Nuclear facility decommissioning management system and method based on generative large model

Abstract

The application discloses a nuclear facility decommissioning management system and method based on a generative large model, wherein the method comprises the following steps: acquiring field sensor data and historical decommissioning records, extracting initial parameter values of internal elements of each process, including radiation source item contribution and structure change index, and obtaining a standardized element data set after removing noise by using a data cleaning method; the standardized element data set is randomly sampled multiple times by using Monte Carlo simulation to simulate the mutual restraint relationship between internal elements of the process, and the coupling strength distribution between the elements is obtained; according to the coupling strength distribution between the elements, a cause-and-effect chain diagram model is constructed, the amplification effect of key micro operations on subsequent links is identified, and a dynamic weight distribution matrix is determined; the application aims to solve the problem of how to accurately capture the mutual restraint and amplification relationship between the elements on the basis of fully identifying and analyzing the internal constituent elements of each decommissioning process in the prior art.

Description

Technical Field

[0001] This invention relates to the technical field of information technology, specifically to a nuclear facility decommissioning management system and method based on a generative large model. Background Technology

[0002] Nuclear facility decommissioning and management is the final stage of the nuclear energy industry's full life-cycle management, directly impacting environmental safety, public health, and sustainable resource utilization. As a large number of early-stage nuclear facilities enter the decommissioning phase, the importance of this area is increasingly evident, becoming an unavoidable key task for ensuring the sustainable development of nuclear energy.

[0003] Current decommissioning projects generally rely on empirical formulas, historical analogies, and relatively crude overall estimation methods. These methods often fail to accurately reflect the reality of complex nuclear facilities structures, uneven distribution of radioactive contamination, and high coupling between processes. Especially when a single major process contains multiple sub-operations of completely different natures, and these sub-operations affect the final radiation dose, personnel man-hours, and waste generation in different ways, traditional holistic, packaged predictions are prone to systematic biases, leading to resource imbalances or insufficient safety margins.

[0004] The most critical practical challenge in the decommissioning of nuclear facilities lies in the complex interrelationships among the constituent elements of the process. On the one hand, each specific operational step contributes its own radiation source term, while simultaneously being directly constrained by contamination residues and structural changes from preceding operations. On the other hand, the weights of different sub-steps on the overall process outcome vary greatly and are highly dynamic; some critical micro-operations can significantly amplify the dose load or time consumption of subsequent steps. This strong coupling between internal elements and the constantly changing weight distribution with site conditions make traditional linear superposition or average allocation methods inaccurate in depicting the real process. For example, in the dismantling of reactor pressure vessels, if the support and reinforcement process does not adequately consider the degree of metal fatigue, leading to localized stress concentration, the diffusion range and concentration of trace radioactive aerosols generated during subsequent main body cutting may increase exponentially, directly increasing the radiation dose to operators and the workload of cleanup. This causal chain is often ignored or severely underestimated in current methods.

[0005] Therefore, how to accurately capture the mutual constraints and amplification relationships between the constituent elements of each decommissioning process, and thereby achieve multi-level, dynamic coupling numerical prediction from individual micro-operations to overall macro-processes, has become a key issue that urgently needs to be addressed in the decommissioning management of nuclear facilities. Summary of the Invention

[0006] This invention provides a nuclear facility decommissioning management system and method based on a generative large model. The purpose is to solve the problem in the prior art of how to accurately capture the mutual constraints and amplification relationships between the internal components of each decommissioning process, based on fully identifying and analyzing these components.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A nuclear facility decommissioning management method based on a generative large model includes: acquiring on-site sensor data and historical decommissioning records; extracting initial parameter values ​​for internal elements of each process, including radiation source contributions and structural change indicators; and obtaining a standardized element dataset after noise removal using data cleaning methods. The standardized element dataset is then subjected to multiple random samplings using Monte Carlo simulations to simulate the mutual constraints between internal elements and obtain the coupling strength distribution between elements. Based on the coupling strength distribution, a causal chain graph model is constructed to identify the amplification effect of key micro-operations on subsequent stages and determine the dynamic weight allocation matrix. Finally, a neural network is used to apply the dynamic weight allocation matrix. The training process involves inputting node features from a causal chain graph model and outputting an adjusted sequence of weight values ​​to capture changes in weights under site conditions. If the weight value sequence exceeds a preset threshold, the amplified relationship analysis module is triggered to extract high-impact sub-operations from the weight value sequence, resulting in a micro-operation influence vector. By mapping the micro-operation influence vector to macro-process objectives, a multi-level coupled prediction model is generated, integrating pollution distribution and structurally complex parameters to determine the overall radiation dose estimate. The overall radiation dose estimate is then extended to predict working hours and waste volume. An iterative optimization algorithm is used to update the parameters of the multi-level coupled prediction model, yielding the final numerical prediction result for decommissioning and remediation decisions.

[0008] In one aspect of the invention, the acquisition of on-site sensor data and historical decommissioning records, extraction of initial parameter values ​​for internal elements of each process, including radiation source contribution and structural change indicators, and the generation of a standardized element dataset after noise removal using data cleaning methods, includes: Acquire on-site sensor data and historical decommissioning records, and record and store relevant information for each process in real time through a data acquisition system to obtain a complete raw dataset; For the original dataset, data cleaning methods are used to process outliers and missing values ​​in sensor data and historical records. Valid data is filtered out through a preset threshold range to determine the initial cleaned dataset. Initial parameter values ​​of internal elements of each process are extracted from the cleaned preliminary dataset, covering radiation source contribution and structural change indicators. The data is then classified and organized using field mapping methods to obtain the classified parameter dataset. For the classified parameter dataset, standardization processing techniques are applied to normalize the numerical values ​​of radiation source contribution and structural change indicators. A standardized element dataset is constructed through a unified dimensional conversion method. Based on the standardized element dataset, the correlation between elements within each process is analyzed using the support vector machine algorithm. By calculating the correlation matrix between elements, the dependency relationship between elements is determined. By analyzing the dependencies, the radiation source items contribution and structural change indicators in the standardized feature dataset are grouped. If the correlation of a certain feature is lower than the preset threshold, it is classified as an independent feature, and the grouped feature dataset is obtained. Based on the grouped element dataset, a distribution model of elements within each process is generated. By recording the distribution characteristics of independent and related elements, the final element analysis results are determined.

[0009] In one aspect of the invention, the step of performing multiple random samplings of a standardized element dataset using Monte Carlo simulation to simulate the mutual constraints between elements within a process and obtain the distribution of coupling strength between elements includes: The standardized element dataset was randomly sampled using the Monte Carlo simulation method, and multiple simulations were performed on the distribution characteristics of elements within the process to obtain a preliminary set of constraints between elements. For the initial set of constraints between elements, data processing techniques are used to classify and organize the sampling analysis results, focusing on the mutually constraining internal elements, and determining the initial coupling strength distribution between elements; Based on the preliminary coupling strength distribution, the strength analysis data between each element within the process is obtained. If the coupling strength of a certain element is lower than the preset threshold, it is classified as a weakly correlated element, and the strength distribution results after classification are obtained. Based on the intensity distribution results after classification, the strongly correlated and weakly correlated elements are grouped using relationship construction techniques to construct a detailed mapping of mutual constraints between elements, thereby obtaining a set of grouped element relationships. Based on the grouped set of element relationships, statistical tools are used to further refine the distribution results of coupling strength, focusing on the significant constraints within the process, and determining the final strength distribution model; By using the final intensity distribution model, the mutual constraints of each internal element under different process scenarios are obtained. The distribution results are then verified using simulation methods to obtain a complete dataset for element coupling analysis.

