lifecycle management, the 3D model itself is often not the final deliverable. Key aspects such as project bidding, material procurement, workshop processing, on-site construction, and even final cost settlement still heavily rely on a series of traditional, standardized 2D drawings and quantity and price lists. If, after generating the 3D model, designers still need to manually measure dimensions, count materials, draw processing diagrams, and estimate costs from the model, then the problem of the disconnect between design and cost management mentioned in the background technology is not fundamentally solved, and the overall efficiency improvement will be greatly reduced. In order to fully connect the entire data link from initial design to final project delivery and achieve a high degree of synergy between design, quantity calculation and cost estimation, the engineering associated data generation module 150 makes full use of the single and reliable data source of the generated 3D engineering model. Through intelligent back projection and attribute extraction, it automatically generates engineering associated data that is from the same source and highly correlated for all downstream links.

[0042] Figure 4 This is a schematic diagram of the data flow in the AI ​​self-generating system for converting 2D engineering drawings to 3D drawings according to an embodiment of this application, specifically the module for generating engineering-related data. Figure 4As shown, in an exemplary embodiment, the engineering-related data generation module 150 includes: a data dimensionality reduction and stripping unit, used to perform view dimensionality reduction contour capture and metadata stripping on the three-dimensional engineering model to obtain two-dimensional processing view data and model material attribute set; a quantity and price accounting list generation unit, used to perform physical engineering quantity statistics and regional price mapping estimation on the model material attribute set to obtain quantity and price accounting list data; and a data association and binning unit, used to perform component icon index penetration association and standard format serialization binning on the two-dimensional processing view data and quantity and price accounting list data to obtain engineering-related data.

[0043] The relevant operational details are as follows: The data dimensionality reduction and stripping unit first reads the input 3D engineering model file (e.g., an IFC format file). Within the 3D model, each entity object, such as a valve or a section of pipe, possesses a globally unique identifier (GUID) assigned during model creation; this identifier is like an identity card for each object. The data dimensionality reduction and stripping unit traverses the entity feature tree within the model based on this GUID, performing a forced data decoupling operation on each entity. For example, for a 3D entity representing a gate valve, its data structure contains both complex 3D geometric mesh data describing its shape and a series of non-geometric metadata describing its engineering attributes, such as {"tag number":"V-101","specification":"DN150","material":"carbon steel"}. The decoupling operation completely separates these two parts of data; the 3D geometric mesh is sent to the view dimensionality reduction calculation process, while the metadata is sent to the BOM (Bill of Materials) attribute extraction process. For the stripped 3D geometric mesh structure, the data dimensionality reduction and stripping unit activates a virtual orthogonal camera array. This array consists of six virtual cameras, each located at the center of one of the six faces of a giant cube surrounding the entire model, directly facing the center of the model. This simulates the six basic view directions in mechanical drawing: front view, top view, left view, etc. For each view direction, the data dimensionality reduction and stripping unit performs a reverse view dimensionality reduction calculation. The core of this calculation process utilizes the depth buffer (Z-Buffer) algorithm from computer graphics. This algorithm, from the camera's perspective, determines which contour lines of the model are directly visible, which are invisible due to occlusion by other parts, and which are internal structural lines of the object. According to preset mechanical drawing standards, such as using solid thick lines to represent visible contours and dashed lines to represent hidden contours, the algorithm projects these edges in three-dimensional space onto a two-dimensional plane, generating a series of two-dimensional vector line segments. After completing the projection calculations for all views, the data dimensionality reduction and stripping unit also automatically encapsulates these two-dimensional views with bounding boxes and intelligently adds necessary dimension lines, such as automatically annotating the total length, width, and height of the equipment, as well as the positioning dimensions of key interfaces. Ultimately, this standardized 2D vector graphics data is output as 2D machining view data (e.g., AutoCAD DWG format files) that can be directly used in workshop manufacturing. Simultaneously, while the view is being reduced in dimensionality, a data parser processes all the BOM metadata decoupled from the model. The parser iterates through this metadata, extracting all discrete tags describing material characteristics, such as the precise length of a pipe (measured directly from its 3D geometry), the specific model of a valve, the pressure rating and specifications of a flange, etc. These extracted attribute tags are aggregated to form a list-like, but not yet categorized and summarized, raw material list.This list is defined as a set of model material attributes and is output as the second output of this data dimensionality reduction stripping unit.

