Human error probability assessment method based on integrated SF-DEMATEL, SF-SWARA and BN-SLIM

Abstract

This invention discloses a method for assessing the probability of human error based on the integration of SF-DEMATEL, SF-SWARA, and BN-SLIM. The method includes: structurally decomposing the target task, identifying performance-forming factors, and calculating expert weight factors; using the SF-DEMATEL method to process expert verbal evaluations, identifying the causal topology among performance-forming factors by calculating the total influence matrix and centrality and causality; using the SF-SWARA method to aggregate expert ratings of factor importance, determining the weight of each factor through defuzzification and comparison coefficient calculation; constructing a BN-SLIM model based on the causal topology and weights, using Bayesian network forward inference to predict the baseline probability of human error, calculating the average variance through backward diagnostic inference to identify key performance-forming factors, and calculating the ratio variance to identify key risk levels. This invention integrates spherical fuzzy sets, multi-criteria decision-making, and the bidirectional inference capabilities of Bayesian networks, providing a systematic analytical tool for quantifying human error risk in high-risk, small-sample environments.

Description

Technical Field

[0001] This invention belongs to the field of human reliability analysis technology, specifically involving a human error probability assessment method based on the integration of SF-DEMATEL, SF-SWARA and BN-SLIM, which is particularly suitable for the quantification and diagnosis of human error risk in complex operational scenarios such as extravehicular activities that are high-risk, data-scarce, and have strong coupling of multiple factors. Background Technology

[0002] Extravehicular activity (EVA) is one of the most risky and complex critical operations in manned space missions. During EVA maintenance, astronauts must perform high-precision operations in extreme environments such as microgravity, high vacuum, and strong radiation. Human error can lead to mission failure or even endanger the lives of astronauts. Therefore, a systematic and interpretable assessment of the risks of human error in EVA maintenance has become a core research issue in manned spaceflight safety engineering.

[0003] Among existing Human Reliability Analysis (HRA) methods, the Success Likelihood Index Methodology (SLIM) can comprehensively consider multiple Performance Shaping Factors (PSFs) and has good applicability under conditions of data scarcity. However, its core limitation lies in assuming that each PSF is independent, which contradicts the strong coupling reality in EVA tasks and makes it difficult to characterize the nonlinear relationships between PSFs and the risk transmission path.

[0004] In recent years, some studies have attempted to extend SLIM by introducing Bayesian networks (BN-SLIM) to leverage its causal modeling and bidirectional reasoning capabilities to compensate for the shortcomings of SLIM. However, existing BN-SLIM research still has significant limitations in extreme application scenarios such as EVA. First, existing methods struggle to effectively address the hesitant judgments commonly observed by experts in high-risk space missions, and the connection mechanism between fuzzy information and probabilistic reasoning remains unclear. Second, given the unique characteristics of EVA missions—strong coupling and multi-stage risk propagation—existing research lacks a systematic methodological framework that moves from fuzzy cognitive modeling to causal probabilistic reasoning.

[0005] Therefore, there is an urgent need for a method that can accurately handle expert multidimensional uncertainty judgments, explicitly characterize the complex causal dependencies between PSFs, and achieve human error risk assessment from qualitative analysis to quantitative probabilistic reasoning. Summary of the Invention

[0006] The purpose of this invention is to provide a method for assessing the probability of human error based on the integration of SF-DEMATEL, SF-SWARA, and BN-SLIM, in order to overcome the shortcomings of existing HRA methods in that they cannot characterize the nonlinear correlation of PSFs (Performance Shaping Factors) and are difficult to handle the uncertainty of expert cognition.

[0007] To achieve the above objectives, the present invention employs the following technical solution: Human error probability assessment methods based on the integration of SF-DEMATEL, SF-SWARA, and BN-SLIM include: Step 1, Assessment Foundation Construction: The target task is decomposed in a structured manner, performance formation factors that affect human error are identified, an assessment panel composed of multi-domain experts is established, the qualifications of the experts are scored, and the weight factor of each expert is calculated based on the scores. Step 2, Interaction Analysis of Performance Formation Factors: Using SF-DEMATEL, we processed the experts' verbal evaluations of the degree of influence among performance formation factors. By calculating the total influence matrix and centrality and causality, we identified the causal topology among performance formation factors and screened out the core causal edges. Step 3, Weighting of performance formation factors: Using SF-SWARA, we process the experts' verbal evaluations of the importance ranking of performance formation factors. By aggregating expert opinions, defuzzifying, and calculating comparison coefficients, we determine the relative importance weight of each performance formation factor under different tasks. Step 4, Probabilistic Modeling and Dynamic Evaluation: Based on the causal topology and relative importance weights, construct BN-SLIM, use forward inference through Bayesian networks to predict the probability of human error, perform backward diagnostic inference with the overall probability of human error as a certain event, measure the contribution of each performance formation factor to the overall probability of human error by average variation to identify key performance formation factors, and calculate the ratio variation of each rating level to identify key risk levels.

[0008] A terminal device includes a processor, a memory, and a computer program stored in the memory; when the processor executes the computer program, it implements the human error probability assessment method based on the integration of SF-DEMATEL, SF-SWARA, and BN-SLIM.

[0009] A computer-readable storage medium storing a computer program; when executed by a processor, the computer program implements the human error probability assessment method based on the integration of SF-DEMATEL, SF-SWARA, and BN-SLIM.

[0010] Compared with the prior art, the present invention has the following technical features: 1. This invention introduces spherical fuzzy sets, which simultaneously characterize membership, non-membership, and hesitation in the two core stages of PSFs weight determination and causal identification. Its three-dimensional expressive ability more realistically reflects the cognitive state of experts in high-risk, data-scarce tasks, effectively reducing subjective bias.

[0011] 2. This invention uses the spherical fuzzy DEMATEL method to quantify the mutual influence between PSFs and constructs an explicit causal network by combining task analysis, which makes up for the shortcomings of the traditional SLIM assumption that factors are independent and accurately reveals the risk propagation path.

