3D Site Nuclear Radiation Dose Analysis System and Method Based on Large Model Technology

Abstract

The present application provides a three-dimensional site nuclear radiation dose analysis system and method based on large model technology. The system includes an input module, a large model processing module, a nuclear software interaction module, and an output module. The input module is used to receive the operation intention input by the user in the form of natural language or the information input in the form of a nuclear radiation field data file. The large model processing module is used to, based on the large model, when receiving the operation intention input in the form of natural language, convert it into an operation instruction executable by the nuclear software. The nuclear software interaction module is used to establish a communication mechanism between the nuclear software and the large model. The output module uses a visualization tool to help the user understand and track the operation process of the nuclear software through the animation demonstration of the operation steps or the text prompt of the operation instruction. The present application enables the user to operate the nuclear software without in-depth understanding of the complex operation process of the nuclear software, realizes the operation of the nuclear software through natural language description, reduces the operation difficulty, and improves the work efficiency.

Description

Technical Field

[0001] This application relates to the field of nuclear radiation analysis technology. Specifically, it relates to a three-dimensional site nuclear radiation dose analysis system and method based on large model technology. Background Art

[0002] Currently, nuclear software consists of traditional function buttons, which are manipulated through keys to obtain corresponding output results. However, the existing technology has the following problems: Since nuclear software contains a large number of professional parameters and vocabulary, as well as complex operation interfaces and operation processes, users need to undergo professional training and learning, which incurs certain learning and usage costs; in terms of data analysis, traditional nuclear software can often only perform preset and relatively simple analyses, and it is difficult to handle complex and changeable data situations; moreover, the functions of nuclear software are usually fixed. When encountering unmet required functions, it is usually necessary to modify and iterate from the source, and it cannot be used normally for a long time. Therefore, a new nuclear software mode is needed to solve the above problems. Summary of the Invention

[0003] In view of this, this application provides a three-dimensional site nuclear radiation dose analysis system and method based on large model technology to solve the problems of strong professionalism, complex operation, and the need for continuous iteration and update of nuclear software in the existing technology.

[0004] To achieve the above object, the technical solutions adopted in this application are as follows:

[0005] A three-dimensional site nuclear radiation dose analysis system based on large model technology includes an input module, a large model processing module, a nuclear software interaction module, and an output module;

[0006] The input module is used to receive the operation intention input by the user in the form of natural language or the information input in the form of a nuclear radiation field data file, perform semantic understanding on the input in the form of natural language, extract key information, and perform preprocessing on the input in the form of a nuclear radiation field data file, including data cleaning and format conversion, to meet the input requirements of the large model;

[0007] The large model processing module is used to, based on the large model, when receiving the operation intention input in the form of natural language, convert it into an operation instruction executable by the nuclear software according to the pre-trained language model and the pre-established operation mapping relationship, and when receiving a nuclear radiation field data file, use machine learning algorithms to perform feature extraction, pattern recognition, and trend prediction on the data;

[0008] The nuclear software interaction module is used to establish a communication mechanism between the nuclear software and the large model to ensure the accurate transmission of operation instructions and the timely feedback of the running state of the nuclear software;

[0009] Output module: For the operation instructions executable by the kernel software generated by the large model processing module, a visualization tool is used to help users understand and track the operation process of the kernel software through an animated demonstration of operation steps or a text prompt of the operation instructions; for the analysis results of nuclear radiation field data, they are presented in the form of 3D images, charts or reports, and detailed explanations and descriptions are provided.

[0010] Furthermore, the semantic understanding of the natural language form input by the input module and the extraction of key information are specifically as follows: An open-source natural language processing toolkit is used to perform word segmentation, part-of-speech tagging, named entity recognition, etc. on the natural language input by the user, and key operation intentions and data requirements are extracted; the preprocessing of the data file form input by the input module, including data cleaning and format conversion, is specifically as follows: A data file parser is designed, and corresponding parsing codes are written for nuclear data formats to implement data reading, cleaning and format conversion.

[0011] Furthermore, the large model on which the large model processing module is based is TensorFlow or PyTorch; the pre-trained language model is a GPT, BERT or DeepSeek language model fine-tuned according to a corpus related to the nuclear field.

[0012] Furthermore, when the large model processing module receives a nuclear radiation field data file, it uses machine learning algorithms to perform feature extraction, pattern recognition and trend prediction on the data, specifically as follows:

