Equipment performance evaluation index screening method based on Bayesian network

Through Bayesian network analysis of the performance evaluation indicators of complex equipment, the problem of screening complex system and group equipment evaluation indicators in the existing technology is solved, and efficient and economical equipment efficiency evaluation is achieved, reducing the evaluation cost and cycle.

CN118965146APending Publication Date: 2024-11-15UNIT 63892 OF PLA
View PDF 0 Cites 3 Cited by

Patent Information

Application Number
CN202410850681.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-06-27
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

It is difficult for the prior art to effectively screen the effectiveness evaluation indicators of complex systems and group equipment under system task conditions, resulting in high evaluation costs, long cycles and limited data acquisition.

Method used

Using the Bayesian network-based equipment performance evaluation index screening method, the initial index set is determined through functional analysis, characteristic analysis and system task analysis, and the experimental data is obtained using simulation and semi-physical systems, data preprocessing and discrete quantization are performed, and domain knowledge is combined. Perform Bayesian network structure and parameter learning, analyze indicator sensitivity, and determine the final evaluation indicator set.

Benefits of technology

It realizes reasonable screening of the performance evaluation index system of complex equipment systems, reduces the evaluation cost and cycle, and improves the credibility and economicality of the evaluation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118965146A_ABST
    Figure CN118965146A_ABST
Patent Text Reader

Abstract

An equipment performance evaluation index screening method based on a Bayesian network comprises the following steps: preliminarily determining a performance evaluation index initial set ZB through tested analysis; the method comprises the following steps: acquiring system test data by using a simulation system and a semi-physical system, and performing preprocessing and index calculation on the data to acquire a test and evaluation original sample data set YS; performing discrete quantization processing on each index set in the test evaluation original sample data set YS to obtain a training sample set TZ; performing Bayesian network structure learning based on domain knowledge and the sample data set to obtain a structure model BNS of a Bayesian network; parameter learning is carried out to acquire a completed Bayesian network model BNall; and analyzing the sensitivity between each initial index and efficiency, and determining a final evaluation index set ZB through sensitivity sorting, so as to assist in constructing a complex equipment system efficiency evaluation index system.
Need to check novelty before this filing date? Find Prior Art

Description

Technical field:

[0001] The present invention belongs to the technical field of equipment effectiveness evaluation, and mainly relates to a method for screening equipment effectiveness evaluation indicators based on a Bayesian network. Background technology:

[0002] Equipment effectiveness evaluation is an objective and scientific evaluation of whether the equipment can play its due role and realize its own value. In the process of equipment effectiveness evaluation, the construction of the evaluation index system is a basic and key task, and is the core factor affecting the feasibility, economy and credibility of the entire evaluation. For complex systems, the evaluation of their effectiveness often requires the construction of a system task that includes the equipment under test, a limited number of tests based on the system test design, and then the evaluation results are finally obtained by calculating and aggregating step by step according to the evaluation index system.

[0003] Since system testing involves many factors, high costs, long cycles, limited number of tests and available data, how to screen out key evaluation factors and determine a reasonable indicator system is of great theoretical and practical significance for equipment effectiveness evaluation.

[0004] Traditional evaluation indicators are mainly determined by experts based on the characteristics of the object being measured, domain knowledge and historical experience. This evaluation method is suitable for the evaluation of single equipment and simple tasks, but it is difficult to implement the performance evaluation of complex systems and group equipment under system task conditions. In view of this, the present invention proposes a method for screening equipment effectiveness evaluation indicators based on Bayesian networks to assist in the construction of a complex equipment system effectiveness evaluation indicator system. Summary of the invention:

[0005] In order to overcome the above-mentioned shortcomings, the present invention provides a method for screening equipment effectiveness evaluation indicators based on a Bayesian network.

