Method and system for constructing medical accelerator quality control knowledge base by using large language model

Abstract

The invention discloses a system for constructing a medical accelerator quality control knowledge base by using a large language model. The system comprises an acquisition module, a two-stage preprocessing module, an abnormal value detection module, a large language training module, a knowledge graph module and a knowledge base indexing and querying module. According to the method, related data of the medical accelerator are collected through multiple channels, information of the medical accelerator in multiple aspects of equipment performance, faults, quality control and the like can be covered, and the constructed knowledge base has comprehensiveness. Intelligent screening and hierarchical processing are performed on the data, so that more accurate processing of the data in different states is facilitated; the preset rule set comprises an equipment parameter range rule, a data format and integrity rule, a data frequency rule, a historical data statistical rule and the like, various characteristics of the medical accelerator data are comprehensively considered, and the accuracy of data screening and blocking is improved; and the method can better adapt to the complexity and specialty of medical accelerator data.

Description

Technical Field

[0001] The present invention relates to the field of medical accelerator quality control, and in particular to a method and system for constructing a medical accelerator quality control knowledge base using a large language model. Background Art

[0002] A medical accelerator is a device that uses high-energy rays (such as photons, electrons, protons or heavy ions) for radiation therapy. These high-energy rays are directed to tumor tissue to damage the DNA of cancer cells, thereby inhibiting tumor growth or eliminating cancer cells.

[0003] Medical accelerators are important equipment for radiotherapy, and their quality control is directly related to the safety and effectiveness of patient treatment. Traditional quality control methods rely on manual inspection and record keeping, which has problems such as heavy workload, prone to errors, and difficulty in knowledge transfer. With the development of big data and artificial intelligence technologies, especially the widespread application of large language models in the field of natural language processing, new solutions have been provided for the quality control of medical accelerators. However, there is currently no large language model knowledge base specifically for the field of medical accelerator quality control. Summary of the invention

[0004] The purpose of the present invention is to provide a method and system for constructing a medical accelerator quality control knowledge base using a large language model, which solves the problems of heavy workload and easy errors in the above-mentioned medical accelerator quality control method.

[0005] To achieve the above object, the present invention provides a method for constructing a medical accelerator quality control knowledge base using a large language model, comprising the following steps:

[0006] S1. Collect medical accelerator-related data through multiple channels and format the data from different channels in a unified manner;

[0007] S2. Perform two-stage preprocessing on the collected data, intelligently screen the data, analyze the screened data according to the set standards and divide it into multiple data blocks, check the integrity of the analyzed data blocks and divide them into normal data blocks, potential abnormal data blocks and data blocks that need further inspection, generate normal data tags, potential abnormal data tags and data tags that need further inspection, and mark the corresponding data blocks.

[0008] The normal data mark indicates that the data integrity is good and within the normal range, the potential abnormal data mark indicates that the data is close to the normal range boundary, and the data mark that needs further inspection indicates that the data has integrity problems;

[0009] S3, using autoencoders (AE), variational autoencoders (VAE) and generative adversarial networks (GAN) to build an analysis model to detect outliers in potential abnormal data blocks and data blocks that require further inspection involving equipment performance parameters.

[0010] At the same time, the Z-score method and IQR method are used to detect outliers for each data in the data block determined to be abnormal by the analysis model:

[0011] If the Z-score of the data exceeds ±3 and the IQR range, it is determined to be an outlier, and the analysis model removes the outlier. If the Z-score of the data does not exceed ±3 / does not exceed the IQR range, it is determined to be a value to be processed and analyzed and processed according to the physical meaning of the data and historical data: correction / elimination / human intervention processing;

[0012] S4. Automated data verification. Using a stratified random partitioning method, the data after preprocessing and outlier detection is divided into a training set, a validation set, and a test set. The training set is used to build the training model, the validation set is used to adjust the model, and the test set is used for performance evaluation.

[0013] A monitoring mechanism is built to analyze the test results in detail when they do not meet the predetermined standards, and to feed the analysis results back to the analysis model or model training model to make targeted adjustments to the model structure.

[0014] S5. Divide the processed data into three sub-libraries based on quality control knowledge: equipment performance sub-library, equipment problem and treatment sub-library, and quality control standard sub-library.

[0015] All text data in the three sub-databases are merged to build a vocabulary. All text data are traversed to count the number of co-occurrences of each word in the vocabulary with other words, and a co-occurrence matrix is ​​built. The GloVe algorithm is used for training to obtain the vector representation of each word.

[0016] A graph database (Neo4j) was established and three types of nodes were created in it corresponding to the three sub-databases. The relationships were defined and the graph structure was constructed based on the entities in the equipment performance sub-database, equipment problem and treatment sub-database, and quality control standard sub-database. The graph structure was input into the R-GCN model and trained in the model using random negative sampling to obtain the vector representation of each node.

[0017] By calculating the cosine similarity between two vectors, the relationship weights between each relationship in the graph database are determined.

[0018] According to nodes, relationships and relationship weights, a knowledge graph is established in the graph database to build a quality control knowledge base;

[0019] S6. Integrate and optimize the constructed knowledge graph, and semantically annotate the entities and relationships therein.

[0020] Establish a knowledge base indexing mechanism to classify and index the nodes in the knowledge graph.

[0021] Develop a knowledge base query interface and establish corresponding query interfaces based on the three sub-bases: equipment performance query, fault diagnosis query, and quality control standard query. When one of the sub-base data is mobilized, the knowledge base will mobilize the other two sub-base data based on the relationship weight and arrange and display them according to the size of the relationship weight.

[0022] And build a knowledge graph visualization interface to support users to interact with nodes and relationships in the graph.

[0023] Preferably, the two-stage pretreatment in S2 comprises the following steps:

[0024] S21, a processing step, based on a preset set of rules, using machine learning algorithms to learn the normal patterns of data, intelligently filter the data, and divide the filtered data into multiple data sets,

[0025] S22, second stage processing, setting block standards according to different data sets, dividing multiple data sets into different data blocks by analyzing the data, generating normal data block markers, potential abnormal data block markers and missing data block markers according to the block standards, and marking them with the corresponding data blocks,

[0026] The preset rule set includes equipment parameter range rules, data format and integrity rules, data frequency rules and historical data statistical regularity rules.

[0027] Preferably, the data set includes equipment performance parameter data set, performance stability data set, performance and environment relationship data set, equipment failure data set, problem warning data set, problem processing history data set, quality control indicator data set, quality control result data set and quality control deviation analysis data set.

[0028] Preferably, the combination of the analysis models in S3 comprises the following steps:

[0029] S31, the autoencoder (AE) includes an encoder and a decoder. For data in a potential abnormal data block and a data block that needs further inspection, the encoder maps high-dimensional input data to a low-dimensional feature space, and the decoder reconstructs the original data from the low-dimensional feature space. When the difference between the encoder's reconstructed structure and the original data exceeds a set threshold, the data is determined to be suspected abnormal data;

[0030] S32: Randomly use variational autoencoder (VAE) or / and generative adversarial network (CAN) to reanalyze the data that is suspected to be abnormal by the autoencoder (AE).

[0031] The variational autoencoder (VAE) performs probability modeling on potential abnormal data, assuming that the suspected abnormal data follows a prior distribution, and calculates the probability of the suspected abnormal data under the prior distribution. If the obtained probability structure is lower than the set threshold, the data is judged to be abnormal data.

