Artificial intelligence power grid building imaging method and engine based on large language model

Abstract

The invention provides an artificial intelligence power grid building imaging method and engine based on a large language model, and relates to the technical field of power grid building imaging, and the method comprises the steps: reading project information and a power grid building design standard through a large language model, so as to obtain a building demand; detecting and correcting an abnormal value in the data by utilizing a large language model; and step-by-step reasoning is carried out by adopting a three-dimensional reconstruction model according to the two-dimensional CAD image data after data preprocessing, the three-dimensional shape of the power grid building object is recovered, building elements are identified based on a large language model, and whether the building elements conform to the current scene and construction conditions or not is judged. The artificial intelligence power grid building imaging engine comprises a three-dimensional image reconstruction module and a large language model processing module. The method can achieve the precise modeling of a power grid building from two dimensions to three dimensions, improves the data quality and modeling efficiency, and optimizes the building layout.

Description

Technical Field

[0001] The present invention relates to the technical field of power grid building imaging, and in particular to an artificial intelligence power grid building imaging method and engine based on a large language model. Background Art

[0002] In the field of traditional power construction, there are many drawbacks in the drawing and application of building structure drawings. For a long time, the two-dimensional drawing software AutoCAD has dominated the process of drawing building structure drawings, and its drawing results are mainly stored in formats such as DWG and DXF. These formats are significantly different from the formats used by three-dimensional modeling software such as Revit, SketchUp, 3dsMax, etc., resulting in obstacles to data sharing and interaction between the two, and a gap between two-dimensional design and three-dimensional space.

[0003] This gap has brought many problems in practical applications. On the one hand, during the design phase, designers design based on two-dimensional drawings, which may not fully consider the actual situation of the building in three-dimensional space, resulting in defects in the design plan in three-dimensional space. For example, some parts of the building may collide or conflict in three-dimensional space, affecting the structural safety and construction feasibility of the building. On the other hand, during the construction phase, construction workers mainly refer to two-dimensional drawings for construction. Due to the lack of intuitive guidance from three-dimensional models, construction workers may find it difficult to accurately understand the design intent, which may easily lead to construction errors or substandard construction quality.

[0004] In addition, the separation of two-dimensional design and three-dimensional space also brings many hidden dangers to the safety assessment of power buildings. For example, in terms of dual power supply settings, design and construction based only on two-dimensional drawings may not accurately determine whether the dual power supply settings are reasonable and whether they comply with relevant standards and specifications, thus affecting the reliability and safety of power supply. In terms of the construction space of emergency power supply equipment, due to the lack of intuitive display of three-dimensional models, it may not be possible to accurately reserve enough space, resulting in the inability to install or operate emergency power supply equipment normally, affecting the emergency power supply capacity of the building. In terms of building space layout, unreasonable layout may lead to the inability to ensure the safety of grounding equipment, such as the distance between the grounding equipment and other parts of the building does not meet safety requirements, increasing the risk of electric shock or short circuit. At the same time, problems such as repeated grounding and equipotential bonding may also occur, affecting the overall safety and stability of power buildings.

[0005] In summary, the separation of two-dimensional design and three-dimensional space in the traditional power construction field not only affects the efficiency and quality of design and construction, but also brings many hidden dangers to the safety assessment of power buildings. There is an urgent need for a method that can effectively integrate two-dimensional design and three-dimensional modeling to solve these problems. Summary of the invention

[0006] The present invention aims to solve at least one of the above-mentioned technical problems existing in the prior art.

[0007] To this end, the first aspect of the present invention provides an artificial intelligence power grid building imaging method based on a large language model.

[0008] A second aspect of the present invention provides an artificial intelligence power grid building imaging engine based on a large language model.

[0009] The present invention provides an artificial intelligence power grid building imaging method based on a large language model, comprising: Acquire basic data, the basic data including at least two-dimensional CAD image data and project information for describing a power grid building; use a large language model to read the project information and power grid building design standards to obtain building requirements; Performing data preprocessing on the two-dimensional CAD image data; the data preprocessing includes format conversion, data cleaning and data enhancement; wherein, in data cleaning, a large language model is used to detect and correct outliers in the data; A three-dimensional reconstruction model is constructed, wherein the three-dimensional reconstruction model gradually infers and restores the three-dimensional shape of the power grid building object according to the two-dimensional CAD image data after data preprocessing; wherein the power grid building site information and the two-dimensional feature information of various building units in the site are obtained by feature extraction, wherein the building unit is a power grid functional unit composed of at least one power equipment; the extracted two-dimensional feature information is classified and labeled by semantic segmentation, and the building elements are identified based on the large language model to determine whether the building elements meet the current scene and construction conditions; According to the two-dimensional feature information, a matching basic building unit model is selected in the prefabricated model database; the basic building unit model is placed in the three-dimensional coordinates of the power grid construction site according to the coordinate information and the layout information of the plan; and based on the six-sided view rendering engine, a mapping relationship is established between the two-dimensional image and its corresponding basic building unit model to perform spatial three-dimensional imaging and modeling tasks.

[0010] The artificial intelligence power grid building imaging method based on a large language model according to the above technical solution of the present invention may also have the following additional technical features: In the above technical solution, the two-dimensional CAD image data records the power grid building site information and the image information of various building units in the site, and the image information includes the two-dimensional coordinates, geometric shape and surface features of the power grid building; The project information includes at least one of a project name, a project type, a special situation description file, a risk description file, a CAD file description list, a building review file list, and voxel points; The special situation description file records the special information of the designated building, and the special information includes the location, size and angle of the designated building; The risk description document is used to restrict certain buildings that are prohibited from construction; The CAD file description list is used to describe the function of each two-dimensional CAD image data; The building audit document list is used to supplement and verify building parameters.

[0011] In the above technical solution, the large language model adopts the large language model specialized in CPM-Bee architecture; the method of using the large language model to detect and correct outliers in the data includes: The CPM-Bee building-specific large language model identifies the height information, distance information and angle information in the two-dimensional CAD image data that deviate from the required range based on its own knowledge base and the building review document list, and marks and corrects the deviated data to limit it to the required range; For data information of designated buildings that do not meet the requirements of the CPM-Bee building-specific large language model knowledge base and building audit document list, but are recorded in the special situation description file, the data will be checked and corrected based on the requirement range specified in the special situation description file.

