Parametric simulation design method for carbon dioxide pipeline construction and installation

Abstract

The application discloses a carbon dioxide pipeline construction installation parameterized simulation design method, and belongs to the technical field of pipeline construction simulation design. The method constructs an interactive simulation engine, receives a natural language query sentence input by a user, performs semantic analysis and parameter mapping processing to obtain a construction parameter constraint set, calls a pre-trained generative model to perform a multi-stage question and answer conversation, gradually refines a parameterized simulation model of each construction link, and instantiates and configures the parameterized simulation model through the interactive simulation engine to perform iterative optimization, thereby generating an optimal construction scheme of carbon dioxide pipeline construction installation. The method gradually constructs a refined simulation model through natural language interaction and multi-stage question and answer, and realizes optimal decision of the construction scheme in combination with multi-physical field simulation and multi-objective optimization, thereby improving design efficiency and scheme quality.

Description

Technical Field

[0001] This invention relates to the field of pipeline construction simulation design technology, specifically a parametric simulation design method for carbon dioxide pipeline construction and installation. Background Technology

[0002] The construction and installation of carbon dioxide pipelines involves multiple constraints, including complex terrain, materials, welding processes, and fluid transport. The design process requires balancing numerous parameters. Existing construction and installation design methods largely rely on manual experience and fixed templates. Designers determine parameters such as pipe diameter, wall thickness, installation angle, welding process, and backfilling scheme by estimating or simplifying formulas based on specifications and past project cases. This approach is highly dependent on the designer's personal experience and makes it difficult to systematically translate the specific needs of the project site into precise construction parameter constraints. This is especially true when facing the special physical property requirements of supercritical carbon dioxide transport, where traditional design methods struggle to guarantee the global optimality of the solution. The deficiency in existing technologies lies in the lack of an interactive mechanism to automatically transform vague, unstructured user requirements into a structured set of construction parameters. Users often cannot provide all key parameters completely at once during the initial design phase, and existing methods lack the ability to support multi-round, progressive parameter refinement. This results in inefficient parameter input, prone to omissions, and subsequent simulation modeling is based on an incomplete parameter set, affecting design reliability. Furthermore, existing methods typically separate and independently verify pipeline flow characteristics and welding mechanical response when generating construction plans. They fail to perform joint simulation optimization of fluid mechanics, welding process, and construction procedures in a unified interactive environment. Consequently, they cannot automatically search for the optimal construction decision while taking into account transportation efficiency, construction cost, and safety constraints, resulting in final plans that are often conservative or have potential risks.

[0003] To overcome the above-mentioned shortcomings, it is necessary to solve the problem of how to use natural language interaction to gradually acquire and refine construction parameters in the design and installation of carbon dioxide pipelines, and automatically build parameterized simulation models for each construction stage; at the same time, it is also necessary to solve the problem of how to link multiphysics simulation with construction parameters for iterative optimization to generate the optimal construction scheme that meets multiple objective constraints. Summary of the Invention

[0004] This invention provides a parametric simulation design method for carbon dioxide pipeline construction and installation. It aims to refine user queries into a complete set of construction parameters through natural language interaction and a multi-stage question-and-answer mechanism, automatically construct a parametric simulation model, and generate a comprehensive optimal construction scheme through multidisciplinary joint simulation and multi-objective optimization via an interactive simulation engine. This reduces reliance on human experience and improves the degree of design automation and scheme quality.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a parametric simulation design method for carbon dioxide pipeline construction and installation, comprising: An interactive simulation engine is constructed to receive natural language queries from users regarding the construction and installation of carbon dioxide pipelines. Semantic parsing and parameter mapping are performed on the queries to obtain a set of construction parameter constraints. Based on this set, a pre-trained generative model is invoked to execute a multi-stage question-and-answer session, progressively refining and generating parameterized simulation models for each construction stage. The interactive simulation engine is then used to instantiate, configure, and iteratively optimize the parameterized simulation models, generating the optimal construction plan for the carbon dioxide pipeline installation.

[0006] As a technical solution of this invention, the process of obtaining a set of construction parameter constraints by performing semantic parsing and parameter mapping on natural language query statements specifically involves: calling a pre-trained semantic parsing model to perform dependency parsing and entity recognition on the natural language query statements, extracting construction parameter entities and parameter relation expressions; performing semantic matching processing on the construction parameter entities with a preset construction parameter ontology library to obtain standardized parameter names and parameter value ranges; and performing logical constraint reasoning processing on the parameter value ranges based on the parameter relation expressions to generate a set of construction parameter constraints including pipe diameter, wall thickness, material grade, installation angle, and welding process type. This processing method can transform the user's colloquial natural language input into structured and standardized parameter constraints, providing a reliable data foundation for the accurate construction of subsequent simulation models.

[0007] Preferably, the process of calling a pre-trained generative model to execute a multi-stage question-and-answer session based on a set of construction parameter constraints includes: using the pipe diameter and wall thickness in the construction parameter constraint set as the initial state, calling the generative model to generate the first round of questions and answers, asking the user about the first stage of questions regarding pipe material selection and anti-corrosion layer type; receiving the user's first answer text to the first stage of questions, merging the first answer text with the set of construction parameter constraints, updating the context state of the generative model, and generating a first-stage refined parameter set containing pipe material attributes and anti-corrosion layer parameters; based on the first-stage refined parameter set, calling the generative model to generate a second round of questions and answers, asking the user about the second stage of questions regarding welding process specifications and welding material type; iteratively executing the above question-and-answer session until the generative model determines that all construction parameters are clear, generating a parameterized simulation model containing excavation parameters, assembly parameters, welding parameters, non-destructive testing parameters, and backfill parameters. This method, through multiple rounds of interactive question-and-answer, can gradually uncover and refine the user's deeper needs, ensuring that the simulation model fully covers the key parameters of the entire construction process.

[0008] In the above technical solution, the process of merging the first response text with the construction parameter constraint set to generate the first-stage refined parameter set further includes: performing entity extraction and intent recognition processing on the first response text to extract the pipeline material name and anti-corrosion layer type identifier; performing association query processing between the pipeline material name and a pre-trained material performance knowledge base to obtain the elastic modulus, yield strength, and coefficient of thermal expansion of the pipeline material properties; performing matching processing between the anti-corrosion layer type identifier and a pre-trained anti-corrosion process parameter library to obtain the coating thickness, curing temperature, and adhesion requirements of the anti-corrosion layer parameters; and performing cross-constraint verification processing between the pipeline material properties and anti-corrosion layer parameters and the pipeline wall thickness in the construction parameter constraint set to generate the first-stage refined parameter set. This cross-validation mechanism can effectively avoid conflicts between user input and basic physical constraints, improving the rationality and feasibility of parameter combinations.

[0009] As a preferred implementation method for iteratively executing question-and-answer sessions until a parameterized simulation model is generated, this method includes: calling the parameter integrity detection module of the generative model to identify missing parameters in the currently accumulated parameter set, generating missing parameter category identifiers and priority rankings; based on the priority ranking of missing parameters, calling the generative model to generate question text corresponding to the category, guiding the user to supplement the missing construction parameters; after receiving the user's supplementary answer, performing parameter validity verification on the supplementary answer; if the verification passes, the parameter set is updated; if the verification fails, the generative model is called to generate clarification questions and re-request input; when the parameter integrity detection module determines that the integrity score of the parameter set exceeds a preset threshold, the generative model is called to generate a parameterized simulation model containing the logical relationships of construction procedures. This process ensures that all necessary parameters are complete and verified before automatically building the simulation model, guaranteeing the smooth progress of subsequent simulation calculations.

[0010] As another technical solution of the present invention, the process of instantiating, configuring, and iteratively optimizing a parametric simulation model using an interactive simulation engine to generate the optimal construction scheme includes: inputting the pipe diameter and material properties from the parametric simulation model into a fluid dynamics simulation engine, performing coupled calculations of pressure drop and temperature drop for supercritical carbon dioxide transport, and obtaining pressure distribution curves and temperature distribution curves along the pipeline; inputting welding parameters and non-destructive testing parameters from the parametric simulation model into a welding process simulation engine, performing coupled calculations of temperature field and stress field for multi-layer, multi-pass welding processes, and obtaining residual stress distribution maps and heat-affected zone widths of the welded joints; using the pressure distribution curves, temperature distribution curves, residual stress distribution maps, and heat-affected zone widths as constraints, and calling a multi-objective optimization algorithm to search and optimize the installation angle and backfill parameters in the construction parameter constraint set, generating the optimal construction scheme. This scheme comprehensively optimizes the construction scheme from two dimensions: media transport safety and structural integrity, through the combined simulation of fluid dynamics and welding processes.

[0011] Preferably, the specific steps for using a multi-objective optimization algorithm to perform search and optimization include: constructing a constraint boundary that includes the maximum permissible pressure of the pipeline, the minimum delivery temperature, the maximum residual stress of the weld, and the maximum width of the heat-affected zone; initializing the population of individuals with the installation angle and backfill parameters as decision variables, and minimizing pipeline construction costs and maximizing delivery efficiency as optimization objectives; performing joint simulation calculations using a fluid dynamics simulation engine and a welding process simulation engine on each individual to obtain the corresponding pressure distribution curve, temperature distribution curve, residual stress distribution map, and heat-affected zone width; comparing the joint simulation calculation results with the constraint boundary, selecting feasible solutions that satisfy all constraints, and performing Pareto front analysis on the feasible solutions to obtain the optimal construction scheme. In this way, the best balance point can be found among multiple mutually constraining engineering objectives, achieving the optimal match between the economy and performance of the construction scheme.

