Nuclear power virtual construction method and system based on artificial intelligence

Abstract

The invention discloses a nuclear power virtual construction method and system based on artificial intelligence. The method comprises the following steps: identifying and extracting drawing information data; performing digital twin modeling based on the extracted drawing information data to generate a completion prediction model; comparing and verifying the completion prediction model with a historical empirical model in the reference model library, if abnormity exists, positioning an abnormal point in the completion prediction model, associating the number of a drawing where the abnormal point is located, and feeding back abnormal information; if no abnormity exists, the model is judged to be a correct completion prediction model; based on the correct completion prediction model, an optimal construction path is searched, and a construction scheme is optimized; and on the basis of the correct completion prediction model, predicting items and workload information required by completion. According to the method, the prediction time spans from the drawing stage to the completion state, the time difference between drawing publishing and problem feedback is reduced, the problem prediction is more comprehensive, the abnormal positioning error can be reduced, and the method has a more intelligent modeling capability, a more timely problem identification capability and a more energy-saving and efficient construction capability.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of nuclear power engineering construction, and particularly relates to a nuclear power virtual construction method and system based on artificial intelligence. BACKGROUND

[0002] The statements herein are provided only to complement the present application and are not necessarily to constitute the prior art.

[0003] In recent years, the nuclear power industry is in the ascendant period of high-quality development, and the nuclear power construction is in the key period of intelligent transformation. The intelligent level of the nuclear power engineering design and construction process plays a key role in the economy and efficiency of the construction stage.

[0004] The traditional nuclear power construction technology has been at a relatively advanced level, but there is still room for improvement in terms of work efficiency. For example, the traditional nuclear power unit construction stage needs to be constructed according to the design drawings. Only when problems occur will the design be fed back. It is easy to cause rework, delay the construction period and increase the cost due to the problems of the drawings. For another example, the three-dimensional laser scanning technology of the traditional nuclear power unit construction stage only reflects the current construction situation, and does not use artificial intelligence algorithms to predict the future and even the completed state at the stage when the drawings are designed and have not been constructed, so as to timely find out the problems that may occur in the design and construction. For another example, the construction collision simulation technology of the traditional nuclear power unit construction stage only reflects the collision errors of the current model, and does not verify other problems such as position deviation. In addition, the traditional digital twin technology only contains the models of the virtual and real scenes, and does not use artificial intelligence to train the experience model as a third reference scene.

[0005] In summary, the inventors found that the traditional nuclear power construction technology has the following problems: (1) The traditional nuclear power construction relies on a linear process. After the design drawings are published, the construction needs to be manually interpreted. The problem feedback cycle is as long as several weeks, the rework rate is high, the problem solving efficiency is low, and the design-construction fault is easy to occur.

[0006] (2) The existing three-dimensional laser scanning only reflects the current construction state statically, lacks dynamic completion prediction ability based on historical data, and the collision detection only verifies the geometric conflict, and does not accurately compare and verify the model to three coordinates, resulting in that the prediction of the traditional simulation has limitations.

[0007] (3) The traditional digital twin technology only contains the virtual and real double model comparison, and does not have the reference of the experience model, so the abnormal positioning error is high.

[0008] (4) The construction sequence of the traditional nuclear power construction relies on manual experience, and cannot quantitatively evaluate the multi-objective optimal solution of the construction period-cost-safety.

[0009] (5) The traditional nuclear power construction cannot predict the required material resources and man-hours at the start of construction, and cannot predict the manpower and material resources required for completion. SUMMARY

[0010] The present application aims to overcome the deficiencies in the prior art, and provides a nuclear power virtual construction method and system based on artificial intelligence, which can make the nuclear power unit construction scheme have more intelligent modeling capability, more timely problem identification capability and more energy-efficient construction capability by intelligently identifying drawing information and performing digital twin modeling to predict the completion state, comparing and verifying the historical experience model as an experience model, and performing multi-objective path optimization based on the correct prediction model.

[0011] To achieve the above-mentioned purpose, the present application is realized by the following technical scheme: On the one hand, the technical scheme of the present application provides a nuclear power virtual construction method based on artificial intelligence, comprising: identifying and extracting drawing information data; based on the extracted drawing information data, performing digital twin modeling to generate a completion prediction model; comparing and verifying the completion prediction model with the historical experience model in the reference model library, if there is an anomaly, locating the abnormal point in the completion prediction model and associating the drawing number where the abnormal point is located, and feeding back the abnormal information; if there is no anomaly, determining it as a correct completion prediction model; based on the correct completion prediction model, finding the best construction path and optimizing the construction scheme; based on the correct completion prediction model, predicting the material and workload information required for completion.

