Engineering amount calculation method, device and equipment based on drawing AI intelligent identification

Abstract

The invention discloses an engineering quantity calculation method, device and equipment based on drawing AI intelligent identification, and relates to the field of the cross technology of the constructional engineering technology and artificial intelligence, and the metering method comprises the following steps: S1, carrying out the preprocessing and multi-format analysis of a drawing; s2, multi-modal component feature extraction is carried out, wherein spatial features and semantic features of the standardized drawing are extracted through a CNN-TransformerFusionModel (CNN-TransformerFusionModel) model; s3, cross-scale component identification and classification: performing cross-scale identification on the multi-modal fusion features based on a CSA-YOLOv8 model, and outputting component categories, bounding box coordinates and confidence; s4, performing component correlation reasoning and parameter completion; and S5, engineering quantity intelligent calculation and result output. According to the engineering quantity calculation method, device and equipment based on drawing AI intelligent identification, the efficiency is remarkably improved, full-process automation from drawing import to engineering quantity output is achieved, the calculation efficiency is effectively improved compared with manual work, and the problem that a traditional method is long in time consumption is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of building engineering and artificial intelligence, in particular to an engineering quantity calculation method, device and equipment based on AI intelligent identification of drawings. BACKGROUND

[0002] Engineering quantity calculation is the core link of building engineering bidding, cost accounting and construction management, and its accuracy and efficiency directly determine the effect of engineering cost control. Traditional engineering quantity calculation relies on manual reading and manual statistics, and there are three major problems: first, the efficiency is low, and it often takes several days or even weeks for cost personnel to complete the engineering quantity calculation of a set of complex building drawings, which is difficult to meet the engineering progress demand; second, the error rate is high, manual identification is easily affected by factors such as drawing clarity, individual experience difference and visual fatigue, and the problems of omission and miscalculation occur frequently, which may cause the engineering cost deviation to be more than 10%; third, the compatibility is poor, and different formats of drawings such as CAD, PDF and BIM need to be converted and processed by multiple tools, and information loss may occur in the data flow process.

[0003] In the prior art, although AI-based drawing identification technology has appeared, there are still obvious limitations: first, the drawing feature extraction is rough, and traditional CNN (convolutional neural network) is usually used to identify only simple graphic elements (such as straight lines and circles), which cannot accurately distinguish between building components (such as beams, columns and walls) and label information (such as size and material), and it is even more difficult to understand the spatial relationship between components; second, the model generalization ability is weak, and different types of buildings (such as residential buildings, factory buildings and bridges) require separate model training, and the migration adaptability is poor; third, the engineering quantity calculation logic is disconnected from the AI identification, and the identification results need to be manually imported into the cost software for secondary conversion, and the integrated process of "identification-calculation-output" has not been formed. For example, the patent with publication number CN114548652A proposes a drawing identification method based on deep learning, but only realizes component contour extraction and does not involve automatic calculation of engineering quantity; the patent with publication number CN115115258B focuses on the engineering quantity statistics of BIM model and relies on preset model parameters, which cannot directly process two-dimensional drawings.

[0004] Therefore, in view of the existing structure and deficiencies, an engineering quantity calculation method, device and equipment based on AI intelligent identification of drawings are proposed. SUMMARY

[0005] The purpose of the present application is to provide an engineering quantity calculation method, device and equipment based on AI intelligent identification of drawings to solve the problems raised in the background art.

[0006] To achieve the above purpose, the present application provides the following technical scheme: an engineering quantity calculation method based on AI intelligent identification of drawings, the calculation method comprising the following steps: S1. Preprocessing and multi-format parsing of drawings: Obtain the engineering drawings to be processed, and perform differentiated parsing for CAD, PDF, and picture formats to output standardized drawing data. S2. Multi-modal component feature extraction: Extract spatial and semantic features of standardized drawings through a CNN-TransformerFusionModel model, and output multi-modal fusion features through attention gate units. S3. Cross-scale component recognition and classification: Perform cross-scale recognition on multi-modal fusion features based on a CSA-YOLOv8 model to output component categories, bounding box coordinates, and confidence. S4. Component association reasoning and parameter completion: Construct a component association graph through a GNN association reasoning model, complete missing parameters, and verify them combined with building specifications to output a standardized component parameter table. S5. Intelligent engineering quantity calculation and result output: Build differentiated calculation sub-models based on component types, and output engineering quantity lists and visual reports combined with random forest correction models.