[0012] As a further technical solution of the present invention, the method further includes: after the optimal construction plan is generated, calling a pre-trained construction risk prediction model to perform risk assessment processing on the optimal construction plan, obtaining the risk level and risk triggering conditions of each construction process; based on the risk level, calling a generative model to generate risk mitigation questions and answers for high-risk processes, guiding users to input risk handling parameters; integrating the risk handling parameters into the optimal construction plan to generate a risk-controllable construction plan including emergency plans. This technical solution enables the final output construction plan to not only possess theoretical optimality but also the ability to cope with actual on-site risks, improving the robustness and practicality of the plan.

[0013] Regarding the specific implementation of risk assessment and processing, the preferred scheme is as follows: Extract the temporal logical relationships and construction environment parameters of each construction procedure from the optimal construction plan to construct a Bayesian risk network for the construction process; input the Bayesian risk network into the construction risk prediction model, perform risk propagation probability calculation based on graph convolution to obtain the failure probability of each construction node; based on the failure probability and a preset risk level threshold, determine the risk level of each construction procedure, and extract risk triggering conditions according to the conditional probability table in the Bayesian risk network. This risk quantification process can accurately locate weak links in the construction chain and provide precise basis for risk management.

[0014] As a preferred implementation method for integrating risk management parameters into the optimal construction plan, the process includes: parsing the emergency response type and resource allocation quantity from the risk management parameters; mapping the emergency response type to a preset emergency operation procedure template; performing spatiotemporal association binding processing between the emergency operation procedure template and the construction procedures in the optimal construction plan to generate the trigger nodes and response sequence of the emergency plan; overlaying and merging the resource allocation quantity with the material list of the construction procedures to generate a construction resource summary table including emergency materials; and embedding the emergency plan and the construction resource summary table into the data structure of the optimal construction plan to generate a risk-controllable construction plan. This integration method ensures seamless connection between emergency measures and normal construction procedures, enabling the plan to respond to sudden risks in a timely and orderly manner during implementation.

[0015] The technical effects and advantages provided by the present invention in the above technical solution are as follows: By constructing an interactive simulation engine and calling a pre-trained generative model to execute a multi-stage question-and-answer session, the natural language query is parsed into a set of construction parameter constraints. A step-by-step question-and-answer approach guides the user to supplement detailed parameters such as pipe material, anti-corrosion coating type, welding process specifications, and welding material type. The generative model dynamically updates the context state based on each round of user responses, extracts entities from the response text, identifies intent, and correlates queries with the material performance knowledge base and anti-corrosion process parameter database to obtain attributes such as elastic modulus, yield strength, coefficient of thermal expansion, coating thickness, and curing temperature. These attributes are then cross-validated with existing parameter constraints, automatically determining parameter completeness and identifying missing items. Follow-up questions are generated according to priority until all construction parameters are clarified, ultimately generating a parameterized simulation model that includes excavation parameters, assembly parameters, welding parameters, non-destructive testing parameters, and backfill parameters. This approach transforms users' unstructured descriptive requirements into a structured, computable set of construction parameters. Parameter conflicts or omissions are identified instantly during interactive dialogue, avoiding repeated modifications and rework caused by incomplete or ambiguous parameters in traditional design. This significantly shortens the time from requirement input to simulation model generation and improves the accuracy and completeness of parameter settings. The parametric simulation model is instantiated, configured, and iteratively optimized using an interactive simulation engine. Pipe diameter and material properties are input into the fluid mechanics simulation engine to perform coupled calculations of pressure and temperature drops in supercritical carbon dioxide transport, yielding pressure and temperature distribution curves along the pipeline. Welding parameters and non-destructive testing parameters are input into the welding process simulation engine to perform coupled calculations of the temperature and stress fields in a multi-layer, multi-pass welding process, resulting in residual stress distribution maps and heat-affected zone widths of the weld joints. Based on this, constraint boundaries are constructed using the pipeline's maximum permissible pressure, minimum transport temperature, maximum residual stress in the weld, and maximum heat-affected zone width. Installation angle and backfill parameters are used as decision variables, and the optimization objectives are minimizing pipeline construction cost and maximizing transport efficiency. A multi-objective optimization algorithm is invoked to perform joint simulation calculations on each individual population, screening feasible solutions that satisfy all constraints and conducting Pareto front analysis to obtain the optimal construction scheme. This process places fluid dynamics characteristics and welding mechanical response within a unified iterative optimization framework, ensuring that the selection of construction parameters is no longer based on isolated single-point checks, but rather simultaneously weighs cost, efficiency, and safety within the global feasible domain. It automatically generates a comprehensive optimal design that surpasses human experience, significantly improving the global optimality and engineering practicality of the scheme. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0017] Figure 1 This is a flowchart of the parametric simulation design method for carbon dioxide pipeline construction and installation; Figure 2 This is a flowchart of semantic parsing and constraint reasoning processing for carbon dioxide pipeline construction parameters; Figure 3 It is a flowchart for the deduction of construction parameters and generation of parameterized simulation models based on multi-turn dialogue interaction; Figure 4 It is a flowchart of the interactive acquisition, verification and simulation model generation of construction parameters; Figure 5 This is a flowchart of the joint simulation optimization of construction parameters for supercritical carbon dioxide pipelines; Figure 6 This is a flowchart of the process for generating construction risk prediction and risk-controllable construction schemes based on graph convolution. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] See Figure 1 This invention provides a parametric simulation design method for carbon dioxide pipeline construction and installation, comprising: constructing an interactive simulation engine to receive natural language query statements input by users for a carbon dioxide pipeline construction and installation scenario; performing semantic parsing and parameter mapping processing on the natural language query statements to obtain a set of construction parameter constraints; based on the set of construction parameter constraints, calling a pre-trained generative model to execute a multi-stage question-and-answer session, progressively refining and generating parametric simulation models for each construction stage; and using the interactive simulation engine to instantiate, configure, and iteratively optimize the parametric simulation models to generate the optimal construction scheme for the carbon dioxide pipeline construction and installation.

[0020] Example 1: In specific implementation, please refer to Figure 2The pre-trained semantic parsing model employs a Transformer-based encoder-decoder architecture, where the encoder is used for feature extraction and the decoder for dependency parsing and entity recognition. The pre-trained semantic parsing model comprises an input embedding layer, a twelve-layer Transformer encoder layer, a dependency parsing output layer, and an entity recognition output layer. The input embedding layer converts each word in the natural language query into an embedding vector consisting of a word vector, a position vector, and a segment vector. Each of the twelve Transformer encoder layers contains a multi-head self-attention sublayer and a feedforward fully connected sublayer. The multi-head self-attention sublayer uses 12 attention heads, and the feedforward fully connected sublayer has a hidden dimension of 3072. The twelve Transformer encoder layers output a context representation vector for each word. The dependency parsing output layer uses a dual affine attention mechanism, calculating the dependency arc score matrix and relation label score matrix based on the context representation vectors of every two words, and obtaining the dependency parsing tree through a maximum spanning tree algorithm. The entity recognition output layer consists of a linear mapping layer and a conditional random field layer. The linear mapping layer maps the context representation vector of each word to a space with a dimension equal to the number of entity label categories. The conditional random field layer learns the transition probabilities of the label sequence and outputs the optimal entity label sequence during decoding using the Viterbi algorithm.

[0021] The training process for the pre-trained semantic parsing model is as follows: Construction specifications, design drawings, and construction records related to carbon dioxide pipeline construction are collected to construct a corpus of 8500 sentences. Dependency syntax is performed on each sentence in the corpus, annotating the headword index and dependency relation labels for each word. The dependency relation label set includes 45 relations such as "nsubj", "dobj", "amod", and "advmod". Simultaneously, entity annotation is performed on each sentence, annotating the starting position and category labels of nine types of construction parameter entities, including pipe diameter, wall thickness, material grade, installation angle, and welding process type. The corpus is divided into training, validation, and test sets in a 7:2:1 ratio. During training, training sentences are input into the pre-trained semantic parsing model in batches, with a batch size of 16. The input embedding layer of the pre-trained semantic parsing model is initialized using parameters from a pre-trained BERT-base Chinese model. The BERT-base Chinese model contains 12 Transformer layers with a hidden dimension of 768. The loss function during training consists of the sum of dependency parsing loss and entity recognition loss. The dependency parsing loss is calculated using the cross-entropy loss function, which sums the losses for dependency arc classification and relation label classification. The entity recognition loss uses the negative log-likelihood loss from the conditional random field layer. The optimizer used is AdamW, with a learning rate of 5e-5 and a weight decay coefficient of 0.01. After each complete training epoch on the training set, the UAS and LAS scores for dependency parsing and the F1 score for entity recognition are calculated on the validation set. Training stops when the entity recognition F1 score on the validation set no longer increases for three consecutive epochs, the model parameters are saved, and the pre-trained semantic parsing model is obtained.

