AI-based building drawing to BIM model automatic generation method and system

Abstract

The application relates to the technical field of data processing, and discloses an AI-based building drawing-to-BIM model automatic generation method and system, wherein the method comprises the following steps: filtering invalid layers to generate a candidate layer set; performing multi-model concurrent classification and maintaining historical confidence; calculating a Jaccard difference degree and a voting information entropy to detect conflicts; using weighted voting and a double adaptive dynamic threshold to resolve differences and extract a consensus result; generating a three-dimensional BIM model with attributes from the consensus result and calibrating the historical confidence. Compared with the prior art, when the multi-model outputs are inconsistent, simple intersection determination leads to component missing detection, and simple union determination introduces a large amount of noise, so that the technical problem that precision and recall rate cannot be considered cannot be solved. Due to the construction of the weighted voting and double adaptive dynamic threshold conflict resolution mechanism based on the three-dimensional historical confidence, quantitative evaluation and dynamic arbitration of the difference result are realized, and the accuracy of the drawing-to-model automatic generation is improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an AI-based method and system for automatically generating BIM models from architectural drawings. Background Technology

[0002] Currently, in the field of Building Information Modeling (BIM) automated reconstruction, semantic classification of layers based on 2D CAD drawings is a crucial prerequisite for accurate 3D reconstruction. With the widespread adoption of deep learning technology, the industry has widely adopted multiple heterogeneous AI models for concurrent semantic classification of drawing layers to reduce the instability caused by training data bias or image noise in a single model. Conventional multi-model result fusion strategies mainly focus on simple set operations: when high-confidence results are needed, the intersection of the layer classification subsets output by all models is directly calculated; when pursuing more comprehensive coverage, a union operation is used. However, this rigid, either-or approach ignores the inherent cognitive differences between models and the performance fluctuations of different layers on different models, making it difficult to effectively address the challenges posed by diverse drawing formats and professional drafting habits in real-world engineering scenarios.

[0003] For example, in practical applications, when multiple AI models classify layers of the same building structural plan as "shear walls," different models often output inconsistent subsets of classifications due to differences in architecture, training data distribution, and sensitivity to line type and color features. While using intersection as the consensus result ensures that all selected layers are consistently recognized by multiple models, this stringent consensus condition is highly susceptible to errors by a single model on a particular layer or blind spots caused by local occlusion. This can lead to the complete omission of correctly classified shear wall layers, resulting in missing key load-bearing components in the final BIM model and severely impacting the reliability of the structural analysis. Conversely, while using union to determine the classification includes all possible correct outputs, it also introduces numerous misclassified layers due to model illusions, such as misclassifying leader lines or fill patterns as shear walls, into the result set. These noise layers directly interfere with the component generation logic of the modeling engine, causing chaotic application of wall components and annotation lines. This results in the automatically generated BIM model losing usable accuracy, forcing users to perform a large number of tedious manual deletions and modifications, which essentially negates the efficiency improvement brought by the automation process.

[0004] Therefore, there is an urgent need for a consensus extraction mechanism that can dynamically balance precision and recall when multiple AI models return inconsistent layer classification results during concurrent inference. This mechanism would represent a leap from simple set operation logic to intelligent conflict resolution algorithms, thereby resolving the technical contradiction between component omissions and noise introduction, and comprehensively improving the overall robustness of automated drawing-to-model generation and the reliability of engineering delivery. Summary of the Invention

[0005] To address the aforementioned technical shortcomings, the purpose of this invention is to propose an AI-based method for automatically generating BIM models from architectural drawings. This method aims to solve the technical problems in existing technologies where inconsistent outputs from multiple models lead to missed component detection due to simple intersection determination and introduce a large amount of noise through simple union determination, making it impossible to achieve both accuracy and recall.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The present invention provides an AI-based method for automatically generating BIM models from architectural drawings.

[0007] The AI-based method for automatically generating BIM models from architectural drawings includes: Step S10: Obtain the 2D CAD drawing file uploaded by the user, and perform layer parsing and candidate set generation tasks based on the 2D CAD drawing file using a preset invalid layer rule filtering mechanism, and output the candidate layer set for inference; Step S20: Based on the natural language query command input by the user and the candidate layer set participating in the inference, the layer classification task is performed using a multi-model concurrent inference and a three-dimensional historical confidence matrix maintenance mechanism, and the layer classification result subset of each model and its original confidence score are output. Step S30: Based on the subset of layer classification results of each model and its original confidence score, a conflict detection task is performed using a dual difference measurement mechanism based on Jaccard difference and voting information entropy, and a conflict detection result package is output. Step S40: Based on the conflict detection result package, a consensus extraction task is performed using a weighted voting mechanism based on three-dimensional historical confidence and a dual adaptive dynamic threshold conflict resolution mechanism, and the final consensus result set is output. Step S50: Based on the final consensus result set, the final consensus result set is used as the structured layer classification result, mapped to the component type and layer correspondence that the BIM modeling engine can recognize, and the modeling engine is driven to call the Revit API to parse the geometric entities in the corresponding layer into three-dimensional component models with dimensions, elevation and material attributes, and generate a three-dimensional BIM model with attributes.