[0007] Further, in S1, CAD format drawings are extracted through OpenCADSDK to extract vector data and convert them into standardized geometric matrices; PDF format drawings are fused with vector layer parsing and OCR text extraction results; and picture format drawings are normalized in resolution and processed through Retinex light correction.

[0008] Further, in S2, in the CNN-TransformerFusionModel model, the CNN branch is an improved ResNet-50 network, and the 3rd and 4th convolution layers are introduced into the model; the Transformer branch adopts a BERT-base model to generate text word embedding vectors, which are processed through a 6-layer Transformer encoder; and feature fusion is dynamically calculated through an attention gate unit.

[0009] Further, in S3, the CSA-YOLOv8 model constructs an FPN+PAN cross-scale feature pyramid, introduces channel-spatial dual attention modules at each scale feature layer, adopts an improved CIoU loss function for training, and realizes the recognition and classification of 12 types of building components.

[0010] Further, in S4, the GNN association reasoning model uses the GraphSAGE algorithm to iteratively update the features of the component association graph, the component association graph takes components as nodes and spatial associations as edges, the node features include category, size parameter, and position information, and the edge features include spatial distance and connection type.

[0011] Further, the entity component in S5 is calculated based on a volume formula, the reinforcement component is calculated using the "number x single length x theoretical weight" mode, and the plan component is calculated by converting the pixel area and deducting the opening. The calculation result is corrected by a random forest regression model.

[0012] Further, the engineering quantity calculation device based on AI intelligent recognition of drawings comprises: A drawing analysis module configured to perform drawing preprocessing and multi-format analysis steps, supports standardized processing of CAD, PDF, picture multi-format drawings, and outputs standardized drawing data; A feature extraction module: built-in CNN-TransformerFusionModel model, configured to perform multi-modal feature extraction on standardized drawing data, and output multi-modal fusion features of fusion spatial features and semantic features; A component recognition module: equipped with a CSA-YOLOv8 model, configured to realize cross-scale component recognition and classification based on multi-modal fusion features, and output component category and bounding box coordinate basic information; An associated reasoning module: containing a GNN component association reasoning model and a specification rule library, configured to realize component association analysis, parameter completion and verification, and output a standardized component parameter table; An engineering quantity calculation module: integrating multi-type component calculation sub-models and random forest correction models, configured to perform engineering quantity calculation, result fusion correction, and output standardized engineering quantity list and visual report.

[0013] Further, the engineering quantity calculation device based on AI intelligent recognition of drawings comprises a processor, a memory and a computer program stored in the memory, and the processor executes the computer program to realize the steps of the method.

[0014] The present application provides an engineering quantity calculation method, device and equipment based on AI intelligent recognition of drawings, which has the following advantages: Efficiency is significantly improved: the whole process from drawing import to engineering quantity output is automated, the calculation efficiency is effectively improved compared with manual calculation, and the problem of long time consumption in traditional methods is solved; High recognition and calculation accuracy: multi-modal feature extraction and cross-scale attention mechanism can significantly improve the accuracy of component recognition, combined with GNN association reasoning and specification verification, the engineering quantity calculation error rate is much lower than that of manual calculation; Strong generalization ability: the CNN-Transformer fusion model and the cross-scale recognition architecture are suitable for different building type drawings such as residential buildings, factories and bridges, and the model does not need to be retrained when migrated, only the rule library needs to be updated; Good compatibility: support multi-format drawing analysis, output results can be directly connected to mainstream cost software, break down data flow barriers, reduce the operation threshold of cost personnel. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 The flowchart of the engineering quantity calculation method based on drawing AI intelligent recognition of the application; Figure 2 The schematic diagram of the engineering quantity calculation device based on drawing AI intelligent recognition of the application. DETAILED DESCRIPTION

[0016] Example 1 Please refer to Figure 1 The application provides a technical solution: an engineering quantity calculation method based on drawing AI intelligent recognition, which includes the following steps: S1. Drawing preprocessing and multi-format analysis: obtain the engineering drawing to be processed, perform differential analysis for CAD, PDF, and picture formats, and output standardized drawing data; The CAD format drawing extracts vector data through OpenCADSDK and converts it into a standardized geometric matrix (including component contour coordinates, layer information, and label text); The PDF format drawing fuses vector layer analysis and OCR text extraction results, uses PyPDF2 to analyze the vector layer, combines OCR technology (Tesseract5.0) to extract text information from the raster layer, and realizes the fusion of vector and raster data; The picture format drawing is normalized in resolution and corrected in Retinex lighting, specifically through a bilinear interpolation algorithm for resolution normalization (uniformly adjusted to 300dpi), and uses the Retinex algorithm to eliminate uneven lighting interference, and outputs the preprocessed standardized drawing data.