[0012] In the above technical solution, the two-dimensional characteristic information of the power grid building site information and various types of power equipment in the site includes plane building location characteristic information, plane building unit characteristic information and building unit connection relationship characteristic information; The two-dimensional feature information of the power grid construction site and various building units in the site are obtained by feature extraction, including: A plurality of two-dimensional convolutional layers are used to perform convolution processing on the preprocessed two-dimensional CAD image data; each two-dimensional convolutional layer is equipped with a convolution kernel of a set size and step length, and these convolution kernels slide in a two-dimensional space to extract different features in the two-dimensional CAD image data; a nonlinear activation function is added after each two-dimensional convolutional layer to introduce a nonlinear transformation, and a feature map to be processed is output; Perform a maximum pooling operation on the feature map to be processed, select the maximum eigenvalue in each pooling window as the output, and generate a local feature vector; The local feature vectors are subjected to feature fusion and transformation through a fully connected layer, the connection relationship and layout rules between different building units are learned, and the local feature vectors are combined into a global feature vector.

[0013] In the above technical solution, the classification and labeling of the two-dimensional feature information by semantic segmentation includes: In the decoder, a deconvolution layer is used to perform an upsampling operation on the global feature vector, so as to expand the low-resolution feature map into a higher-resolution feature map, and restore the original spatial resolution of the data with the decoded feature map; For each pixel or voxel in the decoded feature map, the probability distribution of each building unit type is calculated through the Softmax function to obtain the most likely type label of each pixel or voxel, including: For the global feature vector output by the fully connected layer without activation function , n is the total number of categories, The sample belongs to The score of the class, the Softmax function calculation process is:

[0014] represents the conditional probability symbol, y is the true category label, It refers to the probability that the output category is t under the condition of a given input feature x; r represents the category index (traversing all categories), ∈{1,2,…,n}; Indicates the inactive feature vector The nth element of , this score is calculated by the last layer of the neural network; the total number of categories n, that is, the total number of elements in the input vector, uses the number of building unit types defined by the power industry standard, which determines the number of probabilities that the Softmax function needs to calculate; According to the classification results, a corresponding segmentation mask is generated for each category, and the distribution of building units in three-dimensional space is displayed through the segmentation mask; In the three-dimensional reconstruction model, the plane building position feature information, the plane building unit feature information and the building unit connection relationship feature information are classified respectively.

[0015] In the above technical solution, the classifying and labeling the two-dimensional feature information by semantic segmentation further includes: A skip connection is introduced in the decoder to fuse the feature maps of the corresponding levels in the encoder with the upsampled feature maps in the decoder.

[0016] In the above technical solution, the identification of architectural elements based on the large language model to determine whether the architectural elements meet the current scene and construction conditions includes: The large language model uses the plane building location feature information to determine whether the building elements conform to the current scene according to the building requirements read based on the project information, that is, to determine whether the location of each building unit is reasonable; the building elements that do not conform to the current scene are marked and recorded, and the error information can be directly output for reference by designers; The large language model determines whether the current building unit meets the construction needs of the corresponding building unit based on the building needs read based on the project information based on the feature information of the planar building unit; After completing the determination of the current scene and construction conditions, the large language model supplements the building unit with detailed drawing information based on the knowledge base and the precautions entered in the project information.

[0017] In the above technical solution, the matching basic building unit model is selected from the prefabricated model database according to the two-dimensional feature information, and the basic building unit model is placed in the three-dimensional coordinates of the power grid building site according to the coordinate information and the layout information of the plan; and based on the six-sided view rendering engine, a mapping relationship is established between the two-dimensional image and its corresponding three-dimensional voxel point model, and the spatial three-dimensional imaging and modeling tasks are performed, including: Selecting a matching basic building unit model in a prefabricated model database according to the extracted planar building unit feature information; wherein the prefabricated model database includes three-dimensional voxel point models of power grid functional units of different types of power stations, and defining any three-dimensional voxel point model as a basic building unit model; The basic building unit model is placed in the three-dimensional coordinates of the power grid construction site according to the coordinate information and the layout information of the plan; After receiving the processed input image sequence, the six-view rendering engine first encodes these image sequences into a series of hidden states. These hidden states capture the temporal dependencies in the image sequence and contain spatial features; these hidden states are then transmitted to the decoder part. When processing these hidden states, the decoder first applies 3D convolution operations to extract and fuse spatial-temporal features. Next, nonlinear characteristics are introduced through nonlinear activation functions (sigmoid), allowing the model to learn more complex mapping relationships. The resolution of the hidden states is gradually increased until the target output resolution is reached.

[0018] The above technical solution also includes: The generated segmentation mask is smoothed to remove possible isolated noise points and small misclassified areas. The holes that may exist in the segmentation mask are repaired using a hole filling algorithm based on neighborhood information. The final semantic segmentation results are presented to power designers in an intuitive 3D visualization form; the segmentation mask is superimposed on the original 3D imaging data to identify different power equipment categories with different colors or textures.

[0019] The present invention provides an artificial intelligence power grid building imaging engine based on a large language model, which is applied to the artificial intelligence power grid building imaging method based on a large language model as described in any of the above technical solutions, and the artificial intelligence power grid building imaging engine includes: a three-dimensional image reconstruction module and a large language model processing module; the three-dimensional image reconstruction module receives two-dimensional CAD image data for describing power grid buildings; the large language model processing module is used to receive project information, and use the large language model to read the project information and power grid building design standards to obtain building requirements; The three-dimensional image reconstruction module comprises: A data preprocessing unit performs data preprocessing on the two-dimensional CAD image data; the data preprocessing includes format conversion, data cleaning and data enhancement; wherein, in data cleaning, a large language model is called to detect and correct abnormal values ​​in the data; A feature extraction module is used to obtain the two-dimensional feature information of the power grid building site and various types of power equipment in the site through feature extraction; A semantic segmentation module is used to classify and label the two-dimensional feature information through semantic segmentation; the large language model processing module uses a large language model to identify building elements according to the processing results of the feature extraction module and the semantic segmentation module, and determines whether the building elements meet the current scene and construction conditions; The three-dimensional imaging module selects a matching basic building unit model in the prefabricated model database according to the two-dimensional feature information; places the basic building unit model into the three-dimensional coordinates of the power grid building site according to the coordinate information and the layout information of the plan; and establishes a mapping relationship between the two-dimensional image and its corresponding basic building unit model based on the six-sided view rendering engine to perform spatial three-dimensional imaging and modeling tasks.