[0012] 3. This invention organically integrates the spherical fuzzy multi-criteria decision-making method with the BN-SLIM framework, constructing a systematic method from expert fuzzy cognitive modeling to probabilistic risk quantification and posterior diagnosis, solving the problem of poor connection between fuzzy information and probabilistic reasoning in existing methods.

[0013] 4. This invention makes full use of expert knowledge, transforms qualitative judgments into computable causal structures through a structured approach, and combines the bidirectional reasoning capabilities of Bayesian networks to achieve accurate prediction of the probability of human error and effective diagnosis of key risk factors even when historical data is scarce. Attached Figure Description

[0014] Figure 1 This is a framework diagram of the method of the present invention; Figure 2 This is a visual representation of the spherical fuzzy set in this invention; Figure 3 This is a schematic diagram of the Bayesian network-success likelihood index method model structure in this invention; Figure 4 This is a flowchart of the complete extravehicular activity (EVA) maintenance mission in an embodiment of the present invention; Figure 5 This is an initial hierarchical diagram of performance formation factors in an embodiment of the present invention; Figure 6 This is a diagram of the performance formation factor system established in the embodiments of the present invention; Figure 7 This is a correlation analysis diagram of task and performance formation factors in an embodiment of the present invention; Figure 8 This is a causal relationship diagram of performance formation factors in an embodiment of the present invention; Figure 9 This is a Bayesian network structure diagram in an embodiment of the present invention; Figure 10 This is a visualization summary of the probability of human error in tasks in an embodiment of the present invention; Figure 11 This is a prior probability diagram of the probability of human error in an embodiment of the present invention; Figure 12 This is a graph showing the average variation of performance formation factors in an embodiment of the present invention; Figure 13 This is a ratio variation graph of key risk indicators in an embodiment of the present invention. Detailed Implementation

[0015] This invention provides a method based on the integration of SF-DEMATEL (Spherical Fuzzy Decision-Making Trial and Evaluation Laboratory), SF-SWARA (Spherical Fuzzy Step-wise Weight Assessment Ratio Analysis), and BN-SLIM (Bayesian Network - Success Likelihood Index Method). This invention presents a Bayesian network-success likelihood index method for assessing the probability of human error. It constructs a performance-forming factor identification system and task decomposition structure, utilizes the spherical fuzzy DEMATEL method to handle multidimensional uncertainty in expert evaluation, identifies the causal topological structure among performance-forming factors, and then combines the spherical fuzzy SWARA method to determine the relative importance weights of each factor. Finally, it constructs a BN-SLIM model to achieve probabilistic inference and posterior diagnosis of human error probability, thereby identifying key error factors and their key error states. This method, by integrating the bidirectional reasoning capabilities of spherical fuzzy sets, multi-criteria decision-making, and Bayesian networks, provides a systematic analytical tool for quantifying human error risk in high-risk, small-sample environments, improving the accuracy and interpretability of the assessment results. The specific steps of this invention are as follows: Step 1, assess the basic infrastructure.

[0016] The target task (such as astronaut extravehicular activity maintenance) is decomposed in a structured manner, performance factors that affect human error are identified, and an evaluation team composed of multi-domain experts is established to calculate the weight factor of each expert based on their qualifications.

[0017] (1) Construct an expert group and calculate weights.

[0018] Basic data was obtained using an expert-heuristic approach; to reduce individual subjective bias, differentiated weighting was applied to the experts participating in the evaluation. Expert qualifications were scored across three dimensions: space mission experience, ground simulation training experience, and education level, as shown in Table 1; the overall score for each expert was... It is obtained by summing up the scores of each item, where Number the expert; Weighting factors of experts The weighting factor is determined by the proportion of its score to the total score of all experts, as shown in Equation (1); this weighting factor will be applied to the subsequent aggregation of all expert opinions to ensure the consensus and reliability of the evaluation results.

[0019] (1); in, The total number of experts, The total weighted score for all experts.

[0020] Table 1 Expert Classification Parameters

[0021] (2) Task identification, analysis and decomposition.

[0022] The hierarchical task analysis method is used to decompose the target task in a structured manner. The granularity of the task decomposition is based on the standard that the probability of human error can be independently assessed. That is, each sub-task has a clear start and end state and corresponds to a specific set of performance formation factors.

[0023] Performance Shaping Factor (PSF) refers to the collective term for various factors that influence the probability of an operator successfully completing a mission. In high-risk space missions such as extravehicular activities (EVAs), PSF encompasses multi-dimensional factors affecting astronaut operational performance, such as: microgravity adaptation level, muscle fatigue and physical condition, concentration, situational awareness, the effects of microgravity / vacuum environment, lighting conditions, joint freedom of movement, tool usability, clarity of intravehicular and extravehicular communication, communication latency and stability, clarity of EVA operation procedures, realism of simulation training, complexity of mission operation procedures, and mission time pressure.

[0024] This plan includes... There are several performance-forming factors, each of which is rated in step four to reflect its degree of advantage in the current task context. These performance-forming factors are not independent of each other, but rather have complex causal relationships, which are quantified in step two using the spherical fuzzy DEMATEL method.

[0025] (3) Identification of performance formation factors.

[0026] By combining a systematic literature review with the Delphi method, we integrated the performance formation factor classification system in mainstream human reliability analysis methods. After multiple rounds of anonymous review and screening, we established a performance formation factor system applicable to the target work scenario.

[0027] Step 2, interaction analysis of performance formation factors.

[0028] The spherical fuzzy decision laboratory method SF-DEMATEL is used to process experts' verbal evaluations of the degree of influence among performance-forming factors. By calculating the total influence matrix and centrality and causality, the causal topology among performance-forming factors is identified, and the core causal edges are screened out.

[0029] (1) Collect language terminology questionnaires.

[0030] Experts use a predefined spherical fuzzy language evaluation set to assess the degree of direct influence between any two performance-forming factors, with each evaluation represented as a spherical fuzzy number. ;in Membership degree represents the degree of support experts have for the proposition that "there is an influence relationship among performance-forming factors"; The degree of non-membership indicates the degree of opposition from experts to the proposition; The degree of hesitation represents the level of uncertainty among experts regarding the proposition; the above three components satisfy the following constraints: ; in For spherical fuzzy numbers In elements Membership value on The non-membership value, The degree of hesitation. Elements in a set of language terms.