[0017] S2. Multi-modal component feature extraction: extract the spatial and semantic features of the standardized drawing through the CNN-TransformerFusionModel model, and output the multi-modal fusion features through the attention gate unit; In the CNN-TransformerFusionModel model in S2, the CNN branch is an improved ResNet-50 network, and the first 3 and 4 convolution layers are introduced into the hollow convolution; the Transformer branch uses the BERT-base model to generate text word embedding vectors, which are processed by the 6-layer Transformer encoder; the feature fusion is dynamically calculated by the attention gate unit; Specifically: a multi-modal feature extraction model that combines CNN and Transformer, namely CNN-TransformerFusionModel, is constructed to extract features from standardized drawing data. The core formula and application method are as follows: Improved ResNet-50 Dilation Convolution Formula and Application: Dilation Convolution expands the receptive field by introducing holes (zero padding) in the convolution kernel. The formula is: where, is the output feature image pixel value, is the input feature, is the convolution kernel weight, is the convolution kernel size, is the hole rate (3rd convolution layer , 4th convolution layer ), , is the convolution kernel size; Application Method: Standardized drawing data (size 256x256x3) is input into the improved ResNet-50. The first two convolution layers use regular convolution ( ) to extract basic texture features. The 3rd convolution layer uses dilation convolution to capture medium-sized component outlines such as beams and columns. The 4th convolution layer uses dilation convolution to cover large-sized component areas such as shear walls and floors. Finally, 1x1 convolution is used to compress the channel number, with the compression formula: where, is the output of the 4th convolution layer, is the convolution kernel with size 1x1x1024x256 (realizing 1024-dimensional to 256-dimensional channel dimension reduction), is the bias term, and the final output dimension is 64x64x256 spatial feature map, is the spatial feature map output by the CNN branch, provides high-quality spatial feature support for subsequent multi-modal fusion.

[0018] Transformer Branch Word Embedding and Self-Attention Formula: Word Embedding: The labeled text (e.g. "KL1300x600") is split into word units ("KL1" "300" "x" "600"), which are converted into word embedding vectors through the BERT-base model. The formula is: where, is the word embedding vector (combining word table and position information), is the word table embedding (dimension 768), is the position encoding, which uses the sine position encoding formula: where, is the token position, is the dimension index, is the embedding dimension; Self-attention calculation: Capture the correlation between tokens by 8 heads of self-attention, and the single-head attention formula is: where, , , , (single-head dimension 96 = 768 / 8), is the Q / K dimension. Multi-head attention outputs by concatenating the results of 8 heads: where, is the output projection matrix, and the final Transformer branch outputs semantic features , is the number of tokens.

[0019] To achieve dynamic fusion of spatial and semantic features, an attention gating unit is introduced, and the weight calculation and fusion formula is: , where, is the spatially aligned feature of the semantic feature after transposed convolution (size 64x64x256), , is the weight matrix, is the bias, is the element-wise multiplication. The weight matrix G in the interval [0, 1] is generated by the Sigmoid function, which dynamically allocates the contribution of the two types of features, and the final output is .

[0020] Transformer branch: Convert the text annotations in the drawing (such as "KL1300x600") into word embedding vectors (use BERT-base model, output 768-dimensional vectors), combine position encoding, and input into the Transformer encoder (6-layer encoder, 8 heads of self-attention) to capture text semantics and position correlation features; Feature fusion: Weighted fusion of spatial features output by CNN and semantic features output by Transformer through feature attention gating unit (Attention Gating Unit), weight coefficient is dynamically calculated by sigmoid function, output multi-modal fusion features (size 64x64x512).

