A medical accelerator fault warning system and method based on a large language model

Abstract

A medical accelerator fault warning system based on a large language model, comprising a data acquisition module, a data processing module, a large language model prediction module, and a user interaction module. The large language model prediction module includes an equipment historical maintenance unit, a maintenance processing aggregation unit, a fault prediction unit, and a fault analysis unit. The large language model has powerful data analysis and processing capabilities and can better process complex medical data and equipment parameters. By using the large language model, a large amount of historical data can be learned and analyzed to discover the rules and relationships hidden in the data, thereby improving the accuracy of fault warning. Compared with traditional fault warning systems, the fault warning system of the present invention has higher intelligence and adaptability.

Description

Technical Field

[0001] The present invention relates to the technical field of medical accelerator warning, and particularly to a medical accelerator fault warning system and method based on 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 radiotherapy. It directs these high-energy rays onto tumor tissues to damage the DNA of cancer cells, thereby inhibiting tumor growth or eliminating cancer cells.

[0003] [[ID=eleven]]As a key device for radiotherapy, the operating state of a medical accelerator directly affects the treatment effect and safety of patients. Traditional fault detection methods mainly rely on manual inspections and regular maintenance, which have problems such as lag and subjectivity. With the rapid development of artificial intelligence technology, especially the wide application of large language models in the field of natural language processing, new solutions have been provided for the fault prediction and warning of medical accelerators. However, there is currently no mature large language model system applied to the fault prediction and warning of medical accelerators on the market. Summary of the Invention

[0004] The purpose of the present invention is to provide a medical accelerator fault warning system and method based on a large language model, which solves the problems of lag and subjectivity in the fault detection of medical accelerators mentioned above.

[0005] To achieve the above purpose, the present invention provides a medical accelerator fault warning system based on a large language model, including:

[0006] A data acquisition module for obtaining the medical plan information of patients in the hospital information system, as well as the current device parameters and historical maintenance parameters of the medical accelerator;

[0007] A data processing module for preprocessing the medical plan information of patients and the operating parameters of the medical accelerator and removing abnormal data points to obtain processed patient medical data and medical accelerator parameter data;

[0008] The large language model prediction module includes:

[0009] An equipment historical maintenance unit for analyzing the fault causes of the medical accelerator through the equipment historical maintenance process, marking the medical accelerator as the analysis object, analyzing the frequently faulty parts of the analysis object, dividing the analysis object into K fault areas, classifying the parts of the analysis object into high-frequency fault parts, medium-frequency fault parts, and low-frequency fault parts, generating high-frequency fault marks, medium-frequency fault marks, and low-frequency fault marks, and marking the corresponding fault areas;

[0010] A maintenance processing collection unit is used to regularize the historical maintenance records and equipment parameters of the analysis object, analyze the causes of faults for the maintenance processing solutions in each fault area of the analysis object, classify the maintenance processing solutions into occasional maintenance solutions and specific maintenance solutions through analysis, generate occasional maintenance marks and specific maintenance marks, and collect the corresponding maintenance solutions;

[0011] A fault prediction unit is used to bring the obtained patient medical data and medical accelerator parameter data into the aging rate calculation formula, and predict the equipment aging situation of the medical accelerator after adopting the patient's medical plan according to the calculation result of the aging rate formula, and output the aging rate result.

[0012] Its aging rate (OL rate ) calculation formula is: ; ; ; ; ; Among them, Freq is the usage frequency of the medical accelerator, Time is the total usage time of the medical accelerator, DoseType is the effectiveness factor after classification of the irradiation dose, Cone is the dose size influence factor after standardization, Load is the treatment dose executed within the set time, Repairs is the number of fault repairs of the medical accelerator, FR is the fault incidence rate of the medical accelerator within the set time. 、 、...... are the corresponding weight coefficients respectively, α, β, γ, δ, η are the corresponding non-linear function parameters respectively, Thresh1, Thresh2, Thresh3 are the thresholds of the variables, X>Thresh is a binary function, when X is greater than Thresh, the value is 1, otherwise it is 0; A fault analysis unit is used to predict the fault conditions of each component of the medical accelerator after adopting the patient medical plan and generate a recommended plan for the fault conditions according to the aging rate (OL rate ) of each part of the analysis object, the number of fault repairs (Repairs) and the fault incidence rate (FR); A user interaction module is used to interact with the user through the interface with the fault prediction result and the recommended plan, and provide functions such as prediction report generation, recommended plan query and prediction reminder.

[0013] Preferably, the data processing module constructs a database of the corresponding rules between the irradiation dose and the effectiveness, sets an effectiveness range for each type of dose, and uses algorithms such as decision tree or support vector machine (SVM) or neural network to match the irradiation dose name in the medical plan as the input feature and the known effectiveness factor after classification of the irradiation dose as the output label, and then assigns the corresponding effectiveness factor according to the category.

[0014] The data processing module parses the medical plan text, searches for keywords, accurately extracts the magnitude value of the irradiation dose, and calculates the minimum value C of the irradiation dose in the historical data min and the maximum value C max , and normalizes the new irradiation dose C new to obtain the standardized dose magnitude influence factor. The calculation formula is: .

[0015] Preferably, the data processing module uses the box plot method to identify and remove outliers in the medical accelerator parameters, calculates its quartiles Q1 and Q3, and determines the upper and lower limits of the data points. The calculation formula is as follows: Lower limit = , Upper limit = , where IQR = .