[0031] (2) Calculate the average matrix.

[0032] The evaluations of all experts are aggregated using a spherically weighted arithmetic mean, with the following settings: n Evaluation indicators and K An expert, an expert K Evaluation indicators right The degree of influence is the spherical fuzzy number The spherical fuzzy average matrix is ​​obtained. The aggregation formula is shown in equation (2): (2); in: This represents the Spherical Weighted Arithmetic Mean operator. It is based on Each expert expressed performance formation factors Factors influencing performance formation A spherical fuzzy number vector of degree; Indicates the first Experts on performance formation factors Factors influencing performance formation A spherical fuzzy evaluation of the degree of spherical shape. ; The total number of performance formation factors (i.e., criteria). ; The specific calculation formula is as follows: ; in, For the first The weighting factor of each expert is calculated by formula (1); For the first Membership component evaluated by experts; For the first The non-membership degree component evaluated by experts; For the first The degree of hesitation is evaluated by several experts.

[0033] (3) Calculate the normalized initial direct relation matrix: The average matrix... Decomposed into a membership matrix Non-membership matrix and hesitation matrix (extract separately) The matrix contains membership, non-membership, and hesitation, which together form these three matrices, which are then normalized.

[0034] Among them, the membership matrix Represented as: (3); in, Membership matrix The Middle Line number Column elements, , This represents the total number of performance-forming factors.

[0035] membership matrix Normalization is performed as follows:

[0036] in, Normalized membership matrix The Middle Line number Column elements; This represents the maximum sum of all rows in the membership matrix; This represents the maximum sum of all columns in the membership matrix.

[0037] Similarly, the normalized non-membership matrix can be obtained. and hesitation matrix .

[0038] (4) Determine the overall influence matrix .

[0039] First, analyze the normalized membership matrix. Non-membership matrix and hesitation matrix Solve for the corresponding total influence submatrix , and Then, the three sub-matrices are aggregated into a complete spherical fuzzy total influence matrix. .

[0040] Among them, the normalized membership matrix Solve for the corresponding total influence submatrix The process is as follows: (4); in I Represents the identity matrix; The variable is a power variable; the formula expands the direct relationship into a total influence relationship that includes all indirect paths by summing the matrix power series.

[0041] Similarly, they can be used separately. and Obtain the corresponding total influence submatrix and .

[0042] Finally, the three total influence sub-matrices will be analyzed. , and Aggregated into a total influence matrix :

[0043] in, , , These are the total influence matrix The Middle Line number The membership, non-membership, and hesitation components of column elements. .

[0044] (5) Calculate the sum of the rows and columns of the total influence matrix.

[0045] Row sum of the total influence matrix and column sum They are respectively: (5); in, The addition operator for spherical fuzzy numbers. Total number of performance-forming factors; rows and Indicating performance formation factors The extent of the combined impact on all other performance-forming factors; columns and Indicating performance formation factors The degree to which it is affected by the combined effects of all other performance-forming factors. The specific calculation formula is as follows: remember , and They are rows and The membership components, non-membership components, and hesitation components are calculated as follows: (6); in , Indicates the number of performance-forming factors; , , These are the total influence matrix The Middle Line number Membership, non-membership, and hesitation components of column elements; and For different The column index variable is used to calculate the interaction terms between different columns; The symbol for multiplication; Indicates the exclusion Summing all columns except those in the previous column; Indicates the exclusion and Sum all columns except those in the column.

[0046] Similarly, we can obtain the sum of the series. Membership components, non-membership components, and hesitation components ( The calculation results.

[0047] (6) Deblurring.

[0048] Based on rows and , column and The calculation results are used to represent the performance formation factors using spherical fuzzy numbers (SF). The extent to which other performance-forming factors influence performance-forming factors The degree to which it is influenced by other performance-forming factors is specified as follows: and ;in It is the overall influence matrix The sum of the values ​​in each row represents the performance-forming factors. iThe degree of influence; It is the overall influence matrix The sum of each column represents the performance-forming factors. j The degree to which it is affected.

[0049] (7); (8); in The score function is the spherical fuzzy number. , , They are rows and The membership degree, non-membership degree, and hesitation degree components; , , Columns and The membership degree, non-membership degree, and hesitation degree components; and They are and The deblurring score.

[0050] (7) Calculate centrality and causality.

[0051] Centrality Factors reflecting performance formation i Importance within the system; causality Factors reflecting performance formation i The impact on other performance-forming factors. If >0, performance formation factor i It is a causal factor, meaning that this performance-forming factor actively influences other performance-forming factors; if If the value is less than 0, the performance formation factor is an outcome factor, meaning that the performance formation factor is mainly influenced by other performance formation factors.

[0052] (9); (10); (8) Constructing causal networks and screening core causal edges.

[0053] Based on centrality and causal degree Construct a causal network among performance-forming factors; the nodes of this causal network are the performance-forming factors themselves, and directed edges represent the causal relationships between these factors, with arrows pointing from causal factors to outcome factors; causal degree... The performance-forming factors are parent nodes (actively influencing other factors), and the degree of causation... The performance-forming factors are sub-nodes (affected by other factors).

[0054] To ensure that edges in a causal network are statistically significant, a strength threshold is set. The process involves filtering, retaining only edges whose influence intensity exceeds a threshold. The intensity threshold is calculated using the following formula: (11); in, Total Influence Matrix The Middle Line number The defuzzy element values ​​of the column represent the performance forming factors. Factors contributing to performance The overall intensity of the impact; This represents the total number of performance-forming factors; this threshold represents the average of the total influence among all performance-forming factors, used to determine whether the influence relationship is significant.

[0055] The filtering rules are as follows: (12); in, Indicating performance formation factors Factors contributing to performance There is a significant influence relationship, so we retain the following: point to The directed edges are used as the core causal edges; If no significant influence relationship exists, the causal edge is removed. Simultaneously, a strength-inclusion dual-screening strategy is employed to ensure that the task occurrence set of the child node is a subset of the task occurrence set of the parent node; that is, all tasks involved in the child node must also be involved in the parent node, thus guaranteeing the logical consistency of the causal relationship.

