A method for estimating medical behavior preferences of lung cancer screening patients
A Bayesian network with latent variables addresses the accuracy issues of existing models, enabling precise patient behavior predictions and informed resource allocation in lung cancer screening.
Patent Information
- Application Number
- CN202311095915.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-08-28
- Publication Date
- 2024-12-20
- Estimated Expiration
- 2043-08-28
AI Technical Summary
Existing prediction models for lung cancer screening patient behavior lack accuracy, especially in large and complex feature spaces, and suffer from underfitting and poor interpretability, leading to inefficient allocation of medical resources.
A method using a Bayesian network with latent variable structure to integrate prior knowledge and data, optimizing latent variable states and training the network with expectation maximization to infer medical behavior preferences, identifying sensitive controllable variables.
The method provides accurate predictions and policy recommendations, improving resource allocation by understanding complex relationships and preferences, enhancing the predictive power in multivariate systems and sparse data scenarios.
Smart Images
Figure CN117114111B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of lung cancer data processing, and in particular to a method for inferring medical behavior preference of lung cancer screening patients. Background Art
[0002] In recent years, the screening effect of lung cancer screening technology based on low-dose spiral CT has been further confirmed, and significant progress has been made in the selection of high-risk groups, screening intervals, and lung nodule management. However, due to the low penetration rate and the misallocation of limited resources, the number of people undergoing this screening is low, which in turn leads to poor lung cancer screening results. Accurate speculation of patients' medical behavior preferences for lung cancer screening will help managers understand patient behavior and further promote the optimal allocation of medical resources for lung cancer screening, thereby fundamentally helping to solve the above problems.
[0003] At present, there are few prediction models for patients' lung cancer screening and medical treatment behavior, and most studies only consider factors such as the flow of people within a single medical institution, the number of visits, and the queue. For patients' preferences for hospitals, the traditional binary logistic regression model is only suitable for small-scale predictions. When the feature space is large and the feature variables are complex, the prediction performance of this method is poor and underfitting is prone to occur. Traditional machine learning models such as artificial neural networks and support vector machines have poor interpretability, which brings great difficulties to the deep understanding of multivariate systems. In actual situations, the influencing factors are complex and changeable, and the data space is huge. Therefore, these traditional models cannot accurately measure and understand the multivariate feature variables of the system and make more accurate predictions. In contrast, Bayesian networks provide a modeling framework for probability graphs, which represents the complex relationship between multiple variables through directed acyclic graphs, and are suitable for solving nonlinear and uncertain problems. This type of model can effectively infer missing data values, integrate multi-source data, capture and deeply understand the complex causal relationship between feature variables, and provide a more potential method for inferring patients' lung cancer screening and medical treatment behavior. Summary of the invention
[0004] The purpose of the present invention is to provide a method for inferring the medical behavior preference of lung cancer screening patients, which solves the problem that ordinary models cannot make accurate predictions, and uses current data information to form a network association relationship, organically combines prior information with sample knowledge, promotes the integration of prior knowledge and data, and ensures that accurate inferences can still be made when sample data is sparse or difficult to obtain, thereby solving the problems of inaccurate inferences due to small number of data samples obtained in traditional models and poor explanatory power for multivariate systems.
[0005] To achieve the above object, the present invention provides a method for estimating the medical behavior preference of lung cancer screening patients, comprising the following steps:
[0006] S1. Obtain data through questionnaire surveys, pre-process the collected data, and screen variables that affect medical behavior preferences;
[0007] S2, defining the variables that affect the medical behavior preference screened out in step S1 to form a data set;
[0008] S3, determining the latent variables of the Bayesian network model according to the variables affecting the medical behavior preference defined in step S2, and optimizing the number of latent variable states;
[0009] S4, constructing a Bayesian network with improved latent variable structure using the variables in step S2 and the latent variables determined in step S3;
[0010] S5, using the latent variable structured improved Bayesian network constructed in step S4 and actual survey data to train a trained model, and perform inference and prediction based on the model;
[0011] S6. Use the improved Bayesian network with latent variables constructed in step S4 and the model trained with actual data in step S5 to further identify sensitive controllable variables that affect patients' medical behavior preferences.