[0010] In one aspect of the invention, the step of constructing a causal chain diagram model based on the coupling strength distribution between elements, identifying the amplification effect of key micro-operations on subsequent stages, and determining a dynamic weight allocation matrix includes: The causal relationship is extracted from the coupling strength distribution using the causal discovery method to obtain the causal chain diagram model within the process. Based on the pointing relationships between nodes in the causal chain graph model, the number of outgoing edges and incoming edges of each node are counted to obtain the outgoing degree set and the incoming degree set of the nodes. For a set of nodes with high out-degree values, determine whether their downstream path length exceeds a preset level. If it does, mark them as key micro-operation nodes to obtain a set of key micro-operation nodes. Extract all directed paths from key micro-operation nodes to downstream nodes from the causal chain graph model to obtain the path set; For each path, the transmission strength values ​​of each edge on the path are accumulated to obtain the set of accumulated transmission strengths of the key micro-operation node to each downstream node. Based on the strength values ​​of each downstream node in the cumulative transmission strength set, all downstream nodes of the same key micro-operation node are sorted by strength to obtain the sorted downstream node sequence. A dynamic weight allocation matrix for the key micro-operation node is constructed using the sorted downstream node sequence and the corresponding cumulative transmission intensity value. The matrix rows correspond to the key nodes, the columns correspond to the downstream links, and the elements are normalized weight coefficients.

[0011] In one aspect of the invention, the method of training the dynamic weight allocation matrix using a neural network, inputting node features of a causal chain graph model, and outputting an adjusted weight value sequence, is used to capture changes in weights as a function of field conditions, including: Obtain the feature vectors of all nodes in the causal chain graph model; The above feature vectors are received by a neural network and processed to obtain a sequence of weight values ​​calculated by the neural network. Establish a correspondence between each element in the weight value sequence and the on-site operating status to determine which category the on-site operating status belongs to; If it belongs to the first type of state, the first set of pre-stored weight adjustment coefficients will be used; If it belongs to the second type of state, the second set of pre-stored weight adjustment coefficients will be used; After determining the applicable weight adjustment coefficients, the original weight value sequence is corrected to obtain the corrected final weight value sequence; Write the corrected final weight value sequence back to the corresponding position in the dynamic weight allocation matrix.

[0012] In one aspect of the invention, if the weight value sequence exceeds a preset threshold, the amplification relationship analysis module is triggered to extract high-impact sub-operations from the weight value sequence to obtain a micro-operation influence vector, including: Obtain the feature vectors of all current nodes in the causal chain graph; The weighted sequence is obtained by processing the feature vectors using a neural network model. A numerical comparison is performed between each element in the weighted adjustment sequence and a preset threshold. If an element exceeding a preset threshold appears in the weighted adjustment sequence, the amplification relationship analysis module will be executed. A set of high-impact sub-operations is obtained by extracting high-value elements exceeding a preset threshold from the weighted adjustment sequence. The influence vectors of the corresponding micro-operations are obtained by analyzing each of the high-influence sub-operations; The micro-operation influence vector is stored in a specified location to complete the amplification relationship processing flow.

[0013] In one aspect of the invention, the step of generating a multi-level coupled prediction model by mapping the microscopic operational influence vector to the macroscopic process target, and fusing pollution distribution and structurally complex parameters to determine the overall radiation dose estimate, includes: Obtain relevant data on micro-operations, extract the corresponding influence vector for each micro-operation, store it as an initial vector set, and obtain a preliminarily organized vector dataset; By using an initial vector set and combining it with the target data corresponding to the macro-process, and employing pre-established mapping rules, the vector dataset is associated with the process target to determine the basic framework of multi-layer coupling. Based on the multi-layer coupling framework, relevant information about pollution distribution is introduced. If the pollution distribution data exceeds the preset threshold range, the vectors within the framework are weighted and adjusted to obtain the adjusted coupling structure. For the adjusted coupling structure, combined with the complex parameter data, the parameter data is integrated into the coupling structure through a logical matching method to determine the weight of the impact of structural complexity on the overall estimation. Starting from the influence weights, a prediction model is constructed, and the support vector machine algorithm is used to process the adjusted coupling structure and weight data to obtain a preliminary estimate of the radiation dose. Based on the preliminary estimation results and combined with the overall estimation business requirements, the data is calibrated a second time. If the calibrated data deviates from the preset range, the estimation results are fine-tuned to determine the final radiation dose estimate. The final radiation dose estimate is used to store the processed data in a designated database, and the storage path and index information are obtained to complete the recording of the entire estimation process.

[0014] In one aspect of the invention, the expansion from overall radiation dose estimation to prediction of working hours and waste volume, the use of an iterative optimization algorithm to update the parameters of the multi-level coupled prediction model, and the resulting final numerical prediction results for decommissioning and remediation decision-making include: By extracting basic information from radiation dose data, and for each data record, obtaining the correlation information between the corresponding working hours prediction and the total amount of waste, a preliminary comprehensive dataset is determined. Based on the preliminary comprehensive dataset, the radiation dose, working hour prediction, and total waste amount within the dataset are classified and organized using a pre-established hierarchical structure to obtain structured data groups. For structured data grouping, relevant business rules for decommissioning management are introduced. If the data records in a group do not match the preset threshold range, the data in the group is redistributed, and the adjusted data grouping is determined. Based on the adjusted data grouping, an iterative optimization algorithm is used to process the parameter updates within the coupled model and obtain the updated model parameter set. By using the updated model parameter set, calculations are performed on the predicted working hours and total waste data within the prediction range to determine the comprehensive numerical prediction results. Based on the comprehensive numerical prediction results and combined with the decision support needs of decommissioning management, the results data are formatted and stored to obtain the final decision reference data.

[0015] In another aspect, the present invention also relates to a nuclear facility decommissioning management system based on a generative large model, the system comprising: The data acquisition and standardization module is used to acquire on-site sensor data and historical decommissioning records, extract the initial parameter values ​​of internal elements of each process, including radiation source contribution and structural change indicators, and obtain a standardized element dataset after removing noise using data cleaning methods. The Monte Carlo simulation module is used to perform multiple random samplings on a standardized element dataset through Monte Carlo simulation, simulate the mutual constraints between elements within a process, and obtain the distribution of coupling strength between elements. The causal chain construction module is used to construct a causal chain diagram model based on the distribution of coupling strength between elements, identify the amplification effect of key micro-operations on subsequent links, and determine the dynamic weight allocation matrix. The neural network training module is used to train the dynamic weight allocation matrix using a neural network. It takes the node features of the causal chain graph model as input and outputs the adjusted weight value sequence to capture the changes in weights as the field conditions change. The amplified relationship analysis module is used to extract high-impact sub-operations from the weight value sequence if the weight value sequence exceeds a preset threshold, thereby obtaining the micro-operation influence vector. The multi-level coupled prediction module is used to generate a multi-level coupled prediction model by mapping the micro-operational influence vector with the macro-process target, and to determine the overall radiation dose estimate by integrating pollution distribution and structurally complex parameters. The iterative optimization and decision-making module is used to expand from the overall radiation dose estimate to the prediction of working hours and waste volume. It uses an iterative optimization algorithm to update the parameters of the multi-level coupled prediction model and obtain the final numerical prediction results for decommissioning and management decisions.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention constructs an element dataset by cleaning and standardizing on-site and historical data, and uses Monte Carlo simulation to reveal the coupling strength between elements. It then builds a causal chain graph model to identify the amplification effect of key operations. This invention trains a dynamic weight allocation matrix using a neural network to capture the characteristics of weight changes under conditions, and triggers amplification relationship analysis when weights exceed a threshold, extracting high-impact operation vectors. Finally, it generates a multi-level coupled prediction model, fusing pollution distribution and structural parameters to accurately predict radiation dose, working hours, and waste volume. This invention continuously updates model parameters through iterative optimization algorithms to ensure the adaptability and accuracy of prediction results, providing a scientific basis for decommissioning and remediation decisions, and significantly improving remediation efficiency and safety. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is one of the flowcharts for the nuclear facility decommissioning management method based on a generative large model according to the present invention.