[0013] The large model processing module is based on a deep learning model. First, according to the characteristics that different nuclides have different characteristic peaks in the gamma energy spectrum, the deep learning model is trained for nuclide identification. Through the obtained energy spectrum data of the measurement points, the characteristic peaks of the energy spectrum are found and matched with the nuclides in the database for identification training. Finally, the effect of accurately predicting the nuclides contained in the radiation source is achieved; then, according to the relationship between radiation propagation attenuation and spatial distance, the deep learning model is trained for nuclear radiation field prediction. Through the nuclear dose data received at the measurement points and the nuclear radiation propagation algorithm, radiation is calculated for the space and scene, and the position of the nuclear radiation source in the scene and the nuclear radiation dose distribution in the scene space are predicted. The results are presented in a 3D scene manner and are accompanied by an explanation of the scene; secondly, according to the differences in nuclear radiation dose and the radiation received by different human organs, the deep learning model is trained for human nuclear radiation analysis, predicting and analyzing the nuclear radiation received by each organ part of the human body, and predicting and analyzing the current human state, giving specific explanations, state evaluations and specific protection suggestions; finally, according to the geometric feature method and the self-attention mechanism to capture long-range dependencies, the deep learning model is trained for point cloud repair to achieve the effect of repairing and supplementing the 3D scene, ensuring the complete presentation of the 3D radiation scene. The geometric feature method is a technical means based on the geometric characteristics of objects to describe, analyze and process objects.

[0014] Further, the nuclear software interaction module includes a nuclear software interface module and a secure communication module. The nuclear software interface module defines clear interface specifications and data formats in the form of an API interface (Application Programming Interface). The secure communication module uses the TCP / IP protocol (Transmission Control Protocol / Internet Protocol) for network communication and encrypts and performs security checks on the data transmitted between the nuclear software and the large model processing module.

[0015] Further, the secure communication module uses an AI-driven data security supervision method to preprocess the data stream, call an adaptive AI model, and automatically adjust the parameters of the AI model based on the changes in the data type. When an abnormal data stream is detected, it determines the type, source, and potential threat level of the abnormality. When a high-risk abnormality is detected, it automatically triggers an event-driven response mechanism. It feeds back the event response result and processing status to the adaptive AI model to optimize the learning parameters and decision logic of the model.

[0016] Further, the secure communication module also uses the principles of non-clonability and uncertainty in quantum mechanics to encrypt the data to ensure absolute security of data transmission. When sending data, it constructs a hash tree (Merkle tree), calculates the root hash value, and sends the root hash value and the data together. When receiving, it recalculates the hash tree of the data and compares it with the received root hash value to verify the integrity of the data.

[0017] Further, the visualization tool in the output module is OpenGL (Open Graphics Library), Matplotlib, or Plotly. The output module specifically also executes an end-to-end Transformer network-based conditional text generation algorithm to generate a detailed report and explanatory notes based on the analysis results.

[0018] A three-dimensional site nuclear radiation dose analysis method based on large model technology. The method is based on the three-dimensional site nuclear radiation dose analysis system based on large model technology, and the method includes:

[0019] S1: Receive the operation intention input by the user in natural language form or the information input in the form of a nuclear radiation field data file through the input module, perform semantic understanding on the input in natural language form, extract key information, and preprocess the input in the form of a nuclear radiation field data file, including data cleaning and format conversion, to meet the input requirements of the large model;

[0020] S2: The large model processing module, based on the large model, when receiving the operation intention input in natural language form, converts it into operation instructions executable by the nuclear software according to the pre-trained language model and the pre-established operation mapping relationship. When receiving the nuclear radiation field data file, it uses machine learning algorithms to perform feature extraction, pattern recognition, and trend prediction on the data;

[0021] S3: The nuclear software interaction module establishes a communication mechanism between the nuclear software and the large model to ensure the accurate transmission of operation instructions and the timely feedback of the running state of the nuclear software;

[0022] S4: For the operation instructions executable by the nuclear software generated by the large model processing module, a visualization tool is used to help users understand and track the operation process of the nuclear software through an animated demonstration of the operation steps or a text prompt of the operation instructions. For the analysis results of the nuclear radiation field data, they are presented in the form of three-dimensional images, charts, or reports, and detailed explanations and descriptions are provided.

[0023] Further, when receiving the nuclear radiation field data file in step S2, the specific process of using machine learning algorithms to perform feature extraction, pattern recognition, and trend prediction on the data is as follows:

[0024] S2.1: The large model processing module, based on the deep learning model, first conducts nuclide identification training on the deep learning model according to the characteristics that different nuclides have different characteristic peaks in the gamma energy spectrum. By obtaining the energy spectrum data of the measurement points, it searches for the characteristic peaks of the energy spectrum and performs matching and identification training with the nuclides in the database. Finally, it achieves the effect of accurately predicting the nuclides contained in the radiation source;

[0025] S2.2: According to the relationship between radiation propagation attenuation and spatial distance, it conducts nuclear radiation field prediction training on the deep learning model. Through the nuclear dose data received at the measurement points and the nuclear radiation propagation algorithm, it performs radiation calculation for the space and the scene, predicts the position of the nuclear radiation source in the scene, and the nuclear radiation dose distribution in the scene space. The results are presented in a three-dimensional scene manner and are accompanied by an explanation of the scene;

[0026] S2.3: According to the differences in nuclear radiation dose and the radiation received by different organs of the human body, it conducts human body nuclear radiation analysis training on the deep learning model, predicts and analyzes the nuclear radiation received by each organ part of the human body, and conducts prediction analysis of the current human body state, giving specific explanations, state evaluations, and specific protection suggestions;

[0027] S2.4: Capture long-range dependencies according to the geometric feature method and the self-attention mechanism, and perform point cloud repair training on the deep learning model to achieve the effect of repairing and supplementing the three-dimensional scene, ensuring the complete presentation of the three-dimensional radiation scene. The geometric feature method is a technical means for describing, analyzing, and processing objects based on the geometric characteristics of objects.

