A method and system for assessing potential risks of damage to building exterior walls

Abstract

This invention discloses a method and system for assessing potential risks of damage to building exterior walls. The assessment method includes the following steps: establishing an evaluation index system using a three-level assessment process: single damage → single-type damage to exterior wall surface → comprehensive damage to exterior wall surface; constructing an expert knowledge base, inputting the evaluation index system into a fine-tuned basic large language model to output an initial index relationship matrix; constructing a fuzzy direct comparison matrix, calculating the fuzzy comprehensive degree value of each index, and determining the initial weight vector of each index based on importance; constructing a fuzzy direct influence matrix, decomposing and standardizing the fuzzy direct influence matrix to obtain a standardized matrix; processing the causality degree to obtain the influence weight vector, thus obtaining the influence weight vector of each index; fusing the importance weight vector and the influence weight vector to obtain the comprehensive weight vector of each index; calculating the spatial distance between the target scene and the template scene and selecting the optimal template.

Description

Technical Field

[0001] This invention relates to the fields of building safety assessment, drone detection, and artificial intelligence, and in particular to a method and system for assessing potential risks of damage to building exterior walls. Background Technology

[0002] Building exterior walls, as a crucial component of urban infrastructure, pose a direct threat to public safety due to damage such as cracks and peeling, necessitating the establishment of scientific methods for assessing potential risks. Currently, research on risk assessment for building exterior wall damage is relatively limited. Most technical solutions focus on data collection methods such as manual inspections or drone photography, while a few scholars evaluate single factors like material aging and crack severity based on methods like the analytic hierarchy process (AHP) and entropy weighting. However, existing technical solutions fail to overcome numerous shortcomings, particularly the lack of assessment schemes specifically addressing the potential risks of building exterior wall damage. Specifically: 1. Existing methods rely excessively on expert experience for indicator selection, failing to fully utilize the vast amounts of professional textual knowledge from technical specifications, maintenance reports, and academic papers. This results in a lack of data-driven support in the indicator system construction process, hindering automated construction and dynamic updates. 2. The interrelationships between indicators are difficult to quantify objectively. The traditional DEMATEL method relies on direct expert scoring to construct the initial influence matrix, leading to inefficiency, poor consistency in judgments, and difficulty in ensuring the objectivity and reliability of the analysis results, failing to truly reflect the complex relationships between indicators. 3. The construction of ANP network structure lacks data-driven basis. Existing ANP methods lack objective basis when constructing network structure, the accuracy of weight calculation needs to be improved, and it cannot fully reflect the nonlinear correlation between evaluation indicators, affecting the scientificity and credibility of evaluation results. 4. Existing solutions lack automated and intelligent evaluation processes that are deeply integrated with UAV detection data, and cannot achieve a closed loop from data collection to risk rating, making it difficult to meet the needs of intelligent urban management. Summary of the Invention

[0003] This solution addresses the problems and needs raised above by proposing a method and system for assessing potential risks of damage to building exterior walls. Due to the adoption of the following technical features, it is able to achieve the aforementioned technical objectives and bring about several other technical benefits.

[0004] One object of the present invention is to provide a method for assessing the potential risk of damage to building exterior walls, comprising the following steps: S10: Construct an evaluation index system for potential risks of building exterior wall damage: Based on building industry standards, expert interviews and literature reviews, adopt a three-level assessment process of single damage → single type of damage to exterior wall surface → comprehensive damage to exterior wall surface, and establish an evaluation index system. S20: Intelligent Assisted Construction of Indicator Relationships Based on Large Language Models: Construct an expert knowledge base and use it to fine-tune the selected basic large language model. Input the evaluation index system into the fine-tuned basic large language model to output the initial index relationship matrix and provide the confidence level of each relationship. Experts correct and evaluate the relationships output by the basic large language model and fuzzify the evaluation. S30: Calculation of indicator importance weights based on Fuzzy-ANP: Construct a fuzzy direct comparison matrix, calculate the fuzzy comprehensive degree value of each indicator, calculate the probability of any two fuzzy comprehensive degree values ​​of the same expert and determine the initial weight vector of each indicator based on importance, calculate the priority weight vector of each indicator and perform consistency verification. S40: Calculation of mutual influence weights of indicators based on Fuzzy-DEMATEL laboratory analysis: Construct a fuzzy direct influence matrix, decompose and standardize the fuzzy direct influence matrix to obtain a standardized matrix; calculate the comprehensive influence matrix of the lower bound, median, and upper bound to obtain the fuzzy comprehensive influence matrix, calculate the influence degree, influenced degree, centrality, and causation degree of the fuzzy comprehensive influence matrix; process the causation degree to obtain the influence weight vector, calculate the comprehensive influence weight of all experts on the indicators to obtain the influence weight vector of each indicator; S50: Select the optimal template: Combine the importance weight vector obtained by the Fuzzy-ANP method with the influence weight vector obtained by the Fuzzy-DEMATEL method to obtain the comprehensive weight vector of each indicator; calculate the spatial distance between the target scene and the template scene and select the optimal template.

[0005] In addition, the method and system for assessing potential risks of damage to building exterior walls according to the present invention may also have the following technical features: In one example of the present invention, step S10 specifically includes the following steps: S11: Develop a fuzzy language scale: Use triangular fuzzy numbers to quantify expert language evaluations, where the triangular fuzzy number is represented as... , where l, m and u represent the minimum possible value, the most likely value and the maximum possible value of the fuzzy event, respectively; S12: Design an independent questionnaire for each building exterior wall scenario to be evaluated. The building exterior wall scenarios include the target scenario and multiple candidate template scenarios. The questionnaire consists of three parts: the first part is a basic description of the exterior wall to be evaluated, corresponding to the specific background of the research question; the second part is the direct comparison matrix C of the importance between indicators, which is used for subsequent Fuzzy-ANP analysis; the third part is the direct influence relationship matrix R between indicators, which is used for subsequent Fuzzy-DEMATEL laboratory analysis.

[0006] In one example of the present invention, step S20 specifically includes the following steps: S21: Construct an expert knowledge base: Collect unstructured text data related to building exterior wall inspection, including technical specifications, maintenance reports, academic papers, and expert interview records, and perform cleaning, deduplication, and formatting to form a domain knowledge base; S22: Fine-tuning of the basic large language model: Select a basic large language model and fine-tune it using a domain knowledge base; the input is a text fragment containing indicator descriptions, and the output is the relational labels between indicators; S23: Initial Relationship Matrix Generation: The indicator system and the definition of each indicator are input into the fine-tuned basic large language model. By designing specific prompt words, the basic large language model analyzes the potential relationships between indicators one by one. S24: Expert Correction and Fuzziness Processing: The initial relation matrix and confidence results generated by the basic large language model are provided to the expert group. Experts review, correct, and confirm the relations output by the basic large language model based on their own experience. For the relations between indicators confirmed by the experts, Fuzzy-DEMATEL linguistic variables are used to make a final evaluation of the strength of the relations. At the same time, experts use Fuzzy-ANP linguistic variables to compare the importance of all indicators pairwise. Finally, all the evaluation languages ​​of the experts are converted into corresponding triangular fuzzy numbers to construct a fuzzy direct comparison matrix and a fuzzy direct influence matrix.

[0007] In one example of the present invention, step S30 specifically includes the following steps: S31: Constructing a fuzzy direct comparison matrix: Organize the fuzzy evaluation results of the k-th expert on the importance of the indicators into a fuzzy direct comparison matrix. ; Elements in the array This indicates the importance of the i-th indicator relative to the j-th indicator. The result is determined by experts based on preset linguistic variables and converted into a corresponding triangular fuzzy number. Among them, the fuzzy direct comparison matrix The expression is as follows: in, For triangular fuzzy numbers; S32: Calculate the fuzzy generalization degree value: The fuzzy generalization degree value of each index i is calculated using the extended analysis method; the expression for calculating the fuzzy generalization degree value is: S33: Calculate the likelihood and determine the weight vector: First, for any two fuzzy comprehensive degree values ​​of the same expert k and ,calculate The probability degree; this probability degree is defined as the ordinate of the highest point of the intersection of the membership functions of two fuzzy numbers, that is: This expression is equivalent to the following piecewise function: Among them, when When the degree is 1, it corresponds to the case where the ordinate of the highest point of the intersection of the membership functions of the two fuzzy numbers is 1. Then, calculate the weight vector of expert k, a fuzzy generalization degree value. Greater than all others The degree of possibility of (j≠i) is defined as: right Normalization is performed to obtain the initial weight vector of each indicator based on importance given by the k-th expert. S34: Calculate the priority weight vector and perform consistency verification: First, for the k-th expert, calculate the fuzzy comprehensive degree value for each index i. Greater than all other indicators The probability of (j≠i), i.e.: Then, to After normalization, we obtain the relative importance weight vector of each indicator given by the k-th expert using the fuzzy ANP method: Repeat the above steps to obtain the weight vector of all experts; randomly divide the evaluation results of all experts into two groups with an equal number of people in each group, and use the F test to verify the validity of the data; if the significance level P value is greater than 0.05, it indicates that there is no significant difference between the two groups of data and the expert evaluation results have a high degree of consistency. S35: Calculate the priority weight vector: Based on step S34, calculate the comprehensive weight of all experts for indicator i. The geometric mean method is used to express it as follows: Finally, the overall weights are normalized to obtain the final weights of each indicator. Its expression is: This yields the priority weight vector for this evaluation scenario. .