[0012] Preferably, in step S1, non-basic features in the collected data are eliminated, and the variables affecting the preference for medical treatment behavior are divided into two categories: basic personal features and medical health features;
[0013] Among them, personal basic characteristics include five influencing variables: age, education level, family income, gender, and current residence; medical and health characteristics include six influencing variables: family history, lung cancer awareness, CT screening, screening reasons, medical insurance, and health examinations;
[0014] For the inferred target, there are two variables, screening institution and treatment hospital. The treatment hospital is the final target variable, which is used to express the patient's preference for medical behavior.
[0015] Preferably, in step S2, the variables that are screened out and affect the preference for medical treatment behavior are defined to form a data set, including the following steps:
[0016] S21, define the 13 types of variable data selected in step S1, and select a variable state for each variable; in order to reduce the space complexity, each variable state is divided into 2 to 4 levels according to the actual situation;
[0017] The specific variable status division method is as follows: gender is male and female; current residence, screening institution, and hospital are divided into provincial and non-provincial levels by region; family history, medical insurance, lung cancer awareness, and CT screening are judged as yes and no, among which, the option of forgetting is added to CT screening; screening reasons are divided into four states: family and friends' suggestions, doctor's suggestions, physical examination results, and smoking influence; age, education level, and family income are divided into high, medium, and low levels by stage; health check-ups are divided into three states according to frequency: never, occasionally, and frequently;
[0018] S22, the data defined in step S21 are used to form a modeling data set, the modeling data set includes data set 1 and data set 2, wherein data set 1 is used for subsequent improved Bayesian network learning, and data set 2 is used to test the inference accuracy of the improved network.
[0019] Preferably, in step S3, determining the hidden variables and optimizing the number of hidden variable states include the following steps:
[0020] S31, based on the two major influencing factors of personal basic characteristics and medical health characteristics selected in step S1, introduce two latent variables: personal condition latent variable and medical condition latent variable;
[0021] S32. Optimize the number of latent variable states and use KLD as the target criterion to measure the difference between two probability distributions. The KLD divergence is expressed as:
[0022]
[0023] Among them, R is the number of states of the latent variable, K is the index of different states of the latent variable; ρ K is the probability of state K appearing in the historical data distribution; P K Based on the probability of occurrence of state K in the posterior probability distribution of the data; in formula (1), the smaller the KLD corresponding to the different states of the latent variable, the more the posterior probability distribution of the target variable matches the historical data distribution. The number of latent variable states with the smallest KLD is selected as the optimized number of latent variable states.
[0024] Preferably, in step S4, a latent variable structured improved Bayesian network is constructed to determine the multivariate correlation between the two major influencing factors of personal basic characteristics and medical health characteristics, and the multivariate is connected together according to the correlation and the latent variables determined in step S31 to form a latent variable structured Bayesian network structure; wherein the multivariate with correlation is family history, health examination, degree of understanding of lung cancer and reasons for screening and other variables.
[0025] Preferably, in step S5, the constructed model structure is trained with actual data to obtain an optimal model, and reasoning and prediction are performed with the help of the model, including the following steps:
[0026] S51, using the expectation maximization algorithm to learn network parameters: first select all nodes, randomize all conditional probability distribution tables, and compile the network; import the data set 1 prepared in step S2 to learn the network parameters, and after iteration, record the likelihood value;
[0027] S52, repeating step S51 to execute the expectation maximization algorithm with multiple restarts, comparing the log-likelihood values of all records, and selecting the model with the largest log-likelihood value as the optimal model, which is the trained model;
[0028] S53, with the help of the model trained in step S52, infer the posterior probability P(I=i|E=e), which is expressed as:
[0029]
[0030]
[0031] As for the target variable, E is the evidence variable, that is, multiple influencing factors in the network, e is the instantiation of E, I represents the target variable, the real choice of hospital for medical treatment of the population, i represents the instantiation of the target variable, and R represents the maximum number of states of the target variable; through the Bayesian network improved by latent variable structure, new data is introduced for learning and inference to obtain the population's preference for hospital treatment.