[0012] In at least one embodiment, the identification and extraction of drawing information data comprises: based on computer vision technology, automatically identifying and extracting geometric elements, text elements, symbol elements and building information model parameters on the drawing; based on the association relationship between the elements, reconstructing the two-dimensional drawing elements into a three-dimensional model with spatial topological relationship.

[0013] In at least one embodiment, the digital twin modeling based on the extracted drawing information data to generate a completion prediction model comprises: based on a space-time attention mechanism, constructing a long short-term memory network-generative adversarial network joint model, inputting the extracted drawing information data and the current construction progress into the long short-term memory network-generative adversarial network joint model for digital twin modeling, and outputting the completion prediction model.

[0014] In at least one embodiment, the long short-term memory network-generative adversarial network joint model is trained based on historical engineering data in a reference model library, specifically including: extracting drawing design parameters of historical engineering and determining feature fields and target value fields; selecting space-time labels and performing space-time weight distribution; based on the feature fields and target value fields in the historical engineering data set, a data set is constructed, and a long short-term memory network-generative adversarial network joint model is constructed, the long short-term memory network-generative adversarial network joint model is pre-trained based on the data in the historical engineering data set; after the model training is completed, the drawing information data and the current construction progress are taken as inputs for dynamic simulation prediction to obtain predicted target values, and a three-dimensional model is reconstructed according to the predicted target values, and a finished product prediction model integrating time dimension is output.

[0015] In at least one embodiment, the construction process of a nuclear power plant is modularized and decomposed into discrete tasks, drawing information is fused, spatial constraints and time dependencies between discrete tasks are established, safety requirements are integrated, and a finished product state is predicted in a virtual space.

[0016] In at least one embodiment, the finished product prediction model is presented in three-dimensional visualization.

[0017] In at least one embodiment, the comparison and verification of the finished product prediction model with the historical experience model in the reference model library includes: constructing a comparison and verification model based on space-time multi-modal fusion comparison; first, a specified formula is used to select models for comparison and verification from the reference model library; then, cross-modal alignment is performed for the same space-time grid or graph node; subsequently, the multi-modal features of the finished product prediction model and the selected historical experience model at each grid node are input into the space-time multi-modal fusion comparison model, and abnormality degree is calculated after time series convolution, graph neural network and cross-modal attention extraction fusion embedding; if the following condition is met: then it is determined that the finished product prediction model has no abnormality, otherwise the node is marked as an abnormal point, and the position of the abnormal point and its corresponding drawing number are located through the drawing system; wherein μ and σ represent the mean and standard deviation of the historical embedding difference value respectively; k represents the region type, k=2 represents the containment region, and k=3 represents the conventional region.

[0018] In at least one embodiment, if an abnormal point is detected in the finished product prediction model, a plurality of historical experience models are selected again from the reference model library, and whether the abnormal point is abnormal compared with the plurality of historical experience models is calculated respectively, if the abnormal probability of the abnormal point exceeds a certain percentage, the abnormal point is determined as a must-revised abnormal point, if the abnormal probability of the abnormal point does not exceed the certain percentage, the abnormal point is determined as a revisable abnormal point.

[0019] In at least one embodiment, based on the correct completion prediction model, the best construction path is found, specifically: a multi-objective genetic algorithm (NSGA-II) is used for construction path optimization, the comprehensive minimization of duration, cost and risk is taken as the optimization goal, the genetic algorithm iterative optimization is carried out under the constraint conditions of resource allocation, process logic and safety, and the Pareto optimal solution set is output as the best construction path.

[0020] In at least one embodiment, the initial path or initial node of the multi-objective genetic algorithm is artificially selected or randomly selected.

