Distribution network engineering material image-text intelligent checking method based on large model

Abstract

The application discloses a kind of based on big model's distribution network engineering material figure and text intelligent checking method, it is related to distribution network engineering figure and text checking technical field, comprising the following steps: using pre-training big model to the multi-modal engineering data of to-be-checked distribution network engineering is parsed, and engineering characteristic information is extracted;According to the topographic and geomorphic features of engineering site, generate construction environment constraint, according to engineering characteristic information, construct node connectivity capacity constraint and safety distance threshold value;Based on the geometric primitives in engineering characteristic information, construct engineering topology, and according to construction environment constraint, node connectivity capacity constraint and safety distance threshold value, each geometric primitive in the engineering topology is verified, and conflict set and conflict label are generated;Based on the conflict set and conflict label, generate multidimensional review opinion.The application is used to solve the problem that path evaluation is separated from actual environment, equipment connection lacks capacity verification and review dimension is single.

Description

Technical Field

[0001] This invention relates to the field of graphic verification technology for distribution network engineering, and more specifically, to an intelligent graphic verification method for distribution network engineering materials based on a large model. Background Technology

[0002] Current power distribution network engineering design reviews primarily employ parameter extraction from drawings based on preset rules, or combine multimodal large-scale models for auxiliary graphic and textual analysis. For example, invention application CN122369051A discloses a method and system for identifying illegal topological relationships in power drawings based on a large-scale model. This method performs joint semantic analysis of power drawings and accompanying text, extracts coordinates for geometric modeling, and verifies whether the actual spatial parameters between equipment meet preset thresholds. Another example is invention application CN122153041A, which discloses a data processing method and system based on a multimodal large-scale model. This method constructs an undirected graph by identifying electrical components in the drawings, uses graph algorithms to detect connectivity and calculate electrical distances, and combines knowledge retrieval and reasoning to generate review results containing problem location and modification suggestions.

[0003] However, the aforementioned existing technologies still have some technical problems in practical applications: First, route feasibility and safety assessments are mostly based on geometric conditions on drawings. There is a lack of effective joint verification methods for the matching relationship between factors such as terrain undulations and surface cover in the actual areas crossed by the route and the mechanical parameters of materials. The rationality of the span configuration and construction feasibility are difficult to fully confirm during the design phase. Second, the correspondence between the equipment configuration scheme in the design drawings and the actual engineering resources is difficult to verify. There is a lack of a comparison mechanism between the degree of matching between the number of equipment, node connection relationships, and the upper limit of physical resources, which can easily lead to a disconnect between the design scheme and the on-site material conditions. Third, when faced with multiple types of complex anomalies, it is difficult to conduct structured sorting and comprehensive analysis of various conflicts, thus making it impossible to directly generate comprehensive review opinions for multi-dimensional collaborative assessment. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method for intelligent verification of distribution network engineering materials based on large models. This method generates construction environment constraints through terrain features, verifies the engineering topology by combining safety distance thresholds and node connectivity capacity constraints, and then generates conflict labels and multi-dimensional review opinions to solve the problems of path assessment being detached from the actual environment, lack of capacity verification for equipment connections, and single review dimensions in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A method for intelligent verification of materials in power distribution network engineering based on large-scale models includes the following steps: A pre-trained large model is used to analyze the multimodal engineering data of the power distribution network project to be verified, and the engineering feature information is extracted. Construction environment constraints are generated based on the topographic features of the engineering site, and node connectivity capacity constraints and safety distance thresholds are constructed based on engineering feature information. An engineering topology is constructed based on geometric primitives in the engineering feature information. Constraints are verified on each geometric primitive in the engineering topology according to construction environment constraints, node connectivity capacity constraints and safety distance thresholds, and conflict sets and conflict labels are generated. Based on the aforementioned conflict set and conflict tags, multi-dimensional review opinions are generated.

[0006] In a preferred embodiment, the multimodal engineering data includes design documents, engineering drawings, and corresponding standard equipment and material parameters; the engineering feature information includes engineering semantic information, discrete geometric primitives, and a set of standard parameters.

[0007] In a preferred embodiment, generating construction environment constraints includes: acquiring elevation data and surface cover data of the engineering site, extracting slope and surface cover type at each spatial location, and constructing a spatial cost matrix; calculating the maximum allowable span based on the material mechanical performance parameters corresponding to the engineering semantic information, and converting the maximum allowable span into a path cost threshold; the spatial cost matrix and the path cost threshold together constitute the construction environment constraints.

[0008] In a preferred embodiment, constructing the spatial cost matrix includes: obtaining the basic cost value corresponding to each surface cover type based on a preset engineering construction quota; multiplying the basic cost value by the slope as a weighting coefficient to obtain the construction cost value of each spatial location, thereby forming the spatial cost matrix.

[0009] In a preferred embodiment, the construction of node connectivity capacity constraints and safety distance thresholds includes: using the number of various types of equipment recorded in the equipment ledger in the engineering semantic information as the upper limit of the number of nodes of the same type in the engineering topology; extracting the number of interfaces of the corresponding equipment from the standard parameter set according to the specifications and models of the equipment in the engineering semantic information as the upper limit of node degree; and using the upper limit of quantity and the upper limit of node degree to constitute the node connectivity capacity constraint; and scaling the electrical spacing standard value in the standard parameter set to the coordinate system of the discrete geometric primitives according to the scale of the engineering drawings to obtain the safety distance threshold.

[0010] In a preferred embodiment, the constraint verification of each geometric element in the engineering topology includes: establishing a mapping relationship between each geometric element and equipment entities in the engineering feature information; in the path analysis of the engineering topology, mapping each geometric element to the spatial cost matrix, and calculating the cumulative construction cost of the path corresponding to each geometric element; obtaining the current number of connections of the associated nodes of the geometric element, the number of nodes mapped to the current equipment model, and the map spacing between adjacent geometric elements; comparing and verifying the cumulative construction cost, current number of connections, number of mapped nodes, and map spacing with the path cost threshold, node degree upper limit, number of similar nodes upper limit, and safety distance threshold, respectively, and calculating the deviation data of the geometric elements that fail the verification in the corresponding verification item.

[0011] In a preferred embodiment, generating the conflict set and corresponding conflict labels includes: taking the geometric primitives that failed the verification as conflict primitives, extracting the engineering semantic information that has an entity mapping relationship with the conflict primitives, and constructing a conflict set; binding conflict labels to the conflict primitives according to the type of constraint violated, the conflict labels including connectivity exceedance label, cost exceedance label, and spacing violation label.

[0012] In a preferred embodiment, generating multi-dimensional review opinions based on the conflict set and conflict labels includes: extracting corresponding connectivity deviation data, spacing deviation data, and cumulative construction cost deviation data based on the conflict labels; calculating the semantic similarity between technical parameters and standard specification parameters in the engineering semantic information, and the difference between the current operating years and the preset retirement years recorded in the design documents in the multimodal engineering data; calculating the spatial coverage of the engineering topology after topological connectivity and the spatial overlap between the area to be renovated recorded in the design documents; and inputting the deviation data, semantic similarity, age difference, and spatial overlap into the pre-trained large model to generate the multi-dimensional review opinions.