[0008] In one example of the present invention, step S31 further includes: performing a consistency check on the matrix by first defuzzifying the fuzzy matrix to obtain a clear judgment matrix. Then calculate the consistency ratio. Its expression is as follows: The formula for calculating the consistency index (CI) is as follows: In the formula, For matrix The largest eigenvalue, where n is the matrix order; RI is the random consistency index, which is determined by looking up a table based on the matrix order. If CR ≤ 0.1, the matrix is ​​considered to have passed the consistency test. If CR > 0.1, feedback to experts is required for correction.

[0009] In one example of the present invention, step S40 specifically includes the following steps: S41: Constructing a fuzzy direct influence matrix: The k-th expert, based on the actual condition of the building's exterior wall, determines the interrelationships between various indicators; expert evaluation values. This indicates the degree of direct influence of indicator i on indicator j, and is filled in the matrix. At the intersection of the i-th row and j-th column; Convert to triangular fuzzy number Among them, the fuzzy direct influence matrix constructed by the k-th expert As shown in the following formula: in, For triangular fuzzy numbers; S42: The direct impact matrix of decomposition and standardized fuzzy model: The fuzziness directly affects the matrix Decomposed into a lower bound matrix Median matrix and upper bound matrix , respectively by Composition, where the lower bound matrix Median matrix and upper bound matrix The expression is: Standardize each matrix individually; the standardization factor is the maximum value of the sum of each row: The standardized matrix is ​​then: S43: Calculate the fuzzy comprehensive influence matrix: Based on the standardized matrix, the lower bound comprehensive influence matrix is ​​calculated respectively. Median Comprehensive Influence Matrix Upper bound comprehensive influence matrix The calculation formula is as follows: In the formula, I is an n×n identity matrix; This yields the fuzzy comprehensive influence matrix. ,in ; S44: Calculate influence, affectedness, centrality, and causality: Fuzzy Comprehensive Influence Matrix The influence is obtained by calculating the sum of each row. The sum of each column yields the degree of influence. Among them, influence And the degree of influence The expression is: To facilitate comparison, the centroid method is used to fuzzify the fuzzy numerical solutions into clear values: Further calculate the centrality of each indicator and causal degree : Among them, centrality Centrality reflects the importance of the indicator in the system; a higher centrality indicates a more critical indicator. Causality... If the value is positive, it indicates that the indicator is a causal factor, meaning it actively influences other indicators; if the value is negative, it is an outcome factor, meaning it is influenced by other indicators. S45: Calculate the weight vector based on influence relationships: The causal degree is transformed into a positive value index for comprehensive weighting to obtain the influence weight vector of all experts; the evaluation results of all experts are randomly divided into two groups with equal numbers of people in each group, and the F test is used to verify the validity of the data; if the p value is greater than 0.05, it indicates that there is no significant difference between the two groups of data and the expert evaluation results have a high degree of consistency. Based on this, calculate the combined influence weight of all experts on indicator i. The geometric mean method is used to express it as follows: Finally, the overall impact weights are normalized to obtain the final impact weight vectors of each indicator. Its expression is: This yields the influence weight vector for this evaluation scenario. .

[0010] In one example of the present invention, step S45, obtaining the influence weight vector of all experts, includes the following steps: The causality degree is converted into a positive value that can be used for comprehensive weighting. The causality degree is then shifted using the following formula: right After normalization, we obtain the influence weights of each indicator given by the k-th expert based on the Fuzzy-DEMATEL method: Repeat the above steps to obtain the influence weight vector of all experts.

[0011] In one example of the present invention, step S50 specifically includes the following steps: S51: Calculate the comprehensive weight vector for each scenario: The importance weight vector obtained by the Fuzzy-ANP method and the influence weight vector obtained by the Fuzzy-DEMATEL method are fused to obtain the comprehensive weight vector of each indicator. For the building exterior wall scene to be evaluated, denoted as target scene O, and for each candidate template scene, denoted as template scene G, the comprehensive weight vector W is calculated using the following formula: In the formula, Let be the Fuzzy-ANP importance weight of the i-th indicator; Let be the Fuzzy-DEMATEL influence weight of the i-th indicator; let be the comprehensive weight vector of the target scenario O. The comprehensive weight vector of each template scenario G is denoted as... ; S52: Calculate the spatial distance between the target scene and the template scene and select the optimal template: The target scene O and each candidate template scene are calculated using the following formula. Spatial distance between This distance reflects the difference between the target scene and the template scene in terms of the overall importance of the indicators, and its expression is: In the formula, n is the total number of evaluation indicators, and m is the number of candidate template scenarios; Wherein, the distance value d(O, The smaller the value, the closer the target scene is to the Gth template scene in terms of the overall importance of various indicators; the candidate template with the smallest distance is selected as the optimal template. Its expression is: The building exterior wall type or inspection area corresponding to the optimal template has the most similar damage risk characteristics to the target scenario. Its assessment results, risk level classification or maintenance strategy can be referenced to provide a basis for risk assessment and maintenance decisions for the target scenario.

[0012] Another object of the present invention is to provide a potential risk assessment system for damage to building exterior walls, comprising: A risk assessment indicator system module was constructed, configured to establish an evaluation indicator system based on building industry standards, expert interviews, and literature reviews, using a three-level assessment process: single damage → single type of damage to exterior wall surface → comprehensive damage to exterior wall surface. The intelligent auxiliary construction module is configured to build an expert knowledge base and use the expert knowledge base to fine-tune the selected basic large language model. The evaluation index system is input into the fine-tuned basic large language model, thereby outputting an initial index relationship matrix and providing the confidence level of each relationship. Experts correct and evaluate the relationships output by the basic large language model and perform fuzzy processing on the evaluation. The importance weight calculation module is configured to construct a fuzzy direct comparison matrix, calculate the fuzzy comprehensive degree value of each indicator, calculate the probability of any two fuzzy comprehensive degree values ​​of the same expert and determine the initial weight vector of each indicator based on importance, calculate the priority weight vector of each indicator and perform consistency verification. The mutual influence weight calculation module is configured to construct a fuzzy direct influence matrix, decompose and standardize the fuzzy direct influence matrix to obtain a standardized matrix; calculate the comprehensive influence matrix of the lower bound, median and upper bound to obtain a fuzzy comprehensive influence matrix; calculate the influence degree, influenced degree, centrality and causation degree of the fuzzy comprehensive influence matrix; process the causation degree to obtain the influence weight vector; calculate the comprehensive influence weight of all experts on the indicator to obtain the influence weight vector of each indicator. Select the optimal template module and configure it to fuse the importance weight vector obtained by the Fuzzy-ANP method with the influence weight vector obtained by the Fuzzy-DEMATEL method to obtain the comprehensive weight vector of each indicator; calculate the spatial distance between the target scene and the template scene and select the optimal template.

[0013] In one example of the present invention, the intelligent auxiliary construction module includes: An expert knowledge base unit is constructed and configured to collect unstructured text data related to building exterior wall inspection, including technical specifications, maintenance reports, academic papers, and expert interview records. This data is then cleaned, deduplicated, and formatted to form a domain knowledge base. The basic large language model fine-tuning unit is configured to select a basic large language model and fine-tune it using a domain knowledge base; the input is a text fragment containing indicator descriptions, and the output is the relational labels between indicators. The initial relation matrix generation unit is configured to input the indicator system and the definition of each indicator into the fine-tuned basic large language model. By designing specific prompt words, the basic large language model can analyze the potential relationships between indicators one by one. The expert correction and fuzzification unit is configured to provide the initial relation matrix and confidence results generated by the basic large language model to the expert group. Experts, based on their own experience, review, correct, and confirm the relations output by the basic large language model. For the relations between indicators confirmed by the experts, Fuzzy-DEMATEL linguistic variables are used to make a final evaluation of the strength of the relations. At the same time, experts use Fuzzy-ANP linguistic variables to compare the importance of all indicators pairwise. Finally, all the evaluation languages ​​of the experts are converted into corresponding triangular fuzzy numbers to construct a fuzzy direct comparison matrix and a fuzzy direct influence matrix.

[0014] Compared with the prior art, the present invention has the following beneficial effects: Firstly, in terms of constructing the evaluation index system, traditional methods heavily rely on expert experience, resulting in highly subjective index extraction and difficulty in dynamic updates. In contrast, this invention introduces a large language model to deeply mine and learn from massive amounts of professional texts (including technical specifications, maintenance reports, academic papers, expert interview records, and other unstructured data), achieving automated generation of the initial index relationship matrix. This provides experts with reference relationships with a confidence level as high as 0.82 (as shown in the embodiment). This process significantly reduces the subjectivity and cognitive burden of manually constructing the index system, making the establishment of the index system more objective and scientific. At the same time, it has the ability to be dynamically updated and can be continuously optimized and improved with the accumulation of domain knowledge.

[0015] Secondly, regarding the calculation of indicator weights, existing technologies often employ a single weighting method, making it difficult to simultaneously consider both the importance of the indicator itself and the mutual influence between indicators. This invention innovatively combines Fuzzy Analytic Hierarchy Process (FuzzyANP) with Fuzzy Dematel (Fuzzy Decision Experimentation and Evaluation Laboratory) to comprehensively calculate the weights of each evaluation indicator from two dimensions: "indicator importance" and "mutual influence between indicators," resulting in more comprehensive and accurate weight calculations. As shown in the examples, the F-test (all P-values ​​> 0.05) verified the high consistency of expert evaluation results (consistency coefficient reaching 0.89), significantly outperforming the traditional expert scoring method (consistency coefficient 0.67). In terms of automation and intelligence in the evaluation process, this invention constructs a three-level assessment process: "single damage → single-type damage to exterior wall surface → comprehensive damage to exterior wall surface," deeply integrated with UAV detection data. Compared to the traditional method requiring a 3-day evaluation cycle, this invention's method takes only 1 day from data collection to template selection, significantly improving evaluation efficiency.