[0056] Through the above steps, a causal topological structure of the Bayesian network is formed. This structure explicitly characterizes the causal dependencies between various performance-forming factors, providing a structural foundation for constructing the Bayesian network-success likelihood index model in step 4.

[0057] Step 3: Weight allocation of performance formation factors.

[0058] The spherical fuzzy stepwise weighted ratio analysis method (SF-SWARA) was used to process the linguistic evaluations of experts' ranking of the importance of performance formation factors. By aggregating expert opinions, defuzzifying, and calculating comparison coefficients, the relative importance weights of each performance formation factor under different tasks were determined.

[0059] (1) Construct the decision matrix.

[0060] Experts used predefined spherical fuzzy linguistic variables to assign importance levels to each performance-forming factor.

[0061] (2) Integrate expert opinions.

[0062] The spherical weighted arithmetic mean is used to aggregate the spherical fuzzy language ratings of the importance of each performance factor by all experts, as shown in Equation (13); (13); in, Represents the spherical weighted arithmetic mean operator; It is all the experts' opinion on the first A spherical fuzzy number set for the importance rating of each performance formation factor; For the 1st to the 1st The expert commented on the first The spherical fuzzy number of the importance rating of each performance factor is aggregated to obtain the i-th result. A comprehensive spherical fuzzy importance rating of each performance-forming factor. The expansion of the operator is defined in formula (2).

[0063] Let the aggregation result of this step be the spherical fuzzy number. .

[0064] (3) Deblurring.

[0065] The aggregated spherical blur value is converted into a fractional value using the following formula: (14); in, The spherical fuzzy number after aggregation Membership components; The spherical fuzzy number after aggregation The non-membership component; The spherical fuzzy number after aggregation The degree of hesitation component; For the first The defuzzy score of each performance factor indicates that the higher the score, the more important the performance factor.

[0066] (4) Ranking of importance.

[0067] Performance formation factors are ranked by score. The factors are arranged in descending order to determine the relative importance of each performance-forming factor.

[0068] (5) Determine the relative importance .

[0069] For each performance formation factor Calculate its score Compared with the previous performance formation factor The ratio of score values ​​is used as a measure of comparative importance. .

[0070] (6) Calculate each performance formation factor Comparison coefficient .

[0071] (15); in, The comparison coefficient of the performance formation factor that ranks first (most important) ; Time Comparison Coefficient .

[0072] (7) Calculate each performance formation factor Corrected weights .

[0073] (16); in, For the first The adjusted weights of each performance factor.

[0074] (8) Calculate each performance formation factor relative importance weights .

[0075] (17); in, This represents the total number of performance-forming factors.

[0076] Step 4, Probabilistic Modeling and Dynamic Evaluation: Based on the causal topology among the performance formation factors obtained in Step 2 and the relative importance weight of each performance formation factor obtained in Step 3, a Bayesian Network-Success Likelihood Index Model (BN-SLIM) is constructed. The forward inference of the Bayesian network is used to predict the probability of human error, and the backward diagnosis is used to identify key risk performance formation factors and their key risk levels.

[0077] Batch Normalization-Syllable Modeling (BN-SLIM) is used for probabilistic modeling and dynamic evaluation. The model construction process comprises two parts: SLIM and BN. The specific steps are as follows: (1) Determine the probability and probability distribution of performance formation factors.

[0078] Rating levels of performance formation factors For integers between 1 and 9, reflecting the first... The degree of advantage of each performance-forming factor in the current task context, among which Indicates the most unfavorable conditions, This represents the optimal condition. The probability distribution of each score is determined based on existing data and expert experience. , This represents the total number of performance-forming factors.

[0079] (2) Calculate the success likelihood index SLI.

[0080] (18); in Indicates the first The relative importance weights of each performance formation factor are calculated using formula (17); For the first The scores of performance formation factors; SLI is the success likelihood index, and a higher value indicates that the current task situation is more conducive to successful execution.

[0081] (3) Estimate the probability of human error (HEP) when performing the task.

[0082] A linear mapping relationship between the success likelihood index (SLI) and the human error probability (HEP) is established through two anchor tasks: (19); Where, constant and Solve the problem using the known HEP ​​anchor point task. For the slope parameter, This is the intercept parameter. Anchor tasks should be selected from typical high-risk and low-risk tasks, and their HEP values ​​can be set based on domain experience or expert judgment.

[0083] (4) Construct the BN structure.

[0084] In a Bayesian network structure, performance formation factors, SLI (Self-Limited Instructions), and HEP (Human Error Probability) are considered as three types of nodes. Each task corresponds to a set of performance formation factor nodes as parent nodes that directly point to the SLI node for that task. The SLI node serves as the sole parent node for the HEP node of that task. The HEP nodes of all tasks collectively point to the overall HEP node. The overall HEP node represents the total probability of human error occurring in the entire evaluated task. In high-risk tasks such as extravehicular activities, a task can typically be decomposed into multiple sub-tasks (such as equipment disassembly, maintenance operations, and equipment installation), each with its own HEP ​​node. The overall HEP node summarizes the probability of human error from all sub-tasks, comprehensively reflecting the final risk level of the entire task.

[0085] The directed edge connections between nodes are determined by the causal topology constructed in step (8) of step 2; when At that time, a Bayesian network was established based on performance-forming factors. Performance formation factors The directed edge.

[0086] (5) Assigning values ​​to BN parameters.

[0087] For the performance formation factor that serves as the root node, its probability distribution is determined by combining historical data with expert experience; for non-independent performance formation factor nodes, their conditional probability tables are calculated based on the rating distribution of their parent nodes through logical relationships; the conditional probability tables between SLI nodes and their corresponding HEP nodes are established based on the linear mapping relationship of Equation (19) using the anchor task calibration method; the conditional probability tables of the overall HEP nodes are calculated using the Noisy-OR method, as shown in Equation (20). (20); in, Represents conditional probability; For events only When the target event occurs while all other events do not occur, The conditional probability of occurrence was obtained through expert opinion; For the set of parent nodes When all events occur, the target event The conditional probability of occurrence; Indicates the first The occurrence of events corresponding to each concurrent parent node; Indicates the first The events corresponding to the concurrent parent nodes will not occur; Represents a set of concurrent parent nodes; For the target node event; This indicates the number of parent nodes pointing to the overall HEP node.