[0013] In a preferred embodiment, the extraction of engineering semantic information, discrete geometric primitives, and standard parameter sets includes: performing natural language parsing on design documents and standard equipment material parameters in multimodal engineering data to obtain engineering semantic information and standard parameter sets; and performing vectorization processing on engineering drawings in multimodal engineering data to obtain discrete geometric primitives.

[0014] In a preferred embodiment, the training process of the pre-trained large model includes: acquiring historical multimodal engineering data in the field of power distribution network engineering, combining the corresponding structured annotation results, and performing cross-modal understanding training on the basic large model; using historical conflict sets and historical conflict labels as inputs and historical multidimensional review opinions as the target output, performing supervised training on the basic large model after cross-modal understanding training to obtain the pre-trained large model.

[0015] The technical effects and advantages of the intelligent verification method for distribution network engineering materials based on large models in this invention are as follows: 1. This invention generates construction environment constraints by acquiring the topographic features of the engineering site, and introduces these constraints into the constraint verification of the engineering topology. It establishes a joint verification mechanism between the drawing design data and physical environment elements, ensuring the feasibility of the distribution network design scheme and improving the scientific nature and construction safety of power grid infrastructure engineering planning.

[0016] 2. This invention constructs node connectivity capacity constraints by parsing engineering semantic information and engineering attributes in the standard parameter set, and verifies the engineering topology based on these constraints. It establishes a correlation and comparison mechanism between the network characteristics of nodes in the drawings and the upper limit of the physical attributes of engineering equipment, reducing the risk of disconnect between the logic of the drawings and physical resources, and ensuring the accuracy of the allocation of power grid engineering materials and the efficient progress of on-site construction.

[0017] 3. This invention generates multi-dimensional review opinions based on conflict sets and conflict tags, transforming the underlying physical conflicts and rule verification results into structured comprehensive guidance information, thereby improving the system's ability to intelligently diagnose and sort out complex design defects, and providing reliable decision support for the quality control and digital review of power grid engineering design. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of a method for intelligent verification of materials in power distribution network engineering based on a large model, provided in an embodiment of the present invention.

[0019] Figure 2 This is a schematic diagram of geometric primitive bounding box prediction and semantic binding based on a large model, provided for an embodiment of the present invention.

[0020] Figure 3 This is a schematic diagram of the mapping of geometric primitives to the spatial cost matrix and the corresponding path discretization provided in the embodiments of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0022] Example 1, Figure 1 This invention presents a method for intelligent verification of materials in power distribution network engineering based on a large model, comprising the following steps: Step 1: Use a pre-trained large model to analyze the multimodal engineering data of the power distribution network project to be verified and extract the engineering feature information.

[0023] The multimodal engineering data includes design documents, engineering drawings, and corresponding standard equipment and material parameters; the engineering feature information includes engineering semantic information, discrete geometric primitives, and a set of standard parameters.

[0024] 1.1 The training process of a pre-trained large model includes: 1.1.1 The architecture of the basic large model is as follows: The underlying network architecture of this basic large model adopts a two-stream Transformer structure, which includes a visual encoder, a text encoder, a cross-modal cross-attention layer, and an autoregressive decoding network composed of multiple layers (e.g., 12 or 24 layers) of Transformer decoders stacked together. Among them, the encoder and decoder both introduce multi-head self-attention mechanism, feedforward neural network (FFN), layer normalization module and GeLU activation function to ensure the non-linear fitting ability of feature extraction and gradient propagation stability.

[0025] The data flow process of this network architecture is as follows: the visual encoder is responsible for mapping the input engineering drawing image blocks into a sequence of visual feature vectors, and the text encoder lexicalizes the input design documents and corresponding standard equipment material parameters and maps them into a sequence of text semantic feature vectors of fixed dimensions. Subsequently, the vector sequences of these two modalities are input into the cross-modal attention layer. By calculating the attention weight distribution between text lexical units and image blocks, a deep association and fusion expression of engineering primitive features and text semantic features are established in the shared high-dimensional vector space. Finally, the fused multimodal features are input into the autoregressive decoding network to generate and output the engineering semantic information matrix, the standard parameter set, and the predicted two-dimensional continuous value coordinate vectors of discrete geometric primitives (i.e., bounding box coordinates).

[0026] 1.1.2 Obtain historical multimodal engineering data in the field of power distribution network engineering, and combine it with the corresponding structured annotation results to train the basic large model for cross-modal understanding, as follows: Collect historical multimodal engineering data in the field of power distribution network engineering, including more than 500,000 historical design documents, historical engineering drawings, corresponding standard equipment and material parameters, and corresponding structured annotation results pre-marked by experts. The structured annotation results include the text of the actual graphic element category, the corresponding actual bounding box coordinate data, and the equipment commissioning date, the survey boundary coordinates of the area to be renovated, the coordinates of the global control points on the drawings, the global scale, the text of the specific technical parameters of the equipment, and the asset's full life cycle retirement standards manually marked from the historical data.

[0027] The above raw data are preprocessed separately, specifically: (1) For historical engineering drawings, render and crop them into a matrix of image segments with a fixed resolution (e.g., 512×512 pixels), and cut them into non-overlapping segments according to a preset image segment size (e.g., 16×16 pixels). After flattening, map them into a sequence of visual feature vectors with a sequence length of 1024 through a linear projection layer. At the same time, perform pixel value normalization processing to make each pixel value within a uniform value range.

[0028] (2) For the historical design documents and the corresponding standard equipment material parameters, after denoising and cleaning, the two are spliced ​​into a complete engineering description text using a preset delimiter (such as [SEP]). Then, the Byte-Pair Encoding (BPE) algorithm is used to perform word segmentation and word embedding preprocessing on the engineering description text. The algorithm takes characters as the initial unit, iteratively merges adjacent character pairs in the corpus to generate a dictionary, and converts the text into a word sequence. The [CLS] identifier is inserted at the beginning of the word sequence after word segmentation. According to the maximum input length of the encoder, the maximum sequence length of the word sequence is set to 1024 identifiers. When the number of identifiers after text conversion exceeds 1024, the first 1024 are truncated and retained. When there are fewer than 1024, a 0-value placeholder is used at the end of the sequence to fill in the gaps, thereby converting each set of associated historical design documents and standard equipment material parameters into a 1×1024-dimensional text input matrix.

[0029] (3) For the structured annotation results, the real primitive category text is converted into a one-hot encoded vector matrix; the real bounding box coordinate data (including the pixel coordinates of the upper left and lower right corners) is divided by the image resolution and converted into a continuous value coordinate matrix distributed in the interval [0,1]; the manually calibrated equipment commissioning date, the survey boundary coordinates of the area to be modified, the coordinates of the global control points on the drawing, the global scale, the specific technical parameters of the equipment, and the asset life cycle retirement standards are converted into a continuous value coordinate matrix distributed in the interval [0,1] according to the preset JSON key-value pair template (the template example is: {"commissioning date":"[Date]","survey boundary":"[Coords]","control points":"[C The target attribute text is assembled from the following: trl_Coords]", "Scale":"[Scale]", "Technical Parameters":"[Tech_Spec]", "Retirement Standard":"[Retire_Std]"}. The byte-pair encoding (BPE) algorithm is also used to convert the text into a target attribute lexical sequence. The maximum length of the target attribute lexical sequence is set to 512 identifiers. If the number of identifiers exceeds 512, it is truncated; if it is insufficient, meaningless placeholders are used to pad the end of the sequence. The one-hot encoded vector matrix, the continuous value coordinate matrix, and the target attribute lexical sequence together serve as the supervision signal matrix for model training.