[0020] In summary, due to the adoption of the above technical features, the beneficial effects of the present invention are: The present invention integrates two-dimensional CAD data and large language models to achieve accurate modeling of power grid buildings from two-dimensional to three-dimensional, improve data quality and modeling efficiency, optimize building layout, enhance safety and compliance, break down data barriers between software, reduce manual intervention, and promote the intelligent development of power construction.

[0021] Additional aspects and advantages of the present invention will become apparent from the following description or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which: Figure 1 is a flow chart of an artificial intelligence power grid building imaging method based on a large language model according to an embodiment of the present invention; Figure 2 It is a schematic diagram of a process in which a 2D convolutional neural network layer converts an input image into a feature map to be processed in one embodiment of the present invention; Figure 3 This is a schematic diagram of a 3D-LSTM network structure in one embodiment of the present invention; Figure 4 This is a schematic diagram of a preliminary 3D-LSTM rendering effect in one embodiment of the present invention; Figure 5 This is a schematic diagram of the GRU unit assisted drawing effect in one embodiment of the present invention; Figure 6 is a schematic diagram of the effect after mean square error correction in one embodiment of the present invention; Figure 7 It is a schematic diagram of the effect of smoothing and hole filling in one embodiment of the present invention. DETAILED DESCRIPTION

[0023] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0024] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0025] Refer to the following Figures 1 to 7 To describe the artificial intelligence power grid building imaging method and engine based on a large language model provided according to some embodiments of the present invention.

[0026] Some embodiments of the present application provide an artificial intelligence power grid building imaging method based on a large language model.

[0027] The first embodiment of the present invention proposes an artificial intelligence power grid building imaging method based on a large language model, including the following steps S1-S4.

[0028] S1. Acquire basic data, wherein the basic data at least includes two-dimensional CAD image data and project information for describing power grid buildings; use a large language model to read the project information and power grid building design standards to obtain building requirements.

[0029] Specifically, the two-dimensional CAD image data records the information of the power grid building site and the image information of various building units in the site, and the image information includes the two-dimensional coordinates, geometric shapes and surface features of the power grid building. In some embodiments, the two-dimensional CAD image data can be two-dimensional imaging data from different channels, such as CAD files formed by laser radar (LiDAR) scanning data, aerial photogrammetry data, etc., which contain rich spatial information of power facilities, such as the two-dimensional coordinates, geometric shapes and surface features of transmission lines, substations, main plant buildings, dust removal devices, induced draft sites and chimney flues.

[0030] The project information includes at least one of the project name, project type, special situation description file, risk description file, CAD file description list, building review file list and voxel points; the richer the content of the project information, the more beneficial it is for the large language model to obtain clear building requirements, and then analyze and judge the rationality and safety of the project. In the present disclosure, it is taken as an example that the project information includes all the above information. Among them, the special situation description file, risk description file, CAD file description list, and building review file list can all be input into the large language model using .csv format files.

[0031] The project name is used to generally describe the current project and supports multiple languages, such as Chinese and English. The project type is used to distinguish the category of the current project to facilitate subsequent model selection and other work. In a specific embodiment, the project types include: thermal power projects, new energy projects, nuclear power projects, power grid projects, substation projects, power transmission projects, system planning, design and scheduling projects, communication projects, thermal projects, desulfurization projects, construction projects, environmental projects, survey projects and other projects. The above project types cover the common types of power grid projects.

[0032] The special situation description file records the special information of the specified building, including the location, size and angle of the specified building; the risk description file is used to restrict some buildings that are prohibited from construction; the CAD file description list is used to describe the role of each two-dimensional CAD image data to facilitate the engine to carry out orderly modeling; the building audit file list is used to supplement and verify the building parameters. Voxel points are a cubic unit in three-dimensional space, which are used to represent discrete points in three-dimensional images or volume data.

[0033] The power grid building design standards refer to the design standards and design specifications of each building unit in the current power grid. By reading the power grid building design standards, the large language model can have the corresponding data processing capabilities.

[0034] In a specific embodiment, the large language model analyzes the composition of the two-dimensional CAD image data according to the CAD file description list, clarifies the various elements and attributes contained in the drawing, and uses the CAD file description list to match the names of the currently entered CAD image data to ensure the accuracy of the drawing information. Subsequently, the CAD drawings are sorted according to their types (such as architecture, machinery, electrical, etc.) for subsequent processing.

[0035] In a specific embodiment, the large language model uses the CPM-Bee architecture-specific large language model. It should be noted that CPM-Bee is a completely open source, commercially available, 10 billion parameter Chinese and English base model that uses a Transformer autoregressive architecture and can be used for the semantic library of lightweight power grid construction documents and conventional formula judgment of architectural problems.

[0036] S2. Perform data preprocessing on the two-dimensional CAD image data; the data preprocessing includes format conversion, data cleaning and data enhancement; wherein, in data cleaning, a large language model is used to detect and correct outliers in the data.

[0037] In some embodiments, step S2 converts the two-dimensional CAD image data of various formats into a standard processable format, including unified drawing size and format, scale, drawing lines, fonts, symbols and other requirements, to ensure the consistency and compatibility of the data for subsequent processing and analysis, and deletes some unprocessable data information and records it in the error log. In a specific embodiment, the standard data format standard adopts the content provided by the "National Standard of the People's Republic of China Electrical Engineering CAD Drawing Rules" GB / T18135-2000 file. In addition, all two-dimensional CAD image data can also be converted into the same length unit and the building orientation can be uniformly adjusted in this process.

[0038] In some embodiments, since the imaging process may be affected by environmental factors, there may be noise points in the data, so data cleaning is required.