[0028] Compared with the prior art, the beneficial effects of this application are as follows:

[0029] 1. Improved operation convenience: Users do not need to deeply understand the complex operation process of the nuclear software. They can operate the nuclear software through natural language description, which reduces the operation difficulty and improves work efficiency.

[0030] 2. Enhanced data analysis ability: The introduction of the large model enables more in-depth and comprehensive analysis of nuclear data, and can discover potential problems and laws that are difficult to detect by traditional methods, providing more powerful support for decision-making in the nuclear field.

[0031] 3. Expanded application scope: This technical solution enables the nuclear software to better meet the needs of different users, including non-professionals and scientific researchers, and expands the application scope of the nuclear software. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the technical solutions of the embodiments of this application, the following will briefly introduce the drawings required in the embodiments. It should be understood that the following drawings only show some embodiments of this application, so they should not be regarded as limiting the scope. For those of ordinary skill in the art, without creative efforts, other related drawings can also be obtained based on these drawings.

[0033] Figure 1 It is the structural block diagram of the three-dimensional site nuclear radiation dose analysis system based on the large model technology of this application;

[0034] Figure 2 It is the flowchart of the three-dimensional site nuclear radiation dose analysis method based on the large model technology of this application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0035] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the following will clearly and completely describe the technical solutions in the embodiments of this application with reference to the drawings in the embodiments of this application. Obviously, the described embodiments are part of the embodiments of this application, rather than all of them.

[0036] As Figure 1 shown, the three-dimensional site nuclear radiation dose analysis system based on the large model technology includes an input module 110, a large model processing module 120, a nuclear software interaction module 130, and an output module 140;

[0037] An input module 110, which is used to receive the operation intention input by the user in the form of natural language or the information input in the form of a nuclear radiation field data file, perform semantic understanding on the natural language input, extract key information, and perform preprocessing on the nuclear radiation field data file input, including data cleaning and format conversion, to meet the input requirements of the large model;

[0038] Specifically, an open-source natural language processing toolkit can be used to perform word segmentation, part-of-speech tagging, named entity recognition, etc. on the natural language input by the user, and extract key operation intentions and data requirements. For example, when the user inputs "display the abnormal areas in the scene", the system can identify key information such as "display", "scene", and "abnormal areas". For the input in the form of a data file, it supports nuclear software-related data formats, such as nuclear radiation site three-dimensional point cloud data files, measurement point data files, etc., and corresponding parsing codes are written for the nuclear data format to perform data preprocessing, including data cleaning, format conversion, etc., to meet the input requirements of the large model. The nuclear data format is a custom.mpt data format; the nuclear data is the data jointly measured by a Geiger tube and a spectrometer, and contains key data such as measurement integration time, radiation source position, radiation direction vector, dose value, and energy spectrum.

[0039] A large model processing module 120, which is used to, based on the large model, when receiving the operation intention input in the form of natural language, convert it into an operation instruction executable by the nuclear software according to the pre-trained language model and the pre-established operation mapping relationship, and when receiving the nuclear radiation field data file, use machine learning algorithms to perform feature extraction, pattern recognition, and trend prediction on the data;

[0040] Furthermore, the large model on which the large model processing module 120 is based is TensorFlow or PyTorch; the pre-trained language model is a GPT, BERT, or DeepSeek language model fine-tuned according to a nuclear domain-related corpus.

[0041] In specific implementation, the fine-tuned GPT, BERT, or DeepSeek language model is loaded into the large model (TensorFlow or PyTorch) for specific prediction applications. The fine-tuned GPT, BERT, or DeepSeek language model is a deep learning model designed and specially trained according to the characteristics of nuclear data. For example, a large amount of nuclear data is collected, data annotation and feature engineering are performed, and neural network algorithms are used for model training and optimization. The fine-tuning specifically includes parameter freezing, learning rate adjustment, and layer adapters. The fine-tuned GPT, BERT, or DeepSeek language model is deployed and applied after model evaluation (specifically including accuracy, perplexity, etc.).