[0030] The preprocessed image segment matrix, fixed-dimensional text input matrix, and supervision signal matrix are mapped and aligned one-to-one to obtain a power distribution network engineering image-text alignment sample set. Based on this power distribution network engineering image-text alignment sample set, the basic large model is trained for cross-modal understanding. Specifically, the preprocessed image segment matrix and text input matrix are input into the visual encoder and text encoder, respectively, to extract normalized features. The network outputs predicted categories and bounding boxes, and the autoregressive decoding network outputs attribute text sequences. The training objective function is composed of a cross-modal image-text contrast loss term, a bounding box localization regression loss term, a primitive category classification loss term, and an attribute text generation loss term, as shown in the following formula: (1) (2) (3) (4) (5) in, The total loss function for joint training; For cross-modal image-text contrast loss function, This represents the total number of image-text sample pairs in the current training batch. For the first The normalized visual feature vector output by a visual encoder from each image sample. For the first The normalized text feature vector output by a text encoder from a text sample. , These are the first in this training batch. Each text feature vector and image feature vector This is a temperature hyperparameter used to control the scaling factor of the vector inner product discrimination. The bounding box localization regression loss function is calculated using mean squared error (MSE). This represents the number of subsamples containing labeled bounding boxes within this training batch. For the supervision signal matrix corresponding to the first [number] in the current batch The true continuous value coordinate vector of each sample The bounding box coordinate vector is the output of the large model prediction. The loss function for primitive category classification is calculated using cross-entropy; The total number of preset primitive categories, The one-hot encoded vector matrix of the true primitive category in the supervision signal matrix is ​​used to define the category. The label value (0 or 1) on it. To predict the category to which the sample belongs in a large model The probability value; A loss function is generated for the attribute text, and it is calculated using autoregressive cross-entropy. For the first The total length of the target attribute word sequence for each sample. For the first term in the target attribute word sequence Each word element, To generate the first The target attribute lexical sequence that has been generated before each lexical element; , and These are all preset hyperparameters used to balance the weights of each loss term, for example, with values ​​of 1.5, 1.0, and 1.0 respectively.

[0031] This training phase uses the AdamW optimizer, with the initial and minimum learning rates set to 1e-4 and 1e-6 respectively, and a cosine annealing learning rate decay strategy configured. The training continues until the total loss function value decreases by less than a preset debouncing threshold (e.g., set to 10) for five consecutive training epochs. -4 When convergence is reached, the model parameters are updated.

[0032] 1.1.3 Using the set of historical conflicts and their labels as input, and multi-dimensional historical review opinions as the target output, supervised training is performed on the base model trained on cross-modal understanding to obtain a pre-trained model, as detailed below: It should be noted that the dictionary data structure (such as key-value pairs containing primitive IDs, deviation values, and associated entities) and generation logic of the historical conflict set and historical conflict labels used as the training basis data are isomorphic to the conflict results generated in real time based on physical constraints in the subsequent online verification step, so as to ensure data consistency between model training and inference applications.

[0033] Collect over 100,000 historical review records of distribution network experts, which have undergone anonymization and include expert step-by-step logical deduction processes. Perform feature extraction calculations on the data in these historical review records to ensure consistency with subsequent online verification. Following a preset prompt template, concatenate the calculated deviation data with global engineering features to form an input instruction sequence. The prompt template is structured as follows: "[Input]: Please perform logical reasoning based on the following constraint violation features and global engineering features: <connectivity deviation, spacing deviation, and cumulative construction cost deviation data>, <difference between the current operating years of equipment and the preset retirement years>, <spatial overlap between the topology coverage area and the area to be renovated>, and <semantic similarity between engineering technical parameters and standard parameters>; [Output]: Please perform step-by-step logical deduction to generate a multi-dimensional comprehensive review opinion encompassing completeness, safety, cost rationality, standardization, necessity, and accuracy." Use the corresponding historical multi-dimensional review opinions as the target output sequence, constructing one-to-one instruction fine-tuning sample pairs in the format of <instruction, reply>. Based on the fine-tuned sample pairs according to the instructions, supervised training is performed on the base model after cross-modal understanding training. The objective function of the supervised training is as follows: (6) in, For the autoregressive cross-entropy loss function during the supervised training phase, The total length of the text terms for the target output (i.e., historical multi-dimensional review comments); The complete prompt instruction input sequence is obtained by concatenating all global features such as historical deviation data, year difference, spatial overlap, and semantic similarity; For the target output sequence, the first... Each word element, In order to generate the first All historical output word sequences that have been generated before each word. These are the model network weight parameters that are updated in real time during backpropagation. To add weight parameters to the model network The conditional probability of generating the current word is based on the input instruction and the preceding word.

[0034] This stage also uses the AdamW optimizer, with the initial learning rate set to 2e-5, and adopts a linear warm-up strategy (i.e., the learning rate is linearly increased from 0 to 2e-5 in several steps in the early stage of training). After three training rounds of iterative training, the final construction of the pre-trained large model is completed.

[0035] 1.2 Extract engineering semantic information, discrete geometric primitives, and standard parameter sets, including: 1.2.1 Natural language parsing is performed on the design documents and standard equipment and material parameters to obtain engineering semantic information and a set of standard parameters, as follows: During actual verification, the design documents and standard equipment material parameters undergo BPE segmentation and fixed-length padding preprocessing, consistent with the training phase, and are then input as text prompts into the text encoder of the pre-trained large model. Based on its supervised training and established domain knowledge representation capabilities, the large model extracts core entities and their associated attributes from unstructured text through its internal autoregressive decoding network, outputting a structured engineering semantic information matrix and a set of standard parameters.

[0036] The engineering semantic information matrix, in the form of JSON dictionary groups, structurally records the entity category, specifications, quantity, and mechanical design specifications related to cable selection for each piece of equipment in the design documents, as well as the commissioning date, technical parameter text of individual equipment, and the coordinates of the boundary of the area to be modified, the global scale, and the coordinates of global control points for the entire project. For example, the record for a regular piece of equipment is: {'Entity':'Ring Main Unit','Model':'One-in-Three-out Ring Main Unit','Quantity':5,'Commissioning Date':'2008-05-12','Technical Parameter Text':'Rated Voltage 12kV, Rated Current 630kV'} A'}; For overhead lines, the record is: {'Entity':'Overhead Conductor','Specification Model':'JKLYJ-10-240','Maximum Allowable Tension':'8000N','Design Maximum Sag':'1.2m','Comprehensive Load per Unit Length':'35.5N / m'}; For global spatial information, the record is: {'Global Information':{'Coordinates of the Boundary of the Area to be Modified':'[(x1,y1),(x2,y2)...]','Global Scale':'1:500','Coordinates of Global Control Points on the Drawing':'[(u1,v1),(u2,v2)...]'}}.