[0039] Specifically, the data is denoised by filtering algorithm (Gaussian filtering) to improve the quality of the data and reduce the interference of noise on subsequent analysis. For each pixel point in the two-dimensional CAD image data ( ), and its value after Gaussian filtering is , the calculation formula is:

[0040] in, Indicates that at position ( ) at Gaussian filtering; Coordinates representing the original pixel values; represents the Gaussian standard deviation, The smaller the size, the weaker the smoothing effect and the more image details are retained. The value is 0.13, and the filtering strength is moderate; e represents the base of the natural logarithm; represents a double integral over the entire two-dimensional plane, that is, for every possible and The values ​​are summed.

[0041] During the data cleaning process, the CPM-Bee building-specific large language model identifies the height information, distance information and angle information in the two-dimensional CAD image data that deviates from the required range based on its own knowledge base and building audit document list, and marks and corrects the deviated data to limit it to the required range; for data information of designated buildings that do not meet the requirements of the CPM-Bee building-specific large language model knowledge base and building audit document list but are recorded in the special situation description file, the data is detected and corrected based on the required range specified in the special situation description file.

[0042] In some embodiments, data enhancement is performed by rotation, translation, scaling, and cropping.

[0043] Specifically, random rotation and translation operations are performed on the 2D CAD image data to simulate observation perspectives at different angles and positions, increase data diversity, and improve the generalization ability of the model, so that it can better cope with various changes in actual scenarios. The data is scaled and cropped according to the preset scale range, retaining key areas related to power design and removing irrelevant background information, while ensuring the consistency of the data at different scales, which is convenient for model learning and feature extraction.

[0044] To achieve the above data enhancement function, the graphic edge detection calculation formula used in this embodiment is gradient calculation. First, the input image is defined as I, x and y are the horizontal and vertical coordinates of the pixel in the image, respectively, then the value of I(x, y) represents the gray value of the pixel, and the Sobel operator consists of two 3×3 convolution kernels, which are used to calculate the gradient component in the horizontal direction. and the vertical gradient component , the specific calculation formula is as follows:

[0045] The calculation formula used for the convolution operation is:

[0046]

[0047] in, Represents the location index coordinates of the output feature map; In pixels The gradient component in the horizontal direction; Represents the gradient component in the vertical direction at the pixel (x, y); I(x+i, y+j) represents the grayscale value of the corresponding position (x+i, y+j) in the input image.

[0048] After calculating the gradient component, the gradient modulus (or gradient amplitude) is calculated. In this embodiment, the maximum value between the two is taken as the gradient modulus, which is expressed by the following formula:

[0049] After the gradient modulus is calculated, the image edge is judged as follows:

[0050] in, represents the set threshold value of the gradient modulus, in this disclosure The value of is set to 127, and pixels with gradient modulus greater than the threshold are marked as edge pixels: Edge(x,y) is the output binary image, where pixels with a value of 1 represent edges and pixels with a value of 0 represent non-edges.

[0051] S3. Construct a three-dimensional reconstruction model (3D-EPowerNet model), which gradually infers and restores the three-dimensional shape of the power grid building object based on the two-dimensional CAD image data after data preprocessing; wherein, the power grid building site information and the two-dimensional feature information of various building units in the site are obtained through feature extraction, and the building unit is a power grid functional unit composed of at least one power equipment; the extracted two-dimensional feature information is classified and labeled through semantic segmentation, and the building elements are identified based on the large language model to determine whether the building elements meet the current scene and construction conditions.

[0052] Specifically, the two-dimensional feature information of the power grid building site information and various types of power equipment in the site includes plane building position feature information, plane building unit feature information and building unit connection relationship feature information; the plane building position feature information at least records the coordinate positions of all building units; the plane building unit feature information at least records the appearance shape of a single building unit; the building unit connection relationship feature information at least records the connection relationship between connected building units.

[0053] In some embodiments, the step of obtaining the grid building site information and the two-dimensional feature information of various building units in the site through feature extraction includes the following steps S311-S313.

[0054] S311. In the 2D convolutional neural network layer (2D-CNN), multiple two-dimensional convolutional layers are used to perform convolution processing on the pre-processed two-dimensional CAD image data; each two-dimensional convolutional layer is equipped with a convolution kernel of a set size and step length, and these convolution kernels slide in the two-dimensional space to extract different features in the two-dimensional CAD image data; a nonlinear activation function is added after each two-dimensional convolutional layer to introduce nonlinear transformation, and the feature map to be processed is output. For example, a shallow convolutional layer can extract smaller local features, such as the detailed structure of the power supply room; while a deep convolutional layer can capture a wider range of contextual information, such as the overall direction and layout of the transmission line.

[0055] In a specific embodiment, the expression of a 2D convolutional neural network layer is as follows:

[0056] in, Represents the output feature map to be processed; Represents the convolution kernel weight matrix; Represents the input feature map, i.e., the preprocessed 2D CAD image data; The output feature map to be processed can be updated and optimized through training for advanced feature extraction. The number of rows of the convolution kernel; Indicates the number of columns of the convolution kernel; Indicates the row stride of the convolution kernel; Indicates the column stride of the convolution kernel. In this design, the stride length is 1. Since CAD drawings are usually monochrome drawings, the number of channels can be set to 3. The process of the 2D convolutional neural network layer converting the input image into the feature map to be processed is as follows: Figure 2 shown.

[0057] A nonlinear activation function (ReLU) is added after each convolutional layer to introduce nonlinear transformations, enhance the model's expressiveness, and enable the network to learn more complex feature mapping relationships, thereby better fitting the diverse forms and characteristics of power facilities. The formula for the ReLU activation function is:

[0058] in, Represents the output value after being processed by the activation function. Indicates the input value. This means that when the input value When greater than 0, the output value is the input value itself; when the input value When it is less than or equal to 0, the output value is zero.

[0059] S312: Perform a maximum pooling operation on the feature map to be processed, select the maximum eigenvalue in each pooling window as output, and generate a local feature vector.

[0060] Specifically, the maximum pooling operation is performed on the feature map, and the maximum eigenvalue in each pooling window is selected as the output, which effectively reduces the spatial dimension of the feature map, reduces the amount of calculation and the number of parameters, and retains the most important feature information. This helps prevent overfitting, improves the generalization performance of the model, and enables the model to focus on learning and identifying key features.