[0016] Preferably, a medical accelerator fault warning method based on a large language model runs on the above-mentioned medical accelerator fault warning system based on a large language model. The method includes:

[0017] S1. Obtain the medical plan information of patients, the current device parameters of the medical accelerator, and the historical maintenance parameters in the hospital information system;

[0018] S2. Preprocess the obtained data, clean and standardize the medical plan information of patients to obtain a unified data format and parameters; for the current device parameters and historical maintenance parameters of the medical accelerator, use the box plot method to identify and remove outliers, calculate the quartiles Q1 and Q3 of the outliers, and determine the upper and lower limits of the data points. The calculation formula is as follows: Lower limit = , Upper limit = , where IQR = ;

[0019] S3. Calculate the fault frequency values of each fault area of the medical accelerator, and classify the fault areas by analyzing the fault conditions: high-frequency fault parts, medium-frequency fault parts, and low-frequency fault parts, and rank the levels of each fault area according to the fault frequency values and generate corresponding marks. The frequency value Fr of the fault part is calculated by the formula: ; where, T0 is the average usage time of each part of the medical accelerator, N is the number of fault repairs of each part of the medical accelerator, I0 is the longest idle time allowed for each part of the medical accelerator under normal maintenance conditions, M0 is the average repair time of each part of the medical accelerator for this type of fault (the average repair time for this type of fault), D0 is the regular detection period of the medical accelerator, and a, b, c, d, e are the corresponding weight coefficients;

[0020] S4. Regularize the historical maintenance records and equipment parameters of the medical accelerator, analyze the causes of faults for the historical maintenance treatment plans in each fault area of the medical accelerator, classify the maintenance treatment plans into occasional maintenance plans and specific maintenance plans through analysis, generate occasional maintenance marks and specific maintenance marks, and collect the corresponding maintenance plans;

[0021] S5. Arrange the high-frequency fault parts, medium-frequency fault parts, and low-frequency fault parts classified in S3 in accordance with the level, and calculate the equipment aging rate (OL rate ). Substitute the data preprocessed in S2 into the aging rate (OL rate ). The calculation formula of the aging rate (OL rate ) is as follows: ; ; ; ;

[0022] ; where Freq is the usage frequency of the medical accelerator, Time is the total usage time of the medical accelerator, DoseType is the effectiveness factor after classification of the irradiation dose, Cone is the dose size influence factor after standardization, Load is the treatment dose executed within the set time, Repairs is the number of fault repairs of the medical accelerator, FR is the fault incidence rate of the medical accelerator within the set time, 、 、...... are the corresponding weight coefficients respectively, α, β, γ, δ, η are the corresponding non-linear function parameters respectively, Thresh1, Thresh2, Thresh3 are the thresholds of the variables, X>Thresh is a binary function, when X is greater than Thresh, the value is 1, otherwise it is 0; Calculate the equipment aging rate of each part. If the aging rate (OL rate ) ≥ the set threshold (threshold), then perform S6; if the aging rate (OL rate ) < the set threshold (threshold), then perform S8, where the set threshold (threshold) is the numerical value of the equipment performance aging warning line of the medical accelerator. When the aging rate is lower than the set threshold, it means that the equipment performance is seriously affected, otherwise it means that the equipment performance is sufficient to meet the usage requirements;

[0023] S6. According to the aging rate (OLrate), the number of fault repairs (Repairs), and the fault incidence rate (FR) of each part of the analysis object, predict the fault conditions of each component after the medical accelerator adopts the patient medical plan and generate a recommended plan for this fault condition;

[0024] S7. Interact with users through the interface on fault prediction results and recommended solutions, providing functions such as prediction report generation, recommended solution query, and prediction reminder;

[0025] S8. Save the fault prediction result and generate a historical record.

[0026] Preferably, after the fault frequency value Fr of the part is calculated in S3, the fuzzy membership function is calculated for the part. The fuzzy membership function calculation formula is: high frequency fault membership μHF(f) ;Medium frequency fault membership μMF(f) ; Low-frequency fault membership μLF(f)=1-μHF(f)-μMF(f); where j and k are preset parameters. j indicates that when the number of faults at a part exceeds the threshold within a set period, the part enters a high-risk state. k indicates how fast the part fault transitions from normal to abnormal. When μHF(f)>μMF(f) and μHF(f)>μLF(f), the part will be classified as a high-frequency fault part and a label will be generated: high-frequency fault part [membership μHF(f)]. When μMF(f)>μHF(f) and μMF(f)>μLF(f), the part will be classified as a medium-frequency fault part and a label will be generated: medium-frequency fault part [membership μMF(f)]. When μLF(f)>μHF(f) and μLF(f)>μMF(f), the part will be classified as a low-frequency fault part and a label will be generated: low-frequency fault part [membership μLF(f)]. The parts are ranked according to the size of the membership value.

[0027] Preferably, the S6 further analyzes the historical maintenance records and equipment parameters of the medical accelerator, and the steps are as follows:

[0028] S61. Calculate the number of fault repairs per unit frequency of use r1 = Repairs / Freq, and the number of fault repairs per unit treatment dose r2 = Repairs / Load. If r1 > z1, and r2 <z2,则判断此维修方案为偶发维修方案,其中z1、z2为预设值,若条件不满足则进入下一步判断;

[0029] S62. Plot the collected data on the failure rate (FR) and the treatment dose (Load) into a scatter plot, with the treatment dose (Load) and the failure rate (FR) as the vertical axis. Mark the points corresponding to each set of data on the plot. Analyze the distribution of the points using the scatter plot. If the distribution of the points follows a straight line with a clear linear relationship, determine that a specific maintenance solution is being used. If the distribution of the points is scattered with no clear linear relationship, proceed to the next step of determination.

[0030] S63. Plot the collected data on failure rate (FR) and frequency of use (Freq) into a scatter plot, with frequency of use (Freq) and failure rate (FR) as the vertical axis. Mark the points corresponding to each set of data on the plot, and analyze the distribution of the points through the scatter plot. If the distribution of the points is in the shape of a straight line with an obvious linear relationship, it is judged to be a specific maintenance plan. If the distribution of the points is scattered and does not have an obvious linear relationship, it is judged to be an occasional maintenance plan.

[0031] Preferably, the calculation formula of the set threshold is: set threshold (threshold) ,in, Represents the average value of historical aging rate data, represents the standard deviation of historical aging rate data, is the coefficient, The value range is [1,2].

[0032] Beneficial effects of the present invention:

[0033] 1. The present invention utilizes a large language model to learn and analyze large amounts of historical data, uncovering patterns and relationships hidden within the data, thereby improving the accuracy of fault warnings. Compared to traditional fault warning systems, the present invention's fault warning system is more intelligent and adaptable.

[0034] 2. The aging rate formula of the present invention integrates multiple factors that affect equipment aging, such as frequency of use, duration of use, and type of irradiation dose, making the evaluation more comprehensive. Through weight coefficients (w1 to w7) and nonlinear function parameters (α, β, γ, δ, η), the influence of each factor can be adjusted according to the characteristics of different equipment or environments, making the formula more targeted. By setting thresholds (Thresh1, Thresh2, Thresh3), the significant impact of certain factors on the aging rate after exceeding specific values can be highlighted, enabling the model to reflect the accelerated aging phenomenon under critical conditions.

