Evaluation index system construction method and system of emergency rescue intelligent command decision-making system

Abstract

The invention provides an emergency rescue intelligent command decision system evaluation index system construction method and system, and belongs to the technical field of artificial intelligence. The method objectively screens out the key evaluation indexes in the emergency rescue intelligent command decision-making system through the interpretability analysis of the SHAP model, the measurement decision-making result and the influence degree of the evaluation indexes by the data characteristics, thereby revealing the decision-making mechanism in the model and the importance degree of each evaluation index, and meanwhile, adopting a Bayesian causal reasoning method to improve the evaluation accuracy of the emergency rescue intelligent command decision-making system. And constructing a causal relationship model among the evaluation indexes, and quantifying the causal influence among the evaluation indexes. According to the method, the problem that a black box model is difficult to explain is solved, the orthogonal independence between evaluation indexes is improved, the rationality and effectiveness of an evaluation index system are ensured, and the defect that a causal relationship cannot be revealed through traditional correlation analysis is overcome.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence, and particularly relates to a method and system for constructing an evaluation index system for an intelligent command and decision-making system for emergency rescue. Background Art

[0002] In recent years, with the rapid development of artificial intelligence technology, intelligent decision-making systems have been widely applied in various fields, especially playing an important role in emergency rescue. Constructing a scientific and reasonable evaluation index system and conducting comprehensive and objective evaluations are crucial for ensuring the normal and effective operation of intelligent decision-making systems. However, due to the complexity, diversity, and uncertainty of intelligent decision-making systems (the input of the emergency rescue intelligent command and decision-making system is not homogeneous, and the evaluation is multi-dimensional), the construction of the evaluation index system faces many challenges. With the rapid development of machine learning technology, the "data-driven" automatic construction method of the evaluation index system for intelligent decision-making systems has become the mainstream.

[0003] The data-driven automatic evaluation index system construction method can customize and generate evaluation indexes according to the specific requirements and application fields of intelligent decision-making systems. For example, in the field of emergency rescue, evaluation indexes such as response speed, resource scheduling efficiency, rescue success rate, and collaborative command ability can be automatically extracted. Through this method, the problems of high cost, low efficiency, interference of human subjective factors in traditional construction methods, and the difficulty of accurately and timely reflecting the performance changes of the emergency rescue intelligent command and decision-making system due to easy omission of key indexes are effectively solved. Therefore, driven by massive data, automatically screening key indexes from the data set can generate an objective and comprehensive evaluation index system. This can not only ensure the comprehensiveness and accuracy of the index system but also significantly reduce the cost of designing indexes and adapt to the dynamic changes of the emergency rescue intelligent command and decision-making system.

[0004] The evaluation index system refers to an organic whole composed of a series of indexes that can comprehensively measure and reflect the characteristics of the evaluated object under the guidance of a specific evaluation goal, based on scientific and reasonable evaluation principles. The evaluation index system construction method refers to the whole process of determining the selection of evaluation indexes, weight assignment, and evaluation standard setting according to the evaluation purpose and requirements, using scientific methods and procedures. Its purpose is to establish an objective, fair, and scientific index system to ensure the accuracy and credibility of evaluation results. The evaluation index system construction method for the emergency rescue intelligent command and decision-making system is different from that of other intelligent systems. The emergency rescue intelligent command and decision-making system faces a large number of uncertain decision-making factors, involves complex decision-making strategies, and has an opaque algorithm structure. Its evaluation index system needs to deeply consider aspects such as the intelligent characteristics, decision-making effects, learning ability, interpretability, and robustness of the system. Therefore, its construction method needs to pay more attention to the internal working mechanism and decision-making logic of the system.