[0008] Preferably, step S10, which involves obtaining the user-uploaded 2D CAD drawing file, performing layer parsing and candidate set generation tasks based on the 2D CAD drawing file using a preset invalid layer rule filtering mechanism, and outputting a set of candidate layers for inference, specifically includes: Step S101: Receive the 2D CAD drawing file uploaded by the user through the front-end interface or CAD plugin, call the Teigha parsing engine compatible with DWG / DXF format to perform structured reading of the 2D CAD drawing file, and generate a complete set of candidate layers. ; Step S102: Call the preset invalid layer rule library and select from the entire set of candidate layers. In the process, the standard non-printing layers named "0" or "Defpoints" are removed in sequence, the layers marked as locked or frozen in the layer table are removed, and blank layers without any geometric primitives or text entities are removed by parsing the drawing's graphic object database. Step S103: Output the filtered set of valid layers as the candidate layer set for inference. .

[0009] Preferably, step S20, based on the user-input natural language query command and the candidate layer set participating in the inference, employs a multi-model concurrent inference and a 3D historical confidence matrix maintenance mechanism to perform the layer classification task, and outputs a subset of the layer classification results of each model and its original confidence score, specifically includes: Step S201: Receive natural language query commands input by the user through the interactive interface or API. and query command Parsed as target component category ; Step S202: From the preset three-dimensional historical confidence matrix Query the BIM model of the i-th item corresponding to the current target component category. Time-decay weighted historical accuracy and field expertise correction factor ,in, Indicates the corresponding timestamp; Step S203: Category the target component With candidate layer set The layer names, line types, and color metadata of each layer are standardized and assembled using Prompt to form... One input data packet, of which... The total number of models; and through asynchronous concurrency... The input data packet is pushed to An independent deep learning semantic matching model to Each deep learning semantic matching model independently outputs its pair of candidate layers. Does each layer belong to the target component category? The binary classification results are recorded, along with the original confidence score for each result. This forms a subset of the layer classification results for each model. .

[0010] Preferably, in step S202, the domain expertise correction factor It is obtained by recalibrating and fine-tuning the F1-score of the deep learning semantic matching model in each category.

[0011] Preferably, step S30, which involves performing a conflict detection task based on a subset of the layer classification results of each model and their original confidence scores, using a dual difference measurement mechanism based on Jaccard difference and voting information entropy, and outputting a conflict detection result package, specifically includes: Step S301: Aggregate the subset of layer classification results from each model And for the current target component category Calculate the intersection of all subsets. Union ; Step S302: Calculate the set conflict dissimilarity based on the intersection and union of all subsets. Meanwhile, for union Each layer in ,calculate Information entropy of the model's judgment result on this layer The calculation formula is: ; in, for The percentage of models that voted "yes" in each model. The percentage of models that voted "no"; Step S303: If and only if the set conflict dissimilarity degree Greater than the preset difference trigger threshold and union There is at least one layer Information entropy Information entropy exceeding the preset trigger threshold When both conditions are met, a significant conflict is determined in the current reasoning round, and a set containing the intersection is generated. Union The conflict detection result package includes the information entropy of each layer and the conflict markers.

[0012] Preferably, step S40, which involves performing a consensus extraction task based on the conflict detection result package using a weighted voting mechanism based on three-dimensional historical confidence and a dual adaptive dynamic threshold conflict resolution mechanism, and outputting the final consensus result set, specifically includes: Step S401: For each decision layer in the conflict detection result package, from the preset 3D historical confidence matrix... Query the BIM model of the i-th item corresponding to the current target component category. Time-decay weighted historical accuracy and field expertise correction factor The voting weight in this round The calculation formula is: ; in, The temperature coefficient is used to control the sharpness of the weight distribution. For the total number of models, The weighted historical accuracy of the m-th round of voting; Step S402: Based on the voting weights of this round Gathering Weighted voting by individual deep learning semantic matching models is used to calculate the decision layer in a linear weighted manner. Weighted consensus score ; Step S403: Obtain Standard deviation of the weights of an independent deep learning semantic matching model At the same time, obtain the layer to be decided. Corresponding information entropy Generate dual adaptive dynamic thresholds , ,in, Based on the threshold, As a weighted divergence adjustment factor, It is an information entropy adjustment factor; Step S404: Calculate the weighted consensus score With dual adaptive dynamic threshold If a comparison is made, Then the decision layer will be... Included in the final consensus result set Otherwise, the layer Drop into the conflict suspension queue.

[0013] This invention also provides an AI-based automatic generation system for architectural drawings to BIM models, including: The layer parsing and candidate set generation module is used to obtain the two-dimensional CAD drawing file uploaded by the user, and perform the layer parsing and candidate set generation task based on the two-dimensional CAD drawing file using a preset invalid layer rule filtering mechanism, and output the candidate layer set for inference. The multi-model concurrent inference module is used to perform layer classification tasks based on the natural language query command input by the user and the candidate layer set participating in the inference, and adopts a multi-model concurrent inference and a three-dimensional historical confidence matrix maintenance mechanism to output a subset of the layer classification results of each model and its original confidence score. The dual-difference conflict detection module is used to perform conflict detection tasks based on the subset of layer classification results of each model and its original confidence scores, and adopts a dual-difference measurement mechanism based on Jaccard difference and voting information entropy, and outputs a conflict detection result package. The weighted voting and dynamic threshold resolution module is used to perform consensus extraction tasks based on the conflict detection result package, using a weighted voting mechanism based on three-dimensional historical confidence and a dual adaptive dynamic threshold conflict resolution mechanism, and output the final consensus result set. The structured output and calibration module is used to perform BIM model generation and online calibration tasks based on the final consensus result set, and output a three-dimensional BIM model with attributes.