[0022] After receiving a natural language query from a user regarding the construction and installation of carbon dioxide pipelines, the query is segmented into words, and the segmentation results are input into a pre-trained semantic parsing model. The pre-trained semantic parsing model's input embedding layer converts the segmented results into an embedding vector sequence, which is then processed by a twelve-layer Transformer encoder to extract a context representation vector sequence. The dependency parsing output layer calculates the context representation vector sequence, outputting the probability distribution of each word as the center word of the dependency arc and the probability distribution of the dependency relation label, constructing a dependency parsing tree. Each word node in the dependency parsing tree records its pointed-to center word node and dependency relation label. The entity recognition output layer's linear mapping layer maps the context representation vector sequence, and the conditional random field layer performs sequence labeling and decoding, outputting the entity label corresponding to each word. Words with the same entity label and consecutive positions are merged to obtain construction parameter entities, which retain their original expression in text form. Simultaneously, the node containing the construction parameter entity is located in the dependency syntax tree. A sequence of all words along the shortest dependency path from the first construction parameter entity node to the second is extracted. Comparison keywords and numerical modification words on the path are identified. The construction parameter entity is then combined with the relational words on the path to form a parameter relation expression. The parameter relation expression is stored in a structured triple format. Each triple contains a left operand, a relational operator, and a right operand. Both the left and right operands are construction parameter entities or modified numerical expressions. The relational operators include "greater than," "less than," "equal to," and "between."

[0023] The construction parameter ontology is constructed as follows: National standards, industry standards, and design manuals related to carbon dioxide pipeline construction are collected. Standardized parameter names are extracted from these standard documents, along with a list of units, minimum values, maximum values, and enumerated values ​​for each parameter name. Simultaneously, commonly used parameter aliases are extracted from historical construction plans and material lists, establishing a mapping table between standardized parameter names and parameter aliases. This information is stored in a graph database. Each node in the graph database represents a standardized parameter concept, with node attributes including standardized parameter name, parameter alias set, numerical type, minimum value, maximum value, enumerated value set, and unit. Attribute indexing supports rapid retrieval of standardized parameter names and aliases. The pre-built construction parameter ontology is this graph database.

[0024] When performing semantic matching between construction parameter entities and a pre-defined construction parameter ontology, for each construction parameter entity, the text similarity between the construction parameter entity's text and the standardized parameter name and alias set of each standardized parameter node in the construction parameter ontology is calculated. Text similarity is calculated using cosine similarity based on pre-trained word vectors, which are 200-dimensional word vectors trained using the Tencent Chinese Word Vector Corpus. After segmenting the construction parameter entity text, the average of the word vectors is taken as the text vector. The standardized parameter name or alias is also converted into a text vector, and the cosine similarity between the two text vectors is calculated. The maximum value among all calculated results is taken. If the maximum value is greater than or equal to a pre-defined threshold of 0.82, the construction parameter entity is matched to the standardized parameter node with the highest similarity, the standardized parameter name of that node is obtained, and the minimum and maximum values ​​of that node are extracted as the parameter value range for the construction parameter entity. If the maximum value is less than the pre-defined threshold of 0.82, a prompt message is generated requiring the user to confirm or re-enter the standardized parameter corresponding to the construction parameter entity.

[0025] When performing logical constraint reasoning on the parameter value range based on the parameter relation expression, each parameter relation expression is parsed sequentially. During parsing, the left operand is mapped to the corresponding standardized parameter name, and the parameter value is represented by a variable. The numerical value or variable expression in the right operand is extracted, and inequality constraints are constructed based on the relation operator. For the constraints implied in the parameter relation expression, when the relation operator is "greater than" or "less than", one inequality is directly generated; when the relation operator is "equal to", an equality constraint is generated; and when the relation operator is "between", two inequality constraints are generated. All constants involved in the constraints are directly obtained from the numerical part of the parameter relation expression. After the constraints are generated, the original value range of each standardized parameter in the preset construction parameter ontology is intersected with all inequality constraints involving that standardized parameter to obtain the updated value range. Intersection calculation is achieved through a numerical interval intersection algorithm, which overlaps the current value range of each standardized parameter with the value range defined by the constraint inequality. If the intersection is empty, it is marked as a conflict constraint, and a clarification request is sent to the user. Finally, after logical constraint reasoning, the converged parameter value ranges and definite relationships between the parameters are obtained. The resulting set of construction parameter constraints is organized in a key-value pair list structure, including the value range and unit of pipe diameter, the value range and unit of wall thickness, an enumerated list of material grades, the value range and unit of installation angle, and an enumerated list of welding process types.

[0026] In some embodiments, the following constraint formulas are involved in logical constraint reasoning processing: in, This indicates the value of the pipe diameter, in millimeters. This indicates the value of the pipe wall thickness, in millimeters. This is the diameter-to-wall-thickness ratio factor, whose value is obtained from the ratio factor mapping table based on the material grade. The ratio factor mapping table is a preset table that records the ratios corresponding to each material grade. Values, for example, when the material grade is X52. The value is 8 when the material grade is X70. The value is 11; The diameter compensation constant is 3 mm, and its value is determined according to the minimum machining allowance specified in the pipeline manufacturing standard. The pipeline diameter value must simultaneously satisfy the original value range obtained from the construction parameter library and the lower bound condition specified by the above constraint formula.

[0027] Example 2: In specific implementation, please refer to Figure 3 The pre-trained generative model employs an autoregressive language model based on a Transformer decoder architecture. The model comprises a word embedding layer, forty Transformer decoder layers, and an output projection layer. The word embedding layer maps each word in the input text sequence to a word vector of dimension 4096 and adds it to a positional encoding using rotational positional encoding. Each of the forty Transformer decoder layers contains a masked multi-head self-attention sublayer, a cross-attention sublayer, and a feedforward fully connected sublayer. The masked multi-head self-attention sublayer uses 32 attention heads to perform self-attention calculations on the word sequence preceding the current time step. The cross-attention sublayer, also using 32 attention heads, focuses on historical context information in multi-turn dialogues; its query vector comes from the input of the current decoder layer, while the key and value vectors come from externally stored contextual memory representations. The feedforward fully connected sublayer has a hidden dimension of 16384 and uses the SwiGLU activation function. The output projection layer is a linear transformation layer that maps the representation vector output from the last decoder layer to a dimension equal to the vocabulary size, and then uses a softmax function to generate the probability distribution of the next word. The total number of parameters in the pre-trained generative model is 130 billion.

[0028] The initial pre-training phase of the pre-trained generative model involved collecting publicly available internet text, technical documents, and dialogue data containing 220 billion words. Autoregressive pre-training was performed using standard language model training methods, with the optimization objective being to maximize the log-likelihood of the next word. After initial pre-training, the generative model was fine-tuned on construction-related data. The fine-tuning dataset was constructed by extracting multi-turn question-and-answer dialogues between engineers and designers from historical carbon dioxide pipeline construction project technical briefing records, design change communication records, and expert review meeting minutes. Each dialogue contained consecutive questions and answers, covering all construction stages, including pipeline material selection, anti-corrosion layer selection, welding process specifications, welding material selection, non-destructive testing methods, and determination of excavation and backfill parameters. The extracted dialogues were organized into multi-turn conversation sequences in chronological order, with each sequence containing 4 to 10 rounds of question-and-answer. A total of 12,500 multi-turn conversation sequences were constructed and divided into a fine-tuning training set, a fine-tuning validation set, and a fine-tuning test set in an 8:1:1 ratio. During fine-tuning, each multi-turn conversation sequence in the training set is expanded by turn. For each turn, the input is the concatenated text of all previous turns' dialogues and the question text of the current turn, while the target is the response text of the current turn. The loss for generating the response text is calculated using the cross-entropy loss function. The optimizer used is AdamW, with a learning rate of 2e-5, a batch size of 4 multi-turn conversation sequences, and 8 gradient accumulation steps. The perplexity metric is monitored on the validation set. Fine-tuning stops when the perplexity no longer decreases after two consecutive turns of training, the model parameters are saved, and the pre-trained generative model with fine-tuning completed is obtained.

[0029] After obtaining the set of construction parameter constraints, the pipe diameter and wall thickness in the set are used as the initial state to generate an initial state text sequence. This sequence is formatted as "Pipe diameter: [lower limit of value range] mm to [upper limit of value range] mm, wall thickness: [lower limit of value range] mm to [upper limit of value range] mm". A pre-trained generative model is invoked, appending a special question generation instruction "<|ask|>" to the initial state text sequence. This concatenated text is then fed into the pre-trained generative model. The pre-trained generative model internally maintains a context memory representation list to store processed dialogue history representations. In the first round, the context memory representation list is initially empty. After the input text is processed by the lexical embedding layer and forty Transformer decoder layers, attention calculations are performed in each cross-attention sublayer by reading historical key-value pairs and value vectors from the context memory representation list. Since the list is empty, cross-attention degenerates into self-attention calculations. The output projection layer generates the output sequence until the end-of-line character "<|end|>" is encountered. During the decoding process, a temperature-based sampling strategy with a temperature parameter of 0.8 is used to sample terms from the output distribution. The generated output sequence is the question text in the first round of question-and-answer pairs, which asks the user for specific information about pipe material selection and anti-corrosion coating type, such as material grade and anti-corrosion coating material code.