[0042] Further, when the large model processing module receives a nuclear radiation field data file, the use of machine learning algorithms for data feature extraction, pattern recognition, and trend prediction is specifically as follows:

[0043] Based on the deep learning model, the large model processing module first conducts nuclide identification training on the deep learning model according to the characteristics of different nuclides having different characteristic peaks in the gamma energy spectrum. By obtaining the energy spectrum data of the measurement points, it searches for the characteristic peaks of the energy spectrum and performs matching identification training with the nuclides in the database. Finally, it achieves the effect of accurately predicting the nuclides contained in the radiation source. Then, according to the relationship between radiation propagation attenuation and spatial distance, it conducts nuclear radiation field prediction training on the deep learning model. Through the measurement points receiving nuclear dose data and the nuclear radiation propagation algorithm, it performs radiation calculation for the space and scene, predicts the position of the nuclear radiation source in the scene, as well as the nuclear radiation dose distribution in the scene space. The result is presented in a three-dimensional scene manner and is accompanied by an explanatory note of the scene. Secondly, according to the differences in nuclear radiation dose and the radiation reception of different human organs, it conducts human nuclear radiation analysis training on the deep learning model. On the one hand, it predicts and analyzes the nuclear radiation received by each organ part of the human body. On the other hand, it conducts prediction and analysis of the current human body state and gives specific explanations, state evaluations, and specific protection suggestions. Finally, according to the geometric feature method and the self-attention mechanism to capture long-range dependencies, it conducts point cloud repair training on the deep learning model to achieve the effect of repairing and supplementing the three-dimensional scene, ensuring the complete presentation of the three-dimensional radiation scene. The geometric feature method is a technical means based on the geometric characteristics of objects to describe, analyze, and process objects.

[0044] For example, in the scenario of storing nuclear waste in a factory, instruments are equipped to measure real-time nuclear dose data and location information at multiple positions. After obtaining the data, based on the relationship between radiation propagation attenuation and spatial distance, a deep learning model is trained. The location information of the measurement points, the measured nuclear dose data, and the distance information between different positions in the scene space and the measurement points are used as input data, and the output data is the predicted nuclear dose values at other positions in the space. Through multiple iterative trainings, the deep learning model learns the relationship between nuclear dose data, spatial location, and radiation propagation attenuation. For example, the model learns the rule that the farther away from the measurement point, the lower the nuclear dose value, and can predict the nuclear dose value at that position based on the input measurement point information and spatial location information. According to the results of radiation calculation, the change trend of nuclear radiation dose is analyzed, and algorithms (such as gradient ascent algorithm, etc.) are used to find the areas with higher nuclear dose values and map them to the positions of scene entities, thereby predicting the location of the nuclear radiation source. Subsequently, the predicted nuclear radiation source location and nuclear radiation dose distribution are presented in a three-dimensional scene manner. In the point cloud scene, different colors and brightness are used to represent the magnitude of nuclear radiation dose. For example, red represents high-dose areas and blue represents low-dose areas. The attached scene explanation details the location of the measurement points and the measurement data, the principle and parameter settings of the nuclear radiation propagation algorithm, and how to obtain the results of the nuclear radiation source location and dose distribution through analysis. For example, "Measurement point A is located at (location or coordinates) in the scene, and the measured nuclear dose is xx μSv / h; an exponential decay model is used for radiation calculation, and the decay coefficient is set to 0.1. Based on the measurement point data and spatial location information, the nuclear radiation dose distribution of the entire scene is calculated. Through analysis, it is found that the nuclear dose value in the xx area is relatively high, and it is initially judged that the nuclear radiation source may be located at xx in the warehouse."

[0045] The geometric feature method is a technical means for describing, analyzing, and processing objects based on the geometric characteristics of objects. By detecting point features such as corner points, image registration and stitching are achieved. Then, the image is divided into different regions through surface features, and each region is analyzed and processed separately. Finally, a three-dimensional model of the object is constructed. For example, through equipment acquisition, and by referring to public three-dimensional model libraries such as ShapeNet and ModelNet, defective point cloud data and complete point cloud data are obtained respectively. Then, based on PointNet++ and the geometric feature method, block-by-block local comparison and repair training are performed on the two sets of data. Then, according to the results, it is evaluated whether the effect is achieved, and the model is fine-tuned.

[0046] The nuclear software interaction module 130 is used to establish a communication mechanism between the nuclear software and the large model to ensure the accurate transmission of operation instructions and the timely feedback of the running state of the nuclear software;

[0047] After the large model generates an operation instruction, it is sent to the nuclear software through the nuclear software interaction module 130. After the nuclear software executes the operation, it feeds back the execution result and the current status to the large model.

[0048] Furthermore, the nuclear software interaction module 130 includes a nuclear software interface module and a secure communication module. The nuclear software interface module defines clear interface specifications and data formats in the form of an API interface (Application Programming Interface). The secure communication module uses the TCP / IP protocol (Transmission Control Protocol / Internet Protocol) for network communication and encrypts and performs security checks on the data transmitted between the nuclear software and the large model processing module.

[0049] Specifically, the nuclear software interface module can be a module after appropriate modification of the nuclear software. This module can receive and execute the operation instructions sent by the large model and provide necessary interfaces and data outputs so that the large model can obtain the data generated by the operation of the nuclear software.