[0014] The present invention also provides an AI-based automatic generation device for architectural drawings to BIM models. The AI-based automatic generation device for architectural drawings to BIM models includes: a memory, a processor, and an AI-based automatic generation program for architectural drawings to BIM models stored in the memory and executable on the processor. When the AI-based automatic generation program for architectural drawings to BIM models is executed by the processor, it implements the above-mentioned method.

[0015] The present invention also provides a computer program product, the computer program product including an AI-based automatic generation program for building drawings to BIM models, the AI-based automatic generation program for building drawings to BIM models implementing the above method when executed by a processor.

[0016] The beneficial effects of this invention are as follows: 1. This invention employs a dual discrepancy measurement mechanism based on Jaccard discrepancy and voting information entropy. When disagreements arise in the results of concurrent inference by multiple models, it can accurately identify the severity of conflicts, avoiding the computational waste caused by indiscriminately initiating forced arbitration in traditional methods. Simultaneously, a weighted voting mechanism based on a three-dimensional historical confidence matrix and a dual adaptive dynamic threshold resolution mechanism dynamically integrates the historical performance of each model with its current output. This allows consensus decisions to incorporate both historical experience and real-time feedback, effectively balancing recall while ensuring that component identification accuracy meets engineering requirements, thus resolving the contradiction between missed component detection and noise introduction.

[0017] 2. This invention tightly integrates conflict detection with dynamic threshold resolution, forming a complete closed-loop reasoning process from layer candidate set generation, multi-model classification, conflict quantification to adaptive consensus extraction. This closed loop enables the system to consistently output highly reliable layer classification consensus results even when faced with real-world engineering conditions such as inconsistent drawing quality and non-standard layer naming. It directly drives the downstream modeling engine to generate detailed 3D BIM models with attributes such as dimensions, elevation, and material, thus improving the reliability and engineering practicality of automated architectural drawing modeling. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the first embodiment of the AI-based method for automatically generating BIM models from architectural drawings according to the present invention. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0020] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1: As Figure 1 The diagram shown is a flowchart of the first embodiment of the AI-based automatic generation method for building drawings to BIM models according to the present invention. The first embodiment of the AI-based automatic generation method for building drawings to BIM models according to the present invention is presented.

[0022] In the first embodiment, the AI-based method for automatically generating BIM models from architectural drawings includes: Step S10: Obtain the 2D CAD drawing file uploaded by the user, and perform layer parsing and candidate set generation tasks based on the 2D CAD drawing file using a preset invalid layer rule filtering mechanism, and output the candidate layer set for inference; In step S10, the object being processed is a user-uploaded 2D CAD drawing file, typically in DWG or DXF format, containing geometric shapes, text, layer tables, and status attributes such as frozen and locked. The preset invalid layer rule filtering mechanism is a predefined set of judgment rules, the core components of which include: matching rules for layer names to identify "0" or "..." generated by the CAD system by default. Non-printable layers; traversal rules for layer table status attributes to mark and remove locked or frozen non-editable layers; and deep entity check rules for the graphic object database to remove blank layers containing no geometric primitives or text entities. This step calls a DWG / DXF compatible... The parsing engine performs structured reading of the file and generates a complete set of candidate layers. After further filtering through the aforementioned rule base, a set of candidate layers for inference is output. The output set eliminates invalid data that does not contribute to subsequent semantic classification or introduces interference, providing a clean base of layer metadata input for step S20, enabling concurrent inference to focus on layer entities with real architectural semantics.

[0023] After performing this step, the layer information in the original drawing undergoes a cleanup process, no longer containing system default layers, frozen layers, or empty layers. The internal effect of this change is that the candidate layer set... Each layer in the process corresponds to a real geometric primitive, and its metadata, such as line type and color, is clearly and effectively associated with the graphic object. This makes step S20 possible during... During standardized assembly, the signal-to-noise ratio of the input data is significantly improved. Multiple deep learning models can avoid ambiguity or invalid calculations caused by invalid layers when extracting features, thereby reducing the randomness differences between the output results of each model. In the subsequent step S30, when calculating the intersection, union, and information entropy based on the purified candidate set, the measurement results can more accurately reflect the differences in the model's perception of valid building components, rather than the false height conflicts caused by interfering layers, providing a reliable data foundation for the conflict detection and resolution process.

[0024] Compared to traditional techniques that directly traverse and identify all layers of a drawing or simply filter by layer names using keywords, this step combines layer status attributes with entity checks in a graphic object database to more effectively handle complex situations such as non-standard layer naming, the presence of temporary auxiliary lines, or locked reference maps. Traditional methods often include system layers like "Defpoints" or blank layers in the recognition model. These layers lack geometric information but consume computational resources and may increase Jaccard discrepancies at the set level due to low-confidence outputs from the model, leading to unnecessary conflict arbitration. This step, without increasing user interaction costs, filters out layers containing temporary auxiliary lines but lacking long-term semantics through entity checks, avoiding "ghost components" or misclassification mappings. This improves the accuracy of concurrent inference across multiple models and reduces the probability of false triggers in subsequent conflict resolution stages.