[0032] Preferably, in step S6, a variance reduction method is used to perform variable sensitivity analysis and identify sensitive controllable variables that affect the patient's medical behavior preference, including the following steps:
[0033] S61. Variable sensitivity analysis, where variable X n For an instantiation x i The squared error cost C(X n |x i )Use the posterior probability distribution P(X n |x i ) to express:
[0034]
[0035] in, For X n The posterior mean is calculated by averaging all X i , average variance reduction C(X n |X i )for:
[0036]
[0037] Among them, x n Yes X n An example of xi Yes X i An example of is a constant, the larger the average variance reduction value C(X n |X i ) means X n X i The higher the sensitivity, μ Xn|xi Yes X n The posterior mean of ; Yes X n The average value of
[0038] S62, Identification of sensitive controllable variables, based on the controllability perspective, distinguish the controllable and uncontrollable variables in the two major categories of influencing variables: personal basic characteristics and medical health characteristics. Among them, five influencing variables, namely, lung cancer awareness, CT screening, screening reasons, health examinations, and medical insurance, are controllable. By distinguishing the controllable variables X c and the uncontrollable variable X uc , and then according to the variable sensitivity ranking identified in step S61, the variable sensitivity order of the controllable variables is given again; among them, the controllable variables are sorted according to the sensitivity level in the order of screening reasons, CT screening, medical insurance, health examination, and lung cancer understanding.
[0039] Therefore, the present invention adopts the above-mentioned method for estimating the medical behavior preference of lung cancer screening patients, and its technical effects are as follows:
[0040] (1) The present invention describes the relationship between data, which is easy to understand and can intuitively give prediction results. At the same time, it can see the influence of each characteristic variable on the target variable, so as to make targeted policy recommendations and intervene in related variables through policies, ultimately improving the patient's medical experience.
[0041] (2) Compared with the general Bayesian network with a comparable network size, the Bayesian network based on latent variable structured improvement has a smaller parameter space, smaller computational complexity and better generalization performance, and can cope with complex scenarios such as uncertainty, nonlinearity and multivariate relationships. By using the current data information to form network associations, the prior information and sample knowledge are organically combined, which promotes the integration of prior knowledge and data, ensuring that accurate inferences can still be made when sample data is sparse or difficult to obtain, solving the problem of inaccurate inferences due to the small number of data samples obtained in traditional models and poor explanatory power for multivariate systems.
[0042] (3) Using this network, we found that income is an important variable that affects residents’ choice of hospital, which is reflected in the level of medical expenses paid by residents. Therefore, it is recommended to include this examination in medical insurance to reduce the burden on residents and facilitate the promotion of new lung cancer screening technologies. In addition, using this network, we can recognize that the degree of understanding of lung cancer is an important characteristic variable that affects residents’ medical preferences. Therefore, the popularization of lung cancer-related knowledge and the promotion of new technologies have become key directions.
[0043] (4) The prediction results show that a large number of residents prefer to undergo lung cancer screening in non-provincial hospitals, and there is a certain gap between primary medical care and the advanced technology and equipment of provincial hospitals. Therefore, it is suggested to strengthen policy support for primary medical institutions, implement corresponding medical assistance for low-income and remote areas, increase resource investment in primary medical institutions, and popularize new technologies to primary hospitals.
[0044] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] Figure 1 This is a schematic diagram of the overall process of a method for inferring the medical behavior preference of lung cancer screening patients;
[0046] Figure 2 Bayesian network modeling diagram for inferring the medical preferences of Chengdu residents. DETAILED DESCRIPTION
[0047] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.
[0048] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.
[0049] Embodiment 1
[0050] like Figure 1 As shown, the present invention provides a method for estimating the medical behavior preference of lung cancer screening patients, comprising the following steps:
[0051] S1. Obtain data through questionnaire surveys, pre-process the collected data, and screen variables that affect medical behavior preferences;
[0052] Eliminate non-basic characteristics from the collected data and divide the variables that affect medical behavior preferences into two categories: basic personal characteristics and medical and health characteristics;
[0053] Among them, personal basic characteristics include five influencing variables: age, education level, family income, gender, and current residence; medical and health characteristics include six influencing variables: family history, lung cancer awareness, CT screening, screening reasons, medical insurance, and health examinations;
[0054] For the inferred target, there are two variables, screening institution and treatment hospital. The treatment hospital is the final target variable, which is used to express the patient's preference for medical behavior.