[0088] (6) Identify key performance formation factors.

[0089] First, forward inference is performed using a parameterized Bayesian network. That is, without any evidence input, the initial probability distribution of each performance-forming factor node is calculated as the prior probability. This process can simultaneously predict the probability of human error under the baseline state; subsequently, backward diagnostic reasoning is performed with the overall probability of human error (overall HEP) as a certain event, updating the probability distribution of each performance-forming factor node to obtain the posterior probability. The contribution of each performance factor to the overall probability of human error is measured by the mean variation. The formula for calculating the mean variation is shown in equation (21). (twenty one); in, For the first The performance formation factor in the first The rating level for each probability value, for example Indicates the first The current performance formation factor is at rating level 5; For the first The performance formation factor rating level is: The prior probability is derived from forward inference using a Bayesian network; For the first The performance formation factor rating level is: The posterior probability is the conditional probability that the performance factor is at that rating level given the probability of overall human error occurring. The probability is the number of times it takes a value; For the first The average variation of each performance factor indicates that the larger the value, the greater the contribution of that performance factor to the overall probability of human error.

[0090] The average variance of all performance-forming factors was calculated. Then, sort them in descending order of value to obtain the sequence. ; Subsequently, calculation before k Cumulative contribution rate of each factor : , k =1, 2, ..., n (twenty two); in, Indicates the order before k Performance formation factors The sum of values; Indicates all n Performance formation factors MV The sum of values; n This represents the total number of performance-forming factors. k This is a temporary variable, indicating that the top rankings are considered. k Performance-forming factors of a position.

[0091] Set a cumulative contribution rate threshold The number of key performance formation factors K Defined as the smallest integer that causes the cumulative contribution rate to first reach or exceed the threshold. k ,Right now: (twenty three); in, This is a pre-set cumulative contribution rate threshold (e.g., 70%–80%) based on engineering experience. It indicates that when the cumulative contribution rate reaches or exceeds this value, the previous... K Several factors can be considered as major sources of risk. Ultimately, the top [factors] after ranking... KPerformance formation factors These are the identified key performance indicators (KPIs). All identified KPIs will be used as inputs for step (7) to conduct the next step of risk level analysis.

[0092] (7) Identify key risk levels.

[0093] For the key performance formation factors identified in step (6), the ratio variation of each rating level is further calculated. The formula for calculating the ratio variation is shown in equation (24).

[0094] (twenty four); in, This indicates that the level is a high-risk state, meaning that the performance formation factor is at the rating level. At that time, the overall probability of human error increases significantly.

[0095] After predicting the probability of human error and identifying key performance formation factors and their high-risk levels, the assessment results can be applied to the following aspects: First, risk warning; for performance formation factors at high risk levels, take early intervention measures to reduce the probability of human error; Second, training optimization; based on the distribution characteristics of key performance formation factors, strengthen targeted training for operators' weak links; Third, task redesign; optimize processes or add protective measures for task links with excessively high risks.

[0096] In summary, this invention introduces spherical fuzzy sets to simultaneously characterize membership, non-membership, and hesitation in the two core stages of performance formation factor weight determination and causal identification. Its three-dimensional representation more realistically reflects the cognitive state of experts in high-risk, data-scarce tasks, effectively reducing subjective bias. Through the spherical fuzzy DEMATEL method, the mutual influence between performance formation factors is quantified, and an explicit causal network is constructed in conjunction with task analysis, overcoming the shortcomings of traditional SLIM's assumption of independent factors and accurately revealing risk propagation paths. By organically integrating the spherical fuzzy multi-criteria decision-making method with the BN-SLIM framework, a systematic approach is constructed from expert fuzzy cognitive modeling to probabilistic risk quantification and posterior diagnosis, solving the problem of poor connection between fuzzy information and probabilistic reasoning in existing methods. This invention fully utilizes expert knowledge, transforming qualitative judgments into computable causal structures through a structured method, and combining the bidirectional reasoning capabilities of Bayesian networks, enabling accurate prediction of human error probability and effective diagnosis of key risk factors even under conditions of scarce historical data.

[0097] Example: The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment takes an astronaut extravehicular maintenance mission as an example to illustrate the specific implementation process of the present invention.

[0098] I. Assessment of the basic infrastructure.

[0099] 1.1 Expert Panel Construction and Weight Calculation.

[0100] This embodiment invited 10 aerospace experts to participate in the evaluation. The experts' qualifications were scored based on three dimensions: space mission experience, ground simulation training experience, and education level (scoring criteria are shown in Table 1). Each expert's overall score was obtained by summing the scores of each dimension. k The weight factor of each expert is determined by the proportion of their score to the total score of all experts, and the calculation formula is shown in Equation (1). The results of the expert weight calculation are shown in Table 2.

[0101] Table 2. Weighting of Consulting Experts

[0102] 1.2 Task identification and analysis.

[0103] A hierarchical task analysis method was used to structurally decompose the extravehicular activity (EVA) maintenance operation for astronauts. The EVA maintenance process was broken down into main task phases and their sub-tasks, identifying a total of 3 phases, 9 main tasks, and 21 sub-tasks. The complete task flow is as follows: Figure 4 As shown.

[0104] 1.3 Identification of performance formation factors.

[0105] Through a systematic literature review, seven HRA methods and five classic studies were analyzed, resulting in the extraction of 331 initial PSF factors. These factors were then merged according to the principle of non-redundancy, forming an initial system encompassing six dimensions and 42 sub-factors. Figure 5 The Delphi method was used to organize an expert panel for two rounds of anonymous review and screening. Factors with an average score below 3.0 and a coefficient of variation greater than 0.25 were eliminated based on relevance principles, ultimately establishing a PSF system comprising 6 categories and 14 specific factors. Figure 6 The correlation analysis between each task and PSF is as follows: Figure 7 As shown.