[0035] 3. The early warning method of the present invention utilizes multi-factor fusion analysis methods in multiple stages, such as fault frequency calculation, equipment aging rate calculation, and fault prediction. Compared with traditional single-factor analysis methods, the present invention can more comprehensively and accurately reflect the operating status, fault risk, and aging of medical accelerators. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art.

[0037] Figure 1It is the structural block diagram of the medical accelerator fault warning system of the present invention.

[0038] Figure 2 It is the structural block diagram of the large language model prediction module of the present invention.

[0039] Figure 3 It is the flowchart of the medical accelerator fault warning method of the present invention. Detailed implementation manners

[0040] The embodiments of the present invention will be described in detail below. The examples of the embodiments are shown in the accompanying drawings. The embodiments described below by referring to the accompanying drawings are exemplary and are intended to explain the present invention, and should not be construed as a limitation to the present invention.

[0041] As Figure 1 shown, a medical accelerator fault warning system based on a large language model includes a data acquisition module, a data processing module, a large language model prediction module, and a user interaction module. The large language model prediction module includes an equipment historical maintenance unit, a maintenance processing aggregation unit, a fault prediction unit, and a fault analysis unit. The large language model has powerful data analysis and processing capabilities and can better process complex medical data and equipment parameters. By using the large language model, a large amount of historical data can be learned and analyzed to discover the rules and relationships hidden in the data, thereby improving the accuracy of fault warning. Compared with traditional fault warning systems, the fault warning system of the present invention has higher intelligence and adaptability.

[0042] The data acquisition module is used to obtain the medical plan information of patients in the hospital information system, as well as the current equipment parameters and historical maintenance parameters of the medical accelerator. Through the data acquisition module, the medical plan information of patients in the hospital information system, the current equipment parameters of the medical accelerator, and the historical maintenance parameters can be obtained. It ensures that the system has a comprehensive data source and provides rich basic data for subsequent fault warning. By integrating various types of data, the system can more comprehensively consider various factors affecting the faults of the medical accelerator and improve the accuracy of fault prediction.

[0043] The data processing module is used to preprocess the medical plan information of patients and the working parameters of the medical accelerator and eliminate abnormal data points to obtain the processed medical data related to patients and the parameter data of the medical accelerator. Through the data processing module, the medical plan information of patients and the working parameters of the medical accelerator can be preprocessed and abnormal data points can be eliminated. The preprocessing operation can remove the data noise that may interfere with the analysis, making the subsequent analysis more accurate. The processed data is more reliable and can more truly reflect the operating state of the medical accelerator and the factors related to patient treatment, which helps to improve the prediction reliability of the entire system.

[0044] The device maintenance history unit analyzes the causes of medical accelerator failures based on the device's maintenance history. The unit identifies the medical accelerator as an analysis target, analyzes the frequently failing locations within the target, and divides the target into K fault zones. The analysis then categorizes the locations into high-frequency, medium-frequency, and low-frequency fault zones. High-frequency, medium-frequency, and low-frequency fault markers are generated and labeled with the corresponding fault zones. The device maintenance history unit of this early warning system analyzes the medical accelerator's maintenance history to divide the device into K fault zones, further categorizing the locations into high-frequency, medium-frequency, and low-frequency fault zones. This classification approach provides a clearer understanding of the fault distribution of the medical accelerator. Different maintenance and prevention strategies can be implemented for fault locations with different frequencies. For example, high-frequency fault locations can be subject to enhanced monitoring and preventive maintenance, thereby improving the overall reliability of the medical accelerator, reducing the likelihood of failure, lowering maintenance costs, and extending the device's service life.

[0045] The maintenance processing collection unit is used to organize the historical maintenance records and equipment parameters of the analysis object, analyze the causes of the maintenance processing plans for each fault area of the analysis object, divide the maintenance processing plans into occasional maintenance plans and specific maintenance plans through analysis, generate occasional maintenance tags and specific maintenance tags, and collect the corresponding maintenance plans. This system organizes the historical maintenance records and equipment parameters of the analysis object, analyzes the causes of the maintenance processing plans for each fault area, and divides the maintenance processing plans into occasional maintenance plans and specific maintenance plans. It can help maintenance personnel quickly identify different types of maintenance situations and improve maintenance efficiency. For occasional maintenance plans, their accidental causes can be further analyzed, and for specific maintenance plans, more in-depth research and improvements can be conducted on their specific failure causes, which helps to optimize the maintenance process and improve the overall performance of the equipment.

[0046] The fault prediction unit is used to input the obtained patient medical data and medical accelerator parameter data into the aging rate calculation formula, calculate the equipment aging situation after the medical accelerator adopts the patient's medical plan according to the calculation result of the aging rate formula, and output the aging rate result. The fault prediction unit of the present invention predicts the equipment aging situation by inputting the patient medical data and medical accelerator parameter data into the aging rate calculation formula. This aging rate calculation formula takes into account multiple factors, such as the usage frequency (Freq) of the medical accelerator, the total usage time (Time), the effectiveness factor (DoseType) after irradiation dose classification, the dose size impact factor (Cone) after standardization, the treatment dose (Load) executed within the set time, the number of fault repairs (Repairs), and the fault incidence rate (FR), etc. These factors cover all aspects of the operation of the medical accelerator, including the usage situation of the equipment itself, the irradiation dose factors related to treatment, and the maintenance history, etc. By comprehensively considering these factors, the system can more accurately predict the equipment aging situation, discover potential fault risks in advance, and provide a strong basis for preventive maintenance.

[0047] Its aging rate (OL rate ) calculation formula is:

[0048] ;

[0049] ;

[0050] ;

[0051] ;

[0052] ;

[0053] Among them, Freq is the usage frequency of the medical accelerator, Time is the total usage time of the medical accelerator, DoseType is the effectiveness factor after irradiation dose classification, Cone is the dose size impact factor after standardization, Load is the treatment dose executed within the set time, Repairs is the number of fault repairs of the medical accelerator, FR is the fault incidence rate of the medical accelerator within the set time, , ,...... are the corresponding weight coefficients respectively, α, β, γ, δ, η are the corresponding non-linear function parameters respectively, Thresh1, Thresh2, Thresh3 are the thresholds of the variables, X>Thresh is a binary function, when X is greater than Thresh, the value is 1, otherwise it is 0.

