Method and system for establishing solid waste high-value nuclear engineering concrete material knowledge base

Abstract

The application relates to a solid waste high-value nuclear engineering concrete material knowledge base establishment method and system, solves the problem that a research and development process is highly dependent on experience-based fixed mixing amount trial and error, which leads to a long research and development cycle and difficulty in adapting to SMR rapid construction requirements, and the method comprises the following steps: a nuclear engineering concrete material database is constructed by integrating double-source data and standardizing processing; a performance mapping model is established by adopting machine learning to analyze component-process-performance structure-effect relationships; a plurality of optimization schemes are generated by inversely deducing according to target performance, and the optimal solution is output after verification and scoring, and finally a multi-layer knowledge base containing data, models and schemes is constructed. The application has the following effects: a leap from experience-based trial and error to data-driven intelligent generation of a nuclear engineering concrete material formula is realized, and precise and efficient optimal design is realized.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of nuclear power engineering, in particular to a method and system for establishing a solid waste high-value nuclear engineering concrete material knowledge base. BACKGROUND

[0002] Global nuclear power is accelerating the transformation to small modular reactors (SMR), and nuclear engineering concrete, as a core structural material of nuclear facilities, accounts for more than 30% of the total carbon emissions of nuclear power in the whole life cycle, which is in conflict with the environmental friendly goal of SMR. Using industrial solid waste such as steel slag and fly ash for the preparation of nuclear engineering concrete can not only reduce the pollution of solid waste storage and the consumption of natural raw materials, but also take into account the radiation shielding performance, which has become a key path to promote the greenization of nuclear engineering.

[0003] The existing solid waste-based nuclear engineering concrete technology is mostly focused on "solid waste replacing traditional components", such as using steel slag to replace fine aggregate, mixing waste glass in barite concrete, or using iron tailings to replace coarse aggregate and mixing fly ash as cementitious materials, to realize the resource utilization of solid waste and the optimization of concrete performance.

[0004] The most core limitation of the existing technology is that the research and development process highly depends on the experience-based fixed dosage trial and error, and lacks systematic data integration and quantitative model guidance on the correlation of solid waste composition, ratio, process and performance, which leads to not only a long research and development cycle that is difficult to adapt to the rapid construction demand of SMR, but also an inability to precisely control the proportion of solid waste and performance stability, making it difficult to break through the conventional upper limit of 25% of solid waste, and easily affecting the radiation shielding and strength reliability of nuclear engineering concrete due to parameter matching deviation. SUMMARY

[0005] In order to realize the leap from experience-based trial and error to data-driven intelligent generation of nuclear engineering concrete material formula, and realize precise and efficient optimal design, the application provides a method and system for establishing a solid waste high-value nuclear engineering concrete material knowledge base.

[0006] In the first aspect, the application provides a method for establishing a solid waste high-value nuclear engineering concrete material knowledge base, which adopts the following technical scheme:

[0007] A method for establishing a solid waste high-value nuclear engineering concrete material knowledge base, comprising:

[0008] External data is collected through a pre-set standardized retrieval term library and an intelligent retrieval engine, internal data is collected synchronously, and double-source heterogeneous data integration and classification are completed;

[0009] The data after integration and classification is structurally cleaned based on a pre-set tree-shaped four-level parameter architecture, classified according to parameter categories, subcategories and index levels, abnormal values are removed and data units are unified, and a standardized data set is formed;

[0010] The standardized data set is imported into a preset database framework to build an initial database, a preset machine learning model is used to learn the correlation between components, processes and performances, a supplementary sample is generated to realize data expansion, and a perfect database is obtained;

[0011] The composition-ratio-process-performance structure-activity relationship in the perfect database is mined through a preset gene encoding model, and the quantitative structure-activity relationship between performance and solid waste composition, ratio and process is established;

[0012] The target performance requirement of the nuclear engineering concrete is input, and based on the above-mentioned quantitative structure-activity relationship and preset derivation constraints, multiple sets of initial parameter combinations are automatically generated and reverse derivation is performed to generate multiple initial optimization schemes;

[0013] Each initial optimization scheme is substituted into the quantitative structure-activity relationship to verify the performance compliance: if it is compliant, it is included in the candidate optimization scheme set; if it is not compliant, the parameters are corrected according to the preset adjustment rule and then re-derived until a preset number of candidate schemes are generated;

[0014] Based on a preset screening index system, a preset weighted scoring model is used to score the candidate scheme set, and the highest scoring scheme is selected as the optimal scheme;

[0015] The perfect database, the quantitative structure-activity relationship, the candidate scheme set, the optimal scheme and the iteration process data are stored in association according to a preset knowledge graph structure to build a knowledge base including a data layer, a model layer and a scheme layer.

[0016] In the second aspect, the present application provides a solid waste high-value nuclear engineering concrete material knowledge base establishment system, which adopts the following technical scheme:

[0017] A solid waste high-value nuclear engineering concrete material knowledge base establishment system includes a memory, a processor and a program stored on the memory and executable on the processor. The program can be loaded and executed by the processor to implement the solid waste high-value nuclear engineering concrete material knowledge base establishment method of the first aspect. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 It is the overall flowchart of the solid waste high-value nuclear engineering concrete material knowledge base establishment method of the embodiment of the present application. DETAILED DESCRIPTION

[0019] The present application will be further described in detail below with reference to the accompanying drawings.

[0020] REFERENCE Figure 1 A solid waste high-value nuclear engineering concrete material knowledge base establishment method disclosed by the present application includes:

[0021] Step S100, through the preset standardized retrieval term library and intelligent retrieval engine, external data is collected, internal data is synchronously collected, and double-source heterogeneous data integration classification is completed.

[0022] The database constructed by the application and all data processing processes thereof are based on public academic literature, patent data, and material experiment and production data within the field. These data do not involve any personal identity information or sensitive personal information, and the collection, storage and use thereof strictly comply with relevant regulations and will not violate social morality or harm public interests.

[0023] External data: refers to data obtained through channels such as public literature, patents, industry reports, etc., usually stored in formats such as PDF, Excel, etc. These data sources are extensive, covering research results and practical experience within the industry. Internal data: refers to data collected directly in the experiment and production process, usually stored in formats such as Origin, Word, etc. These data include solid waste material parameters (such as element abundance), process parameters (such as stirring rate), and characterization data (such as gamma-ray attenuation coefficient).

[0024] Standardized retrieval term library: a retrieval word library constructed based on national standards and industry terms, used for accurate collection of external data. For example, a keyword system with “GB / T5072-2022 Nuclear Power Concrete” as the root node is constructed, and specific parameters such as “steel slag-moisture content” and “fly ash-ignition loss” are expanded.

[0025] Intelligent retrieval engine: a retrieval tool combining Boolean logic and semantic association, capable of automatically identifying and supplementing synonyms, improving retrieval efficiency and accuracy. For example, Python-Whoosh combined with cosine similarity algorithm is used to realize joint retrieval of “steel slag OR converter slag”.

[0026] The acquisition method and process are described as follows:

[0027] 1. External data collection: 1.1, input retrieval formula: combine industry standards and specific needs to construct a retrieval formula, such as “(“small modular reactor” OR “nuclear containment”) AND (“solid waste” OR “steel slag” OR “fly ash”) AND (“gamma attenuation coefficient” OR “shielding”)”. 1.2, retrieval engine: use Python-Whoosh combined with semantic association algorithm to automatically parse key parameters in literature, patents and industry reports, and output standardized Excel templates. For example, incrementally grab data from CNKI, USPTO, IAEA-TECDOC, etc. databases every day, parse DOI, abstract and key parameters. 1.3, data formatting: format the collected external data into Excel tables, align the fields with the four-level parameter architecture of internal data, and ensure the consistency and operability of the data.

[0028] 2. Internal data collection: 2.1. Data sources: Collect data from experiments and production sites, including solid waste material properties, process parameters, and characterization data. For example, upload stirring rate and temperature data every 30 seconds through PLC, automatically return XRF element abundance and gamma-ray spectrometer attenuation values from laboratory LIMS. 2.2. Data storage: Store collected data in a local MongoDB database for subsequent processing and analysis.

[0029] 3. Dual-source data alignment: 3.1. Primary key matching: Use "solid waste category + chemical composition" as the joint primary key, and use the Levenshtein distance algorithm for fuzzy matching to combine external data and internal data into the same feature record. For example, combine "steel slag - MnO 8.2%" in external data with "steel slag - MnO 8.0%" in internal data. 3.2. Quality preliminary screening: Run box plot rules on the combined data to remove outliers. For example, if the gamma attenuation coefficient exceeds Q3 + 1.5IQR, it is marked as an outlier and uniformly processed in subsequent steps.

[0030] Step S200, based on the preset tree four-level parameter architecture, the integrated classified data is structured and cleaned, classified according to parameter categories, subcategories and index levels, outliers are removed and data units are unified to form a standardized data set.

[0031] Tree four-level parameter architecture: a hierarchical data classification method, which divides data into solid waste materials, physical properties, chemical properties, process parameters, etc. Each category is further divided into subcategories and specific indicators. For example, solid waste materials are divided into subcategories such as steel slag and fly ash, and steel slag is further divided into physical properties (such as moisture content) and chemical properties (such as MnO content). Structured cleaning: standardize the collected data, including unifying units, removing outliers, filling missing values, etc., to make the data meet the preset format and quality requirements. Feature deconstruction: extract useful feature information for model construction from raw data, such as extracting key parameters affecting concrete performance from solid waste material physical and chemical properties.

[0032] Acquisition method and process:

[0033] 1. Data classification and hierarchy construction: 1.1. Classification basis: Based on the types and properties of solid waste materials, combined with the actual needs of nuclear power engineering concrete, a tree-shaped four-level parameter framework is constructed. For example, solid waste materials are divided into steel slag, fly ash, iron tailings, etc.; physical properties include moisture content, density, etc.; chemical properties include element abundance (such as MnO, SiO2 content), etc. 1.2. Hierarchy construction: The data is divided into four levels: solid waste materials (major category), physical properties / chemical properties (sub-category), specific indicators (such as moisture content, MnO content), and data values. For example, steel slag (major category) → physical properties (sub-category) → moisture content (indicator) → 6.2% (data value).

[0034] 2. Structured cleaning: 2.1. Unit unification: unify the units of all data to standard units, such as density to g / cm³ and moisture content to percentage (%). For example, convert density data from kg / m³ to g / cm³ from different sources. 2.2. Outlier removal: use the box plot method to detect and remove outliers. For example, for the loss on ignition data, if the loss on ignition of a batch of steel slag exceeds 5%, it is marked as an outlier and removed. 2.3. Missing value processing: for missing data, use mean or median to fill in. For example, if the loss on ignition data of a batch of fly ash is missing, use the median of known data to fill in.

[0035] 3. Feature deconstruction: 3.1. Key feature extraction: extract key features from the physical and chemical properties of solid waste materials that have a significant impact on concrete performance. For example, extract moisture content, MnO content, etc. from steel slag, and extract loss on ignition, SiO2 content, etc. from fly ash. 3.2. Feature correlation: establish the correlation between features, such as the correlation between moisture content and compressive strength, and the correlation between MnO content and gamma-ray attenuation coefficient. For example, through correlation analysis, it is found that for every 1% increase in moisture content of steel slag, the compressive strength decreases by 2%.

[0036] Step S300, import the standardized data set into the preset database framework to build an initial database, use a preset machine learning model to learn the correlation between composition-process-performance, generate supplementary samples to realize data expansion, and obtain a perfect database.