[0006] The technical solution adopted by the present invention to solve its technical problem is:

[0007] A method for screening equipment effectiveness evaluation indicators based on Bayesian networks, the specific steps are as follows:

[0008] Step 1: Through functional analysis, characteristic analysis, and system task analysis of the object under test, domain experts preliminarily determine the initial set of performance evaluation indicators ZB;

[0009] Step 2: Based on the initial set of performance evaluation indicators ZB, use the simulation system and semi-physical system to obtain system test data, pre-process the data, and calculate indicators to obtain the original sample data set YS for test evaluation;

[0010] Step 3: Perform discrete quantization processing on each indicator set in the test evaluation original sample data set YS to obtain the training sample set TZ for discrete Bayesian network learning;

[0011] Step 4: Based on the training sample TZ obtained in step 3, Bayesian network structure learning is performed based on domain knowledge and sample data sets to obtain the Bayesian network structure model BN_S;

[0012] Step 5: Perform parameter learning on the basis of the Bayesian network structure model BN_S obtained in step 4 to obtain the completed Bayesian network model BN_All;

[0013] Step 6: Use the constructed Bayesian network model BN_All to analyze the sensitivity between each initial indicator and performance, and determine the final evaluation indicator set ZB through sensitivity sorting.

[0014] The object to be measured in the step 1 is the reconnaissance equipment, and the target tracking task of the reconnaissance equipment is evaluated for performance. The initial set of performance evaluation indicators includes: single target tracking accuracy, single target tracking success rate, multi-target tracking precision, multi-target tracking accuracy, recognition accuracy, recognition recall rate and average ratio.

[0015] In the step 2, the test data obtained by the simulation system or the semi-physical system is used to calculate the indicators of each sample data according to the calculation formula of each indicator in the initial indicator set to obtain the original sample data set;

[0016] The Dixon criterion is used to judge the outliers of the original sample data, eliminate unreasonable abnormal data, and form the final test evaluation original sample data set YS.

[0017] The Dixon Criteria evaluation method is as follows:

[0018] Assume the sample is x0,x0,…,x0, and its order statistic is

[0019] x (0) <x (0) <… <x (0)

[0020] Among them, x (0) is the minimum value, x (0) is the maximum value. For different sample sizes n, the choice of statistics is different. When the order statistic x (0) When the data obey the normal distribution, Dixon gives the statistic R for different sample sizes n: ij The calculation formula of

[0021] In a one-sided test, the statistic R ij The choices are as follows:

[0022] When 3≤n≤7, (or )

[0023] When 8≤n≤10, (or )

[0024] When 11≤n≤13, (or )

[0025] When 14≤n≤30, (or )

[0026] Extending to the general case,

[0027]

[0028] Assuming the significance level is α, Dixon gives its critical value D 000 (n);

[0029] If the statistic R of a sample 00 >D 000 (n) or R000>D 000 (n), then judge x (0) or x (0) is an outlier;

[0030] In a two-sided test, when R 00 >R000, Judge x (0) is an outlier;

[0031] When R000>R 00 , Judge x (0) is an outlier.

[0032] The steps of performing linear discrete quantization processing on each indicator set in the test evaluation original sample data set YS in step 3 are as follows:

[0033] For index X, let Ω be the index space, K max is the maximum value of the indicator, K min is the minimum value, N is the number of states, then, for the indicator value x, the state it belongs to is S x :

[0034]

[0035] In step 4, the method for learning the Bayesian network structure based on domain knowledge and sample data is as follows:

[0036] After the training data D is given, the posterior probability of the Baye network G is used as the score;

[0037] It is easy to get from the Bayesian formula:

[0038]

[0039] In formula (2), P(D) will not affect different structures G, so it can be ignored. Ignoring this term and taking the logarithm, we can get:

[0040] log P(G|D)∝log P(G,D)=log P(D|G)+log P(G) (3)

[0041] The P(D|G) term in the above formula is called the marginal likelihood and is calculated as follows:

[0042] P(D|G)=∫P(D|G,Θ)P(Θ|G)dΘ (4)

[0043] Taking a uniform distribution for Θ, we can get the CH score:

[0044]

[0045] Taking the Dirichlet distribution, we can get the BDe score:

[0046]