[0030] In the first round of question-and-answer, the question text is displayed to the user through an interactive simulation engine. After the user inputs the first answer text, the interactive simulation engine receives it. The process of merging the first answer text with the construction parameter constraint set and updating the context state of the pre-trained generative model is as follows: The first answer text and the construction parameter constraint set are converted into a unified structured information block. The construction parameter constraint set retains its original key-value pair format. The first answer text is extracted to identify candidate pipe material names and anti-corrosion layer type identifiers, and these candidate values ​​are merged with the existing parameters in the construction parameter constraint set. Subsequently, the structured information block is serialized into text, inserted at the end of the current dialogue history, and appended with a state update separator "<|update|>". This complete updated text sequence is input into the pre-trained generative model, but instead of generating new question text, the pre-trained generative model internally updates the context memory representation list: the vector corresponding to the separator "<|update|>" position in the representation vector of each word in the last layer obtained after the updated text has been processed by the forty layers of Transformer decoder is used as a new context summary representation and appended to the context memory representation list. In subsequent question-and-answer rounds, each cross-attention sublayer reads key-value pairs and value vectors from this list, thus enabling continuous referencing of historical information. After the context state is updated, the pre-trained generative model generates a first-stage refined parameter set based on the updated state. This first-stage refined parameter set contains, in a structured form, the elastic modulus, yield strength, and coefficient of thermal expansion of the pipe material properties, as well as the coating thickness, curing temperature, and adhesion requirements of the anti-corrosion layer parameters. The specific values ​​of the pipe material properties and anti-corrosion layer parameters are derived through an internally stored knowledge mapping table. This knowledge mapping table has been embedded in the model parameters during the fine-tuning stage of the pre-trained generative model, allowing the model to directly generate corresponding attribute values ​​based on the pipe material name and anti-corrosion layer type identifier.

[0031] Based on the refined parameter set from the first stage, a pre-trained generative model is invoked to generate a second round of question-and-answer pairs. A textual description of the refined parameter set from the first stage is appended to the end of the current dialogue history, along with the question generation instruction "<|ask|>", and input into the pre-trained generative model. The pre-trained generative model's contextual memory representation list has saved the previous state. The cross-attention sublayer reads the historical key-value pairs and value vectors, combining them with the new input to generate the question text for the second round of question-and-answer pairs. The question text for the second round of question-and-answer pairs asks the user for specific details regarding welding process specifications and welding material types, such as welding method codes, welding material grades, and welding parameter ranges.

[0032] After receiving the user's answer to the second-stage question, the same fusion, context state update, and parameter set refinement processes as described above are repeated, sequentially covering all subsequent construction parameter categories that need to be clarified. The pre-trained generative model includes a parameter integrity detection module, which is a multilayer perceptron classifier. It receives the current dialogue state summary vector output from the last layer decoder of the pre-trained generative model as input. This summary vector is obtained by average pooling the representation vectors of all words in the last layer. The output of the parameter integrity detection module is a construction parameter completeness score, ranging from 0 to 1. The parameter integrity detection module contains two hidden layers: the first hidden layer has a dimension of 2048, the second hidden layer has a dimension of 1024, and the output layer has a dimension of 1, using a sigmoid activation function to output the score. The parameter integrity detection module is trained simultaneously with the pre-trained generative model during the fine-tuning phase. During training, dialogue data with labeled integrity is used as the supervision signal. The integrity labeling rule is as follows: when the dialogue explicitly includes parameter entities and parameter values ​​from eight categories—pipe material, anti-corrosion layer, welding process, welding material type, non-destructive testing method, excavation parameters, pairing parameters, and backfill parameters—the integrity label is 1; otherwise, it is 0. The training loss of the parameter integrity detection module is a binary cross-entropy loss, which is added to the generation loss of the generative model with a weight of 0.1 and backpropagated together. After each question-answer pair is generated and the context state is updated, the parameter integrity detection module calculates the current dialogue state summary vector to obtain the current construction parameter integrity score. When the construction parameter integrity score is less than a preset threshold of 0.92, the parameter integrity detection module determines that the construction parameters are not fully clear, and the iteration continues; when the construction parameter integrity score is greater than or equal to the preset threshold of 0.92, the parameter integrity detection module determines that all construction parameters are clear.

[0033] Once the parameter integrity detection module determines that all construction parameters are clear, it calls the pre-trained generative model to generate a parametric simulation model. The complete dialogue history and the final accumulated parameter set are transcribed into text, and the generation instruction "<|generate_model|>" is added as input. The pre-trained generative model outputs a predefined structured text describing the parametric simulation model. The parametric simulation model comprises five main components: excavation parameter set, pairing parameter set, welding parameter set, non-destructive testing parameter set, and backfill parameter set. The excavation parameter set includes trench depth, trench width, slope ratio, and support method; the pairing parameter set includes joint gap, allowable misalignment, and pairing method; the welding parameter set includes welding method, welding current range, welding voltage range, welding speed range, and interpass temperature range; the non-destructive testing parameter set includes testing method, testing ratio threshold, and acceptance level; and the backfill parameter set includes layer thickness, compaction requirements, and backfill material type. Each parameter value in the model is derived from information extracted from the dialogue history and the internal parameter knowledge of the pre-trained generative model. The logical relationships between the parameters in the construction process are represented in the structured text as a directed acyclic graph (DAG). The nodes of the DAG represent the construction process, and the edges represent the sequential dependencies between the processes. After output, a parameterized simulation model containing all the parameters of the construction process is obtained.

[0034] In some embodiments, when updating the context state, the context summary representation recorded in the context memory representation list is calculated in the following manner: in, Indicates the first The context summary representation vector stored in the context memory representation list after the next update has a dimension of 4096. This represents the hidden state vector corresponding to the position of the delimiter "<|update|>" output by the last Transformer decoder layer of the pre-trained generative model, with a dimension of 4096; This indicates the number of feedforward processing branches, with a value of 3. Indicates the first The mapping function of each feedforward processing branch contains a linear mapping layer with dimensions from 4096 to 1024, a GELU activation function, and a linear mapping layer with dimensions from 1024 to 4096. Indicates the first The weight coefficients of each feedforward processing branch are obtained by normalizing the learnable parameter vector through softmax. Initially, the weights of each feedforward processing branch are equal. Representation layer normalization operation. In this way, the information from each dialogue interaction is compressed into a fixed-dimensional summary representation vector, which can be used by the cross-attention sublayer in subsequent rounds.

[0035] Example 3: In specific implementation, please refer to Figure 4 The system performs entity extraction and intent recognition on the first response text using an entity extraction model composed of a bidirectional long short-term memory (LSTM) network and a conditional random field (CRF). The bidirectional LSTM network consists of a forward LSTM layer and a backward LSTM layer, each with a 512-dimensional hidden state. The input is a sequence of vectors mapped from the first response text using pre-trained Chinese word vectors. The forward LSTM layer reads the word vector sequence from left to right and updates the hidden state at each time step; the backward LSTM layer reads the word vector sequence from right to left and updates the hidden state at each time step. The hidden states of the forward and backward LSTM layers at the same time step are concatenated to obtain a context encoding vector for each word, with a 1024-dimensional vector. The CRF layer receives the context encoding vector sequence and maps the 1024-dimensional context encoding vector to an emission score vector with a dimension equal to the number of entity label categories through a linear transformation, while maintaining a trainable label transition score matrix. During the decoding phase, the Conditional Random Field (CRF) layer uses the Viterbi algorithm to calculate the optimal label sequence and outputs the entity label corresponding to each word. The entity label categories include the starting word of the pipe material name, the middle word of the pipe material name, the starting word of the anti-corrosion layer type identifier, the middle word of the anti-corrosion layer type identifier, and non-entity words. Starting words and middle words with the same entity label that are adjacent are combined to extract the pipe material name and anti-corrosion layer type identifier.

[0036] The entity extraction model employs supervised training. The training dataset consists of 8000 user response texts annotated by labelers at the entity level, highlighting the specific locations of words or phrases corresponding to pipe material names and anti-corrosion layer types. The training dataset is divided into a 7:3 ratio to create a training set and a validation set for the entity extraction model. During training, the batch size is set to 32 texts, the optimizer is Adam, and the learning rate is 1e-3. The F1 score is monitored on the validation set, and training stops when the F1 score fails to improve for three consecutive rounds, at which point the entity extraction model parameters are saved.

[0037] Intent recognition processing is performed using a text classification model based on a convolutional neural network. The text classification model consists of an input embedding layer, three convolutional layers, and a classification output layer. The input embedding layer uses the same pre-trained Chinese word vectors as the entity extraction model to convert the first response text into a sequence of word vectors. The three convolutional layers employ one-dimensional convolutional kernels of sizes 3, 4, and 5, with 128 kernels of each size, performing one-dimensional convolution operations on the word vector sequences. Each convolutional kernel slides along the sequence direction, calculating the inner product of the kernel and the local word vector window to generate a feature map. Each feature map undergoes global max pooling, extracting the maximum value as the output feature of that convolutional kernel. The 384 output features from the three convolutional kernel sizes are concatenated into a 384-dimensional feature vector. The classification output layer contains a fully connected layer and a softmax activation function. The fully connected layer maps the 384-dimensional feature vector to an output vector with a dimension equal to the number of intent categories, and the softmax activation function transforms the output vector into a probability distribution for each intent category. The intent categories include four types: material type confirmation, anti-corrosion layer type confirmation, material attribute query, and anti-corrosion process parameter query. The text classification model is trained using the cross-entropy loss function, with Adam as the optimizer and a learning rate of 5e-4. Training is performed on 6000 question-and-answer texts labeled with intent categories, and the model parameters with the highest accuracy on the validation set are saved.