[0044] After obtaining the set of material attributes from the model, the quantity and price accounting bill of quantities generation unit begins to execute its core tasks of quantity statistics and cost estimation. This unit first performs a hash-based clustering of the somewhat messy original material list using national standard codes for material classification. National standard codes for material classification are a nationally unified system that assigns unique codes to each category of engineering materials. The unit matches each material item in the list with its unique classification code in the national standard code library based on its model, specifications, material, and other attributes. For example, all seamless steel pipes with a specification of DN200 and made of carbon steel, regardless of how many segments they are divided into in the model, will be assigned the same national standard code. Then, by hashing and grouping these national standard codes, all materials belonging to the same category are quickly grouped together. Based on this, the quantity and price accounting bill of quantities generation unit can accurately perform physical quantity statistics. For example, it can sum up the lengths of all pipe segments belonging to the DN200 carbon steel seamless pipe category to obtain a total length, such as 85.5 meters; simultaneously, it can count the total number of centrifugal pumps of a specific model, ISG-80-160, such as 2 units. Along with in-depth statistical analysis of macroscopic physical engineering quantities, the system can also refine and identify physical components and automatically generate multi-dimensional professional engineering lists, including pipe material lists, valve and pump equipment lists, and support and hanger material lists. After generating an accurate and clearly categorized list of material quantities, the quantity and price calculation list generation unit will input it into a dynamically updated historical price database in the system's backend to perform automated cost calculations. This historical price database is a long-term maintained database that continuously collects historical purchase prices of various engineering materials by connecting with the company's own procurement system (ERP) or by periodically scraping data from publicly available building materials information websites. Each price record in the database includes information such as the national standard code of the material, the unit price, the purchase time, and the supplier. For each type of material in the current bill of quantities, the quantity and price calculation unit will perform a dynamic cost forecast. Its calculation logic follows the formula below to calculate the predicted total price of the k-th type of material. : In this formula, This is the total physical quantity of the k-th type of material, calculated from the previous step, such as 85.5 meters of steel pipe. The core part is the unit price prediction. It represents the weighted average unit price over a specified historical time window (e.g., the past 6 months, i.e., T=6). This is the average purchase price per unit of the material in the past month t, retrieved from the historical price database. It is a weighting factor that decays over time. This weighting factor is designed to ensure that the more recent the purchase price, the greater its impact on the current unit price forecast. Its calculation method can be, for example, linear decay. Or it could be exponential decay. This allows the predicted unit price to more sensitively reflect the latest market price dynamics. This is an absolute amount representing the estimated logistics and loss redundancy based on the total quantity and category of materials. For example, for materials like pipes that are prone to generating scrap, the system first calculates a fixed percentage (e.g., 2%) of the estimated total cost, and then uses this absolute amount as... The costs are directly added into the parentheses within the total price forecast formula to scientifically calculate logistics premiums and material loss costs. Finally, the sum of all basic costs needs to be multiplied by a regional price fluctuation index for the current project location. This index, for example, might be 1.05 for a certain region, and is obtained from a macroeconomic database to adjust for price differences in building materials across different regions. By performing this calculation on each type of material in the list, the quantity and price accounting list generation unit ultimately generates and outputs a detailed, accurate, and dynamically reflective quantity and price accounting list of data.