[0025] For example, in a task involving the conversion of HVAC and Plumbing drawings for a large commercial complex, the DWG file uploaded by the user was collaboratively drawn by multiple subcontractors. It contained 212 layers, including four standard layers named "0," three old reference layers locked by the designers, and twelve empty shell layers that, after thorough examination, were confirmed to contain no geometric elements. Step S10 calls... The engine analyzes and obtains the complete set. Then, the rule base removes "0" and "..." from the layer names. "Layers: Read the layer table to remove locked and frozen layers, then scan the graphics object database to remove empty layers, finally outputting a candidate set of 193 valid layers." This filtering means that in subsequent step S20, when the AI ​​model encounters target components such as "fire hydrants" or "smoke exhaust valves," it no longer consumes resources to judge these irrelevant layers. As a result, the original confidence distribution of each model in the classification task is more concentrated on real building components, improving the reliability of subsequent conflict detection.

[0026] Step S20: Based on the natural language query command input by the user and the candidate layer set participating in the inference, the layer classification task is performed using a multi-model concurrent inference and a three-dimensional historical confidence matrix maintenance mechanism, and the layer classification result subset of each model and its original confidence score are output. The core processing object of step S20 includes the natural language query command input by the user. and the candidate layer set output in step S10. First, the natural language understanding module will execute the instructions. Parsed as target component category For example, "shear wall" or "fire door". Then, from the preset three-dimensional historical confidence matrix... Query each model Corresponding to the current category Time-decay weighted historical accuracy and field expertise correction factor .in The exponential moving average method is used for maintenance, and the update formula is as follows: , This is the performance score of the model in the previous task for this category. Then, the target category... Compare with the metadata of each layer in the candidate layer set, such as layer name, line type, and color. Standardized assembly, forming Each input data packet is pushed to the system asynchronously and concurrently. An independent deep learning semantic matching model to Each model independently outputs a binary classification result and records the original confidence score. This forms a subset of the classification results for each model. The model subsets and confidence scores output in this step, together with the historical confidence score entries that have been queried, provide multi-dimensional decision information with prior capabilities for the conflict detection in step S30.

[0027] Through this step, the user's ambiguous natural language instructions are transformed into computable target component categories. Furthermore, each model involved in the inference process obtains historical performance records corresponding to that category. This internal change transforms layer classification from point-based determination by a single model to a system encompassing multiple models. A multidimensional decision matrix of group independent opinions and confidence levels. Each layer The attribution is no longer an isolated label, but is accompanied by the original confidence scores given by each model. This includes the historical credibility information implicit in each model for this task. These rich opinion expressions are used in step S30 to calculate the intersection and union of the sets, as well as the voting information entropy of each layer. Complete input was provided. Historical accuracy was also recorded. The introduction of this allows the subsequent step S40 to allocate weights based on the model's historical performance during weighted voting, thereby more accurately identifying which disagreements originate from high-confidence models and which are merely accidental fluctuations of low-quality models.

[0028] Traditional techniques typically rely on single-model predictions or fixed-weighted multi-model voting, failing to adapt to the varying capabilities of different models across different component categories. When a single model makes a catastrophic misjudgment due to training data blind spots and specific layer naming conventions, subsequent processes are powerless to correct it. Fixed-weighted ensemble methods ignore fluctuations in model proficiency across different categories; for example, a model might perform exceptionally well in identifying shear walls but historically have low accuracy in identifying curtain wall mullions, yet its voting weight remains constant across different tasks. This step utilizes a three-dimensional historical confidence matrix... By maintaining independent historical accuracy and domain correction factors for each model across specific component categories, the output of concurrent inference not only reflects the features of the current layer but also embeds prior knowledge of capabilities specific to the task. This provides a precise quantitative basis for the subsequent step S30 to discover structural cognitive conflicts, avoiding the direct substitution of persistent model misjudgments into the final consensus result.

[0029] For example, in a high-rise residential project, a user enters the natural language command "find all layers of shear walls," and the system parses it as the target category "shear wall." From the matrix... Model found in the search Due to its expertise in geometric profile analysis, its weighted historical accuracy is high. The domain correction factor is 0.95. The model is 1.1. Its ability to identify curved variable cross-section walls is slightly weaker. It is 0.78. The calibration was set to 0.9. The system then packaged the "shear wall" suggestion along with the candidate layer's name and linetype and sent it. (Model) Layers “STR-SHEAR-WALL-P” and “S-WALL-BOUNDARY” were identified as shear walls with high confidence levels of 0.97 and 0.93, respectively, while the model... The former is judged as yes with a confidence level of 0.88, while the latter is given a low confidence level of 0.41 for rejection. This cluster of concurrent outputs with differential characteristics provides a typical input scenario for the conflict detection in step S30.

[0030] Step S30: Based on the subset of layer classification results of each model and its original confidence score, a conflict detection task is performed using a dual difference measurement mechanism based on Jaccard difference and voting information entropy, and a conflict detection result package is output. Step S30 performs a conflict detection task based on a dual difference measurement mechanism of Jaccard difference and voting information entropy, with the input being a subset of the model classification results output in step S20. to and the corresponding original confidence score First, aggregate all subsets, targeting the current component category. Calculate the intersection Union Based on this, calculate the set conflict dissimilarity degree. This is used to measure the degree of macroscopic divergence of the model set on the overall classification boundary. Meanwhile, for the union... Each layer ,calculate Information entropy of the model's judgment results ,in and The entropy value represents the proportion of models that voted "yes" and "no," respectively, reflecting the degree of local divergence on a specific layer. If and only if Greater than the preset trigger threshold And there exists at least one layer. of Greater than the information entropy trigger threshold If a significant conflict is determined in the current round, a conflict detection result package containing the intersection, union, information entropy of each layer, and conflict markers is generated; otherwise, the intersection is... The result is directly adopted as the consensus result with high confidence, and step S40 is skipped to proceed to step S50.