[0106] II. Interaction analysis of performance formation factors.

[0107] 2.1 Calculation process of spherical fuzzy DEMATEL.

[0108] Step 2.1.1: Experts use the spherical fuzzy language evaluation set shown in Table 3 to assess the degree of direct influence between any two PSFs.

[0109] Table 3. Evaluation Set of DEMATEL Language for Spherical Fuzzy

[0110] Step 2.1.2: Use SWAM to aggregate the evaluations of all experts to obtain the spherical fuzzy mean matrix, as shown in Table 4.

[0111] Table 4 Average Matrix

[0112]

[0113] Step 2.1.3: Decompose the average matrix into membership degree, non-membership degree and hesitation degree submatrices, normalize them respectively and calculate the total influence submatrix, as shown in Table 5.

[0114] Table 5 Normalized Direct Relationship Matrix

[0115]

[0116] Step 2.1.4: Calculate the total influence matrix of the spherical fuzzy matrix. The results are shown in Table 6.

[0117] Table 6 Overall Influence Matrix

[0118]

[0119] Step 2.2.5: Calculate the sum of the rows and columns of the total influence matrix. The results are shown in Table 7.

[0120] Table 7 Matrix The sum of the actions sum of columns

[0121]

[0122] Step 2.2.6: After defuzzification, calculate the influence, affectedness, centrality, and causation degree of each PSF. The results are shown in Table 8. Factors with a causation degree greater than 0 are causal factors, and those with a causation degree less than 0 are outcome factors. The causal relationship is as follows: Figure 8 As shown.

[0123] Table 8. Deblurred and and centrality and causality

[0124]

[0125] Step 2.2.7: Set the intensity threshold and combine the intensity-inclusion dual screening strategy to extract 7 core causal edges (Table 9) to form the causal structure of the Bayesian network.

[0126] Table 9. Core causal edges of BN retained by double screening

[0127] III. Weighting of performance formation factors.

[0128] The weights of each PSF in the nine tasks were determined using the Spherical Fuzzy SWARA method.

[0129] 3.1 Spherical fuzzy SWARA calculation process.

[0130] In step 3.1.1, the experts assigned importance levels to each PSF using the spherical fuzzy linguistic variables shown in Table 10.

[0131] Table 10 SWARA Language Evaluation Set for Spherical Fuzzy

[0132] Step 3.1.2: Use SWAM to aggregate expert opinions, as shown in Equation (13). The results are shown in Table 11.

[0133] Step 3.1.3: Deblurring is performed using equation (14) to obtain the sharpness score. The results are shown in Table 11.

[0134] Table 11 Integration and De-ambiguation of Expert Opinions

[0135] Step 3.1.4: Sort the factors in descending order of their scores.

[0136] Step 3.1.5: Calculate the comparison coefficient, as shown in equation (15). The results are shown in Table 12.

[0137] Step 3.1.6: Calculate the corrected weights, as shown in equation (16). The results are shown in Table 12.

[0138] Step 3.1.7: Calculate the final normalized weights, as shown in equation (17). The results are shown in Table 12.

[0139] Table 12 PSF Factor Weights

[0140]

[0141] 3.2 Weight allocation results.

[0142] Based on the causal edges identified in step 2, the local weights of each parent node to its child nodes are calculated, and the results are shown in Table 13.

[0143] Table 13 Local Weights of PSF Parent Nodes Based on Causality

[0144] IV. Probabilistic Modeling and Dynamic Evaluation 4.1 Determining SLIM parameters.

[0145] By combining existing data and expert experience, rating levels (1-9 points) and their probability distributions are defined for each PSF. The probability distribution of the root node PSF is shown in Table 14. The probability distributions of non-independent PSFs (PSF2, PSF4, PSF7, PSF14) depend on their parent nodes and are obtained through weighted sum mapping.

[0146] Table 14 Probability and probability distribution of PSF

[0147] The weights and probability distributions of each PSF were input into the GeNIe software to construct a BN-SLIM model, and the SLI values ​​for each task were calculated. A boundary task calibration method was used, selecting the best and worst operating conditions as anchor points, and the HEP was set to 10. -4 With 10 -3 Substituting into the logarithmic linear mapping equation (19), the calibration coefficients are obtained. a =-3.58, b = 25.085.

[0148] 4.2 Bayesian Network Construction and Inference.

[0149] Based on the aforementioned causal structure and weights, a Bayesian network is constructed, such as... Figure 9 As shown in Table 14, the probability distribution of PSF nodes is based on the data from Table 14. The HEP of each task is calculated using forward inference through a Bayesian network, and the results are as follows: Figure 10 and Figure 11 As shown.

[0150] The calculation results show that T4 (extension rod insertion and interface docking) has the highest HEP (4.2%), followed by T8 (EPA docking and locking) (4.1%), and T1 (cabin tooling and assembly preparation) has the lowest (1.9%). The overall HEP is 14%.

[0151] 4.3 Identification of key risk factors.

[0152] Diagnostic reasoning was performed with the overall HEP as a certain event, and the prior and posterior means of each PSF were calculated. The results are shown in Table 15.

[0153] Table 15 Prior and Posterior Means of PSF

[0154] The mean variation (MV) of each PSF is calculated using the formula shown in equation (21), and the results are as follows. Figure 12 As shown. Then the cumulative contribution rate of the MV values ​​of all PSFs was calculated and filtered, as shown in formulas (22) and (23). The threshold was set to 70% based on the Pareto principle and expert experience. The calculation results showed that the top 4 PSFs after sorting were listed as key PSFs, namely: PSF5 (microgravity / vacuum environment influence), PSF2 (muscle fatigue and physical condition), PSF7 (joint mobility freedom) and PSF14 (task time pressure).

[0155] Further calculations were performed on the ratio variation (RV) of each rating level of the key PSF mentioned above, as shown in Equation (22). The results are as follows: Figure 13 As shown, the RV values ​​of each PSF in the worst-case state (R1) are all much greater than 1, indicating that these levels are high-risk states.