[0005] Traditional methods for constructing evaluation index systems include the analytic hierarchy process, Delphi method, fuzzy comprehensive evaluation method, grey relational analysis method, entropy value method, and expert scoring method, etc. These methods use a combination of qualitative and quantitative approaches to construct evaluation indicators and determine indicator weights, and usually rely on expert experience and subjective judgment. However, traditional methods have many problems, such as strong subjectivity of evaluation indicators, lack of adaptability to dynamic environments, and difficulty in reflecting the complex internal operation mechanism of the emergency rescue intelligent command and decision-making system. Since traditional methods often lack an in-depth understanding of the internal working mechanism of the model and cannot accurately identify the key factors affecting decision-making results, the evaluation results are not comprehensive and accurate enough, and it is difficult to objectively and comprehensively reflect the effectiveness of the emergency rescue intelligent command and decision-making system. Currently, there are some studies on constructing evaluation index systems driven by data, as follows: In Chinese patent application document CN118114853A, a method and device for dividing an effectiveness evaluation index system based on correlation analysis are disclosed, and a method for dividing an effectiveness evaluation index system applied to pipeline integrity management is disclosed. This method calculates the correlation coefficient of each element in the index system of the level to be divided, establishes a corresponding relationship table for the correlation analysis of the index system, and determines the indicator weights according to the elements in the relationship table. Finally, the effectiveness evaluation index system is divided according to the determined indicator weights. Specifically, this method first calculates the correlation coefficient of each element in the index system of the level to be divided, and this step is to quantify the correlation degree between each element. Then, a corresponding relationship table for correlation analysis is established according to the correlation coefficient of each element, and this table can intuitively show the correlation situation between each element. However, the correlation analysis method in this solution only considers the correlation and co-occurrence among many variables, and does not really solve the problem of depicting and measuring the causal relationship and influence degree of variables, and cannot objectively and accurately screen out reasonable evaluation indicators.

[0006] In Chinese Patent Application Document CN107563596A, a method for analyzing the equilibrium state of evaluation indicators based on Bayesian causal network is disclosed. This method first obtains relevant information of the system, establishes an evaluation index system of the system, and determines the exogenous factors affecting the system; corresponding endogenous variable sets are obtained through the evaluation indicators, an exogenous input variable set is obtained from the exogenous influencing factors of the system, and an output variable set is obtained based on the evaluation results. Then, a three-layer Bayesian causal network structure is constructed according to the endogenous variable set, the exogenous input variable set, and the output variable set, and conditional independence tests are used to discover the causal relationships between variables. Next, based on the Bayesian causal network structure and the causal relationships between variables, system dynamics modeling is performed on the system, and the equilibrium states of each variable are obtained through simulation calculations. Finally, the equilibrium states of each variable are mapped to the evaluation indicators to obtain the equilibrium states of each evaluation indicator under exogenous condition constraints. However, this invention only considers the causal independence between candidate variables, does not consider the influence degree of data characteristics on evaluation indicators, and does not construct an index system.

[0007] In Chinese Patent Application Document CN117291463A, a comprehensive evaluation method and device for the interpretability technology of intelligent models are disclosed, which are used to evaluate the interpretability technology of intelligent models. This method mainly includes the following steps: First, an initial model is constructed according to the task requirements, and appropriate interpretability technologies are selected to explain the model results. Next, multi-view evaluation is carried out, and indicators are calculated for the selected interpretability technologies and their interpretation results to quantify their performance. Then, subjective evaluation is carried out, user feedback is collected through questionnaires, and the questionnaire results are converted into data indicators by combining the analytic hierarchy process and the fuzzy comprehensive evaluation method. Finally, the subjective and objective evaluation results are summarized to obtain the final comprehensive evaluation result. The innovation of this method lies in combining the analytic hierarchy process and the fuzzy comprehensive evaluation method to convert the subjective questionnaire results into quantifiable measurement indicators, so as to obtain comprehensive subjective and objective evaluation results. This method not only provides a clear comprehensive evaluation process for the interpretability technology of intelligent models, but also presents the subjective and objective evaluation results in data form, making the evaluation results more intuitive and easy to understand. However, this solution still requires the design and implementation of questionnaires, which takes a lot of time and resources, and the accuracy and representativeness of the questionnaire results may be affected by the subjective biases of the respondents. In addition, the application of the analytic hierarchy process and the fuzzy comprehensive evaluation method requires certain professional knowledge and experience, which may pose relatively high requirements for the evaluators. It is used to evaluate the model itself and does not construct an evaluation index system.

[0008] It can be seen from this that there are still several limitations in the existing methods. First, from the perspective of the emergency rescue intelligent command and decision-making system, there is a lack of measurement of the influence degree of data on evaluation indicators. Second, from the perspective of the independence of evaluation indicators, there is a lack of reasonable measurement of the influence degree of the causal relationship of evaluation indicators and the index system structure. Summary of the Invention