[0031] Through this dual-measurement mechanism, the system distinguishes between simple inconsistencies and severe structural conflicts among multiple models. Macroeconomic indicators This reflects the model's consistency with the overall boundary of "which layers the target category includes," a micro-indicator. This depicts the divergence of opinions on each specific layer. When only a few models show slight skewness towards a few layers, The impact of a conflict might be small. Even if the entropy of a certain layer is slightly high, it might be considered a non-significant conflict because the dual triggering conditions haven't been met. In this case, the intersection can be directly used as the final result for quick processing, avoiding the computational waste caused by subsequent complex resolution. Conversely, when there is a systematic cognitive bias between models, a wide range of result differences can drive up the impact of a conflict. Furthermore, the high-entropy states of multiple layers trigger micro-thresholds, at which point a conflict detection result package is generated and proceeds to step S40 for fine-grained arbitration. This hierarchical response concentrates computational resources on the points of divergence that truly affect the quality of model generation, providing clear arbitration trigger signals and the range of data to be decided for subsequent steps.

[0032] Traditional multi-model fusion methods typically employ simple intersection or union determination, or blind majority voting, when outputs are inconsistent. Directly taking the intersection implicitly requires... Furthermore, the extremely low information entropy of all layers can lead to the omission of numerous correct judgments due to stringent consistency requirements, resulting in missed component inspections. Directly taking the union ignores the severity of conflicts, incorporating misclassifications caused by model illusions and introducing significant noise. Forcing majority voting without distinguishing conflict levels may lead to erroneous conclusions in boundary situations where almost all models are uncertain. This step, without intervening in how to resolve disagreements, accurately defines the subsequent processing path through difference quantification. This allows the system to adaptively select the most efficient processing strategy when facing fluctuations in drawing quality and differences in model performance, avoiding unnecessary computational delays and parameter drift risks introduced by forcibly initiating a weighted arbitration process when model opinions are actually consistent.

[0033] For example, in a HVAC system conversion task for an office building, the current target category is "fresh air duct". Five models are assigned to a layer named "M-HVAC-DUCT-FRESH". Four models identify it as a fresh air duct with a high confidence score exceeding 0.85, while the other model, due to similar labels in the training data often representing return air ducts, identifies it as "no" with a confidence score of 0.6. This results in the layer being in the union of the target categories. In but not in intersection In the middle, there is also a leader layer named "A-ANNO-LEADER". Two models incorrectly identified it as a fresh air duct, while the other three models correctly identified it. At this point, the system calculates... The entropy is 0.43, exceeding the preset threshold of 0.35; the information entropy of layer "M-HVAC-DUCT-FRESH" is... The value is as high as 0.72, exceeding the threshold of 0.5. Since both conditions are met, the system determines that a significant conflict exists, generates a conflict detection result package, and marks it for step S40, instead of incorrectly using the intersection or union directly to generate the BIM model.

[0034] Step S40: Based on the conflict detection result package, a consensus extraction task is performed using a weighted voting mechanism based on three-dimensional historical confidence and a dual adaptive dynamic threshold conflict resolution mechanism, and the final consensus result set is output. Step S40 processes the conflict detection result packet output from step S30, activating a conflict resolution mechanism based on weighted voting and dual adaptive dynamic thresholds using a three-dimensional historical confidence matrix. Query each model Corresponding to the current target component category Weighted historical accuracy and field expertise correction factor The voting weight in this round The calculation formula is: ,in The temperature coefficient controls the sharpness of the weight distribution. Then, for the union... Each decision layer in Gathering The weighted voting of each model is used to calculate the weighted consensus score. ,in For indicator functions, This is the original confidence score. Next, the standard deviation of the model weights for the current round is calculated. And obtain the layer from the collision detection result package. Information entropy Generate dual adaptive dynamic thresholds ,in Based on the threshold, As a weighted divergence adjustment factor, This is an information entropy adjustment factor. and In comparison, if Then it will be included in the final consensus set. Otherwise, it will be dropped into the conflict suspension queue.

[0035] Through the weighted voting and dynamic thresholding mechanism in this step, consensus extraction no longer relies on static rules, but evolves into a precise arbitration that dynamically adjusts according to the task and model state. When faced with a high-value correct layer, even if half of the models vote against it, as long as the models that vote in favor are those with excellent historical performance in that component category and high current output confidence, their weighted consensus score will still be valid. It may still be high. Meanwhile, due to significant disagreements among models, the standard deviation of the weights... Increase information entropy Also relatively high, dual dynamic threshold The scores will rise synchronously, forming an admission threshold that adaptively adjusts according to the degree of disagreement. This linkage between scores and thresholds allows the system to adopt a more conservative admission strategy when there is extreme uncertainty, while adopting a relatively lenient admission strategy when opinions are highly consistent and dominated by high-weight models. This achieves a dynamic balance between precision and recall, rescuing key layers that are mistakenly rejected by low-quality models.

[0036] Traditional conflict resolution methods often employ simple majority voting or fixed thresholds. Simple majority voting assigns equal voting power to each model, ignoring differences in their capabilities on a specific task. A novice model's opinion has the same weight as an expert model's, potentially leading to incorrect inclusion or exclusion when the majority of models are weak. Fixed threshold methods, on the other hand, cannot adjust to fluctuations in task difficulty and model status. When the quality of input drawings deteriorates, causing a general low confidence level across all models, a predefined high threshold may incorrectly reject classifications that should pass, resulting in missed component detections. Conversely, when models generally provide high confidence levels but there is ambiguity, a fixed threshold may easily allow misclassifications to pass. This step integrates historical model capability metrics with real-time, local task divergence metrics into the decision boundary, making the arbitration threshold... It can adaptively adjust based on weight divergence and information entropy, which strengthens the dominance of high historical accuracy models, while maintaining caution for highly divergent layers, effectively improving the problem of difficulty in balancing component missed detection and false detection in cases of ambiguous boundaries.