[0047] In the scoring function, α ij· is the hyperparameter to be set for node i, and the hyperparameter settings are evenly distributed;

[0048] m ijk For data set D that satisfies X i = k and Pa(X i ) = the number of samples of j, r i For node X i The number of state values;

[0049] Use the BDe score, which requires a hyperparameter α ijk When the default structural prior information follows a uniform distribution, the value of α ij k takes uniform distribution:

[0050] α ijk =αP(X i =k,π(X i )=j|G)=α / (r i q i ) (7)

[0051] Substituting formula (7) into the BDe score, the BDeu score can be obtained.

[0052] In the step 5, the method for performing parameter learning based on the Bayesian network structure model BN_S obtained in step 4 is as follows:

[0053] In Bayesian parameter learning, the parameter Θ is regarded as a random variable, and the final value can be obtained by calculating the posterior probability of Θ; the calculation formula is as follows:

[0054] p(Θ|D)=p(D|Θ)·p(Θ) / p(D) (8)

[0055] The prior distribution P(Θ) in formula (8) follows the Dirichlet distribution:

[0056]

[0057] Then substitute into formula (10) to get the posterior probability of Θ:

[0058]

[0059] In formula (9) and (10) α1,...,α r It is a hyperparameter, which contains prior knowledge;

[0060] Finally, we can get θ k Estimates:

[0061]

[0062] In step 6, the method of analyzing the sensitivity between each initial indicator and the performance using the constructed Bayesian network model BN_All is as follows:

[0063] The calculation method of sensitivity index based on the degree of change of conditional probability, taking X1 as an example,

[0064]

[0065] Where:

[0066] q——Number of states of Y

[0067] r x1 ——X i The number of value states

[0068] Through parameter learning, we can obtain P(X1), P(X i |π(X i )) and P(Y|X1,X2,…,X n ), where i = 2,…,n; from the conditional probability formula and direct inference algorithm, we know that:

[0069]

[0070] Substituting formulas (15) and (16) into formula (14) yields the contribution of index X1 to performance Y.

[0071] In step 6, the method of analyzing the sensitivity between each initial indicator and the performance using the constructed Bayesian network model BN_All is as follows:

[0072] The calculation method of sensitivity index based on the degree of state change, taking X1 as an example,

[0073]

[0074] q——the number of states of Y, ——X i The number of value states of ;

[0075] Through parameter learning, the size of P(Y=a|X1=k) and P(Y=a|X1=b) can be obtained, and then the quantitative sensitivity index can be obtained through formula (17).

[0076] Due to the adoption of the above-mentioned technical solution, the present invention has the following advantages:

[0077] The present invention provides a method for screening equipment effectiveness evaluation indicators based on a Bayesian network. The method preliminarily determines an initial set ZB of effectiveness evaluation indicators by analyzing the tested object; obtains system test data by using a simulation system and a semi-physical system, and performs preprocessing and indicator calculation on the data to obtain a test evaluation original sample data set YS; performs discrete quantization processing on each indicator set in the test evaluation original sample data set YS to obtain a training sample set TZ; performs Bayesian network structure learning based on domain knowledge and sample data sets to obtain a Bayesian network structure model BN_S; performs parameter learning to obtain a completed Bayesian network model BN_All; analyzes the sensitivity between each initial indicator and effectiveness, and determines the final evaluation indicator set ZB by ranking the sensitivity, thereby assisting in constructing a complex equipment system effectiveness evaluation indicator system. Description of the drawings:

[0078] Figure 1 It is an equipment effectiveness evaluation index screening process based on Bayesian network. Specific implementation method:

[0079] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0080] Combined with Figure 1 A method for screening equipment effectiveness evaluation indicators based on a Bayesian network is shown, and the specific steps are as follows:

[0081] Step 1: Through the function analysis, characteristic analysis and system task analysis of the object to be measured, the domain experts preliminarily determine the initial set of performance evaluation indicators ZB Ini ={x0,x1,…,x n}, n is the number of initial indicators;

[0082] This step needs to be specifically analyzed in combination with the characteristics of the equipment being tested and the tasks. Different equipment and tasks have different initial indicator sets. Taking the target tracking task of reconnaissance equipment as an example, the following initial indicator sets can be considered:

[0083] ①Single target tracking accuracy

[0084] The distance between the true target and the tracking centerline point.