[0016] Furthermore, this invention introduces an optimal template matching mechanism. By calculating the spatial distance between the target scene and multiple template scenes on the comprehensive weight vector of indicators, the most similar template scene is automatically selected as the evaluation reference. In the embodiment, the template scene G2 (spatial distance 0.064) that best matches the target scene was successfully identified, and a targeted maintenance plan was formulated based on this. After the plan was implemented, no new serious damage was found on the exterior wall of the target scene in the subsequent year of monitoring, verifying the accuracy and reliability of the evaluation results.

[0017] Finally, from the perspective of practical application effects, this invention, through the systematic integration of large language model-assisted construction and fuzzy multi-criteria decision-making method, significantly shortens the evaluation cycle, reduces the time and manpower costs of expert participation, and improves the consistency and reliability of evaluation results. It provides strong technical support for the safety risk management of urban building facades and has good engineering promotion value and economic and social benefits.

[0018] The preferred embodiments of the invention will be described in more detail below with reference to the accompanying drawings, so as to facilitate an understanding of the features and advantages of the invention. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. The drawings are merely illustrative of some embodiments of the present invention and are not intended to limit the scope of the present invention to all embodiments.

[0020] Figure 1 This is a flowchart of a method for assessing potential risks of damage to building exterior walls according to an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The same reference numerals in the drawings represent the same components. It should be noted that the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0022] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms “first,” “second,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, “an” or “a” and similar terms do not necessarily indicate a quantity limitation. Terms such as “comprising” or “including” mean that the element or object preceding the word encompasses the element or object listed following the word and its equivalents, without excluding other elements or objects. Terms such as “connected” or “linked” are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as “upper,” “lower,” “left,” and “right” are used only to indicate relative positional relationships; these relative positional relationships may change accordingly when the absolute position of the described object changes.

[0023] According to a method for assessing potential risks of damage to building exterior walls based on a first aspect of the present invention, such as Figure 1 As shown, it includes the following steps: S10: Constructing a Potential Risk Assessment Index System for Building Exterior Wall Damage: Based on building industry standards, expert interviews, and literature reviews, a three-tiered assessment process is adopted: single damage → single type of exterior wall damage → comprehensive exterior wall damage. An evaluation index system is established to cover four major exterior wall defect types: cracks, hollow areas, leaks, and detachments. Specific evaluation indicators are set for each defect type, and each indicator is quantified through risk level classification, as shown in Table 1. Table 1. Evaluation Index System for Potential Risk of Damage to Building Exterior Walls After completing the construction of the indicator system, the next stage is to calculate the comprehensive weight of the indicators based on the improved LLM-ANP-DEMATEL. S20: Intelligent Assisted Construction of Indicator Relationships Based on Large Language Models: Construct an expert knowledge base and use it to fine-tune the selected basic large language model. Input the evaluation index system into the fine-tuned basic large language model to output the initial index relationship matrix and provide the confidence level of each relationship. Experts correct and evaluate the relationships output by the basic large language model and fuzzify the evaluation. S30: Calculation of indicator importance weights based on Fuzzy-ANP: Construct a fuzzy direct comparison matrix, calculate the fuzzy comprehensive degree value of each indicator, calculate the probability of any two fuzzy comprehensive degree values ​​of the same expert and determine the initial weight vector of each indicator based on importance, calculate the priority weight vector of each indicator and perform consistency verification. S40: Calculation of mutual influence weights of indicators based on Fuzzy-DEMATEL laboratory analysis: Construct a fuzzy direct influence matrix, decompose and standardize the fuzzy direct influence matrix to obtain a standardized matrix; calculate the comprehensive influence matrix of the lower bound, median, and upper bound to obtain the fuzzy comprehensive influence matrix, calculate the influence degree, influenced degree, centrality, and causation degree of the fuzzy comprehensive influence matrix; process the causation degree to obtain the influence weight vector, calculate the comprehensive influence weight of all experts on the indicators to obtain the influence weight vector of each indicator; S50: Select the optimal template: Combine the importance weight vector obtained by the Fuzzy-ANP method with the influence weight vector obtained by the Fuzzy-DEMATEL method to obtain the comprehensive weight vector of each indicator; calculate the spatial distance between the target scene and the template scene and select the optimal template.

[0024] In terms of constructing the evaluation index system, traditional methods rely heavily on expert experience, resulting in highly subjective index extraction and difficulty in dynamic updates. This invention, however, introduces a large language model to deeply mine and learn from massive amounts of professional text (including technical specifications, maintenance reports, academic papers, expert interview records, and other unstructured data), achieving automated generation of the initial index relationship matrix. This provides experts with reference relationships with a confidence level as high as 0.82 (as shown in the embodiment). This process significantly reduces the subjectivity and cognitive burden of manually constructing the index system, making the establishment of the index system more objective and scientific, while also possessing dynamic updating capabilities, continuously optimizing and improving with the accumulation of domain knowledge.

[0025] Regarding the weight calculation dimension of indicators, existing technologies often employ a single weighting method, making it difficult to simultaneously consider both the importance of the indicators themselves and the mutual influence between indicators. This invention innovatively combines Fuzzy Analytic Hierarchy Process (FuzzyANP) with Fuzzy Dematel (Fuzzy Decision Experimentation and Evaluation Laboratory) to comprehensively calculate the weight of each evaluation indicator from two dimensions: "individual importance of the indicator" and "mutual influence between indicators," resulting in more comprehensive and accurate weight calculation results. As shown in the example, the F-test (all P-values ​​> 0.05) verifies the high consistency of expert evaluation results (consistency coefficient reaches 0.89), significantly outperforming the traditional expert scoring method (consistency coefficient 0.67). In terms of automation and intelligence in the evaluation process, this invention constructs a three-level assessment process: "single damage → single-type damage to exterior wall surface → comprehensive damage to exterior wall surface," deeply integrated with UAV detection data. Compared to the traditional method requiring a 3-day evaluation cycle, this invention's method takes only 1 day from data collection to template selection, significantly improving evaluation efficiency.

[0026] This evaluation method introduces an optimal template matching mechanism. By calculating the spatial distance between the target scene and multiple template scenes on the comprehensive weight vector of indicators, it automatically selects the most similar template scene as the evaluation reference. In the example, the template scene G2 (spatial distance 0.064) that best matches the target scene was successfully identified, and a targeted maintenance plan was formulated based on this. After the plan was implemented, no new serious damage was found on the exterior wall of the target scene in the subsequent year of monitoring, verifying the accuracy and reliability of the evaluation results.

[0027] From the perspective of practical application, this evaluation method significantly shortens the evaluation cycle, reduces the time and manpower costs of expert participation, and improves the consistency and reliability of evaluation results by using the auxiliary construction of a large language model and the systematic integration of fuzzy multi-criteria decision-making methods. It provides strong technical support for the safety risk management of urban building facades and has good engineering promotion value and economic and social benefits.

[0028] In one example of the present invention, step S10 specifically includes the following steps: S11: Developing a fuzzy language scale: Using triangular fuzzy numbers to quantify expert language evaluations. A triangular fuzzy number can be represented as... , where l, m and u represent the minimum possible value, the most likely value and the maximum possible value of the fuzzy event, respectively, and their membership functions satisfy the definition of equation (1). Triangular fuzzy numbers, with their characteristics of being easy to calculate and understand, have become a commonly used tool for dealing with uncertainty problems.

[0029] This study employs two fuzzy multi-criteria decision-making methods: Fuzzy-ANP (Fuzzy Analytic Hierarchy Process) and Fuzzy-DEMATEL (Fuzzy Dematel). Fuzzy-ANP measures the relative importance of evaluation indicators, while Fuzzy-DEMATEL analyzes the degree of mutual influence between indicators. To facilitate semantic judgment by experts, corresponding sets of linguistic variables were established and assigned triangular fuzzy numbers, as shown in Table 2. In Fuzzy-ANP, the linguistic variables include "equally important," "weakly important," "strongly important," "significantly important," and "absolutely important"; in Fuzzy-DEMATEL, the linguistic variables include "irrelevant," "weakly correlated," "correlated," "significantly correlated," and "linearly correlated." Each linguistic variable corresponds to a triangular fuzzy number and its reciprocal, the latter used to construct an inverse comparison matrix.

[0030] Table 2 Fuzzy linguistic variables and their corresponding triangular fuzzy numbers S12: A separate questionnaire was designed for each building exterior wall scenario to be evaluated (including the target scenario and multiple candidate template scenarios). The building exterior wall scenarios include the target scenario and multiple candidate template scenarios; in this study, a total of 1 target scenario and 5 template scenarios were involved, totaling 6 scenarios to be evaluated. The questionnaire consists of three parts: the first part is a basic description of the exterior wall to be evaluated, corresponding to the specific background of the research question; the second part is a direct comparison matrix C of the importance between indicators, used for subsequent Fuzzy-ANP analysis; the third part is a direct influence relationship matrix R between indicators, used for subsequent Fuzzy-DEMATEL analysis. Except for the first part, the questionnaires for different scenarios are identical.