[0021] S3. Cross-scale component recognition and classification: Based on the CSA-YOLOv8 model, the multi-modal fusion features are recognized across scales, and the component class, bounding box coordinates and confidence are output; In S3, the CSA-YOLOv8 model constructs a FPN+PAN cross-scale feature pyramid, introduces a channel-space dual attention module at each scale feature layer, uses an improved CIoU loss function for training, and realizes the identification and classification of 12 types of building components. Based on the multi-modal fusion features, a cross-scale attention target detection model (CSA-YOLOv8) is used to realize component recognition, and the core formula and application method are as follows: Cross-scale feature pyramid (FPN+PAN) feature fusion formula: FPN is responsible for passing down high semantic features, PAN is responsible for supplementing high-resolution details from bottom to top, and the fusion formula is: FPN up-sampling fusion: ; PAN down-sampling fusion: ; Wherein, is the feature after FPN up-sampling fusion, is the feature after PAN down-sampling fusion, is the feature of the layer feature map (corresponding to 16x16, 32x32, 64x64 scale), DownSample uses a 3x3 convolution with a step of 2, and Conv is a 1x1 convolution (the number of channels is unified to 256). When applied, first up-sample the 16x16 high semantic features to 32x32 and 64x64 through FPN, and fuse them with the corresponding scale features; then down-sample the 64x64 high-resolution features back through PAN, and finally output three scales of enhanced feature maps (S=16, 32, 64), is the enhanced feature map output by FPN+PAN.

[0022] Channel-space dual attention formula: SE channel attention: strengthen key channels through squeeze-excitation mechanism, formula: , , Wherein, is the global feature (dimension 256), , are the dimension reduction and dimension increase matrices, is the channel weight, is the feature after SE processing; CBAM spatial attention: focus on the outline area of the component, formula: , , Wherein, is the channel concatenation (output dimension 2), Conv is a 3x3 convolution (output dimension 1), For the spatial weight map. When applied, first perform SE channel weighting on the PAN output features, then strengthen the spatial contour features through CBAM, and output the cross-scale attention features ; Improved CIoU loss function: Introduce a class weight coefficient to optimize the class imbalance problem, the formula is: , where, is the intersection over union of the predicted box and the real box, is the Euclidean distance between the predicted box b and the real box is the diagonal length of the minimum enclosing rectangle of the two boxes, is the balance coefficient, is the aspect ratio consistency parameter, is the class weight (small components , large components ). When applied, use as the target function to train the model, minimize the loss through the SGD optimizer (momentum 0.9, weight decay 1e-4), and realize precise positioning and classification of 12 types of components.

[0023] Introduce channel-spatial dual attention modules at each scale feature layer: channel attention strengthens key feature channel weights through squeeze-excitation (SE) mechanism, and spatial attention highlights component contour regions through convolution attention module (CBAM); Use the improved CIoU loss function (introduce a class weight coefficient) to train the model, realize precise identification of 12 types of core building components (beams, columns, walls, plates, foundations, steel bars, doors and windows, stairs, lintels, eaves, canopies, and embedded parts), and output component categories, bounding box coordinates, and confidence (confidence threshold set to 0.85).

[0024] S4. Component association reasoning and parameter completion: build a component association graph through a GNN association reasoning model, complete missing parameters and verify them combined with building codes, and output a standardized component parameter table; The GNN association reasoning model in S4 uses the GraphSAGE algorithm to iteratively update the component association graph features. The component association graph takes components as nodes and spatial associations as edges. Node features include category, size parameters, and location information, and edge features include spatial distance and connection type. Specifically, the GraphSAGE node feature update formula is: , which uses the mean aggregation function: where, is the node feature at the layer, is​ a set of neighbor points of the current point, a first layer weight matrix, a first layer weight matrix, a ReLU activation function. When applied, the components are taken as nodes (initial feature The components with a category, a known size, and the like 5-dimensional information, a spatial distance <1 m, and an existing connection relationship are neighbors, and an undirected graph is constructed; the node features are iteratively updated through 2 layers of GraphSAGE, and the updated input a full connection layer (the output dimension is equal to the number of missing parameters), to realize parameter completion (for example, the span of a beam = the spacing of neighbor columns); The specification rule library checks the formula: a parameter legality check function is constructed based on the “Code for Design of Concrete Structures”, and a beam section size check is taken as an example: wherein is the beam width, is the beam height, is the beam net span. When applied, the completed parameters are executed one by one to the corresponding check function, and if an exception handling is triggered: similar component parameters are matched in the historical database through a K-nearest neighbor algorithm (K=5), the abnormal value is replaced, and a correction log is recorded, and finally a standardized parameter table is output; Correlation reasoning: the GraphSAGE algorithm is used to iteratively update the component correlation graph, and the missing parameters are completed by aggregating the neighbor node features (for example, the unmarked beam section height can be derived through the height of the associated column and the specification constraint); Parameter checking: a rule library is constructed in combination with the building engineering specifications (such as the “Code for Design of Concrete Structures” GB50010-2010), the completed parameters are checked, the abnormal values exceeding the specification range are removed, and a standardized component parameter table (including component ID, category, material, size parameter, and position information) is output.