[0085]

[0086] Where x0, y0 represent the real point position, x0, y0 represent the tracking center point coordinates

[0087] ②Single target tracking success rate

[0088] The success rate is calculated by calculating the intersection of pixels in the predicted box and the true value box. The predicted area is recorded as a, and the true area is recorded as b. The overlap rate is defined as:

[0089]

[0090] Then, the success rate can be obtained by calculating the AUC (Area under curve) of the overlap rate.

[0091] ③Multi-target tracking accuracy (MOTA)

[0092] The score MOTA of multi-target tracking accuracy is calculated as follows:

[0093]

[0094] m0 represents the number of false positives (FN) in the entire video sequence, fp0 represents the number of false negatives (FP) in the entire video sequence, mme0 represents the number of identity exchanges (ID Sw.) during tracking, and g0 represents the number of real targets in the entire video sequence.

[0095] ④Multi-target tracking accuracy (MOTP)

[0096] The score of multi-target tracking precision MOTP is calculated as follows:

[0097]

[0098] C0 represents the total number of matches at the tth frame, and d represents the distance between the hypothesized bounding box and the true bounding box.

[0099] ⑤Identification Precision (IDP)

[0100] To evaluate the quality of the tracker as a whole, the score of the identification precision IDP is calculated as follows:

[0101]

[0102] Among them, IDTP can be regarded as the number of correctly assigned detection targets in the entire video, IDFN is the number of missed detection targets in the entire video, and IDFP is the number of incorrectly assigned detection targets in the entire video.

[0103] ⑥Identification recall rate (IDR)

[0104] It is the ratio of the number of correctly predicted targets to the number of true targets when the IDF1-score is the highest. The IDR score is calculated as follows:

[0105]

[0106] Among them, IDTP can be regarded as the number of correctly assigned detection targets in the entire video, IDFN is the number of missed detection targets in the entire video, and IDFP is the number of incorrectly assigned detection targets in the entire video.

[0107] ⑦ Average ratio (IDF1)

[0108] IDF1 refers to the ratio of the number of correct target detections to the average of the true number and the calculated number of detections. Here, the IDF1 score is calculated as follows:

[0109]

[0110] Among them, IDTP can be regarded as the number of correctly assigned detection targets in the entire video, IDFN is the number of missed detection targets in the entire video, and IDFP is the number of incorrectly assigned detection targets in the entire video.

[0111] Step 2: Based on the initial set of performance evaluation indicators ZB, use the simulation system and semi-physical system to obtain system test data, pre-process the data, and calculate the indicators to obtain the original sample data set YS = {x0, x1, ..., x m}, m is the number of simulation sample data;

[0112] According to the calculation formula of each indicator, the indicators of each sample data are calculated to obtain the original data sample. At this time, it is necessary to use the Dixon criterion to judge the outliers of the data, eliminate unreasonable abnormal data, and form the final evaluation sample data set. The Dixon criterion evaluation method is as follows:

[0113] Assume the sample is x0,x0,…,x0, and its order statistic is

[0114] x (0) <x (0) <… <x (0)

[0115] Among them, x (0) is the minimum value, x (0) is the maximum value. For different sample sizes n, the choice of statistics is different. When the order statistic x (0) When the data obey the normal distribution, Dixon gives the statistic R for different sample sizes n: ij The calculation formula for .

[0116] In a one-sided test, the statistic R ij The choices are as follows:

[0117] When 3≤n≤7, (or )

[0118] When 8≤n≤10, (or )

[0119] When 11≤n≤13, (or )

[0120] When 14≤n≤30, (or )

[0121] Extending to the general case,

[0122]

[0123] Assuming the significance level is α, Dixon gives its critical value D 000 (n). If the statistic R of a sample 00 >D 000 (n) or R000>D 000 (n), then judge x (0) or x (0) is an outlier.