[0032] Generative adversarial networks (GANs) consist of a generator and a discriminator. The generator generates normal-like data based on the patterns learned from normal data blocks, and the discriminator determines abnormal data by judging the difference between suspected abnormal data and generated normal data.

[0033] S33, calculate and analyze abnormal data using Z-score method and IQR method,

[0034] If the Z-score of a data point exceeds the range of ±3 and also exceeds the upper and lower limits of the IQR method, the data point is judged to be abnormal data and is removed.

[0035] If only the Z-score of a data point exceeds the range of ±3 / exceeds the upper and lower limits of the IQR method, the data point is analyzed based on its physical meaning and historical data, and attempts are made to reasonably correct, eliminate, or manually intervene in the data point: if it can be corrected, the corrected data is retained; if it cannot be reasonably corrected, the data is eliminated; if it is difficult to determine whether to correct or eliminate based on the physical meaning and historical data, it is marked as a "value to be noted" and pushed to the operator, who then performs manual intervention.

[0036] Preferably, the calculation formula of the Z-score method is: , where x is the value of the data point, is the mean of the data block, is the standard deviation of the data block;

[0037] Sort the data in the data block from small to large and calculate the first quartile (i.e. the value at the 25th percentile in the data) and the third quartile (i.e. the value at 75% of the data),

[0038] The calculation method of IQR method is , where the lower limit is , the upper limit is .

[0039] Preferably, in S5, a distributed computing method is used to calculate the vector representation of each word and each node, specifically:

[0040] The text data of the three sub-databases are divided into multiple data sets, and the GloVe algorithm is used to construct the co-occurrence matrix and calculate the word vector on multiple data sets at the same time, and finally the results are merged.

[0041] The graph structure data is divided into multiple subgraphs, each subgraph is trained with the R-GCN model at the same time, and then the results of each subgraph are merged through the All-Reduce algorithm.

[0042] Preferably, S6 also includes constructing a knowledge base update mechanism: regularly collecting data and knowledge according to a set time, evaluating the importance and necessity of the newly collected data and knowledge,

[0043] Update the new data and knowledge in a small scale, analyze the impact on the overall structure and relationship weights of the knowledge graph, evaluate the importance of the new data, and update the overall data or abandon the update according to the level of importance.

[0044] The levels range from small to large: no association, slight impact, and severe impact.

[0045] If the new data is assessed as not relevant, the update will be abandoned. If the new data is assessed as having a slight impact, a unified update will be performed according to the set time. If the new data is assessed as having a severe impact, the update will be given priority.

[0046] A system for constructing a medical accelerator quality control knowledge base using a large language model, using the above-mentioned method for constructing a medical accelerator quality control knowledge base using a large language model, including an acquisition module, a two-stage preprocessing module, an outlier detection module, a large language training module, a knowledge graph module and a knowledge base index and query module,

[0047] The acquisition module is used to collect data from different data sources and uniformly format them.

[0048] The two-stage preprocessing module is used to intelligently screen the data, divide the data into multiple data blocks, generate normal data markers, potential abnormal data markers and data markers that need further inspection, and mark the corresponding data blocks.

[0049] The outlier detection module is used to build an analysis model and perform outlier detection on data involving equipment performance parameters in potential abnormal data blocks and data blocks that need further inspection.

[0050] The large language training module is used to build a large language model and use the processed data to perform model training verification.

[0051] The knowledge graph module is used to construct sub-libraries and graph databases, and combines the trained large language model with the knowledge graph according to nodes, relationships and relationship weights to construct a quality control knowledge base.

[0052] The knowledge base indexing and query module is used to build a data indexing mechanism, develop a knowledge base query interface, and construct a knowledge graph visualization interface.

[0053] Preferably, the two-stage pre-processing module comprises a first-stage processing unit and a second-stage processing unit.

[0054] The first processing unit is used to filter and divide the data into multiple data sets.

[0055] The second-stage processing unit is used to analyze multiple data sets and divide them into normal data blocks, potential abnormal data blocks and data blocks that need further inspection.

[0056] Preferably, the knowledge graph module includes an equipment performance sub-library, an equipment problem and treatment method sub-library, a quality control standard sub-library and a graph database.

[0057] The equipment performance sub-library is used to store medical accelerator parameter performance data, medical accelerator performance stability data, and medical accelerator performance and environment related data.

[0058] The equipment problem and treatment method sub-library is used to store medical accelerator fault data, medical accelerator problem warning data and problem treatment history data.

[0059] The quality control standard sub-library is used to store medical accelerator quality control indicator data, medical accelerator quality control result data and medical accelerator quality control deviation analysis data.

[0060] Beneficial effects of the present invention:

[0061] 1. The present invention collects medical accelerator-related data through multiple channels, which can cover information on medical accelerators from equipment performance, failures to quality control, etc., making the constructed knowledge base comprehensive. Intelligent screening and layered processing of data are adopted to help process data in different states more accurately; and the pre-set rule set includes equipment parameter range rules, data format and integrity rules, data frequency rules and historical data statistical law rules, etc., which fully considers various characteristics of medical accelerator data, helps to improve the accuracy of data screening and segmentation, and can better adapt to the complexity and professionalism of medical accelerator data.

[0062] 2. The present invention uses an autoencoder (AE), a variational autoencoder (VAE) and a generative adversarial network (GAN) to construct an analysis model for outlier detection. The combination of multiple models can make full use of the advantages of different models, detect data involving equipment performance parameters in potential abnormal data blocks and data blocks that need further inspection from multiple angles, improve the accuracy of outlier detection, and the three models complement each other and can more comprehensively detect various types of abnormal data. At the same time, for each data in the data block determined to be abnormal by the analysis model, an outlier detection method (Z-score method and IQR method) is used for detection. The accuracy of outlier judgment is further ensured through a double verification mechanism.

[0063] 3. The present invention divides the processed data into three sub-databases, builds a vocabulary by merging text data, builds a co-occurrence matrix, uses the GloVe algorithm to train to obtain word vector representation, then establishes a graph database and constructs a graph structure, and finally inputs it into the R-GCN model to train to obtain node vector representation and other steps, which can better dig into the semantic information in the data and lay the foundation for building an accurate knowledge graph. And by calculating the cosine similarity between vectors to determine the relationship weights between each relationship in the graph database, the importance of the relationship can be quantified, making the relationship representation in the knowledge graph more accurate. At the same time, in the process of building the knowledge graph, a distributed computing method is used to calculate the vector representation of each word and each node, which improves the computing efficiency and is particularly suitable for the processing of large-scale data.

[0064] 4. The present invention establishes a knowledge graph in a graph database based on nodes, relationships and relationship weights, constructs a quality control knowledge base, and integrates data and knowledge from all aspects of medical accelerators, including equipment performance, equipment problem handling and quality control standards, to form a comprehensive knowledge base. A knowledge base indexing mechanism is established to classify and index the nodes in the knowledge graph, and query interfaces corresponding to different sub-libraries are developed. When querying a sub-library data, the remaining sub-library data can be mobilized according to the relationship weight and displayed in order of weight, which helps users to obtain relevant knowledge more conveniently. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art are briefly introduced below.

[0066] Figure 1 It is a flow chart of the method for constructing a medical accelerator quality control knowledge base of the present invention.

[0067] Figure 2 It is a flow chart of the steps of the two-stage pretreatment of the present invention.