[0054] This aging rate calculation formula is in the form of weighted summation, and each factor is multiplied by the corresponding weight coefficient ( , , etc.), and adjusted according to different non - linear functions (such as , , etc.). For example, for the usage frequency (Freq), the higher the usage frequency, the faster the device ages, and it affects the aging rate through this term. When the usage frequency Freq is greater than the threshold Thresh1, this term will affect the aging rate in the way of . This function increases with the increase of Freq, and the growth rate is determined by α, which means that the higher the usage frequency, the greater the impact on the aging rate, but the growth mode is non - linear. Therefore, adding non - linear functions to the formula can better describe these relationships and be more in line with the actual situation. Similarly, other factors such as the total usage time (Time) affect the aging rate through , reflecting the quadratic growth relationship of the total usage time on the aging rate (when Time is greater than Thresh2), indicating that as time goes by, the aging speed of the device will accelerate. This formula combines multiple factors affecting device aging, such as usage frequency, usage time, irradiation dose type, etc., which makes the evaluation more comprehensive; through the weight coefficients (w1 to w7) and non - linear function parameters (α, β, γ, δ, η), the influence intensity of each factor can be adjusted according to the characteristics of different devices or environments, making the formula more targeted; by setting thresholds (Thresh1, Thresh2, Thresh3), the significant impact of some factors on the aging rate after exceeding specific values can be highlighted, enabling the model to reflect the aging acceleration phenomenon under critical conditions.

[0055] For example, in an example, the parameters of the medical accelerator part are as follows:

[0056] Freq = 300 times / week, Thresh1 = 200, Time = 3 years, Thresh2 = 2 years, DoseType = 0.8 (the impact factor of the standardized irradiation dose type), Cone = 0.7 (the standardized dose size factor), Load = 200 times / week, Repairs = 5 times / year, Thresh3 = 3, FR = 0.05 times / month.

[0057] Let = 0.2, = 0.15, = 0.1, = 0.15, = 0.1, = 0.15, = 0.15; α = 0.005, β = 0.05, γ = 0.2, δ = 0.1, η = 0.002. Then, substituting into the aging rate formula, we get: OLrate = 0.2×(1−e −0.005×300 )×[300 > 200] + 0.15×(0.05×3 2 )×[3 > 2] + 0.1×(1−0.2×0.8) + 0.15×(0.1×0.7 1.5 ) + 0.1×(0.002×200 0.5 ) + 0.15×5×[5 > 3] + 0.15×0.05. OLrate ≈ 1.241.

[0058] In this example, the final calculated result of OLrate indicates that the aging rate of this part is quite high and requires key attention and possible maintenance upgrades.

[0059] In addition, for the method of obtaining the non - linear function parameters in the present invention, first, historical data related to the performance of medical accelerators is collected, and these data should include variables such as usage frequency, usage time, irradiation dose type, irradiation dose, treatment load, number of fault repairs, and failure rate. Secondly, by plotting scatter plots and calculating correlation coefficients, the non - linear trend of the relationship between variables can be preliminarily understood. According to the results of data analysis, non - linear function forms such as exponential functions, power functions, logarithmic functions, or Sigmoid functions are selected. Then, a statistical software package (such as R, scikit - learn library of Python, or SPSS) is used for non - linear regression analysis to estimate the parameters of the non - linear function for finding the best - fitting parameters. Next, cross - validation, residual analysis, and coefficient of determination (R²) are used to evaluate the goodness of fit and prediction ability of the model to confirm the reliability of the model and the accuracy of parameter estimation. Finally, sensitivity analysis is performed to evaluate the impact of parameter changes on the calculation result of the aging rate, to identify which parameters are most sensitive to the model output, and to ensure the robustness of the model.

[0060] Regarding the setting of the threshold, the present invention first collects data related to equipment aging, including usage frequency, time, irradiation dose type, size, treatment load, number of repairs, failure rate, etc. Secondly, statistical analysis is performed on the collected data, including calculating the mean, median, standard deviation, maximum value, minimum value, etc., to understand the data distribution, and methods such as correlation coefficients are used to analyze the correlation between different variables. Then, according to the data distribution, a certain quantile (such as 25th, 50th, 75th, etc.) is selected as the threshold. Next, a part of the data is used to establish an aging model, and then another part of the data is used to verify the accuracy of the model, and the threshold is adjusted according to the performance of the model. Finally, through continuous iteration, the model performances under different thresholds are compared to select the best threshold.

[0061] A fault analysis unit is used to predict the fault conditions of each component of a medical accelerator after adopting a patient medical plan and generate a recommended plan for such fault conditions based on the aging rate (OL rate ) of each part of the analysis object, the number of fault repairs (Repairs), and the fault incidence rate (FR). By incorporating key factors such as the aging rate (OLrate), the number of fault repairs (Repairs), and the fault incidence rate (FR) of each part of the analysis object, the fault analysis unit of the present invention can more comprehensively predict the fault conditions of each component of the medical accelerator after adopting the patient medical plan. This comprehensive consideration of multiple factors avoids the limitations of single-factor prediction, making the prediction results more accurate and reliable. For example, only considering the fault incidence rate may not accurately reflect the actual fault risk of a component. By adding the aging rate and the number of repairs, it is possible to evaluate from multiple perspectives such as the usage wear and tear of the component and its historical repair situation, thus more accurately predicting the fault conditions. Based on the prediction of component fault conditions, corresponding recommended plans can be generated, providing direct guidance for the maintenance and management of the medical accelerator. For medical institutions, they can take timely measures according to these recommended plans, such as preventive maintenance, component replacement, etc., thereby extending the service life of the medical accelerator, reducing maintenance costs, and at the same time ensuring the safety and reliability of medical equipment, which is conducive to improving the efficiency of medical services.

[0062] A user interaction module is used to interact with the user through an interface about the fault prediction results and recommended plans, providing functions such as prediction report generation, recommended plan query, and prediction reminder. The user interaction module of the present invention displays information such as possible fault types and occurrence probabilities in a clear table or chart form. For example, the prediction report may show that the probability of an accelerating tube fault is 30%. The recommended plans corresponding to different fault types are listed separately to facilitate users to quickly locate. At the same time, the module can list the recommended plans corresponding to different fault types separately to facilitate users to quickly locate; for urgent faults (such as an impending accelerating tube fault), the most urgent solution is ranked first, enabling users to take measures preferentially. When it is predicted that a fault is about to occur, the module will pop up a reminder window on the user interface, briefly showing the fault type and the degree of urgency. And users can set the reminder frequency according to their own needs, such as reminding once an hour, or only reminding when the fault occurrence probability exceeds a certain threshold (such as 50%).