[0021] In another aspect, the technical scheme of the present application also provides an artificial intelligence-based nuclear power virtual construction system, comprising: The drawing information intelligent extraction module is configured to identify and extract drawing information data; The digital twin modeling module is configured to perform digital twin modeling based on the extracted drawing information data, and generate a completion prediction model; The comparison and verification module is configured to compare and verify the completion prediction model with the historical experience model in the reference model library, if there is an anomaly, locate the abnormal point in the completion prediction model and associate the abnormal point with the drawing number, and feed back the abnormal information; if there is no anomaly, it is determined as a correct completion prediction model; The path optimization module is configured to find the best construction path based on the correct completion prediction model, and optimize the construction scheme; The work load prediction module is configured to predict the material and work load information required for completion based on the correct completion prediction model.

[0022] The technical scheme of the present application has the following beneficial effects: The present application belongs to the technical field of nuclear power construction and artificial intelligence, and predicts the completion state by intelligently identifying drawing information and performing digital twin modeling. The prediction model time can span from the drawing stage to the completion state, greatly reducing the time difference between design drawing publication and construction problem feedback. The historical experience model is used as an experience model for comparison and verification, the prediction problem is more comprehensive, and the abnormal positioning error is effectively reduced. Based on the correct prediction model, multi-objective path optimization is carried out to realize the intuitiveness and economic efficiency of the nuclear power construction process. The present application can make the nuclear power unit construction scheme have more intelligent modeling capability, more timely problem identification capability and more energy-efficient construction capability. BRIEF DESCRIPTION OF DRAWINGS

[0023] The drawings accompanying the specification of the present application form part of the present application and serve to provide a further understanding of the present application, the exemplary embodiments of the present application and their descriptions serve to explain the present application and do not constitute an improper limitation of the present application.

[0024] Figure 1 is a schematic diagram of a nuclear power virtual construction method based on artificial intelligence disclosed by embodiment 1 of the present application; Figure 2 is an implementation schematic diagram for identifying and extracting drawing information data disclosed by embodiment 1 of the present application; Figure 3 is a schematic diagram of a completion prediction model construction disclosed by embodiment 1 of the present application; Figure 4 is a schematic diagram of a long short-term memory network-generative adversarial network (LSTM-GAN) joint model disclosed by embodiment 1 of the present application; Figure 5 is a schematic diagram of a model comparison and verification process disclosed by embodiment 1 of the present application; Figure 6 is a schematic diagram of a comparison and verification algorithm architecture disclosed by embodiment 1 of the present application; Figure 7 is an abnormal point judgment schematic diagram disclosed by embodiment 1 of the present application; Figure 8 is a schematic diagram of a construction path optimization process disclosed by embodiment 1 of the present application. DETAILED DESCRIPTION

[0025] It should be pointed out that the following detailed description is exemplary and is intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used in the present application have the same meaning as generally understood by those skilled in the art to which the present application belongs.

[0026] As introduced in the background art, the purpose of the present application is to overcome the shortcomings of the prior art described above, and to provide a nuclear power virtual construction method and system based on artificial intelligence, which identifies drawing information intelligently and performs digital twin modeling to predict the completion state, compares and verifies the historical experience model as an experience model, and optimizes the multi-objective path based on the correct prediction model, so that the nuclear power unit construction scheme has more intelligent modeling capability, more timely problem identification capability and more energy-saving and efficient construction capability.

[0027] Embodiment 1 In a typical embodiment of the present application, as shown in Figure 1 The present embodiment discloses a nuclear power virtual construction method based on artificial intelligence, comprising the following steps: S1. Identify and extract drawing information data; S2. Perform digital twin modeling based on the extracted drawing information data to generate a completion prediction model; S3. Compare and verify the completion prediction model with the historical experience models in the reference model library. If there are any anomalies, locate the anomaly points in the completion prediction model and associate them with the drawing numbers of the anomaly points, and report the anomaly information. If there are no anomalies, the model is determined to be the correct completion prediction model. S4. Based on the correct completion prediction model, find the best construction path and optimize the construction plan; S5. Based on the correct completion prediction model, predict the items and workload required for completion.

[0028] The above-mentioned artificial intelligence-based nuclear power virtual construction method will be described in detail below with reference to specific implementation methods.

[0029] S1. Identify and extract drawing information data.

[0030] The construction of nuclear power plants is based on design drawings. Therefore, the BIM drawings or 2D CAD drawings of nuclear power plants contain all the information required for construction. Based on this, in this step, intelligent computer vision technology is used to automatically identify and extract the information data on the drawings, and upload it to the database offline. The SM4 national cryptographic algorithm is used for editing to ensure that all drawing data is in a physically isolated environment throughout the process, thus ensuring the security of drawing data and information.