[0156] V. Method Comparison.

[0157] This section compares the proposed framework with two representative BN-SLIM methods. The first is the basic BN-SLIM framework proposed by Abrishami et al., referred to as the baseline method. The second is the STPA-IDAC extension method proposed by Liao et al. Table 16 summarizes the main differences between these three methods in seven key aspects, including handling of expert opinion uncertainty, expert weighting, PSF causal relationship identification, causal network construction, weight determination method, probabilistic inference, and output interpretability. The following section will elaborate on the theoretical basis and impact of these differences from three representative dimensions: handling of expert judgment uncertainty, causal dependency modeling, and output analysis.

[0158] (1) Handling of Uncertainty in Expert Judgment. In the baseline method, PSF weights are directly determined by expert scores without any processing of cognitive ambiguity. Expert opinions are aggregated through a simple weighted average. This method implicitly assumes that experts have complete confidence in and are consistent with their judgments. In high-risk EVA tasks with incomplete knowledge, potentially divergent opinions, and scarce data, this assumption rarely holds. The STPA-IDAC extension improves upon this by employing a traditional fuzzy linguistic aggregation method that captures membership information. However, it still ignores two other key dimensions in expert judgment: non-membership (the degree to which experts are skeptical of a statement) and hesitation (the degree of uncertainty or indecisiveness). By discarding these two dimensions, traditional fuzzy aggregation methods inevitably lose information crucial for risk assessment.

[0159] To overcome these limitations, the proposed framework employs a preprocessing method based on spherical fuzziness. Unlike traditional fuzzy sets, spherical fuzzy sets retain membership, non-membership, and hesitation values ​​simultaneously while satisfying the sum-of-squares constraint, thus fully preserving the richness of expert judgment. Furthermore, by incorporating expert qualification weighting factors based on experience and educational background, the spherical weighted arithmetic mean (SWAM) is used to aggregate multi-expert opinions. This ensures that more credible experts have a greater impact on the final result, while explicitly reflecting the hesitation dimension throughout the weight determination and causal identification stages. This feature is not present in previous methods.

[0160] (2) Causal Dependency Modeling. Baseline methods primarily rely on expert subjective experience to construct Bayesian Network (BN) structures, introducing potential biases and a lack of transparency. The STPA-IDAC extension identifies the mapping between Unsafe Control Actions (UCAs) and PSFs through STPA, thereby establishing causal relationships in system theory analysis and enhancing the rigor of the structure. However, this method remains qualitative. It can only determine whether a causal relationship exists, but cannot quantify the strength or extent of the interaction between PSFs. This limitation hinders the ability to prioritize risk factors or eliminate weak relationships when constructing a simplified BN.

[0161] To address this gap, this paper proposes the SF-DEMATEL method. This method constructs a direct influence matrix based on expert evaluation under spherical fuzzy uncertainty, and then calculates the total influence matrix. From this matrix, both the influence (centrality) of each PSF on other factors and the degree to which it is influenced by other factors (causality) can be derived. The strength of each causal edge is explicitly quantified, and negligible influences are filtered out using a threshold. An innovative "strength-inclusion" dual filtering strategy further ensures that only edges with causal significance and logical consistency are retained in the final BN structure. Therefore, the proposed framework not only identifies causal relationships but also provides a clear quantitative basis for constructing sparse and interpretable BNs. This advantage is neither provided by baseline methods nor fully realized by the STPA-IDAC extension method.

[0162] (3) Analysis Output. The baseline method simultaneously outputs HEP values, mean variation (MV), and ratio variation (RV) analysis, enabling diagnostic identification of critical PSFs and their critical states. The STPA-IDAC extension focuses on HEP ​​prediction and MV analysis but does not include RV analysis. The proposed framework provides complete HEP prediction, MV analysis, and RV analysis capabilities, and additionally reveals explicit risk propagation paths between PSFs. This richer output provides more actionable evidence for safety intervention measures.

[0163] In summary, this framework complements existing BN-SLIM methods in terms of expert judgment handling and causal dependency modeling. By preserving complete three-dimensional information of expert judgment, quantifying the strength of causal influence, and providing comprehensive diagnostic risk analysis including RV, it offers a more systematic, accurate, and easily interpretable analytical tool for quantifying human error risks in high-risk, small-sample mission scenarios such as extravehicular activities.

[0164] Table 16 Comparison of Methods

[0165] VI. Conclusion.

[0166] The evaluation framework proposed in this invention, which integrates SF-DEMATEL, SF-SWARA, and BN-SLIM, can effectively quantify the probability of human error in EVA tasks, identify key risk factors and their risk levels, and reveal the propagation path of risk in the PSF network. It provides a systematic quantitative tool for assessing human error risk in high-risk, small-sample environments.

[0167] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A human error probability assessment method based on the integration of SF-DEMATEL, SF-SWARA, and BN-SLIM, characterized in that, Includes the following steps: Step 1, Assessment Foundation Construction: The target task is decomposed in a structured manner, performance formation factors that affect human error are identified, an assessment panel composed of multi-domain experts is established, the qualifications of the experts are scored, and the weight factor of each expert is calculated based on the scores. Step 2, Interaction Analysis of Performance Formation Factors: Using SF-DEMATEL, we processed the experts' verbal evaluations of the degree of influence among performance formation factors. By calculating the total influence matrix and centrality and causality, we identified the causal topology among performance formation factors and screened out the core causal edges. Step 3, Weighting of performance formation factors: Using SF-SWARA, we process the experts' verbal evaluations of the importance ranking of performance formation factors. By aggregating expert opinions, defuzzifying, and calculating comparison coefficients, we determine the relative importance weight of each performance formation factor under different tasks. Step 4, Probabilistic Modeling and Dynamic Evaluation: Based on the causal topology and relative importance weights, construct BN-SLIM, use forward inference through Bayesian networks to predict the probability of human error, perform backward diagnostic inference with the overall probability of human error as a certain event, measure the contribution of each performance formation factor to the overall probability of human error by average variation to identify key performance formation factors, and calculate the ratio variation of each rating level to identify key risk levels.