[0050] To prevent potential security risks (such as data leakage, malicious attacks, etc.) that the large model may bring, a security supervision system (secure communication module) is set up between the large model and the nuclear software. This system adopts an AI-driven data security supervision method to preprocess the data stream, call an adaptive AI model, and automatically adjust the parameters of the AI model based on the change of data types. When the system detects an abnormal data stream, it judges the type, source, and potential threat level of the abnormality; when a high-risk abnormality is detected, it automatically triggers an event-driven response mechanism; the event response result and processing status are fed back to the adaptive AI model to optimize the learning parameters and decision-making logic of the model. The large model runs under the supervision system and is isolated from the nuclear software and nuclear data. The supervision system will strictly monitor and filter the input and output of the large model, and only allow data and operations that conform to the security policy to pass. When the nuclear software needs to call the large model for analysis and prediction of nuclear data, the relevant data is encrypted and preprocessed before being passed into the large model. After the large model processes the data, the result is returned to the supervision system, and the supervision system then performs a security check on the result and passes it to the nuclear software.

[0051] The security communication module also establishes a confidential communication protocol. To ensure the reliable transmission of highly sensitive nuclear data between the large model and nuclear software, it encrypts the data using the principles of non-clonability and uncertainty in quantum mechanics, ensuring the absolute security of data transmission. When sending data, it constructs a hash tree (Merkle tree), calculates the root hash value, and sends the root hash value together with the data. When receiving, it recalculates the hash tree of the data and compares it with the received root hash value to verify the integrity of the data. This method has high efficiency for verifying large-scale nuclear data.

[0052] The output module 140 uses visualization tools for the operation instructions executable by the nuclear software generated by the large model processing module to help users understand and track the operation process of the nuclear software through animated demonstrations of operation steps or text prompts of operation instructions; for the analysis results of the nuclear radiation field data, it presents them in the form of three-dimensional images, charts or reports, and provides detailed explanations and descriptions.

[0053] Furthermore, the visualization tools in the output module 140 are OpenGL (Open Graphics Library), Matplotlib or Plotly; the output module 140 is specifically further used to execute an end-to-end Transformer network-based conditional text generation algorithm for generating detailed reports and explanatory notes according to the analysis results.

[0054] Using visualization tools such as OpenGL, Matplotlib, Plotly, etc., the analysis results generated by the large model are presented in the form of charts. At the same time, develop an end-to-end Transformer network-based conditional text generation algorithm. On the one hand, by adding a branched convolutional structure, convolutional networks with different convolutional kernels are used to extract different semantic information in the input text to enrich semantic features. On the other hand, by adding a special convolutional branch, targeted reinforcement training is carried out, and through nuclear technology professional words and input purpose semantic information, to improve professionalism and accuracy.

[0055] For example, for the prediction results of the radiation area of the scene and the simulated radiation dose received by the human body in this area, not only the simulation data is given, but also the risk level of the data is analyzed, and radiation protection and treatment suggestions are provided.

[0056] Precautions during the implementation process:

[0057] (1) During the training and use of the large model, attention should be paid to data privacy protection, and sensitive nuclear data should be encrypted to avoid data leakage.

[0058] (2) Since large models consume a large amount of computing resources, it is necessary to reasonably allocate computing devices, such as using high-performance GPU clusters, to ensure the operating efficiency of the system.

[0059] (3) Conduct sufficient testing and verification on the system, including functional testing, performance testing, and security testing, etc., to ensure the stability and reliability of the system.

[0060] As Figure 2 shown, a three-dimensional site nuclear radiation dose analysis method based on large model technology, the method is based on the three-dimensional site nuclear radiation dose analysis system based on the large model technology, and the method includes:

[0061] S1: Receive the operation intention input by the user in the form of natural language through the input module or the information input in the form of a nuclear radiation field data file, perform semantic understanding on the natural language form input, extract key information, and perform preprocessing on the nuclear radiation field data file form input, including data cleaning and format conversion, to meet the input requirements of the large model;

[0062] Specifically, open-source natural language processing toolkits can be used to perform word segmentation, part-of-speech tagging, named entity recognition, etc. on the natural language input by the user, and extract key operation intentions and data requirements. For example, when the user inputs "display the abnormal areas in the scene", the system can identify key information such as "display", "scene", and "abnormal areas". For the input in the form of a data file, it supports nuclear software-related data formats, such as three-dimensional point cloud data files of nuclear radiation sites, measurement point data files, etc., and corresponding parsing codes are written for the nuclear data format to perform data preprocessing, including data cleaning, format conversion, etc., to meet the input requirements of the large model. The nuclear data format is a custom.mpt data format; the nuclear data is data jointly measured by Geiger tubes and spectrometers, and contains key data such as measurement integration time, radiation source position, radiation direction vector, dose value, and energy spectrum.

[0063] S2: The large model processing module is based on the large model. When receiving the operation intention input in the form of natural language, according to the pre-trained language model and the pre-established operation mapping relationship, it is converted into an operation instruction executable by the nuclear software. When receiving the nuclear radiation field data file, machine learning algorithms are used to extract features, identify patterns, and predict trends from the data;

[0064] The large model is TensorFlow or PyTorch; the pre-trained language model is a GPT, BERT, or DeepSeek language model fine-tuned according to a corpus related to the nuclear field.