[0037] For example, when processing an old underground parking garage drawing with inconsistent layer naming, step S30 detected a significant conflict for the "fire sprinkler pipe" category. (Union) There is a layer named "FP_Pipeline" that needs to be identified. Query the matrix. Discovery, model Because the training data contains a large number of old drawings, its weighted historical accuracy is... The domain correction factor is 0.92. The weight was 1.15, and the weight was generated. The weights of the other two general models are only 0.12 and 0.08, respectively. The model with a high confidence level of 0.89 voted "yes," while the other two models voted "no." The high-weighted model was dominant, resulting in a weighted consensus score. The value reached 0.73. Due to significant discrepancies in voting ratios, the weighted standard deviation... Larger, and with higher information entropy High, dynamic threshold The threshold was raised from a base value of 0.5 to 0.68. Despite the increase in the threshold due to disagreement, a consensus score of 0.73 still met the requirement, and the layer was successfully included in the consensus set. It was correctly identified as a fire sprinkler pipe in subsequent modeling, avoiding the omission of key components due to being abandoned in the hanging queue by a simple majority vote of 2 to 1.

[0038] Step S50: Based on the final consensus result set, the final consensus result set is used as the structured layer classification result, mapped to the component type and layer correspondence that the BIM modeling engine can recognize, and the modeling engine is driven to call the Revit API to parse the geometric entities in the corresponding layer into three-dimensional component models with dimensions, elevation and material attributes, and generate a three-dimensional BIM model with attributes.

[0039] Step S50 processes the final consensus result set obtained from step S40 or step S30. The structured output mechanism transforms the layer-component category mapping relationship in the consensus set into a sequence of instructions recognizable by downstream BIM modeling engines (such as APIs or self-developed geometry engines), extracts the geometric entities within the corresponding layers, parses them into 3D component models with dimensions, elevations, and material attributes, and generates a complete BIM model. The exponential moving average historical confidence refresh mechanism, based on the true / false labels obtained after manual calibration and confirmation of the conflict suspension queue, updates each model with the corresponding labels. Calculate in this target category Performance score (1 for correct, 0 for incorrect), then use a time decay factor. The exponential moving average formula updates the three-dimensional historical confidence matrix. Entries in This step not only solidifies the preliminary reasoning results into attributed digital assets, but also completes the online dynamic calibration of the system's core decision-making capabilities.

[0040] This step bridges all the semantic reasoning and conflict arbitration results from the previous steps to the generation of specific 3D components. This transforms each consensus layer from an abstract category label into a data source driving the generation of BIM components with detailed attributes, achieving lossless information transfer. Simultaneously, the exponential moving average refresh mechanism constructs a continuously self-optimizing closed loop: after each task processing or manual verification, the model's historical accuracy for a specific component category is updated. This approximates the actual capabilities in the current real-world environment. This online update allows for a more comprehensive understanding of the historical confidence matrix. As the system continues to evolve, when subsequent tasks call step S20 again to perform model prior queries, or step S40 to perform weighted voting weight calculations, the capability parameters that better reflect the current application scenario and drawing distribution will be used. This will cause the entire system to gradually stabilize in terms of the accuracy of identifying similar components and the rationality of conflict resolution after multiple tasks.

[0041] Traditional BIM automated generation systems typically employ a one-time deployment and fixed model parameter model, which cannot adapt to the continuous evolution of drawing sources and design standards. Different design institutes and different years have significantly different drawing habits, and the recognition ability of static models gradually deteriorates over time, leading to missed component inspections or increased noise introduction rates. Traditional solutions require re-collecting annotation data and retraining the model offline, which is time-consuming and costly. This step constructs an online incremental calibration pathway that does not rely on complete retraining. It automatically updates the model's capability profile using an exponential moving average method each time a task is processed or manual verification is performed. (Time decay factor) The introduction of this feature gives greater weight to recent performance, allowing the system to automatically detect and adapt to new data distributions. For example, when a design institute starts using a new layer naming convention, it can more quickly identify models with new patterns.

[0042] For example, when an architectural firm was implementing a digital archiving project, the system needed to continuously process its historical and newly created drawings. Initially, when processing drawings from 2019, the model... The accuracy rate of identifying "curtain wall mullions" is not high, and its initial... Only 0.7. In each subsequent task, the user calibrates and confirms the conflict queue through the interactive interface. For vertical mullion layers that are missed or incorrectly identified, the system obtains a score based on their actual performance. When When correctly identifying irregularly shaped mullions with locally varying cross-sections, ,according to The formula was updated, and the entry gradually migrated to higher values. This was done after more than twenty curtain wall-related projects. The historical accuracy for this firm's specific mullion representation has smoothly improved to 0.91. When processing a newly completed 2022 construction drawing, based on the updated... , In step S20, the original confidence score of the mullion layer is more decisive and accurate, which reduces the conflict rate in step S30 and obtains a higher voting weight in step S40. Ultimately, this drives the size and spatial positioning of the curtain wall mullion components in the generated BIM model to be closer to reality, reducing the workload of manual correction in the later stage.