[0009] In view of the problems existing in the above-mentioned prior art, the present invention provides a method and system for constructing an evaluation index system for an emergency rescue intelligent command and decision-making system. The evaluation index system is constructed based on the causal relationship of model interpretability. This method analyzes the interpretability of the SHAP (SHapley Additive exPlanations) model to measure the degree to which the decision results and their evaluation indexes are affected by data features, objectively screens out the key evaluation indexes in the emergency rescue intelligent command and decision-making system, thereby revealing the internal decision-making mechanism of the model and the importance of each evaluation index, and overcoming the problem that the black-box model is difficult to explain. At the same time, the Bayesian causal reasoning method is adopted to construct a causal relationship model between evaluation indexes, quantify the causal influence between each evaluation index, improve the orthogonal independence between evaluation indexes, ensure the rationality and effectiveness of the evaluation index system, and make up for the deficiency that traditional correlation analysis cannot reveal causal relationships. By combining model interpretability and causal relationship analysis, the present invention reduces the workload of constructing the evaluation index system, improves the credibility of the evaluation results, and can obtain a high-quality, scientific, reasonable and comprehensive evaluation index system for the emergency rescue intelligent command and decision-making system. At the same time, it effectively solves the problems of difficult selection of evaluation indexes and complex construction of index systems in the evaluation of emergency rescue intelligent command and decision-making systems.

[0010] To solve the above technical problems, the present invention provides a method for constructing an evaluation index system for an emergency rescue intelligent command and decision-making system, including the following steps: Step S01: Adjust the input data features, and use model interpretability to preliminarily screen out, through the SHAP algorithm, the evaluation indexes with a contribution degree higher than a preset contribution degree threshold from the potential evaluation index set of the emergency rescue intelligent command and decision-making system to form a candidate evaluation index set. The data features are the limiting conditions affecting the emergency rescue intelligent command and decision-making, and the potential evaluation index set includes the indexes for evaluating the advantages and disadvantages of the emergency rescue intelligent command and decision-making system; Step S02: Construct a Bayesian causal relationship graph model, and use the Bayesian causal relationship graph model to determine the causal relationship strength and direction between the evaluation indexes in the candidate evaluation index set; Step S03: Verify the accuracy of the causal relationship strength and direction between the evaluation indexes in the candidate evaluation index set, optimize the causal relationship strength and direction between the evaluation indexes in the candidate evaluation index set, and screen out the key evaluation indexes; Step S04: Intervene in and analyze the causal relationship between the key evaluation indexes to determine the causal relationship effect value of the key evaluation indexes; Step S05: Integrate the contribution degree and the causal relationship effect value of the key evaluation indexes, calculate the weights of each key evaluation index, and generate an evaluation index system.

[0011] Preferably, in step S01, by adjusting the input data features, the SHAP value algorithm is used to interpret the intelligent decision-making results of the emergency rescue intelligent command and decision-making system. The SHAP values of the data features for the decision-making results and their evaluation indicators are calculated respectively as the contribution degrees of the data features to the decision-making results and their evaluation indicators. The evaluation indicators are sorted according to the SHAP values of the data features for the evaluation indicators, and the evaluation indicators with contribution degrees higher than the preset contribution threshold are initially screened out to form a candidate evaluation indicator set for the data features in a sample of the emergency rescue intelligent command and decision-making system The calculation formula of the SHAP value for the decision-making result is as follows: ; Where: N is the set of all data features; S is N a subset of, excluding the data feature i ; represents using only the data feature vectors in the data feature subset S ; ; is the prediction function; represents the model prediction value using only the data feature vectors in the data feature subset S ; ; represents the contribution degree SHAP value of the data feature to the decision-making result in a sample of the emergency rescue intelligent command and decision-making system; The SHAP values for all samples of the emergency rescue intelligent command and decision-making system are averaged to obtain the contribution degree i of each data feature , and the calculation formula is: ; Where, is the total number of samples; is the j th sample, and i is the SHAP value of the feature

[0012] Preferably, in step S02, by analyzing and screening the evaluation indicators in the candidate evaluation indicator set, the evaluation indicators with causal relationships are determined as potential key evaluation indicators. Taking the potential key evaluation indicators as nodes and the causal relationships as edges, a Bayesian causal relationship graph model is constructed using a structure learning algorithm. By performing parameter estimation on each node and edge in the Bayesian causal relationship graph model, the strength and direction of each causal relationship between the potential key evaluation indicators are determined.

[0013] Preferably, in step S04, different intervention measures for the key evaluation indicators are simulated to generate an intervention dataset. These intervention measures are executed on the Bayesian causal relationship graph model, and the impact on the target key evaluation indicator is observed. By comparing the key evaluation indicators before and after the intervention, the degree of influence of different intervention measures on the causal relationship between the key evaluation indicators is determined, and the causal relationship effect value between the key evaluation indicators is determined.

[0014] Preferably, by fusing the contribution degree and the causal relationship effect value of the key evaluation indicators, the weight of each key evaluation indicator is calculated. Specifically, the following formula is used for calculation:

[0015] Where: is the comprehensive weight of the key evaluation indicator i; is the causal relationship effect value of the key evaluation indicator i after standardized processing; is the importance degree of the key evaluation indicator i after standardized processing; γ and δ are parameters that adjust the influence of the causal relationship effect value and the contribution degree; n is the total number of key evaluation indicators.