[0037] Wherein, the initial database: a preliminary database constructed based on the standardized data set, containing information of physical properties, chemical properties, process parameters of solid waste materials. Machine learning model: used to learn the correlation between solid waste material composition, process parameters and concrete performance, and generate supplementary samples to expand the data set. Data expansion: generate new sample data through machine learning model to increase the diversity and coverage of data set and improve the generalization ability of model. The specific process of the above steps can be referred to steps S310 to S340, which will not be repeated here.

[0038] Step S400, by presetting the gene encoding model to mine the relationship between solid waste composition, ratio, process and performance in the perfect database, and to establish the quantitative structure-activity relationship between performance and solid waste composition, ratio and process.

[0039] Gene encoding model: a mathematical model used to analyze the structure-activity relationship between solid waste composition, ratio, process parameters and concrete performance. It establishes the quantitative structure-activity relationship between input parameters (solid waste composition, ratio, process) and output performance (such as compressive strength, gamma-ray attenuation coefficient) through quantitative analysis. Structure-activity relationship: the law describing the interaction and influence between solid waste composition, ratio, process parameters and concrete performance. For example, how the MnO content of steel slag affects the compressive strength of concrete, how the SiO2 content of fly ash affects the gamma-ray attenuation coefficient, etc.

[0040] The specific process of the above steps can be referred to steps S410 to S450, which will not be repeated here.

[0041] Step S500, input the target performance requirements of nuclear engineering concrete, based on the above quantitative structure-activity relationship and preset derivation constraints, automatically generate multiple initial parameter combinations and perform reverse derivation to generate multiple initial optimization schemes.

[0042] Among them, the target performance requirement: according to the specific application scene of nuclear engineering concrete, the performance index of concrete is set, such as compressive strength, gamma-ray attenuation coefficient, freeze-thaw cycle resistance, etc. Reverse derivation: based on the target performance requirement, using the established gene encoding model, from the performance index to derive the solid waste composition, ratio and process parameters that meet these performances. Initial optimization scheme: the preliminary scheme generated by reverse derivation, which needs to be further verified and optimized to meet the actual application requirements. The specific process of the above steps can be referred to steps S510 to S550, which will not be repeated here.

[0043] Step S600, substitute each initial optimization scheme into the quantitative structure-activity relationship to verify the performance compliance: if it meets the standard, it is included in the candidate optimization scheme set; if it does not meet the standard, the parameters are corrected according to the preset adjustment rules and then re-derived until a preset number of candidate schemes are generated.

[0044] Among them, performance verification: through experimental or simulation means, verify whether the initial optimization scheme meets the target performance requirement. Quantitative structure-activity relationship is the quantitative correlation between solid waste composition, ratio, process parameters and concrete performance indicators determined after analyzing and mining the improved database through the preset gene coding model, a mathematical model. It mathematically represents the interaction between parameters and performance, so it is computable. "Substitute quantitative structure-activity relationship verification" technically embodies: according to the quantitative correlation, calculate the specific parameter combination of any scheme, and thus obtain the corresponding performance index estimate value.

[0045] The specific process is as follows:

[0046] 1. Performance prediction based on quantitative structure-activity relationship: calculate the specific parameters (solid waste composition, ratio, process parameters) of each initial optimization scheme according to the quantitative structure-activity relationship, and obtain the estimated value of its key performance indicators (such as compressive strength, gamma ray attenuation coefficient).

[0047] 2. Empirical verification of experiment and / or simulation:

[0048] Experimental verification: for the scheme that meets the prediction, prepare concrete test pieces according to its parameters, and measure the performance indicators. For example, the compressive strength of a scheme test piece is 47MPa, and the gamma ray attenuation coefficient is 0.53 / cm.

[0049] Simulation verification: numerical simulation methods such as finite element analysis can be used to predict and verify the performance of the scheme. For example, the software simulation gets the predicted value of the compressive strength of a scheme as 46MPa, and the predicted value of the gamma ray attenuation coefficient as 0.54 / cm.

[0050] 3. Comprehensive compliance judgment: compare the performance estimate value based on the quantitative structure-activity relationship, the experimental test results and / or the simulation prediction results with the target performance requirement (for example: compressive strength ≥45MPa, gamma ray attenuation coefficient ≥0.55 / cm). If the scheme is verified by one or more of the above methods, and the results all meet or exceed the target threshold, it is determined that the scheme meets the standard and is directly included in the candidate optimization scheme set; otherwise, it is determined as not meeting the standard and enters the parameter adjustment stage.

[0051] 4. Parameter adjustment and recalculation verification:

[0052] Adjustment rule application: according to the preset adjustment rule for the non-compliant performance indicators, modify the specific parameters of the scheme. For example, if the gamma ray attenuation coefficient does not meet the standard, increase the dosage of barite powder according to the rule.

[0053] Adjustment method: optimization can be carried out by step-by-step adjustment method (adjusting one parameter at a time) or comprehensive adjustment method (adjusting multiple parameters at the same time).

[0054] Recalculation and verification: The adjusted new parameter combination is recalculated based on the quantitative structure-property relationship to obtain a new performance estimate. The empirical verification and comprehensive judgment process in steps 2 to 3 is repeated until the performance meets the standard.

[0055] 5. Generate a candidate optimization scheme set: Summarize all validated compliant schemes (including those that initially meet the standards and those that have been adjusted to meet the standards) to form a candidate optimization scheme set. Completely record the solid waste composition, proportions, process parameters, and corresponding performance estimates, experimental and / or simulation validation data for each candidate scheme.

[0056] Step S700: Based on the preset screening index system, the candidate solution set is scored using a preset weighted scoring model, and the solution with the highest score is selected as the optimal solution.

[0057] The weighted scoring model is a mathematical model that calculates the comprehensive score of candidate solutions by assigning weights to each scoring indicator. The comprehensive score reflects the overall performance and applicability of the solution. The optimal solution is the one with the highest comprehensive score among all candidate solutions, which is considered to best meet the performance requirements of nuclear engineering concrete.

[0058] In step S700, we comprehensively evaluate the candidate optimization schemes based on a comprehensive scoring index system to determine the optimal scheme. The specific process is as follows: First, a comprehensive scoring index system is established based on the performance requirements and actual application conditions of nuclear engineering concrete. For example, compressive strength has a weight of 30%, gamma-ray attenuation coefficient has a weight of 25%, cost has a weight of 20%, durability has a weight of 15%, and construction convenience has a weight of 10%. These weights are determined based on actual engineering needs and expert experience, ensuring that the importance of each index is reasonably reflected in the overall evaluation. Next, the performance index data of each candidate scheme are standardized to eliminate the influence of different dimensions and orders of magnitude. For example, the formula: Standardized Score = (Actual Value - Minimum Value) ÷ (Maximum Value - Minimum Value) is used to convert indices such as compressive strength and gamma-ray attenuation coefficient into standardized scores. This process ensures the comparability between different indices and provides a basis for subsequent weighted scoring.

[0059] Then, the standardized performance indicators of each candidate solution are weighted and summed according to the preset weights to calculate the comprehensive score. For example, if the compressive strength standardized score of a certain solution is 0.9, the gamma-ray attenuation coefficient standardized score is 0.85, the cost standardized score is 0.7, the durability standardized score is 0.8, and the construction convenience standardized score is 0.9, then the comprehensive score is 0.3 x 0.9 + 0.25 x 0.85 + 0.2 x 0.7 + 0.15 x 0.8 + 0.1 x 0.9 = 0.8225. This calculation process ensures that the comprehensive performance of each solution is fully evaluated.

[0060] In step S800, the perfect database, quantitative structure-activity relationship, candidate solution set, optimal solution, and iteration process data are stored in association according to the preset knowledge graph structure, and a knowledge base containing data layer, model layer, and solution layer is constructed.

[0061] Among them, the knowledge graph: a structured semantic knowledge base, describes the relationship between entities (such as solid waste materials and performance indicators) through a graph structure, facilitating information retrieval and knowledge discovery. Associated storage: different types of data and models are organized and stored according to the preset knowledge graph structure to ensure data consistency and traceability.

[0062] The specific process is as follows: First, the verified and optimized database, structure-activity relationship, candidate solution set, and optimal solution data are systematically integrated. For example, the database expanded in step S300, the structure-activity relationship established in step S400, the candidate solution set generated in step S600, and the optimal solution selected in step S700 are summarized to form a comprehensive data set. This process ensures the integrity and consistency of all related data, providing a solid foundation for subsequent knowledge base construction. Next, according to the preset knowledge graph structure, these data are stored in association. For example, the physical properties, chemical properties, and process parameters of solid waste materials are associated with the corresponding concrete performance indicators to form a structured knowledge graph. Through the knowledge graph, the relationship between the composition, ratio, and process parameters of solid waste materials and the performance of concrete can be clearly displayed, facilitating users to quickly retrieve and understand the internal relationship between data. Finally, the constructed knowledge base and knowledge graph are stored and managed. For example, use professional database management systems (such as MySQL or MongoDB) and knowledge graph tools (such as Neo4j) for storage and management to ensure data security and accessibility. At the same time, establish a data update and maintenance mechanism to regularly update and optimize the knowledge base to adapt to new research progress and actual application needs.

[0063] The completion of the integration and classification of double-source heterogeneous data includes:

[0064] Step S110, the collected external data is standardized in format, key entities in unstructured text are extracted using natural language processing algorithms, and the data representation format is unified through regular expressions. Key entities include solid waste types, performance indicators, and process descriptions.

[0065] The specific process is as follows: First, the collected external data is standardized in format. The PyMuPDF library of Python is used to extract text content from PDF documents, and the pandas library is used to convert it into a structured Excel table. For example, performance descriptions of steel slag and fly ash are extracted from the literature and converted into table form for subsequent processing and analysis. Next, natural language processing algorithms are used to extract key entities from unstructured text. Through the Named Entity Recognition (NER) algorithm, key information such as solid waste types, performance indicators, and process descriptions is identified from literature and patent text. For example, "steel slag" is extracted from the text as a solid waste type, "compressive strength 45MPa" as a performance indicator, and "mixing rate 100rpm" as a process description. Then, regular expressions are used to unify the data representation format. For example, regular expressions are used to unify different representations of performance indicators into a standard format, such as "compressive strength 45MPa" or "compressive strength of 45MPa" into "compressive strength: 45MPa". This process ensures the consistency of data representation, facilitating subsequent data processing and analysis. Finally, the extracted key entities are matched and stored with the pre-set fields. For example, the extracted key entities "steel slag", "compressive strength: 45MPa", "mixing rate: 100rpm" are stored in the pre-set fields to form structured data records.

[0066] Step S120, the collected internal data is structured and converted, discrete data in experimental records and production logs is reorganized according to the pre-set fields of solid waste name-detection date-property parameter-process parameter-characterization data, and the reorganized data is preliminarily screened using the Z-score method to eliminate invalid data deviating from the pre-set normal distribution range.