[0031] This study invited no fewer than 10 experts to participate in the questionnaire survey. Each expert received 6 questionnaires (corresponding to 1 target scenario and 5 template scenarios). Based on their professional knowledge and the scenario descriptions provided in Part 1 of the questionnaire, experts independently completed Matrix C and Matrix R. All questionnaires were distributed and collected either in person or online, and the collected questionnaires were numbered and organized to ensure data integrity and traceability.

[0032] It should be noted that the preceding step includes: selecting an expert panel: inviting several experts in fields such as resource-depleted area redevelopment, urban planning, natural resource management, and commercial planning to participate in the research. Invited experts should work in government departments, natural resource enterprises, commercial planning companies, urban planning companies, etc., and possess extensive practical experience. Following Willis's suggestion, the experts were randomly divided into two groups, and the validity of the data was verified by comparing the differences between the two groups. This study invited several experts, divided into two equal groups, to improve the reliability of the evaluation results.

[0033] In one example of the present invention, step S20 specifically includes the following steps: S21: Construct an expert knowledge base: Collect unstructured text data related to building exterior wall inspection, including technical specifications, maintenance reports, academic papers, and expert interview records, and perform cleaning, deduplication, and formatting to form a domain knowledge base; S22: Fine-tuning of the Basic Large Language Model: Select a basic large language model and fine-tune it using a domain knowledge base. The goal of fine-tuning is to enable the model to understand and generate professional content related to building exterior wall damage risk assessment, especially to accurately identify potential relationships between indicators. Supervised learning is used during the fine-tuning process, where the input is a text fragment containing indicator descriptions, and the output is the relationship labels between indicators. S23: Initial Relationship Matrix Generation: The indicator system and the definitions of each indicator are input into the fine-tuned basic large language model. Specific prompts are designed, such as "Please analyze whether indicator A (e.g., crack width) has a direct impact on indicator B (e.g., hollow area) in the risk assessment of building exterior wall damage? Is the impact positive or negative?" This allows the basic large language model to analyze the potential relationships between indicators pairwise. Based on its knowledge learned from professional texts, the LLM outputs an initial indicator relationship matrix and provides the confidence level for each relationship. This process can automatically identify a large number of potential and complex indicator associations, providing a scientific reference basis for subsequent expert scoring.

[0034] S24: Expert Correction and Fuzzification: The initial relation matrix and confidence results generated by the basic large language model are provided to the expert group. Experts review, correct, and confirm the relations output by the basic large language model based on their own experience. For the relations between indicators confirmed by the experts, the Fuzzy-DEMATEL linguistic variables in Table 2 are used to make a final evaluation of the strength of the relations. At the same time, the experts use the Fuzzy-ANP linguistic variables to compare the importance of all indicators pairwise. Finally, all the evaluation languages ​​of the experts are converted into the corresponding triangular fuzzy numbers according to Table 2 to construct the fuzzy direct comparison matrix and the fuzzy direct influence matrix. In one example of the present invention, step S30 specifically includes the following steps: S31: Constructing a fuzzy direct comparison matrix: Organize the fuzzy evaluation results of the k-th expert on the importance of the indicators into a fuzzy direct comparison matrix. ; Elements in the array This indicates the importance of the i-th indicator relative to the j-th indicator. The result is determined by experts based on preset linguistic variables and converted into a corresponding triangular fuzzy number. Among them, the fuzzy direct comparison matrix The expression is as follows: in, For triangular fuzzy numbers; S32: Calculate the fuzzy generalization degree value: The fuzzy generalization degree value of each index i is calculated using the extended analysis method; the expression for calculating the fuzzy generalization degree value is: S33: Calculate the likelihood and determine the weight vector: First, for any two fuzzy comprehensive degree values ​​of the same expert k and ,calculate The probability degree; this probability degree is defined as the ordinate of the highest point of the intersection of the membership functions of two fuzzy numbers, that is: This expression can be equivalently represented as the following piecewise function: Among them, when When the degree is 1, it corresponds to the case where the ordinate of the highest point of the intersection of the membership functions of the two fuzzy numbers is 1. Then, calculate the weight vector of expert k, a fuzzy generalization degree value. Greater than all others The degree of possibility of (j≠i) is defined as: right After normalization, the initial weight vector of each indicator based on importance given by the k-th expert can be obtained. S34: Calculate the priority weight vector and perform consistency verification: First, for the k-th expert, calculate the fuzzy comprehensive degree value for each index i. Greater than all other indicators The probability of (j≠i), i.e.: Then, to After normalization, we obtain the relative importance weight vector of each indicator given by the k-th expert using the fuzzy ANP method: Repeat the above steps to obtain the weight vector of all experts; randomly divide the evaluation results of all experts into two groups with an equal number of people in each group, and use the F test to verify the validity of the data; if the significance level P value is greater than 0.05, it indicates that there is no significant difference between the two groups of data and the expert evaluation results have a high degree of consistency. S35: Calculate the priority weight vector: Based on step S34, calculate the comprehensive weight of all experts for indicator i. The geometric mean method is used to express it as follows: Finally, the overall weights are normalized to obtain the final weights of each indicator. Its expression is: This yields the priority weight vector for this evaluation scenario. This vector reflects the relative importance of each indicator; the larger the weight value, the higher the importance of that indicator.

[0035] In one example of the present invention, step S31 further includes: to ensure the logical consistency of expert judgment, a consistency check needs to be performed on the matrix. First, the fuzzy matrix is ​​defuzzified to obtain a clear judgment matrix. Then calculate the consistency ratio. Its expression is as follows: The formula for calculating the consistency index (CI) is as follows: In the formula, For matrix The largest eigenvalue, where n is the matrix order; RI is the random consistency index, which is determined by looking up a table based on the matrix order. If CR ≤ 0.1, the matrix is ​​considered to have passed the consistency test. If CR > 0.1, feedback to experts is required for correction.

[0036] In one example of the present invention, step S40 specifically includes the following steps: S41: Constructing a fuzzy direct influence matrix: The k-th expert, based on the actual condition of the building's exterior wall, determines the interrelationships between various indicators; expert evaluation values. This indicates the degree of direct influence of indicator i on indicator j, and is filled in the matrix. At the intersection of the i-th row and j-th column; through the correspondence shown in Table 2, Convert to triangular fuzzy number Among them, the fuzzy direct influence matrix constructed by the k-th expert As shown in the following formula: in, For triangular fuzzy numbers; S42: The direct impact matrix of decomposition and standardized fuzzy model: The fuzziness directly affects the matrix Decomposed into a lower bound matrix Median matrix and upper bound matrix , respectively by Composition, where the lower bound matrix Median matrix and upper bound matrix The expression is: Standardize each matrix individually; the standardization factor is the maximum value of the sum of each row: The standardized matrix is ​​then: S43: Calculate the fuzzy comprehensive influence matrix: Based on the standardized matrix, the lower bound comprehensive influence matrix is ​​calculated respectively. Median Comprehensive Influence Matrix Upper bound comprehensive influence matrix The comprehensive influence matrix represents the total influence relationship between indicators after considering indirect transmission, and its calculation formula is as follows: In the formula, I is an n×n identity matrix; This yields the fuzzy comprehensive influence matrix. ,in ; S44: Calculate influence, affectedness, centrality, and causality: Fuzzy Comprehensive Influence Matrix The influence is obtained by calculating the sum of each row. The sum of each column yields the degree of influence. Among them, influence And the degree of influence The expression is: To facilitate comparison, the centroid method is used to fuzzify the fuzzy numerical solutions into clear values: Further calculate the centrality of each indicator and causal degree : Among them, centrality Centrality reflects the importance of the indicator in the system; a higher centrality indicates a more critical indicator. Causality... If the value is positive, it indicates that the indicator is a causal factor, meaning it actively influences other indicators; if the value is negative, it is an outcome factor, meaning it is influenced by other indicators. S45: Calculate the weight vector based on influence relationships: The causality degree is converted into a positive value that can be used for comprehensive weighting. The causality degree is then shifted using the following formula: Then to After normalization, we obtain the influence weights of each indicator given by the k-th expert based on the Fuzzy-DEMATEL method: Repeat the above steps to obtain the influence weight vector of all experts; randomly divide the evaluation results of all experts into two groups with equal numbers of people in each group, and use the F test to verify the validity of the data; if the P value is greater than 0.05, it indicates that there is no significant difference between the two groups of data and the expert evaluation results have a high degree of consistency. Based on this, calculate the combined influence weight of all experts on indicator i. The geometric mean method is used to express it as follows: Finally, the overall impact weights are normalized to obtain the final impact weight vectors of each indicator. Its expression is: This yields the influence weight vector for this evaluation scenario. This vector reflects the relative importance of each indicator in their interrelationships; a larger weight value indicates a stronger influence of that indicator within the system. The same method was used to process expert questionnaire data from other scenarios to be evaluated.

[0037] In one example of the present invention, step S45, obtaining the influence weight vector of all experts, includes the following steps: The causality degree is converted into a positive value that can be used for comprehensive weighting. The causality degree is then shifted using the following formula: right After normalization, we obtain the influence weights of each indicator given by the k-th expert based on the Fuzzy-DEMATEL method: Repeat the above steps to obtain the influence weight vector of all experts.