[0068] Figure 3It is a flow chart of the steps of combining the analysis models of the present invention.

[0069] Figure 4 It is a structural block diagram of the medical accelerator quality control knowledge base of the present invention. DETAILED DESCRIPTION

[0070] Embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, but should not be construed as limiting the present invention.

[0071] like Figure 1-Figure 3 As shown, a method for constructing a medical accelerator quality control knowledge base using a large language model includes the following steps:

[0072] S1. Collect medical accelerator-related data through multiple channels and format the data from different channels in a unified manner;

[0073] S2. Perform two-stage preprocessing on the collected data, intelligently screen the data, analyze the screened data according to the set standards and divide it into multiple data blocks, check the integrity of the analyzed data blocks and divide them into normal data blocks, potential abnormal data blocks and data blocks that need further inspection, generate normal data tags, potential abnormal data tags and data tags that need further inspection, and mark the corresponding data blocks.

[0074] The normal data mark indicates that the data integrity is good and within the normal range, the potential abnormal data mark indicates that the data is close to the normal range boundary, and the data mark that needs further inspection indicates that the data has integrity problems;

[0075] S3, using autoencoders (AE), variational autoencoders (VAE) and generative adversarial networks (GAN) to build an analysis model to detect outliers in potential abnormal data blocks and data blocks that require further inspection involving equipment performance parameters.

[0076] At the same time, outlier detection methods (Z-score method and IQR method) are used to detect outliers for each data in the data block determined to be abnormal by the analysis model:

[0077] If the Z-score of the data exceeds ±3 and the IQR range, it is determined to be an outlier, and the analysis model removes the outlier. If the Z-score of the data does not exceed ±3 / does not exceed the IQR range, it is determined to be a value to be processed and analyzed and processed (corrected / eliminated) according to the physical meaning of the data and historical data;

[0078] S4. Automated data verification. Using a stratified random partitioning method, the data after preprocessing and outlier detection is divided into a training set, a validation set, and a test set. The training set is used to build the training model, the validation set is used to adjust the model, and the test set is used for performance evaluation.

[0079] A monitoring mechanism is built to analyze the test results in detail when they do not meet the predetermined standards, and to feed the analysis results back to the analysis model or model training model to make targeted adjustments to the model structure.

[0080] S5. Divide the processed data into three sub-libraries based on quality control knowledge: equipment performance sub-library, equipment problem and treatment sub-library, and quality control standard sub-library.

[0081] All text data in the three sub-databases are merged to build a vocabulary. All text data are traversed to count the number of co-occurrences of each word in the vocabulary with other words, and a co-occurrence matrix is ​​built. The GloVe algorithm is used for training to obtain the vector representation of each word.

[0082] A graph database (Neo4j) was established and three types of nodes were created in it corresponding to the three sub-databases. The relationships were defined and the graph structure was constructed based on the entities in the equipment performance sub-database, equipment problem and treatment sub-database, and quality control standard sub-database. The graph structure was input into the R-GCN model and trained in the model using random negative sampling to obtain the vector representation of each node.

[0083] By calculating the cosine similarity between two vectors, the relationship weights between each relationship in the graph database are determined.

[0084] According to nodes, relationships and relationship weights, a knowledge graph is established in the graph database to build a quality control knowledge base;

[0085] S6. Integrate and optimize the constructed knowledge graph, and semantically annotate the entities and relationships therein.

[0086] Establish a knowledge base indexing mechanism to classify and index the nodes in the knowledge graph.

[0087] Develop a knowledge base query interface and establish corresponding query interfaces based on the three sub-bases: equipment performance query, fault diagnosis query, and quality control standard query. When one of the sub-base data is mobilized, the knowledge base will mobilize the other two sub-base data based on the relationship weight and arrange and display them according to the size of the relationship weight.

[0088] And build a knowledge graph visualization interface to support users to interact with nodes and relationships in the graph.

[0089] The present invention uses a combination of autoencoders (AE), variational autoencoders (VAE) and generative adversarial networks (GAN) to detect outliers. This combination can give full play to the advantages of each model, and can provide more comprehensive and accurate outlier detection results when detecting potential abnormal data blocks and data involving equipment performance parameters in data blocks that need further inspection. At the same time, combined with the traditional Z-score method and IQR method, each data in the data block determined to be abnormal by the analysis model is re-tested. The traditional method has the advantages of being simple, intuitive and easy to understand, while the modern method has stronger adaptability and accuracy. The combination of the two can ensure the accuracy of outlier detection while also facilitating the understanding and operation of personnel with different technical backgrounds.

[0090] In addition, in the process of knowledge graph construction, the semantic information of the data is deeply mined through multiple steps, from data sub-library division to vocabulary construction, co-occurrence matrix calculation, vector representation obtained by using GloVe algorithm and R-GCN model, and finally the relationship weight is determined to construct the knowledge graph. Multi-step semantic mining can accurately construct a knowledge graph that reflects the various knowledge relationships of medical accelerators.

[0091] Furthermore, the two-stage preprocessing in S2 includes the following steps:

[0092] S21, a processing step, based on a preset set of rules, using machine learning algorithms to learn the normal patterns of data, intelligently filter the data, and divide the filtered data into multiple data sets,

[0093] S22, second stage processing, setting block standards according to different data sets, dividing multiple data sets into different data blocks by analyzing the data, generating normal data block markers, potential abnormal data block markers and missing data block markers according to the block standards, and marking them with the corresponding data blocks,

[0094] Among them, the preset rule set includes equipment parameter range rules, data format and integrity rules, data frequency rules and historical data statistical law rules.

[0095] The two-stage preprocessing of the present invention can process the data comprehensively and meticulously. First, the first stage of processing uses a machine learning algorithm to learn the normal pattern of the data based on a preset rule set and perform intelligent screening to divide the data into multiple data sets. This step can remove data that obviously does not meet the requirements based on a variety of rules, thereby improving the quality and availability of the data. Secondly, the second stage of processing sets the block standard according to different data sets, further divides the data set into different data blocks, and generates normal data block markers, potential abnormal data block markers, and missing data block markers to mark the corresponding data blocks. The hierarchical labeling method enables the data to be clearly classified, which facilitates the subsequent adoption of different processing strategies for different types of data blocks. Compared with the traditional single preprocessing method, it can more accurately identify the state of the data (normal, potentially abnormal, missing, etc.), providing a higher quality data foundation for subsequent data processing links.

[0096] Furthermore, the equipment parameter range rule of the present invention refers to the reasonable range of various parameters of the equipment when it is operating normally, including but not limited to the operating temperature, operating pressure, operating frequency, operating current, etc. of the equipment. For example, the operating temperature range of a medical accelerator may be 20-25 degrees Celsius, and the common operating frequency may be 3000MHz. These parameter ranges are determined by the design and manufacture of the equipment, and exceeding these ranges may cause the performance of the equipment to degrade or be damaged. The data format and integrity rules refer to the format and integrity requirements that data should follow during the collection, storage and transmission process. The data format rules generally include the type of data (such as integer, floating point number, string, etc.), the length of the data, the precision of the data, etc. For example, a parameter of a medical accelerator may require storage in floating point format and retain two decimal places. The data integrity rule requires that the data cannot be lost or damaged during the collection and transmission process, otherwise it may affect the subsequent data analysis and processing. Similarly, the data frequency rule refers to the rule that the frequency of data collection should follow. In the construction of the quality control knowledge base of the medical accelerator, the data frequency rule may require that certain key parameters need to be collected once every certain time (such as every minute, every hour, etc.) in order to monitor the operating status of the equipment in a timely manner. The rules of historical data statistics refer to finding the changing rules and trends of data through statistical analysis of historical data, so as to predict and analyze future data. For example, by statistically analyzing the operating data of medical accelerators in the past year, we can find out the operating rules of the equipment in different seasons and time periods, as well as the changing trends of equipment performance parameters. These rules and trends can help medical personnel better manage and maintain equipment and improve the efficiency and safety of equipment use.