[0031] In this embodiment, the information data on the drawing includes four types of elements: (1) Geometric elements: The element information in the drawing used to describe the physical shape and spatial relationship, including outlines (such as equipment outlines, building structure lines, etc.), connection elements (such as pipeline centerlines, flange mating surfaces, etc.), spatial positioning (such as equipment installation coordinates, elevation symbols, etc.), geometric tolerances (such as flatness / roundness, coaxiality / perpendicularity, etc.), and section details (such as weld groove shape, concrete reinforcement arrangement, etc.).

[0032] (2) Text elements: The elements in the drawings used to mark technical parameters and specifications, including dimension markings (such as linear dimensions, angular dimensions, etc.), material specifications (such as metal material grades, concrete grades, etc.), process requirements (such as welding processes, non-destructive testing, etc.), safety markings (such as nuclear safety levels, radiation zones, etc.), technical instructions (such as installation instructions, acceptance standards, etc.) and other text information.

[0033] (3) Symbol elements: The element information in the drawings used to convey engineering semantics and operation instructions, including welding symbols (such as weld type, flaw detection location, etc.), fluid symbols (such as valve type, pipeline medium, etc.), electrical symbols (such as Class 1E equipment, explosion-proof level, etc.), mechanical symbols (such as load direction, constraint type, etc.), radiation protection (such as shielding mark, pollution control, etc.) and other symbol information.

[0034] (4) Building Information Modeling Parameters: The element information in the drawings used to specify the accuracy of the model, including geometric expression accuracy, information depth, time parameters, cost parameters, safety parameters and other parameter information.

[0035] In this embodiment, computer vision technology employs a fusion architecture of Convolutional Neural Network (CNN) and Graph Neural Network (GNN). Taking BIM drawings or 2D CAD drawings as input, the CNN and GNN fusion network identifies and extracts geometric elements, text elements, symbolic elements, and building information model parameters from the drawings. Based on the relationships between these elements (such as the adjacency relationship between pipes and valves, the inclusion relationship of equipment located in a room, and the mechanical relationship between supports and pipes), the feature information extracted from the 2D drawings is reconstructed, and a 3D model with spatial topological relationships is output.

[0036] Specifically, such as Figure 2 As shown, using BIM drawings or 2D CAD drawings as input, the system first reads the drawing information through two parallel residual networks (ResNet-50), forming a "visual + semantic" dual-channel feature. Channel one uses a deformable convolutional network (Deformable CNN) to extract spatial geometric elements (such as pipe routing and equipment outlines), while channel two uses an OCR-BiLSTM fusion network to extract textual elements (such as valve type and equipment number). Then, each element in the drawing (such as a pipe section or a valve) is treated as a "node," and the physical relationships (such as connection, subordination, and adjacency) and logical relationships (such as being in the same system or functional dependence) between elements are treated as "edges." A heterogeneous graph is constructed based on the principle of "node = element, edge = relationship," thus achieving simultaneous expression of geometric and semantic associations. Finally, inference is performed on the constructed heterogeneous graph using a multi-relationship graph neural network (MR-GCN). The three-dimensional topological relationships of elements (such as pipe connection sequence, equipment hierarchy, etc.) are used to complete the accurate depth information of the two-dimensional drawings. Based on the inferred three-dimensional topological relationships, a parametric three-dimensional model is generated through the open-source geometry kernel (OpenCASCADE), which supports subsequent modifications and is output as a STEP / GLTF model. At the same time, the topological structure of the three-dimensional model is abstracted into a lightweight .ntop file and output as a topology database for easy storage and fast loading. Based on the STEP / GLTF model and the topology database, a three-dimensional digital twin of the nuclear power plant is generated, integrating geometric, topological, and real-time data information to facilitate subsequent operation and maintenance simulation or fault prediction.

[0037] S2. Based on the extracted drawing information data, perform digital twin modeling to generate a completion prediction model.