[0124] In a two-sided test, when R 00 >R000, Judge x (0) is an abnormal value; when R000>R 00 , Judge x (0) is an outlier.

[0125] Step 3: Perform discrete quantization processing on each indicator set in the test evaluation original sample data set YS to obtain the training sample set TZ = {S0, S1, ..., S m}, m is the number of simulation sample data;

[0126] The indicator feature discretization method adopts a linear discretization method, which is as follows:

[0127] For index X, let Ω be the index space, K max is the maximum value of the indicator, K min is the minimum value, N is the number of states, then, for the indicator value x, the state it belongs to is S x :

[0128]

[0129] Step 4: Based on the training sample TZ obtained in step 3, Bayesian network structure learning is performed based on domain knowledge and sample data sets to obtain the Bayesian network structure model BN_S;

[0130] The specific algorithm for learning the Bayesian network structure that integrates domain knowledge is as follows:

[0131] After the training data D is given, the posterior probability of the Baye network G is used as the score.

[0132] It is easy to get from the Bayesian formula:

[0133]

[0134] In formula (2), P(D) will not affect different structures G, so it can be ignored. Ignoring this term and taking the logarithm, we can get:

[0135] logP(G|D)∝logP(G,D)=logP(D|G)+logP(G) (14)

[0136] The P(D|G) term in the above formula is called the marginal likelihood and is calculated as follows:

[0137] P(D|G)=∫P(D|G,Θ)P(Θ|G)dΘ (15)

[0138] Taking a uniform distribution for Θ, we can get the CH score:

[0139]

[0140] Taking the Dirichlet distribution, we can get the BDe score:

[0141]

[0142] In the scoring function, α ij· It is the hyperparameter to be set for node i. The setting of hyperparameter is difficult and is generally taken as uniform distribution. ijk For data set D that satisfies X i = k and Pa(X i ) = the number of samples of j, r i For node X i The number of status values.

[0143] Currently, Bayesian scoring generally uses the BDe score, which requires a hyperparameter α ijk When the default structural prior information follows a uniform distribution, the value of α ijk Take a uniform distribution:

[0144]

[0145] Substituting formula (7) into the BDe score, the BDeu score can be obtained.

[0146] Step 5: Perform parameter learning on the basis of the Bayesian network structure model BN_S obtained in step 4 to obtain the completed Bayesian network model BN_All;

[0147] The Bayesian network parameter learning algorithm used is as follows:

[0148] Similar to the Bayesian scoring function in structure learning, in Bayesian parameter learning, the parameter Θ is regarded as a random variable, and the final value can be obtained by calculating the posterior probability of Θ. The calculation formula is as follows:

[0149] p(Θ|D)=p(D|Θ)·p(Θ) / p(D) (19)

[0150] The prior distribution P(Θ) in formula (8) obeys the Dirichlet distribution:

[0151]

[0152] Then substitute into formula (10) to get the posterior probability of Θ:

[0153]

[0154] In formula (9) and (10) α1,...,α r is a hyperparameter that contains prior knowledge.

[0155] Finally, we can get θ k Estimates:

[0156]

[0157] Step 6: Use the constructed Bayesian network model BN_All to analyze the sensitivity between each initial indicator and performance, and determine the final evaluation indicator set ZB={x0,x1,…,x k}, k is the number of indicators after screening.

[0158] The present invention adopts the following two sensitivity analysis methods based on Bayesian networks:

[0159] (1) Sensitivity index calculation method based on the degree of change of conditional probability. Take X1 as an example, as shown in formula (14). The mathematical meaning of formula (14) is: the contribution of the index to the performance is measured by the degree to which the conditional probability of the performance node changes with the change of the values ​​of other related indicators.