[0038] The pre-trained material property knowledge base is implemented using a relational database. The database contains a pipe property table, whose fields include pipe material name, elastic modulus, yield strength, coefficient of thermal expansion, tensile strength, and Poisson's ratio. The pipe material name field is the primary key. The data in the pipe property table comes from the mechanical property parameters of pipe steel specified in national standards GB / T1591 and GB / T20801. The process of relational query processing is as follows: the extracted pipe material name is used as the query condition to construct a structured query statement, and records in the pipe property table that exactly match the query condition are retrieved. If a matching record is found, the values ​​of the elastic modulus, yield strength, and coefficient of thermal expansion fields are read as the three specific values ​​of the pipe material attribute; if no matching record is found, a material not recognized prompt is generated, and the user is requested to reconfirm the pipe material name.

[0039] The pre-trained anti-corrosion process parameter library is stored in a structured document format (JSON). Each document maintains a mapping between anti-corrosion layer type identifiers and anti-corrosion process parameters. Each anti-corrosion layer type identifier serves as a key, and its corresponding value includes a recommended coating thickness, a curing temperature range, and an adhesion requirement. The data in the pre-trained anti-corrosion process parameter library originates from the anti-corrosion layer construction process parameters specified in the industry standard SY / T0414. The matching process involves performing a string-based exact match between the extracted anti-corrosion layer type identifier and the key in the pre-trained anti-corrosion process parameter library. Upon successful matching, the recommended coating thickness is read as the coating thickness for the anti-corrosion layer parameter, the median of the curing temperature range is read as the curing temperature for the anti-corrosion layer parameter, and the adhesion requirement is read as the adhesion requirement for the anti-corrosion layer parameter.

[0040] The cross-constraint verification process is as follows: The yield strength in the pipe material properties is correlated with the pipe wall thickness in the construction parameter constraint set; the coating thickness in the anti-corrosion layer parameters is correlated with the pipe diameter in the construction parameter constraint set. For the yield strength and pipe wall thickness constraint, a lower limit function for the pipe wall thickness is set, calculated as the design pressure inside the pipe multiplied by the pipe diameter divided by twice the yield strength. The design pressure inside the pipe is obtained from the construction parameter constraint set. If the current pipe wall thickness is less than the calculated lower limit, the pipe wall thickness is updated to the lower limit value, and a wall thickness adjustment notification is sent to the user. For the coating thickness and pipe diameter constraint, an allowable range for the coating thickness is set, with the lower limit and upper limit being the pipe diameter multiplied by 0.0005 and 0.003, respectively. If the coating thickness in the anti-corrosion layer parameters exceeds the allowable range, the coating thickness is adjusted to the endpoint value within the allowable range that is closest to the original coating thickness, and an adjustment notification is sent to the user. After completing the above inspections and adjustments, the elastic modulus, yield strength, and coefficient of thermal expansion in the pipe material properties are combined with the coating thickness, curing temperature, adhesion requirements, and verified pipe wall thickness in the anti-corrosion layer parameters to form a dataset, generating the first-stage refined parameter set.

[0041] During the iterative question-answering session, the parameter integrity detection module operates as follows: Stored within the pre-trained generative model, it receives the dialogue state summary vector obtained by average pooling all hidden word states output from the last decoder layer of the pre-trained generative model. The module's architecture is a multilayer perceptron classifier. The first layer is the input layer, receiving a 4096-dimensional dialogue state summary vector. The second layer is the first hidden layer, containing 2048 neurons with the GELU activation function. The third layer is the second hidden layer, containing 1024 neurons with the GELU activation function. The fourth layer is the output layer, containing 9 neurons, corresponding to binary classification missing value determination for 9 parameter categories: pipe material, anti-corrosion layer, welding process, welding material type, non-destructive testing method, excavation parameters, pairing parameters, backfill parameters, and welding parameters. Each neuron independently outputs the confidence score of whether the parameter category has been acquired using the sigmoid activation function. The parameter integrity detection module is trained simultaneously with the pre-trained generative model during the fine-tuning phase. The training loss is the sum of the binary cross-entropy losses of the nine output neurons. The training label for each neuron is either 1 or 0, based on whether the parameter entity of that parameter category and the parameter value appear in the dialogue history. During training, the loss of the parameter integrity detection module is added to the text generation loss of the pre-trained generative model with a weight of 0.1 and then backpropagated together.

[0042] When performing missing parameter identification on the currently accumulated parameter set, the current dialogue state summary vector is input into the parameter integrity detection module to obtain 9 output confidence values. For each parameter category, if the corresponding output confidence value is less than the preset independent threshold of 0.5, the parameter category is marked as a missing parameter category and added to the missing parameter category identifier list. After all the marked missing parameter categories are processed, they are sorted according to the dependency order of construction procedures, with the following sorting rules: pipe material first, anti-corrosion layer second, welding process third, welding material type fourth, welding parameters fifth, non-destructive testing method sixth, excavation parameters seventh, assembly parameters eighth, and backfilling parameters ninth. The missing parameter categories in the missing parameter category identifier list are then rearranged according to the above order to generate a missing parameter priority sort.

[0043] Based on the priority of missing parameters, a pre-trained generative model is invoked to generate question text for the corresponding category. The pre-trained generative model maintains a mapping between parameter categories and question templates, in key-value pairs. The key is the parameter category name, and the value is a natural language question template guiding the user to complete the corresponding parameters. This mapping data is embedded into the model as parameterized knowledge during the fine-tuning phase of the pre-trained generative model. For the highest-priority missing parameter category after ranking, the pre-trained generative model is invoked with the current dialogue history and the question template corresponding to the highest-priority missing parameter category as context to generate complete question text. The generated complete question text is then displayed to the user through an interactive simulation engine.

[0044] After receiving the user's supplementary response, parameter validity verification is performed. This verification is accomplished jointly by invoking an entity extraction model and a pre-defined construction parameter ontology. The entity extraction model extracts construction parameter entities from the supplementary response, maps these entities to standardized parameter names in the pre-defined ontology, and compares the extracted standardized parameter names with the standardized parameter names corresponding to the currently missing parameter category. If the names match, the supplementary response is further verified to contain valid parameter values, which must fall within the range of the corresponding standardized parameters in the pre-defined ontology. If the names match and the parameter values ​​are valid, the parameter validity verification passes, and the parameter entities and values ​​from the supplementary response are added to the currently accumulated parameter set, updating the parameter set. If the names do not match or the parameter values ​​are invalid, the parameter validity verification fails, a pre-trained generative model is invoked to generate a clarification question. This clarification question includes the standardized parameter name and value range description for the currently missing parameter category, and the user is again requested to input parameter values ​​that meet the constraints.

[0045] When the parameter integrity detection module calculates the updated dialogue state summary vector and determines that the parameter set integrity score exceeds a preset threshold, the parameter set integrity score is taken as the arithmetic mean of the nine output confidence values ​​of the parameter integrity detection module. The preset threshold is set to 0.92, which is based on the mean integrity score calculated from all dialogue samples labeled as complete in the fine-tuning validation set minus one standard deviation, ensuring that the threshold is statistically significant. When the integrity score exceeds 0.92, a pre-trained generative model is invoked to generate a parameterized simulation model containing the logical relationships of construction procedures. When generating the parameterized simulation model, a structured output instruction is appended to the input of the pre-trained generative model, requiring the model to output a JSON structured document containing each parameter category and its specific value, and to describe the sequential relationship between construction procedures in the form of nodes and edges in the procedure dependency field of the document. The pre-trained generative model generates text conforming to the instruction format based on all parameter information obtained from the dialogue history and the internally stored construction procedure logic knowledge. This text, after parsing, becomes the parameterized simulation model containing the logical relationships of construction procedures.

[0046] In some embodiments, the lower limit of the pipe wall thickness in the cross-constraint verification process is calculated based on the following formula: in, This indicates the lower limit of the pipe wall thickness, in millimeters. This represents the design pressure inside the pipeline, expressed in megapascals (MPa), and is obtained from the set of construction parameter constraints. The nominal diameter of the pipe, in millimeters, is obtained from the set of construction parameter constraints. This represents the yield strength, measured in megapascals (MPa), retrieved from a pre-trained material properties knowledge base. This represents the safety factor, with a value of 0.72. This value is based on the design factor for gas transmission pipelines specified in GB50251, and is selected as 0.72 for gas transmission pipelines in Class I areas. If the current pipeline wall thickness is less than... At that time, the wall thickness adjustment operation is triggered.

[0047] Example 4: In specific implementation, please refer to Figure 5The fluid dynamics simulation engine is built upon the continuity, momentum, and energy equations of one-dimensional steady-state pipe flow, and is used to perform coupled calculations of pressure drop and temperature drop in supercritical carbon dioxide transport. The input to the fluid dynamics simulation engine is the pipe diameter and material properties from the parametric simulation model. The pipe diameter specifically includes the pipe inner diameter and pipe wall thickness, while the material properties include the elastic modulus, yield strength, coefficient of thermal expansion, and absolute wall roughness. The pipe inner diameter is obtained by subtracting twice the pipe wall thickness from the nominal pipe diameter. The absolute wall roughness is obtained from a preset wall roughness table based on the pipe material and manufacturing process. The preset wall roughness table stores reference values ​​for the absolute wall roughness corresponding to different pipe materials; for example, the reference value for the absolute wall roughness of X70 pipeline steel is 0.045 mm.