[0025] S5. Intelligent calculation of engineering quantities and output of results: a differential calculation submodel is constructed based on the component type, and an engineering quantity list and a visual report are output in combination with a random forest correction model.

[0026] In S5, the entity components are calculated based on the volume formula, the reinforcement components are calculated in the “number x single length x theoretical weight” mode, the planar components are calculated through pixel area conversion and opening deduction, and the calculation results are corrected by a random forest regression model.

[0027] A differential calculation submodel and an error correction model are constructed, and the core formula and the application method are as follows: Core formula for component engineering quantity calculation: the volume of an entity component: for a regular section, the formula is used, is the cross-sectional area, is the length, For regular cross-sectional areas, S is calculated using the coordinate method for irregular cross-sections (such as irregularly shaped columns). ,in, The coordinates of the cross-section vertex (sorted clockwise) ); Reinforcement work quantity Among them, the number of roots ( For the length of the component, For the thickness of the protective layer, (Reinforcement spacing), single bar length ( For anchorage length, (This refers to the overlap length, taken according to GB50010-2010), and the theoretical weight. ( (The diameter of the reinforcing bar is in mm). Area of ​​planar components: ,in, This is the scale conversion factor. Area of ​​door and window openings The pixel area of ​​the component; Random Forest Error Correction Formula: This formula corrects systematic errors through ensemble learning of multiple decision trees. ,in, For the number of decision trees, The calculation results for the sub-model It is a feature vector (including component type, size, material, etc.). Let be the prediction function of the t-th decision tree. This is the corrected bill of quantities. In application, the model is trained using the "sub-model result - actual result - feature vector" data from historical data. Gini coefficients are used to select split nodes, and the final output is the error-corrected bill of quantities. Reinforcing steel members: The calculation model of "number of bars × length of a single bar × theoretical weight" is adopted. The information of reinforcing steel bars marked in the drawings (such as "Φ12@200") is analyzed by NLP technology, and the number of bars is calculated in combination with the length of the member. The length of a single bar is derived from the member size and anchorage length specifications. For planar components (floor slabs, roofs): the pixel area conversion method is used, and the identified component pixel area is converted into the actual area by combining the drawing scale. The net area is obtained after deducting the area of ​​door and window openings. Results fusion and correction: A random forest regression model is introduced to fuse the calculation results of each sub-model. The model is trained with historical engineering data (100,000+ samples) to learn error correction rules and reduce systematic errors. Output format: Generates bill of quantities (Excel format) and visual reports (including component distribution heat map and quantity percentage pie chart), and supports direct import into mainstream cost estimation software such as Glodon and Luban.

[0028] Example 2 Please see Figure 2 A quantity calculation device based on AI-powered drawing recognition, comprising: Drawing parsing module: Configured to perform drawing preprocessing and multi-format parsing steps, supporting standardized processing of CAD, PDF, and image-based drawings, and outputting standardized drawing data; Feature extraction module: Built-in CNN-TransformerFusionModel, configured to perform multimodal feature extraction on standardized drawing data, outputting multimodal fusion features that integrate spatial and semantic features; Component recognition module: Equipped with CSA-YOLOv8 model, configured to achieve cross-scale component recognition and classification based on multimodal fusion features, and output component category and bounding box coordinates. The associative reasoning module includes a GNN component associative reasoning model and a standardized rule base, configured to perform component associative analysis, parameter completion and verification, and output a standardized component parameter table. The quantity calculation module integrates multi-type component calculation sub-models and random forest correction models, and is configured to perform quantity calculations, result fusion and correction, and output standardized quantity lists and visualization reports.

[0029] Example 3 An engineering quantity calculation device based on AI-powered drawing recognition includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it implements the steps of the method.

[0030] The embodiments of the present invention are given for illustrative and descriptive purposes only, and are not intended to be exhaustive or to limit the invention to the forms disclosed. Many modifications and variations will be apparent to those skilled in the art. The embodiments were chosen and described in order to better illustrate the principles and practical application of the invention, and to enable those skilled in the art to understand the invention and to design various embodiments with various modifications suitable for a particular purpose.