[0019] Figure 2 This is the second flowchart of the nuclear facility decommissioning management method based on a generative large model according to the present invention.

[0020] Figure 3 This is the third flowchart of the nuclear facility decommissioning management method based on a generative large model according to the present invention. Detailed Implementation

[0021] The present invention will be further described below with reference to embodiments. These embodiments are merely some, not all, of the embodiments of the present invention. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the protection scope of the present invention.

[0022] Please see Figures 1-3 As shown in the figure, this embodiment discloses a method and system for the decommissioning and management of nuclear facilities based on a generative large model, wherein the method may specifically include: Step 101: Obtain on-site sensor data and historical decommissioning records, extract the initial parameter values ​​of internal elements of each process, including radiation source contribution and structural change indicators, and obtain a standardized element dataset after removing noise using data cleaning methods.

[0023] Data acquisition involves obtaining on-site sensor data and historical decommissioning records. A data acquisition system records and stores relevant information from each process in real time, resulting in a complete raw dataset. For this raw dataset, data cleaning methods are used to remove outliers and missing values ​​from the sensor data and historical records. Valid data is filtered through a preset threshold range to determine the cleaned preliminary dataset. Initial parameter values ​​for internal elements of each process are extracted from the cleaned preliminary dataset, covering radiation source contributions and structural change indicators. The data is then classified and organized using field mapping methods to obtain a categorized parameter dataset. For the categorized parameter dataset, standardization techniques are applied to normalize the values ​​of radiation source contributions and structural change indicators. A standardized element dataset is constructed using a unified dimensional transformation method. Based on the standardized element dataset, the correlation between internal elements of each process is analyzed using a support vector machine algorithm. The correlation matrix between elements is calculated to determine their dependencies. Based on the obtained dependencies, the radiation source contributions and structural change indicators in the standardized element dataset are grouped. If the correlation of a certain element is below a preset threshold, it is classified as an independent element, resulting in a grouped element dataset. Based on the grouped element dataset, a distribution model of elements within each process is generated. By recording the distribution characteristics of independent and related elements, the final element analysis results are determined.

[0024] Specifically, in acquiring on-site sensor data and historical decommissioning records, the process begins by using a sensor network deployed at the nuclear facility decommissioning site to collect radiation dose rate data in real time. For example, the average radiation value recorded by 10 sensors in a certain area over 24 hours is 0.25 microsieverts per hour. Simultaneously, decommissioning records from the past five years are extracted from the database, including data such as equipment dismantling time and radiation source term distribution. For instance, when a certain piece of equipment was decommissioned, the cobalt-60 source term contribution was recorded as 60%. Next, the initial parameter values ​​of each process element are extracted, and the radiation source term contribution is decomposed using an algorithm. Assuming the total radiation source term is 100%, the contribution ratio of each nuclide is calculated using a weighted average method. For example, cobalt-60 accounts for 60%, cesium-137 accounts for 30%, and other nuclides account for 10%. Combined with structural change indicators, such as a 20% volume reduction rate after equipment dismantling, regression analysis predicts a structural stability impact factor of 0.85. Subsequently, data cleaning methods were employed to remove noise. For outliers in the sensor data, such as a sudden spike to 10 microsieverts per hour (µSv) significantly deviating from the mean, a median filtering algorithm was used to replace it with the median of 0.24 µSv per hour from the five nearest time points. Simultaneously, missing values ​​in historical data were filled using linear interpolation; if a record for a particular day was missing, interpolation was performed based on data from the two days before and after it. Finally, a standardized feature dataset was obtained. Z-score standardization was used to convert radiation values ​​and structural change indicators into dimensionless data with a mean of 0 and a standard deviation of 1. For example, the original radiation value of 0.25 µSv per hour was standardized to 0.3, and the 20% structural reduction rate was standardized to 0.5, ensuring consistency in subsequent analyses. This entire process—data acquisition, cleaning, and standardization—was automated using scripts. Combined with business logic, such as the correlation analysis between radiation source items and structural changes, a complete data processing chain was formed, ensuring the logical rigor from raw data to the standardized dataset.

[0025] Step 102: The standardized element dataset is randomly sampled multiple times using Monte Carlo simulation to simulate the mutual constraints between elements within the process and obtain the distribution of coupling strength between elements.

[0026] A Monte Carlo simulation method was used to randomly sample a standardized element dataset. Multiple simulations were performed to analyze the distribution characteristics of elements within a process, resulting in a preliminary set of constraints between elements. Data processing techniques were then used to classify and organize the sampling analysis results, focusing on mutually constraining internal elements to determine the initial coupling strength distribution. Based on this initial coupling strength distribution, strength analysis data between elements within the process was obtained. If the coupling strength of an element was below a preset threshold, it was classified as a weakly correlated element, resulting in a classified strength distribution. For the classified strength distribution, relationship construction techniques were used to group strongly correlated and weakly correlated elements, constructing a detailed mapping of mutual constraints between elements, resulting in a grouped set of element relationships. Based on this grouped set of element relationships, statistical tools were used to further refine the distribution of coupling strength, focusing on significant constraints within the process, and determining the final strength distribution model. Using the final strength distribution model, the mutual constraint characteristics of each internal element under different process scenarios were obtained. The distribution results were then validated using simulation methods, resulting in a complete element coupling analysis dataset.

[0027] Specifically, in the risk assessment of nuclear facility decommissioning procedures, a standardized dataset of elements was sampled 50,000 times using Monte Carlo simulation. In each iteration, key variables such as radiation dose rate, source term contribution ratio, and structural degradation rate were subjected to random perturbations consistent with their distribution characteristics. For example, the radiation dose rate followed a log-normal distribution (mean taken as the original value after standardization, 0.28 μSv / h, standard deviation 0.07), the source term contribution ratio adopted a Dirichlet distribution (α parameter set to 12 for Cobalt-60, 8 for Cesium-137, and 3 for others based on historical frequency), and the structural degradation rate used a triangular distribution (minimum 0.12, mode 0.18, maximum 0.27). After each sampling, the coupling effect strength between elements was calculated, specifically using a combination of Pearson correlation coefficient and partial correlation coefficient. Emphasis was placed on the constraint relationship between increased radiation dose rate and accelerated structural degradation rate; for instance, when the radiation dose rate perturbation exceeded 1.5 standard deviations, the average increase in structural degradation rate reached 1.62 times the original predicted value. By summarizing all iteration results, the coupling strength distribution characteristics were obtained: the 95% confidence interval for the coupling strength between radiation and structural degradation was [0.71, 0.94], with a median of 0.83, exhibiting a right-skewed distribution, indicating that the constraint effect is more easily amplified under extreme conditions. Simultaneously, sensitivity ranking was extracted, revealing that the contribution ratio of Cobalt-60 had the highest weight on the overall coupling strength, reaching 38.6%, followed by the initial radiation dose rate at 29.2%. The entire simulation process was automatically completed using a parallel computing framework, forming a reliable probabilistic profile of the dynamic constraint relationships between elements, providing a quantitative basis for subsequent decommissioning process optimization.

[0028] Step 103: Based on the distribution of coupling strength among elements, construct a causal chain diagram model, identify the amplification effect of key micro-operations on subsequent links, and determine the dynamic weight allocation matrix.