[0097] Similarly, the sources of data can be diverse, including but not limited to the following:

[0098] Various sensors of medical accelerators collect real-time data (such as temperature sensors, pressure sensors, dose monitors, etc.); interface with the hospital information system to obtain the patient's basic information, treatment plan, and equipment operation log according to a preset schedule; obtain the patient's imaging data from PACS, such as CT, MRI, etc., periodically collect imaging data according to the patient's treatment plan, and compress the imaging data to reduce the amount of data transmission; obtain the fault record of the medical accelerator from the equipment maintenance management system, including the time of the fault, the fault phenomenon (such as equipment error code, abnormal sound, etc.), the troubleshooting process and the corresponding maintenance plan; obtain the quality control standards of medical accelerators from industry standard documents and internal quality control specification documents of the hospital, and update the relevant content in the knowledge base in a timely manner when the quality control standards are updated (such as the industry releases new standards or the hospital adjusts the specifications internally).

[0099] In addition, to ensure the privacy and security of the knowledge base data, the present invention anonymizes the patient's privacy data during data collection and storage. An irreversible hash function (such as SHA-256) is used to process the patient's identity (such as name, ID number, etc.) to ensure that the patient's identity cannot be directly identified during data use. At the same time, a symmetric encryption algorithm (such as AES algorithm) is used to encrypt and store sensitive data in the knowledge base (such as patient privacy information, key parameters of equipment operation, etc.).

[0100] A role-based access control (RBAC) model is established to define different user roles, such as administrators, doctors, researchers, etc. Each role has different access rights. The administrator has the highest authority and can fully manage the knowledge base (including data updates, user permission settings, etc.); doctors can query information related to their patients; researchers can query anonymized data for research purposes. In addition, all user access and operation logs to the knowledge base are recorded, including user login time, login IP address, operation type (such as query, update, delete, etc.), operation object (such as specific data table or entity), and other information.

[0101] In one solution, the data set includes equipment performance parameter data set, performance stability data set, performance and environment relationship data set, equipment failure data set, problem warning data set, problem processing history data set, quality control indicator data set, quality control result data set and quality control deviation analysis data set. By constructing a comprehensive data set, the present invention can simultaneously cover multiple aspects of data such as equipment performance, stability, relationship with the environment, failure, warning, processing history, quality control, etc. The integration of all-round multi-dimensional data provides a solid data foundation for the comprehensive management of equipment, helps to improve the efficiency and quality of equipment management, reduce equipment operating costs, and enhance the reliability and safety of equipment.

[0102] Furthermore, the combination of the analysis model in S3 includes the following steps:

[0103] S31, the autoencoder (AE) includes an encoder and a decoder. For data in a potential abnormal data block and a data block that needs further inspection, the encoder maps high-dimensional input data to a low-dimensional feature space, and the decoder reconstructs the original data from the low-dimensional feature space. When the difference between the encoder's reconstructed structure and the original data exceeds a set threshold, the data is determined to be suspected abnormal data;

[0104] S32: Randomly use variational autoencoder (VAE) or / and generative adversarial network (CAN) to reanalyze the data that is suspected to be abnormal by the autoencoder (AE).

[0105] The variational autoencoder (VAE) performs probability modeling on potential abnormal data, assuming that the suspected abnormal data follows a prior distribution, and calculates the probability of the suspected abnormal data under the prior distribution. If the obtained probability structure is lower than the set threshold, the data is judged to be abnormal data.

[0106] Generative adversarial networks (GANs) consist of a generator and a discriminator. The generator generates normal-like data based on the patterns learned from normal data blocks, and the discriminator determines abnormal data by judging the difference between suspected abnormal data and generated normal data.

[0107] S33, calculate and analyze abnormal data using Z-score method and IQR method,

[0108] If the Z-score of a data point exceeds the range of ±3 and also exceeds the upper and lower limits of the IQR method, the data point is judged to be abnormal data and is removed.

[0109] If only the Z-score of a data point exceeds the range of ±3 / exceeds the upper and lower limits of the IQR method, the data point is analyzed based on its physical meaning and historical data, and attempts are made to reasonably correct, eliminate, or manually intervene in the data point: if it can be corrected, the corrected data is retained; if it cannot be reasonably corrected, the data is eliminated; if it is difficult to determine whether to correct or eliminate based on the physical meaning and historical data, it is marked as a "value to be noted" and pushed to the operator, who then performs manual intervention.

[0110] The present invention constructs a multi-dimensional and multi-level analysis model by combining autoencoders (AE), variational autoencoders (VAE), generative adversarial networks (GAN), Z-scores and IQR methods. Autoencoders (AE), variational autoencoders (VAE) and generative adversarial networks (GAN) perform anomaly detection from different angles such as data feature extraction, probability modeling and comparison with normal data, and then further verify and process them in combination with traditional Z-scores and IQR methods. Multi-level detection logic evaluates data from different theoretical and methodological perspectives to further improve the accuracy of anomaly detection and reduce the possibility of missed judgments and misjudgments. For data in a critical state (only the Z-score exceeds the range of ±3 or only exceeds the upper and lower limits in the IQR method), analysis is performed based on the physical meaning of the data and historical data, and attempts are made to reasonably correct or eliminate it. This fully considers the particularity of the data and avoids generalizing the data. The multi-factor processing method can more accurately process critical data and maximize the value of data.

[0111] Specifically, if the data is determined to be abnormal by both the Z-score method and the IQR method, it is determined to be an abnormal value and removed. If it is determined to be abnormal by only one method, it is marked as a pending value and further analyzed and processed according to the physical meaning and historical data trend.

[0112] When data is input into the autoencoder, the encoder maps the high-dimensional input data x to the low-dimensional feature space z, and the decoder reconstructs the low-dimensional feature space to obtain x1. The difference between the original data x and the reconstructed data x1 is calculated, and the mean square error (MSE) is used as the measurement indicator. . If the MSE exceeds the set threshold T1, the data is judged to be suspected abnormal data. The suspected abnormal data is input into the variational autoencoder and the generative adversarial network for re-analysis. The variational autoencoder calculates the probability p(x|z) of the suspected abnormal data under the prior distribution. If it is lower than the threshold T2, it is judged to be abnormal data; the generative adversarial network generates similar normal data x′ through the generator, and the discriminator judges the difference between the suspected abnormal data x and x′. If the discriminator outputs the probability D(x) lower than the threshold T3, it is judged to be abnormal data.

[0113] A physical model is established for the key performance parameters of medical accelerators. For example, a relationship model is established for parameters such as radiation energy E, acceleration voltage V, and electron beam current I. , (k is a constant determined according to the accelerator structure). Then, according to the physical model and equipment design specifications, determine the reasonable data range of each parameter. When the data deviates from the normal range but is within the correctable range (such as the deviation does not exceed 10% 10%), make corrections according to the physical model. For example, if the radiation energy E measured Deviation from theoretical value Etheoretical Within 10%, adjust the relevant parameters (such as V or I) according to the physical model for correction. If the data deviates from the normal range beyond the correctable range (such as more than 10%), or violates the laws of physics (such as energy increases as voltage decreases), the data will be discarded.