[0016] The present invention also provides a system for constructing an evaluation index system of an emergency rescue intelligent command and decision-making system, which is used to implement the above-mentioned method for constructing an evaluation index system of an emergency rescue intelligent command and decision-making system, and includes: A candidate evaluation indicator screening module, which adjusts the input data features, and uses model interpretability to preliminarily screen out the indicators with a contribution degree higher than a preset contribution degree threshold from the potential evaluation indicator set of the emergency rescue intelligent command and decision-making system through the SHAP algorithm to form a candidate evaluation indicator set; A Bayesian causal relationship graph construction module, which constructs a Bayesian causal relationship graph model, and uses the Bayesian causal relationship graph model to determine the strength and direction of the causal relationship between the evaluation indicators in the candidate evaluation indicator set; verifies the accuracy of the strength and direction of the causal relationship between the evaluation indicators in the candidate evaluation indicator set, optimizes the strength and direction of the causal relationship between the evaluation indicators in the candidate evaluation indicator set, and screens out the key evaluation indicators; A causal reasoning module, which intervenes in and analyzes the causal relationship between the key evaluation indicators to determine the causal relationship effect value of the key evaluation indicators; The index system generation module integrates the contribution degrees and causal relationship effect values of key evaluation indexes, calculates the weights of each key evaluation index, and generates an evaluation index system.

[0017] Preferably, the candidate evaluation index screening module includes an intelligent decision-making agent model training and inference sub-module and a SHAP index contribution degree calculation sub-module; The intelligent decision-making agent model training and inference sub-module includes an intelligent decision-making agent model training unit and an intelligent decision-making agent model inference unit; the intelligent decision-making agent model training unit is responsible for selecting an intelligent decision-making model with an equivalent interpretable internal structure as an agent model for the intelligent decision-making target model, obtaining a training data set, and training the intelligent decision-making agent model; the intelligent decision-making agent model inference unit uses the trained intelligent decision-making model for inference, generates an intelligent decision-making result, and selects all possible calculated evaluation indexes from the set of potential evaluation indexes; The SHAP index contribution degree calculation sub-module: includes a data feature SHAP calculation unit, an evaluation index contribution degree calculation and sorting unit, and an index contribution degree screening unit; the data feature SHAP calculation unit is used to calculate the contribution degree SHAP values of all data features to the decision result, and the SHAP values of all data features are used as the contribution degrees; the evaluation index contribution degree calculation and sorting unit is responsible for calculating the contribution degree SHAP values of data features to the evaluation indexes through model ablation analysis, and sorting the evaluation indexes according to the SHAP values of data features to the evaluation indexes. The index contribution degree screening unit preliminarily screens out the evaluation indexes with contribution degrees higher than the preset contribution threshold to form a set of candidate evaluation indexes.

[0018] Preferably, the Bayesian causal relationship graph construction module includes: a key index determination unit, a Bayesian causal relationship graph model construction unit, a Bayesian model parameter estimation unit, and a Bayesian model verification unit; The key index determination unit is used to determine the candidate evaluation indexes with causal relationships as the key evaluation indexes for constructing the Bayesian causal relationship graph by screening and analyzing the candidate evaluation indexes in the set of candidate evaluation indexes; The Bayesian causal relationship graph model construction unit: uses the determined key evaluation indexes as nodes and the causal relationships as edges, and uses a structure learning algorithm to establish the network structure of the evaluation indexes to form a Bayesian causal relationship graph model; by analyzing the causal relationships between the nodes, a directed acyclic graph is constructed to represent the causal relationships between the key evaluation indexes; The Bayesian model parameter estimation unit determines the strength and direction of each causal relationship by parameter estimating each node and edge in the Bayesian causal relationship graph model; The Bayesian model verification unit is used to verify the accuracy of the Bayesian model; by verifying the Bayesian causal relationship graph model and adjusting and optimizing the model. Preferably, the causal inference module includes a causal intervention analysis sub-module, specifically including a causal intervention data generation unit, a causal intervention execution unit, and a causal effect evaluation unit; The causal intervention data generation unit generates causal intervention data and simulates different intervention measures; The causal intervention execution unit executes the intervention measures simulated by the causal intervention data generation unit and records the effects on the results in the causal relationship; The causal effect evaluation unit analyzes the data generated by the causal intervention execution unit to evaluate the effects of different intervention measures and determine the causal relationship effect values of the key evaluation indicators.