[0029] A causal discovery method is used to extract directed causal relationships from the coupling strength distribution, resulting in a causal chain graph model within the process. Based on the pointing relationships of nodes in the causal chain graph model, the number of outgoing and incoming edges for each node is counted, yielding the out-degree and in-degree sets of nodes. For nodes with high out-degree values, it is determined whether their downstream path length exceeds a preset level; if so, they are marked as critical micro-operation nodes, resulting in a set of critical micro-operation nodes. All directed paths from the critical micro-operation nodes to their downstream nodes are extracted from the causal chain graph model, resulting in a path set. For each path, the transmission strength values ​​of each edge are accumulated, resulting in a cumulative transmission strength set from the critical micro-operation node to each downstream node. Based on the strength values ​​corresponding to each downstream node in the cumulative transmission strength set, all downstream nodes of the same critical micro-operation node are sorted by strength, resulting in a sorted downstream node sequence. Using the sorted downstream node sequence and the corresponding cumulative transmission strength values, a dynamic weight allocation matrix for the critical micro-operation node is constructed, where rows correspond to critical nodes, columns correspond to downstream links, and elements are normalized weight coefficients.

[0030] Specifically, in the field of risk analysis for nuclear facility decommissioning procedures, based on the analysis results of the coupling strength distribution between elements, the system identifies the amplification effect of key micro-operations on subsequent stages by constructing a causal chain graph model and determines a dynamic weight allocation matrix. First, the system uses a Bayesian network algorithm, taking the coupling strength distribution data as input, to automatically construct a causal chain graph. Nodes represent key elements in the process, such as equipment aging index (initial value set at 0.35), environmental corrosion rate (mean value 0.22 mm / year), and operating load fluctuation (standard deviation 0.15). Edges represent the causal dependencies between elements, and weights are calculated using historical data; for example, the influence weight of the equipment aging index on the corrosion rate is 0.48. Next, the system identifies the amplification effect through a path analysis algorithm, calculating the influence transmission path from micro-operations to subsequent stages. It finds that when the operating load fluctuation exceeds 0.2, the corrosion rate can increase by up to 1.35 times the original value, with a path influence coefficient of 0.67. Subsequently, the system employed principal component analysis to extract key influencing factors and determine a dynamic weight allocation matrix. The calculation results showed that the equipment aging index accounted for 42.3% of the weight, operating load fluctuation for 31.5%, and environmental corrosion rate for 26.2%. This matrix was stored as the basis for subsequent analysis. To ensure logical rigor, the system also correlated the temperature control variable in the process (set to a range of 22-28 degrees Celsius). Regression analysis revealed that for every 1 degree Celsius increase in temperature, the corrosion rate increased by 0.03 mm / year, thus supplementing the completeness of the causal chain. The entire process was automated, with data processing and model building completed on a distributed computing platform, ensuring the efficiency and accuracy of the analysis.

[0031] Step 104: The dynamic weight allocation matrix is ​​trained using a neural network. The node features of the causal chain graph model are input, and the adjusted weight value sequence is output to capture the changes in weights as the field conditions change.

[0032] The feature vectors of all nodes in the causal chain graph model are obtained. These feature vectors are then processed by a neural network to obtain a weight value sequence calculated by the neural network. For each element in the weight value sequence, a correspondence is established with the on-site operating state to determine which category the on-site operating state belongs to. If it belongs to the first category, the first set of pre-stored weight adjustment coefficients is used; if it belongs to the second category, the second set of pre-stored weight adjustment coefficients is used. After determining the currently applicable weight adjustment coefficients, the original weight value sequence is corrected to obtain the corrected final weight value sequence. The corrected final weight value sequence is then written back to the corresponding position in the dynamic weight allocation matrix. Specifically, in the field of risk analysis for nuclear facility decommissioning procedures, the system employs a deep neural network to optimize and train the dynamic weight allocation matrix. First, feature vectors for each node are extracted from the constructed causal chain graph, including equipment aging index (current value 0.41), cumulative radiation dose rate (average 0.78 mSv / h), vibration amplitude standard deviation (0.09 mm / s), and humidity fluctuation range (32%-68%), generating a total of 168-dimensional feature input vectors. Subsequently, the system constructs a fully connected neural network with three hidden layers. The first hidden layer has 128 neurons using the ReLU activation function, the second layer has 64 neurons, and the third layer uses softmax to output the dynamic weight value sequence corresponding to each element. During training, the probability of actual risk events occurring under historical decommissioning conditions is used as the supervision label. The Adam optimizer (learning rate 0.001) is used, the cross-entropy loss function is selected, the batch size is set to 32, and a total of 120 epochs are performed. During training, it was found that when the cumulative radiation dose rate exceeded 0.95 mSv / h, the model's sensitivity to the vibration amplitude weights increased significantly. After convergence, the adjusted dynamic weight sequence was obtained: equipment aging index weight 0.393, cumulative radiation dose rate weight 0.362, vibration amplitude weight 0.158, and humidity fluctuation weight 0.087. To verify the model's adaptability, the system was additionally input with a set of real-time field data (radiation dose rate suddenly rising to 1.12 mSv / h). The network output weights were immediately adjusted to 0.374, 0.418, 0.139, and 0.069, reflecting the changes in field conditions where radiation was the dominant risk. This weight sequence was directly used in the subsequent risk quantification calculation module, achieving accurate allocation that adapts to environmental and operating conditions.

[0033] Step 105: If the weight value sequence exceeds the preset threshold, the amplification relationship analysis module is triggered to extract high-influence sub-operations from the weight value sequence and obtain the micro-operation influence vector.

[0034] The process involves obtaining the feature vectors of all current nodes in the causal chain graph, processing these feature vectors using a neural network model to obtain a weight adjustment sequence, comparing each element in the weight adjustment sequence with a preset threshold, and executing the amplification relationship analysis module to extract high-value elements exceeding the preset threshold from the weight adjustment sequence. This yields a set of high-influence sub-operations, which are then analyzed one by one to obtain the corresponding micro-operation influence vectors. These micro-operation influence vectors are then stored in a designated location, completing the amplification relationship processing flow. Specifically, in the risk analysis of nuclear facility decommissioning procedures, the system first collects multi-source data on current operating conditions through a real-time sensor network. This includes data such as the cutting equipment speed fluctuation coefficient (0.27), the peak radon concentration in the air (41.3 Bq / m³), the surface contamination activity measurement (8.9 × 10³ Bq / cm²), and the hoisting cable tension deviation rate (4.2%). After integration by the feature engineering module, a 192-dimensional high-dimensional input vector is formed. Subsequently, a hybrid model combining a bidirectional long short-term memory network and an attention mechanism is constructed. The BiLSTM layer contains two layers with 96 hidden units each, and the attention layer adopts a multi-head self-attention mechanism (8 heads). Finally, the model is mapped to the dynamic weight distribution of each risk element through a fully connected layer, and the output layer uses a softmax function to ensure that the weight sum is 1. The model was trained using an RMSprop optimizer with time decay (initial learning rate 0.0008, decay rate 0.96), with the actual occurrence frequency of segmented risk events in the past three years of decommissioning operations as the supervision signal. A weighted negative log-likelihood loss function was used, with a batch size of 24, and training lasted for 150 epochs. Later in the training process, it was observed that when the radon concentration exceeded 38 Bq / m³, the model's weight for cable tension deviation rapidly increased to above 0.241. The baseline dynamic weight sequence obtained after convergence was: rotational speed fluctuation coefficient weight 0.318, radon concentration weight 0.364, surface contamination activity weight 0.196, and cable tension deviation weight 0.122. When the system received a real-time data update for a particular operation (radon concentration suddenly increased to 52.7 Bq / m³), the model immediately inferred and adjusted the weights: 0.284, 0.429, 0.171, and 0.116. At this point, the radon concentration weight in the weight sequence has exceeded the preset threshold of 0.40. The system automatically triggers the amplified relationship analysis module to extract the top two high-impact sub-operations (radon concentration-related cutting and ventilation sub-operations) from the current weight vector, generating an impact vector containing 12 micro-operation dimensions for subsequent refined risk propagation path deduction and intelligent recommendation of key control points.