[0114] Furthermore, the trend of the equipment performance parameter P is calculated using the moving average method, and the formula is: , ( P MAN is the n-period moving average P i is the parameter value of the i-th period). And according to the fluctuation of historical data, the threshold of trend change is determined. For example, for the moving average of equipment performance parameters, if the change rate of two adjacent moving averages exceeds 20%, it is determined that the trend has changed significantly. The present invention sets appropriate trend determination thresholds for different types of data sets according to their data characteristics and historical fluctuation ranges. When the newly collected data deviates from the trend of historical data, if the deviation is within a certain range (such as less than 15%) and similar deviations in history are mostly normal fluctuations, the data is corrected, and the correction method can be adjusted according to the average deviation of historical data; if the deviation exceeds a certain range (such as greater than 15%) or similar deviations in history are mostly related to equipment failures, consider eliminating the data.

[0115] When the judgment result of the analysis model (the judgment of the autoencoder, variational autoencoder, and generative adversarial network) conflicts with the judgment results of the Z-score method and the IQR method, manual intervention is triggered. For example, in the data detection of medical accelerators, the analysis model (such as the autoencoder) determines that a certain electron beam data is normal data based on internal mechanisms such as the reconstruction error of the data. However, from a physical point of view, according to the established physical model , (I is the electron beam current, E is the radiation energy, V is the acceleration voltage, and k is a constant). When the current radiation energy E and acceleration voltage V are known, it is calculated that the electron beam current data deviates from the theoretical value by more than 10% (according to the physical meaning, it should be abnormal data). In this case, manual intervention is performed. The operator makes decisions based on his or her own professional knowledge and experience, combined with the information provided by the system. If the data is corrected, the basis for the correction needs to be clear, such as correction based on the theoretical model of the equipment or the known normal data range. If the data is excluded, the reason for the exclusion should be recorded, such as the data obviously does not conform to the normal operating logic of the equipment or the data is abnormal due to the failure of the acquisition equipment. The system records the process of manual intervention in detail, including the identity information of the operator, detailed information of the data, the original judgment result of the model, the judgment result after manual intervention, and the basis for intervention.

[0116] In one embodiment, the Z-score is calculated as , where x is the value of the data point, is the mean of the data block, is the standard deviation of the data block;

[0117] Sort the data in the data block from small to large and calculate the first quartile (i.e. the value at the 25th percentile in the data) and the third quartile (i.e. the value at 75% of the data),

[0118] The calculation method of IQR method is , where the lower limit is , the upper limit is .

[0119] The Z-score method and the IQR method of the present invention are used in combination to give play to their respective advantages. The Z-score method is based on the mean and standard deviation, and is more sensitive to the overall distribution characteristics of the data; the IQR method is based on quartiles, and has a better grasp of the distribution of the middle part and both ends of the data. When the two methods are applied simultaneously, if a data point is determined to be abnormal by both methods (such as the Z-score of the data point exceeds the range of ±3, and also exceeds the upper and lower limits in the IQR method), then the possibility that the data point is abnormal data is higher. This combination can more accurately identify abnormal data, overcome the limitations of a single method, thereby improving data quality, and providing a more reliable data basis for subsequent data-based analysis, decision-making, etc.

[0120] In one embodiment, the present invention performs model evaluation on the analysis model, and adopts a cross-validation method to divide the data set into K (such as K = 5) mutually exclusive subsets. Each time, one of the subsets is selected as the test set, and the remaining K-1 subsets are used as the training set. Repeat K times to obtain K evaluation results, and finally take the average value as the final evaluation result of the model. During the evaluation process, the value of each evaluation index is recorded, and a curve (such as an accuracy-recall curve) is drawn to intuitively display the performance of the model. At the same time, the performance of different models (individual autoencoders, VAEs, GANs, and combined models) under the same evaluation indicators is compared, and the model with the best performance is selected for practical application. Among them, the evaluation indicators include accuracy, recall, F1 value, mean square error, and root mean square error.

[0121] In one solution, the present invention establishes a correction algorithm library, which includes linear interpolation, mean replacement and decision tree correction algorithms. The data of the medical accelerator are classified according to the data type. For example, for the temperature data of the device, a smoothing algorithm based on time series can be used for correction; for the power data of the device, when an abnormal value appears, a correction algorithm based on the average power of similar devices can be used. Secondly, a suitable algorithm is selected according to the severity of the abnormal situation and the distribution characteristics of the data. For example, if the abnormal value deviates from the normal range by a small amount and the data is normally distributed, the mean replacement algorithm can be preferred; if the abnormal value deviates greatly and the data has obvious seasonal or periodic characteristics, such as the energy consumption data of the device in a specific time period, a correction algorithm based on time series analysis is used. Similarly, taking the voltage data of the medical accelerator as an example, if the voltage value of a certain data point is found to be abnormally high. First, analyze the characteristics of the abnormal value. If it is an isolated abnormal point and the surrounding data is approximately normally distributed. At this time, the mean replacement algorithm can be selected from the algorithm library. Calculate the mean of the normal voltage values ​​around the data point, replace the abnormal value with this mean, and complete the data correction.

[0122] In one solution, the present invention has established a strict recording and auditing mechanism to record in detail the relevant information of data exclusion, including the source of the data (which medical accelerator, which sensor collected it), the time of data collection, the specific value of the abnormal value, and the reason for exclusion (how many times it exceeds the normal range, whether it is related to the known failure mode of the device, etc.). Secondly, a multi-level audit process is established. For example, the data acquisition personnel first make a preliminary judgment and propose exclusion suggestions, then the data quality management personnel conduct a secondary review, and finally a professional medical equipment engineer conducts a final review. Only data that has passed the three-level review can be eliminated. Finally, a manual intervention backtracking of specific data is constructed, and the conditions and processes of the backtracking are clarified. For example, when it is found that a certain excluded data may have an important impact on the subsequent equipment performance analysis or fault diagnosis, the backtracking mechanism can be started. Specifically, the backtracking process includes submitting a backtracking application, stating the reason for the backtracking and the expected impact; after review and approval, the data is restored from the backup data and re-analyzed. For example, in the fault diagnosis of the medical accelerator, a seemingly abnormal device parameter data was previously eliminated. However, in subsequent analysis, it was found that this data might be related to a new failure mode. At this time, manual intervention was performed to recover the data.

[0123] In one solution, in S5, a distributed computing method is used to calculate the vector representation of each word and each node, specifically:

[0124] The text data of the three sub-databases are divided into multiple data sets, and the GloVe algorithm is used to construct the co-occurrence matrix and calculate the word vector on multiple data sets at the same time, and finally the results are merged.

[0125] The graph structure data is divided into multiple subgraphs, each subgraph is trained with the R-GCN model at the same time, and then the results of each subgraph are merged through the All-Reduce algorithm.