[0067] The specific process is as follows: first, the internal data collected is structured and converted. The discrete data in the experimental records and production logs is reorganized according to the preset fields of solid waste name, detection date, physical property parameter, process parameter and characterization data. For example, the water content of steel slag, loss on ignition of fly ash and other data in the experimental records are reorganized according to the format of solid waste name-detection date-physical property parameter-process parameter-characterization data to form a structured data table. This process ensures the uniformity of the format of internal data, facilitating subsequent processing and analysis. Next, the Z-score method is used to preliminarily screen the reorganized data. By calculating the Z-score value of each data point, it is judged whether it deviates from the preset normal distribution range. For example, for the water content of steel slag, calculate its Z-score value, if the Z-score value of a certain data point exceeds 3 (usually considered as the threshold value of abnormal value), mark it as invalid data and eliminate it. This process ensures the accuracy and consistency of the data, providing a reliable data basis for subsequent data processing and analysis. Finally, the data after preliminary screening is stored in the local database for subsequent query and use. For example, the screened data is stored in the local MongoDB database to ensure the long-term availability and reliability of the data. At the same time, a data updating and maintaining mechanism is established to regularly check and update the data to ensure the timeliness and accuracy of the data.

[0068] Step S130, the cosine similarity algorithm is used to calculate the feature matching degree of the key entities extracted from the external data and the preset fields of the internal data, and a bidirectional association mapping of literature performance index-experimental detection performance data and patent process description-production process parameter is established.

[0069] Among them, the bidirectional association mapping: establishes the mapping relationship between the key entities extracted from the external data and the preset fields of the internal data, ensuring that the information between the two can correspond and associate with each other.

[0070] The specific process is as follows: first, the external data key entities extracted in step S110 and the structured converted internal data fields in step S120 are preprocessed. For example, the "steel slag compressive strength 45 MPa" in the external data and the "steel slag-compressive strength-45 MPa" in the internal data are text cleaned and standardized to ensure consistency in format. This process provides a basis for subsequent feature matching. Next, the cosine similarity algorithm is used to calculate the similarity between the external data key entities and the internal data fields. For example, the cosine similarity between "steel slag compressive strength 45 MPa" in the external data and "steel slag-compressive strength-45 MPa" in the internal data is calculated. By converting the text into vector form and calculating the cosine value of the angle between these vectors, a similarity score is obtained. If the similarity score exceeds the preset threshold (such as 0.8), it is considered that the two are successfully matched. Then, based on the matching results, a bidirectional association mapping between the literature mentioned performance indicators and the experimental detection performance data, and the patent process description and the production process parameters is established. For example, the "steel slag compressive strength 45 MPa" mentioned in the literature is associated with "steel slag-compressive strength-45 MPa" in the experimental record, and at the same time, "stirring rate 100 rpm" in the patent is associated with "stirring rate-100 rpm" in the production log.

[0071] Step S140, according to the association mapping results, the double-source data is classified according to the solid waste type and the nuclear engineering concrete application scenario, the redundant data with a feature matching degree lower than the preset feature matching degree is removed, and finally a classified and effective double-source integrated data set is formed.

[0072] Feature matching degree: calculated by the cosine similarity algorithm, indicating the similarity between the key entities extracted from the external data and the preset fields of the internal data.

[0073] The specific process is as follows: first, according to the bidirectional association mapping result established in step S130, the external data and the internal data are classified according to the solid waste type. For example, all data related to "steel slag" is classified as a class, and data related to "fly ash" is classified as another class. This process ensures clear classification of data, facilitating subsequent analysis and application. Next, the classified data is further subdivided according to the nuclear engineering concrete application scenario. For example, steel slag data and fly ash data related to "nuclear containment construction" are classified separately to ensure the relevance of data to specific application scenarios. This process further refines data classification, improving the relevance and practicality of data. Then, the classified data is evaluated for feature matching degree. According to the pre-set feature matching degree threshold (such as 0.8), redundant data with a feature matching degree lower than the threshold is removed. For example, if the feature matching degree of a certain data is 0.7, which is lower than the pre-set threshold, it is marked as redundant data and removed. This process ensures the effectiveness and consistency of the data, reducing the interference of redundant data on subsequent analysis. Finally, the data after classification and removal of redundant data is integrated to form a dual-source integrated data set. For example, all valid steel slag data and fly ash data are integrated into a data set to ensure data integrity and consistency.

[0074] When the collected internal data is structured, it also includes blockchain traceability processing, including:

[0075] Step S121: The discrete data in the experimental records and production logs is reorganized according to the pre-set fields of solid waste name-blockchain traceability code-detection date-property parameters-process parameters-characterization data, and a blockchain traceability code is added to the original fields.

[0076] Among them, the blockchain traceability code: a unique identifier, used to record and track the whole link data of solid waste from the production place to the use process, to ensure the traceability and non-tamperability of the data.

[0077] The specific process is as follows: first, the discrete data in the experimental records and production logs is reorganized according to the preset fields of solid waste name, blockchain traceability code, detection date, physical property parameters, process parameters, and characterization data. A blockchain traceability code field is added based on the original fields to ensure that each data record is associated with a unique blockchain traceability code. For example, the data of water content in steel slag, loss on ignition of fly ash, etc. in the experimental records is reorganized according to the format of solid waste name-blockchain traceability code-detection date-physical property parameters-process parameters-characterization data to form a structured data table. Next, the blockchain traceability code is associated with the solid waste full-link data, including the origin identification, transportation process information, detection raw data, and storage information. The full-link data is encrypted and stored by the smart contract to ensure the security and tamper resistance of the data. For example, for a batch of steel slag, the blockchain traceability code is associated with the origin identification (such as "steel slag mine A"), transportation process information (such as "transportation vehicle B, transportation time C"), detection raw data (such as "water content 6.2%"), and storage information (such as "storage time D, storage location E"). These information is encrypted and stored by the smart contract to ensure the authenticity and integrity of the data. Finally, the reorganized data is stored in the local database for subsequent query and use. For example, the reorganized data is stored in the local MongoDB database to ensure the long-term availability and reliability of the data.

[0078] In step S122, the blockchain traceability code is associated with the solid waste full-link data, including the origin identification, transportation process information, detection raw data, and storage information. The full-link data is encrypted and stored by the smart contract to ensure the tamper resistance.

[0079] The specific process is as follows: first, the blockchain traceability code is associated with the full-link data of solid waste. The full-link data includes the origin source identification of solid waste, transportation process information, detection raw data and storage information. For example, for a batch of steel slag, its blockchain traceability code is associated with the origin identification (such as "steel slag mine A"), transportation process information (such as "transportation vehicle B, transportation time C"), detection raw data (such as "moisture content 6.2%") and storage information (such as "storage time D, storage location E"). These information is recorded through blockchain technology, ensuring that the data of each link can be traced and verified. Then, the full-link data is encrypted and stored by smart contract. Smart contract is a kind of automatic execution contract clause, which realizes the encryption and storage of data through blockchain technology, ensuring the security and non-tamperability of data. For example, using Ethereum smart contract platform, the full-link data of solid waste is written into smart contract, once the data is written, it cannot be tampered with. Finally, the encrypted and stored data is stored on the blockchain, ensuring the long-term availability and traceability of the data. For example, the full-link data of solid waste is stored on the Ethereum blockchain, each data record has a unique blockchain hash value, which can quickly trace and verify the authenticity of the data.

[0080] In step S123, when the recombined data is preliminarily screened by Z-score method, the internal data without blockchain traceability code or incomplete traceability data is simultaneously removed, and only the structured data with complete traceability information is retained.

[0081] The specific process is as follows: first, the Z-score method is used to preliminarily screen the recombined internal data. Z-score method is a statistical method used to detect outliers in data. By calculating the Z-score value of each data point, it is judged whether it deviates from the preset normal distribution range. For example, for the water content data of steel slag, calculate its Z-score value, if the Z-score value of a certain data point exceeds 3 (usually considered as the threshold value of outliers), mark it as invalid data and eliminate it. This process ensures the accuracy and consistency of the data, providing a reliable data basis for subsequent data processing and analysis. Then, the integrity of the blockchain traceability information is verified synchronously. In the preliminary screening process, check whether each data record contains valid blockchain traceability code and ensure the integrity of the traceability data. For example, check whether the blockchain traceability code is associated with complete solid waste full-link data, including origin source identification, transportation process information, detection original data and warehouse information. If a certain data record lacks blockchain traceability code or traceability data is incomplete, it will be marked as invalid data and eliminated. Finally, only the structured data with complete traceability information is retained. The data that has passed the preliminary screening and traceability integrity verification is stored in the local database for subsequent query and use. For example, the screened data is stored in the local MongoDB database to ensure the long-term availability and reliability of the data.

[0082] The standardized data set is imported into the preset database framework to build an initial database, a preset machine learning model is used to learn the correlation between composition, process and performance, and a supplementary sample is generated to realize data expansion, and a perfect database is obtained, including:

[0083] In step S310, the standardized data set is classified according to four-level parameters of solid waste attributes, production process, material performance and service environment, and is imported into a preset relational database framework; the framework defines the field type, dimension unit and cross-parameter association rule of each parameter through a data dictionary, and realizes fast data calling through an index structure, so as to build an initial database with standard structure that can be directly used for model training.

[0084] The specific process is as follows: first, the standardized data set is classified according to the four-level parameters of solid waste properties, production process, material performance and service environment. For example, the physical properties (such as moisture content), chemical properties (such as MnO content), process parameters (such as stirring rate), material performance (such as compressive strength) and service environment (such as temperature, humidity) of solid waste materials such as steel slag and fly ash in the data set are classified. Then, the classified data is imported into the preset relational database framework. The framework defines the field type, dimension unit and cross-parameter association rule of each parameter through the data dictionary. For example, the data dictionary defines the "moisture content" field as a floating point type, with a unit of percentage (%), and specifies its association with the solid waste material name. Through the index structure, the data can be quickly called, ensuring efficient storage and retrieval of data. For example, an index is established for the "solid waste name" field to quickly query all data records related to a specific solid waste. Finally, an initial database with a structured specification that can be directly used for model training is constructed. Through the above classification and import process, the data structure in the initial database is clear, the field type is clear, the dimension unit is consistent, and the association between parameters is clear. For example, each record in the initial database contains the fields of solid waste name, moisture content, MnO content, stirring rate, compressive strength, etc., and the association between these fields is defined by the data dictionary and index structure.

[0085] In step S320, the data in the initial database is used as a training sample, a BP neural network is used as a preset machine learning model, the solid waste composition parameters and process parameters are used as the input layer, the material performance parameters are used as the output layer, the network weight is iteratively adjusted through the error back propagation algorithm to learn the nonlinear correlation between composition-process-performance, and the model is formed until the prediction error of the validation set is lower than the preset threshold, and the trained model is formed which can accurately map the parameter correlation.

[0086] The BP neural network model includes an input layer, at least one hidden layer and an output layer. The number of input layer nodes corresponds to the number of selected solid waste composition parameters and process parameters (for example, steel slag MnO content, fly ash SiO2 content, stirring rate, etc., a total of n). The number of output layer nodes corresponds to the number of material performance parameters to be predicted (for example, compressive strength, gamma-ray attenuation coefficient, etc., a total of m). The number of layers and the number of nodes of the hidden layer can be determined by conventional cross-validation according to the size of the initial database, for example, two hidden layers are set, and the number of nodes is 64 and 32 respectively. When training the model, the mean square error (MSE) is used as the loss function, the Adam optimizer is used, the learning rate is set to 0.001, the batch size is set to 32, and the number of iterations (epochs) is set until the validation set error no longer decreases significantly.