[0019] Preferably, the index system generation module includes an index weight calculation unit and an index hierarchy division unit; the index weight calculation unit integrates the contribution degrees and causal relationship effect values of the key evaluation indicators to calculate the weights of each key evaluation indicator; the index hierarchy division unit divides the evaluation indicators according to the Bayesian network hierarchy to generate an evaluation index system.

[0020] Compared with the prior art, the present invention has at least the following beneficial effects: 1. The evaluation index system construction method of the present invention calculates the contribution degree based on SHAP, aiming to deeply identify the key variables and evaluation indicators in the emergency rescue intelligent command and decision-making system, solves the problems of difficult interpretation of the black-box model and difficult variable selection, and reveals the decision-making mechanism inside the intelligent model and the importance of each variable.

[0021] 2. The evaluation index system construction method of the present invention is based on the variable causal relationship modeling technology of Bayesian causal inference, constructs a causal relationship model between variables, effectively solves the problem that traditional correlation analysis cannot reveal causal relationships, and quantifies the causal influences between each variable and evaluation indicator.

[0022] 3. The evaluation index system construction method of the present invention innovatively proposes a dual index weight method, jointly weights the average index by combining the causal relationship weight and the SHAP contribution degree weight of the data characteristics of the evaluation indicator, more reasonably depicts the independence and contribution of the evaluation indicator, and can more scientifically evaluate the emergency rescue intelligent command and decision-making system. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 is the technical route diagram of the evaluation index system construction method for the emergency rescue intelligent command and decision-making system of the present invention.

[0024] Figure 2 is the flow chart of the evaluation index system construction method for the emergency rescue intelligent command and decision-making system of the present invention.

[0025] Figure 3 This is the schematic diagram for constructing the evaluation index system of the emergency rescue intelligent command and decision-making system of the present invention.

[0026] Figure 4 This is the overall structure diagram of the system for constructing the evaluation index system of the emergency rescue intelligent command and decision-making system of the present invention.

[0027] Figure 5 This is an example of a directed acyclic graph of the present invention. Detailed implementation manners

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0029] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed present invention, but merely represents selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0030] It should be noted that similar reference numerals and letters denote similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0031] In the present invention, unless otherwise clearly defined and limited, the terms "installed", "connected", "connected", "fixed", etc. shall be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or integrated; it can be directly connected, or indirectly connected through an intermediate medium, and can be the internal connection of two components or the interaction relationship between two components. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0032] As Figure 1-2 shown, the present invention provides a method for constructing an evaluation index system for an emergency rescue intelligent command and decision-making system, including the following steps: Step S01: Adjust the input data features, and initially screen out the evaluation indicators with a contribution degree higher than the preset contribution degree threshold from the potential evaluation indicator set of the emergency rescue intelligent command and decision-making system by using the SHAP algorithm with model interpretability to form a candidate evaluation indicator set; the data features are the limiting conditions affecting the emergency rescue intelligent command and decision-making, including: limiting conditions such as personnel, environment, and equipment, and the potential evaluation indicator set includes indicators for evaluating the advantages and disadvantages of the emergency rescue intelligent command and decision-making system. Step S02: Construct a Bayesian causal graph model, and use the Bayesian causal graph model to determine the causal relationship strength and direction among the evaluation indicators in the candidate evaluation indicator set. Step S03: Verify the accuracy of the causal relationship strength and direction among the evaluation indicators in the candidate evaluation indicator set, optimize the causal relationship strength and direction among the evaluation indicators in the candidate evaluation indicator set, and screen out the key evaluation indicators. Step S04: Intervene in and analyze the causal relationship among the key evaluation indicators to determine the causal relationship effect value of the key evaluation indicators. Step S05: Integrate the contribution degree and the causal relationship effect value of the key evaluation indicators, calculate the weights of each key evaluation indicator, and generate an evaluation indicator system.

[0033] According to a specific implementation of the present invention, in step S01, an intelligent decision-making model with an equivalent interpretable internal structure is selected as the proxy model of the target emergency rescue intelligent command and decision-making system, and after training, it is used as the emergency rescue intelligent command and decision-making system, and the trained intelligent decision-making model is used for reasoning to generate an intelligent decision result.