Claims

1. A method for calculating engineering quantities based on AI-powered intelligent recognition of drawings, characterized in that: The measurement method includes the following steps: S1. Drawing Preprocessing and Multi-Format Parsing: Obtain the engineering drawings to be processed, perform differentiated parsing for different formats such as CAD, PDF, and images, and output standardized drawing data; S2. Multimodal component feature extraction: Spatial and semantic features of standardized drawings are extracted using the CNN-TransformerFusionModel, and multimodal fusion features are output through attention gating units; S3. Cross-scale component identification and classification: Based on the CSA-YOLOv8 model, cross-scale identification is performed on multimodal fusion features, and the component category, bounding box coordinates and confidence score are output. S4. Component Association Reasoning and Parameter Completion: Construct a component association graph through the GNN association reasoning model, complete the missing parameters, and combine them with building code verification to output a standardized component parameter table; S5. Intelligent Calculation and Output of Quantities: Based on component type, a differentiated calculation sub-model is constructed, and combined with the random forest correction model, the bill of quantities and visualization reports are output.

2. The engineering quantity calculation method based on AI intelligent recognition of drawings according to claim 1, characterized in that, In S1, CAD format drawings extract vector data using OpenCADSDK and convert it into a standardized geometric matrix; PDF format drawings integrate vector layer parsing and OCR text extraction results. Image format drawings have been processed with resolution normalization and Retinex illumination correction.

3. The method for calculating engineering quantities based on AI intelligent recognition of drawings according to claim 1, characterized in that, In the CNN-TransformerFusionModel model in S2, the CNN branch is an improved ResNet-50 network, with dilated convolutions introduced in the 3rd and 4th convolutional layers; the Transformer branch uses the BERT-base model to generate text word embedding vectors, which are then processed by a 6-layer Transformer encoder. Feature fusion dynamically calculates weights through attention gating units.

4. The engineering quantity calculation method based on AI intelligent recognition of drawings according to claim 3, characterized in that, In S3, the CSA-YOLOv8 model constructs an FPN+PAN cross-scale feature pyramid, introduces a channel-spatial dual attention module in each scale feature layer, and uses an improved CIoU loss function for training to achieve the recognition and classification of 12 types of building components.

5. The method for calculating engineering quantities based on AI intelligent recognition of drawings according to claim 1, characterized in that, The GNN association inference model in S4 uses the GraphSAGE algorithm to iteratively update the component association graph features. The component association graph has components as nodes and spatial associations as edges. Node features include category, size parameters and location information, and edge features include spatial distance and connection type.

6. The method for calculating engineering quantities based on AI intelligent recognition of drawings according to claim 1, characterized in that, In S5, solid components are calculated based on volume formulas, steel reinforcement components are calculated using the "number of bars × single bar length × theoretical weight" model, and planar components are calculated through pixel area conversion and opening deduction. The calculation results are corrected by a random forest regression model.

7. A quantity calculation device based on AI-powered drawing recognition, applied to the quantity calculation method based on AI-powered drawing recognition as described in any one of claims 1-6, characterized in that, The device includes: Drawing parsing module: Configured to perform drawing preprocessing and multi-format parsing steps, supporting standardized processing of CAD, PDF, and image-based drawings, and outputting standardized drawing data; Feature extraction module: Built-in CNN-TransformerFusionModel, configured to perform multimodal feature extraction on standardized drawing data, outputting multimodal fusion features that integrate spatial and semantic features; Component recognition module: Equipped with CSA-YOLOv8 model, configured to achieve cross-scale component recognition and classification based on multimodal fusion features, and output component category and bounding box coordinates. The associative reasoning module includes a GNN component associative reasoning model and a standardized rule base, configured to perform component associative analysis, parameter completion and verification, and output a standardized component parameter table. The quantity calculation module integrates multi-type component calculation sub-models and random forest correction models, and is configured to perform quantity calculations, result fusion and correction, and output standardized quantity lists and visualization reports.

8. A quantity calculation device based on AI-powered drawing recognition, applied to the quantity calculation method based on AI-powered drawing recognition as described in any one of claims 1-6, characterized in that, It includes a processor, a memory, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method according to any one of claims 1-6.

Images