[0063] In one embodiment, the data processing module constructs a database of the correspondence rules between irradiation dose and efficacy, sets an efficacy range for each type of dose, and uses algorithms such as decision trees, support vector machines (SVMs), or neural networks to match the irradiation dose names in the medical plan as input features with the efficacy factors after classification of known irradiation doses as output labels, and then assigns corresponding efficacy factors according to the categories;

[0064] The data processing module parses the medical plan text, searches for keywords, accurately extracts the magnitude value of the irradiation dose, and obtains the minimum value C of the irradiation dose in the historical data through statistics min and the maximum value C max , and performs normalization processing on the new irradiation dose C new to obtain the standardized dose magnitude impact factor.

[0065] Calculation formula: .

[0066] The data processing module of the present invention constructs a database of the corresponding rules between the irradiation dose and the efficacy, and sets an efficacy range for each type of dose. Using advanced algorithms such as decision trees, support vector machines (SVMs), or neural networks, with the irradiation dose name in the medical plan as the input feature, the efficacy factor after classifying the irradiation dose is used as the output label for matching and the corresponding efficacy factor is assigned. In this way, the efficacy factor can be accurately determined according to the irradiation dose name, improving the accuracy of the irradiation dose efficacy evaluation. For example, in a complex medical plan, the efficacy of different irradiation doses may be affected by multiple factors, and through this module, its efficacy range can be accurately determined, helping medical staff better understand the effect of the irradiation dose.

[0067] Furthermore, the data processing module parses the medical plan text to search for keywords and accurately extracts the magnitude value of the irradiation dose. Then, by statistically calculating the minimum value Cmin and the maximum value Cmax of the irradiation dose in the historical data, normalization processing is performed on the new irradiation dose Cnew to obtain the standardized dose magnitude impact factor. This processing method can convert irradiation doses with different value ranges into a unified standard scale, facilitating comparison and analysis between different medical plans or irradiation dose data of different batches. For example, when studying the impact of the irradiation dose on the treatment effect, the standardized dose magnitude impact factor can more intuitively reflect the relative impact degree of the irradiation dose, helping medical researchers better analyze and optimize medical plans.

[0068] In one embodiment, the data processing module uses the box plot method to identify and eliminate outliers of the medical accelerator parameters, calculates its quartiles Q1 and Q3, and determines the upper and lower limit ranges of the data points.

[0069] The calculation formula is as follows: Lower limit = , Upper limit = , where IQR = .

[0070] The calculation method of quartiles in the present invention does not depend on the distribution pattern of data (such as normal distribution, skewed distribution, etc.), and is applicable to the working parameter data of medical accelerators of various distribution types. For example, the energy output data of medical accelerators may not have an ideal normal distribution due to reasons such as equipment aging and sample differences between different batches, but the box plot method can still well identify outliers. 1.5 times the IQR has a certain robustness while ensuring that true outliers can be captured. If the multiple is set too small, some normal data that is slightly deviated from the center may be misjudged as outliers. For example, in the working parameters of medical accelerators, small fluctuations in parameters caused by occasional environmental perturbations (such as minor voltage fluctuations) are normal. If the multiple is too small, these normally fluctuating data may be wrongly excluded. If the multiple is set too large, some true outliers may be retained, thus affecting the subsequent judgment of the normal working range of medical accelerators and the formulation of medical plans based on these data. For example, if a key parameter of a medical accelerator suddenly shows a large deviation beyond the normal range (possibly due to equipment failure), if the multiple is too large, this outlier may not be recognized as an anomaly and be used for analysis, leading to wrong conclusions.

[0071] The upper and lower limits determined in this way can reflect the overall structure of the data. In the dataset of the working parameters of medical accelerators, data points outside the upper and lower limits are regarded as outliers, which helps to maintain the "purity" of the data and makes subsequent analysis based on these data (such as determining the normal working range of the equipment, analyzing the matching with medical plans, etc.) more accurate. For example, when analyzing the parameter stability of medical accelerators in different working states, accurately removing outliers can more clearly show the distribution range and change trend of parameters in the normal working state, helping to timely detect potential problems that the equipment may have or optimize the part related to equipment parameters in the medical plan.

[0072] A medical accelerator fault warning method based on a large language model, comprising:

[0073] S1. Obtain the medical plan information of patients, the current equipment parameters of the medical accelerator, and the historical maintenance parameters from the hospital information system;

[0074] S2. Preprocess the obtained data. Clean and standardize the medical plan information of patients to obtain a unified data format and parameters. For the current equipment parameters and historical maintenance parameters of the medical accelerator, use the box plot method to identify and remove outliers, calculate its quartiles Q1 and Q3, and determine the upper and lower limit ranges of the data points.

[0075] The calculation formula is as follows: Lower limit = Upper limit = , where IQR= ;

[0076] S3. Calculate the fault frequency value of each fault area of the medical accelerator, and classify the fault areas into high-frequency fault areas, medium-frequency fault areas, and low-frequency fault areas by analyzing the fault conditions. Then, rank the fault areas according to the fault frequency values and generate corresponding labels. The frequency value Fr of the fault area is calculated as follows: Wherein, T0 is the average usage time of each part of the medical accelerator, N is the number of times each part of the medical accelerator is repaired, I0 is the maximum idle time allowed under normal maintenance conditions for each part of the medical accelerator, M0 is the average repair time for each part of the medical accelerator (the average repair time for this type of fault), D0 is the regular inspection cycle of the medical accelerator, and a, b, c, d, and e are the corresponding weight coefficients respectively;

[0077] S4. Collating historical maintenance records and equipment parameters of the medical accelerator, analyzing the causes of the failures in the historical maintenance solutions for each fault area of the medical accelerator, classifying the maintenance solutions into occasional maintenance solutions and specific maintenance solutions through analysis, generating occasional maintenance tags and specific maintenance tags, and grouping the corresponding maintenance solutions;