[0061] In a specific embodiment, the expression of the maximum pooling operation is:

[0062] in, Represents the value of the local feature vector at position (i, j), that is, the maximum value after pooling; Indicates the pooling step size, that is, the number of pixels the window slides each time, ranging from 1 to 3, depending on the required feature detail; Indicates the number of rows of the pooling window; Indicates the number of columns of the pooling window; , Indicates selecting the maximum value of the pooling window within the given index range m∈[1,P], n∈[1,Q].

[0063] In this application, the maximum pooling formula is used to slide a fixed-size pooling window on the feature map to be processed and select the maximum eigenvalue in each window as the output, thereby reducing the spatial dimension of the data while retaining the most important feature information.

[0064] In a specific embodiment, the pooling layer downsamples the feature map output by the convolution layer through a sliding window (such as 2×22×2), which greatly reduces the amount of data. For example, a 1000×10001000×1000 power transmission line plan may be compressed to 500×500500×500 after pooling, significantly reducing the computational complexity of subsequent modules (such as fully connected layers, 3D-LSTM).

[0065] S313, performing feature fusion and conversion on the local feature vectors through a fully connected layer, learning the connection relationship and layout rules between different building units, and combining the local feature vectors into a global feature vector.

[0066] Specifically, after multiple layers of convolution and pooling operations, the two-dimensional feature tensor is flattened into a one-dimensional vector, and further feature fusion and transformation are performed through the fully connected layer. The fully connected layer combines local features into a global feature representation, providing a more abstract and representative feature vector for subsequent classification or regression tasks. In this process, the model can learn the complex relationships and patterns between different power building units, such as the connection between different types of boilers, condensers, high and low pressure heaters, condensate pumps, main transformers, distribution devices, batteries, control panels, etc.

[0067] In a specific embodiment, the expression of the fully connected layer is as follows:

[0068] in, Represents the output vector of the fully connected layer, that is, the global feature vector, and its value range is ; Represents the input vector of the fully connected layer, that is, the local feature vector, and its value range is ; represents the weight matrix, ; Represents the bias vector, and its value range is [ ; N represents the number of elements in the output vector; For example, suppose there is a fully connected layer whose input vector = [1, 2] (dimension = 2), output vector There need to be 3 elements (ie N=3).

[0069]

[0070] The output vector It can be obtained by the following calculation:

[0071] This output vector It can be used as input for the next layer.

[0072] In some embodiments, the classifying and labeling of the two-dimensional feature information by semantic segmentation includes the following steps S321-S323.

[0073] S321. In the decoder, a deconvolution layer is used to perform an upsampling operation on the global feature vector, so as to expand the low-resolution feature map into a higher-resolution feature map, and restore the original spatial resolution of the data with the decoded feature map.

[0074] Specifically, in order to restore the original spatial resolution of the data, the deconvolution layer is used to upsample the output of the fully connected layer. The deconvolution layer expands the low-resolution feature map into a high-resolution feature map by transposing the convolution kernel, and gradually restores it to the same or similar size as the input data for pixel-level classification and labeling. It can be understood that the encoder part corresponds to the feature extraction part, and the decoder is the corresponding content of the above deconvolution layer.

[0075] In some embodiments, a skip connection is introduced in the decoder to fuse the feature map of the corresponding level in the encoder with the upsampled feature map in the decoder. This connection method can pass the coarse-grained feature information extracted from the encoder to the decoder, assisting the decoder to better restore detail information, especially when dealing with areas with blurred boundaries or occlusions of power facilities, which can improve the accuracy and completeness of semantic segmentation.

[0076] S322. For each pixel or voxel in the decoded feature map, the probability distribution of each pixel or voxel belonging to each building unit type is calculated by the Softmax function to obtain the type label that each pixel or voxel most likely belongs to.

[0077] Specifically, the Softmax function converts the raw scores output by the model into probability values ​​so that the sum of the probabilities of all categories is 1, thereby obtaining the category label to which each pixel or voxel most likely belongs.

[0078] In a specific embodiment, for the global feature vector output by the fully connected layer without the activation function, , n is the total number of categories, The sample belongs to The score of the class, the Softmax function calculation process is:

[0079] represents the conditional probability symbol, y is the true category label, It refers to the probability that the output category is t under the condition of a given input feature x; r represents the category index (traversing all categories), ∈{1,2,…,n}; Indicates the inactive feature vector The nth element of , this score is calculated by the last layer of the neural network; the total number of categories n, that is, the total number of elements in the input vector, uses the number of building unit types defined by the power industry standard, which determines the number of probabilities that the Softmax function needs to calculate.

[0080] The following is an example of step S322: Input data: a pixel feature in the CAD drawing; The model does not have eigenvalues ​​activated: =[3.5,1.2,-0.5], which correspond to the scores of the boiler, power distribution device, and load-bearing wall of the pixel point; The Softmax function is calculated as:

[0081] This indicates that the pixel has an 88% probability of belonging to a boiler, which can be used to guide the subsequent generation of a 3D model.

[0082] Through the above two-dimensional CAD drawing processing, the information of the power grid construction site and the information of various power building units in the site can be obtained, providing basic conditions for the next three-dimensional structure processing.

[0083] S323. Generate a corresponding segmentation mask for each category based on the classification results, and display the distribution of building units in three-dimensional space through the segmentation mask; specifically, the segmentation mask is a binary image or volume data with the same dimension as the input data, in which the position corresponding to the target category is marked as 1, and the background and other categories are marked as 0. These segmentation masks can intuitively display the distribution of power facilities in 3D space, providing accurate geographic information and facility layout reference for power designers.

[0084] After the above steps S2 to S3, the unprocessed two-dimensional CAD image data (usually the two-dimensional CAD image data in the same project at least includes six views of the imaging target, including the top, the bottom and four sides) is first subjected to position rotation, size ratio adjustment and perspective switching to obtain the corrected two-dimensional CAD image data, and then subjected to clarity enhancement and noise reduction processing to obtain the pre-processed two-dimensional CAD image data. After the feature extraction and classification annotation described in step S3 are performed based on the pre-processed two-dimensional CAD image data, the plane building position feature information, the plane building unit feature information and the building unit connection relationship feature information are respectively classified.