[0038] In this step, digital twin modeling adopts a long short-term memory network-generative adversarial network (LSTM-GAN) joint model based on a spatio-temporal attention mechanism to dynamically simulate and predict the output of a completion prediction model that fuses the time dimension, realizes digital twin mapping of the construction scheme and the completion state, and outputs the completion prediction model, as shown in Figure 3 and Figure 4 The specific process is as follows: (1) First, the historical engineering data in the reference model library is used as the data set for training the LSTM-GAN joint model. The design parameters of the historical engineering drawings are extracted using the method in S1. These design parameters are directly used as feature values of the data set, and on the other hand, they are processed to generate parameterized BREP three-dimensional model entities in S1, and then the geometric coordinates (X, Y, Z) are extracted using OpenCASCADE, and the GNN data is extracted as feature values, and the historical daily completed engineering quantity (since it is historical information, all data is known), resource input (manpower / machinery, etc.), and environmental indicators are extracted synchronously.

[0039] When the historical engineering data is used as the data set to establish and train the machine learning model, the feature fields used include: historical drawing design parameters, three-dimensional model GNN data at different time sequences, historical daily completed engineering quantity, resource input, and environmental indicators. When the historical engineering data is used as the data set to establish and train the machine learning model, the target value fields used include: three-dimensional model geometric coordinates (X, Y, Z) at different time sequences, and grid coding completion status (continuous probability value 0~1) of each construction area at different time sequences.

[0040] (2) Selection of spatio-temporal labels: the construction area grid code and geometric coordinates (X, Y, Z) are used as spatial labels, and the timestamp is used as a temporal label.

[0041] (3) Spatio-temporal weight distribution: time attention module and space attention module are used to adjust the spatio-temporal weight distribution. (4) Target value selection: the data set constructed in (1) has two target fields, and the selection methods include: the continuous probability value (non-discrete category) of the construction area grid code at a specific time sequence is used as the first target value, representing the completion degree of a certain area at a specific time point (such as the completion degree of 87% of the reactor building on the 120th day); the geometric coordinates (X, Y, Z) of the center point of each component at the next time (continuous value) are used as the second target value (the actual position at the next time can be obtained).

[0042] (5) Model establishment: the above feature values and target values derived from the historical engineering data are used as the data set, and the LSTM-GAN regression algorithm is used to establish the model.

[0043] (6) Model "pre-training": divide the data in the historical engineering dataset into a training set and a test set according to 80% and 20%, and train and test the established LSTM-GAN regression model.

[0044] (7) After the model training is completed, the drawing information data extracted in S1 and the current construction progress are taken as inputs for dynamic simulation prediction, to obtain the predicted target value, i.e., the three-dimensional model geometric coordinates (X, Y, Z) at the completion time and the completion state of each construction area grid code (continuous probability value 0~1) at the completion time. According to the three-dimensional model geometric coordinates (X, Y, Z) and the GNN, the three-dimensional model is reconstructed, and the current construction state can be visually seen. The output is a completion prediction model that integrates the time dimension, realizing the digital twin mapping of the construction scheme and the completion state.

[0045] As a further implementation, in the dataset modeled using historical engineering data, the feature fields of total resource quantity and total workload are added. The completion prediction model generated in this way can automatically capture item information from the three-dimensional model generated by S1 identification of drawings, and can predict the resource quantity of pipe supports and other bulk materials required for the project until completion and the workload required for the project until completion (i.e., the quantity of materials such as reinforcing steel and pipes that need to be installed) even before the construction begins.

[0046] As a further implementation, in this embodiment, the modular construction process of a nuclear power plant, such as reactor pressure vessel pre-assembly and main pipe module hoisting, is decomposed into discrete tasks. The drawing information of each discrete task is parsed and extracted through a CNN and GNN fusion network. When digital twin modeling is performed, spatial constraints such as hoisting radius and collision boundary and time dependence are established, and safety requirements such as mechanical simulation (e.g., module hoisting stress analysis) are integrated. The completion state is predicted in a virtual space, and a completion prediction model is generated.

[0047] In this embodiment, the generated completion prediction model is presented in three dimensions for viewing by construction personnel and for subsequent problem identification.

[0048] S3. Compare and verify the completion prediction model with the historical experience model in the reference model library. If there is an anomaly, locate the abnormal point in the completion prediction model and associate it with the drawing number where the abnormal point is located, and feed back the abnormal information. If there is no anomaly, it is determined to be a correct completion prediction model, as shown in Figure 5

[0049] To reduce the abnormal positioning error, in this embodiment, a reference model library is constructed using all historical engineering data of nuclear power plants and their corresponding historical experience models. The historical experience model in the reference model library is used as an experience model to compare and verify the completion prediction model generated in S2.