[0078] S5. Arrange the high-frequency fault parts, medium-frequency fault parts and low-frequency fault parts classified in S3 according to their levels and calculate the equipment aging rate (OL rate ), the data pre-processed by S2 is brought into the aging rate (OL rate ) calculation formula,

[0079] Its aging rate (OL rate ) is calculated as: ; ; ; ; ; Among them, Freq is the frequency of use of the medical accelerator, Time is the total use time of the medical accelerator, DoseType is the efficiency factor after irradiation dose classification, Cone is the standardized dose size influencing factor, Load is the treatment dose performed within the set time, Repairs is the number of fault repairs of the medical accelerator, and FR is the failure rate of the medical accelerator within the set time. 、 、...... are the corresponding weight coefficients, α, β, γ, δ, η are the corresponding non-linear function parameters, Thresh1, Thresh2, Thresh3 are the thresholds of the variables respectively, X>Thresh is a binary function, when X is greater than Thresh, the value is 1, otherwise it is 0; the aging rate of each part of the equipment is calculated. If the aging rate (OL rate )≥ the set threshold (threshold), then S6 is performed; if the aging rate (OL rate )< the set threshold (threshold), then S8 is performed, where the set threshold (threshold) is the numerical value of the aging warning line of the performance of the medical accelerator equipment. When the aging rate is lower than the set threshold, it means that the equipment performance is seriously affected, otherwise it means that the equipment performance is sufficient to meet the use requirements;

[0080] S6. According to the aging rate (OLrate), the number of fault repairs (Repairs) and the failure rate (FR) of each part of the analysis object, predict the fault conditions of each component after the medical accelerator adopts the patient medical plan and generate a recommended plan for this fault condition;

[0081] S7. Interact with the user through the interface with the fault prediction results and the recommended plan, and provide functions such as prediction report generation, recommended plan query, and prediction reminder;

[0082] S8. Save the fault prediction results and generate a historical record.

[0083] The rationality of the formula for calculating the frequency value Fr of the fault location in this early warning method lies in that it comprehensively considers multiple key parameters of each part of the medical accelerator. Among them, the average usage time T0, the longest idle time I0, the average repair time M0, and the regular detection period D0 are all important factors affecting the fault frequency. By multiplying different weight coefficients a, b, c, d, e, the influence degree of each factor on the fault frequency can be adjusted according to the actual situation. For example, if in the actual use of a certain medical accelerator, the average repair time M0 has a greater impact on the fault frequency, then the value of c can be appropriately increased so that the formula can more accurately reflect the fault frequency situation of the device. This method adopts a multi-factor fusion analysis method in multiple links, such as fault frequency calculation, equipment aging rate calculation, fault prediction, etc. Compared with the traditional single-factor analysis method, the early warning method of the present invention can more comprehensively and accurately reflect the operating state, fault risk, and aging situation of the medical accelerator. The large language model has powerful semantic understanding and data processing capabilities, and this method adopts a large language model-based approach, which can better process and analyze complex medical plan information, equipment parameters and other data. Compared with the traditional fault early warning method, the present invention can dig deeper information, improve the accuracy and reliability of fault early warning; and historical fault data is set, and the recent prediction can be compared with historical similar situations to show trend changes, which is convenient for medical staff or maintenance personnel to analyze the performance of the medical accelerator and control the overall health.

[0084] In one embodiment, after calculating the part fault frequency value Fr in S3 above, a fuzzy membership function calculation is performed on the part again.

[0085] The formula for its fuzzy membership function is: high-frequency fault membership degree μHF(f) ; medium-frequency fault membership degree μMF(f) ; low-frequency fault membership degree μLF(f) = 1 - μHF(f) - μMF(f); where j and k are preset parameters, j means that after the number of faults of the part exceeds the threshold within the set period, the part enters the high-risk state; k means the speed at which the part fault changes from normal to abnormal.

[0086] When μHF(f) > μMF(f) and μHF(f) > μLF(f), then this part will be classified as a high-frequency fault part and a mark will be generated: high-frequency fault part [membership degree μHF(f)].

[0087] When μMF(f) > μHF(f) and μMF(f) > μLF(f), then this part will be classified as a medium-frequency fault part and a mark will be generated: medium-frequency fault part [membership degree μMF(f)].

[0088] When μLF(f)> μHF(f) and μLF(f)> μMF(f), the part will be classified as a low-frequency fault part and a label will be generated: low-frequency fault part [membership μLF(f)];

[0089] Arrange the parts according to their membership values.

[0090] This method calculates the fault frequency Fr of a part and then applies a fuzzy membership function. This method classifies faulty parts by comprehensively considering factors such as pre-set parameters j (a part enters a high-risk state when the number of faults exceeds a threshold within a set period) and k (the speed at which a part's fault transitions from normal to abnormal). This method is more comprehensive and accurate than classification based solely on fault frequency. For example, even if a part's fault frequency is not the highest, it may transition quickly from normal to abnormal, or if the number of faults within a set period approaches a threshold, indicating a high-risk state. Fuzzy membership function calculations can more accurately classify it as a high-frequency fault part. Ranking the faulty parts based on the membership values calculated by the fuzzy membership function provides a visual representation of the relative risk of each fault part within different fault frequency ranges (high, medium, and low). This helps prioritize measures for faulty parts of different levels during maintenance and management of medical accelerators, improving the rationality and effectiveness of resource allocation.

[0091] In one embodiment, the above S6 further analyzes the historical maintenance records and device parameters of the medical accelerator, and the steps are as follows:

[0092] S61. Calculate the number of fault repairs per unit frequency of use r1 = Repairs / Freq, and the number of fault repairs per unit treatment dose r2 = Repairs / Load. If r1 > z1, and r2 <z2,则判断此维修方案为偶发维修方案,其中z1、z2为预设值,若条件不满足则进入下一步判断;

[0093] S62. Plot the collected data on the failure rate (FR) and the treatment dose (Load) into a scatter plot, with the treatment dose (Load) and the failure rate (FR) as the vertical axis. Mark the points corresponding to each set of data on the plot. Analyze the distribution of the points using the scatter plot. If the distribution of the points follows a straight line with a clear linear relationship, determine that a specific maintenance solution is being used. If the distribution of the points is scattered with no clear linear relationship, proceed to the next step of determination.