[0037] The standard parameter set records the interface specifications corresponding to different equipment models (such as "{'one-in-three-out ring main unit':'number of interfaces=4'}"), the asset life cycle specification dictionary (such as "{'one-in-three-out ring main unit':{'preset retirement age':'25 years','standard specification text':'rated voltage 12kV, phase-to-phase center distance...'}}"), and the standard values ​​of electrical spacing for each voltage level.

[0038] 1.2.2 The engineering drawings are vectorized to obtain discrete geometric primitives, as follows: For engineering drawings, a CAD graphics rendering engine (such as the Teigha library) is called to render them into raster images based on a preset viewport (i.e., the display boundary range and zoom center when the CAD drawing is exported as a raster image), and input them into the visual encoder of the pre-trained large model. The large model uses the core entities extracted from the text encoder as query conditions and outputs the predicted bounding box coordinates of the target object in the raster image through a cross-modal cross-attention mechanism.

[0039] Using a CAD parsing script, all entity primitives and their vertex coordinate data are extracted from the DXF / DWG group code file of the engineering drawing. Based on the minimum bounding rectangle (MBR) of each entity primitive, an R-tree spatial index structure is constructed from bottom to top. The scale matrix of the preset viewport is obtained, and the predicted bounding box coordinate vector is multiplied by the inverse of the scale matrix to convert the pixel-level predicted bounding box coordinates into the corresponding CAD world coordinates. The R-tree nodes are traversed, and entity primitives whose minimum bounding rectangles overlap with the CAD world coordinate query box are selected to extract the underlying CAD primitive data that falls within the coordinate query box.

[0040] The extracted underlying CAD drawing data (such as polylines, circles, blocks, and other semantically meaningless pure graphic objects) are attribute-bound to the predicted category labels output by the large model, thereby converting the aforementioned pure graphic objects into a set of discrete vector geometric primitives carrying spatial coordinate bounding boxes and category labels. For example... Figure 2 As shown in the figure, the large model outputs the predicted bounding box of the target object using the cross-modal cross-attention mechanism and binds it to the attribute of the underlying CAD geometric primitives. The black solid line in the figure represents the underlying CAD entities (such as ring main units, utility poles and wires) extracted by the R-tree spatial index, and the red dashed box represents the continuous value bounding box (AABB) predicted by the large model. This figure intuitively shows the deep association and binding process between primitive features and text semantics.

[0041] This step utilizes a pre-trained large model to perform cross-modal joint analysis of design documents and engineering drawings, enabling structured extraction of engineering graphic features. This avoids the information omissions that occur during traditional manual verification and provides standardized underlying data support for subsequent constraint verification.

[0042] Step 2: Generate construction environment constraints based on the topographic features of the engineering site, and construct node connectivity capacity constraints and safe distance thresholds based on engineering feature information.

[0043] 2.1 Generate construction environment constraints based on the topographical features of the project site, including: 2.1.1 Obtain elevation and land cover data from the engineering site, and extract the slope and land cover type for each spatial location, as detailed below: The data interface of the power grid geographic information system (GIS) is invoked to obtain the digital elevation model (DEM) raster data and remote sensing land cover layer of the construction area of ​​the distribution network project to be verified. A third-order finite difference algorithm (such as the Horn algorithm) is used, with a 3×3 neighborhood window centered on the current spatial grid node as input. By calculating the elevation change rate of the central node in the X and Y directions, the slope value of each spatial grid is obtained. The spatial coordinates of the remote sensing land cover layer are registered with the DEM raster, and the land cover type (such as plain, woodland, marshland, hard rock, etc.) of the corresponding coordinate grid is extracted.

[0044] 2.1.2 Based on the preset engineering construction quota, the basic cost value corresponding to each surface cover type is obtained. The basic cost value is multiplied by the slope as a weighting coefficient to obtain the construction cost value of each spatial location, forming a spatial cost matrix, as follows: Based on the budget quota standard library for power construction projects, the labor and machinery consumption for foundation pit excavation, material handling, and foundation pouring under different geological conditions in the quota standard library is normalized, and a mapping dictionary between surface cover type and basic cost value is constructed. Specifically, the basic cost value of plains and grasslands is set at 1.0 based on the consumption of plains and grasslands, and the basic cost value of forest land is determined to be 1.5, and the basic cost value of swamps and water bodies is 2.5, etc., according to the relative increase ratio of quota consumption, thereby transforming quota data into quantified basic cost value.

[0045] The construction cost at each spatial location is obtained by multiplying the slope value as a weighting factor by the foundation cost value, thus forming a spatial cost matrix. The calculation formula for the construction cost value at each location in the spatial cost matrix is ​​as follows: (7) in, Spatial coordinates The construction value of the site; The sensitivity coefficient, set based on engineering experience, is recommended to be 0.5. This represents the slope angle at that coordinate location; The type of land cover at this coordinate location; This represents the basic cost value corresponding to this land cover type.

[0046] 2.1.3 Based on the corresponding material mechanical property parameters in the engineering semantic information, calculate the maximum allowable span and convert the maximum allowable span into a path cost threshold, as follows: For underground cable engineering verification, the span calculation process is ignored, and the path cost threshold is set to infinity. For distribution network engineering involving overhead line laying, since the pre-trained large model has already associated and matched the physical limits scattered in the standard equipment and material parameters with the meteorological control conditions in the design documents during the analysis of design documents and standard equipment and material parameters in step 1, the mechanical design specifications related to conductor selection are extracted as technical parameters and structured into the matrix. Therefore, here, the conductor specification corresponding to the current geometric connection is directly used as the query key to index and extract the maximum allowable tension, maximum design sag, and comprehensive load per unit length of the corresponding cable entity from the engineering semantic information matrix. The maximum allowable span is calculated based on the mechanical performance parameters, as follows: (8) in, The maximum permissible gap is measured in meters. To design the maximum sag; This represents the maximum allowable tension of the conductor; here, the horizontal component of the allowable tension of the conductor is taken. To encompass the conductor's self-weight and pre-set additional loads such as wind deflection and ice cover, the vector composite value of the loads in each direction is taken here.

[0047] Subsequently, this maximum allowable gap is converted into a path cost threshold for a single path segment in the topology search, calculated using the following formula: (9) in, The path cost threshold; The dimensionless baseline cost weighting coefficient under standard plain terrain is set to, for example, 1.0, and used as the baseline conversion ratio under standard construction conditions.

[0048] 2.1.4 The spatial cost matrix and the path cost threshold together constitute the construction environment constraints, as detailed below: The construction cost value of each grid node in the spatial cost matrix is ​​used as the basic weight attribute of the spatial grid node in the underlying graph data structure; the path cost threshold is used as the physical extreme value upper limit of the weight of the connected edge between adjacent spatial nodes; the two are jointly encapsulated by data structure binding to form a set of construction environment constraints for subsequent graph theory search and constraint verification.

[0049] 2.2 Based on engineering characteristic information, construct node connectivity capacity constraints and safe distance thresholds, including: 2.2.1 The quantity of each type of equipment recorded in the equipment ledger in the engineering semantic information is used as the upper limit of the number of nodes of the same type in the engineering topology. Based on the specifications and models of the equipment in the engineering semantic information, the number of interfaces of the corresponding equipment is extracted from the engineering feature information as the upper limit of the node degree. The upper limit of the quantity and the upper limit of the node degree constitute the node connectivity capacity constraint, as follows: Based on the engineering semantic information matrix, the number of records corresponding to various equipment entities in the ledger is extracted as the upper limit of the number of such nodes (i.e., equipment entities with the same equipment name and specifications) in the global topology, which is used to establish a hard boundary between the number of equipment in the drawings and the actual project list.