[0050] ​In this embodiment, a comparison and verification model is constructed based on Spatio-Temporal Multi-Modal Fusion Comparison (ST-MFC), such as... Figure 6 As shown, the specific process is as follows: (1) Select models for comparison and verification from the reference model library by using specified screening: First, perform metadata filtering based on the core attributes of the project to be tested (such as nuclear power plant unit selection, installed capacity, construction technology, geographical environment, major construction units, etc.). Perform label filtering in the reference model library and retain only historical models of "same type" projects. Sort these candidate models by similarity. Based on the attribute vectors they construct in the metadata space or feature space (project duration, key milestones, etc.), calculate the Euclidean distance / cosine similarity with the attribute vector of the current project. Sort by similarity from high to low and select the Top-3 most similar models as the "specified comparison set".

[0051] (2) Cross-modal alignment: To compare information from different modalities on the same spatiotemporal cell / graph node, cross-modal alignment is first performed, which specifically includes: First, perform spatial registration: using camera intrinsic and extrinsic parameters and depth maps or point clouds, the captured images / point clouds are registered to the BIM coordinate system using inverse projection and ICP algorithm; Second, voxelization mapping: Divide the Nx×Ny×Nz grid with the origin O and voxel sizes Δx, Δy, Δz, and map the spatial points to the grid indices (i, j, k); Third, component mapping: establish a component-grid correspondence table to associate BIM components with the corresponding voxels; Fourth, time discretization: for any timestamp ,according to Divided to the number One time slot, of which Indicates the starting time point; Indicates the length of the time slice; Fifth, perform feature standardization by using Z-Score to standardize the dimensions of the features.

[0052] After completing cross-modal alignment, a unified spatiotemporal graph input containing node features, graph structure, and time series is obtained. The aligned spatiotemporal graph is then fed into the ST-MFC engine.

[0053] (3) Perform anomaly comparisons, such as Figure 7 As shown, the multimodal features (including time-series progress, resource input, grid topology, sensor data, 3D coordinates, etc.) of the completion prediction model and the selected historical experience model at each grid node are input into the ST-MFC engine, and then extracted, fused, embedded, and the difference is calculated through temporal convolution, graph neural network and cross-modal attention extraction.

[0054] The comparative verification model identifies the degree of deviation of the main data (spatial grid topology, time of construction progress, regional completion probability distribution, and three-dimensional model geometric coordinates (X, Y, Z)) of both at each node, i.e., the abnormality, if it satisfies: If the completion prediction model is determined to be abnormal, the position of the abnormal point and its corresponding drawing number are located through the drawing system; wherein μ and σ represent the mean and standard deviation of the historical embedding difference, respectively; k represents the region type, and k=2 represents the containment region, and k=3 represents the conventional region.

[0055] Further, if it is detected that the completion prediction model has an abnormal point compared to the currently selected historical experience model, a plurality of historical experience models are selected again from the reference model library, and whether the abnormal point exists compared to the plurality of historical experience models is calculated, if the abnormal probability of the abnormal point exceeds a certain percentage, the abnormal point is determined to be a must-revised abnormal point, and the position of the abnormal point and the drawing number where it is located are fed back to the designer for modification; if the abnormal probability of the abnormal point does not exceed the certain percentage, the abnormal point is determined to be a revisable abnormal point, and the position of the abnormal point and the drawing where it is located are fed back to the designer, who can choose to revise or ignore according to the actual situation, and if ignored, it is marked on the drawing.

[0056] The drawing analyzed and the digital twin modeling are performed again after the designer processes the drawing, a new completion prediction model is generated, and the comparative verification is performed again until it is determined that the completion prediction model is not abnormal, and a correct completion prediction model is obtained.

[0057] S4. Based on the correct completion prediction model, the best construction path is found, and the construction scheme is optimized.

[0058] As shown in Figure 8 , in this step, based on the correct completion prediction model obtained in S3, the cost, duration, and abnormal risk of each path are judged, and the best construction path is used to optimize the construction scheme.