[0038] In one example of the present invention, step S50 specifically includes the following steps: S51: Calculate the comprehensive weight vector for each scenario: The importance weight vector obtained by the Fuzzy-ANP method and the influence weight vector obtained by the Fuzzy-DEMATEL method are fused to obtain the comprehensive weight vector of each indicator. For the building exterior wall scene to be evaluated, denoted as target scene O, and for each candidate template scene, denoted as template scene G, the comprehensive weight vector W is calculated using the following formula: In the formula, Let be the Fuzzy-ANP importance weight of the i-th indicator; Let be the Fuzzy-DEMATEL influence weight of the i-th indicator; let be the comprehensive weight vector of the target scenario O. The comprehensive weight vector of each template scenario G is denoted as... ; S52: Calculate the spatial distance between the target scene and the template scene and select the optimal template: The target scene O and each candidate template scene are calculated using the following formula. Spatial distance between This distance reflects the difference between the target scene and the template scene in terms of the overall importance of the indicators, and its expression is: In the formula, n is the total number of evaluation indicators, and m is the number of candidate template scenarios; Wherein, the distance value d(O, The smaller the value, the closer the target scene is to the Gth template scene in terms of the overall importance of various indicators; the candidate template with the smallest distance is selected as the optimal template. Its expression is: The building exterior wall type or inspection area corresponding to the optimal template has the most similar damage risk characteristics to the target scenario. Its assessment results, risk level classification or maintenance strategy can be referenced to provide a basis for risk assessment and maintenance decisions for the target scenario.

[0039] According to a second aspect of the present invention, a potential risk assessment system for damage to building exterior walls includes: A risk assessment indicator system module was constructed, configured to establish an evaluation indicator system based on building industry standards, expert interviews, and literature reviews, using a three-level assessment process: single damage → single type of damage to exterior wall surface → comprehensive damage to exterior wall surface. The intelligent auxiliary construction module is configured to build an expert knowledge base and use the expert knowledge base to fine-tune the selected basic large language model. The evaluation index system is input into the fine-tuned basic large language model, thereby outputting an initial index relationship matrix and providing the confidence level of each relationship. Experts correct and evaluate the relationships output by the basic large language model and perform fuzzy processing on the evaluation. The importance weight calculation module is configured to construct a fuzzy direct comparison matrix, calculate the fuzzy comprehensive degree value of each indicator, calculate the probability of any two fuzzy comprehensive degree values ​​of the same expert and determine the initial weight vector of each indicator based on importance, calculate the priority weight vector of each indicator and perform consistency verification. The mutual influence weight calculation module is configured to construct a fuzzy direct influence matrix, decompose and standardize the fuzzy direct influence matrix to obtain a standardized matrix; calculate the comprehensive influence matrix of the lower bound, median and upper bound to obtain a fuzzy comprehensive influence matrix; calculate the influence degree, influenced degree, centrality and causation degree of the fuzzy comprehensive influence matrix; process the causation degree to obtain the influence weight vector; calculate the comprehensive influence weight of all experts on the indicator to obtain the influence weight vector of each indicator. Select the optimal template module and configure it to fuse the importance weight vector obtained by the Fuzzy-ANP method with the influence weight vector obtained by the Fuzzy-DEMATEL method to obtain the comprehensive weight vector of each indicator; calculate the spatial distance between the target scene and the template scene and select the optimal template.

[0040] In constructing the evaluation index system, traditional methods rely heavily on expert experience, resulting in highly subjective index extraction and difficulty in dynamic updates. This invention, however, introduces a large language model to deeply mine and learn from massive amounts of professional text (including technical specifications, maintenance reports, academic papers, expert interview records, and other unstructured data), achieving automated generation of the initial index relationship matrix. This provides experts with reference relationships with a confidence level as high as 0.82 (as shown in the embodiment). This process significantly reduces the subjectivity and cognitive burden of manually constructing the index system, making the establishment of the index system more objective and scientific, while also possessing dynamic updating capabilities, continuously optimizing and improving with the accumulation of domain knowledge.

[0041] Regarding the calculation of indicator weights, existing technologies often employ a single weighting method, making it difficult to simultaneously consider both the importance of the indicators themselves and the mutual influence between them. This invention innovatively combines Fuzzy Analytic Hierarchy Process (FuzzyANP) with Fuzzy Dematel (Fuzzy Decision Experimentation and Evaluation Laboratory) to comprehensively calculate the weights of each evaluation indicator from two dimensions: "individual importance of the indicator" and "mutual influence between indicators," resulting in more comprehensive and accurate weight calculations. As shown in the examples, the F-test (all P-values ​​> 0.05) verified the high consistency of the expert evaluation results (consistency coefficient reaching 0.89), significantly outperforming the traditional expert scoring method (consistency coefficient 0.67). In terms of automation and intelligence in the evaluation process, this invention constructs a three-level assessment process: "single damage → single-type damage to exterior wall surface → comprehensive damage to exterior wall surface," deeply integrated with UAV detection data. Compared to the traditional method requiring a 3-day evaluation cycle, this invention's method takes only 1 day from data collection to template selection, significantly improving evaluation efficiency.

[0042] This evaluation system incorporates an optimal template matching mechanism. By calculating the spatial distance between the target scene and multiple template scenes on the comprehensive weight vector of indicators, it automatically selects the most similar template scene as the evaluation reference. In the example, the template scene G2 (spatial distance 0.064) that best matches the target scene was successfully identified, and a targeted maintenance plan was formulated based on this. After the plan was implemented, no new serious damage was found on the exterior wall of the target scene during the subsequent year of monitoring, verifying the accuracy and reliability of the evaluation results.

[0043] From the perspective of practical application, this evaluation system significantly shortens the evaluation cycle, reduces the time and manpower costs of expert participation, and improves the consistency and reliability of evaluation results by using the auxiliary construction of a large language model and the systematic integration of fuzzy multi-criteria decision-making methods. It provides strong technical support for the safety risk management of urban building facades and has good engineering promotion value and economic and social benefits.

[0044] In one example of the present invention, the intelligent auxiliary construction module includes: An expert knowledge base unit is constructed and configured to collect unstructured text data related to building exterior wall inspection, including technical specifications, maintenance reports, academic papers, and expert interview records. This data is then cleaned, deduplicated, and formatted to form a domain knowledge base. The basic large language model fine-tuning unit is configured to select a basic large language model and fine-tune it using a domain knowledge base. The goal of fine-tuning is to enable the model to understand and generate professional content related to building exterior wall damage risk assessment, particularly to accurately identify potential relationships between indicators. Supervised learning is employed during the fine-tuning process, where the input is a text fragment containing indicator descriptions, and the output is the relationship labels between indicators. The initial relation matrix generation unit is configured to input the indicator system and the definitions of each indicator into the fine-tuned basic large language model. By designing specific prompts, such as "Please analyze whether indicator A (e.g., crack width) has a direct impact on indicator B (e.g., hollow area) in the risk assessment of building exterior wall damage? Is the impact positive or negative?", the basic large language model analyzes the potential relationships between indicators pairwise. Based on its knowledge learned from professional texts, the LLM outputs an initial indicator relation matrix and provides the confidence level for each relationship. This process can automatically identify a large number of potential and complex indicator associations, providing a scientific reference basis for subsequent expert scoring.

[0045] The expert correction and fuzzification unit is configured to provide the initial relation matrix and confidence results generated by the basic large language model to the expert group. Experts, based on their own experience, review, correct, and confirm the relations output by the basic large language model. For the relations between indicators confirmed by the experts, the Fuzzy-DEMATEL linguistic variables in Table 2 are used to make a final evaluation of the strength of the relations. At the same time, the experts use Fuzzy-ANP linguistic variables to compare the importance of all indicators pairwise. Finally, all the evaluation languages ​​of the experts are converted into corresponding triangular fuzzy numbers according to Table 2 to construct the fuzzy direct comparison matrix and the fuzzy direct influence matrix.

[0046] It should be noted that the building exterior wall damage potential risk assessment system of the present invention can also perform any of the processes described in the previously described building exterior wall damage potential risk assessment method, the specific details of which will not be repeated here.

[0047] Specific examples: In this embodiment, the exterior wall of an old residential building in a city is taken as the evaluation object (hereinafter referred to as target scenario O). Five typical building exterior wall scenarios (hereinafter referred to as template scenarios G1~G5) with established risk assessment templates are selected as candidate templates. The method of this invention is used to evaluate the potential risk of exterior wall damage in the target scenario and select the optimal reference template.

[0048] Based on industry standards such as the *Technical Specification for Waterproofing Engineering of Building Exterior Walls* (JGJ / T 235-2011) and the *Technical Specification for Detection and Treatment of Building Cracks* (CECS 293:2011), and combined with expert interviews and literature reviews, a three-level assessment process of "single damage → single type of damage to exterior wall surface → comprehensive damage to exterior wall surface" was adopted, establishing an evaluation index system as shown in Table 1. This system covers four major defect types: cracks, hollow areas, leaks, and peeling. Each defect type has specific secondary indicators and is divided into five risk levels (none, low, medium, high, and extremely high). For ease of subsequent matrix representation, each secondary indicator is assigned a unique code C01~C22, as shown in Table 3.

[0049] Table 3 shows the evaluation index system. Step S1: Data Collection Step S1.1 Select the expert group No fewer than 10 experts in building safety testing, urban renewal, and structural engineering were invited to participate in the research. These experts came from government quality inspection departments, architectural design institutes, testing organizations, and university research institutions, and all had more than ten years of professional experience. The experts were randomly divided into two groups, A and B, with equal numbers in each group, for subsequent data consistency verification.