[0034] According to a specific implementation of the present invention, in step S01, by adjusting the input data features, the SHAP value algorithm is used to interpret the intelligent decision result of the emergency rescue intelligent command and decision-making system, and the SHAP values of the data features for the decision result and for the evaluation indicators are calculated respectively as the contribution degrees of the data features for the decision result and for the evaluation indicators, and the evaluation indicators are sorted according to the SHAP values of the data features for the evaluation indicators, and the evaluation indicators with a contribution degree higher than the preset contribution threshold are initially screened out to form a candidate evaluation indicator set. For the data features in a sample of the emergency rescue intelligent command and decision-making system The formula for calculating the SHAP value of the decision result is as follows: ; Where: N is the set of all data features; S is N a subset of i ; Indicates using only a subset of data features S in the data feature vector ; is the prediction function; Indicates using only a subset of data features S in the data feature vector of the model prediction value; Indicates the contribution degree SHAP value of the data feature in a sample of the emergency rescue intelligent command and decision-making system to the decision-making result; Average the SHAP values of all samples of the emergency rescue intelligent command and decision-making system to obtain the contribution degree of each data feature i of , and the calculation formula is: ; Among them, is the total number of samples; is the j th sample, and the feature i is the SHAP value.

[0035] Generally, there is a linear relationship between the decision-making result and the evaluation index. However, the relationship between the data feature and the decision-making result and the evaluation index is a non-linear relationship. Therefore, through the interpretability method of SHAP, the contribution rate of the data feature to the decision-making result can be calculated. According to the linear propagation characteristics, the contribution rate of the data feature to the evaluation index can also be calculated, and it does not affect the sorting order.

[0036] According to a specific implementation of the present invention, in step S02, by analyzing and screening the evaluation indexes in the candidate evaluation index set, the evaluation indexes with causal relationships are determined as potential key evaluation indexes. Taking the potential key evaluation indexes as nodes and the causal relationships as edges, a Bayesian causal relationship graph model is constructed using the structure learning algorithm. By estimating the parameters of each node and edge in the Bayesian causal relationship graph model, the strength and direction of each causal relationship between the potential key evaluation indexes are determined, and the potential key evaluation indexes are used to screen out the key evaluation indexes in step S03.

[0037] According to a specific implementation of the present invention, Bayesian estimation is performed using the prior distribution, and the posterior estimation of the parameters is performed using the following formula:

[0038] Among them, is the prior count, usually taken as 1.

[0039] According to a specific embodiment of the present invention, in step S04, different intervention measures for the key evaluation indicators are simulated to generate an intervention dataset, and these intervention measures are executed on the Bayesian causal relationship diagram model to observe the impact on the target key evaluation indicators. By comparing the key evaluation indicators before and after the intervention, the degree of influence of different intervention measures on the causal relationship between the key evaluation indicators is determined, and the causal relationship effect value between the key evaluation indicators is determined.

[0040] The interference measure is a specific step in causal inference. There are many methods of data intervention, such as data augmentation, data amplification, data synthesis, data semantic replacement, etc.

[0041] The main interference measure is to perform a certain noise perturbation on the input data. For example, for numerical data, to perform an intervention operation, a value is added or subtracted on the original value. For image data, a certain gradient perturbation can be added to intervene on the original image, etc.

[0042] The calculation of the causal effect is the key to causal intervention analysis. By comparing the system performance before and after the intervention, the effects of different intervention measures are evaluated to determine the degree of causal influence. For example, if the target variable Y changes significantly after intervening on a certain indicator X, it indicates that X has a significant causal influence on Y.

[0043] According to a specific embodiment of the present invention, the causal relationship effect value is calculated by the following method: two key evaluation indicators are respectively set as the cause evaluation indicator X and the result evaluation indicator Y, and ACE (Average Causal Effect) average causal effect is defined as the difference in the expectation of the result evaluation indicator Y when performing intervention measures and on multiple samples; where, ; among them, represents the data vector before interference, including k data of samples, k represents the measure-specific value vectors respectively performed on the data vector before interference including k samples, CE represents the expectation, Y represents the result evaluation indicator vector, including

[0044] the result evaluation indicators of

[0045] By calculating the expectation of the result evaluation indicator Y under different intervention measures implemented on the cause evaluation indicator X, the causal relationship effect value of the cause evaluation indicator X on the result evaluation indicator Y is obtained. According to a specific embodiment of the present invention, in step S03, the Bayesian causal relationship diagram model is cross-validated, and the Bayesian causal relationship diagram model is adjusted and optimized according to the verification result, and the key evaluation indicators are screened out.

[0046] According to a specific embodiment of the present invention, by integrating the contribution degree and causal relationship effect value of key evaluation indicators, the weights of each key evaluation indicator are calculated, and the specific calculation is carried out using the following formula:

[0047] Where: is the comprehensive weight of the key evaluation indicator i; is the causal relationship effect value of the key evaluation indicator i after standardized processing; is the importance degree of the key evaluation indicator i after standardized processing; γ and δ are parameters that adjust the influence of the causal relationship effect value and contribution degree; n is the total number of key evaluation indicators.