[0050] For a single device node, its node degree (i.e., the actual number of branch lines connected) is limited by the total number of hardware terminals when the device leaves the factory. Therefore, using the device specification model in the engineering semantic information matrix as the association key, the corresponding interface specification in the standard parameter set is indexed to extract the actual number of interfaces for that device model, which serves as the upper limit of the node degree for that node instance in the engineering topology. For example, if a device is a "10kV one-in-three-out ring main unit," its number of interfaces is 4, meaning the upper limit of the node degree in the topology diagram is 4. This constraint can directly identify wiring errors in the diagram where the number of lines drawn exceeds the number of interfaces.

[0051] The upper limit on the number of nodes of the same type and the upper limit on the degree of a node together constitute the node connectivity capacity constraint.

[0052] 2.2.2 The electrical spacing standard values ​​in the standard parameter set are scaled to the coordinate system of the discrete geometric elements in the engineering drawings of the multimodal engineering data to obtain the safety distance threshold, as follows: From the standard parameter set, extract the standard values ​​of electrical spacing corresponding to each voltage level of the current distribution network project (for example, the standard value of the phase-to-phase safety distance for 10kV bare conductors is 0.7 meters, and for 0.4kV it is 0.4 meters). Then, in combination with the global scale of the engineering drawings (such as 1:500), convert the standard values ​​of electrical spacing in the physical dimension into the equivalent pixel distance in the pixel coordinate system of the drawings (i.e., the drawing coordinate system) as the safety distance threshold. After this conversion, collision and clearance detection can be directly performed using the geometric distance of the drawing, thereby quantitatively identifying electrical discharge safety hazards caused by the excessively close laying of lines.

[0053] This step establishes gridded construction environment constraints by acquiring geographic information data, and constructs node connectivity capacity constraints and safety distance thresholds by analyzing the ledger attributes in the semantic information matrix and electrical specifications. It establishes three-dimensional boundaries for limiting the generation of topological connectivity edges, physical capacity boundaries for limiting the degree of topological nodes, and safety clearance boundaries for limiting the spatial location of primitives, providing a quantitative basis for identifying map topological connectivity anomalies and spatial compliance defects.

[0054] Step 3: Construct an engineering topology based on discrete geometric primitives in the engineering feature information, and perform constraint verification on each geometric primitive in the engineering topology according to construction environment constraints, node connectivity capacity constraints and safety distance thresholds, and generate conflict sets and conflict labels.

[0055] 3.1 Perform constraint verification on each geometric element in the engineering topology, including: 3.1.1 Establish the mapping relationship between each geometric element and the equipment entity in the engineering feature information, as follows: For discrete geometric primitives without established topological connections, a geometric snapping tolerance threshold (e.g., 5 pixels) is set for the map coordinates. The endpoint coordinates of polyline primitives representing wires and cables, and the bounding box boundaries of block primitives representing electrical equipment or poles are extracted. The shortest Euclidean distance from the polyline endpoints to the bounding box boundary of the equipment is calculated. Primitives whose spatial locations directly intersect or whose shortest Euclidean distance is within the tolerance threshold are fused into topological nodes to establish data associations between points and lines. After completing the above data association, an engineering topology is constructed using an adjacency list. Specifically, blocks, circles, and other primitives corresponding to electrical equipment or poles are instantiated as nodes in the graph, and polyline primitives corresponding to wires and cables are instantiated as edges in the graph, completing the structured transformation from discrete geometric primitives to a topologically connected network model.

[0056] In the above structured transformation process, when the large model performs cross-modal joint analysis in step 1, each discrete geometric primitive is bound to a unique category label based on the bounding box predicted by the cross-modal model. The equipment ledger node (i.e. the equipment entity in the engineering feature information) pointed to by the category label is written into the attribute field of the corresponding topology node, thereby establishing a mapping relationship between each geometric primitive and the equipment entity in the engineering feature information. Based on this mapping relationship, each node and connecting edge in the topology can be back-indexed for its corresponding equipment entity name, specifications, and ledger details.

[0057] 3.1.2 In the path analysis of the engineering topology, each geometric element is mapped to a spatial cost matrix, and the cumulative construction cost of the path corresponding to each geometric element is calculated, as follows: Centered on the coordinates of the extracted crosshair control points of the engineering drawing (i.e., the global control point coordinates), a local image region of a preset size (e.g., 256×256 pixels) is extracted, and the ORB algorithm is used to perform feature matching between the engineering drawing and the GIS map within this region; the specific steps are as follows: (1) In order to improve the stability of cross-modal image feature matching, a matrix structure element of a preset size (such as 3×3 pixels) is used to perform morphological dilation processing (i.e. local maximum filtering) on ​​the local image region to make the fine line segment features thicker and wider. (2) Traverse each pixel in the local image area after processing, and compare the gray value difference between the current pixel and the pixels on the surrounding ring with a preset radius (e.g., 3 pixels, the circumference contains 16 pixels) with the current pixel as the center. If there are a preset number (e.g., 9) or more consecutive pixels on the ring, and their gray values ​​are all greater than the sum of the gray value of the center and the preset deviation threshold (e.g., 20), or all are less than the difference between the gray value of the center and the preset deviation threshold, then the current pixel is determined to be a feature point (i.e., a corner point). (3) Randomly select multiple pairs (e.g., 256 pairs) of pixel points within the local neighborhood of the feature point (e.g., a 31×31 pixel image window) in the preset range of the feature point. Record the two pixels in each pair as the first pixel point and the second pixel point respectively according to the sampling order. Compare the gray values ​​of the two. If the first pixel point is greater than the second pixel point, encode it as 1; otherwise, encode it as 0. Combine them to form a fixed-length binary feature descriptor. Calculate the Hamming distance between the feature point descriptor in the engineering drawing and each feature point descriptor in the GIS map, that is, the number of bits that are different at corresponding positions in the two binary feature descriptors. Select the feature point pair with the smallest Hamming distance as the initial candidate point pair.

[0058] The RANSAC algorithm is used to remove mismatched point pairs from the initial candidate point pairs, resulting in a set of identical point pairs. Specifically, the maximum number of iterations for the RANSAC algorithm is set (e.g., 1000 times), and iterative loops are performed. In each iteration, three pairs of non-collinear candidate point pairs are randomly selected, and a system of linear equations for a two-dimensional affine transformation is constructed, in the form of... and ,in and These are the planar coordinates of feature points on the drawing and feature points in the GIS, respectively. , , , , , For the unknown coefficients to be solved, the coordinate values ​​of three randomly selected pairs of geometric primitive point pairs are substituted into the system of equations to construct a full-rank system of equations consisting of six linear equations and solved to obtain a unique temporary unknown coefficient. This temporary unknown coefficient is then substituted into the system of linear equations and used as the temporary transformation model for the current iteration. The remaining candidate point pairs are substituted into this temporary transformation model to calculate the predicted coordinates, and point pairs whose deviation between the actual coordinates and the predicted coordinates is less than a preset pixel distance threshold (e.g., 3 pixels) are retained as interior points. After satisfying the maximum number of iterations, the set of interior points corresponding to the temporary transformation model with the most interior points is selected as the set of point pairs with the same name.