[0126] The present invention adopts a distributed computing method in S5, by dividing the text data of the three sub-libraries into multiple data sets, and then simultaneously performing the construction of the co-occurrence matrix and word vector calculation of the GloVe algorithm on multiple data sets. This parallel processing method can make full use of computing resources and greatly shorten the calculation time. For example, if distributed computing is not adopted, according to the traditional sequential computing method, the entire text data needs to be processed in sequence. When the amount of data is very large, the traditional centralized computing may not be able to be effectively processed due to memory limitations or long computing time. Distributed computing allows the calculation of multiple data sets to be performed simultaneously, and finally the results are merged, which improves the overall computing efficiency. For graph structure data, it is divided into multiple sub-graphs for processing, so that the system can cope with large-scale graph data. As the data scale grows, the difficulty of processing graph structure data will also increase. The distributed processing method can improve the system's processing ability for large-scale graph data without sacrificing calculation accuracy.

[0127] In one solution, during the construction of the knowledge graph of the present invention, multiple domain experts and data analysts can be organized to review entity relationships. Through multiple rounds of discussion and voting, consensus on entity relationships can be reached, thereby reducing subjectivity. For example, different experts may have different opinions on the relationship between "fault type" and "repair method" in the equipment problem and treatment method library. The most reasonable relationship definition can be determined through multi-person review.

[0128] Furthermore, S6 also includes building a knowledge base update mechanism: regularly collecting data and knowledge according to a set time, evaluating the importance and necessity of the newly collected data and knowledge,

[0129] Update the new data and knowledge in a small scale, analyze the impact on the overall structure and relationship weights of the knowledge graph, evaluate the importance of the new data, and update the overall data or abandon the update according to the level of importance.

[0130] The levels range from small to large: no association, slight impact, and severe impact.

[0131] If the new data is assessed as not relevant, the update will be abandoned. If the new data is assessed as having a slight impact, a unified update will be performed according to the set time. If the new data is assessed as having a severe impact, the update will be given priority.

[0132] In one solution, an incremental learning method is used to further improve the knowledge base update mechanism. Specifically, new text data related to equipment performance, equipment problem handling, and quality control standards are continuously collected. These data sources can include new equipment inspection reports, equipment maintenance records, quality control standard update documents, etc. The newly collected data is preprocessed, such as cleaning (removing noise, erroneous data, etc.) and standardization (unifying data formats, etc.) operations to keep it consistent with the data format when the knowledge graph was previously constructed.

[0133] Add the words in the new data to the existing vocabulary. At the same time, recalculate the number of co-occurrences of the new words with other words in the vocabulary and update the co-occurrence matrix. For example, if the word "smart device" appears in the new data, add it to the vocabulary and count its co-occurrences with existing words such as "device performance" and "device failure", and update the co-occurrence matrix accordingly.

[0134] Extract entity relationships from new data and update the graph structure in the graph database based on the new entity relationships. If a relationship is found between a new device performance indicator and an existing device problem, create the corresponding node and relationship in the graph database. For example, if a new device inspection report indicates that a new performance indicator is associated with a specific device failure, add a node representing the performance indicator to the graph and establish a relationship with the node representing the device failure.

[0135] The GloVe algorithm is retrained using the updated co-occurrence matrix to obtain updated word vector representations. The updated graph structure is input into the R-GCN model and retrained using random negative sampling to obtain updated node vector representations.

[0136] Evaluate the importance of new data based on the degree of data change and the potential impact on the entire knowledge graph. For very important data (such as updates involving key equipment performance indicators or new discoveries of serious equipment problems), give higher priority to knowledge updates. For example, if new data shows that the key performance indicators of a certain equipment have changed significantly, this will have a profound impact on the operation, maintenance and quality control of the equipment, and these data should be processed first to update the knowledge graph. In addition, selective updates are performed based on the importance and frequency of occurrence of words. For words with extremely low frequency of occurrence and little relevance to existing knowledge, they can be temporarily excluded from the vocabulary to reduce unnecessary calculations. For example, some uncommon words that appear occasionally in new data can be excluded from the vocabulary if they are not directly related to equipment performance, problems and quality control standards. Secondly, when updating the graph structure, adopt a progressive approach. For newly discovered entity relationships, first perform a small-scale test update to evaluate the impact on the overall structure and relationship weights of the knowledge graph. If the impact is within an acceptable range, then perform a comprehensive graph structure update. For example, when a possible relationship between a new device problem and an existing device performance indicator is discovered, this relationship is first added to the local graph structure, the changes in the weights of the relationship with the surrounding nodes are observed, and then a decision is made whether to generalize to the entire graph structure. Finally, for the GloVe and R-GCN models, a fine-tuning approach is adopted. Using the existing model parameters, only the parts related to the new data are adjusted. For example, when updating the GloVe model, only the model parameters related to new words or words with significant changes in co-occurrence relationships are adjusted according to the new co-occurrence matrix; for the R-GCN model, only the model parameters related to the nodes and edges involving the new relationships are fine-tuned according to the new graph structure changes.

[0137] In one solution, the knowledge base update mechanism can also regularly check the consistency of the knowledge graph, including the logical consistency of node relationships, the rationality of relationship weights, etc. If inconsistencies are found, they are repaired in a timely manner. For example, check whether the relationship weight between device performance degradation and device problem occurrence is consistent with the actual situation. If there is an unreasonable weight, re-evaluate and adjust it.

[0138] In one solution, this knowledge base update mechanism can also establish a user feedback channel, allowing equipment operators, quality control personnel and other users who use the knowledge graph to feedback problems or unreasonable aspects they find during use. For example, a user may find that the description of the method for handling a certain equipment problem in the knowledge graph does not match the actual operation. Based on user feedback, the knowledge graph is adjusted and improved in a timely manner. If multiple users report that a certain relationship is wrong, the relevant vocabulary, graph structure and model are re-evaluated and updated.

[0139] In the knowledge base update mechanism of S6, the present invention can effectively screen out valuable data by evaluating the importance and necessity of newly collected data and knowledge. It avoids blindly updating the knowledge base and ensures that only data that has a positive meaning to the knowledge base will be considered for updating. For example, for those data that are not associated with the knowledge base (assessed as not associated), the update is directly abandoned, thereby saving computing resources and storage space and preventing irrelevant information from being mixed into the knowledge base. For data with a slight impact, it can be uniformly updated according to the set time. While ensuring the update of the knowledge base, the stability of the knowledge base will not be affected by frequent small updates, which effectively balances the stability and timeliness of the knowledge base structure. For data with a heavy impact, priority is given to updating, and data that has a significant impact on the knowledge base structure and relationship weights can be included in the knowledge base in a timely manner to ensure the timeliness and accuracy of the knowledge base. The stability of the knowledge base structure is very important for applications that rely on the knowledge base for query, analysis and other operations. If the update is too frequent or random, the structural integrity of the knowledge base may be destroyed, resulting in problems such as inaccurate query results. However, if it is not updated in time, the knowledge base will lose its timeliness. The update strategy of this system treats different types of data updates differently by evaluating the impact level of new data. It can not only maintain the relative stability of the knowledge base structure, but also update important information in a timely manner, thereby improving the overall effectiveness of the knowledge base.

[0140] like Figure 4 As shown, a system for constructing a medical accelerator quality control knowledge base using a large language model adopts the above-mentioned method for constructing a medical accelerator quality control knowledge base using a large language model, including an acquisition module, a two-stage preprocessing module, an outlier detection module, a large language training module, a knowledge graph module and a knowledge base index and query module.

[0141] The acquisition module is used to collect data from different data sources and format them uniformly.