[0094] S63. Plot the collected data on failure rate (FR) and frequency of use (Freq) into a scatter plot, with frequency of use (Freq) and failure rate (FR) as the vertical axes. Mark the points corresponding to each data set on the plot. Analyze the distribution of the points using the scatter plot. If the distribution of the points follows a straight line with a clear linear relationship, it is determined to be a specific maintenance plan. If the distribution of the points is scattered and does not show a clear linear relationship, it is determined to be an occasional maintenance plan. Calculate the number of failure repairs per unit of frequency of use, i.e., Repairs / Freq. If this ratio is very small, for example, Repairs / Freq < 0.001, it means that failure repairs are rare during extensive device use. Calculate the number of failure repairs per unit of treatment load, i.e., Repairs / Load. If Repairs / Load < 0.01, it means that failure repairs rarely occur despite performing many treatments. This indicates that failures may be random, unpredictable, and low-probability events, and the corresponding maintenance plan is therefore an occasional maintenance plan. For devices like medical accelerators, usage frequency and treatment dose are important indicators of their operational status. The advantages of these two formulas are their simplicity, intuitiveness, and ease of calculation. Medical personnel or maintenance staff can easily access data on the number of repairs, usage frequency, and treatment dose. Using simple division, they can then determine the values of r1 and r2, and use these values to make a preliminary assessment of the appropriate repair plan. This simplicity enables rapid analysis of repair plans in practice, improving the efficiency of equipment maintenance management.

[0095] In one embodiment, the threshold calculation formula is: ,in, Represents the average value of historical aging rate data, represents the standard deviation of historical aging rate data, is the coefficient, The value range is [1,2].

[0096] The calculation formula for setting the threshold in the present invention comprehensively considers the average value of historical aging rate data. , standard deviation and coefficients ( The value range is [1,2]. Average Reflects the central trend of historical aging rate data, standard deviation It reflects the degree of dispersion of the data. The threshold value can be set flexibly according to the characteristics of historical data. For example, if the historical aging rate data fluctuates greatly (the standard deviation is large), the coefficient can be adjusted appropriately. , a more reasonable threshold can be obtained to adapt to the changes in this kind of data. The advantage of this formula lies in its simplicity and interpretability. The data required for calculation (average value and standard deviation) are relatively easy to obtain, and the formula structure is simple and easy to understand. Medical staff can reasonably select the coefficient value according to their understanding of the historical aging rate data of the equipment, so as to obtain the set threshold that meets the actual requirements. This simple and effective formula can quickly and accurately determine the threshold, provide a basis for the equipment performance evaluation of medical accelerators, make full use of the historical operation data of the equipment, make the threshold setting more scientific and reasonable, and help improve the accuracy of the equipment performance evaluation of medical accelerators.

[0097] The above-disclosed are only one or more preferred embodiments of the present application, and the scope of rights of the present application cannot be limited thereby. Those of ordinary skill in the art can understand all or part of the processes of implementing the above embodiments, and the equivalent changes made according to the claims of the present application still fall within the scope covered by the present application.

Claims

1. A medical accelerator fault warning system based on a large language model, comprising: A data acquisition module is used to obtain the patient's medical plan information, the current equipment parameters, historical operating parameters and historical maintenance parameters of the medical accelerator from the hospital information system; A data processing module is used to pre-process the patient's medical plan information and the working parameters of the medical accelerator and eliminate abnormal data points to obtain processed patient medical data and medical accelerator parameter data; Large language model prediction module, including: An equipment history maintenance unit is used to analyze the cause of a medical accelerator's failure based on the equipment's historical maintenance process, mark the medical accelerator as an analysis object, analyze the frequently faulty parts of the analysis object, divide the analysis object into K fault zones, and divide the parts of the analysis object into high-frequency fault parts, medium-frequency fault parts, and low-frequency fault parts. High-frequency fault markers, medium-frequency fault markers, and low-frequency fault markers are generated, and the corresponding fault zones are marked. The maintenance processing collection unit is used to organize the historical maintenance records and equipment parameters of the analysis object, analyze the fault causes of the maintenance processing plans for each fault area of the analysis object, classify the maintenance processing plans into occasional maintenance plans and specific maintenance plans through analysis, generate occasional maintenance tags and specific maintenance tags, and collect the corresponding maintenance plans; The fault prediction unit is used to input the acquired patient medical data and medical accelerator parameter data into the aging rate calculation formula, and predict the aging of the medical accelerator after the patient's medical plan is adopted according to the calculation result of the aging rate formula, and output the aging rate result. Its aging rate OL rate The calculation formula is as follows: ； ； ； ； ； Among them, Freq is the usage frequency of the medical accelerator, Time is the total usage time of the medical accelerator, DoseType is the effectiveness factor after classification of the irradiation dose, Cone is the dose size influence factor after standardization, Load is the treatment dose executed within the set time, Repairs is the number of fault repairs of the medical accelerator, and FR is the fault incidence rate of the medical accelerator within the set time. and ...... are the corresponding weight coefficients respectively, α, β, γ, δ, η are the corresponding non-linear function parameters respectively, and Thresh1, Thresh2, Thresh3 are the thresholds of the variables. A fault analysis unit, which is used to predict the fault conditions of each component after the medical accelerator adopts the patient medical plan and generate a recommended plan for the fault conditions according to the aging rate OL of each part of the analysis object rate , the number of fault repairs Repairs, and the fault incidence rate FR The user interaction module is used to interact with users through the interface of fault prediction results and recommended solutions, and provides functions such as prediction report generation, recommended solution query and prediction reminder; The data processing module constructs a database of radiation dose and efficacy correspondence rules, sets efficacy ranges for each dose category, and uses a decision tree or support vector machine (SVM) algorithm to take the radiation dose in the medical plan as input features and the efficacy factors after known dose classification as output labels for matching, and then assigns corresponding efficacy factors according to categories; The data processing module parses the medical plan text, searches for keywords, accurately extracts the irradiation dose, and obtains the minimum value C and the maximum value C of the irradiation dose in the historical data, and normalizes the new irradiation dose C to obtain the standardized dose size impact factor. The calculation formula is: min and the maximum value C max , for the new irradiation dose C new to perform normalization processing to obtain the standardized dose size impact factor. The calculation formula is: .

2. The medical accelerator fault warning system based on the large language model according to claim 1, characterized in that, The data processing module uses the box plot method to identify and eliminate abnormal values of medical accelerator parameters, calculate their quartiles Q1 and Q3, and determine the upper and lower limits of the data points. The calculation formula is as follows: Lower limit = , upper limit = , where IQR = .