[0048] The present invention also provides a system for constructing an evaluation index system of an emergency rescue intelligent command and decision-making system, which is used to implement the above-mentioned method for constructing an evaluation index system of an emergency rescue intelligent command and decision-making system, including: A candidate evaluation indicator screening module, which adjusts the input data features, and uses model interpretability to preliminarily screen out indicators with a contribution degree higher than a preset contribution degree threshold from the potential evaluation indicator set of the emergency rescue intelligent command and decision-making system through the SHAP algorithm to form a candidate evaluation indicator set; A Bayesian causal relationship diagram construction module, which constructs a Bayesian causal relationship diagram model, and uses the Bayesian causal relationship diagram model to determine the causal relationship strength and direction among the evaluation indicators in the candidate evaluation indicator set; verifies the accuracy of the causal relationship strength and direction among the evaluation indicators in the candidate evaluation indicator set, optimizes the causal relationship strength and direction among the evaluation indicators in the candidate evaluation indicator set, and screens out key evaluation indicators; A causal reasoning module, which intervenes in and analyzes the causal relationship among the key evaluation indicators to determine the causal relationship effect value of the key evaluation indicators; An index system generation module, which integrates the contribution degree and causal relationship effect value of the key evaluation indicators, calculates the weights of each key evaluation indicator, and generates an evaluation index system.

[0049] According to a specific embodiment of the present invention, the candidate evaluation indicator screening module includes an intelligent decision-making agent model training and inference sub-module and a SHAP index contribution degree calculation sub-module; The intelligent decision-making agent model training and inference sub-module includes an intelligent decision-making agent model training unit and an intelligent decision-making agent model inference unit; the intelligent decision-making agent model training unit is responsible for selecting an equivalent intelligent decision-making model with interpretable internal structure as the agent model for the intelligent decision-making target model, obtaining the training data set, and training the intelligent decision-making agent model; the intelligent decision-making agent model inference unit uses the trained intelligent decision-making model for inference, generates intelligent decision-making results, and selects all possible calculated evaluation indicators from the set of potential evaluation indicators. SHAP index contribution calculation sub-module: includes a data feature SHAP calculation unit, an evaluation index contribution calculation and sorting unit, and an index contribution screening unit; the data feature SHAP calculation unit is used to calculate the contribution SHAP values of all data features to the decision result, and the SHAP values of all data features are used as contributions; the evaluation index contribution calculation and sorting unit is responsible for calculating the contribution SHAP values of data features to the evaluation index through model ablation analysis, and sorting the evaluation indexes according to the SHAP values of data features to the evaluation index. The index contribution screening unit initially screens out the evaluation indexes with contributions higher than the preset contribution threshold to form a candidate evaluation index set.

[0050] Furthermore, the data feature SHAP calculation unit interprets the intelligent decision-making results of the emergency rescue intelligent command and decision-making system by adjusting the input data features and using the SHAP value algorithm, calculates the SHAP values of data features to the decision result and to the evaluation index respectively, and uses them as the contributions of data features to the decision result and to the evaluation index. For the data features in a sample of the emergency rescue intelligent command and decision-making system The formula for calculating the SHAP value of the decision result is as follows: ; Where: N is the set of all data features; S is N a subset of i ; represents using only the data feature subset S in the data feature vector ; is the prediction function; represents the model prediction value using only the data feature vector S in the data feature subset ; represents the data feature SHAP value of the contribution to the decision result; Average the SHAP values of all samples of the emergency rescue intelligent command decision-making system to obtain the contribution of each data feature i degree , and the calculation formula is: ; Among them, is the total number of samples; is the j th sample, and the feature i SHAP value.

[0051] According to a specific implementation of the present invention, the Bayesian causal relationship graph construction module includes: a key index determination unit, a Bayesian causal relationship graph model construction unit, a Bayesian model parameter estimation unit, and a Bayesian model verification unit; The key index determination unit is used to determine the candidate evaluation indexes with causal relationships as the key evaluation indexes for constructing the Bayesian causal relationship graph by screening and analyzing the candidate evaluation indexes in the candidate evaluation index set; The Bayesian causal relationship graph model construction unit: uses the determined key evaluation indexes as nodes and the causal relationships as edges, and uses the structure learning algorithm to establish the network structure of the evaluation indexes to form a Bayesian causal relationship graph model; by analyzing the causal relationships between the nodes, constructs a directed acyclic graph to represent the causal relationships between the key evaluation indexes; The Bayesian model parameter estimation unit determines the strength and direction of each causal relationship by parameter estimating each node and edge in the Bayesian causal relationship graph model; The Bayesian model verification unit is used to verify the accuracy of the Bayesian model; verifies the Bayesian causal relationship graph model and adjusts and optimizes the model. Furthermore, the Bayesian model parameter estimation unit uses the prior distribution for Bayesian estimation and uses the following formula for posterior estimation of parameters:

[0052] Among them, is the prior count, usually taking 1.