[0055] S2, defining the variables that affect the medical behavior preference screened out in step S1 to form a data set;
[0056] S21, define the 13 types of variable data selected in step S1, and select a variable state for each variable; in order to reduce the space complexity, each variable state is divided into 2 to 4 levels according to the actual situation;
[0057] The specific variable status division method is as follows: gender is male and female; current residence, screening institution, and hospital are divided into provincial and non-provincial levels by region; family history, medical insurance, lung cancer awareness, and CT screening are judged as yes and no, among which, the option of forgetting is added to CT screening; screening reasons are divided into four states: family and friends' suggestions, doctor's suggestions, physical examination results, and smoking influence; age, education level, and family income are divided into high, medium, and low levels by stage; health check-ups are divided into three states according to frequency: never, occasionally, and frequently;
[0058] S22, the data defined in step S21 are used to form a modeling data set, the modeling data set includes data set 1 and data set 2, wherein data set 1 is used for subsequent improved Bayesian network learning, and data set 2 is used to test the inference accuracy of the improved network.
[0059] S3, determining the latent variables of the Bayesian network model according to the variables affecting the medical behavior preference defined in step S2, and optimizing the number of latent variable states;
[0060] S31, based on the two major influencing factors of personal basic characteristics and medical health characteristics selected in step S1, introduce two latent variables: personal condition latent variable and medical condition latent variable;
[0061] S32. Optimize the number of latent variable states and use KLD as the target criterion to measure the difference between two probability distributions. KLD is expressed as:
[0062]
[0063] Among them, R is the number of states of the latent variable, K is the index of different states of the latent variable; ρ K is the probability of state K appearing in the historical data distribution; P KBased on the probability of occurrence of state K in the posterior probability distribution of the data; in formula (1), the smaller the KLD corresponding to the different states of the latent variable, the more the posterior probability distribution of the target variable matches the historical data distribution. The number of latent variable states with the smallest KLD is selected as the optimized number of latent variable states.
[0064] S4, constructing a Bayesian network with improved latent variable structure using the variables in step S2 and the latent variables determined in step S3;
[0065] In step S4, a Bayesian network with improved structure of latent variables is constructed to determine the multivariate correlation among the two major influencing factors of basic personal characteristics and medical health characteristics. The multivariate is connected together according to the correlation and the latent variables determined in step S31 to form a Bayesian network structure with latent variables structured; among them, the multivariate with correlation is family history, health examination, degree of understanding of lung cancer and reasons for screening.
[0066] S5, using the latent variable structured improved Bayesian network constructed in step S4 and the actual survey data to train a trained model, and perform inference and prediction based on the model;
[0067] In step S5, the constructed model structure is trained with actual data to obtain the best model, and the model is used to perform inference prediction, including the following steps:
[0068] S51, using the expectation maximization algorithm to learn network parameters: first select all nodes, randomize all conditional probability distribution tables, and compile the network; import the data set 1 prepared in step S2 to learn the network parameters, and after iteration, record the likelihood value;
[0069] S52, repeating step S51 to execute the expectation maximization algorithm with multiple restarts, comparing the log-likelihood values of all records, and selecting the model with the largest log-likelihood value as the optimal model, which is the trained model;
[0070] S53, with the help of the model trained in step S52, infer the posterior probability P(I=i|E=e), which is expressed as:
[0071]
[0072]
[0073] As for the target variable, E is the evidence variable (which can be a combination of multiple variables), that is, multiple influencing factors in the network, e is the instantiation of E, I represents the target variable, the real choice of hospital for medical treatment of the population, i represents the instantiation of the target variable, and R represents the maximum number of states of the target variable; through the Bayesian network improved by latent variable structure, new data is introduced for learning and inference to obtain the population's preference for hospital treatment.
[0074] S6. Use the improved Bayesian network with latent variables constructed in step S4 and the model trained with actual data in step S5 to further identify sensitive controllable variables that affect patients' medical behavior preferences.