[0059] When performing digital twin modeling to generate a completion prediction model, multiple construction paths are generated based on the standards and specifications of the construction scheme. For several construction paths, a multi-objective genetic algorithm (NSGA-II) is used for construction path optimization, with the comprehensive minimization of duration, cost, and risk as the optimization objective, and the crane radius, parallel resource conflict, and regulatory mandatory nodes as the constraints for genetic algorithm iteration optimization. The Pareto optimal solution set is output as the best construction path, and the construction scheme is optimized.

[0060] For example, in the process of hoisting the main pipeline module, the hoisting sequence of the main pipeline in traditional construction caused a delay of 23 days. In this embodiment, the completion prediction model is built based on the drawings of the main pipeline module, and the completion state of different hoisting sequences is predicted: (1) Path A: hoist the pressure vessel first → completion time 64 days, collision probability 12%; (2) Path B: install the support frame first → completion time 58 days, risk of stress exceeding the standard.

[0061] Generate the optimal path C: segmented hoisting + temporary support through the multi-objective genetic algorithm, construction period 52 days, risk <5%, which can reduce the idle cost of machinery by 1.27 million yuan.

[0062] As an alternative embodiment, the initial path or initial node of the multi-objective genetic algorithm can be manually selected. By manually selecting according to experience, the time required for path optimization can be further reduced, and the demand for computing power can be reduced.

[0063] S5. Based on the correct completion prediction model, predict the material and work quantity information required for completion.

[0064] In this step, based on the correct completion prediction model obtained in S3, the resource quantity of pipeline supports required for the project until completion and the work quantity (the form of work quantity is how many tons of steel reinforcement need to be installed, how many hundreds of meters of pipeline need to be installed, etc.) required for the project until completion are automatically extracted.

[0065] Based on the correct completion prediction model, the material and work quantity information required for completion can be predicted before the construction starts, which can take a step ahead in engineering settlement and resource allocation. The specific process includes: (1) Based on the BIM attributes of the correct completion prediction model, export the three-dimensional model to an IFC standard file; (2) Use Python's ifcOpenShell to read the exported IFC file and output a CSV file of "material summary", to obtain the resource quantity information contained in the three-dimensional model; (3) According to the "work quantity calculation library" of the project department engineer (the work quantity calculation method is different for each material, such as pipeline and support), obtain the work time (work quantity) information contained in the three-dimensional model.

[0066] In this embodiment, by converting the BIM attributes of the three-dimensional model into an IFC file, and then parsing the file, the resource quantity and work quantity of the three-dimensional model are read. When the three-dimensional model is a completion prediction model, the output is "predicted resource quantity, predicted work quantity".

[0067] Embodiment 2 In an exemplary embodiment of the present application, the embodiment discloses a nuclear power virtual construction system based on artificial intelligence, comprising: The drawing information intelligent extraction module is configured to identify and extract drawing information data; The digital twin modeling module is configured to perform digital twin modeling based on the extracted drawing information data, and generate a completion prediction model; The comparison and verification module is configured to compare and verify the completion prediction model with historical experience models in the reference model library, if there is an anomaly, locate the abnormal point in the completion prediction model and associate the abnormal point with the drawing number, and feed back the abnormal information, if there is no anomaly, determine it as a correct completion prediction model; The path optimization module is configured to find the best construction path based on the correct completion prediction model, and optimize the construction scheme; The work load prediction module is configured to predict the material and work load information required for completion based on the correct completion prediction model.

[0068] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for virtual construction of nuclear power plants based on artificial intelligence, characterized in that, include: Identify and extract drawing information data; Digital twin modeling is performed based on the extracted drawing information data to generate a completion prediction model; The completion prediction model is compared and verified with historical experience models in the reference model library. If anomalies are found, the anomaly points in the completion prediction model are located and associated with the drawing numbers of the anomaly points, and the anomaly information is fed back. If no anomalies are found, the completion prediction model is determined to be correct. Based on the correct completion prediction model, find the best construction path and optimize the construction plan; Based on the correct completion prediction model, predict the items and workload required for completion.

2. The artificial intelligence-based virtual construction method for nuclear power plants as described in claim 1, characterized in that, The process of identifying and extracting drawing information data includes: automatically identifying and extracting geometric elements, text elements, symbol elements, and building information model parameters on the drawings based on computer vision technology; and reconstructing the two-dimensional drawing elements into a three-dimensional model with spatial topological relationships based on the relationships between the elements.