[0142] The two-stage preprocessing module is used to intelligently screen the data and divide the data into multiple data blocks, generate normal data markers, potential abnormal data markers, and data markers that need further inspection, and mark the corresponding data blocks.

[0143] The outlier detection module is used to build an analysis model and perform outlier detection on data involving equipment performance parameters in potential abnormal data blocks and data blocks that need further inspection.

[0144] The large language training module is used to build a large language model and use the processed data to perform model training and verification.

[0145] The knowledge graph module is used to build sub-libraries and graph databases, and combines the trained large language model with the knowledge graph according to nodes, relationships, and relationship weights to build a quality control knowledge base.

[0146] The knowledge base indexing and query module is used to build a data indexing mechanism, develop a knowledge base query interface, and construct a knowledge graph visualization interface.

[0147] When constructing the medical accelerator quality control knowledge base, the combination of the large language model and the knowledge graph can integrate the semantic processing ability of the large language model and the knowledge structure representation ability of the knowledge graph. In the medical field, there is a large amount of natural language information (such as medical reports, operating manuals, etc.). The large language model can effectively process this information, and the knowledge graph can organize this information in a structured way. At the same time, the system covers a complete range of functional modules from data acquisition, preprocessing, outlier detection, model training to knowledge base construction and index query, which can ensure the accuracy, completeness and ease of use of the knowledge base. The various functional modules cooperate with each other to form an organic whole.

[0148] Furthermore, the two-stage pre-processing module includes a first-stage processing unit and a second-stage processing unit.

[0149] A processing unit is used to filter and divide the data into multiple data sets.

[0150] The second-stage processing unit is used to analyze multiple data sets and divide them into normal data blocks, potential abnormal data blocks and data blocks that need further inspection.

[0151] The first-stage processing unit of the present invention screens the data and divides it into multiple data sets, which can realize the preliminary arrangement of the original data. In this way, the large-scale original data is decomposed into multiple relatively small data sets, which helps to reduce the complexity of data processing and improve the efficiency of subsequent processing. For example, when processing a large amount of medical data, dividing it into multiple data sets can make each data set easier to manage in subsequent processing, reducing processing time and resource consumption. On the basis of the first-stage processing, the second-stage processing unit further analyzes and divides the multiple data sets into normal data blocks, potential abnormal data blocks and data blocks that need further inspection. This precise classification can process different types of data more targetedly. For normal data blocks, they can be directly used for subsequent normal process operations; for potential abnormal data blocks and data blocks that need further inspection, more in-depth detection and analysis can be performed. The two-stage processing of the two-stage preprocessing module is synergistic, and the preliminary screening of the first-stage processing unit provides a basis for the precise classification of the second-stage processing unit; the classification results of the second-stage processing unit can be fed back to the entire system, so that the system can adopt different processing strategies according to different types of data blocks. This collaborative performance improves the efficiency of the data processing flow of the entire system, avoids unnecessary duplication of operations, and makes the conversion of data from its original state to a state that can be used for subsequent processes more reasonable and efficient.

[0152] Furthermore, the knowledge graph module includes the equipment performance sub-library, the equipment problem and treatment method sub-library, the quality control standard sub-library and the graph database.

[0153] The equipment performance sub-library is used to store medical accelerator parameter performance data, medical accelerator performance stability data, and medical accelerator performance and environment related data.

[0154] The equipment problem and solution sub-library is used to store medical accelerator fault data, medical accelerator problem warning data and problem processing history data.

[0155] The quality control standard sub-library is used to store medical accelerator quality control indicator data, medical accelerator quality control result data and medical accelerator quality control deviation analysis data.

[0156] The equipment performance sub-library, equipment problem and treatment method sub-library and quality control standard sub-library in the knowledge graph module of the present invention store different types of data respectively. This classified storage method can improve the efficiency of data management. For example, when it is necessary to query the performance-related data of the medical accelerator, it can be directly obtained from the equipment performance sub-library without searching in the entire data set, which greatly shortens the query time. For different types of data, such as medical accelerator fault data stored in the equipment problem and treatment method sub-library, quality control indicator data stored in the quality control standard sub-library, etc., this classification makes the data organization more orderly. In the case of a large amount of data, orderly data organization helps to reduce the complexity of data storage and retrieval and improve the operating efficiency of the entire system. The data in the equipment performance sub-library, the equipment problem and treatment method sub-library and the quality control standard sub-library are used to construct a graph database to form a knowledge graph. By integrating different types of medical accelerator-related data to construct a knowledge graph, the potential relationship between the data can be mined. For example, the relationship between the performance data of the medical accelerator and the quality control result data, or the relationship between the fault data and the quality control deviation analysis data, etc. can be found through the knowledge graph.

[0157] What is disclosed above is only one or more preferred embodiments of the present application, and cannot be used to limit the scope of rights of the present application. Ordinary technicians in this field can understand that all or part of the processes of implementing the above embodiments and equivalent changes made according to the claims of the present application are still within the scope covered by the present application.

Claims

1. A method for constructing a medical accelerator quality control knowledge base using a large language model, characterized in that: The following steps are involved: S1. Collect medical accelerator-related data through multiple channels and format the data from different channels in a unified manner; S2. Perform two-stage preprocessing on the collected data, intelligently screen the data, analyze the screened data according to the set standards and divide it into multiple data blocks, check the integrity of the analyzed data blocks and divide them into normal data blocks, potential abnormal data blocks and data blocks that need further inspection, generate normal data tags, potential abnormal data tags and data tags that need further inspection, and mark the corresponding data blocks. The normal data mark indicates that the data integrity is good and within the normal range, the potential abnormal data mark indicates that the data is close to the normal range boundary, and the data mark that needs further inspection indicates that the data has integrity problems; S3, using autoencoders, variational autoencoders and generative adversarial networks to build an analysis model, and detect outliers in potential abnormal data blocks and data blocks that need further inspection involving equipment performance parameters. At the same time, the Z-score method and IQR method are used to detect outliers for each data in the data block determined to be abnormal by the analysis model: If the Z-score of the data exceeds ±3 and the IQR range, it is determined to be an outlier, and the analysis model removes the outlier. If the Z-score of the data does not exceed ±3 / does not exceed the IQR range, it is determined to be a value to be processed and analyzed and processed according to the physical meaning of the data and historical data: correction / elimination / human intervention processing; S4. Automated data verification. Using a stratified random partitioning method, the data after preprocessing and outlier detection is divided into a training set, a validation set, and a test set. The training set is used to build the training model, the validation set is used to adjust the model, and the test set is used for performance evaluation. A monitoring mechanism is built to analyze the test results in detail when they do not meet the predetermined standards, and to feed the analysis results back to the analysis model or model training model to make targeted adjustments to the model structure. S5. Divide the processed data into three sub-libraries based on quality control knowledge: equipment performance sub-library, equipment problem and treatment sub-library, and quality control standard sub-library. All text data in the three sub-databases are merged to build a vocabulary. All text data are traversed to count the number of co-occurrences of each word in the vocabulary with other words, and a co-occurrence matrix is ​​built. The GloVe algorithm is used for training to obtain the vector representation of each word. A graph database Neo4j was established and three types of nodes were created in it corresponding to the three sub-databases. The relationships were defined and the graph structure was constructed according to the entities in the equipment performance sub-database, equipment problem and treatment sub-database, and quality control standard sub-database. The graph structure was input into the R-GCN model and trained in the model using random negative sampling to obtain the vector representation of each node. By calculating the cosine similarity between two vectors, the relationship weights between each relationship in the graph database are determined. According to nodes, relationships and relationship weights, a knowledge graph is established in the graph database to build a quality control knowledge base; S6. Integrate and optimize the constructed knowledge graph, and semantically annotate the entities and relationships therein. Establish a knowledge base indexing mechanism to classify and index the nodes in the knowledge graph. Develop a knowledge base query interface and establish corresponding query interfaces based on the three sub-bases: equipment performance query, fault diagnosis query, and quality control standard query. When one of the sub-base data is mobilized, the knowledge base will mobilize the other two sub-base data based on the relationship weight and arrange and display them according to the size of the relationship weight. And build a knowledge graph visualization interface to support users to interact with nodes and relationships in the graph.