[0065] During specific implementation, the fine-tuned GPT, BERT, or DeepSeek language model is loaded into the large model (TensorFlow or PyTorch) for specific prediction applications. The fine-tuned GPT, BERT, or DeepSeek language model is a deep learning model designed and specifically trained according to the characteristics of nuclear data. For example, a large amount of nuclear data is collected, data annotation and feature engineering are carried out, and neural network algorithms are used for model training and optimization. The fine-tuning specifically includes parameter freezing, learning rate adjustment, and layer adapters. The fine-tuned GPT, BERT, or DeepSeek language model is deployed and applied after model evaluation (specifically including accuracy, perplexity, etc.).

[0066] Furthermore, when receiving the nuclear radiation field data file in step S2, the use of machine learning algorithms to perform feature extraction, pattern recognition, and trend prediction on the data specifically is as follows:

[0067] S2.1: Based on the deep learning model, the large model processing module first conducts nuclide identification training on the deep learning model according to the characteristics that different nuclides have different characteristic peaks in the gamma energy spectrum. By obtaining the energy spectrum data of the measurement points, it searches for the characteristic peaks of the energy spectrum and performs matching identification training with the nuclides in the database. Finally, it achieves the effect of accurately predicting the nuclides contained in the radiation source;

[0068] S2.2: According to the relationship between radiation propagation attenuation and spatial distance, the deep learning model is trained for nuclear radiation field prediction. Through the nuclear dose data received at the measurement points and the nuclear radiation propagation algorithm, radiation is calculated for the space and scene, predicting the position of the nuclear radiation source in the scene and the nuclear radiation dose distribution in the scene space. The results are presented in a three-dimensional scene manner and accompanied by an explanation of the scene;

[0069] S2.3: According to the differences in nuclear radiation dose and the radiation received by different human organs, the deep learning model is trained for human nuclear radiation analysis, predicting and analyzing the nuclear radiation received by each organ part of the human body, and conducting prediction analysis of the current human state, giving specific explanations, state evaluations, and specific protection suggestions.

[0070] S3: The nuclear software interaction module establishes a communication mechanism between the nuclear software and the large model to ensure the accurate transmission of operation instructions and the timely feedback of the operating status of the nuclear software;

[0071] After the large model generates an operation instruction, it is sent to the nuclear software through the nuclear software interaction module. After the nuclear software executes the operation, it feeds back the execution result and the current state to the large model.

[0072] Furthermore, the nuclear software interaction module uses the API interface (Application Programming Interface) to define clear interface specifications and data formats; and uses the TCP / IP protocol (Transmission Control Protocol / Internet Protocol) for network communication, and encrypts and performs security checks on the data transmitted between the nuclear software and the large model processing module.

[0073] To prevent potential security risks (such as data leakage, malicious attacks, etc.) that may be brought by the large model, a layer of security supervision system is set up between the large model and the nuclear software. This system uses an AI-driven data security supervision method to preprocess the data stream, call an adaptive AI model, and automatically adjust the parameters of the AI model based on the change of data types. When the system detects an abnormal data stream, it judges the type, source, and potential threat level of the abnormality; when a high-risk abnormality is detected, it automatically triggers an event-driven response mechanism; and feeds back the event response result and processing status to the adaptive AI model to optimize the learning parameters and decision-making logic of the model. The large model runs under the supervision system and is isolated from the nuclear software and nuclear data. The supervision system will strictly monitor and filter the input and output of the large model, and only allow data and operations that conform to the security policy to pass. When the nuclear software needs to call the large model for nuclear data analysis and prediction, the relevant data is encrypted and preprocessed and then transmitted to the large model. After the large model processes it, the result is returned to the supervision system, and the supervision system performs a security check on the result and then transmits it to the nuclear software.

[0074] In addition, a confidential communication protocol can be established. To ensure the reliable transmission of highly sensitive nuclear data between the large model and the nuclear software, the data is encrypted using the principles of non-clonability and uncertainty in quantum mechanics to ensure the absolute security of data transmission. When sending data, a hash tree (Merkle tree) is constructed, the root hash value is calculated, and the root hash value and the data are sent together. When receiving, the hash tree of the data is recalculated and compared with the received root hash value to verify the integrity of the data. This method has high efficiency for the verification of large-scale nuclear data.

[0075] S4: For the operation instructions executable by the nuclear software generated by the large model processing module, the output module uses a visualization tool to help users understand and track the operation process of the nuclear software through an animated demonstration of the operation steps or a text prompt of the operation instructions; for the analysis results of the nuclear radiation field data, they are presented in the form of three-dimensional images, charts, or reports, and detailed explanations and descriptions are provided.

[0076] Use visualization tools such as OpenGL, Matplotlib, Plotly, etc. to display the analysis results generated by the large model in the form of charts. At the same time, develop a conditional text generation algorithm based on the end-to-end Transformer network. On the one hand, by adding a branch convolution structure, convolutional networks with different convolution kernels are used to extract different semantic information in the input text to enrich semantic features. On the other hand, by adding a special convolution branch, targeted reinforcement training is carried out, and through kernel technology professional words and sentences and input purpose semantic information, to improve professionalism and accuracy.