[0050] Step S1.2: Define the fuzzy linguistic scale To address the ambiguity and uncertainty in the expert evaluation process, triangular fuzzy quantitative linguistic variables are employed. For both the Fuzzy-ANP and Fuzzy-DEMATEL methods, corresponding sets of linguistic variables are established, as shown in Table 4. The Fuzzy-ANP linguistic variables include "equally important," "slightly important," "relatively important," "very important," and "absolutely important," while the Fuzzy-DEMATEL linguistic variables include "no impact," "weak impact," "impact," "strong impact," and "extremely strong impact." Each linguistic variable corresponds to a set of triangular fuzzy numbers and their reciprocals.

[0051] Table 4. Fuzzy linguistic variables and their corresponding triangular fuzzy numbers Step S1.3 Design and distribute questionnaires Design an independent questionnaire for each of the target scenario O and the five template scenarios G1~G5. Each questionnaire consists of three parts: the first part is a description of the basic situation of the scenario (including building type, years of use, environment, etc.); the second part is a direct comparison matrix C of the importance of indicators (for Fuzzy-ANP analysis); and the third part is a matrix R of the direct influence relationship between indicators (for Fuzzy-DEMATEL analysis).

[0052] The designed questionnaire was distributed to all invited experts. Experts independently completed matrices C and R based on their professional knowledge and the descriptions of each scenario. All questionnaires were collected on-site, numbered, and organized to ensure data integrity and traceability. In this embodiment, the calculation process is illustrated using the evaluation data of the first expert on target scenario O as an example. Part of the matrices they completed are shown in Tables 5 and 6 (only the first five indicators C01~C05 are listed; the remaining indicators are similar).

[0053] Table 5. Fuzzy direct comparison matrix of the first expert on the target scene O (partial, C01~C05) Table 6. Matrix of direct fuzzy influence of the first expert on target scene O (partial, C01~C05) Step S2: LLM Model Training and Assisted Modeling Step S2.1 Constructing an expert knowledge base We collected a large amount of unstructured text data related to building exterior wall inspection, including technical specifications, maintenance reports, academic papers, and expert interview records. We then cleaned, deduplicated, and formatted the data to create a high-quality domain knowledge base containing a wealth of information.

[0054] Step S2.2 LLM fine-tuning A lightweight, computationally efficient, and well-designed basic language model with moderate parameters was selected as the foundation, and fine-tuned using the aforementioned domain knowledge base. The fine-tuning employed a supervised learning approach, with input being text fragments containing indicator descriptions and output being labels representing the relationships between indicators. By constructing a sufficient number of high-quality labeled samples for training and validation, the model was ensured to maintain stable high performance in the indicator relationship recognition task, achieving an accuracy of over 85%. This significantly reduced the computational resource requirements for model deployment and inference while maintaining recognition accuracy. In this embodiment, training and validation samples of a suitable scale were constructed, and the model accuracy remained unchanged after fine-tuning.

[0055] Step S2.3 Initial Relationship Matrix Generation The indicator system and definitions of each indicator were input into the fine-tuned LLM. Using the prompt "Please analyze whether indicator A has a direct impact on indicator B in the risk assessment of building exterior wall damage? Is the impact positive or negative? Please provide a confidence level (0-1)," the potential relationships between indicators were analyzed pair by pair. The LLM output an initial relationship matrix and confidence levels, identifying over a hundred pairs of potential indicator relationships with an average confidence level exceeding 0.8. In this embodiment, several pairs of potential relationships were identified, with an average confidence level of 0.82. Some results are shown in Table 7.

[0056] Table 7. Initial relation matrix generated by LLM (partial) Step S2.4 Expert Correction and Blurring Processing The initial relation matrix and confidence levels generated by the LLM were provided to the expert panel. The expert panel, drawing on their experience, reviewed the relations output by the LLM, confirming valid relation pairs and removing those with low confidence or that did not conform to engineering realities. Subsequently, experts were invited to use the linguistic variables in Table 2 to evaluate the strength of the relationships between the confirmed indicators, and to perform pairwise comparisons of indicator importance. The expert evaluation language was converted into triangular fuzzy numbers according to Table 2, and fuzzy direct comparison matrices and fuzzy direct influence matrices were constructed, as shown in Tables 5 and 6.

[0057] Step S3: Calculation of indicator importance weights based on Fuzzy-ANP Step S3.1 Construct a fuzzy direct comparison matrix and verify consistency. The evaluation results of the first expert on the target scenario O are organized into a fuzzy direct comparison matrix with an order of n=22 (corresponding to 22 secondary indicators C01~C22). After defuzzifying the fuzzy matrix, the clear matrix is ​​obtained. The largest eigenvalue and the consistency index are calculated. The random consistency index is obtained by looking up the table. The consistency ratio is calculated and the consistency test is passed.

[0058] Step S3.2 Calculate the fuzzy comprehensive degree value The fuzzy comprehensive degree value of each indicator is calculated using the extended analysis method according to the formula. Taking C01 as an example: Similarly, the fuzzy comprehensive degree values ​​of the other indicators were calculated, and some results are shown in Table 8.

[0059] Table 8. Partial fuzzy comprehensive values ​​of the first expert for the target scene O. S3.3 Calculate the probability and determine the weight vector Calculate the probability of pairwise comparison of the fuzzy comprehensive degree values ​​of each indicator. Taking C01 and C02 as an example, the median value of C01 is 0.0544 and the median value of C02 is 0.0472. Since the median value of C01 is greater than the median value of C02, the probability that C01 is greater than or equal to C02 is 1.

[0060] Taking C01 and C03 as examples, the median value of C01 is 0.0544, and the median value of C03 is 0.0618. Since the median value of C01 is less than the median value of C03, further calculation is needed. The upper bound of C01 is 0.1080, and the lower bound of C03 is 0.0347. Substituting these values ​​into the formula, we can calculate: (0.1080 - 0.0347) / [(0.0618 - 0.0347) + (0.1080 - 0.0544)] = 0.0733 / (0.0271 + 0.0536) = 0.0733 / 0.0807 = 0.908 Calculate all possible degrees and obtain the possible degree matrix as shown in Table 9 (partial).

[0061] Table 9. Partial probability matrix V of the first expert regarding target scenario O Calculate the priority probability of C01, and take the minimum value of all probability probabilities of the row containing C01, i.e., min(1.000, 0.908, 0.842, 1.000, ...) = 0.842.

[0062] Similarly, the priority probability of other indicators is calculated, and after normalization, the importance weight vector of the first expert based on Fuzzy-ANP is obtained, as shown in Table 10.

[0063] Table 10. Fuzzy-ANP Importance Weights of the First Expert for Target Scenario O (Partial) S3.4 Integration of Expert Opinions and Consistency Verification Repeat the above steps to process all expert data, obtaining multiple importance weight vectors. Experts are randomly divided into two groups, A and B, and an F-test is performed on the weights of each indicator. Taking C01 as an example, group A has a mean of 0.0428 and a variance of 0.00011; group B has a mean of 0.0442 and a variance of 0.00013. The calculated F-test result is 0.85, and the p-value is 0.37 > 0.05, indicating no significant difference between the two groups and good consistency in expert evaluation. The F-test results for all indicators are shown in Table 11.

[0064] Table 11. Results of Fuzzy-ANP expert grouping F-test (partial) The overall importance weight of several experts was calculated using the geometric mean method. First, the geometric mean of the priority probability of several experts for each indicator was calculated, and then normalized to obtain the overall importance weight. The Fuzzy-ANP importance weight vector of the target scenario O is shown in Table 12 (partial).

[0065] Table 12. Fuzzy-ANP Comprehensive Importance Weights for Target Scenario O (Partial) Similarly, the expert data for the five template scenarios were processed to obtain the Fuzzy-ANP weight vectors for each template scenario. Some results are shown in Table 13.

[0066] Table 13. Fuzzy-ANP Comprehensive Importance Weights for Each Template Scenario (Partial) Step S4: Calculation of the mutual influence weights of indicators based on Fuzzy-DEMATEL S4.1 Constructing the fuzzy direct influence matrix Taking the first expert's evaluation of the target scenario O as an example, a fuzzy direct influence matrix is ​​constructed, as shown in Table 4.

[0067] S4.2 Decomposition and Standardization The fuzzy direct influence matrix is ​​decomposed into a lower bound matrix, a median matrix, and an upper bound matrix. The maximum value of the row sums is calculated: the maximum value of the row sums of the lower bound matrix is ​​11.8, the maximum value of the row sums of the median matrix is ​​17.5, and the maximum value of the row sums of the upper bound matrix is ​​24.3.

[0068] The three matrices were standardized by dividing each element by the sum of the rows of the corresponding matrix and the maximum value. The standardized median matrices are shown in Table 14.

[0069] Table 14 Standardized Direct Impact Matrix (Median, C01~C05) S4.3 Calculate the fuzzy comprehensive influence matrix Based on the three standardized matrices, the comprehensive influence matrices for the lower bound, median, and upper bound are calculated respectively. The calculation formula is: standardized matrix multiplied by identity matrix minus the inverse of the standardized matrix. The fuzzy comprehensive influence matrix is ​​obtained, and its median is shown in Table 15.

[0070] Table 15 Fuzzy Comprehensive Influence Matrix (Median, C01~C05) S4.4 Calculate influence, degree of influence, centrality, and degree of causation. For the fuzzy comprehensive influence matrix, the sum of each row is used to obtain the influence degree, and the sum of each column is used to obtain the affected degree. The centroid method is used to fuzzify the fuzzy numerical solution, that is, to take the average of the sums of corresponding elements of the three matrices.