[0087] The specific process is as follows: first, the data in the initial database is divided into a training set and a validation set. For example, 80% of the data is used as the training set, and 20% of the data is used as the validation set. The training set is used to train the BP neural network model, and the validation set is used to evaluate the prediction performance of the model. Next, the BP neural network model is constructed. The solid waste composition parameters (such as the MnO content of steel slag and the SiO2 content of fly ash) and process parameters (such as stirring rate and curing temperature) are used as the input layer, and the material performance parameters (such as compressive strength and gamma-ray attenuation coefficient) are used as the output layer. For example, the input layer includes steel slag MnO content, fly ash SiO2 content, stirring rate, etc., and the output layer includes compressive strength and gamma-ray attenuation coefficient. Then, the network weights are iteratively adjusted through the error backpropagation algorithm. During training, the BP neural network calculates the predicted value through forward propagation, then calculates the error between the predicted value and the actual value through the error backpropagation algorithm, and adjusts the network weights according to the error. For example, the mean square error (MSE) is used as the loss function, and the weights are adjusted through gradient descent to minimize the prediction error. Finally, the model is continuously trained until the prediction error of the model on the validation set is below the preset threshold. For example, set the prediction error threshold to 0.05, and when the prediction error of the model on the validation set is below 0.05, consider that the model training is complete. At this time, the model can accurately map the nonlinear correlation between solid waste composition parameters, process parameters and material performance parameters, forming a trained model.

[0088] In step S330, based on the correlation captured by the trained model and the sample distribution characteristics of the initial database, a generative adversarial network algorithm is used to generate supplementary samples: a generator is trained based on real samples from the initial database to generate virtual samples that conform to the composition-process-performance correlation; at the same time, a discriminator is used to verify the consistency of virtual samples and real samples, and only virtual samples that are judged to be real samples with a probability reaching a preset threshold are temporarily stored as candidate supplementary samples.

[0089] In the above process, the generator (Generator) of the generative adversarial network (GAN) adopts a fully connected neural network structure, and its input is a random noise vector (dimension 100) and its output is a simulated solid waste composition-process parameter vector. The discriminator (Discriminator) also adopts a fully connected neural network structure, which is used to distinguish whether the input parameter vector comes from the real database or the generator. During the adversarial training process, the generator and the discriminator are alternately optimized until the discriminator cannot effectively distinguish between real and fake samples (for example, the discrimination accuracy is close to 50%). After the training is stable, the virtual samples generated by the generator and judged by the discriminator to have a probability of being "real" exceeding a preset threshold (such as 0.9) are adopted as candidate supplementary samples.

[0090] The specific process is as follows: first, the nonlinear correlation between the waste composition parameters, process parameters and material performance parameters captured by the trained BP neural network model is captured. These laws will serve as prior knowledge for the generation of adversarial networks (GAN), ensuring that the generated supplementary samples conform to the learned parameter correlation laws. For example, the model has learned the relationship between the MnO content of steel slag and the compressive strength, as well as the effect of stirring rate on the gamma-ray attenuation coefficient. Next, a generative adversarial network (GAN) is constructed. GAN consists of two parts: a generator and a discriminator. The goal of the generator is to generate virtual samples that conform to the correlation between waste composition parameters, process parameters and material performance parameters, while the goal of the discriminator is to distinguish between generated virtual samples and real samples. Through the adversarial training of the two, the generator gradually generates more and more realistic samples. Then, the generator is trained with real samples from the initial database as a benchmark. The generator generates virtual samples that conform to the composition-process-performance correlation by learning the sample distribution characteristics in the initial database. For example, the generator generates new virtual samples based on the steel slag and fly ash samples in the initial database, which have similar distribution characteristics in composition, process and performance to the real samples. At the same time, the consistency of virtual samples and real samples is verified by the discriminator. The discriminator evaluates the generated virtual samples to determine whether they have high consistency with real samples. For example, the discriminator compares the feature distribution of virtual samples and real samples to determine the authenticity of virtual samples. Only when the authenticity of virtual samples reaches the preset accuracy (such as 90%) can they be temporarily stored as candidate supplementary samples.

[0091] Step S340, integrate the initial database and the candidate supplementary samples verified by the discriminator, use the K-means clustering algorithm for consistency verification, remove abnormal samples with clustering deviation values exceeding the preset range, and finally form a perfect database with balanced data distribution and significant correlation.

[0092] The specific process is as follows: first, the real samples in the initial database are integrated with the candidate supplementary samples verified by the discriminator. For example, the steel slag and fly ash samples in the initial database are combined with the generated virtual samples to form a larger data set. Next, the integrated data set is subjected to consistency verification using the K-means clustering algorithm. The K-means clustering algorithm divides the data points into K clusters, calculates the center point (centroid) of each cluster, and evaluates the deviation of each data point from the centroid. For example, the integrated data set is divided into several clusters, each cluster representing a group of data points with similar characteristics. Then, the deviation value of each data point from the centroid of the cluster to which it belongs is calculated. For example, the Euclidean distance is used to calculate the deviation of each data point from the centroid. A preset range (such as a deviation value not exceeding 10% of the centroid) is set, and data points with a deviation value exceeding the preset range are marked as abnormal samples. This process ensures the consistency of data points with cluster centroids and eliminates possible abnormal or outlier data points. Finally, the samples marked as abnormal are removed to form the final perfect database. For example, samples with a deviation value exceeding the preset range are removed from the data set, and data points with high consistency are retained. The final perfect database not only has balanced data distribution, but also has significant correlation rules.

[0093] By presetting the gene encoding model to mine the composition-ratio-process-performance structure-activity relationship of solid waste in the perfect database, the quantitative structure-activity relationship between performance and solid waste composition, ratio, and process is established, including:

[0094] Step S410, extract data with dynamic association tags from the perfect database, and construct a four-dimensional feature matrix according to the dimensions of solid waste composition-ratio parameter-process parameter-performance index, wherein the dynamic association tags record the source information of the data and the causal association information between parameters.

[0095] The specific process is as follows: first, data is extracted from the perfect database. These data not only contain the composition, proportioning and process parameters of solid waste, but also have dynamic correlation tags, which record the source and causal correlation information of the data. For example, the extracted data record may include the MnO content of steel slag, the SiO2 content of fly ash, the proportioning of steel slag and fly ash, the stirring rate, the curing temperature, and the corresponding performance indicators such as compressive strength and gamma-ray attenuation coefficient, and the dynamic correlation tag of "increasing the MnO content of steel slag leads to the improvement of compressive strength". Then, the extracted data is sorted and organized according to the four dimensions of solid waste composition, proportioning parameters, process parameters and performance indicators. For example, the MnO content of steel slag and the SiO2 content of fly ash are classified as the solid waste composition dimension, the proportioning of steel slag and fly ash is classified as the proportioning parameter dimension, the stirring rate and the curing temperature are classified as the process parameter dimension, and the compressive strength and the gamma-ray attenuation coefficient are classified as the performance indicator dimension. This process ensures the structuring and ordering of the data, providing a basis for subsequent feature matrix construction. Then, a four-dimensional feature matrix is constructed. The sorted data is arranged according to the above four dimensions to form a four-dimensional feature matrix. For example, each row of the feature matrix represents a data sample, and each column represents a feature dimension. In this way, the information of each data sample in the four dimensions of solid waste composition, proportioning parameters, process parameters and performance indicators is completely recorded.

[0096] Step S420, based on the four-dimensional feature matrix, enters the parameter coding stage: converts the solid waste composition, proportioning parameters and process parameters into standardized gene sequences, and embeds the causal correlation information in the dynamic correlation tags into the corresponding coding bits to form a gene coding data set with causal markers.

[0097] In this step, the gene sequence refers to the conversion of solid waste composition, proportioning parameters, process parameters and other parameters into standardized numerical sequences, similar to gene coding in biological genetics, used to represent the characteristics of different parameters. Dynamic correlation tag: a tag that records the causal relationship and correlation information between parameters in the data, used to trace and explain the impact of parameter changes on performance in subsequent steps. Causal marker: causal information embedded in the gene sequence, used to identify the specific impact of parameter changes on performance indicators, such as "increasing the MnO content of steel slag leads to the improvement of compressive strength".

[0098] The specific process is as follows: first, the solid waste components, the proportioning parameters, and the process parameters in the four-dimensional feature matrix are standardized. For example, the MnO content of steel slag and the SiO2 content of fly ash, and other component parameters, and the proportioning of steel slag and fly ash, the stirring rate, the curing temperature, and other process parameters are converted into standardized values between 0 and 1 through the normalization method. This process ensures the comparability and consistency between different parameters and provides a basis for subsequent coding work. Next, the standardized parameters are converted into gene sequences. For example, the standardized steel slag MnO content, fly ash SiO2 content, steel slag and fly ash proportioning, stirring rate, curing temperature, and other parameters are converted into gene sequences according to certain rules. Each parameter corresponds to a gene site, and each value in the gene sequence represents the characteristics of the corresponding parameter. This process converts complex parameter information into simple gene sequence form, facilitating subsequent processing and analysis. Then, the causal information is extracted from the dynamic correlation label and embedded into the corresponding gene coding site. For example, if the dynamic correlation label records the causal information that “increasing the MnO content of steel slag leads to an increase in compressive strength”, then in the gene sequence, the gene site corresponding to the MnO content of steel slag will be marked as “positively affecting compressive strength”. This process ensures that the gene sequence not only contains numerical information of parameters, but also contains causal influence information of parameter changes on performance indicators. Finally, the gene sequence with causal markers is integrated into a gene coding dataset. For example, the gene sequences of all samples and their corresponding causal markers are integrated into a dataset to form a gene coding dataset with causal markers.

[0099] In step S430, the encoded gene coding dataset is inputted to enter the relationship mining stage: the ridge regression algorithm is used to control parameter collinearity, and the linear correlation coefficient between a single parameter and a performance indicator is quantified; simultaneously, the random forest algorithm is used to mine the nonlinear relationship of multi-parameter interaction, the linear and nonlinear analysis results are verified and corrected based on the embedded causal correlation information, and the complete structure-activity relationship coefficient is outputted.

[0100] wherein "gene encoding" is the process of mapping continuous solid waste component parameters (such as content percentage), mixing ratio parameters (such as mixing amount ratio), process parameters (such as temperature, rate) into a fixed-length digital sequence through interval discretization and normalization, and each bit in the sequence represents the encoding of a parameter in a specific interval. Dynamic correlation labels (such as "parameter A is positively correlated with performance B") are converted into weight symbols or constraint conditions of the parameter encoding bits and embedded into subsequent regression and random forest analysis. When using ridge regression, the regularization coefficient a is determined by cross-validation in the interval 0.1, 100.1, 10. The number of decision trees in the random forest model is set to 100, and the maximum depth of the tree is dynamically adjusted according to the number of features. Ridge regression algorithm: a linear regression algorithm that controls the multicollinearity problem of parameters by adding an L2 regularization term to the loss function, thereby improving the stability and generalization ability of the model. QSAR coefficient: a coefficient that quantifies the relationship between solid waste components, mixing ratio parameters, process parameters, and material performance indicators, including linear correlation coefficients and nonlinear relationship coefficients.

[0101] The specific process is as follows: First, the encoded gene encoding dataset is input, and the ridge regression algorithm is used to control the parameter multicollinearity. Ridge regression effectively controls the multicollinearity problem between parameters by adding an L2 regularization term to the loss function, thereby improving the stability and generalization ability of the model. For example, when analyzing the linear relationship between the MnO content of steel slag and the compressive strength, ridge regression can provide a stable linear correlation coefficient to quantify the influence of MnO content change on compressive strength. Then, the random forest algorithm is used to simultaneously mine the nonlinear relationship of multiple parameters. Random forest can effectively capture complex nonlinear relationships by constructing multiple decision trees and integrating their prediction results. For example, when analyzing the comprehensive influence of multiple parameters such as the MnO content of steel slag, the SiO2 content of fly ash, and the stirring rate on the gamma-ray attenuation coefficient, random forest can provide a nonlinear relationship coefficient to quantify the influence of the interaction of these parameters on the performance indicator.