[0077] For example, for the prediction results of the scene radiation area and the simulated radiation dose received by the human body in this area, not only the simulation data is given, but also the danger level of the data is analyzed, and radiation protection and treatment suggestions are provided.

[0078] This application constructs an integrated system of nuclear software and a large model. In this system, the large model serves as the intelligent core. On the one hand, it receives the natural language description or relevant data input of the user's operation on the nuclear software, generates corresponding nuclear software operation instructions after processing, and drives the nuclear software to execute specific operations. On the other hand, it deeply analyzes the data generated by the nuclear software and outputs valuable analysis results. This application does not require in-depth understanding of the professional knowledge in the nuclear field and the complex operation process of the nuclear software. The operation of the nuclear software can be realized through natural language description, which reduces the operation difficulty and improves work efficiency. In addition, this application makes data analysis more in-depth and comprehensive, can discover potential problems and laws that are difficult to detect by traditional methods, and provides more powerful support for decision-making in the nuclear field. Then in terms of application, this technical solution enables the nuclear software to better meet the needs of different users, including non-professionals and scientific researchers, and expands the application scope of the nuclear software.

[0079] The above is only the specific implementation manner of this application, but the protection scope of this application is not limited thereto. Any person skilled in the art within the technical scope disclosed by this application can easily think of changes or substitutions, which should all be covered within the protection scope of this application. Therefore, the protection scope of this application should be subject to the protection scope of the claims.

Claims

1. A three-dimensional site nuclear radiation dose analysis system based on large model technology, characterized in that, It includes an input module, a large model processing module, a nuclear software interaction module, and an output module; The input module is used to receive the operation intention input by the user in the form of natural language and the information input in the form of a nuclear radiation field data file, perform semantic understanding on the natural language form input, extract key information, and perform preprocessing on the nuclear radiation field data file form input, including data cleaning and format conversion, to meet the input requirements of the large model; The large model processing module is used to, based on the large model, when receiving the operation intention input in the form of natural language, convert it into an operation instruction executable by the nuclear software according to the pre-trained language model and the pre-established operation mapping relationship, and when receiving the nuclear radiation field data file, use machine learning algorithms to perform feature extraction, pattern recognition, and trend prediction on the data; The nuclear software interaction module is used to establish a communication mechanism between the nuclear software and the large model to ensure the accurate transmission of operation instructions and the timely feedback of the running state of the nuclear software; The output module, for the operation instructions executable by the nuclear software generated by the large model processing module, uses a visualization tool to help the user understand and track the operation process of the nuclear software through an animated demonstration of the operation steps or a text prompt of the operation instructions; For the analysis results of the nuclear radiation field data, they are presented in the form of three-dimensional images, charts, or reports, and detailed explanations and descriptions are provided; When the large model processing module receives the nuclear radiation field data file, the specific implementation of using machine learning algorithms to perform feature extraction, pattern recognition, and trend prediction on the data is as follows: Based on the deep learning model, the large model processing module first performs nuclide identification training on the deep learning model according to the characteristics that different nuclides have different characteristic peaks in the gamma energy spectrum. By obtaining the energy spectrum data of the measurement points, it searches for the characteristic peaks of the energy spectrum and performs matching identification training with the nuclides in the database. Finally, it achieves the effect of accurately predicting the nuclides contained in the radiation source. Then, according to the relationship between radiation propagation attenuation and spatial distance, it performs nuclear radiation field prediction training on the deep learning model. Through the nuclear dose data received at the measurement points and the nuclear radiation propagation algorithm, it performs radiation calculation on the space and scene, predicts the position of the nuclear radiation source in the scene, as well as the nuclear radiation dose distribution in the scene space. The result is presented in a three-dimensional scene manner, accompanied by an explanation of the scene; Secondly, according to the differences in nuclear radiation dose and the radiation received by different organs of the human body, it performs human body nuclear radiation analysis training on the deep learning model, predicts and analyzes the nuclear radiation received by each organ part of the human body, and performs prediction analysis of the current human body state, giving specific explanations, state evaluations, and specific protection suggestions. Finally, according to the geometric feature method and the self-attention mechanism to capture long-range dependencies, it performs point cloud repair training on the deep learning model to achieve the effect of repairing and supplementing the three-dimensional scene, ensuring the complete presentation of the three-dimensional radiation scene. The geometric feature method is a technical means based on the geometric characteristics of objects to describe, analyze, and process objects.

2. The three-dimensional site nuclear radiation dose analysis system based on large model technology according to claim 1, characterized in that, The input module performs semantic understanding on the input in the form of natural language and extracts key information, specifically: using an open-source natural language processing toolkit to perform word segmentation, part-of-speech tagging, and named entity recognition on the natural language input by the user, and extracting key operation intentions and data requirements; the input module preprocesses the input in the form of data files, including data cleaning and format conversion, specifically: designing a data file parser, writing corresponding parsing code for nuclear data formats, and implementing data reading, cleaning, and format conversion.