[0059] To ensure the robustness of coordinate system mapping under extreme modal differences, a threshold for determining the number of interior points is set (e.g., 5 pairs). If the number of interior points in the final output of the above RANSAC algorithm is lower than this threshold, the ORB feature matching is deemed to have failed. At this time, an alternative mechanism is triggered: the system simultaneously presents a partial engineering drawing and a partial GIS map on the graphical interface, and receives the user's click input operation, forcibly extracting at least 3 pairs of user-specified target homonymous point pairs as homonymous point pair sets.

[0060] Based on this, the least squares method is adopted, taking the non-collinear two-dimensional plane pixel coordinates of the drawing and their corresponding GIS plane coordinates of all the corresponding point pairs in the set as input, and reconstructing an overdetermined two-dimensional affine transformation linear equation system containing 6 unknown parameters. The solution is performed with the goal of minimizing the sum of squared errors between the transformed coordinates and the real GIS plane coordinates, and the globally optimal translation, rotation angle, scaling coefficients of each coordinate axis and shear parameters are output to obtain the final affine transformation matrix.

[0061] In the path analysis of the engineering topology, the discrete vertex coordinates of each geometric primitive (i.e., connecting edge) in the CAD world coordinate system are extracted, multiplied by the scale matrix of the preset viewport mentioned in step 1, and converted into two-dimensional pixel coordinates in the drawing pixel coordinate system and constructed as a homogeneous coordinate vector. The homogeneous coordinate vector is multiplied by the affine transformation matrix to obtain the corresponding GIS spatial coordinates, thereby mapping the geometric primitives to the aligned GIS geographic raster coordinate system. Since the geometric trajectory of the primitives in the engineering drawing (i.e., the path corresponding to the geometric primitives) is fixed, the geometric segment clipping algorithm is directly used to discretize the mapped geometric trajectory of the primitives. Specifically, the Cohen-Sutherland segment clipping algorithm is used, taking the geometric trajectory of the geometric primitive in the GIS space as input, and the boundary of each unit grid of the GIS geographic raster as the clipping window, to calculate the actual intercepted line segment of the geometric primitive passing through each grid, and then calculate the physical length of the actual intercepted line segment in each grid based on the coordinates of the two endpoints. Extract the construction cost of all GIS grid nodes covered by the geometric trajectory, multiply the construction cost of each grid by the physical length of the actual intercepted line segment within that grid, and sum the multiplications of all the grids along the route to obtain the cumulative construction cost corresponding to the geometric primitive.

[0062] like Figure 3As shown in the figure, the spatial cost matrix and the discretization of geometric primitives are illustrated. The gradient area in the background represents the construction cost value of each spatial location in the GIS spatial cost matrix. The red broken line trajectory represents the geometric trajectory of the primitive after affine transformation and mapping to GIS space. The yellow cross marks indicate the intersection points of the actual intercepted line segments generated when the geometric trajectory of the primitive passes through the boundaries of each GIS grid using the Cohen-Sutherland line segment clipping algorithm. This figure intuitively demonstrates how to discretize the geometric trajectory of CAD primitives, calculate the actual intercepted line segments passing through each grid and their physical lengths, and accumulate them with weights to achieve a quantitative evaluation of the cumulative construction cost value of geometric primitives.

[0063] It should be noted that when assessing the construction cost and physical feasibility of power distribution network project routes, traditional route cost assessments are usually based on two-dimensional coordinates on drawings or logical topologies detached from the actual geographical environment. They approximate the cost using only two-dimensional geometric lengths or empirical weights, failing to incorporate the three-dimensional terrain undulations and surface cover differences at the project site. This leads to a significant discrepancy between the assessment results and the actual construction environment. Therefore, this step maps the project topology to the actual GIS geospace and performs discretized cost calculations based on the spatial cost matrix: first, the affine transformation matrix is ​​solved to align the geometric primitives on the drawings to the GIS raster coordinate system; then, the actual physical lengths of the intercepted line segments of each geometric primitive within different terrain and elevation grids are extracted, and multiplied and accumulated with the construction cost value of the corresponding grid to obtain the cumulative construction cost value. This method establishes the route cost assessment on a real three-dimensional geographical environment, achieving accurate quantification of the impact of complex terrain on the configuration of engineering materials and actual construction costs.

[0064] 3.1.3 Obtain the current number of connections of the associated nodes of geometric primitives, the number of nodes mapped to the current device model, and the surface spacing between adjacent geometric primitives, as follows: During graph traversal, the current connection count of the nodes associated with the current geometric element is obtained in real time by querying the real-time in-degree statistics of the nodes associated with the current geometric element in the underlying adjacency table. The total number of mapped nodes is obtained by querying the device ledger status hash table maintained in memory beforehand, which counts the total number of nodes that have successfully joined the topology network and whose model labels are completely consistent with the current node. This hash table uses the device model label (e.g., "one-in-three-out ring network cabinet") in the project semantic information as the key and the number of nodes of the corresponding model that have joined the network as the value. During the topology construction initialization phase, the values ​​corresponding to each key are set to 0 by default. When constructing the project topology, each time a geometric element is instantiated as a network node and added to the topology network, the value of the node's model label in the hash table is incremented by 1.

[0065] For adjacent geometric primitives that have not yet established topological connections, axis-aligned bounding boxes (AABBs) are constructed based on their original two-dimensional pixel coordinate data in the drawing pixel coordinate system. A minimum bounding box distance algorithm is used, taking the center point coordinates and the length and width boundaries of the two bounding boxes as input, to calculate the minimum projected distance between the two bounding boxes in the X and Y axes. If the projected distances in both directions are greater than zero, the square root of the sum of the squares of the two projected distances is taken as the drawing distance between adjacent geometric primitives. If the projected distance in either direction is less than or equal to zero, it indicates that the two bounding boxes intersect spatially. In this case, the drawing primitives are further discretized into multiple basic line segments. All line segment combinations are traversed, and the perpendicular distance from the endpoint of one line segment to the line containing another line segment is calculated. If the foot of the perpendicular falls within the line segment, the perpendicular distance is taken; otherwise, the Euclidean distance from the endpoint of one line segment to the endpoint of another line segment is taken. The minimum value among all line segment combinations is taken as the drawing distance between adjacent geometric primitives.

[0066] 3.1.4 Compare and verify the above indicators with their corresponding thresholds, and calculate the deviation data of the geometric primitives that failed the verification for the corresponding verification item, as follows: The cumulative construction cost, current number of connections, number of mapped nodes, and map spacing are compared and verified against the path cost threshold, node degree upper limit, number of similar nodes upper limit, and safety distance threshold, respectively. During the verification process, if any verification item violates the corresponding constraint boundary, the geometric element is deemed to have failed the verification. The specific criteria for failing the verification are: the cumulative construction cost is greater than the path cost threshold, the current number of connections is greater than the node degree upper limit, the number of mapped nodes is greater than the number of similar nodes upper limit, or the map spacing is less than the safety distance threshold.