[0075] Variance reduction method was used to conduct variable sensitivity analysis and identify sensitive controllable variables that affect patients' medical behavior preferences, including the following steps:
[0076] S61. Variable sensitivity analysis, where variable X n For an instantiation x i The squared error cost C(X n |x i )Use the posterior probability distribution P(X n |x i ) to express:
[0077]
[0078] in, For X n The posterior mean is calculated by averaging all X i , average variance reduction C(X n |X i )for:
[0079]
[0080] Among them, x n Yes X n An example of x i Yes X i An example of is a constant, the larger the average variance reduction value C(X n |X i ) means X n X i The higher the sensitivity, μ Xn|xi Yes X n The posterior mean of ; Yes X n The average value of
[0081] S62, Identification of sensitive controllable variables, based on the controllability perspective, distinguish the controllable and uncontrollable variables in the two major categories of influencing variables: personal basic characteristics and medical health characteristics. Among them, five influencing variables, namely, lung cancer awareness, CT screening, screening reasons, health examinations, and medical insurance, are controllable. By distinguishing the controllable variables X c and the uncontrollable variable X uc , and then give the variable sensitivity order to the controllable variables according to the variable sensitivity ranking identified in step S61; wherein the controllable variables are sorted according to the sensitivity in the order of screening reasons, CT screening, medical insurance, health examination, and lung cancer awareness.
[0082] The technical effect of the present invention is described below through specific experiments:
[0083] Experimental conditions: A latent variable structured improved Bayesian network was constructed according to the above steps, and the model effect was verified based on the Qingjiang District dataset of Chengdu and the Chenghua District dataset of Chengdu.
[0084] Experimental process: Using the constructed latent variable structured improved Bayesian network, the data sets of Qingbaijiang District and Chenghua District of Chengdu were used for network learning using the expectation maximization algorithm, and the confusion matrix was used for accuracy calculation.
[0085] Experimental results:
[0086] (1) Figure 2 It can be seen that 37.8% of residents prefer to undergo lung cancer screening in non-provincial hospitals.
[0087] (2) The calculation results of the estimation accuracy of Qingbaijiang District and Chenghua District of Chengdu City show that the accuracy rate of Qingbaijiang District is 93.6%, and the accuracy rate of Chenghua District is 89.5%.
[0088] Therefore, the present invention adopts the above-mentioned method for inferring the medical behavior preference of lung cancer screening patients, which solves the problem that ordinary models cannot make accurate predictions, and uses current data information to form network associations, organically combines prior information with sample knowledge, and promotes the integration of prior knowledge and data, ensuring that accurate inferences can still be made when sample data is sparse or difficult to obtain, and solves the problems of traditional models with inaccurate inferences due to small number of data samples and poor explanatory power for multivariate systems.
[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.
Claims
1. A method for estimating the medical behavior preference of lung cancer screening patients. It is characterized in that The following steps are involved: S1. Obtain data through questionnaire surveys, pre-process the collected data, and screen variables that affect medical behavior preferences; S2, defining the variables that affect the medical behavior preference screened out in step S1 to form a data set; S3, determining the latent variables of the Bayesian network model according to the variables affecting the medical behavior preference defined in step S2, and optimizing the number of latent variable states; S4, constructing a Bayesian network with improved latent variable structure using the variables in step S2 and the latent variables determined in step S3; S5, using the latent variable structured improved Bayesian network constructed in step S4 and the actual survey data to train a trained model, and perform inference and prediction based on the model; S6, using the improved Bayesian network with latent variables constructed in step S4 and the model trained with actual data in step S5 to further identify sensitive controllable variables that affect patients' medical behavior preferences; In step S1, non-basic characteristics in the collected data are eliminated, and the variables that affect the preference for medical treatment behavior are divided into two categories: basic personal characteristics and medical and health characteristics; Among them, personal basic characteristics include five influencing variables: age, education level, family income, gender, and current residence; medical and health characteristics include six influencing variables: family history, lung cancer awareness, CT screening, screening reasons, medical insurance, and health examinations; For the inferred target, there are two variables: screening institution and hospital. The hospital is the final target variable, which is used to express the patient's preference for medical behavior. In step S3, determining the hidden variables and optimizing the number of hidden variable states includes the following steps: S31, based on the two major influencing factors of personal basic characteristics and medical health characteristics selected in step S1, introduce two latent variables: personal condition latent variable and medical condition latent