3. The three-dimensional site nuclear radiation dose analysis system based on large model technology according to claim 1, characterized in that, The large model on which the large model processing module is based is TensorFlow or PyTorch; the pre-trained language model is a GPT, BERT, or DeepSeek language model fine-tuned according to a corpus related to the nuclear field.

4. The three-dimensional site nuclear radiation dose analysis system based on large model technology according to claim 3, wherein The nuclear software interaction module includes a nuclear software interface module and a secure communication module. The nuclear software interface module uses the API interface method to define clear interface specifications and data formats; the secure communication module uses the TCP / IP protocol for network communication and encrypts and performs security checks on the data transmitted between the nuclear software and the large model processing module.

5. The three-dimensional site nuclear radiation dose analysis system based on large model technology according to claim 4, characterized in that, The secure communication module uses an AI-driven data security supervision method to preprocess the data stream, call an adaptive AI model, and automatically adjust the parameters of the AI model based on the change of data types; when detecting an abnormal data stream, judge the type, source, and potential threat level of the abnormality; when detecting a high-risk abnormality, automatically trigger an event-driven response mechanism; feedback the event response result and processing status to the adaptive AI model to optimize the learning parameters and decision-making logic of the model.

6. The three-dimensional site nuclear radiation dose analysis system based on large model technology according to claim 5, characterized in that, The secure communication module is also used to encrypt data using the principles of non-clonability and uncertainty in quantum mechanics to ensure the absolute security of data transmission. When sending data, construct a hash tree, calculate the root hash value, and send the root hash value and the data together. When receiving, recalculate the hash tree of the data and compare it with the received root hash value to verify the integrity of the data.

7. The three-dimensional site nuclear radiation dose analysis system based on large model technology according to claim 6, characterized in that, The visualization tool in the output module is OpenGL, Matplotlib, or Plotly; the output module is specifically further used to execute an end-to-end Transformer network-based conditional text generation algorithm to generate a detailed report and explanatory notes according to the analysis results.

8. A three-dimensional site nuclear radiation dose analysis method based on large model technology, the method being based on the three-dimensional site nuclear radiation dose analysis system based on large model technology according to any one of claims 1-7, characterized in that, The method includes: S1: Receive the operation intention input by the user in the form of natural language through the input module or the information input in the form of a nuclear radiation field data file, perform semantic understanding on the input in the form of natural language, extract key information, and preprocess the input in the form of a nuclear radiation field data file, including data cleaning and format conversion, to meet the input requirements of the large model; S2: The large model processing module is based on the large model. When receiving the operation intention input in the form of natural language, according to the pre-trained language model and the pre-established operation mapping relationship, convert it into an operation instruction executable by the nuclear software. When receiving a nuclear radiation field data file, use machine learning algorithms to perform feature extraction, pattern recognition, and trend prediction on the data; S3: The nuclear software interaction module establishes a communication mechanism between the nuclear software and the large model to ensure the accurate transmission of operation instructions and the timely feedback of the running status of the nuclear software; S4: For the operation instructions executable by the nuclear software generated by the large model processing module, the output module uses a visualization tool to help users understand and track the operation process of the nuclear software through the animation demonstration of operation steps or the text prompt of operation instructions; for the analysis results of nuclear radiation field data, they are presented in the form of 3D images, charts or reports, and detailed explanations and descriptions are provided.

9. The three-dimensional site nuclear radiation dose analysis method based on large model technology according to claim 8, wherein When receiving the nuclear radiation field data file in step S2, the specific operations of using machine learning algorithms to extract features, recognize patterns and predict trends from the data are as follows: S2.1: Based on the deep learning model, the large model processing module first conducts nuclide identification training on the deep learning model according to the characteristics that different nuclides have different characteristic peaks in the gamma energy spectrum. By obtaining the energy spectrum data of the measurement points, it searches for the characteristic peaks of the energy spectrum and conducts matching identification training with the nuclides in the database. Finally, it achieves the effect of accurately predicting the nuclides contained in the radiation source; S2.2: According to the relationship between radiation propagation attenuation and spatial distance, the large model processing module conducts nuclear radiation field prediction training on the deep learning model. Through the nuclear dose data received at the measurement points and the nuclear radiation propagation algorithm, it conducts radiation calculation for the space and the scene to predict the position of the nuclear radiation source in the scene and the nuclear radiation dose distribution in the scene space. The results are presented in a 3D scene manner and are accompanied by explanations of the scene; S2.3: According to the difference between nuclear radiation dose and the radiation received by different organs of the human body, the large model processing module conducts human body nuclear radiation analysis training on the deep learning model to predict and analyze the nuclear radiation received by each organ part of the human body, and conduct prediction analysis of the current human body state, giving specific explanations, state evaluations and specific protection suggestions; S2.4: According to the geometric feature method and the self-attention mechanism to capture long-range dependencies, the large model processing module conducts point cloud repair training on the deep learning model to achieve the effect of repairing and supplementing the 3D scene, ensuring the complete presentation of the 3D radiation scene. The geometric feature method is a technical means based on the geometric characteristics of objects to describe, analyze and process objects.

Images