[0160]

[0161] Where:

[0162] q——Number of states of Y

[0163] ——X i The number of value states

[0164] Through parameter learning, we can obtain P(X1), P(X i |π(X i )) and P(Y|X1,X2,…,X n ), where i = 2,…,n. From the conditional probability formula and direct inference algorithm, we can know that:

[0165]

[0166] Substituting formula (15) and (16) into formula (14) yields the contribution of indicator X1 to performance Y:

[0167] (2) The calculation method of sensitivity index based on the degree of state change, taking X1 as an example, is shown in formula (17). The mathematical meaning of formula (17) is: the contribution of relevant indicators to the efficiency is measured by using the average change rate of the efficiency node state value as the state values ​​of other relevant indicator variables change.

[0168]

[0169] q——the number of states of Y, ——Xi Through parameter learning, the size of P(Y=a|X1=k) and P(Y=a|X1=b) can be obtained, and then the quantitative sensitivity index can be obtained through formula (17).

[0170] The parts not described in detail in the above content are prior art, so they are not described in detail.

Claims

1. A method for screening equipment effectiveness evaluation indicators based on Bayesian networks, characterized in that: The specific steps are as follows: Step 1: Through functional analysis, characteristic analysis, and system task analysis of the object under test, domain experts preliminarily determine the initial set of performance evaluation indicators ZB; Step 2: Based on the initial set of performance evaluation indicators ZB, use the simulation system and semi-physical system to obtain system test data, pre-process the data, and calculate indicators to obtain the original sample data set YS for test evaluation; Step 3: Perform discrete quantization processing on each indicator set in the test evaluation original sample data set YS to obtain the training sample set TZ for discrete Bayesian network learning; Step 4: Based on the training sample TZ obtained in step 3, Bayesian network structure learning is performed based on domain knowledge and sample data sets to obtain the Bayesian network structure model BN_S; Step 5: Perform parameter learning on the basis of the Bayesian network structure model BN_S obtained in step 4 to obtain the completed Bayesian network model BN_All; Step 6: Use the constructed Bayesian network model BN_All to analyze the sensitivity between each initial indicator and performance, and determine the final evaluation indicator set ZB through sensitivity sorting.

2. The equipment effectiveness evaluation index screening method based on Bayesian network according to claim 1 is characterized in that: The object to be measured in the step 1 is the reconnaissance equipment, and the target tracking task of the reconnaissance equipment is evaluated for performance. The initial set of performance evaluation indicators includes: single target tracking accuracy, single target tracking success rate, multi-target tracking precision, multi-target tracking accuracy, recognition accuracy, recognition recall rate and average ratio.

3. The equipment effectiveness evaluation index screening method based on Bayesian network according to claim 2 is characterized in that: In the step 2, the test data obtained by the simulation system or the semi-physical system is used to calculate the indicators of each sample data according to the calculation formula of each indicator in the initial indicator set to obtain the original sample data set; The Dixon criterion is used to judge the outliers of the original sample data, eliminate unreasonable abnormal data, and form the final test evaluation original sample data set YS.

4. The equipment effectiveness evaluation index screening method based on Bayesian network according to claim 3 is characterized in that: The Dixon Criteria evaluation method is as follows: Let the samples be x1,x2,...,x n , whose order statistic is x (1) <x (2) <… <x (n) Among them, x (1) is the minimum value, x (n) is the maximum value. For different sample sizes n, the choice of statistics is different. When the order statistic x (i) When the data obey the normal distribution, Dixon gives the statistic R for different sample sizes n: ij The calculation formula of In a one-sided test, the statistic R ij The choices are as follows: When 3≤n≤7, (or ) When 8≤n≤10, (or ) When 11≤n≤13, (or ) When 14≤n≤30, (or ) Extending to the general case, Assuming the significance level is α, Dixon gives its critical value D 1-α (n); If the statistic R of a sample ij >D 1-α (n) or R i ' j >D 1-α (n), then judge x (n) or x (1) is an outlier; In a two-sided test, when R ij >R′ ij , Judge x (n) is an outlier; When R′ ij >R ij , Judge x (1) is an outlier.