[0102] Then, the preliminary linear and nonlinear relationships are checked and corrected using the embedded causal relationship information. Specifically: 1. Causal consistency check: Compare the parameter-performance relationship (including the direction and size of the impact) obtained by ridge regression and random forest analysis with the causal relationship labels embedded in the gene code. For example, if the causal label indicates that "increasing the MnO content of steel slag has a positive impact on compressive strength", but the analysis result shows a negative correlation or no significant impact, it is marked as a pending verification item. 2. Conflict analysis and correction: For items with conflicts, trace back to the original experimental data or production records in the refined database to check data quality and boundary conditions. If the original data is confirmed to be reliable, re-examine the model assumptions or adjust the algorithm parameters; if it is found that the data has limitations or biases, according to the priority of causal knowledge, the analysis results are weighted and corrected or constraints are introduced to recalculate. 3. Knowledge integration and confirmation: Integrate the linear correlation coefficients and nonlinear relationship coefficients that have been checked and corrected to form a set of structure-activity relationship coefficients that are consistent with the field causal knowledge and self-consistent on the basis of data-driven.

[0103] Finally, the complete structure-activity relationship coefficients are output to provide a basis for the subsequent mapping generation stage. These structure-activity relationship coefficients not only include the linear relationship between a single parameter and the performance indicator, but also include the nonlinear relationship of multiple parameter interactions, and have passed the verification of causal knowledge, providing more reliable and interpretable quantitative basis for building accurate function mapping models.

[0104] Step S440, based on the structure-activity relationship coefficients obtained in the relationship mining stage, enter the mapping generation stage: take the solid waste composition, ratio, process parameters as independent variables, and the performance indicators as dependent variables, build a function mapping model, through which the quantitative correspondence between each parameter combination and the corresponding performance indicators is determined, forming a quantitative structure-activity relationship that integrates data sources and causal relationships between performance and solid waste composition, ratio, and process.

[0105] Among them, the function mapping model: a mathematical model used to describe the quantitative relationship between input parameters (such as solid waste composition, ratio, process parameters) and output performance indicators (such as compressive strength, gamma-ray attenuation coefficient). This model predicts performance indicators through input parameters, providing a scientific basis for the design and optimization of nuclear engineering concrete materials.

[0106] The specific process is as follows: first, taking the solid waste components, proportioning parameters, and process parameters as independent variables, and the performance indicators as dependent variables, a function mapping model is constructed. For example, taking the MnO content of steel slag, the SiO2 content of fly ash, the proportion of steel slag and fly ash, the stirring rate, the curing temperature and other parameters as independent variables, and the compressive strength and gamma-ray attenuation coefficient and other performance indicators as dependent variables, a mathematical model is constructed. This model can predict performance indicators by inputting parameters. Then, using the structure-activity relationship coefficients obtained in step S430, the quantitative correspondence between each parameter combination and the corresponding performance indicators is determined. For example, according to the linear correlation coefficients obtained by ridge regression algorithm and the nonlinear relationship coefficients obtained by random forest algorithm, the contribution of each parameter is quantified into the model. Then, through model training and verification, the accuracy and reliability of the model are ensured. For example, use the training data set to train the model, and then use the validation data set to validate the model, to ensure that the prediction error of the model is within an acceptable range. This process ensures that the model can accurately predict performance indicators, providing a reliable tool for subsequent applications. Finally, the mapping relationship between performance and solid waste components, proportioning, and process is formed. For example, through model prediction, the quantitative correspondence between the MnO content of steel slag, the SiO2 content of fly ash, the proportion of steel slag and fly ash, the stirring rate, the curing temperature and other parameter combinations and the compressive strength and gamma-ray attenuation coefficient and other performance indicators is determined. This mapping relationship provides a scientific basis for the design and optimization of nuclear engineering concrete materials.

[0107] In step S450, based on the formed quantitative structure-activity relationship, the application boundary is calibrated first: combined with the industry standards of nuclear engineering concrete, the effective application range of each parameter in the quantitative structure-activity relationship is determined, and the invalid parameter association beyond the boundary is eliminated to obtain the effective quantitative structure-activity relationship within the boundary.

[0108] Among them, the application boundary: refers to the effective range of parameters in a specific application scenario. For nuclear engineering concrete materials, the application boundary is usually defined by industry standards, design specifications and actual engineering requirements to ensure that the material performance meets the safety and functional requirements. Industry standards: refer to standardized regulations in the field of nuclear engineering regarding material performance, design and construction, which ensure the safety and reliability of materials and structures.

[0109] The specific process is as follows: First, determine the effective application range of each parameter in combination with the industry standards of nuclear engineering concrete. For example, according to the International Atomic Energy Agency (IAEA) and national standards, determine the effective range of parameters such as the MnO content of steel slag, the SiO2 content of fly ash, stirring rate, curing temperature, etc. These ranges are usually based on the physical and chemical properties of materials, engineering experience and safety requirements. Next, calibrate the application boundaries of the established quantitative structure-property relationships. Compare the value of each parameter in the quantitative structure-property relationship with the effective range defined in the industry standard. For example, if the industry standard specifies that the MnO content of steel slag should be between 5% and 15%, then the parameter value in the quantitative structure-property relationship that exceeds this range will be considered invalid. Then, eliminate invalid parameter correlations that exceed the application boundaries. For example, if the MnO content of steel slag in the quantitative structure-property relationship is 20%, which exceeds the range of 5% to 15% specified by the industry standard, then this parameter correlation will be eliminated. This process ensures that all retained quantitative structure-property relationships are effective and reliable in actual engineering. Finally, obtain the effective quantitative structure-property relationships within the boundaries. These relationships not only meet the industry standards, but also have application value in actual engineering. For example, the calibrated quantitative structure-property relationships can be used to design and optimize nuclear engineering concrete materials to ensure that their performance meets safety and functional requirements.

[0110] In step S460, the effective quantitative structure-property relationships within the boundaries are subjected to causal explainability verification: SHAP values are used to quantify the influence weight of each parameter on performance, LIME is used to generate regularized explanations, and the explanations are compared with the pre-set rules in the nuclear engineering expert library.

[0111] SHAP value: SHAP (SHapley Additive exPlanations) value is a method based on game theory, which is used to quantify the contribution of each feature (parameter) to the model prediction. It provides a fair and interpretable way to understand the decision-making process of the model. LIME: LIME is a method for explaining individual predictions and their feature importance. It explains the prediction results of complex models by locally fitting a simple model at the prediction point. Causal explainability: refers to the ability of model prediction results to be explained by causal relationships, i.e. the specific impact of parameter changes on performance indicators, rather than just statistical correlations.

[0112] The specific process is as follows: first, the SHAP value is used to quantify the influence weight of each parameter on the performance. For example, for the MnO content of steel slag, the SiO2 content of fly ash, and the stirring rate, the SHAP value of each parameter on the compressive strength and the gamma-ray attenuation coefficient is calculated. These SHAP values represent the contribution of each parameter to the performance index, thereby quantifying the importance of the parameters. Next, LIME is used to generate a regularized explanation. LIME locally fits a simple model to explain the prediction results of a complex model. For example, for a specific concrete sample, LIME can explain why its compressive strength reaches 45MPa and which parameters (such as the MnO content of steel slag and the stirring rate) contribute most to this result. Then, the SHAP value and the explanation generated by LIME are compared with the preset rules of the nuclear engineering expert library. The nuclear engineering expert library contains the experience and knowledge of industry experts and presets the causal relationship rules between parameters and performance indicators. For example, the expert library may preset the rule that "increasing the MnO content of steel slag will increase the compressive strength". Through comparison, it is verified whether the explanation of the quantitative structure-activity relationship is consistent with the expert rules. Finally, if the explanation result conflicts with the expert rules, it will enter the subsequent correction step (step S470). If the explanation result is consistent with the expert rules, it is considered that the quantitative structure-activity relationship has good causal explainability and can be used for practical application.

[0113] Step S470, if the explanation result conflicts with the expert rules, trace back to the original data of the perfect database through dynamic association tags, and correct the conflict items, then re-optimize the quantitative structure-activity relationship until the quantitative structure-activity relationship within the boundary meets the application boundary requirements and the expert causal knowledge.

[0114] Among them, the conflict item refers to the item that the explanation result of the quantitative structure-activity relationship is inconsistent with the preset rules of the nuclear engineering expert library in the causal explainability verification process (step S460). These conflict items may indicate that the quantitative structure-activity relationship has deviation or the expert rules need to be updated. Dynamic association tag tracing: using the data source and causal information recorded by the dynamic association tag, tracing back to the original data of the conflict item to determine the reason for the conflict and make corrections. Quantitative structure-activity relationship optimization: according to the tracing result, adjust and optimize the quantitative structure-activity relationship to ensure that its prediction result is consistent with the expert rules, and improve its reliability and accuracy.

[0115] The specific process is as follows: first, when the causal explainability check (step S460) finds that the explanation result of the quantitative structure-property relationship conflicts with the preset rules of the nuclear engineering expert library, the conflict item is traced back. For example, the quantitative structure-property relationship explanation shows that "increasing the MnO content of steel slag leads to a decrease in compressive strength", while the expert rule is preset as "increasing the MnO content of steel slag leads to an increase in compressive strength", at which time the reason for the conflict needs to be traced back. Then, using the data source and causal information recorded by the dynamic association label, the original data of the conflict item is traced back. The dynamic association label records the data source and causal relationship, and through these labels, the corresponding original data record in the perfect database can be quickly located. For example, it is traced back to an experimental record showing that under certain conditions, the increase in the MnO content of steel slag indeed leads to a decrease in compressive strength, which may be because the experimental conditions are different from the conditions assumed by the expert rules. Then, according to the tracing result, the reason for the conflict is analyzed. Possible reasons include differences in experimental conditions, deviations in quantitative structure-property relationships, limitations of expert rules, etc. For example, it is found that the curing temperature used in the experiment is different from the temperature assumed by the expert rules, resulting in different results. Then, the conflict item is corrected. According to the tracing result, the parameters of the quantitative structure-property relationship are adjusted or the expert rules are updated to solve the conflict. For example, if it is found that the difference in experimental conditions leads to the conflict, the quantitative structure-property relationship can be adjusted to consider the influence of different curing temperatures on performance, or the expert rules can be updated to reflect the new experimental conditions. Finally, the quantitative structure-property relationship is re-optimized. After correcting the conflict item, the mining and construction process of steps S410 to S440 is re-executed to ensure that the newly obtained quantitative structure-property relationship meets the application boundary requirement (step S450) and the expert causal knowledge (step S460). For example, based on the corrected data or rules, the model is retrained to verify whether the newly obtained quantitative structure-property relationship is consistent with the updated expert rules, and to ensure that its prediction results are effective and reliable within the application boundary.

[0116] In step S480, the quantitative structure-property relationship that passes the check is standardized and packaged according to the parameter effective range-performance threshold-value-causal explanation label-source tracing code standard, forming an interpretable quantitative structure-property relationship with boundary constraints, and stored in the knowledge base model layer.