[0043] Example 2: Furthermore, the AI-based automatic generation system for building drawings to BIM models provided by this invention employs the AI-based automatic generation method for building drawings to BIM models described in the above embodiments, and can solve the technical problem of automatic generation of AI-based building drawings to BIM models. The beneficial effects of the AI-based automatic generation system for building drawings to BIM models provided by this invention are the same as those of the AI-based automatic generation method for building drawings to BIM models provided in the above embodiments, and other technical features of the AI-based automatic generation system for building drawings to BIM models are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0044] Example 3: This invention provides an AI-based automatic generation device for architectural drawings to BIM models. The device includes at least one processor and a memory communicatively connected to the processor. The memory stores instructions executable by the processor, which are then executed to enable the processor to perform the AI-based automatic generation method for architectural drawings to BIM models described in Example 1. The AI-based automatic generation device for architectural drawings to BIM models in this invention can include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. This AI-based automatic generation device is merely an example and should not be construed as limiting the functionality or scope of the invention. The device may also include processing devices (e.g., central processing units, graphics processors), which can perform various appropriate actions and processes based on programs stored in read-only memory or loaded from storage devices into random access memory. The random access memory also stores various programs and data required for the operation of the AI-based automatic architectural drawing to BIM model generation device. The processing unit, read-only memory, and random access memory are interconnected via a bus. The I / O interface is also connected to the bus. Typically, the following systems can be connected to the I / O interface: input devices including touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices including liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices including magnetic tapes, hard disks, etc.; and communication devices. The communication device allows the AI-based automatic architectural drawing to BIM model generation device to communicate wirelessly or wiredly with other devices to exchange data. While various systems of AI-based automatic architectural drawing to BIM model generation devices have been described, it should be understood that implementation or possession of all described systems is not required. More or fewer systems can be implemented alternatively.

[0045] Example 4: This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the AI-based automatic generation method for converting architectural drawings to BIM models as described above. The computer program product provided by this invention can solve the technical problem of automatically generating AI-based architectural drawings to BIM models. Compared with the prior art, the beneficial effects of the computer program product provided by this invention are the same as those of the AI-based automatic generation method for converting architectural drawings to BIM models provided in the above embodiments, and will not be repeated here.

[0046] In particular, according to the embodiments disclosed in this invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from a read-only memory. When the computer program is executed by a processing device, it performs the functions defined in the methods of the embodiments disclosed in this invention.

[0047] It should be understood that the various parts disclosed in this invention can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0048] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the present invention and its equivalents, the present invention also intends to include these modifications and variations.

Claims

1. An AI-based method for automatically generating BIM models from architectural drawings, characterized in that: The methods include: Step S10: Obtain the 2D CAD drawing file uploaded by the user, and perform layer parsing and candidate set generation tasks based on the 2D CAD drawing file using a preset invalid layer rule filtering mechanism, and output the candidate layer set for inference; Step S20: Based on the natural language query command input by the user and the candidate layer set participating in the inference, the layer classification task is performed using a multi-model concurrent inference and a three-dimensional historical confidence matrix maintenance mechanism, and the layer classification result subset of each model and its original confidence score are output. Step S30: Based on the subset of layer classification results of each model and its original confidence score, a conflict detection task is performed using a dual difference measurement mechanism based on Jaccard difference and voting information entropy, and a conflict detection result package is output. Step S40: Based on the conflict detection result package, a consensus extraction task is performed using a weighted voting mechanism based on three-dimensional historical confidence and a dual adaptive dynamic threshold conflict resolution mechanism, and the final consensus result set is output. Step S50: Based on the final consensus result set, the final consensus result set is used as the structured layer classification result, mapped to the component type and layer correspondence that the BIM modeling engine can recognize, and the modeling engine is driven to call the Revit API to parse the geometric entities in the corresponding layer into three-dimensional component models with dimensions, elevation and material attributes, and generate a three-dimensional BIM model with attributes.

2. The AI-based automatic generation method for converting architectural drawings into BIM models as described in claim 1, characterized in that, Step S10 involves obtaining the user-uploaded 2D CAD drawing file, performing layer parsing and candidate set generation tasks based on the 2D CAD drawing file using a preset invalid layer rule filtering mechanism, and outputting the candidate layer set for inference. Specifically, this includes: Step S101: Receive the 2D CAD drawing file uploaded by the user through the front-end interface or CAD plugin, call the Teigha parsing engine compatible with DWG / DXF format to perform structured reading of the 2D CAD drawing file, and generate a complete set of candidate layers. ; Step S102: Call the preset invalid layer rule library and select from the entire set of candidate layers. In the process, the standard non-printing layers named "0" or "Defpoints" are removed in sequence, the layers marked as locked or frozen in the layer table are removed, and blank layers without any geometric primitives or text entities are removed by parsing the drawing's graphic object database. Step S103: Output the filtered set of valid layers as the candidate layer set for inference. .