[0035] Step 106: By mapping the micro-operational influence vector to the macro-process target, a multi-level coupled prediction model is generated, which integrates pollution distribution and structural complexity parameters to determine the overall radiation dose estimate.

[0036] Relevant data on micro-level operations are acquired. For each micro-level operation, its corresponding influence vector is extracted and stored as an initial vector set, resulting in a pre-organized vector dataset. Using this initial vector set and the target data corresponding to macro-level processes, a pre-established mapping rule is employed to associate the vector dataset with the process targets, establishing a multi-layered coupling framework. Based on this framework, information on pollution distribution is introduced. If the pollution distribution data exceeds a preset threshold, the vectors within the framework are weighted and adjusted to obtain an adjusted coupling structure. For the adjusted coupling structure, combined with complex parameter data, a logical matching method is used to integrate the parameter data into the coupling structure, determining the influence weight of structural complexity on the overall estimation. Starting from the influence weight, a prediction model is constructed, and a support vector machine algorithm is used to process the adjusted coupling structure and weight data to obtain a preliminary estimation result of the radiation dose. Based on the preliminary estimation result and the overall estimation business requirements, the data undergoes secondary calibration. If the calibrated data deviates from the preset range, the estimation result is fine-tuned to determine the final radiation dose estimate. The final radiation dose estimate is used to store the processed data in a designated database, and the storage path and index information are obtained to complete the recording of the entire estimation process.

[0037] Specifically, in the field of risk analysis for nuclear facility decommissioning procedures, the system performs cross-scale mapping between the micro-level operational impact vectors obtained in the early stages and the overall macro-level operational objectives. First, a multi-level coupled prediction model is constructed using a graph neural network, where nodes represent operational elements at each level, and edge weights are obtained by mapping the micro-level impact vectors using a sigmoid activation function. Inter-layer aggregation employs a GraphSAGE method with a gating mechanism, and the hidden layer dimension is uniformly set to 128. The model simultaneously integrates structural complexity parameters (complexity index calculated to be 3.47) obtained from 3D laser scanning of the decommissioning site and historical contamination distribution grid data (a total of 840 spatial cells, with an average contamination concentration gradient of 12.6 Bq / cm³ / m). Spatial contamination spatial correlation features are extracted through spatial convolutional layers (3×3 kernels, stride 1), and these features are fused with the micro-level operational impact vectors at the feature level in the fourth layer. Finally, the prediction head adopts a three-layer perceptron structure, with the last layer using linear activation to output an estimated overall radiation dose. The training process used the AdamW optimizer (weight decay 0.01, initial learning rate 0.0006), with past actual personal dose equivalent monitoring data (range 0.12~7.84 mSv) as the ground truth. The Huber loss function (δ=1.0) was used to suppress the influence of extreme values. The system was trained for 180 epochs, and the mean absolute error on the validation set converged to 0.19 mSv. In a typical decommissioning and cutting operation, when the micro-influence vector showed a sharp increase in the ventilation failure factor to 0.682, the model predicted that the overall effective dose equivalent would rapidly increase from the baseline value of 0.47 mSv to 1.93 mSv. Based on this, the system determined that the current operating condition had entered a high-risk state and automatically generated a targeted intervention priority ranking.

[0038] Step 107: Expand the overall radiation dose estimate to predict working hours and waste volume, and use an iterative optimization algorithm to update the parameters of the multi-level coupled prediction model to obtain the final numerical prediction results for decommissioning and management decisions.

[0039] By extracting basic information from radiation dose data, and for each data record, the correlation information between its corresponding work hour prediction and total waste amount is obtained to determine a preliminary comprehensive dataset. Based on this preliminary comprehensive dataset, a pre-established hierarchical structure is used to classify and organize the radiation dose, work hour prediction, and total waste amount within the dataset, resulting in structured data groups. For these structured data groups, relevant business rules for decommissioning management are introduced. If a data record within a group does not match a preset threshold range, the data within that group is reassigned to determine the adjusted data group. Based on the adjusted data group, an iterative optimization algorithm is used to update the parameters within the coupled model, obtaining an updated set of model parameters. Using the updated set of model parameters, calculations are performed on the work hour prediction and total waste amount data within the prediction range to determine the comprehensive numerical prediction results. Based on the comprehensive numerical prediction results and the decision support requirements for decommissioning management, the result data is formatted and stored to obtain the final decision reference data.

[0040] Specifically, in the planning and resource estimation of nuclear facility decommissioning procedures, the system first performs cross-scale correlation mapping between the micro-operational influence vector and the overall macro-operational objective, constructing a multi-level coupled prediction model that includes both working hours and waste volume as dual objectives. In this model, graph structure nodes represent work units and material types at each level, and edge weights are generated by normalizing the micro-operational parameters using the tanh activation function. Inter-layer information transfer employs a Graph Attention network with residual connections, and the hidden layer dimension is set to 96. The model also incorporates the structural disassembly difficulty coefficient (calculated value 4.12) obtained from the reconstruction of the 3D point cloud at the site and historical waste classification statistics (a total of 672 record units, with an average metal waste volume ratio of 61.3% and a concrete waste ratio of 34.8%). A one-dimensional temporal convolutional layer (kernel size 7, expansion rate 2) captures the temporal dependency features of the procedures, and in the third layer, it completes channel-level splicing and fusion with the micro-influence vector. The prediction head is designed as a dual-branch structure with a shared underlying layer. The working time prediction branch uses two fully connected layers with ReLU activation, and finally outputs the estimated working time value linearly. The waste quantity prediction branch adds a softmax normalization layer to output the volume ratio of various types of waste, and finally outputs the total waste quantity linearly. Training uses the Adam optimizer (weight decay 0.008, initial learning rate 0.0008), using historical real working time records (range 18.4~156.7 hours) and actual waste weighing data (total mass range 2.13~41.6 tons) as supervised ground truth. Smoothing L1 loss combined with KL divergence is used for joint optimization. A total of 240 iterations were performed. The absolute percentage error of working time on the validation set stabilized at 7.8%, and the average relative error of waste quantity prediction converged to 6.2%. In a typical reactor pressure vessel segmentation and decommissioning task, when the micro-influence vector showed that the equipment aging factor rose to 0.714 and the cutting efficiency decreased significantly, the model iterated and updated the parameters, predicting that the total working time would increase from the baseline value of 42.6 hours to 81.3 hours. At the same time, the total amount of waste was adjusted from the initial estimate of 17.4 tons to 29.8 tons, of which the proportion of high-level radioactive waste increased from 4.1% to 11.7%. Based on this, the system automatically generated suggestions for adjusting the work period and a dynamic allocation plan for temporary waste storage space, providing a quantitative basis for decommissioning and management decisions.

[0041] This invention provides a nuclear facility decommissioning management system based on a generative large model, which mainly includes: The data acquisition and standardization module is used to acquire on-site sensor data and historical decommissioning records, extract the initial parameter values ​​of internal elements of each process, including radiation source contribution and structural change indicators, and obtain a standardized element dataset after removing noise using data cleaning methods. The Monte Carlo simulation module is used to perform multiple random samplings on a standardized element dataset through Monte Carlo simulation, simulate the mutual constraints between elements within a process, and obtain the distribution of coupling strength between elements. The causal chain construction module is used to construct a causal chain diagram model based on the distribution of coupling strength between elements, identify the amplification effect of key micro-operations on subsequent links, and determine the dynamic weight allocation matrix. The neural network training module is used to train the dynamic weight allocation matrix using a neural network. It takes the node features of the causal chain graph model as input and outputs the adjusted weight value sequence to capture the changes in weights as the field conditions change. The amplified relationship analysis module is used to extract high-impact sub-operations from the weight value sequence if the weight value sequence exceeds a preset threshold, thereby obtaining the micro-operation influence vector. The multi-level coupled prediction module is used to generate a multi-level coupled prediction model by mapping the micro-operational influence vector with the macro-process target, and to determine the overall radiation dose estimate by integrating pollution distribution and structurally complex parameters. The iterative optimization and decision-making module is used to expand from the overall radiation dose estimate to the prediction of working hours and waste volume. It uses an iterative optimization algorithm to update the parameters of the multi-level coupled prediction model and obtain the final numerical prediction results for decommissioning and management decisions.