[0071] The influence degree D and the degree of being influenced R of each indicator were calculated. Then, the centrality M=D+R and the causality N=DR were calculated. The results are shown in Table 16.

[0072] Table 16. The influence, degree of influence, centrality, and causation of the first expert on the target scenario O (partial). S4.5 Calculate the weight vector based on influence relationships. The causality is shifted. First, find the minimum causality value of -0.64 (C05). Then, the shifted value of each indicator is: the original causality value minus the minimum value plus 1.

[0073] Taking C01 as an example: 0.47 - (-0.64) + 1 = 0.47 + 0.64 + 1 = 2.11 Similarly, the shifted values ​​of other indicators were calculated and normalized to obtain the first expert's influence weight vector based on Fuzzy-DEMATEL, as shown in Table 17.

[0074] Table 17. Fuzzy-DEMATEL Influence Weights of the First Expert on Target Scenario O (Partial) Repeat the above steps, and after verifying the consensus of experts by the F test (Table 18), the Fuzzy-DEMATEL influence weight vector of the target scenario O is obtained. Some results are shown in Table 19.

[0075] Table 18. Results of the Fuzzy-DEMATEL expert grouping F-test (partial) Table 19. Fuzzy-DEMATEL Combined Influence Weights of Target Scenario O (Partial) Similarly, the expert data for the five template scenarios were processed to obtain the Fuzzy-DEMATEL weight vectors for each template scenario. Some results are shown in Table 20.

[0076] Table 20. Fuzzy-DEMATEL Overall Impact Weights for Each Template Scenario (Partial) Step S5: Optimal Template Selection S5.1 Calculate the composite weight vector The Fuzzy-ANP importance weight and the Fuzzy-DEMATEL influence weight are combined to obtain the comprehensive weight of each indicator. The calculation formula is: Comprehensive weight = sqrt(importance weight × influence weight).

[0077] The comprehensive weight vectors of the target scenario O and each template scenario are shown in Table 21 (only the first 10 indicators C01~C10 are listed due to space limitations).

[0078] Table 21 Comprehensive Weight Vectors of Target Scene and Template Scene (Partial, C01~C10) S5.2 Calculate spatial distance and select the optimal template The spatial distance between the target scene O and each template scene is calculated using the formula: Spatial distance = Sum of squares of the differences in the comprehensive weights of all indicators / Total number of indicators. The calculation results in this embodiment are shown in Table 22.

[0079] Table 22 Spatial distance between the target scene and each template scene The calculation results show that the target scene has the smallest distance to G2 (0.064), the second smallest distance to G5 (0.076), and the largest distance to G3 (0.109). Therefore, template scene G2 is selected as the optimal template, and G5 can be used as an alternative reference template.

[0080] Technical effect verification and application analysis To verify the technical effectiveness of this invention, the evaluation results of this embodiment were compared with those of the traditional expert scoring method. Based on the optimal template selection results, target scenario O can refer to the risk assessment results of template scenario G2 (see Table 23) to formulate a targeted maintenance plan (see Table 24). After the implementation of the plan, no new serious damage was found on the exterior wall of the target scenario during the subsequent year of monitoring, proving that the evaluation results are accurate and reliable.

[0081] Table 23 Risk Assessment Results and Maintenance Strategies for Template Scenario G2 Table 24 Maintenance Recommendations for Target Scenario O In summary, this invention effectively solves the problems of strong subjectivity, poor consistency, and low automation of traditional methods by introducing a large language model to assist in the construction of an index relationship matrix, integrating Fuzzy-ANP and Fuzzy-DEMATEL to calculate index weights from two dimensions, and using a template matching method to select the optimal reference template. This provides a scientific and efficient technical solution for assessing the potential risks of damage to building exterior walls.

[0082] The foregoing description, with reference to preferred embodiments, details an exemplary implementation of the method and system for assessing potential risks of damage to building exterior walls proposed by the present invention. However, those skilled in the art will understand that various modifications and alterations can be made to the above specific embodiments without departing from the concept of the present invention, and various combinations can be made to the various technical features and structures proposed by the present invention without exceeding the protection scope of the present invention, which is determined by the appended claims.

Claims

1. A method for assessing the potential risk of damage to building exterior walls, characterized in that, Includes the following steps: S10: Construct an evaluation index system for potential risks of building exterior wall damage: Based on building industry standards, expert interviews and literature reviews, adopt a three-level assessment process of single damage → single type of damage to exterior wall surface → comprehensive damage to exterior wall surface, and establish an evaluation index system. S20: Intelligent Assisted Construction of Indicator Relationships Based on Large Language Models: Construct an expert knowledge base and use the expert knowledge base to fine-tune the selected basic large language model. Input the evaluation index system into the fine-tuned basic large language model to output the initial index relationship matrix and give the confidence level of each relationship. Experts correct and evaluate the relationships output by the basic large language model, and the evaluation is then fuzzyened. S30: Calculation of indicator importance weights based on Fuzzy-ANP: Construct a fuzzy direct comparison matrix, calculate the fuzzy comprehensive degree value of each indicator, calculate the probability of any two fuzzy comprehensive degree values ​​of the same expert and determine the initial weight vector of each indicator based on importance, calculate the priority weight vector of each indicator and perform consistency verification. S40: Calculation of the mutual influence weights of indicators based on the Fuzzy-DEMATEL laboratory analysis method: Construct a fuzzy direct influence matrix, and decompose and standardize the fuzzy direct influence matrix to obtain a standardized matrix. The fuzzy comprehensive influence matrix is ​​obtained by calculating the comprehensive influence matrix of the lower bound, median, and upper bound respectively. The influence degree, affected degree, centrality, and causal degree of the fuzzy comprehensive influence matrix are then calculated. The influence weight vector is obtained by processing the causal degree, and the influence weight vector of each indicator is obtained by calculating the comprehensive influence weight of all experts on the indicator. S50: Select the optimal template: Combine the importance weight vector obtained by the Fuzzy-ANP method with the influence weight vector obtained by the Fuzzy-DEMATEL method to obtain the comprehensive weight vector of each indicator; calculate the spatial distance between the target scene and the template scene and select the optimal template.

2. The method for assessing potential risks of damage to building exterior walls according to claim 1, characterized in that, Step S10 specifically includes the following steps: S11: Develop a fuzzy language scale: Use triangular fuzzy numbers to quantify expert language evaluations, where the triangular fuzzy number is represented as... , where l, m and u represent the minimum possible value, the most likely value and the maximum possible value of the fuzzy event, respectively; S12: Design an independent questionnaire for each building exterior wall scenario to be evaluated. The building exterior wall scenarios include the target scenario and multiple candidate template scenarios. The questionnaire consists of three parts: the first part is a basic description of the exterior wall to be evaluated, corresponding to the specific background of the research question; the second part is the direct comparison matrix C of the importance between indicators, which is used for subsequent Fuzzy-ANP analysis; the third part is the direct influence relationship matrix R between indicators, which is used for subsequent Fuzzy-DEMATEL laboratory analysis.

3. The method for assessing potential risks of damage to building exterior walls according to claim 1, characterized in that, Step S20 specifically includes the following steps: S21: Construct an expert knowledge base: Collect unstructured text data related to building exterior wall inspection, including technical specifications, maintenance reports, academic papers, and expert interview records, and perform cleaning, deduplication, and formatting to form a domain knowledge base; S22: Fine-tuning of the basic large language model: Select a basic large language model and fine-tune it using a domain knowledge base; the input is a text fragment containing indicator descriptions, and the output is the relational labels between indicators; S23: Initial Relationship Matrix Generation: The indicator system and the definition of each indicator are input into the fine-tuned basic large language model. By designing specific prompt words, the basic large language model analyzes the potential relationships between indicators one by one. S24: Expert Correction and Fuzziness Processing: The initial relation matrix and confidence results generated by the basic large language model are provided to the expert group. Experts review, correct, and confirm the relations output by the basic large language model based on their own experience. For the relations between indicators confirmed by the experts, Fuzzy-DEMATEL linguistic variables are used to make a final evaluation of the strength of the relations. At the same time, experts use Fuzzy-ANP linguistic variables to compare the importance of all indicators pairwise. Finally, all the evaluation languages ​​of the experts are converted into corresponding triangular fuzzy numbers to construct a fuzzy direct comparison matrix and a fuzzy direct influence matrix.

4. The method for assessing potential risks of damage to building exterior walls according to claim 1, characterized in that, Step S30 specifically includes the following steps: S31: Constructing a fuzzy direct comparison matrix: Organize the fuzzy evaluation results of the k-th expert on the importance of the indicators into a fuzzy direct comparison matrix. ; Elements in the array This indicates the importance of the i-th indicator relative to the j-th indicator. The result is determined by experts based on preset linguistic variables and converted into a corresponding triangular fuzzy number. Among them, the fuzzy direct comparison matrix The expression is as follows: in, For triangular fuzzy numbers; S32: Calculate the fuzzy generalization degree value: The fuzzy generalization degree value of each index i is calculated using the extended analysis method; the expression for calculating the fuzzy generalization degree value is: S33: Calculate the likelihood and determine the weight vector: First, for any two fuzzy comprehensive degree values ​​of the same expert k and ,calculate The probability degree; this probability degree is defined as the ordinate of the highest point of the intersection of the membership functions of two fuzzy numbers, that is: This expression is equivalent to the following piecewise function: Among them, when When the degree is 1, it corresponds to the case where the ordinate of the highest point of the intersection of the membership functions of the two fuzzy numbers is 1. Then, calculate the weight vector of expert k, a fuzzy generalization degree value. Greater than all others The degree of possibility of (j≠i) is defined as: right Normalization is performed to obtain the initial weight vector of each indicator based on importance given by the k-th expert. S34: Calculate the priority weight vector and perform consistency verification: First, for the k-th expert, calculate the fuzzy comprehensive degree value for each index i. Greater than all other indicators The probability of (j≠i), i.e.: Then, to After normalization, we obtain the relative importance weight vector of each indicator given by the k-th expert using the fuzzy ANP method: Repeat the above steps to obtain the weight vector of all experts; randomly divide the evaluation results of all experts into two groups with an equal number of people in each group, and use the F test to verify the validity of the data; if the significance level P value is greater than 0.05, it indicates that there is no significant difference between the two groups of data and the expert evaluation results have a high degree of consistency. S35: Calculate the priority weight vector: Based on step S34, calculate the comprehensive weight of all experts for indicator i. The geometric mean method is used to express it as follows: Finally, the overall weights are normalized to obtain the final weights of each indicator. Its expression is: This yields the priority weight vector for this evaluation scenario. .