[0117] Among them, the parameter effective range refers to the reasonable value range of the parameter (such as solid waste composition, ratio, process parameter) in actual application, to ensure that the material performance meets the design requirements. The performance threshold value refers to the minimum standard that the material performance (such as compressive strength, gamma-ray attenuation coefficient) must reach to meet the safety and functional requirements of nuclear engineering concrete. The causal explanation label is used to mark the causal relationship between parameter changes and performance indicators, facilitating the understanding and explanation of the prediction results of the model. The source tracing code is a unique identifier for tracing the data source, ensuring the traceability and transparency of the data.

[0118] The specific process is as follows: first, the verified quantitative structure-property relationship is standardized and packaged. This includes integrating the effective range of each parameter, the corresponding performance threshold, the causal explanation label and the source trace code. For example, for the MnO content of steel slag, the effective range (5% to 15%), the performance threshold (such as compressive strength ≥ 45 MPa), the causal explanation label (such as “positive impact on compressive strength”) and the corresponding data trace code are packaged. Then, based on the above packaging, an interpretable quantitative structure-property relationship with boundary constraints is formed. This relationship not only defines the mathematical correlation between parameters and performance, but also integrates its application boundary, performance target and causal knowledge. Then, the packaged interpretable quantitative structure-property relationship with boundary constraints is stored in the knowledge base model layer. The knowledge base model layer is a structured data storage system for saving verified and standardized model knowledge. This model knowledge can be used by subsequent model training, performance prediction and material design. For example, it is stored in a relational database or knowledge graph for easy querying and application.

[0119] Input the target performance requirements of the nuclear engineering concrete, and automatically generate multiple initial parameter combinations and perform reverse deduction based on the above quantitative structure-property relationship and preset deduction constraints to generate multiple initial optimization schemes, including:

[0120] Step S510, call the interpretable quantitative structure-property relationship with boundary constraints in the knowledge base model layer, and extract the parameter effective range, performance threshold and causal explanation label associated with the target performance requirements.

[0121] The specific process is as follows: first, the target performance requirements of the nuclear engineering concrete are determined. For example, the target performance requirements may include compressive strength ≥ 45 MPa, γ-ray attenuation coefficient ≥ 0.55 / cm, etc. These requirements are set according to specific engineering application scenarios and safety standards. Then, the interpretable quantitative structure-property relationship with boundary constraints in the knowledge base model layer is called. This relationship contains verified and standardized quantitative structure-property information, which can provide parameter effective range, performance threshold and causal explanation label, etc. For example, the relationship may record that the effective range of MnO content of steel slag is 5% to 15%, and the corresponding compressive strength threshold is 45 MPa. Then, the parameter effective range and performance threshold associated with the target performance requirements are extracted from the quantitative structure-property relationship. For example, for the target compressive strength ≥ 45 MPa, the effective range of MnO content of steel slag (5% to 15%) and the corresponding compressive strength threshold (45 MPa) are extracted. At the same time, the causal explanation label such as “increasing the MnO content of steel slag leads to an increase in compressive strength” is extracted. Finally, the extracted parameter effective range, performance threshold and causal explanation label are used for subsequent initial parameter combination generation and reverse deduction. These information provides key boundary conditions and guidance information for generating initial optimization schemes that meet the target performance requirements.

[0122] In step S520, based on the extracted parameter effective range, the Latin hypercube sampling method is used to generate initial parameter combinations: uniformly select the combination of solid waste components, mixing ratio parameters, and process parameters within the effective range of the boundary of the quantitative structure-activity relationship, and each combination needs to meet the safety constraints preset by the nuclear engineering expert library to ensure that all initial combinations fall within the solution space covered by the boundary of the effective quantitative structure-activity relationship. A total of pre-set number of initial parameter combinations are generated.

[0123] Latin hypercube sampling method: a statistical sampling technique used to generate sample points in multi-dimensional space. It divides each dimension into equal probability intervals and randomly selects a sample point in each interval, ensuring that the sample points are uniformly distributed in each dimension, thereby improving the representativeness and efficiency of sampling. Safety constraints: based on the rules preset by the nuclear engineering expert library, ensure that the generated parameter combinations meet the safety and functional requirements of nuclear engineering concrete. These constraints usually include the reasonable range of parameters, the relationship between parameters, etc. Solution space: refers to the space formed by all possible parameter combinations. In this step, the solution space refers to the parameter combination range covered by the boundary of the effective quantitative structure-activity relationship, ensuring that the generated parameter combinations are effective and reliable in practical application.

[0124] The specific process is as follows: First, according to the parameter effective range extracted in step S510, determine the value range of each parameter. For example, the effective range of MnO content of steel slag is 5% to 15%, the effective range of SiO2 content of fly ash is 20% to 30%, and the effective range of stirring rate is 50 to 150 rpm. Then, the Latin hypercube sampling method is used to uniformly select parameter combinations within these effective ranges. The Latin hypercube sampling method divides the value range of each parameter into equal probability intervals and randomly selects a sample point in each interval, ensuring the uniformity of sample points in each parameter dimension. For example, the MnO content range of steel slag is divided into 10 equal probability intervals, and a sample point is randomly selected in each interval; similarly, sample the parameters such as SiO2 content of fly ash and stirring rate. Then, during the sampling process, check whether each generated parameter combination meets the safety constraints preset by the nuclear engineering expert library. For example, the expert library may preset the proportional relationship between the MnO content of steel slag and the SiO2 content of fly ash, or the synergistic relationship between the stirring rate and the curing temperature. Through the check, it is ensured that all initial combinations fall within the solution space covered by the boundary of the effective quantitative structure-activity relationship. Finally, a pre-set number of initial parameter combinations are generated. For example, if 100 initial parameter combinations are needed, 100 parameter combinations are uniformly selected within the parameter effective range by the Latin hypercube sampling method, and it is ensured that each combination meets the safety constraints.

[0125] Step S530, based on the function mapping relationship in the interpretable quantitative structure-activity relationship, perform reverse derivation on the initial parameter combination with the target performance requirement as the output constraint.

[0126] Target performance requirement: performance indicators set according to the specific application scenario of nuclear engineering concrete, such as compressive strength, gamma-ray attenuation coefficient, etc. Quantitative structure-activity relationship: in the interpretable quantitative structure-activity relationship with boundary constraints, the calculable relationship between input parameters (solid waste composition, proportioning, process parameters) and output performance indicators (such as compressive strength, gamma-ray attenuation coefficient) is described.

[0127] The specific process is as follows: First, clarify the target performance requirement. For example, the target performance requirement may include compressive strength ≥ 45 MPa, gamma-ray attenuation coefficient ≥ 0.55 / cm, etc. These requirements are set according to the specific engineering application scenario and safety standards. Then, call the interpretable quantitative structure-activity relationship with boundary constraints. This relationship contains the calculable relationship between input parameters (solid waste composition, proportioning, process parameters) and output performance indicators (such as compressive strength, gamma-ray attenuation coefficient). For example, this relationship may record the specific relationship between parameters such as MnO content of steel slag, SiO2 content of fly ash, and stirring rate and compressive strength and gamma-ray attenuation coefficient. Then, with the target performance requirement as the output constraint, perform reverse derivation on each initial parameter combination. For example, for the target compressive strength ≥ 45 MPa, according to the quantitative structure-activity relationship, derive the specific values of solid waste composition, proportioning, and process parameters that meet this performance requirement from the initial parameter combination. This process may involve numerical optimization methods such as gradient descent or genetic algorithm to find parameter combinations that meet the constraints. Finally, record the reverse derivation results of each initial parameter combination. These results include the specific values of the parameters and the corresponding performance estimates. For example, record that a certain parameter combination (steel slag MnO content 10%, fly ash SiO2 content 25%, stirring rate 100 rpm) corresponds to a compressive strength of 47 MPa and a gamma-ray attenuation coefficient of 0.58 / cm.

[0128] Step S540, after completing the reverse derivation of each initial parameter combination, organize the corresponding initial optimization scheme, which includes parameter specific values, performance compliance estimates, and derivation logic explanations based on causal explanation labels.

[0129] The specific process is as follows:

[0130] First, organize the reverse derivation results of each initial parameter combination. This includes recording the specific values of each parameter and the corresponding performance estimates. For example, for a certain parameter combination (steel slag MnO content 10%, fly ash SiO2 content 25%, stirring rate 100 rpm), record that it corresponds to a compressive strength of 47 MPa and a gamma-ray attenuation coefficient of 0.58 / cm.

[0131] Then, based on the causal explanation tags, the rationality of the selection of each initial parameter combination and the performance estimate value is deduced logically. The causal explanation tags record the specific influence of parameter changes on performance indicators, and these information can be used to explain why a specific parameter combination is selected and how these combinations affect the performance indicators. For example, it is explained that "the selection of steel slag MnO content of 10% is because according to the model prediction, this content can improve the compressive strength to 47 MPa, meeting the target performance requirement".

[0132] Then, the detailed information and deduction logic of each initial parameter combination are sorted into a complete initial optimization scheme. Each scheme not only contains the specific values of the parameters and the performance estimates, but also contains detailed explanations of these values and estimates. For example, a complete initial optimization scheme can be as follows:

[0133] Parameter values:

[0134] Steel slag MnO content: 10%;

[0135] Fly ash SiO2 content: 25%;

[0136] Stirring rate: 100 rpm;

[0137] Performance target estimate value:

[0138] Compressive strength: 47 MPa;

[0139] Gamma-ray attenuation coefficient: 0.58 / cm;

[0140] Deduction logic explanation: The selection of steel slag MnO content of 10% is because according to the model prediction, this content can improve the compressive strength to 47 MPa, meeting the target performance requirement. The selection of fly ash SiO2 content of 25% is because this content can optimize the gamma-ray attenuation coefficient to 0.58 / cm, meeting the target performance requirement. The selection of stirring rate of 100 rpm is because this rate can ensure the uniformity and stability of the concrete, meeting the process requirements of nuclear engineering concrete.

[0141] Step S550, all initial optimization schemes are summarized to form a set of initial optimization schemes with a preset number.

[0142] Based on the function mapping relationship in the interpretable quantitative structure-activity relationship, the initial parameter combination is executed in reverse deduction with the target performance requirement as the output constraint, including:

[0143] Step S531, the target performance requirement is decomposed into nuclear engineering key sub-targets, and the sub-target weight is allocated according to the nuclear engineering safety priority, and the minimum threshold of each sub-target is determined.

[0144] Target performance requirements: Based on the specific application scenarios of nuclear engineering concrete, set performance indicators such as compressive strength, gamma-ray attenuation coefficient, etc. Key sub-targets of nuclear engineering: Decompose the target performance requirements into multiple key sub-targets, each corresponding to a specific performance indicator, such as compressive strength, durability, radiation shielding performance, etc. Sub-target weights: According to the safety priority of nuclear engineering, assign weights to each sub-target to reflect the importance of different sub-targets in the overall target. Minimum threshold: The minimum performance standard that each sub-target must meet to ensure that the nuclear engineering concrete material meets safety and functional requirements.

[0145] The specific process is as follows:

[0146] First, clarify the overall target performance requirements of nuclear engineering concrete. For example, the overall target performance requirements may include compressive strength ≥ 45 MPa, gamma-ray attenuation coefficient ≥ 0.55 / cm, durability ≥ 300 freeze-thaw cycles, etc.

[0147] Next, decompose the overall target performance requirements into multiple key sub-targets. For example, decompose the target performance requirements into the following sub-targets: compressive strength ≥ 45 MPa; gamma-ray attenuation coefficient ≥ 0.55 / cm; durability ≥ 300 freeze-thaw cycles.