5. The equipment effectiveness evaluation index screening method based on Bayesian network according to claim 3 is characterized in that: The steps of performing linear discrete quantization processing on each indicator set in the test evaluation original sample data set YS in step 3 are as follows: For index X, let Ω be the index space, K max is the maximum value of the indicator, K min is the minimum value, N is the number of states, then, for the indicator value x, the state it belongs to is S x :

6. The equipment effectiveness evaluation index screening method based on Bayesian network according to claim 5 is characterized in that: In step 4, the method for learning the Bayesian network structure based on domain knowledge and sample data is as follows: After the training data D is given, the posterior probability of the Baye network G is used as the score; It is easy to get from the Bayesian formula: In formula (2), P(D) will not affect different structures G, so it can be ignored. Ignoring this term and taking the logarithm, we can get: logP(G|D)∝logP(G,D)=logP(D|G)+logP(G) (3) The P(D|G) term in the above formula is called the marginal likelihood and is calculated as follows: P(D|G)=∫P(D|G,Θ)P(Θ|G)dΘ (4) Taking a uniform distribution for Θ, we can get the CH score: Taking the Dirichlet distribution, we can get the BDe score: In the scoring function, α ij· is the hyperparameter to be set for node i, and the hyperparameter settings are evenly distributed; m ijk For data set D that satisfies X i = k and Pa(X i ) = the number of samples of j, r i For node X i The number of state values; Use the BDe score, which requires a hyperparameter α ijk When the default structural prior information follows a uniform distribution, the value of α ijk Take a uniform distribution: a ijk =αP(X i =k,π(X i )=j|G)=α / (r i q i ) (7) Substituting formula (7) into the BDe score, the BDeu score can be obtained.

7. The equipment effectiveness evaluation index screening method based on Bayesian network according to claim 6 is characterized in that: In step 5, the method for performing parameter learning based on the Bayesian network structure model BN_S obtained in step 4 is as follows: In Bayesian parameter learning, the parameter Θ is regarded as a random variable, and the final value can be obtained by calculating the posterior probability of Θ; the calculation formula is as follows: p(Θ|D)=p(D|Θ)·p(Θ) / p(D) (8) The prior distribution P(Θ) in formula (8) follows the Dirichlet distribution: Then substitute into formula (10) to get the posterior probability of Θ: In formula (9) and (10) α1,...,α r It is a hyperparameter, which contains prior knowledge; Finally, we can get θ k Estimates:

8. The equipment effectiveness evaluation index screening method based on Bayesian network according to claim 7 is characterized in that: In step 6, the method of analyzing the sensitivity between each initial indicator and the performance using the constructed Bayesian network model BN_All is as follows: The calculation method of sensitivity index based on the degree of change of conditional probability, taking X1 as an example, Where: q——Number of states of Y r x1 ——X i The number of value states Through parameter learning, we can obtain P(X1), P(X i |π(X i )) and P(Y|X1,X2,…,X n ), where i = 2,…,n; from the conditional probability formula and direct inference algorithm, we know that: Substituting formulas (15) and (16) into formula (14) yields the contribution of index X1 to performance Y.

9. The equipment effectiveness evaluation index screening method based on Bayesian network according to claim 7 is characterized in that: In step 6, the method of analyzing the sensitivity between each initial indicator and the performance using the constructed Bayesian network model BN_All is as follows: The calculation method of sensitivity index based on the degree of state change, taking X1 as an example, ——The number of states of Y, ——X i The number of value states of ; Through parameter learning, the sizes of P(Y=a|X1=k) and P(Y=a|X1=b) can be obtained, and then the quantitative sensitivity index can be obtained through formula (17).

Citation Information

Cited By

  • Adversarial system efficiency evaluation method and device, computer equipment and storage medium

    CN119204755A

  • Method and device for evaluating signal intensity prediction model of maritime wireless communication

    CN119921881A

  • Simulation experiment efficiency evaluation index analysis method and device based on confidence coefficient

    CN120355310A