3. The AI-based automatic generation method for converting architectural drawings into BIM models as described in claim 1, characterized in that, Step S20, based on the user-input natural language query command and the candidate layer set participating in inference, employs a multi-model concurrent inference and a 3D historical confidence matrix maintenance mechanism to perform the layer classification task, outputting a subset of layer classification results for each model and its original confidence score. Specifically, this includes: Step S201: Receive natural language query commands input by the user through the interactive interface or API. and query command Parsed as target component category ; Step S202: From the preset three-dimensional historical confidence matrix Query the BIM model of the i-th item corresponding to the current target component category. Time-decay weighted historical accuracy and field expertise correction factor ,in, Indicates the corresponding timestamp; Step S203: Category the target component With candidate layer set The layer names, line types, and color metadata of each layer are standardized and assembled using Prompt to form... One input data packet, of which... The total number of models; and through asynchronous concurrency... The input data packet is pushed to An independent deep learning semantic matching model to Each deep learning semantic matching model independently outputs its pair of candidate layers. Does each layer belong to the target component category? The binary classification results are recorded, along with the original confidence score for each result. This forms a subset of the layer classification results for each model. .

4. The AI-based automatic generation method for converting architectural drawings into BIM models as described in claim 3, characterized in that, In step S202, the domain expertise correction factor It is obtained by recalibrating and fine-tuning the F1-score of the deep learning semantic matching model in each category.

5. The AI-based automatic generation method for converting architectural drawings into BIM models as described in claim 3, characterized in that, Step S30, which involves performing a conflict detection task based on a subset of the layer classification results from each model and their original confidence scores, using a dual difference measurement mechanism based on Jaccard difference and voting information entropy, and outputting a conflict detection result package, specifically includes: Step S301: Aggregate the subset of layer classification results from each model And for the current target component category Calculate the intersection of all subsets. Union ; Step S302: Calculate the set conflict dissimilarity based on the intersection and union of all subsets. Meanwhile, for union Each layer in ,calculate Information entropy of the model's judgment result on this layer The calculation formula is: ； in, for The percentage of models that voted "yes" in each model. The percentage of models that voted "no"; Step S303: If and only if the set conflict dissimilarity degree Greater than the preset difference trigger threshold and union There is at least one layer Information entropy Information entropy exceeding the preset trigger threshold When both conditions are met, a significant conflict is determined in the current reasoning round, and a set containing the intersection is generated. Union The conflict detection result package includes the information entropy of each layer and the conflict markers.

6. The AI-based automatic generation method for converting architectural drawings into BIM models as described in claim 5, characterized in that, Step S40, based on the conflict detection result packet, employs a consensus extraction task using a weighted voting mechanism based on three-dimensional historical confidence and a dual adaptive dynamic threshold conflict resolution mechanism, and outputs the final consensus result set. This step specifically includes: Step S401: For each decision layer in the conflict detection result package, from the preset 3D historical confidence matrix... Query the BIM model of the i-th item corresponding to the current target component category. Time-decay weighted historical accuracy and field expertise correction factor The voting weight in this round The calculation formula is: ； in, The temperature coefficient is used to control the sharpness of the weight distribution. For the total number of models, The weighted historical accuracy of the m-th round of voting; Step S402: Based on the voting weights of this round Gathering Weighted voting by individual deep learning semantic matching models is used to calculate the decision layer in a linear weighted manner. Weighted consensus score ; Step S403: Obtain Standard deviation of the weights of an independent deep learning semantic matching model At the same time, obtain the layer to be decided. Corresponding information entropy Generate dual adaptive dynamic thresholds , ,in, Based on the threshold, As a weighted divergence adjustment factor, It is an information entropy adjustment factor; Step S404: Calculate the weighted consensus score With dual adaptive dynamic threshold If a comparison is made, Then the decision layer will be... Included in the final consensus result set Otherwise, the layer Drop into the conflict suspension queue.

7. An AI-based automatic generation system for architectural drawings to BIM models, applied to any one of the AI-based automatic generation methods for architectural drawings to BIM models as described in claims 1 to 6, characterized in that, The system includes: The layer parsing and candidate set generation module is used to obtain the two-dimensional CAD drawing file uploaded by the user, and perform the layer parsing and candidate set generation task based on the two-dimensional CAD drawing file using a preset invalid layer rule filtering mechanism, and output the candidate layer set for inference. The multi-model concurrent inference module is used to perform layer classification tasks based on the natural language query command input by the user and the candidate layer set participating in the inference, and adopts a multi-model concurrent inference and a three-dimensional historical confidence matrix maintenance mechanism to output a subset of the layer classification results of each model and its original confidence score. The dual-difference conflict detection module is used to perform conflict detection tasks based on the subset of layer classification results of each model and its original confidence scores, and adopts a dual-difference measurement mechanism based on Jaccard difference and voting information entropy, and outputs a conflict detection result package. The weighted voting and dynamic threshold resolution module is used to perform consensus extraction tasks based on the conflict detection result package, using a weighted voting mechanism based on three-dimensional historical confidence and a dual adaptive dynamic threshold conflict resolution mechanism, and output the final consensus result set. The structured output and calibration module is used to perform BIM model generation and online calibration tasks based on the final consensus result set, and output a three-dimensional BIM model with attributes.

8. An AI-based automatic generation device for architectural drawings into BIM models, characterized in that: The AI-based automatic generation device for architectural drawings to BIM models includes: a memory, a processor, and an AI-based automatic generation program for architectural drawings to BIM models stored in the memory and executable on the processor. When the AI-based automatic generation program for architectural drawings to BIM models is executed by the processor, it implements the AI-based automatic generation method for architectural drawings to BIM models as described in any one of claims 1 to 6.

9. A computer program product, characterized in that, The computer program product includes an AI-based automatic generation program for building drawings to BIM models. When the AI-based automatic generation program for building drawings to BIM models is executed by a processor, it implements the AI-based automatic generation method for building drawings to BIM models as described in any one of claims 1 to 6.

Images