[0045] Finally, after the 2D machining view data and the quantity and price calculation list data are ready, the data association and packing unit will perform the final data integration and packaging delivery task. In the final data distribution and collaborative delivery stage, the system can export accompanying files with varying depths as needed according to the management needs of different business departments. For example, it can output equipment procurement lists with supplier information to the purchasing department, accurately output component manufacturing drawings and pipeline processing drawings generated by the back projection of the 3D model to the construction site, and accurately push project cost analysis tables to the cost consulting department. The data association and packing unit first allocates an integrated workspace in memory and loads the two core data sets output by the first two units simultaneously. Its core innovation lies in using the unique globally unique identifier (GUID) in the original 3D solid model as the primary key for association to break down the data barriers between drawings and lists. When generating 2D machining views, each component symbol projected from the 3D solid (e.g., a 2D symbol for a valve) retains the GUID of its original 3D solid in its data structure. Similarly, when generating the quantity and price calculation list, each material item also retains a list of GUIDs for all the three-dimensional entities that constitute that item. The data association and packing unit traverses this data, establishing a low-level, two-way hyperlink index mechanism between component symbols in the two-dimensional drawings and the corresponding material items in the cost list table. This means that the final delivered document package will have a high degree of interactivity: when a user clicks on a symbol for a pipe or equipment on the electronic drawing, the system can immediately use its embedded GUID to automatically locate and highlight the corresponding row containing that material in the associated cost list table, allowing the user to instantly understand the detailed specifications, quantity, and cost information of the component. Conversely, when a user is reviewing a long cost list, if they have any questions about the material in a particular row, they can directly click on that row, and the system will also use the GUID to highlight or zoom to the location of that material in the two-dimensional drawing, achieving seamless penetration query between drawings and tables. After completing this in-depth indexing and association construction, the data association and packing unit will integrate data according to specific Building Information Modeling (BIM) delivery specifications, such as using the Industrial Foundation Class (IFC) format, or packaging all view files (e.g., DWG), BOM form files (e.g., XLSX), total cost estimates, and their associated index files together into a compressed collaborative delivery package format (e.g., ZIP). Finally, the data association and packing unit will perform a final baking and output of this well-structured, data-interoperable, and directly referenced and interactively queried project-related data package for downstream bidding departments, procurement departments, construction units, and cost consultants.

[0046] In summary, the AI ​​self-generating system 100 for converting 2D engineering drawings to 3D drawings based on embodiments of this application is explained. It addresses technical problems such as over-reliance on manual drawing interpretation, difficulties in extracting complex topological relationships, and a severe disconnect between model construction and subsequent quantity surveying and cost estimation during the conversion from 2D engineering drawings to 3D models. First, it performs deep format parsing and semantic feature extraction on the input 2D engineering drawings, automatically converting high-dimensional graphic symbols into machine-readable structured engineering data, thus overcoming the efficiency and accuracy bottlenecks of manual drawing interpretation at the source. Then, relying on a knowledge graph inference engine, it performs spatial logic reasoning and network flow reconstruction on the structured data, accurately restoring the 3D spatial topological relationships of the pipeline network. Next, the system obtains parametric assembly data through feature comparison and assembly rule mapping, thereby driving the physical pipelines and equipment components to complete coordinate anchoring and automated routing assembly in 3D space, efficiently constructing a high-fidelity 3D engineering model. Finally, by using reverse geometric projection and attribute mapping algorithms, material features are directly extracted from the generated 3D model and the view is reduced in dimensionality. Simultaneously, engineering-related data containing processing documents and quantity and price lists are output, thereby realizing the integrated connection from drawing analysis, topology reasoning, 3D reconstruction to engineering cost, effectively eliminating human error and data silos in traditional methods.

[0047] Various implementations of this disclosure have been described above. The foregoing description is exemplary and not exhaustive. Furthermore, it is not limited to the disclosed implementations, and many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described implementations.