[0085] In some embodiments, after feature extraction and annotation classification are completed, step S3 identifies the building elements based on the large language model to determine whether the building elements meet the current scene and construction conditions, including: The large language model uses the plane building location feature information to determine whether the building elements conform to the current scene according to the building requirements read based on the project information, that is, to determine whether the location of each building unit is reasonable; the building elements that do not conform to the current scene are marked and recorded, and the error information can be directly output for reference by designers; The large language model uses the feature information of the plane building unit to determine whether the current building unit meets the construction needs of the corresponding building unit based on the building requirements read based on the project information; it also marks and records the building units that do not meet the construction conditions, and can directly output the error information for reference by designers; After completing the determination of the current scene and construction conditions, the large language model supplements the building unit with detailed drawing information based on the knowledge base and the precautions entered in the project information.

[0086] In addition, in this step, the erroneous information annotated by the large language model or the building units that do not meet the construction conditions can be marked in red so as to be displayed distinguishably in the final generated model.

[0087] S4. Select a matching basic building unit model in the prefabricated model database according to the two-dimensional feature information; place the basic building unit model into the three-dimensional coordinates of the power grid construction site according to the coordinate information and the layout information of the plan; and establish a mapping relationship between the two-dimensional image and its corresponding basic building unit model based on the six-sided view rendering engine to perform spatial three-dimensional imaging and modeling tasks.

[0088] Specifically, step S4 includes the following steps S41-S43.

[0089] S41. Select a matching basic building unit model in a prefabricated model database according to the extracted planar building unit feature information; wherein the prefabricated model database includes three-dimensional voxel point models of power grid functional units belonging to different types of power stations, and any three-dimensional voxel point model is defined as a basic building unit model.

[0090] Specifically, the basic building unit models included in the prefabricated model database are shown in Table 1: Table 1 Example of basic building unit model

[0091] The basic building unit model records the size description information, 3D modeling information, physical constraint information, etc. of the basic building unit.

[0092] S42. Place the basic building unit model in the three-dimensional coordinates of the power grid construction site according to the coordinate information and the layout information of the plan. Specifically, when we use 2D-CNN to extract the plane building position feature information, the plane building unit feature information and the building unit connection relationship feature information, we can match the corresponding basic building unit model from the prefabricated model library according to the plane building unit feature information, and place the basic building unit model in the three-dimensional coordinates of the construction site according to the coordinate information provided by the large language model and the layout information of the plan.

[0093] S43. After receiving the processed input image sequence, the six-view rendering engine first encodes these image sequences into a series of hidden states; these hidden states capture the temporal dependencies in the image sequence and contain spatial features; then, these hidden states are transmitted to the decoder part. When processing these hidden states, the decoder first applies 3D convolution operations to extract and fuse spatial-temporal features. Next, nonlinear characteristics are introduced through nonlinear activation functions (sigmoid), allowing the model to learn more complex mapping relationships. The resolution of the hidden state is gradually increased until the target output resolution is reached.

[0094] Specifically, the underlying principle of the six-sided view rendering engine is based on the 3D-LSTM network and GRU (Gated Recurrent Unit).

[0095] like Figure 3 As shown in Figure 2, the 3D-LSTM network consists of a set of structured LSTM units with restricted connections forming a three-dimensional grid structure. The six CAD drawings are preprocessed into three-dimensional data and divided into N×N×N grids, where N is the spatial resolution.

[0096] Each 3D-LSTM unit (or 3D-lsmunit) has an independent hidden state and input / output gates to process information from adjacent units and the previous time step. The 3D-LSTM unit state update operation formula is as follows:

[0097]

[0098]

[0099]

[0100]

[0101]

[0102] in, Represents a specific position in 3D space, Represents the index in three dimensions respectively; in 3D-LSTM, such index is used to locate specific input and state; t represents the time step, and t-1 is used to indicate the previous moment relative to the current moment; Represents the input data of the 3D-LSTM network; , , Represent the activation values ​​of the forget gate, input gate, and output gate respectively; and Represent the unit state and candidate unit state respectively; represents the hidden state; σ is the sigmoid function; Represents element-by-element multiplication; tanh represents the hyperbolic tangent function, which is used as the activation function in LSTM to compress the value between -1 and 1; represents the weight, represents the bias parameter, and The subscripts are used to mark the corresponding gates or unit states, where subscript f corresponds to the forget gate, subscript i corresponds to the input gate, subscript o corresponds to the output gate, and subscript c corresponds to the unit state. The weight is set to whether the pixel point exists on the directly read CAD drawing. If it exists, it is 1, and if it does not exist, it is 0. b is set to the weight provided by the large language model. For example, the pixel point at the window position of the power supply room is 0 if there is no window in theory. The drawing effect is as follows Figure 4 shown.

[0103] The GRU layer is placed after the 3D-LSTM layer to assist the 3D-LSTM in processing time series information, especially when long-term dependencies need to be captured, to process time series data and global context information. The update gate and reset gate of the GRU unit can be expressed as:

[0104]

[0105]

[0106]

[0107] in, represents the update gate; Reset gate. Represents the hidden state in the GRU unit; represents the candidate hidden state; Represents the input data of the GRU layer; Represents the weight matrix corresponding to the update gate; Represents the weight matrix corresponding to the reset gate; The weight matrix representing the input vector; represents the weight matrix corresponding to the hidden state; ⊙ represents the element product.

[0108] During the drawing process, the weights are dynamically adjusted according to the characteristic information of the power grid prefabricated model library to make each building more suitable for the power grid. Figure 5 shown.

[0109] Based on the GRU unit-assisted drawing, the mean square error (MSE) calculation method is used as the loss function:

[0110] in, Represents the loss function of the GRU unit; Indicates The true value of the sample is the known observation data; Indicates The predicted value of a sample is the data calculated by the model; Represents the number of samples, that is, the total number of samples contained in the data set. The effect diagram obtained after removing the loss error is as follows: Figure 6 shown.

[0111] In some embodiments, the artificial intelligence power grid building imaging method also includes steps S5 and S6.