[0042] In some different embodiments, the present invention proposes further solutions to the problem of deviations in terminology and specifications in general large models in the prior art; In this embodiment, the specific method for solving the above problem is as follows: Users input their natural language requirements through the front-end interface, and the system automatically categorizes them into types such as solution generation, fault handling, or compliance verification. Specifically, users input natural language questions or requests related to nuclear facility decommissioning and remediation through a dedicated interactive interface (e.g., "Generate a reactor pressure vessel dismantling plan" or "How to handle excessive radiation doses during radioactive soil decontamination"). The system front-end integrates a question classification module, which, based on predefined rules or a lightweight classifier, classifies input requests in real time into categories such as "plan generation," "fault diagnosis and handling," "regulatory and compliance verification," or "knowledge Q&A," providing guidance for subsequent differentiated processing procedures.

[0043] The system performs data cleaning, Chinese word segmentation, part-of-speech tagging, and stop word removal on the input text, and embeds a nuclear decommissioning domain entity thesaurus to ensure accurate segmentation of professional terms; Specifically, this step is crucial for ensuring the accuracy of subsequent analysis. First, the input text undergoes data cleaning, using regular expressions to remove noise such as HTML tags and irrelevant symbols. The core component is Chinese word segmentation. The system employs the Jieba word segmentation tool and deeply integrates a nuclear decommissioning entity lexicon. This lexicon (DEL lexicon) is built based on historical construction logs, environmental impact assessment reports, and national standards (such as GB 18871), covering four main categories of terms: equipment (such as "gamma-ray dose rate detector"), processes (such as "low-level radioactive waste coagulation and sedimentation process"), radionuclides (such as "Cs-137"), and standards. By setting the priority of the DEL lexicon above the general lexicon in the code, it ensures that complex terms such as "non-standard grab bucket long rod sampler" are completely segmented, avoiding incorrect segmentation.

[0044] Specifically, for potential polysemous terms or abbreviations, the Xgboost algorithm is used for word sense disambiguation, and its prediction model is as follows:

[0045] Where t represents the number of decision trees; f k represents the prediction result of the k-th decision tree; represents the predicted value of instance e in ambiguity l after the i-th iteration; Ft(e) is the feature matrix of instance e in ambiguity l.

[0046] By optimizing the objective function:

[0047] Where ye is the actual value; Ω(ft) is the minimum regularization term.

[0048]

[0049] Where J is the number of leaf nodes in each decision tree; wtj is the optimal value of the j-th leaf node; and γ and λ are the corresponding coefficients.

[0050] Performing a second-order Taylor expansion on the above equation and simplifying it, we obtain the objective function:

[0051] And calculate the characteristic split gain:

[0052] Among them, GL and HL are respectively the sum of the first-order and second-order gradients of all samples in the left leaf node after splitting. GR and HR are respectively the sum of the first-order and second-order gradients of all samples in the right leaf node after splitting. When performing node splitting, all the features of the nuclear facility decommissioning governance ambiguous words are traversed, the gain value Gain after feature splitting is calculated, and the feature with the maximum gain is selected as the splitting feature of the current node. Using the forward stepwise algorithm, the initial parameters are initialized, and the optimal model parameters are obtained through continuous iteration. The Xgboost algorithm finally accumulates the prediction values of all decision trees to obtain the prediction result of the ambiguous word l.

[0053] The processed text is transformed into a low-dimensional dense vector, and the bidirectional long short-term memory network is used to extract the context time series features of the text.

[0054] Specifically, the text sequence after word segmentation and annotation is transformed into a numerical form that can be processed by machines. Using distributed representation (such as Word2Vec or pre-trained embeddings), each word is mapped to a low-dimensional, dense real-valued vector (for example, 768 dimensions), so that words with similar semantics are closer in the vector space. Subsequently, the bidirectional long short-term memory network is used to process the word vector sequence. BiLSTM can capture the long-term context dependencies of the text from both the forward and backward directions, effectively learn the sequence associations such as "Fault phenomenon: Jam" and "Fault location: Robot arm", and splice the final features in both directions to form a comprehensive feature representation rich in context information.

[0055] Adopt a named entity recognition model that integrates BERT, BiGRU, attention mechanism and conditional random field to accurately identify entities in the nuclear decommissioning field such as equipment, processes, and standards in the text.

[0056] Specifically, to accurately extract the key domain elements in the text, a BERT-BiGRU-MHA-CRF hybrid neural network model is adopted. First, the pre-trained BERT model is used to obtain the deep context-related vector representation of each character or word. Then, the sequence features are further extracted through the bidirectional gated recurrent unit network. In order to focus on the keywords that contribute more to entity recognition (such as paying attention to "Jam" and "Robot arm" rather than "of" and "了" in the text describing faults), the multi-head attention mechanism is introduced to calculate the attention weights at different positions. Finally, the features are input into the conditional random field layer. CRF learns the transition rules between labels (such as the "B-equipment" label should be followed by "I-equipment" or "E-equipment", rather than "O"), and outputs the globally optimal entity label sequence for the entire sentence, accurately identifying entity types such as equipment, processes, parameters, and standards.

[0057] Based on the pre-trained topic model, the core intention of the user's question is recognized, and it is associated with specific nuclear decommissioning business topics through topic matching.

[0058] Specifically, to address the challenges of potentially lengthy and implicit user questions, the system employs a BERTopic-GMM pre-trained topic model for intent clustering and matching. The process is as follows: 1) Encode the preprocessed question into a sentence vector using Sentence-BERT; 2) Reduce the vector dimension using the UMAP dimensionality reduction algorithm while preserving semantic association; 3) Perform preliminary density clustering using HDBSCAN to obtain several fine-grained topics; 4) For the business logic in the nuclear decommissioning domain, use a Gaussian mixture model to perform secondary aggregation and optimization on the preliminary clustering results, merging similar subtopics (e.g., merging "grab sampling" and "long pole sampling" under the topic "sampling technology") to form a stable and interpretable topic distribution; 5) Extract representative keywords for each topic using the c-TF-IDF algorithm. Finally, user questions are mapped to specific business topics such as "reactor pressure vessel dismantling," "waste liquid treatment process adjustment," or "radiation safety compliance verification."

[0059] Based on the identified intents and entities, relevant structured knowledge such as regulations, cases, and process parameters are retrieved from the domain knowledge base.

[0060] Specifically, the system maintains a structured knowledge base for the nuclear decommissioning domain, including legal provisions, historical cases, equipment parameters, and process flow diagrams. Based on the entities identified in step four and the topics determined in step five, the system automatically constructs Cypher queries (suitable for graph databases) or similarity queries to retrieve relevant fragments from the knowledge base. The retrieval process combines semantic similarity calculation (such as cosine similarity, with a threshold set to 0.7) and entity linking technology to ensure that the returned knowledge not only matches literally but is also highly semantically relevant to the user's question, providing an accurate and reliable basis for generating answers.

[0061] The lightweight generative large model combines retrieved knowledge to generate preliminary answers and performs automatic compliance checks based on domain thesaurus and regulatory database.

[0062] Specifically, the user's original question, the identified entities / intents, and the retrieved relevant knowledge fragments are input into the lightweight DeepSeek-R1 generative model to generate preliminary, linguistically logical answers or document drafts (such as solution paragraphs and processing suggestions). Subsequently, the automated compliance verification engine is activated: the terminology in the generated text is compared with the DEL library to ensure the standardized use of professional terminology; simultaneously, it performs content consistency checks with relevant regulations and clauses in the knowledge base (such as dose limits in GB 18871 and environmental impact assessment requirements in HJ 1188), marking or correcting any potentially non-compliant or inconsistent expressions.