3. The artificial intelligence-based virtual construction method for nuclear power plants as described in claim 1, characterized in that, The process of creating a digital twin model based on the extracted drawing information data and generating a completion prediction model includes: constructing a joint model of a long short-term memory network and a generative adversarial network based on a spatiotemporal attention mechanism; inputting the extracted drawing information data and the current construction progress into the joint model of the long short-term memory network and the generative adversarial network for digital twin modeling; and outputting a completion prediction model.

4. The artificial intelligence-based virtual construction method for nuclear power plants as described in claim 3, characterized in that, The Long Short-Term Memory Network-Generative Adversarial Network (GAN-GAN) joint model is trained based on historical engineering data from a reference model library. Specifically, this involves: extracting design parameters from historical engineering drawings and determining feature fields and target value fields; selecting spatiotemporal labels and assigning spatiotemporal weights; constructing the GAN-GAN joint model using the feature fields and target value fields from the historical engineering dataset; pre-training the GAN-GAN joint model using data from the historical engineering dataset; and performing dynamic simulation prediction using drawing information data and the current construction progress as input to obtain the predicted target value. Based on the predicted target value, a 3D model is reconstructed, and a completion prediction model incorporating the time dimension is output.

5. The artificial intelligence-based virtual construction method for nuclear power plants as described in claim 3, characterized in that, The construction process of nuclear power plants is modularized and decomposed into discrete tasks. Drawing information is integrated, spatial constraints and time dependencies between discrete tasks are established, safety requirements are integrated, and the completion status is predicted in virtual space.

6. The artificial intelligence-based virtual construction method for nuclear power plants as described in claim 1, characterized in that, The comparison and verification of the completion prediction model with historical experience models in the reference model library includes: constructing a comparison and verification model based on spatiotemporal multimodal fusion comparison; firstly, selecting models for comparison and verification from the reference model library using a specified filtering method; then, performing cross-modal alignment for the same spatiotemporal cell or graph node; subsequently, inputting the multimodal features of the completion prediction model and the selected historical experience model at each cell node into the spatiotemporal multimodal fusion comparison model, extracting and fusing them through temporal convolution, graph neural network, and cross-modal attention, and calculating the anomaly degree; if the following conditions are met: If the completion prediction model is found to be free of anomalies, then the node is marked as an anomaly and its location and corresponding drawing number are located through the drawing system. Here, μ and σ represent the mean and standard deviation of the historical embedding difference, respectively; k represents the region type, k=2 represents the containment region, and k=3 represents the normal region.

7. The artificial intelligence-based virtual construction method for nuclear power plants as described in claim 6, characterized in that, If an outlier is detected in the completion prediction model, multiple historical experience models are selected from the reference model library again, and the anomaly is calculated relative to the multiple historical experience models. If the probability of the outlier exceeds a set percentage, the outlier is determined to be an outlier that must be revised. If the probability of the outlier does not exceed the set percentage, the outlier is determined to be an outlier that can be revised.

8. The artificial intelligence-based virtual construction method for nuclear power plants as described in claim 1, characterized in that, Based on the correct completion prediction model, the optimal construction path is found. Specifically, a multi-objective genetic algorithm is used to optimize the construction path. The optimization objective is to minimize the combined time, cost and risk. The genetic algorithm is iteratively optimized under the constraints of resource allocation, process logic and safety, and outputs the Pareto optimal solution set, which is used as the optimal construction path.

9. The artificial intelligence-based virtual construction method for nuclear power plants as described in claim 1, characterized in that, The initial path or initial node of a multi-objective genetic algorithm can be manually selected or randomly selected.

10. A nuclear power plant virtual construction system based on artificial intelligence, characterized in that, include: The intelligent drawing information extraction module is configured to: identify and extract drawing information data; The digital twin modeling module is configured to: perform digital twin modeling based on the extracted drawing information data, and generate a completion prediction model; The comparison and verification module is configured to: compare and verify the completion prediction model with the historical experience models in the reference model library; if there are anomalies, locate the anomaly points in the completion prediction model and associate them with the drawing numbers of the anomaly points, and report the anomaly information; if there are no anomalies, it is determined to be the correct completion prediction model. The path optimization module is configured to: find the best construction path and optimize the construction plan based on the correct completion prediction model; The workload prediction module is configured to predict the items and workload required for completion based on the correct completion prediction model.

Images