[0048] The pressure drop calculation in the fluid dynamics simulation engine employs a piecewise Darcy-Weisbach formula combined with the supercritical carbon dioxide equation of state. The pipeline is divided into multiple segments of 100 meters each. For each segment, the density and viscosity of carbon dioxide are calculated using the Span-Wagner equation of state, based on the inlet pressure, temperature, and carbon dioxide mass flow rate. The Span-Wagner equation of state, based on Helmholtz free energy, fits the thermal properties of carbon dioxide through 62 parameter terms. Temperature and density are input as variables, and the corresponding pressure is obtained through iterative solutions. After obtaining the density, the flow velocity of carbon dioxide within the segment is calculated, obtained by dividing the mass flow rate by the pipe's inner diameter cross-sectional area and density. The Reynolds number is calculated from the density, flow velocity, pipe inner diameter, and viscosity. The relative roughness is then calculated based on the absolute roughness of the pipe wall and the pipe inner diameter. The Reynolds number and relative roughness are substituted into the Colebrook-White implicit equation for iterative solutions to obtain the Darcy friction coefficient. The Darcy friction coefficient is substituted into the Darcy-Weisbach formula to calculate the pressure drop along the pipe segment. The outlet pressure of the pipe segment is obtained by subtracting the pressure drop along the pipe segment from the inlet pressure. The energy equation uses enthalpy balance to calculate the temperature drop. For each pipe segment, the Joule-Thomson effect and heat exchange with the surrounding environment are considered. The ambient temperature is set to a constant value. The heat exchange is calculated from the overall heat transfer coefficient, the pipe outer surface area, and the logarithmic mean temperature difference between the carbon dioxide temperature inside the pipe and the ambient temperature. The overall heat transfer coefficient is derived from the corrosion protection layer parameters of the parametric simulation model, taking into account the thermal conductivity and thickness of the corrosion protection layer. The outlet temperature and outlet pressure of each pipe segment are solved iteratively. The outlet pressure and outlet temperature of the previous pipe segment are used as the inlet conditions of the next pipe segment until the entire pipeline is calculated. Finally, the pressure distribution curve and temperature distribution curve along the pipeline are obtained. The pressure distribution curve is plotted with distance as the horizontal axis and pressure value as the vertical axis, and the temperature distribution curve is plotted with distance as the horizontal axis and temperature value as the vertical axis. The two curves are stored in the form of a two-dimensional coordinate array.

[0049] In practical implementation, the welding process simulation engine is built based on the three-dimensional thermo-mechanical coupled finite element method to perform coupled calculations of temperature and stress fields in multi-layer, multi-pass welding processes. The input to the welding process simulation engine consists of welding parameters and non-destructive testing parameters from the parametric simulation model. Welding parameters include welding method, welding current range, welding voltage range, welding speed range, interpass temperature range, and welding material grade. Non-destructive testing parameters include testing method, testing ratio threshold, and acceptance level; however, the actual finite element calculation primarily uses welding parameters. The welding process simulation engine establishes a three-dimensional finite element mesh model based on the pipe diameter and wall thickness. The mesh element type is an eight-node hexahedral thermo-mechanical coupled element. Mesh refinement is performed in the weld and heat-affected zone regions, with a mesh size of 0.5 mm in the refined region. Material property assignment: Curves showing the change of elastic modulus, yield strength, coefficient of thermal expansion, thermal conductivity, and specific heat capacity with temperature are set according to the pipe material properties. Mechanical and thermophysical property parameters of the welding material are obtained from the welding material performance database based on the welding material grade. The welding material performance database stores parameters such as chemical composition, yield strength, tensile strength, thermal conductivity, and specific heat capacity of commonly used welding materials, and queries can be performed using the welding material grade as the primary key.

[0050] The heat source model adopts a double-ellipsoidal heat source model. The length of the ellipsoid parameter of the first half of the heat source is 3.2 times the welding input energy, and the length of the ellipsoid parameter of the second half is 4.5 times the welding input energy. The horizontal and vertical semi-axes of the ellipsoids are determined according to the welding method as 0.8 times the weld width and 1.2 times the penetration depth. The welding input energy is calculated using welding current, welding voltage, and welding speed. The thermal efficiency coefficient is set according to the welding method, with 0.78 for gas metal arc welding and 0.75 for manual arc welding. The welding process simulation engine activates weld elements layer by layer and pass by pass according to the welding sequence. The movement of each welding path is achieved by combining the birth and death element technology with the moving heat source. The center of the heat source moves along the weld direction at the welding speed, and the temperature field distribution at the current time step is calculated. The temperature field is solved using transient thermal analysis. The solution equations involve thermal conductivity, specific heat capacity, density, and internal heat generation rate. The backward difference method is used for time integration, with a time step of 0.1 seconds.

[0051] After the temperature field calculation is completed, the stress field is solved. A thermo-elasto-plastic constitutive model is used for the stress field, where the total strain equals the sum of elastic strain, plastic strain, and thermal strain. Plastic strain is described using the von Mises yield criterion and an isotropic hardening model, while thermal strain is calculated using the coefficient of thermal expansion and temperature change. The temperature field at all time steps during the welding process is input as a thermal load into the stress calculation model to obtain the displacement, stress, and strain at each node. Post-processing generates a residual stress distribution map and the width of the heat-affected zone (HAZ) of the weld joint. The residual stress distribution map displays a von Mises stress cloud map of the weld and its surrounding area. The width of the HAZ is determined by extracting the width of the interval where the hardness value drops sharply from the hardness distribution curve perpendicular to the weld centerline. The hardness value is estimated from the critical cooling time and chemical composition using the Vickers hardness conversion formula for the material.

[0052] In practical implementation, the inputs to the multi-objective optimization algorithm include pressure distribution curves, temperature distribution curves, residual stress distribution diagrams, and the width of the heat-affected zone, as well as installation angle and backfill parameters from the construction parameter constraint set. The constraint boundaries consist of the following four conditions: the maximum permissible pressure boundary of the pipeline, determined by multiplying the pipeline design pressure by a margin factor of 1.1; the minimum delivery temperature boundary, set as the supercritical carbon dioxide critical temperature, i.e., 304.1 Kelvin; the maximum residual stress boundary of the weld, set as 0.8 times the yield strength of the pipeline material; and the maximum heat-affected zone width boundary, set as 1.2 times the pipeline wall thickness. The decision variables are the installation angle and backfill parameters, specifically including the layer compaction degree and the elastic modulus of the backfill material. The search range for the installation angle is obtained from the construction parameter constraint set, typically between 0 and 25 degrees; the search range for the layer compaction degree is 90% to 98%; and the search range for the elastic modulus of the backfill material is 10 MPa to 80 MPa.

[0053] The multi-objective optimization algorithm employs the non-dominated sorting genetic algorithm NSGA-II. During population initialization, a Latin hypercube sampling method is used to generate an initial population of 160 individuals within the range of decision variable values. Each individual represents a set of decision variable values. For fitness calculation, a joint simulation calculation using a fluid dynamics simulation engine and a welding process simulation engine is performed on each individual. The fluid dynamics simulation engine adjusts the pipeline's spatial orientation and along-path temperature drop environmental conditions based on the individual's installation angle, calculating the pressure and temperature distribution curves for that individual. The welding process simulation engine adjusts the pipeline support stiffness boundary conditions based on the individual's backfill parameters, calculating the residual stress distribution map and heat-affected zone width for that individual. The optimization objectives are to minimize pipeline construction cost and maximize transport efficiency. Pipeline construction cost is calculated from the earthwork excavation and support engineering volume caused by the installation angle, and transport efficiency is represented by the ratio of pipeline outlet pressure to inlet pressure. The joint simulation results are compared with the constraint boundaries. If the pressure at any point in the pressure distribution curve of an individual exceeds the maximum permissible pressure boundary of the pipeline, or the temperature at any point in the temperature distribution curve is lower than the minimum delivery temperature boundary, or the maximum von Mises stress in the weld region in the residual stress distribution diagram exceeds the maximum residual stress boundary of the weld, or the width of the heat-affected zone exceeds the maximum width boundary of the heat-affected zone, then the individual is judged not to meet the constraints and is eliminated. Individuals satisfying all constraints are selected to form a feasible solution set.

[0054] The process of Pareto front analysis on the feasible solution set is as follows: For each individual in the feasible solution set, calculate its objective function value under the two objectives of minimizing pipeline construction cost and maximizing transportation efficiency. A fast non-dominated sorting method is used to divide the feasible solutions into different non-dominated levels, and the congestion distance between individuals within the same level is calculated. Individuals are selected in order of non-dominated level from low to high and congestion distance from large to small to enter the next generation population. Crossover and mutation operations are performed on the next generation population. The crossover operation uses simulated binary crossover with a crossover distribution exponent of 11 and a crossover probability of 0.9; the mutation operation uses multinomial mutation with a mutation distribution exponent of 21 and a mutation probability equal to the reciprocal of the number of decision variables. After repeating this evolutionary iteration for 500 generations, the set of individuals with the lowest non-dominated level is selected from the final population as the Pareto optimal solution set. From the Pareto optimal solution set, an ideal point method is used to select a compromise optimal solution. The ideal point is the point composed of the optimal values ​​achievable by optimizing each objective individually. The individual with the smallest Euclidean distance from the ideal point in the Pareto front is selected as the optimal construction scheme. The optimal construction plan determines the specific values ​​of the installation angle and backfill parameters, and integrates them with other existing parameters in the parametric simulation model for output.