[0067] For any geometric primitive that violates the corresponding constraint boundary check item, calculate its deviation data. Specifically, based on the violation item, calculate the difference between its current number of connections and the upper limit of node degree, and the difference between the number of mapped nodes and the upper limit of the number of nodes of the same type, which together serve as connectivity deviation data. Calculate the difference between the safety distance threshold and the actual map spacing as spacing deviation data, and calculate the difference between the cumulative construction cost and the path cost threshold as cumulative construction cost deviation data.

[0068] 3.2 Generate a conflict set and corresponding conflict labels, including: 3.2.1 Geometric primitives that fail the validation are designated as conflict primitives, and engineering semantic information with entity mapping relationships with the conflict primitives is extracted to construct a conflict set, as follows: Geometric elements that fail the verification during the comparison are designated as conflicting elements. Engineering semantic information with entity mapping relationships to these conflicting elements is extracted to construct a conflict set. It should be noted that the "entity mapping relationship" has already been established during the aforementioned element attribute binding and topology construction process: when elements conflict in the topology, the name, specifications, and ledger details of the conflicting objects can be extracted in reverse based on this mapping relationship.

[0069] 3.2.2 Based on the type of constraint violated, conflict labels are bound to conflicting primitives. The conflict labels include connectivity violation labels, cost violation labels, and spacing violation labels, as detailed below: While constructing the conflict set, the conflict elements are bound with corresponding conflict labels according to the type of constraint violated; specifically: when the cumulative construction cost fails the verification, a cost exceeding the limit label is bound; when the current number of connections or the number of mapped nodes fails the verification, they are uniformly classified and bound with a connectivity exceeding the limit label; when the drawing spacing fails the verification, a spacing violation label is bound.

[0070] In practical implementation, this step can rely on graph databases (such as Neo4j) for topological storage and execute algorithm comparison based on the Python runtime environment. To support the standardized reading and intelligent inference of subsequent large models, the conflict set is serialized into a JSON array format; the structure of this array format is shown below: [{"conflict_ID":"Edge_089","conflict_Type":"cost exceeded","deviation_Value":"cumulative 14500>threshold 12000","related_Entity":{"entity":"overhead conductor","specification":"JKLYJ-10-240"}}).

[0071] This step constructs an engineering topology based on discrete geometric primitives and verifies the constraints of each geometric primitive according to construction environment constraints, node connectivity capacity constraints, and safety distance thresholds. It establishes a comparison mechanism between the nodes in the drawing graph theory and the actual geographical environment and the upper limit of physical equipment interfaces, effectively identifying design defects such as equipment disconnection, material exceeding quotas, and spatial violations, thereby improving the accuracy of power grid engineering material allocation and the physical feasibility of on-site construction.

[0072] Step 4: Generate multi-dimensional review opinions based on the conflict set and conflict tags.

[0073] 4.1 Based on the conflict labels, extract the corresponding connectivity deviation data, spacing deviation data, and cumulative construction cost deviation data, including: From the conflict set output in step 3, extract the connectivity deviation data, spacing deviation data, and cumulative construction cost deviation data corresponding to each conflict label. Specifically, for nodes with connectivity exceeding the limit label, extract the corresponding connectivity deviation data; for primitives with spacing violation label, extract the corresponding spacing deviation data; and for edges with cost exceeding the limit label, extract the corresponding cumulative construction cost deviation data.

[0074] 4.2 Calculate the semantic similarity between technical parameters and standard specification parameters in the engineering feature information, and the difference between the current operating years and the preset retirement years recorded in the design documents of multimodal engineering data, including: For each device entity in the conflict set, the text encoder of the pre-trained large model described in step 1 is reused. The specific technical parameter text (e.g., "silicone rubber insulation material with high-altitude anti-condensation properties") of the device entity in the engineering semantic information of the engineering feature information and the corresponding standard specification text (e.g., "high-altitude anti-condensation composite insulation requirements") in the standard parameter set are input into the text encoder respectively. The [CLS] flag feature vector in the output sequence is extracted and pooled to obtain the corresponding fixed-dimensional global semantic feature vectors (this dimension is limited by the number of nodes in the feature mapping layer of the underlying network of the aforementioned text encoder, for example, uniformly mapped to 768 dimensions). The cosine similarity of the two global semantic feature vectors is calculated as the semantic similarity, and the calculation formula is as follows: (10) in, Semantic similarity is used to characterize the degree of compliance of technical parameters. This refers to the global semantic feature vector corresponding to the technical parameter text in the engineering semantic information. This is the global semantic feature vector corresponding to the standard specification parameters in the standard parameter set.

[0075] From the equipment ledger details in the multimodal engineering data design documents, obtain the equipment commissioning date of each equipment entity, and calculate the current operating years of the equipment based on the current system time; according to the equipment nameplate model, obtain the preset retirement age from the asset life cycle specification dictionary of the standard parameter set, for example, the preset retirement age of a 10kV oil-immersed transformer is 25 years; calculate the difference between the current operating years and the preset retirement age to obtain the age difference of the equipment entity.

[0076] 4.3 Calculate the spatial coverage of the engineering topology after topological connectivity and the spatial overlap between the area to be modified as recorded in the design documents, including: Based on the absolute GIS coordinates of each node and connecting edge in the engineering topology, a GIS buffer analysis algorithm is used. Using the geographic coordinate set of each node and connecting trajectory as the reference geometric primitive, a preset safe working radius (e.g., 5 meters) is set and translated outwards along the primitive normal direction. For line segment endpoints or corners, rounded corners are used for smooth connection, thus generating a buffer region containing a combination of rounded corners and polygons. Finally, a polygon Boolean union operation is used to combine the buffer regions of all nodes and edges in the engineering topology, resulting in a complete spatial coverage polygon. From the survey boundary coordinate sequence provided in the design documents, the coordinates of each boundary inflection point are connected sequentially to generate the target polygon of the area to be modified. A GIS spatial analysis algorithm is used to calculate the ratio of the overlapping area of ​​the two polygons to the area of ​​the area to be modified, which is used as the spatial overlap degree. The calculation formula is as follows: (11) in, Spatial overlap, The spatial coverage polygon of the engineering topology. The target polygon for the area to be modified in the design file; To solve for the function of the geometric area of ​​the corresponding polygon, and ,in These are the sequential coordinates of the polygon's vertices in the GIS projected coordinate system (in meters). The total number of vertices of the polygon, and the coordinates of the closed vertices are set. .

[0077] 4.4 Input the various deviation data, semantic similarity, year difference, and spatial overlap into the pre-trained large model to generate multi-dimensional review opinions, including: The various deviation data, semantic similarity, age difference, and spatial overlap are used as actual quantified values ​​and directly filled into the placeholders corresponding to the prompt word template in step 1.1.3 to form a complete instruction text sequence. This instruction text sequence is then input into the autoregressive decoding network of the pre-trained large model. Since this large model has already undergone supervised training in step 1 using a massive historical conflict set, conflict labels, and historical multi-dimensional review opinions, its network weights fit the mapping distribution from engineering deviation features to review text and internalize the multi-step logical deduction thought chain (CoT) capability. Therefore, the large model uses the autoregressive mechanism to predict probabilities word by word, and based on the structured deviation features in the prompt words, uses its inherent thought chain to guide intermediate logical judgments (for example, first deriving necessity and descriptive material standardization from age difference and semantic similarity, then evaluating safety and cost rationality from spacing deviation and cost deviation), finally outputting structured multi-dimensional review opinions, including completeness review opinions, safety review opinions, cost rationality review opinions, standardization review opinions, necessity review opinions, and accuracy review opinions.