5. The method for assessing potential risks of damage to building exterior walls according to claim 4, characterized in that, Step S31 further includes: performing a consistency check on the matrix, firstly defuzzifying the fuzzy matrix to obtain a clear judgment matrix. Then calculate the consistency ratio. Its expression is as follows: The formula for calculating the consistency index (CI) is as follows: In the formula, For matrix The largest eigenvalue, where n is the matrix order; RI is the random consistency index, which is determined by looking up a table based on the matrix order. If CR ≤ 0.1, the matrix is ​​considered to have passed the consistency test. If CR > 0.1, feedback to experts is required for correction.

6. The method for assessing potential risks of damage to building exterior walls according to claim 1, characterized in that, Step S40 specifically includes the following steps: S41: Constructing a fuzzy direct influence matrix: The k-th expert, based on the actual condition of the building's exterior wall, determines the interrelationships between various indicators; expert evaluation values. This indicates the degree of direct influence of indicator i on indicator j, and is filled in the matrix. At the intersection of the i-th row and j-th column; Convert to triangular fuzzy number Among them, the fuzzy direct influence matrix constructed by the k-th expert As shown in the following formula: in, For triangular fuzzy numbers; S42: The direct impact matrix of decomposition and standardized fuzzy model: The fuzziness directly affects the matrix Decomposed into a lower bound matrix Median matrix and upper bound matrix , respectively by Composition, where the lower bound matrix Median matrix and upper bound matrix The expression is: Standardize each matrix individually; the standardization factor is the maximum value of the sum of each row: The standardized matrix is ​​then: S43: Calculate the fuzzy comprehensive influence matrix: Based on the standardized matrix, the lower bound comprehensive influence matrix is ​​calculated respectively. Median Comprehensive Influence Matrix Upper bound comprehensive influence matrix The calculation formula is as follows: In the formula, I is an n×n identity matrix; This yields the fuzzy comprehensive influence matrix. ,in ; S44: Calculate influence, affectedness, centrality, and causality: Fuzzy Comprehensive Influence Matrix The influence is obtained by calculating the sum of each row. The sum of each column yields the degree of influence. Among them, influence And the degree of influence The expression is: To facilitate comparison, the centroid method is used to fuzzify the fuzzy numerical solutions into clear values: Further calculate the centrality of each indicator and causal degree : Among them, centrality Centrality reflects the importance of the indicator in the system; a higher centrality indicates a more critical indicator. Causality... If the value is positive, it indicates that the indicator is a causal factor, meaning it actively influences other indicators; if the value is negative, it is an outcome factor, meaning it is influenced by other indicators. S45: Calculate the weight vector based on influence relationships: The causal degree is transformed into a positive value index for comprehensive weighting to obtain the influence weight vector of all experts; the evaluation results of all experts are randomly divided into two groups with equal numbers of people in each group, and the F test is used to verify the validity of the data; if the p value is greater than 0.05, it indicates that there is no significant difference between the two groups of data and the expert evaluation results have a high degree of consistency. Based on this, calculate the combined influence weight of all experts on indicator i. The geometric mean method is used to express it as follows: Finally, the overall impact weights are normalized to obtain the final impact weight vectors of each indicator. Its expression is: This yields the influence weight vector for this evaluation scenario. .

7. The method for assessing potential risks of damage to building exterior walls according to claim 6, characterized in that, In step S45, the influence weight vector of all experts is obtained, including the following steps: The causality degree is converted into a positive value that can be used for comprehensive weighting. The causality degree is then shifted using the following formula: right After normalization, we obtain the influence weights of each indicator given by the k-th expert based on the Fuzzy-DEMATEL method: Repeat the above steps to obtain the influence weight vector of all experts.

8. The method for assessing potential risks of damage to building exterior walls according to claim 1, characterized in that, Step S50 specifically includes the following steps: S51: Calculate the comprehensive weight vector for each scenario: The importance weight vector obtained by the Fuzzy-ANP method and the influence weight vector obtained by the Fuzzy-DEMATEL method are fused to obtain the comprehensive weight vector of each indicator; for the building exterior wall scene to be evaluated, it is denoted as target scene O; And for each candidate template scene, denoted as template scene G, the comprehensive weight vector W is calculated using the following formula: In the formula, Let be the Fuzzy-ANP importance weight of the i-th indicator; The Fuzzy-DEMATEL influence weight for the i-th indicator; The comprehensive weight vector of the target scenario O is denoted as The comprehensive weight vector of each template scenario G is denoted as... ; S52: Calculate the spatial distance between the target scene and the template scene and select the optimal template: The target scene O and each candidate template scene are calculated using the following formula. Spatial distance between This distance reflects the difference between the target scene and the template scene in terms of the overall importance of the indicators, and its expression is: In the formula, n is the total number of evaluation indicators, and m is the number of candidate template scenarios; Wherein, the distance value d(O, The smaller the value, the closer the target scene is to the Gth template scene in terms of the overall importance of various indicators; the candidate template with the smallest distance is selected as the optimal template. Its expression is: The building exterior wall type or inspection area corresponding to the optimal template has the most similar damage risk characteristics to the target scenario. Its assessment results, risk level classification or maintenance strategy can be referenced to provide a basis for risk assessment and maintenance decisions for the target scenario.

9. A system for assessing potential risks of damage to building exterior walls, characterized in that, include: A risk assessment indicator system module was constructed, configured to establish an evaluation indicator system based on building industry standards, expert interviews, and literature reviews, using a three-level assessment process: single damage → single type of damage to exterior wall surface → comprehensive damage to exterior wall surface. The intelligent auxiliary construction module is configured to build an expert knowledge base and use the expert knowledge base to fine-tune the selected basic large language model. The evaluation index system is input into the fine-tuned basic large language model, thereby outputting the initial index relationship matrix and giving the confidence level of each relationship. Experts correct and evaluate the relationships output by the basic large language model, and the evaluation is then fuzzyened. The importance weight calculation module is configured to construct a fuzzy direct comparison matrix, calculate the fuzzy comprehensive degree value of each indicator, calculate the probability of any two fuzzy comprehensive degree values ​​of the same expert and determine the initial weight vector of each indicator based on importance, calculate the priority weight vector of each indicator and perform consistency verification. The mutual influence weight calculation module is configured to construct a fuzzy direct influence matrix, and then decompose and standardize the fuzzy direct influence matrix to obtain a standardized matrix. The fuzzy comprehensive influence matrix is ​​obtained by calculating the comprehensive influence matrix of the lower bound, median, and upper bound respectively. The influence degree, affected degree, centrality, and causal degree of the fuzzy comprehensive influence matrix are then calculated. The influence weight vector is obtained by processing the causal degree, and the influence weight vector of each indicator is obtained by calculating the comprehensive influence weight of all experts on the indicator. Select the optimal template module and configure it to fuse the importance weight vector obtained by the Fuzzy-ANP method with the influence weight vector obtained by the Fuzzy-DEMATEL method to obtain the comprehensive weight vector of each indicator; calculate the spatial distance between the target scene and the template scene and select the optimal template.

10. The building exterior wall damage potential risk assessment system according to claim 9, characterized in that, The intelligent auxiliary construction module includes: An expert knowledge base unit is constructed and configured to collect unstructured text data related to building exterior wall inspection, including technical specifications, maintenance reports, academic papers, and expert interview records. This data is then cleaned, deduplicated, and formatted to form a domain knowledge base. The basic large language model fine-tuning unit is configured to select a basic large language model and fine-tune it using a domain knowledge base; the input is a text fragment containing indicator descriptions, and the output is the relational labels between indicators. The initial relation matrix generation unit is configured to input the indicator system and the definition of each indicator into the fine-tuned basic large language model. By designing specific prompt words, the basic large language model can analyze the potential relationships between indicators one by one. The expert correction and fuzzification unit is configured to provide the initial relation matrix and confidence results generated by the basic large language model to the expert group. Experts, based on their own experience, review, correct, and confirm the relations output by the basic large language model. For the relations between indicators confirmed by the experts, Fuzzy-DEMATEL linguistic variables are used to make a final evaluation of the strength of the relations. At the same time, experts use Fuzzy-ANP linguistic variables to compare the importance of all indicators pairwise. Finally, all the evaluation languages ​​of the experts are converted into corresponding triangular fuzzy numbers to construct a fuzzy direct comparison matrix and a fuzzy direct influence matrix.

Images