[0112] Step S5 is a post-processing operation, specifically, edge smoothing of the generated segmentation mask to remove possible isolated noise points and small misclassified areas, so that the edges of the segmentation results are smoother and more natural, in line with the physical form and spatial distribution characteristics of actual power facilities. This helps to improve the quality and usability of the segmentation results and reduce errors in subsequent analysis and design. In addition, for possible holes in the segmentation mask (areas that are not correctly classified due to occlusion or missing data), a hole filling algorithm based on neighborhood information is used to repair them; for example, the reasonable category affiliation of the hole area can be inferred based on the characteristics and spatial relationships of the surrounding classified pixels, thereby further improving the integrity and accuracy of the segmentation results. The effect diagram after step S5 is shown below. Figure 7 shown.

[0113] Step S6 is visualization output. Specifically, the 3D visualization interface output mainly relies on the 3D images generated by the present disclosure and relies on the 3D imaging API to connect other 3D software for output. There are two specific types: Drawing engine display: Through the above steps, a 3D model image in black, white and red is generated for designers to refer to and display.

[0114] External 3D drawing software display: Through the API interface operation of 3Dmax2022, the final semantic segmentation results are presented to power designers in an intuitive 3D visualization form. The segmentation mask is superimposed on the original 3D imaging data, and different power facility categories are identified with different colors or textures, allowing designers to clearly observe and analyze the spatial distribution, mutual relationship and interaction of power facilities with the surrounding environment.

[0115] After processing, a list of problems generated during the conversion of 2D CAD drawings to 3D views is output. At the same time, the CPM-Bee building-specific large language model is also provided to evaluate the characteristics of building units and construction locations during the execution process.

[0116] Other embodiments of the present invention provide an artificial intelligence power grid building imaging engine based on a large language model, which is used to implement the artificial intelligence power grid building imaging method based on a large language model described in any of the above embodiments, comprising: a three-dimensional image reconstruction module and a large language model processing module; the three-dimensional image reconstruction module receives two-dimensional CAD image data for describing the power grid building; the large language model processing module is used to receive project information, and use the large language model to read the project information and power grid building design standards to obtain building requirements.

[0117] Specifically, the three-dimensional image reconstruction module includes: A data preprocessing unit performs data preprocessing on the two-dimensional CAD image data; the data preprocessing includes format conversion, data cleaning and data enhancement; wherein, in data cleaning, a large language model is called to detect and correct abnormal values ​​in the data; A feature extraction module is used to obtain the two-dimensional feature information of the power grid building site and various types of power equipment in the site through feature extraction; A semantic segmentation module is used to classify and label the two-dimensional feature information through semantic segmentation; the large language model processing module uses a large language model to identify building elements according to the processing results of the feature extraction module and the semantic segmentation module, and determines whether the building elements meet the current scene and construction conditions; The three-dimensional imaging module selects a matching basic building unit model in the prefabricated model database according to the two-dimensional feature information; places the basic building unit model into the three-dimensional coordinates of the power grid building site according to the coordinate information and the layout information of the plan; and establishes a mapping relationship between the two-dimensional image and its corresponding basic building unit model based on the six-sided view rendering engine to perform spatial three-dimensional imaging and modeling tasks.

[0118] In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0119] Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An artificial intelligence power grid building imaging method based on a large language model, characterized in that: include: Acquiring basic data, wherein the basic data at least includes two-dimensional CAD image data and project information for describing a power grid building; Use big language models to read project information and grid building design standards to obtain building requirements; Performing data preprocessing on the two-dimensional CAD image data; the data preprocessing includes format conversion, data cleaning and data enhancement; wherein, in data cleaning, a large language model is used to detect and correct outliers in the data; A three-dimensional reconstruction model is constructed, wherein the three-dimensional reconstruction model gradually infers and restores the three-dimensional shape of the power grid building object according to the two-dimensional CAD image data after data preprocessing; wherein the power grid building site information and the two-dimensional feature information of various building units in the site are obtained by feature extraction, wherein the building unit is a power grid functional unit composed of at least one power equipment; the extracted two-dimensional feature information is classified and labeled by semantic segmentation, and the building elements are identified based on the large language model to determine whether the building elements meet the current scene and construction conditions; According to the two-dimensional feature information, a matching basic building unit model is selected in the prefabricated model database; the basic building unit model is placed in the three-dimensional coordinates of the power grid construction site according to the coordinate information and the layout information of the plan; and based on the six-sided view rendering engine, a mapping relationship is established between the two-dimensional image and its corresponding basic building unit model to perform spatial three-dimensional imaging and modeling tasks.

2. The artificial intelligence power grid building imaging method based on a large language model according to claim 1 is characterized in that: The two-dimensional CAD image data records the power grid building site information and the image information of various building units in the site, and the image information includes the two-dimensional coordinates, geometric shape and surface features of the power grid building; The project information includes at least one of a project name, a project type, a special situation description file, a risk description file, a CAD file description list, a building review file list, and voxel points; The special situation description file records the special information of the designated building, and the special information includes the location, size and angle of the designated building; The risk description document is used to restrict certain buildings that are prohibited from construction; The CAD file description list is used to describe the function of each two-dimensional CAD image data; The building audit document list is used to supplement and verify building parameters.

3. The artificial intelligence power grid building imaging method based on a large language model according to claim 2 is characterized in that: The large language model adopts a large language model specialized in CPM-Bee architecture; the method of using the large language model to detect and correct outliers in the data includes: The CPM-Bee building-specific large language model identifies the height information, distance information and angle information in the two-dimensional CAD image data that deviate from the required range based on its own knowledge base and the building review document list, and marks and corrects the deviated data to limit it to the required range; For data information of designated buildings that do not meet the requirements of the CPM-Bee building-specific large language model knowledge base and building audit document list, but are recorded in the special situation description file, the data will be checked and corrected based on the requirement range specified in the special situation description file.

4. The artificial intelligence power grid building imaging method based on a large language model according to claim 2 is characterized in that: The two-dimensional characteristic information of the power grid construction site and various types of power equipment in the site includes two-dimensional building location characteristic information, two-dimensional building unit characteristic information and building unit connection relationship characteristic information; The two-dimensional feature information of the power grid construction site and various building units in the site are obtained by feature extraction, including: A plurality of two-dimensional convolutional layers are used to perform convolution processing on the preprocessed two-dimensional CAD image data; each two-dimensional convolutional layer is equipped with a convolution kernel of a set size and step length, and these convolution kernels slide in a two-dimensional space to extract different features in the two-dimensional CAD image data; a nonlinear activation function is added after each two-dimensional convolutional layer to introduce a nonlinear transformation, and a feature map to be processed is output; Perform a maximum pooling operation on the feature map to be processed, select the maximum eigenvalue in each pooling window as the output, and generate a local feature vector; The local feature vectors are subjected to feature fusion and transformation through a fully connected layer, the connection relationship and layout rules between different building units are learned, and the local feature vectors are combined into a global feature vector.