[0055] In some embodiments, the Darcy-Weisbach formula used in the fluid dynamics simulation engine to calculate the pressure drop along the pipe section is as follows: in, This represents the pressure drop along a single pipe section, measured in Pascals. The Darcy coefficient is dimensionless and is obtained by iteratively solving the Colebrook-White implicit equation. This indicates the length of the pipe section, which is fixed at 100 meters. This indicates the inner diameter of the pipe, in meters. It is obtained by subtracting twice the pipe wall thickness from the nominal diameter of the pipe and then converting it to meters. This represents the density of supercritical carbon dioxide within the pipe section, expressed in kilograms per cubic meter, and is calculated using the Span-Wagner equation of state based on the inlet pressure and temperature of the pipe section. This indicates the flow velocity of supercritical carbon dioxide within the pipe section, expressed in meters per second, and is calculated by dividing the mass flow rate by... The pressure is obtained from the pipe's inner diameter and cross-sectional area. The outlet pressure of the pipe section is obtained by subtracting the inlet pressure of the pipe section. get.

[0056] Example 5: In specific implementation, please refer to Figure 6 After the optimal construction plan is generated, a pre-trained construction risk prediction model is invoked to perform risk assessment on the optimal construction plan, obtaining the risk level and risk triggering conditions for each construction procedure. The pre-trained construction risk prediction model is a Bayesian risk propagation model based on graph convolutional networks. The pre-trained construction risk prediction model contains three graph convolutional layers and a risk prediction output layer. The first graph convolutional layer maps the initial feature dimension of the node from 10 dimensions to 64 dimensions, the second graph convolutional layer maps the 64-dimensional features to 32 dimensions, and the third graph convolutional layer maps the 32-dimensional features to 16 dimensions. The risk prediction output layer contains a fully connected layer that maps the 16-dimensional features to 1 dimension, and then outputs the failure probability of the construction procedure corresponding to the node through a sigmoid activation function.

[0057] The training process for the pre-trained construction risk prediction model is as follows: Construction logs and accident reports from historical carbon dioxide pipeline construction projects are collected. From the construction logs, the temporal logical relationships of construction procedures, construction environmental parameters, and labels indicating whether a quality or safety accident occurred in each construction task are extracted. Construction environmental parameters include seven parameters: soil type code, groundwater level depth, ambient temperature, construction depth, pipe diameter, welding method code, and backfill material type code. These, along with three additional parameters—construction procedure type code, planned procedure duration, and whether anomalies occurred in preceding procedures—are combined to form a 10-dimensional initial feature vector for the nodes. All construction procedures in each task are organized into a directed acyclic graph (DAG) according to their temporal logical relationships. Nodes in the DAG represent construction procedures, and directed edges point from preceding procedures to subsequent procedures. A total of 2800 DAG samples are constructed and divided into a risk model training set, a risk model validation set, and a risk model test set in a 7:2:1 ratio. During training, each DAG from the risk model training set is sequentially input into the pre-trained construction risk prediction model. The three-layer graph convolutional layer propagates node features layer by layer in topological order across the graph. The calculation method for each graph convolutional layer is as follows: for each node, the feature vectors of all its parent nodes are aggregated, transformed by a weight matrix, and then passed through an activation function. The risk prediction output layer calculates the failure probability for each node's feature vector output from the third graph convolutional layer. The loss function uses binary cross-entropy loss, calculating the sum of the predicted failure probabilities of all nodes and the actual accident labels. The optimizer is Adam, with a learning rate of 2e-4 and a weight decay coefficient of 1e-5. After each complete training epoch on the risk model training set, the average AUC value of all nodes is calculated on the risk model validation set. Training stops when the average AUC value no longer increases after four consecutive training epochs, the model parameters are saved, and the pre-trained construction risk prediction model is obtained.

[0058] When calling the pre-trained construction risk prediction model, the temporal logical relationships and construction environment parameters of each construction procedure are first extracted from the optimal construction plan to construct a Bayesian risk network for the construction process. The extraction method is as follows: The procedure dependency field in the optimal construction plan data structure is parsed to obtain the directed dependencies between each construction procedure, generating a list of directed edges; soil type codes, groundwater level depth, and ambient temperature are extracted from the construction environment information field in the optimal construction plan; construction depth, pipe diameter level, welding method codes, and backfill material type codes are extracted from the process parameter fields of each procedure; and the planned duration of each procedure and anomaly markers of preceding procedures are extracted from the construction schedule field, combined into a 10-dimensional initial feature vector for each node. The nodes and directed edges are then combined to form the Bayesian risk network, which is represented in the form of an adjacency matrix and a node feature matrix.

[0059] A Bayesian risk network is input into a pre-trained construction risk prediction model, and risk propagation probability calculation is performed based on graph convolution. On a directed acyclic graph, three graph convolutional layers update node features layer by layer. The output feature vector of the third graph convolutional layer is then used by the risk prediction output layer to calculate the failure probability of each construction node. The failure probabilities of all construction nodes constitute the failure probability set.

[0060] In some embodiments, the node feature update formula for the graph convolutional layer in the pre-trained construction risk prediction model is: in, Represents a node In the The feature vector output by the layered graph convolutional layer, when hour For nodes The 10-dimensional initial feature vector; Represents a node The set of parent nodes, that is, all nodes that have a pointer to a node. The set of nodes with directed edges; Represents a node in-degree, i.e., node The number of parent nodes; Represents a node The out-degree of the node. The number of child nodes; Indicates the first The trainable weight matrix of a layered graph convolutional layer The dimension is , The dimension is , The dimension is ReLU is the modified linear unit activation function. After layer-by-layer calculation using the above formula, the final feature vector of each node is obtained. Then, the failure probability is calculated by the risk prediction output layer.

[0061] After obtaining the set of failure probabilities, the risk level of each construction process is determined based on the failure probabilities and preset risk level thresholds. The risk level thresholds are set as follows: a failure probability greater than or equal to 0.7 indicates a high risk level; a failure probability greater than or equal to 0.3 and less than 0.7 indicates a medium risk level; and a failure probability less than 0.3 indicates a low risk level. The preset risk level thresholds are set based on the distribution of all failure probabilities in the risk model validation set, using the upper tertiary of 0.7 and the lower tertiary of 0.3 as the cutoff values. Simultaneously, risk triggering conditions are extracted based on the conditional probability table in the Bayesian risk network. The conditional probability table in the Bayesian risk network is stored in the statistical module within the pre-trained construction risk prediction model. The conditional probability table records the conditional probability distribution of child node failure under a given combination of parent node states. After training, the conditional probability table is obtained by summarizing the joint frequency of parent node state combinations and child node failure labels in all samples of the risk model training set, followed by Laplace smoothing and normalization. When extracting risk triggering conditions, for process nodes with high risk levels, query all records in the condition probability table that have that node as a child node, take the parent node state combination with the largest condition probability value, and output the specific parent node process name and its state parameter combination in the parent node state combination as the risk triggering condition.

[0062] After obtaining the risk level and risk triggering conditions for each construction process, a generative model is invoked to generate risk mitigation questions and answers for high-risk processes based on the risk level. The generative model used is the same pre-trained generative model used in the parametric simulation model generation stage. The name of the high-risk process, risk level identifier, risk triggering conditions, and a text summary of the current optimal construction plan are concatenated into a prompt text, with a risk mitigation question instruction appended to the end of the prompt text, which is then input into the pre-trained generative model. The pre-trained generative model outputs a series of natural language questions, guiding the user to specify the emergency response type and resource allocation quantity for the high-risk process. The user inputs risk management parameters through the interactive simulation engine. These parameters are recorded in text form as emergency response type codes and resource allocation quantity values. Emergency response type codes include four types: personnel evacuation, on-site reinforcement, equipment replacement, and work stoppage pending inspection. Resource allocation quantity values ​​include the number of emergency personnel and the number of emergency equipment sets.

[0063] After obtaining the risk management parameters, these parameters are integrated into the optimal construction plan to generate a risk-controllable construction plan that includes emergency response measures. The integration process is as follows: The emergency response type and resource allocation quantity in the risk management parameters are parsed, and the emergency response type is mapped to a preset emergency operation procedure template. The preset emergency operation procedure template library is stored in the form of a hash table, where the key is the emergency response type code and the value is a list-represented sequence of emergency operation steps. Each step includes an operation description and standard time consumption. For example, when the emergency response type code is "on-site reinforcement," the corresponding emergency operation procedure template is: Step 1 "Stop adjacent process construction," standard time consumption 0.5 hours; Step 2 "Erect temporary support structure," standard time consumption 2 hours; Step 3 "Check and confirm reinforcement effect," standard time consumption 1 hour. When the emergency operation procedure template is spatiotemporally associated with the construction procedures in the optimal construction plan, the parent node procedure and high-risk procedure corresponding to the risk triggering conditions are identified. The trigger node of the emergency operation procedure template is set to the end time of the parent node procedure, and the response sequence is the offset time sequence of the standard time consumption of each step in the emergency operation procedure template relative to the trigger node. By combining the emergency operation procedure template with its trigger nodes and response sequence, the trigger nodes and response sequence of the emergency plan can be generated.