[0063] The final, verified answer or the generated standardized document is then output to the user to complete the interaction.

[0064] Specifically, the final answer, after compliance verification, is presented to the front-end user interface in the form of structured text, charts, or a complete document (such as a draft breakdown plan in Word format). The system can provide answer sourcing, displaying the cited legal provisions or case evidence, enhancing the credibility and interpretability of the results. This completes a single interaction.

[0065] Supporting Mechanism 1: Dynamic Iterative Update Mechanism for Domain Lexicon and Model Specifically, the DEL library is not static but rather has a monthly update process. It automatically mines candidate new terms from newly added decommissioning project reports and technical documents, which are then reviewed by domain experts before being added to the library. Simultaneously, the system records new expressions and feedback from user interactions, which are used for regular incremental training and fine-tuning of the topic model, NER model, and generative model, enabling the system to adapt to the continuous development of nuclear decommissioning technologies.

[0066] Support Mechanism 1: The domain entity thesaurus has a dynamic update mechanism, which can regularly extract and verify new terms from new project documents to maintain its professionalism and timeliness.

[0067] Support Mechanism 2: The system adapts to the real-time application requirements of low computing power and high confidentiality at nuclear decommissioning sites through lightweight model and localized deployment.

[0068] Specifically, to meet the stringent requirements of "data not leaving the domain and low computing power" at nuclear facilities, lightweight processing such as pruning, quantization, and knowledge distillation was performed on core models like BERT and DeepSeek-R1, enabling them to run efficiently on local servers or dedicated edge devices. All data processing and model inference are completed within an internal secure network, ensuring that sensitive data is encrypted throughout the entire process and complies with relevant confidentiality regulations. A multi-task joint pre-training strategy further optimizes the model's basic capabilities, with its joint loss function being: JointLoss=0.15*Y1_Loss+0.15*Y2_Loss+0.15*Y3_Loss+0.15*Y4_Loss+0.15*Y5_Loss+0.25*Y6_Loss; Wherein, Joint Loss represents the combined loss, Y1_Loss represents the loss from AMR analysis tasks related to nuclear facility decommissioning and waste management, Y2_Loss represents the loss from AMR analysis tasks related to nuclear emergency response and nuclear safety, Y3_Loss represents the loss from AMR analysis tasks related to radiation protection, Y4_Loss represents the loss from AMR analysis tasks related to nuclear emergency response and nuclear safety, Y5_Loss represents the loss from AMR analysis tasks related to engineering physics and nuclear chemistry, and Y6_Loss represents the loss from AMR analysis tasks in English. This aims to simultaneously improve semantic understanding and parsing capabilities across multiple nuclear industry sub-fields and general English tasks.

[0069] If the technical solution of this application involves the collection, processing, or application of personal information, the relevant products have strictly complied with the requirements of relevant laws and regulations before implementing any personal information processing activities, clearly and explicitly informing individuals of the rules for personal information processing and obtaining their independent and voluntary authorization and consent. Specifically, if the information involved is sensitive personal information, the product has not only obtained the individual's separate consent before processing, but this consent is also an explicit consent made on the basis of full knowledge. For example, in areas where personal information collection devices such as cameras are deployed, prominent and eye-catching signs have been set up to clearly inform users that entering the area constitutes consent to the collection of their personal information; or, on the personal information processing interface (such as applications, web pages, etc.), through pop-ups, checkboxes, or active uploads, the user is required to actively authorize the process after clearly displaying key rules such as the identity of the personal information processor, the purpose of processing, the processing method, and the types of information involved.

[0070] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for the decommissioning and management of nuclear facilities based on a generative large model, characterized in that, include: Acquire on-site sensor data and historical decommissioning records, extract initial parameter values ​​of internal elements for each process, including radiation source contribution and structural change indicators, and obtain a standardized element dataset after removing noise using data cleaning methods; Monte Carlo simulation was used to perform multiple random samplings on a standardized element dataset to simulate the mutual constraints between elements within a process and obtain the distribution of coupling strength between elements. Based on the distribution of coupling strength among elements, a causal chain diagram model is constructed to identify the amplification effect of key micro-operations on subsequent links and determine the dynamic weight allocation matrix. A neural network is used to train the dynamic weight allocation matrix. The node features of the causal chain graph model are input, and the adjusted weight value sequence is output to capture the changes in weights as the field conditions change. If the weight value sequence exceeds the preset threshold, the amplification relationship analysis module is triggered to extract high-impact sub-operations from the weight value sequence and obtain the micro-operation influence vector. By mapping the influence vector of micro-operations to the macro-process objectives, a multi-level coupled prediction model is generated, which integrates pollution distribution and structurally complex parameters to determine the overall radiation dose estimate. The overall radiation dose estimate is extended to the prediction of working hours and waste volume. An iterative optimization algorithm is used to update the parameters of the multi-level coupled prediction model to obtain the final numerical prediction results, which are then used for decommissioning and management decisions.

2. The nuclear facility decommissioning management method based on a generative large model according to claim 1, characterized in that, The process involves acquiring on-site sensor data and historical decommissioning records, extracting initial parameter values ​​for each process's internal elements, including radiation source contributions and structural change indicators, and then using data cleaning methods to remove noise to obtain a standardized element dataset, including: Acquire on-site sensor data and historical decommissioning records, and record and store relevant information for each process in real time through a data acquisition system to obtain a complete raw dataset; For the original dataset, data cleaning methods are used to process outliers and missing values ​​in sensor data and historical records. Valid data is filtered out through a preset threshold range to determine the initial cleaned dataset. Initial parameter values ​​of internal elements of each process are extracted from the cleaned preliminary dataset, covering radiation source contribution and structural change indicators. The data is then classified and organized using field mapping methods to obtain the classified parameter dataset. For the classified parameter dataset, standardization processing techniques are applied to normalize the numerical values ​​of radiation source contribution and structural change indicators. A standardized element dataset is constructed through a unified dimensional conversion method. Based on the standardized element dataset, the correlation between elements within each process is analyzed using the support vector machine algorithm. By calculating the correlation matrix between elements, the dependency relationship between elements is determined. By analyzing the dependencies, the radiation source items contribution and structural change indicators in the standardized feature dataset are grouped. If the correlation of a certain feature is lower than the preset threshold, it is classified as an independent feature, and the grouped feature dataset is obtained. Based on the grouped element dataset, a distribution model of elements within each process is generated. By recording the distribution characteristics of independent and related elements, the final element analysis results are determined.

3. The nuclear facility decommissioning management method based on a generative large model according to claim 1, characterized in that, The process involves multiple random samplings of a standardized element dataset using Monte Carlo simulation to simulate the inter-factor relationships within a process, thereby obtaining the distribution of coupling strength between elements. This includes: The standardized element dataset was randomly sampled using the Monte Carlo simulation method, and multiple simulations were performed on the distribution characteristics of elements within the process to obtain a preliminary set of constraints between elements. For the initial set of constraints between elements, data processing techniques are used to classify and organize the sampling analysis results, focusing on the mutually constraining internal elements, and determining the initial coupling strength distribution between elements; Based on the preliminary coupling strength distribution, the strength analysis data between each element within the process is obtained. If the coupling strength of a certain element is lower than the preset threshold, it is classified as a weakly correlated element, and the strength distribution results after classification are obtained. Based on the intensity distribution results after classification, the strongly correlated and weakly correlated elements are grouped using relationship construction techniques to construct a detailed mapping of mutual constraints between elements, thereby obtaining a set of grouped element relationships. Based on the grouped set of element relationships, statistical tools are used to further refine the distribution results of coupling strength, focusing on the significant constraints within the process, and determining the final strength distribution model; By using the final intensity distribution model, the mutual constraints of each internal element under different process scenarios are obtained. The distribution results are then verified using simulation methods to obtain a complete dataset for element coupling analysis.