variable; S32, optimize the number of hidden variable states, using KLD As a target criterion, it is used to measure the difference between two probability distributions. KLD It is expressed as: （1） in, R is the number of states of the hidden variable, K is the index of different states of the latent variable; Is the state in the historical data distribution K Probability of occurrence; Based on the state of the data in the posterior probability distribution K Probability of occurrence; In formula (1), different states of hidden variables correspond to KLD The smaller it is, the more the posterior probability distribution of the target variable matches the historical data distribution. KLD The minimum number of hidden variable states is taken as the optimized number of hidden variable states; In step S4, a Bayesian network with structured improvement of latent variables is constructed to determine the correlation between variables in the two major influencing factors of basic personal characteristics and medical health characteristics. The variables are connected together according to the correlation and the latent variables determined in step S31 to form a Bayesian network structure with structured latent variables; wherein the variables with correlation are family history, health examination, degree of knowledge of lung cancer and reason for screening; In step S5, the constructed model structure is trained with actual data to obtain the best model, and the model is used to perform inference prediction, including the following steps: S51, using the expectation maximization algorithm to learn network parameters: first select all nodes, randomize all conditional probability distribution tables, and compile the network; import the data set 1 prepared in step S2 to learn the network parameters, and after iteration, record the likelihood value; S52, repeating step S51 to execute the expectation maximization algorithm with multiple restarts, comparing the log-likelihood values of all records, and selecting the model with the largest log-likelihood value as the optimal model, which is the trained model; S53: Infer the posterior probability using the model trained in step S52 , its formula is expressed as: （2） （3） As for the target variable, E is the evidence variable, i.e., multiple influencing factors in the network, e is the instantiation of E, and I represents the target variable, i.e., the actual hospital choice of the population. represents the instantiation of the target variable, and R represents the maximum number of states of the target variable. Through the Bayesian network with improved latent variable structure, new data is introduced for learning and inference to obtain the hospital preference of the population. In step S6, a variance reduction method is used to perform variable sensitivity analysis and identify sensitive controllable variables that affect patients' medical behavior preferences, including the following steps: S61. Variable sensitivity analysis, where variables For an instantiation The squared error cost Using the posterior probability distribution To express: （4） in, for The posterior mean is calculated by averaging all , mean variance reduction for: （5） in, yes An example of yes An example of is a constant, the larger the average variance reduction value mean right The higher the sensitivity; yes The posterior mean of ; yes The average value of S62, Identification of sensitive controllable variables, based on the controllability perspective, distinguish between controllable and uncontrollable variables in the two major categories of influencing variables: basic personal characteristics and medical health characteristics. Among them, five influencing variables, namely, lung cancer awareness, CT screening, screening reasons, health examinations, and medical insurance, are controllable. By distinguishing controllable variables and uncontrollable variables , and then according to the variable sensitivity ranking identified in step S61, the variable sensitivity order of the controllable variables is given again; among them, the controllable variables are sorted according to the sensitivity level in the order of screening reasons, CT screening, medical insurance, health examination, and lung cancer understanding.
2. According to claim 1, a method for estimating the medical behavior preference of lung cancer screening patients, It is characterized in that In step S2, the variables that are screened out and affect the preference for medical treatment behavior are defined to form a data set, including the following steps: S21, define the 13 types of variable data selected in step S1, and select a variable state for each variable; to reduce the space complexity, each variable state is divided into 2 to 4 levels according to the actual situation; The specific variable status division method is as follows: gender is male and female; current residence, screening institution, and hospital are divided into provincial and non-provincial levels by region; family history, medical insurance, lung cancer awareness, and CT screening are judged as yes and no, among which, the option of forgetting is added to CT screening; screening reasons are divided into four states: family and friends' suggestions, doctor's suggestions, physical examination results, and smoking influence; age, education level, and family income are divided into high, medium, and low levels by stage; health check-ups are divided into three states according to frequency: never, occasionally, and frequently; S22, the data defined in step S21 are used to form a modeling data set, the modeling data set includes data set 1 and data set 2, wherein data set 1 is used for subsequent improved Bayesian network learning, and data set 2 is used to test the inference accuracy of the improved network.