[0064] When overlaying and merging resource allocation quantities with the bill of materials for construction procedures, the existing bill of materials for each construction procedure is extracted from the optimal construction plan. This bill of materials records the material name, specifications, and quantity in list format. Emergency supplies entries are added to the bill of materials, including emergency personnel and emergency equipment entries. The quantity for the emergency personnel entry is the number of emergency personnel, and the quantity for the emergency equipment entry is the number of sets of emergency equipment. Specifications and names are mapped from the emergency supplies coding library based on the emergency response type. The new entries are then merged with the original bill of materials to generate a master construction resource table containing emergency supplies.

[0065] Finally, the emergency response plan and the construction resource summary table are embedded into the data structure of the optimal construction plan. The data structure of the optimal construction plan is a JSON document, with the top level containing four fields: "Basic Parameters," "Construction Procedure Sequence," "Bill of Materials," and "Construction Plan Attachments." Under the "Construction Plan Attachments" field, two new subfields, "Emergency Response Plan" and "Construction Resource Summary Table," are added. The trigger nodes and response sequences of the emergency response plan are written into the "Emergency Response Plan" subfield in structured text format, and the complete entries of the construction resource summary table are written into the "Construction Resource Summary Table" subfield. The modified JSON document is saved, generating a risk-controlled construction plan.

[0066] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A parametric simulation design method for carbon dioxide pipeline construction and installation, characterized in that, include: Build an interactive simulation engine to receive natural language queries from users regarding carbon dioxide pipeline construction and installation scenarios; The natural language query statement is semantically parsed and parameter mapped to obtain a set of construction parameter constraints; Based on the set of construction parameter constraints, a pre-trained generative model is invoked to execute a multi-stage question-and-answer session, gradually refining and generating parameterized simulation models for each construction stage. The interactive simulation engine is used to instantiate, configure, and iteratively optimize the parametric simulation model, thereby generating the optimal construction plan for the carbon dioxide pipeline installation.

2. The parametric simulation design method for carbon dioxide pipeline construction and installation as described in claim 1, characterized in that, The semantic parsing and parameter mapping processing of the natural language query statement yields a set of construction parameter constraints, including: The pre-trained semantic parsing model is invoked to perform dependency parsing and entity recognition on the natural language query statement, and the construction parameter entities and parameter relationship expressions are extracted. The construction parameter entity is semantically matched with a preset construction parameter ontology library to obtain standardized parameter names and parameter value ranges. Logical constraint reasoning is performed on the parameter value range based on the parameter relationship expression to generate the construction parameter constraint set including pipe diameter, wall thickness, material grade, installation angle, and welding process type.

3. The parametric simulation design method for carbon dioxide pipeline construction and installation as described in claim 1, characterized in that, Based on the set of construction parameter constraints, a pre-trained generative model is invoked to execute a multi-stage question-and-answer session, progressively refining and generating parameterized simulation models for each construction stage, including: Using the pipe diameter and wall thickness in the set of construction parameter constraints as the initial state, the generative model is called to generate the first round of questions and answers, asking the user the first stage questions about pipe material selection and anti-corrosion layer type. Receive the user's first answer text to the first stage question, merge the first answer text with the construction parameter constraint set, update the context state of the generative model, and generate a first stage refined parameter set containing pipeline material properties and anti-corrosion layer parameters; Based on the first stage refined parameter set, the generative model is called to generate a second round of question and answer, asking the user second-stage questions about welding process specifications and welding material types; The above question-and-answer session is executed iteratively until the generative model determines that all construction parameters have been clarified, and the parametric simulation model containing excavation parameters, assembly parameters, welding parameters, non-destructive testing parameters, and backfill parameters is generated.

4. The parametric simulation design method for carbon dioxide pipeline construction and installation as described in claim 3, characterized in that, The process of fusing the first answer text with the construction parameter constraint set, updating the context state of the generative model, and generating a first-stage refined parameter set containing pipe material properties and anti-corrosion layer parameters includes: Entity extraction and intent recognition are performed on the first response text to extract the pipe material name and anti-corrosion layer type identifier; The pipe material name is associated with a pre-trained material property knowledge base to obtain the elastic modulus, yield strength and coefficient of thermal expansion of the pipe material. The anti-corrosion layer type identifier is matched with a pre-trained anti-corrosion process parameter library to obtain the coating thickness, curing temperature and adhesion requirements of the anti-corrosion layer parameters. The pipe material properties and the anti-corrosion layer parameters are cross-constrained with the pipe wall thickness in the construction parameter constraint set to generate the first stage refined parameter set.

5. The parametric simulation design method for carbon dioxide pipeline construction and installation as described in claim 3, characterized in that, The iterative execution of the above question-and-answer session continues until the generative model determines that all construction parameters are clear, generating the parameterized simulation model containing excavation parameters, assembly parameters, welding parameters, non-destructive testing parameters, and backfill parameters, including: The parameter integrity detection module of the generative model is invoked to identify missing parameters in the currently accumulated parameter set, and to generate missing parameter category identifiers and missing parameter priority sorting. Based on the priority of the missing parameters, the generative model is invoked to generate corresponding category question texts to guide users in supplementing the missing construction parameters; After receiving the user's supplementary answer, the supplementary answer is subjected to parameter validity verification. If the verification passes, the parameter set is updated. If the verification fails, the generative model is invoked to generate a clarification question and the input is requested again. When the parameter integrity detection module determines that the integrity score of the parameter set exceeds a preset threshold, it calls the generative model to generate the parameterized simulation model containing the logical relationship of construction procedures.

6. The parametric simulation design method for carbon dioxide pipeline construction and installation as described in claim 1, characterized in that, The step of instantiating, configuring, and iteratively optimizing the parameterized simulation model through the interactive simulation engine to generate the optimal construction plan for the carbon dioxide pipeline installation includes: The pipe diameter and material properties in the parametric simulation model are input into the fluid dynamics simulation engine to perform coupled calculations of pressure drop and temperature drop during supercritical carbon dioxide transport, thereby obtaining the pressure distribution curve and temperature distribution curve along the pipeline. The welding parameters and non-destructive testing parameters in the parametric simulation model are input into the welding process simulation engine to perform coupled calculations of the temperature field and stress field of the multi-layer, multi-pass welding process, and the residual stress distribution map and heat-affected zone width of the weld joint are obtained. Using the pressure distribution curve, temperature distribution curve, residual stress distribution diagram, and heat-affected zone width as constraints, a multi-objective optimization algorithm is invoked to search and optimize the installation angle and backfill parameters in the construction parameter constraint set to generate the optimal construction scheme.

7. The parametric simulation design method for carbon dioxide pipeline construction and installation as described in claim 6, characterized in that, The process involves using the pressure distribution curve, temperature distribution curve, residual stress distribution diagram, and heat-affected zone width as constraints, and employing a multi-objective optimization algorithm to search and optimize the installation angle and backfill parameters within the construction parameter constraint set to generate the optimal construction scheme. This includes: Construct a constraint boundary that includes the maximum permissible pressure of the pipeline, the minimum delivery temperature, the maximum residual stress of the weld, and the maximum width of the heat-affected zone; Using the installation angle and the backfill parameters as decision variables, and minimizing pipeline construction cost and maximizing transportation efficiency as optimization objectives, the population individuals are initialized. For each individual in the population, a joint simulation calculation is performed using the fluid dynamics simulation engine and the welding process simulation engine to obtain the pressure distribution curve, the temperature distribution curve, the residual stress distribution map, and the width of the heat-affected zone corresponding to that individual. The joint simulation results are compared with the constraint boundaries to select feasible solutions that satisfy all constraints. Pareto front analysis is then performed on the feasible solutions to obtain the optimal construction scheme.

8. The parametric simulation design method for carbon dioxide pipeline construction and installation as described in claim 1, characterized in that, The method further includes: After the optimal construction plan is generated, a pre-trained construction risk prediction model is called to perform risk assessment on the optimal construction plan to obtain the risk level and risk triggering conditions of each construction procedure. Based on the risk level, the generative model is invoked to generate risk mitigation questions and answers for high-risk processes, guiding users to input risk handling parameters. The risk management parameters are integrated into the optimal construction plan to generate a risk-controllable construction plan that includes emergency response measures.

9. The parametric simulation design method for carbon dioxide pipeline construction and installation as described in claim 8, characterized in that, The process involves calling a pre-trained construction risk prediction model to perform risk assessment on the optimal construction plan, obtaining the risk level and risk triggering conditions for each construction procedure, including: Extract the temporal logical relationships and construction environment parameters of each construction procedure from the optimal construction plan, and construct a Bayesian risk network for the construction process; The Bayesian risk network is input into the construction risk prediction model, and the risk propagation probability calculation based on graph convolution is performed to obtain the failure probability of each construction node. Based on the failure probability and the preset risk level threshold, the risk level of each construction procedure is determined, and the risk triggering conditions are extracted according to the conditional probability table in the Bayesian risk network.

10. The parametric simulation design method for carbon dioxide pipeline construction and installation as described in claim 8, characterized in that, The process of integrating the risk management parameters into the optimal construction plan to generate a risk-controllable construction plan that includes emergency response measures includes: The emergency response type and resource allocation quantity in the risk management parameters are analyzed, and the emergency response type is mapped to a preset emergency operation process template. The emergency operation process template is spatiotemporally linked with the construction procedures in the optimal construction plan to generate the trigger nodes and response sequence of the emergency plan. The resource allocation quantity is superimposed and merged with the material list of the construction process to generate a construction resource master table including emergency materials; The emergency plan and the construction resource summary table are embedded into the data structure of the optimal construction scheme to generate the risk-controllable construction scheme.

Images