[0078] It should be noted that traditional compliance verification of power distribution network engineering drawings typically relies solely on rigid rules at the micro-level (such as distance or angle comparisons based on drawing coordinates). When a set threshold is exceeded, a binary local violation alarm is directly output, lacking a holistic understanding of the overall rationality of the project (such as the coverage of the upgrade, the necessity of decommissioning, etc.). Furthermore, the output conclusions are rigid and difficult to form a comprehensive evaluation with engineering guidance value. To address this, this step introduces GIS buffer spatial coverage analysis and integrates the multi-step logical deduction capability of the CoT (Cooperation of Thought) internalized by the pre-trained large model: First, a GIS buffer area is generated by translating along the normal direction based on the absolute coordinates of the topology nodes. The spatial overlap between the engineering topology and the area to be upgraded is obtained using polygon Boolean union operations. Then, the spatial overlap, semantic similarity, age difference, and various geometric deviation data are uniformly encapsulated into a structured instruction sequence and input into the large model. The model then deduces the macro-level engineering rationality layer by layer from the micro-physical parameters, ultimately generating a multi-dimensional comprehensive review opinion covering integrity, safety, cost rationality, standardization, necessity, and accuracy. This achieves an improvement from single data alarms to comprehensive engineering review conclusions.

[0079] This step encapsulates conflict bias, semantic similarity, and spatial overlap into a structured prompt word input model. Relying on the contextual understanding and logical reasoning capabilities of the pre-trained model, it transforms the discrete underlying spatial geometry and numerical constraints into global-dimensional engineering language guidance. This improves upon the shortcomings of traditional verification methods that are limited to isolated single rules, and provides a systematic basis for the comprehensive compliance verification of power distribution network engineering design schemes.

[0080] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0081] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0082] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0083] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0084] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0085] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent verification of materials in power distribution network engineering based on large-scale models, characterized in that, Includes the following steps: A pre-trained large model is used to analyze the multimodal engineering data of the power distribution network project to be verified, and the engineering feature information is extracted. Construction environment constraints are generated based on the topographic features of the engineering site, and node connectivity capacity constraints and safety distance thresholds are constructed based on engineering feature information. An engineering topology is constructed based on geometric primitives in the engineering feature information. Constraints are verified on each geometric primitive in the engineering topology according to construction environment constraints, node connectivity capacity constraints and safety distance thresholds, and conflict sets and conflict labels are generated. Based on the aforementioned conflict set and conflict tags, multi-dimensional review opinions are generated.

2. The method according to claim 1, characterized in that, The multimodal engineering data includes design documents, engineering drawings, and corresponding standard equipment and material parameters; the engineering feature information includes engineering semantic information, discrete geometric primitives, and a set of standard parameters.

3. The method according to claim 2, characterized in that, The generated construction environment constraints include: Obtain elevation data and land cover data of the project site, extract the slope and land cover type of each spatial location, and construct a spatial cost matrix; Based on the material mechanical property parameters corresponding to the engineering semantic information, the maximum allowable span is calculated, and the maximum allowable span is converted into a path cost threshold. The spatial cost matrix and the path cost threshold together constitute the construction environment constraints.

4. The method according to claim 3, characterized in that, The construction of the spatial cost matrix includes: Based on the preset engineering construction quota, obtain the basic cost value corresponding to the surface cover type in each region; Using the slope as a weighting coefficient, the construction cost of each spatial location is multiplied by the basic cost to obtain the spatial cost matrix.

5. The method according to claim 3, characterized in that, The node connectivity capacity constraints and safe distance thresholds include: The number of various types of equipment recorded in the equipment ledger in the engineering semantic information is used as the upper limit of the number of nodes of the same type in the engineering topology. Based on the specifications and models of the equipment in the engineering semantic information, the number of interfaces of the corresponding equipment is extracted from the standard parameter set as the upper limit of node degree. The upper limit of quantity and the upper limit of node degree constitute the node connectivity capacity constraint. The electrical spacing standard values ​​in the standard parameter set are scaled to the coordinate system of the discrete geometric elements according to the scale of the engineering drawings to obtain the safety distance threshold.

6. The method according to claim 5, characterized in that, The constraint verification of each geometric element in the engineering topology includes: Establish the mapping relationship between each geometric element and the equipment entity in the engineering feature information; In the path analysis of the engineering topology, each of the geometric primitives is mapped to the spatial cost matrix, and the cumulative construction cost of the path corresponding to each geometric primitive is calculated. Obtain the current number of connections of the associated nodes of the geometric primitives, the number of nodes mapped to the current device model, and the surface spacing between adjacent geometric primitives; The cumulative construction cost, current number of connections, number of mapped nodes, and map spacing are compared and verified with the path cost threshold, node degree upper limit, number of similar nodes upper limit, and safety distance threshold, respectively. The deviation data of geometric elements that fail the verification in the corresponding verification item are calculated.

7. The method according to claim 6, characterized in that, The generation of the conflict set and corresponding conflict labels includes: Geometric primitives that fail the verification are designated as conflict primitives, and engineering semantic information that has an entity mapping relationship with the conflict primitives is extracted to construct a conflict set. Based on the type of constraint violated, a conflict label is bound to the conflicting primitive. The conflict labels include connectivity violation label, cost violation label, and spacing violation label.

8. The method according to claim 7, characterized in that, The generation of multi-dimensional review opinions based on the conflict set and conflict tags includes: Based on the conflict labels, extract the corresponding connectivity deviation data, spacing deviation data, and cumulative construction cost deviation data; The semantic similarity between technical parameters and standard specification parameters in the semantic information of the engineering project is calculated, as well as the difference between the current service life and the preset retirement life recorded in the design documents of the multimodal engineering data. Calculate the spatial coverage of the engineering topology after topological connectivity and the spatial overlap between the area to be modified as recorded in the design documents; The deviation data, semantic similarity, age difference, and spatial overlap are input into the pre-trained large model to generate the multi-dimensional review opinion.

9. The method according to claim 2, characterized in that, The extracted engineering semantic information, discrete geometric primitives, and standard parameter set include: Natural language parsing is performed on design documents and standard equipment and material parameters in multimodal engineering data to obtain engineering semantic information and a set of standard parameters; Vectorization is performed on engineering drawings in multimodal engineering data to obtain discrete geometric primitives.

10. The method according to claim 1, characterized in that, The training process of the pre-trained large model includes: Obtain historical multimodal engineering data in the field of power distribution network engineering, and combine it with the corresponding structured annotation results to train the basic large model for cross-modal understanding; Using a set of historical conflicts and historical conflict labels as input, and multi-dimensional historical review opinions as the target output, supervised training is performed on the basic large model trained on cross-modal understanding to obtain the pre-trained large model.

Images