4. The nuclear facility decommissioning management method based on a generative large model according to claim 1, characterized in that, The process of constructing a causal chain diagram model based on the coupling strength distribution between elements, identifying the amplification effect of key micro-operations on subsequent stages, and determining the dynamic weight allocation matrix includes: The causal relationship is extracted from the coupling strength distribution using the causal discovery method to obtain the causal chain diagram model within the process. Based on the pointing relationships between nodes in the causal chain graph model, the number of outgoing edges and incoming edges of each node are counted to obtain the outgoing degree set and the incoming degree set of the nodes. For a set of nodes with high out-degree values, determine whether their downstream path length exceeds a preset level. If it does, mark them as key micro-operation nodes to obtain a set of key micro-operation nodes. Extract all directed paths from key micro-operation nodes to downstream nodes from the causal chain graph model to obtain the path set; For each path, the transmission strength values ​​of each edge on the path are accumulated to obtain the set of accumulated transmission strengths of the key micro-operation node to each downstream node. Based on the strength values ​​of each downstream node in the cumulative transmission strength set, all downstream nodes of the same key micro-operation node are sorted by strength to obtain the sorted downstream node sequence. A dynamic weight allocation matrix for the key micro-operation node is constructed using the sorted downstream node sequence and the corresponding cumulative transmission intensity value. The matrix rows correspond to the key nodes, the columns correspond to the downstream links, and the elements are normalized weight coefficients.

5. The nuclear facility decommissioning management method based on a generative large model according to claim 1, characterized in that, The method employs a neural network to train the dynamic weight allocation matrix, inputting node features of the causal chain graph model and outputting an adjusted weight value sequence to capture weight changes with field conditions, including: Obtain the feature vectors of all nodes in the causal chain graph model; The above feature vectors are received by a neural network and processed to obtain a sequence of weight values ​​calculated by the neural network. Establish a correspondence between each element in the weight value sequence and the on-site operating status to determine which category the on-site operating status belongs to; If it belongs to the first type of state, the first set of pre-stored weight adjustment coefficients will be used; If it belongs to the second type of state, the second set of pre-stored weight adjustment coefficients will be used; After determining the applicable weight adjustment coefficients, the original weight value sequence is corrected to obtain the corrected final weight value sequence; Write the corrected final weight value sequence back to the corresponding position in the dynamic weight allocation matrix.

6. The nuclear facility decommissioning management method based on a generative large model according to claim 1, characterized in that, If the weight value sequence exceeds a preset threshold, the amplification relationship analysis module is triggered to extract high-influence sub-operations from the weight value sequence, obtaining a micro-operation influence vector, including: Obtain the feature vectors of all current nodes in the causal chain graph; The weighted sequence is obtained by processing the feature vectors using a neural network model. A numerical comparison is performed between each element in the weighted adjustment sequence and a preset threshold. If an element exceeding a preset threshold appears in the weighted adjustment sequence, the amplification relationship analysis module will be executed. A set of high-impact sub-operations is obtained by extracting high-value elements exceeding a preset threshold from the weighted adjustment sequence. The influence vectors of the corresponding micro-operations are obtained by analyzing each of the high-influence sub-operations; The micro-operation influence vector is stored in a specified location to complete the amplification relationship processing flow.

7. The nuclear facility decommissioning management method based on a generative large model according to claim 1, characterized in that, The process involves mapping the influence vector of microscopic operations to the target of macroscopic processes to generate a multi-level coupled prediction model, which integrates pollution distribution and structurally complex parameters to determine the overall radiation dose estimate, including: Obtain relevant data on micro-operations, extract the corresponding influence vector for each micro-operation, store it as an initial vector set, and obtain a preliminarily organized vector dataset; By using an initial vector set and combining it with the target data corresponding to the macro-process, and employing pre-established mapping rules, the vector dataset is associated with the process target to determine the basic framework of multi-layer coupling. Based on the multi-layer coupling framework, relevant information about pollution distribution is introduced. If the pollution distribution data exceeds the preset threshold range, the vectors within the framework are weighted and adjusted to obtain the adjusted coupling structure. For the adjusted coupling structure, combined with the complex parameter data, the parameter data is integrated into the coupling structure through a logical matching method to determine the weight of the impact of structural complexity on the overall estimation. Starting from the influence weights, a prediction model is constructed, and the support vector machine algorithm is used to process the adjusted coupling structure and weight data to obtain a preliminary estimate of the radiation dose. Based on the preliminary estimation results and combined with the overall estimation business requirements, the data is calibrated a second time. If the calibrated data deviates from the preset range, the estimation results are fine-tuned to determine the final radiation dose estimate. The final radiation dose estimate is used to store the processed data in a designated database, and the storage path and index information are obtained to complete the recording of the entire estimation process.

8. The nuclear facility decommissioning management method based on a generative large model according to claim 1, characterized in that, The process extends from overall radiation dose estimation to prediction of working hours and waste volume. It uses an iterative optimization algorithm to update the parameters of the multi-level coupled prediction model, obtaining the final numerical prediction results for decommissioning and remediation decisions. This includes: By extracting basic information from radiation dose data, and for each data record, obtaining the correlation information between the corresponding working hours prediction and the total amount of waste, a preliminary comprehensive dataset is determined. Based on the preliminary comprehensive dataset, the radiation dose, working hour prediction, and total waste amount within the dataset are classified and organized using a pre-established hierarchical structure to obtain structured data groups. For structured data grouping, relevant business rules for decommissioning management are introduced. If the data records in a group do not match the preset threshold range, the data in the group is redistributed, and the adjusted data grouping is determined. Based on the adjusted data grouping, an iterative optimization algorithm is used to process the parameter updates within the coupled model and obtain the updated model parameter set. By using the updated model parameter set, calculations are performed on the predicted working hours and total waste data within the prediction range to determine the comprehensive numerical prediction results. Based on the comprehensive numerical prediction results and combined with the decision support needs of decommissioning management, the results data are formatted and stored to obtain the final decision reference data.

9. A nuclear facility decommissioning management system based on a generative large model, characterized in that, The system includes: The data acquisition and standardization module is used to acquire on-site sensor data and historical decommissioning records, extract the initial parameter values ​​of internal elements of each process, including radiation source contribution and structural change indicators, and obtain a standardized element dataset after removing noise using data cleaning methods. The Monte Carlo simulation module is used to perform multiple random samplings on a standardized element dataset through Monte Carlo simulation, simulate the mutual constraints between elements within a process, and obtain the distribution of coupling strength between elements. The causal chain construction module is used to construct a causal chain diagram model based on the distribution of coupling strength between elements, identify the amplification effect of key micro-operations on subsequent links, and determine the dynamic weight allocation matrix. The neural network training module is used to train the dynamic weight allocation matrix using a neural network. It takes the node features of the causal chain graph model as input and outputs the adjusted weight value sequence to capture the changes in weights as the field conditions change. The amplified relationship analysis module is used to extract high-impact sub-operations from the weight value sequence if the weight value sequence exceeds a preset threshold, thereby obtaining the micro-operation influence vector. The multi-level coupled prediction module is used to generate a multi-level coupled prediction model by mapping the micro-operational influence vector with the macro-process target, and to determine the overall radiation dose estimate by integrating pollution distribution and structurally complex parameters. The iterative optimization and decision-making module is used to expand from the overall radiation dose estimate to the prediction of working hours and waste volume. It uses an iterative optimization algorithm to update the parameters of the multi-level coupled prediction model and obtain the final numerical prediction results for decommissioning and management decisions.

Images