2. A method for constructing a medical accelerator quality control knowledge base using a large language model as claimed in claim 1, characterized in that: The two-stage pretreatment in S2 comprises the following steps: S21, a processing step, based on a preset set of rules, using machine learning algorithms to learn the normal patterns of data, intelligently filter the data, and divide the filtered data into multiple data sets, S22, second stage processing, setting block standards according to different data sets, dividing multiple data sets into different data blocks by analyzing the data, generating normal data block markers, potential abnormal data block markers and missing data block markers according to the block standards, and marking them with the corresponding data blocks, The preset rule set includes equipment parameter range rules, data format and integrity rules, data frequency rules and historical data statistical regularity rules.

3. The method for constructing a medical accelerator quality control knowledge base using a large language model as claimed in claim 2, characterized in that: The data sets include equipment performance parameter data sets, performance stability data sets, performance and environment relationship data sets, equipment failure data sets, problem warning data sets, problem handling history data sets, quality control indicator data sets, quality control result data sets and quality control deviation analysis data sets.

4. The method for constructing a medical accelerator quality control knowledge base using a large language model according to claim 1, characterized in that: The combination method of the analysis model in S3 includes the following steps: S31, the autoencoder includes an encoder and a decoder. For data in a potential abnormal data block and a data block that needs further inspection, the encoder maps the high-dimensional input data to a low-dimensional feature space, and the decoder reconstructs the original data from the low-dimensional feature space. When the difference between the encoder's reconstructed structure and the original data exceeds a set threshold, the data is determined to be suspected abnormal data; S32, randomly adopting variational autoencoder and / or generative adversarial network to reanalyze the data suspected of abnormality determined by the autoencoder, The variational autoencoder performs probability modeling on potential abnormal data, assuming that the suspected abnormal data obeys the prior distribution, and calculates the probability of the suspected abnormal data under the prior distribution. If the obtained probability structure is lower than the set threshold, the data is judged to be abnormal data. The generative adversarial network includes a generator and a discriminator. The generator generates normal-like data based on the patterns learned from the normal data blocks, and the discriminator determines the abnormal data by judging the difference between the suspected abnormal data and the generated normal data; S33, calculate and analyze abnormal data using Z-score method and IQR method, If the Z-score of a data point exceeds the range of ±3 and also exceeds the upper and lower limits of the IQR method, the data point is judged to be abnormal data and is removed. If only the Z-score of a data point exceeds the range of ±3 / exceeds the upper and lower limits of the IQR method, the data point is analyzed based on its physical meaning and historical data, and attempts are made to reasonably correct, eliminate, or manually intervene in the data point: if it can be corrected, the corrected data is retained; if it cannot be reasonably corrected, the data is eliminated; if it is difficult to determine whether to correct or eliminate based on the physical meaning and historical data, it is marked as a "value to be noted" and pushed to the operator, who then performs manual intervention.

5. The method for constructing a medical accelerator quality control knowledge base using a large language model as claimed in claim 4, characterized in that: The calculation formula of the Z-score method is: , where x is the value of the data point, is the mean of the data block, is the standard deviation of the data block; Sort the data in the data block from small to large and calculate the first quartile and the third quartile , The calculation method of IQR method is , where the lower limit is , the upper limit is .

6. The method for constructing a medical accelerator quality control knowledge base using a large language model according to claim 1, characterized in that: In S5, a distributed computing method is used to calculate the vector representation of each word and each node, specifically: The text data of the three sub-databases are divided into multiple data sets, and the GloVe algorithm is used to construct the co-occurrence matrix and calculate the word vector on multiple data sets at the same time, and finally the results are merged. The graph structure data is divided into multiple subgraphs, each subgraph is trained with the R-GCN model at the same time, and then the results of each subgraph are merged through the All-Reduce algorithm.

7. The method for constructing a medical accelerator quality control knowledge base using a large language model according to claim 1, characterized in that: The S6 also includes building a knowledge base update mechanism: regularly collecting data and knowledge according to a set time, evaluating the importance and necessity of the newly collected data and knowledge, Update the new data and knowledge in a small scale, analyze the impact on the overall structure and relationship weights of the knowledge graph, evaluate the importance of the new data, and update the overall data or abandon the update according to the level of importance. The levels range from small to large: no association, slight impact, and severe impact. If the new data is assessed as not relevant, the update will be abandoned. If the new data is assessed as having a slight impact, a unified update will be performed according to the set time. If the new data is assessed as having a severe impact, the update will be given priority.

8. A system for constructing a medical accelerator quality control knowledge base using a large language model, using a method for constructing a medical accelerator quality control knowledge base using a large language model as claimed in any one of claims 1 to 7, characterized in that: It includes acquisition module, two-stage preprocessing module, outlier detection module, large language training module, knowledge graph module and knowledge base indexing and query module. The acquisition module is used to collect data from different data sources and uniformly format them. The two-stage preprocessing module is used to intelligently screen the data, divide the data into multiple data blocks, generate normal data markers, potential abnormal data markers and data markers that need further inspection, and mark the corresponding data blocks. The outlier detection module is used to build an analysis model and perform outlier detection on data involving equipment performance parameters in potential abnormal data blocks and data blocks that need further inspection. The large language training module is used to build a large language model and use the processed data to perform model training verification. The knowledge graph module is used to construct sub-libraries and graph databases, and combines the trained large language model with the knowledge graph according to nodes, relationships and relationship weights to construct a quality control knowledge base. The knowledge base indexing and query module is used to build a data indexing mechanism, develop a knowledge base query interface, and construct a knowledge graph visualization interface.

9. A system for constructing a medical accelerator quality control knowledge base using a large language model as claimed in claim 8, characterized in that: The two-stage pre-processing module includes a first-stage processing unit and a second-stage processing unit. The first processing unit is used to filter and divide the data into multiple data sets. The second-stage processing unit is used to analyze multiple data sets and divide them into normal data blocks, potential abnormal data blocks and data blocks that need further inspection.

10. A system for constructing a medical accelerator quality control knowledge base using a large language model as claimed in claim 8, characterized in that: The knowledge graph module includes an equipment performance sub-library, an equipment problem and treatment method sub-library, a quality control standard sub-library and a graph database. The equipment performance sub-library is used to store medical accelerator parameter performance data, medical accelerator performance stability data, and medical accelerator performance and environment related data. The equipment problem and treatment method sub-library is used to store medical accelerator fault data, medical accelerator problem warning data and problem treatment history data. The quality control standard sub-library is used to store medical accelerator quality control indicator data, medical accelerator quality control result data and medical accelerator quality control deviation analysis data.

Images