5. The artificial intelligence power grid building imaging method based on a large language model according to claim 4 is characterized in that: The classifying and labeling the two-dimensional feature information by semantic segmentation includes: In the decoder, a deconvolution layer is used to perform an upsampling operation on the global feature vector, so as to expand the low-resolution feature map into a higher-resolution feature map, and restore the original spatial resolution of the data with the decoded feature map; For each pixel or voxel in the decoded feature map, the probability distribution of each building unit type is calculated through the Softmax function to obtain the most likely type label of each pixel or voxel, including: For the global feature vector output by the fully connected layer without activation function , n is the total number of categories, The sample belongs to The score of the class, the Softmax function calculation process is: represents the conditional probability symbol, y is the true category label, It refers to the probability that the output category is t under the condition of a given input feature x; r represents the category index, traversing all categories, ∈{1,2,…,n}; Indicates the inactive feature vector The nth element of , this score is calculated by the last layer of the neural network; the total number of categories n, that is, the total number of elements in the input vector, uses the number of building unit types defined by the power industry standard, which determines the number of probabilities that the Softmax function needs to calculate; According to the classification results, a corresponding segmentation mask is generated for each category, and the distribution of building units in three-dimensional space is displayed through the segmentation mask; In the three-dimensional reconstruction model, the plane building position feature information, the plane building unit feature information and the building unit connection relationship feature information are classified respectively.

6. The artificial intelligence power grid building imaging method based on a large language model according to claim 5 is characterized in that: The classifying and labeling the two-dimensional feature information by semantic segmentation further includes: A skip connection is introduced in the decoder to fuse the feature maps of the corresponding levels in the encoder with the upsampled feature maps in the decoder.

7. The artificial intelligence power grid building imaging method based on a large language model according to claim 5 is characterized in that: The identification of building elements based on the large language model to determine whether the building elements meet the current scene and construction conditions includes: The large language model uses the plane building location feature information to determine whether the building elements are in line with the current scene based on the building requirements read based on the project information, that is, to determine whether the location of each building unit is reasonable; the building elements that do not meet the current scene are marked and recorded; The large language model determines whether the current building unit meets the construction needs of the corresponding building unit based on the building needs read based on the project information based on the feature information of the planar building unit; After completing the determination of the current scene and construction conditions, the large language model supplements the building unit with detailed drawing information based on the knowledge base and the precautions entered in the project information.

8. The artificial intelligence power grid building imaging method based on a large language model according to claim 7 is characterized in that: The selecting a matching basic building unit model in the prefabricated model database according to the two-dimensional feature information, and placing the basic building unit model into the three-dimensional coordinates of the power grid construction site according to the coordinate information and the layout information of the plan; Based on the six-view rendering engine, a mapping relationship is established between the two-dimensional image and its corresponding three-dimensional voxel point model to perform spatial three-dimensional imaging and modeling tasks, including: Selecting a matching basic building unit model in a prefabricated model database according to the extracted planar building unit feature information; wherein the prefabricated model database includes three-dimensional voxel point models of power grid functional units of different types of power stations, and defining any three-dimensional voxel point model as a basic building unit model; The basic building unit model is placed in the three-dimensional coordinates of the power grid construction site according to the coordinate information and the layout information of the plan; After receiving the processed input image sequence, the six-view rendering engine first encodes these image sequences into a series of hidden states; these hidden states capture the temporal dependencies in the image sequence and contain spatial features; then, these hidden states are transmitted to the decoder part; when processing these hidden states, the decoder first applies 3D convolution operations to extract and fuse spatial-temporal features; then, nonlinear characteristics are introduced through the nonlinear activation function sigmoid, so that the model can learn more complex mapping relationships; the resolution of the hidden state is gradually improved until the target output resolution is reached.

9. The artificial intelligence power grid building imaging method based on a large language model according to claim 1 is characterized in that: Also includes: Perform edge smoothing on the generated segmentation mask to remove possible isolated noise points and small misclassified areas; For possible holes in the segmentation mask, a hole filling algorithm based on neighborhood information is used to repair them; The final semantic segmentation results are presented to power designers in an intuitive 3D visualization form; the segmentation mask is superimposed on the original 3D imaging data to identify different power equipment categories with different colors or textures.

10. An artificial intelligence power grid building imaging engine based on a large language model, characterized in that: The artificial intelligence power grid building imaging method based on a large language model as described in any one of claims 1 to 9, wherein the artificial intelligence power grid building imaging engine comprises: a three-dimensional image reconstruction module and a large language model processing module; the three-dimensional image reconstruction module receives two-dimensional CAD image data for describing the power grid building; the large language model processing module is used to receive project information, and use the large language model to read the project information and power grid building design standards to obtain building requirements; The three-dimensional image reconstruction module comprises: A data preprocessing unit performs data preprocessing on the two-dimensional CAD image data; the data preprocessing includes format conversion, data cleaning and data enhancement; wherein, in data cleaning, a large language model is called to detect and correct abnormal values ​​in the data; A feature extraction module is used to obtain the two-dimensional feature information of the power grid building site and various types of power equipment in the site through feature extraction; A semantic segmentation module is used to classify and label the two-dimensional feature information through semantic segmentation; the large language model processing module uses a large language model to identify building elements according to the processing results of the feature extraction module and the semantic segmentation module, and determines whether the building elements meet the current scene and construction conditions; The three-dimensional imaging module selects a matching basic building unit model in the prefabricated model database according to the two-dimensional feature information; places the basic building unit model into the three-dimensional coordinates of the power grid building site according to the coordinate information and the layout information of the plan; and establishes a mapping relationship between the two-dimensional image and its corresponding basic building unit model based on the six-sided view rendering engine to perform spatial three-dimensional imaging and modeling tasks.

Images