Claims

1. An AI self-generating system for converting 2D engineering drawings to 3D drawings, characterized in that, include: The engineering drawing structuring module is used to parse the format of two-dimensional engineering drawings and extract semantic features to obtain structured engineering data. The engineering data reasoning and reconstruction module is used to perform spatial logic reasoning and network flow reconstruction on structured engineering data based on the knowledge graph reasoning engine to obtain spatial topology data. The data parameterization assembly module is used to perform feature comparison and assembly rule mapping on structured engineering data to obtain parameterized assembly data. The 3D engineering model construction module is used to perform coordinate positioning and entity pipeline routing assembly on spatial topology data and parametric assembly data to obtain a 3D engineering model. It includes: an initial positioning unit for global virtual coordinate system mounting and absolute position anchoring on spatial topology data and parametric assembly data to obtain initial positioning scene skeleton data; a cost calculation connector assembly unit for pipeline routing cost calculation and connector entity interpolation assembly on the initial positioning scene skeleton data to obtain full initial assembly model data; and a model data detection and recalculation unit for performing physical field boundary expansion and attribute decoupling on the full initial assembly model data to obtain multi-dimensional field entity skeleton data. Specifically, using the full initial assembly model data as input, while extracting the pipeline geometric mesh, it decouples and extracts physical characteristics such as fluid temperature and pipeline amplitude of the contained medium. Based on the above decoupled attributes, a dynamic field expansion radius exceeding the basic geometric boundary is calculated for each entity. The formula for calculating this radius is as follows: ;In the formula, The dynamic field expansion radius characterizes the i-th entity; The basic geometric outer diameter representing the i-th entity is directly extracted from the design specifications of the 3D pipe model. This is the thermal radiation diffusion coefficient, the value of which is determined based on the thermal conductivity characteristics of the air at the site. This represents the temperature difference between the fluid temperature inside the i-th entity and the ambient temperature. The ambient air heat transfer threshold; The vibration radiation coefficient; Refers to the working amplitude of the i-th entity; To characterize the vibrational transmission gain factor of mechanical wave amplification, and Based on the dynamic testing experience of mechanical equipment, the system is pre-set and solidified at the bottom layer; the multi-physics field coupling interference intensity is estimated from the multi-dimensional field entity skeleton data to obtain the field interference heat map matrix; based on the field interference heat map matrix, adaptive gradient repulsion and route correction are performed on the full initial assembly model data to obtain the three-dimensional engineering model; the engineering associated data generation module is used to perform view dimensionality reduction calculation and material feature extraction on the three-dimensional engineering model based on the reverse geometric projection and attribute mapping algorithm to generate engineering associated data.

2. The AI ​​self-generating system for converting 2D engineering drawings to 3D drawings according to claim 1, characterized in that, The engineering drawing structuring module includes: an engineering drawing separation and mapping unit, used to perform layer separation and geometric vectorization mapping on two-dimensional engineering drawings based on a format parsing engine to obtain a basic set of graphic elements; a semantic feature recognition and association unit, used to perform semantic feature recognition and association on the basic set of graphic elements based on an AI model to obtain an independent set of graphic and textual features; and a feature decoding and encapsulation unit, used to perform entity category decoding and tree-structured protocol serialization encapsulation on the independent set of graphic and textual features through an engineering specification dictionary mapping and matching mechanism to obtain structured engineering data.

3. The AI ​​self-generating system for converting 2D engineering drawings to 3D drawings according to claim 1, characterized in that, The engineering data reasoning and reconstruction module includes: an initial knowledge network graph construction unit, used to extract physical entities and map connected logical edges from structured engineering data using a graph parser to obtain an initial knowledge network graph; a connectivity measurement direction correction unit, used to perform semantic connectivity measurement and spatial alignment direction correction on the initial knowledge network graph to obtain a directed topological connection network; and a spatial topology generation unit, used to perform relative elevation calculation and positional relationship matrix construction on the directed topological connection network to obtain spatial topology data.

4. The AI ​​self-generating system for converting 2D engineering drawings to 3D drawings according to claim 1, characterized in that, The data parameterized assembly module includes: a spatial constraint unit, used to query hard spatial constraint conditions and bind attributes to structured engineering data to obtain a basic rule constraint set; a constraint set matching unit, used to perform feature cosine similarity matching and topological deviation penalty calculation on the basic rule constraint set to obtain hybrid assembly strategy data; and an injection unit, used to call standard part templates and inject dynamic deformation of engineering dimensions into the hybrid assembly strategy data to obtain parameterized assembly data.

5. The AI ​​self-generating system for converting 2D engineering drawings to 3D drawings according to claim 1, characterized in that, The engineering-related data generation module includes: a data dimensionality reduction and stripping unit, used to perform view dimensionality reduction contour capture and metadata stripping on the three-dimensional engineering model to obtain two-dimensional processing view data and model material attribute set; a quantity and price accounting list generation unit, used to perform physical engineering quantity statistics and regional price mapping estimation on the model material attribute set to obtain quantity and price accounting list data; and a data association and binning unit, used to perform component icon index penetration association and standard format serialization binning on the two-dimensional processing view data and quantity and price accounting list data to obtain engineering-related data.

Images