2. The method for assessing the probability of human error according to claim 1, characterized in that, The qualifications of experts are scored from three dimensions: experience in space missions, experience in ground simulation training, and level of education. The weighting factor of each expert is determined by the proportion of their comprehensive score to the total score of all experts.

3. The method for assessing the probability of human error according to claim 1, characterized in that, The hierarchical task analysis method is used to decompose the target task into a structured form. Each subtask has a clear start and end state and corresponds to a specific set of performance formation factors. A performance formation factor system suitable for the target task context is established.

4. The method for assessing the probability of human error according to claim 1, characterized in that, Step 2 specifically includes: (1) Experts use a predefined spherical fuzzy language evaluation set to assess the degree of direct influence between any two performance-forming factors. Each evaluation is represented as a spherical fuzzy number. ,in For membership degree, Non-membership degree, For hesitation level, and satisfying ,in For spherical fuzzy numbers In elements Membership value on The non-membership value, The degree of hesitation. Elements in a set of language terms; (2) The evaluations of all experts are aggregated using the spherical weighted arithmetic mean operator (SWAM), and the following is set: Individual performance formation factors and An expert, an expert Assessment of performance formation factors right The degree of influence is the spherical fuzzy number The spherical fuzzy average matrix is ​​obtained. , its first Line number Column elements are ,in ; Represents the spherical weighted arithmetic mean operator; (3) Average matrix Decomposed into a membership matrix Non-membership matrix and hesitation matrix Normalization was performed separately to obtain , and ; (4) Determine the overall influence matrix : respectively , and Solve for the corresponding total influence submatrix , and ,in , The first matrix is ​​the identity matrix, and the three sub-matrices are then aggregated into the total influence matrix. ,in , , These are the total influence matrix The Middle Line number Membership, non-membership, and hesitation components of column elements; (5) rows and Performance formation factors The overall impact, listed and Performance formation factors The overall degree of impact, among which Membership components Non-membership components and hesitation component The components of each element in the total influence matrix are obtained through spherical fuzzy number addition. Similarly, we can obtain... Each component; (6) Through the scoring function and The deblurring score is obtained, where The scoring function for spherical fuzzy numbers. and Performance formation factors Influence and performance formation factors The degree of influence; (7) Calculate centrality and causal degree ,like Performance formation factors As a causal factor, if This is the outcome factor; (8) Set the intensity threshold ,in Total Influence Matrix The Middle Line number The deblurred element values ​​of the column, with the filtering rule being when When the directed edge is retained as the core causal edge, Causal edges are removed in time to form the causal topology of the Bayesian network.

5. The method for assessing the probability of human error according to claim 1, characterized in that, Step 3 specifically includes: (1) Experts used predefined spherical fuzzy linguistic variables to assign importance levels to each performance-forming factor; (2) The spherical weighted arithmetic mean operator (SWAM) is used to aggregate the spherical fuzzy language ratings of the importance of each performance factor from all experts, and the aggregation results are obtained. ,in For membership degree components, For non-membership components, For hesitation component; (3) Through The aggregated spherical fuzzy values ​​are converted into fractional values, where For the first Defuzzified scores of each performance-forming factor; (4) Determine the relative importance order of each performance factor by arranging them in descending order of their scores; (5) For each performance formation factor The ratio of its score to the score of the previous performance factor is used as a measure of its importance. ; (6) Calculate the comparison coefficient :when hour ,when hour ; (7) Calculate the corrected weights :when hour ,when hour ,in For the first Adjusted weights for each performance-forming factor; (8) Calculate the relative importance weights ,in This represents the total number of performance-forming factors.

6. The method for assessing the probability of human error according to claim 1, characterized in that, Step 4, which involves constructing the BN-SLIM model, includes: (1) Determine the rating levels of performance formation factors ; (2) Calculate the success likelihood index ,in Performance formation factors The relative importance weights, For the first The rating levels of each performance formation factor; (3) Establish a linear mapping relationship between SLI and the probability of human error (HEP) through two anchor tasks. ,constant and Solve the problem using the known HEP ​​anchor point task. For the slope parameter, For intercept parameters; (4) Construct a Bayesian network structure: with performance formation factors, SLI and HEP as three types of nodes, each task corresponds to a set of performance formation factor nodes as parent nodes that directly point to the SLI node of the task, the SLI node is the only parent node of the HEP node of the task, and the HEP nodes of each task point to the overall HEP node. (5) The probability distribution of the root node is determined by historical data and expert experience. The conditional probability table of the non-independent performance forming factor node is calculated by logical relationship based on the score distribution of its parent node. The conditional probability table between the SLI node and the HEP node is established by anchor task calibration method based on the linear mapping relationship. The conditional probability table of the overall HEP node is calculated by Noisy-OR method.

7. The method for assessing the probability of human error according to claim 6, characterized in that, Step 4, which involves identifying key performance formation factors, specifically involves using a parameterized Bayesian network for forward inference to calculate the prior probability of each performance formation factor node. ,in For the first The performance formation factor in the first The rating level for each probability value; By performing backward diagnostic reasoning with the overall probability of human error as a certain event, the posterior probability is obtained. ; Calculate the mean variation ,in The probability is the number of times it takes a value; Calculate the cumulative contribution rate of each performance-forming factor after sorting them by average variation from largest to smallest. ,Will Compared with the preset cumulative contribution rate threshold If a comparison is made, Then the rankings are from 1st to 6th. k All performance-forming factors were identified as key performance-forming factors.

8. The method for assessing the probability of human error according to claim 7, characterized in that, Step 4, which involves identifying key risk levels, specifically calculates the ratio variation for each rating level of the identified key performance factors. , This indicates that the rating level is a high-risk status.

9. A terminal device, comprising a processor, a memory, and a computer program stored in the memory; characterized in that, When the processor executes a computer program, it implements the human error probability assessment method as described in any one of claims 1-8.

10. A computer-readable storage medium storing a computer program; characterized in that, When the computer program is executed by a processor, it implements the human error probability assessment method according to any one of claims 1-8.

Images