[0148]

[0149] Then, assign weights to each sub-target according to the safety priority of nuclear engineering. For example, according to the safety requirements of nuclear engineering, assign weights as follows: compressive strength: weight 0.4; gamma-ray attenuation coefficient: weight 0.3; durability: weight 0.3.

[0150] Finally, clarify the minimum threshold of each sub-target. These thresholds are set according to the safety standards and functional requirements of nuclear engineering to ensure that the material performance meets the actual application requirements. For example: set the minimum threshold of compressive strength to 45 MPa; set the minimum threshold of gamma-ray attenuation coefficient to 0.55 / cm; set the minimum threshold of durability to 300 freeze-thaw cycles.

[0151] Step S532, based on the function mapping relationship in the interpretable quantitative structure-activity relationship, establish the correlation equation between the initial parameter combination and each sub-target performance, and embed the parameter influence direction in the causal explanation label in the equation.

[0152] ​The specific process is as follows: first, call the interpretable quantitative structure-activity relationship with boundary constraints. The relationship describes the computable relationship between input parameters (such as solid waste composition, proportioning, process parameters) and output performance indicators (such as compressive strength, gamma-ray attenuation coefficient). For example, the relationship may record the specific relationship between parameters such as MnO content of steel slag, SiO2 content of fly ash, stirring rate and compressive strength and gamma-ray attenuation coefficient. Next, according to the sub-targets decomposed in step S531, establish the correlation equation between the initial parameter combination and the performance of each sub-target. For example, for the compressive strength sub-target, a correlation equation is established to describe the relationship between parameters such as MnO content of steel slag, SiO2 content of fly ash, stirring rate and compressive strength. Similarly, establish the corresponding correlation equation for the gamma-ray attenuation coefficient and durability sub-targets. Then, in each correlation equation, according to the parameter influence direction indicated by the causal explanation label, constrain or give a sign direction to the coefficients or structure of the corresponding parameters in the equation. The causal explanation label records the specific influence of parameter changes on performance indicators, which directly guides the setting of the parameter action direction in the equation. For example, in the correlation equation of compressive strength, if the causal explanation label is "increasing the MnO content of steel slag leads to an increase in compressive strength", then ensure that the coefficient of the term representing the MnO content in the equation is positive. Finally, verify whether the established correlation equation can accurately reflect the relationship between the initial parameter combination and the performance of the sub-target. By substituting the known parameter combination and performance data, check whether the prediction result of the equation is consistent with the actual data. If there is a large deviation between the prediction result of the equation and the actual data, trace back to the establishment process of the quantitative structure-activity relationship or the causal explanation label, and check and correct it.

[0153] Step S533, set parameter adjustment dual constraints, with the effective range of parameters in the interpretable quantitative structure-activity relationship as hard constraints, and set soft constraints in combination with sub-target weights, with the single adjustment amplitude of high-weight sub-target associated parameters ≤ a preset value, and the single adjustment amplitude of low-weight sub-target associated parameters ≤ another preset value.

[0154] wherein, hard constraint: refers to the restriction condition that must be strictly followed in the parameter adjustment process, which is usually the effective range of parameters set based on physical and chemical properties or engineering safety standards. For example, the MnO content of steel slag must be between 5% and 15%. Soft constraint: refers to a relatively flexible restriction condition in the parameter adjustment process, which is usually an adjustment amplitude limit set based on sub-target weights. For example, the single adjustment amplitude of high-weight sub-target associated parameters may be limited to a small range to ensure the stability and reliability of the adjustment. Parameter adjustment amplitude: refers to the change amount of parameter value in each adjustment during the parameter adjustment process. For example, the single adjustment amplitude of the MnO content of steel slag may be limited to within 1%.

[0155] The specific process is as follows: First, set the hard constraints. According to the effective range of parameters extracted from the quantitative structure-activity relationship, set hard constraints for each parameter. These hard constraints ensure that the parameter adjustment process does not exceed the range allowed by physical and chemical properties and engineering safety standards. For example, the effective range of MnO content in steel slag is 5% to 15%, so the MnO content must always be within this range during the adjustment process. Next, set the soft constraints. According to the sub-target weight, set soft constraints for each parameter. The single adjustment range of the parameter associated with the high-weight sub-target is limited to a smaller range to ensure the stability and reliability of the adjustment. For example, for the high-weight compressive strength sub-target, the single adjustment range of the associated parameter (such as the MnO content of steel slag) may be limited to within 1%. The single adjustment range of the parameter associated with the low-weight sub-target can be relatively large, but still needs to be within a reasonable range. For example, for the low-weight durability sub-target, the single adjustment range of the associated parameter (such as the stirring rate) may be limited to within 10 rpm. Then, adjust the parameters in combination with the hard constraints and soft constraints. In each adjustment process, first check whether the adjusted parameter value meets the hard constraint condition. If it meets the hard constraint condition, then determine or verify the adjustment range according to the soft constraint condition. For example, if the adjusted MnO content of steel slag is 16%, which exceeds the hard constraint range (5% to 15%), it needs to be re-adjusted to ensure it is within the effective range. If the adjusted MnO content is 14%, which meets the hard constraint condition, but the single adjustment amount exceeds the soft constraint condition (such as 1%), it needs to be adjusted to meet the soft constraint range. Finally, record the results of each adjustment to ensure the transparency and traceability of the parameter adjustment process. For example, record the parameter value, adjustment range, and adjusted performance estimate of each adjustment to facilitate subsequent analysis and optimization.

[0156] In step S534, the weighted gradient descent algorithm is used to perform parameter iterative optimization, taking the sum of the deviations of each sub-target performance estimate from the minimum threshold multiplied by the sub-target weight as the loss function, and prioritizing the iteration of parameters that significantly affect high-weight sub-targets. After each iteration, the performance estimate is calculated through the associated equation until the loss function value is lower than the preset threshold.

[0157] Weighted Gradient Descent Algorithm: An optimization algorithm that iteratively updates parameters by calculating the gradient of the loss function and combining it with weights to minimize the loss function. In this step, the loss function is the sum of the deviations of each sub-objective performance estimate from the minimum threshold multiplied by the sub-objective weight. Loss Function: A function used to measure the difference between the model's predicted value and the target value. In this step, the loss function is the sum of the deviations of each sub-objective performance estimate from the minimum threshold multiplied by the sub-objective weight, which guides the optimization direction of the parameters. Sub-Objective Weight: A weight assigned according to the priority of nuclear engineering safety, reflecting the importance of different sub-objectives in the overall objective. High-weight sub-objectives have a greater proportion in the loss function, and are prioritized for optimization. Parameter Iterative Optimization: Adjusting parameter values through multiple iterations to gradually reduce the value of the loss function until the preset optimization target or stopping condition is reached.

[0158] The specific process is as follows:

[0159] First, define the loss function. The loss function is the sum of the deviations of each sub-objective performance estimate from the minimum threshold multiplied by the sub-objective weight. For example, for the compressive strength sub-objective, its loss function term is (estimated compressive strength - target threshold) x the weight of this sub-objective. Next, initialize the parameters. Set the initial parameter values based on the initial parameter combination generated in step S520. For example, the initial parameter combination may be steel slag MnO content 10%, fly ash SiO2 content 25%, and stirring rate 100 rpm. Then, calculate the initial loss function value. Substitute the initial parameter values into the correlation equation to calculate the performance estimates of each sub-objective, and then calculate the initial loss function value according to the definition of the loss function. For example, based on the initial parameter combination, calculate the performance estimates, and accordingly obtain the initial loss function value. Next, use the weighted gradient descent algorithm for parameter iterative optimization. Calculate the gradient of the loss function with respect to each parameter, and iteratively update the parameters in combination with the sub-objective weights. In the iteration, according to the gradient information and the weight, the parameters that are more sensitive to the performance of the sub-objective with the highest weight are updated first. For example, if the compressive strength sub-objective has the highest weight, then the parameters that have a significant impact on it, such as the MnO content of steel slag and the SiO2 content of fly ash, are adjusted first. After each iteration, calculate the performance estimates corresponding to the updated parameter combination through the correlation equation, and recalculate the loss function value. For example, the adjusted parameter combination is steel slag MnO content 11%, fly ash SiO2 content 26%, and stirring rate 105 rpm, and the new performance estimates and loss function value are calculated. Repeat the above iteration process until the loss function value is lower than the preset optimization threshold. For example, when the loss function value converges to be lower than the preset threshold, stop the iteration.

[0160] Step S535, call the regularization interpretation to perform parameter interaction verification, if the parameter combination after iteration triggers the synergistic constraint, adjust the associated parameters to comply with the quantitative structure-activity relationship rule.

[0161] Rule-based interpretation: A method for interpreting model prediction results by generating concise rules to describe the decision-making process of the model. In this step, rule-based interpretation is used to verify whether the parameter combination after parameter adjustment conforms to the synergistic rules between parameters revealed by the quantitative structure-activity relationship. Parameter interaction verification: Through rule-based interpretation, the adjusted parameter combination is verified to ensure that the synergistic relationship between parameters conforms to the rules of the quantitative structure-activity relationship. Synergistic constraints: Refers to the interdependent relationships between parameters revealed by the quantitative structure-activity relationship, which need to be considered and satisfied during parameter adjustment. For example, the adjustment of certain parameters may affect the optimal values of other parameters. Coordinated adjustment: When the parameter combination triggers synergistic constraints, the associated parameters are adjusted synchronously to ensure that the parameter combination conforms to the rules of the quantitative structure-activity relationship.

[0162] The specific process is as follows:

[0163] First, call rule-based interpretation to verify the parameter combination after iteration. Rule-based interpretation describes the decision-making process of the quantitative structure-activity relationship by generating concise rules, which can be used to verify the rationality of parameter adjustment. For example, rule-based interpretation may generate the following rule: "If the MnO content of steel slag increases, the SiO2 content of fly ash needs to be reduced accordingly to maintain the stability of compressive strength".

[0164] Next, check whether the parameter combination triggers synergistic constraints. Synergistic constraints refer to the interdependent relationships between parameters, which need to be considered and satisfied during parameter adjustment. For example, if the adjusted MnO content of steel slag is 12%, and the SiO2 content of fly ash is still 25%, this may trigger synergistic constraints, because according to the rules of the quantitative structure-activity relationship, the increase in MnO content requires a corresponding decrease in SiO2 content.

[0165] Then, if the parameter combination triggers synergistic constraints, perform coordinated adjustment. Coordinated adjustment refers to the synchronous adjustment of associated parameters to ensure that the parameter combination conforms to the rules of the quantitative structure-activity relationship. For example, according to the rule generated by rule-based interpretation, adjust the SiO2 content of fly ash from 25% to 24% to maintain the stability of compressive strength.

[0166] Finally, verify if the adjusted parameter combination meets the quantitative structure-activity relationship rules. By substituting into the correlation equation, calculate the performance estimates corresponding to the adjusted parameter combination, ensuring that the adjusted parameter combination not only meets the synergy constraints, but also meets the target performance requirements. For example, the adjusted parameter combination (steel slag MnO content 12%, fly ash SiO2 content 24%, stirring rate 105 rpm) corresponds to the compressive strength estimate of 46 MPa, the gamma-ray attenuation coefficient estimate of 0.57 / cm, and the durability estimate of 315 freeze-thaw cycles. Verify whether these estimates meet the target performance requirements.