3. A medical accelerator fault warning method based on a large language model, which runs on a medical accelerator fault warning system based on a large language model as described in any one of claims 1 to 2, characterized in that, The method comprises: S1. Obtaining the patient's medical plan information, and the current device parameters and historical maintenance parameters of the medical accelerator from the hospital information system; S2. Preprocess the acquired data, clean and standardize the patient's medical plan information, and obtain a unified data format and parameters. Use the box plot method to identify and eliminate outliers for the current equipment parameters and historical maintenance parameters of the medical accelerator, calculate their quartiles Q1 and Q3, and determine the upper and lower limits of the data points. The calculation formula is as follows: Lower limit = , upper limit = , where IQR= ; S3. Calculate the fault frequency value of each fault area of the medical accelerator, and classify the fault areas into high-frequency fault areas, medium-frequency fault areas, and low-frequency fault areas by analyzing the fault conditions. Then, rank the fault areas according to the fault frequency value and generate corresponding labels. The calculation formula for the frequency value Fr of its fault location is as follows: ; Wherein, T0 is the average usage time of each part of the medical accelerator, N is the number of times each part of the medical accelerator is repaired, I0 is the maximum idle time allowed under normal maintenance of each part of the medical accelerator, M0 is the average repair time of each part of the medical accelerator, D0 is the regular inspection cycle of the medical accelerator, and a, b, c, d, and e are the corresponding weight coefficients respectively; After calculating the part failure frequency value Fr, perform fuzzy membership function calculation on the part again. The formula for its fuzzy membership function is: high-frequency failure membership degree μHF(f) ; medium-frequency failure membership degree μMF(f) ; low-frequency failure membership degree μLF(f) = 1 - μHF(f) - μMF(f); where j and k are preset parameters. j indicates that when the number of failures of the part exceeds the threshold within the set period, the part enters a high-risk state; k indicates the speed at which the part failure changes from normal to abnormal When μHF(f)>μMF(f) and μHF(f)>μLF(f), the part will be classified as a high-frequency fault part and a label will be generated: high-frequency fault part [membership μHF(f)], When μMF(f)>μHF(f) and μMF(f)>μLF(f), the part will be classified as a medium frequency fault part and a label will be generated: medium frequency fault part [membership μMF(f)], When μLF(f)> μHF(f) and μLF(f)> μMF(f), the part will be classified as a low-frequency fault part and a label will be generated: low-frequency fault part [membership μLF(f)]; Arrange the parts according to their membership values; S4. Regulating historical maintenance records and equipment parameters of the medical accelerator, analyzing the causes of the failures of historical maintenance solutions for each fault area of the medical accelerator, classifying the maintenance solutions into occasional maintenance solutions and specific maintenance solutions through analysis, generating occasional maintenance tags and specific maintenance tags, and grouping the corresponding maintenance solutions; S5. Rank the high-frequency fault locations, medium-frequency fault locations, and low-frequency fault locations classified in S3 according to their levels, and calculate the equipment aging rate OL respectively rate , and substitute the data preprocessed in S2 into the aging rate OL rate calculation formula Its aging rate OL rate The calculation formula is as follows: ； ； ； ； ； Among them, Freq is the usage frequency of the medical accelerator, Time is the total usage time of the medical accelerator, DoseType is the effectiveness factor after classification of the irradiation dose, Cone is the dose size influence factor after standardization, Load is the treatment dose executed within the set time, Repairs is the number of fault repairs of the medical accelerator, and FR is the fault incidence rate of the medical accelerator within the set time. and ... are the corresponding weight coefficients respectively, α, β, γ, δ, η are the corresponding non-linear function parameters respectively, and Thresh1, Thresh2, Thresh3 are the thresholds of the variables. The aging rate of each part of the equipment is calculated. If the aging rate OL rate ≥ the set threshold value threshold, then S6 is performed; if the aging rate OL rate < the set threshold value threshold, then S8 is performed, where the set threshold value threshold is the numerical value of the aging warning line for the performance of the medical accelerator equipment. When the aging rate is lower than the set threshold value, it means that the equipment performance is seriously affected, and vice versa, it means that the equipment performance is sufficient to meet the usage requirements; S6. Based on the aging rate OLrate, the number of repairs Repairs, and the failure rate FR of each part of the analyzed object, predict the failure conditions of each component of the medical accelerator after the patient's medical plan is adopted and generate a recommended solution for the failure condition; S7. Interact with users through the interface on fault prediction results and recommended solutions, providing functions such as prediction report generation, recommended solution query, and prediction reminder; S8. Save the fault prediction result and generate a historical record.

4. The medical accelerator fault warning method based on a large language model according to claim 3, wherein, The S6 also analyzes the historical maintenance records and equipment parameters of the medical accelerator, and the steps are as follows: S61. Calculate the number of fault repairs per unit frequency of use r1 = Repairs / Freq, and the number of fault repairs per unit treatment dose r2 = Repairs / Load. If r1 > z1, and r2 <z2,则判断此维修方案为偶发维修方案,其中z1、z2为预设值,若条件不满足则进入下一步判断; S62. Plot the collected data of the failure rate FR and the treatment dose Load into a scatter plot, with the treatment dose Load and the failure rate FR as the vertical axis. Mark the points corresponding to each set of data on the plot. Analyze the distribution of the points on the scatter plot. If the distribution of the points shows a straight line with a clear linear relationship, determine that a specific maintenance solution is used. If the distribution of the points is scattered with no clear linear relationship, proceed to the next step of determination. S63. Plot the collected data on the failure rate FR and the usage frequency Freq into a scatter plot, with the usage frequency Freq and the failure rate FR as the vertical axis, and mark the points corresponding to each set of data on the graph. Analyze the distribution of the points by observing the scatter plot. If the distribution of the points is in the shape of a straight line with an obvious linear relationship, it is determined to be a specific maintenance plan. If the distribution of the points is scattered and does not have an obvious linear relationship, it is determined to be an occasional maintenance plan.

5. The medical accelerator fault warning method based on a large language model according to claim 3, wherein, The calculation formula for the set threshold is: threshold , Among them, represents the average value of historical aging rate data, represents the standard deviation of historical aging rate data, is a coefficient, whose value range is [1, 2].

Images