[0167] Step S536, when the parameter combination meets all the sub-target performance estimates ≥ minimum threshold and does not exceed the double constraints, stop iteration and output the parameter combination as the reverse derivation result.

[0168] The specific process is as follows:

[0169] First, verify if the parameter combination meets the performance estimate requirements of all sub-targets. By substituting into the correlation equation, calculate the performance estimates corresponding to the current parameter combination, and check whether these estimates meet the minimum threshold of all sub-targets. For example, check if the compressive strength ≥ 45 MPa, the gamma-ray attenuation coefficient ≥ 0.55 / cm, and the durability ≥ 300 freeze-thaw cycles.

[0170] Next, verify if the parameter combination does not exceed the double constraints. Check if the parameter values are within the hard constraints (parameter effective range) and if the parameter word adjustment amount meets the soft constraints (parameter adjustment amplitude limit). For example, check if the MnO content of steel slag is between 5% and 15%, and if the single adjustment amplitude of stirring rate ≤ 10 rpm.

[0171] Then, if the parameter combination meets the performance estimate requirements of all sub-targets and does not exceed the double constraints, stop iteration. This indicates that the current parameter combination has been optimized to meet the target performance requirements, and the adjustment process meets all the constraints.

[0172] Finally, output the parameter combination as the reverse derivation result. Record the specific values of the parameters, the corresponding performance estimates, and the related derivation logic explanation to form a complete reverse derivation result. For example, output the parameter combination (steel slag MnO content 12%, fly ash SiO2 content 24%, stirring rate 105 rpm), the corresponding performance estimates (compressive strength 46 MPa, gamma-ray attenuation coefficient 0.57 / cm, durability 315 freeze-thaw cycles), and the derivation logic explanation.

[0173] Based on the same inventive concept, the embodiment of the present application provides a solid waste high-value nuclear engineering concrete material knowledge base establishment system, which comprises a storage and a processor. Figure 1 The program of the method is stored in the storage and can be run on the processor to realize the method.

[0174] The embodiments of the specific implementation are the preferred embodiments of the present application, but do not limit the protection scope of the present application, so that: any equivalent changes made according to the structure, shape, principle of the present application should be covered within the protection scope of the present application.

Claims

1. A method for establishing a knowledge base of solid waste high-value nuclear engineering concrete material, characterized in that, The application comprises the following steps: Collecting external data through a preset standardized search term library and an intelligent search engine, and synchronously collecting internal data, to complete the integration and classification of double-source heterogeneous data; Based on the preset tree-shaped four-level parameter architecture, the integrated and classified data are structured and cleaned, classified according to parameter categories, subcategories and index levels, and the abnormal values are removed and the data units are unified to form a standardized data set; The standardized data set is imported into a preset database framework to build an initial database, a preset machine learning model is used to learn the correlation between composition, process and performance, a supplementary sample is generated to realize data expansion, and a perfect database is obtained; The composition-ratio-process-performance structure-activity relationship in the perfect database is mined through a preset gene encoding model, and the quantitative structure-activity relationship between performance and solid waste composition, ratio and process is established; The target performance requirements of nuclear engineering concrete are input, based on the above quantitative structure-activity relationship and preset derivation constraints, a plurality of initial parameter combinations are automatically generated and reverse derivation is performed to generate a plurality of initial optimization schemes; Each initial optimization scheme is substituted into the quantitative structure-activity relationship to verify the performance compliance: if it is compliant, it is included in the candidate optimization scheme set; If it is not compliant, the parameters are corrected according to the preset adjustment rules and re-derived until a preset number of candidate schemes are generated; Based on a preset screening index system, a preset weighted scoring model is used to score the candidate scheme set, and the highest scoring scheme is selected as the optimal scheme; The perfect database, quantitative structure-activity relationship, candidate scheme set, optimal scheme and iteration process data are stored in a preset knowledge graph structure, and a knowledge base including data layer, model layer and scheme layer is constructed.

2. The method according to claim 1, wherein The double-source heterogeneous data integration and classification comprises the following steps: The collected external data is subjected to format standardization processing, natural language processing algorithm is used to extract key entities from unstructured text, and regular expression is used to unify the data expression format, the key entities including solid waste type, performance index and process description; The collected internal data is subjected to structured conversion, discrete data in experiment records and production logs are reorganized according to the preset fields of solid waste name-detection date-physical property parameter-process parameter-characteristic data, and the reorganized data is subjected to preliminary screening by Z-score method to remove invalid data deviating from the preset normal distribution range; The cosine similarity algorithm is used to calculate the feature matching degree of the key entities extracted from the external data and the preset fields of the internal data, and a bidirectional association mapping of literature mentioned performance index-experimentally detected performance data and patent process description-production process parameter is established; According to the association mapping result, the double-source data is classified according to the solid waste type and the nuclear engineering concrete application scenario, the redundant data with a feature matching degree lower than a preset feature matching degree is removed, and finally a double-source integrated data set with clear classification and effective data is formed.

3. The method according to claim 2, wherein, When the collected internal data is subjected to structured conversion, the blockchain traceability processing is also included, which comprises the following steps: The discrete data in the experiment records and production logs are reorganized according to the preset fields of solid waste name-blockchain traceability code-detection date-physical property parameter-process parameter-characteristic data, and the blockchain traceability code is added to the original fields. The blockchain traceability code is associated with the solid waste full-link data, including the origin source identification, transportation process information, detection original data and storage information, and the full-link data is encrypted and stored by the smart contract to ensure that it cannot be tampered with; When the Z-score method is used to preliminarily screen the recombined data, the internal data without the blockchain traceability code or incomplete traceability data is simultaneously removed, and only the structured data with complete traceability information is retained.

4. The method according to claim 1, wherein, The standardized data set is imported into the preset database framework to build an initial database, a preset machine learning model is used to learn the correlation between composition-process-performance, and a supplementary sample is generated to realize data expansion, and a perfect database is obtained, including: The standardized data set is classified according to the four-level parameters of solid waste properties, production process, material performance and service environment, and is imported into the preset relational database framework; the framework defines the field type, dimension unit and cross-parameter correlation rule of each parameter through a data dictionary, and realizes fast data calling through an index structure, so as to build an initial database with a standard structure that can be directly used for model training; The data in the initial database is used as the training sample, the BP neural network is used as the preset machine learning model, the solid waste composition parameters and process parameters are used as the input layer, and the material performance parameters are used as the output layer, the network weight is iteratively adjusted through the error back propagation algorithm, the nonlinear correlation between composition-process-performance is learned, until the prediction error of the model on the verification set is lower than the preset threshold, and the trained model which can accurately map the parameter correlation is formed; Based on the correlation captured by the trained model and the sample distribution characteristics of the initial database, a generative adversarial network algorithm is used to generate supplementary samples: a generator is trained based on the real samples of the initial database to generate virtual samples that meet the composition-process-performance correlation rule; at the same time, the consistency of the virtual samples and the real samples is verified by the discriminator, and only the virtual samples whose probability of being judged as real samples reaches the preset threshold are temporarily stored as candidate supplementary samples; Integrate the initial database and the candidate supplementary samples verified by the discriminator, use the K-means clustering algorithm for consistency verification, remove the abnormal samples with clustering deviation value exceeding the preset range, and finally form a perfect database with balanced data distribution and significant correlation rule.

5. The method according to claim 1, wherein, The composition-ratio-process-performance structure-activity relationship of solid waste in the perfect database is mined through the preset gene coding model, and the quantitative structure-activity relationship between performance and solid waste composition, ratio and process is established, including: Extract the data with dynamic association tags from the perfect database, and construct a four-dimensional feature matrix according to the solid waste composition-ratio parameter-process parameter-performance index dimension, wherein the dynamic association tag records the source information of the data and the causal association information between parameters; Based on the four-dimensional feature matrix, enter the parameter coding stage: convert the solid waste composition, ratio parameter and process parameter into standardized gene sequences, and embed the causal association information in the dynamic association tag into the corresponding coding bit to form a gene coding data set with causal markers; With the coded gene encoding dataset as input, enter the relationship mining stage: adopt the ridge regression algorithm to control the parameter collinearity, and quantify the linear correlation coefficient between single parameter and performance index; simultaneously adopt the random forest algorithm to mine the nonlinear relationship of multi-parameter interaction, verify and correct the linear and nonlinear analysis results based on the embedded causal correlation information, and output the complete structure-activity relationship coefficient; Based on the structure-activity relationship coefficient obtained in the relationship mining stage, enter the mapping generation stage: take the solid waste composition, ratio and process parameters as independent variables, and the performance index as dependent variable, construct a function mapping model, and through the model, the quantitative corresponding relationship between each parameter combination and the corresponding performance index is determined, forming a quantitative structure-activity relationship of performance and solid waste composition, ratio and process which integrates data sources and causal correlation.

6. The method according to claim 5, wherein, After forming a quantitative structure-activity relationship of performance and solid waste composition, ratio and process which integrates data sources and causal correlation, the following steps are included: Based on the formed quantitative structure-activity relationship, first calibrate its application boundary: combined with the industry standard of nuclear engineering concrete, the effective application range of each parameter in the quantitative structure-activity relationship is determined, and invalid parameter correlation beyond the boundary is eliminated, obtaining the effective quantitative structure-activity relationship within the boundary; Start the causal explainability check for the effective quantitative structure-activity relationship within the boundary: use SHAP value to quantify the influence weight of each parameter on performance, generate regularized explanation through LIME, and compare with the pre-set rules of nuclear engineering expert database; If the explanation result conflicts with the expert rules, trace back to the original data in the perfect database through dynamic association label, modify the conflict items, and optimize the quantitative structure-activity relationship again until the quantitative structure-activity relationship within the boundary meets the application boundary requirement and the expert causal knowledge at the same time; Standardize the quantitative structure-activity relationship that passes the check according to the parameter effective range-performance threshold-causal explanation label-source trace code, form the interpretable quantitative structure-activity relationship with boundary constraints, and store it in the knowledge base model layer.

7. The method according to claim 6, wherein the method is characterized by, Input the target performance requirement of nuclear engineering concrete, based on the above quantitative structure-activity relationship and pre-set derivation constraints, automatically generate multiple initial parameter combinations and perform reverse derivation to generate multiple initial optimization schemes, including: Call the interpretable quantitative structure-activity relationship with boundary constraints in the knowledge base model layer, extract the parameter effective range, performance threshold and causal explanation label associated with the target performance requirement from it; Based on the extracted parameter effective range, use Latin hypercube sampling method to generate initial parameter combinations: uniformly select the combination of solid waste composition, ratio parameter and process parameter within the parameter effective range of the boundary effective quantitative structure-activity relationship, and each combination must meet the pre-set safety constraints of nuclear engineering expert database, to ensure that all initial combinations fall within the solution space covered by the boundary effective quantitative structure-activity relationship, a total of pre-set number of initial parameter combinations are generated; Based on the function mapping relationship in the interpretable quantitative structure-activity relationship, perform reverse derivation on the initial parameter combinations with the target performance requirement as the output constraint; After completing the reverse derivation of each initial parameter combination, the corresponding initial optimization scheme is formed, which includes parameter specific value, performance threshold estimation and derivation logic description based on causal explanation label. Aggregate all initial optimization schemes to form a preset number of initial optimization scheme sets.

8. A solid waste high-value nuclear engineering concrete material knowledge base establishment system, characterized in that, The computer readable storage medium stores a program, and the program can be loaded and executed by the processor to implement the method for establishing a solid waste high-value nuclear engineering concrete material knowledge base according to any one of claims 1 to 7.

Images