[0053] According to a specific implementation of the present invention, the causal reasoning module includes a causal intervention analysis sub-module, specifically including a causal intervention data generation unit, a causal intervention execution unit, and a causal effect evaluation unit; The causal intervention data generation unit generates causal intervention data and simulates different intervention measures; The causal intervention execution unit executes the intervention measures simulated by the causal intervention data generation unit and records the effects on the results in the causal relationship; The causal effect evaluation unit analyzes the data generated by the causal intervention execution unit to evaluate the effects of different intervention measures and determine the causal relationship effect values of key evaluation indicators.

[0054] Further, in the causal effect evaluation unit, the causal relationship effect value is calculated by the following method: Set two key evaluation indicators as the cause evaluation indicator X and the result evaluation indicator Y respectively. ACE is defined as the difference in the expectation of the result evaluation indicator Y when performing the intervention measure and ; ; where and represent the expected values of Y when the intervention cause evaluation indicator X is at specific values and respectively; By calculating the expectations of the result evaluation indicator Y under different intervention measures on the cause evaluation indicator X, the causal relationship effect value of the cause evaluation indicator X on the result evaluation indicator Y is obtained According to a specific implementation scheme of the present invention, the index system generation module includes an index weight calculation unit and an index hierarchy division unit; the index weight calculation unit integrates the contribution degree and the causal relationship effect value of the key evaluation indicators to calculate the weights of each key evaluation indicator; the index hierarchy division unit hierarchically divides the evaluation indicators according to the Bayesian network hierarchy to generate an evaluation index system.

[0055] Further, the index weight calculation unit integrates the contribution degree and the causal relationship effect value of the key evaluation indicators to calculate the weights of each key evaluation indicator, and specifically uses the following formula for calculation:

[0056] where: is the comprehensive weight of the key evaluation indicator i; is the causal relationship effect value of the key evaluation indicator i after standardization processing; is the importance degree of the key evaluation indicator i after standardization processing; γ and δ are parameters that adjust the influence of the causal relationship effect value and the contribution degree; n is the total number of key evaluation indicators.

[0057] Example 1 As Figure 1-3 shown, the present invention provides a method for constructing an evaluation index system for an emergency rescue intelligent command and decision-making system, including the following steps: Step S01, adjust the input data features, and use model interpretability to preliminarily screen out evaluation indicators with a contribution degree higher than the preset contribution degree threshold from the potential evaluation indicator set of the emergency rescue intelligent command and decision-making system through the SHAP algorithm, forming a candidate evaluation indicator set. The data features are the limiting conditions affecting the emergency rescue intelligent command and decision-making, including limiting conditions such as personnel, environment, and equipment. The potential evaluation indicator set includes indicators for evaluating the pros and cons of the emergency rescue intelligent command and decision-making system; Step S02, construct a Bayesian causal relationship diagram model, and use the Bayesian causal relationship diagram model to determine the causal relationship strength and direction among the evaluation indicators in the candidate evaluation indicator set; Step S03, verify the accuracy of the causal relationship strength and direction among the evaluation indicators in the candidate evaluation indicator set, optimize the causal relationship strength and direction among the evaluation indicators in the candidate evaluation indicator set, and screen out key evaluation indicators; Step S04, intervene in and analyze the causal relationship among the key evaluation indicators to determine the causal relationship effect value of the key evaluation indicators; Step S05, fuse the contribution degree and causal relationship effect value of the key evaluation indicators, calculate the weights of each key evaluation indicator, and generate an evaluation indicator system.

[0058] Embodiment 2 The present invention provides a method for constructing an evaluation indicator system for an emergency rescue intelligent command and decision-making system, including the following steps: Step S01, by adjusting the input data features, use the SHAP value algorithm to interpret the intelligent decision-making results of the emergency rescue intelligent command and decision-making system, calculate the SHAP values of the data features for the decision-making results and for the evaluation indicators respectively, as the contribution degrees of the data features for the decision-making results and for the evaluation indicators, and sort the evaluation indicators according to the SHAP values of the data features for the evaluation indicators, and preliminarily screen out evaluation indicators with a contribution degree higher than the preset contribution threshold, forming a candidate evaluation indicator set. The data features are the limiting conditions affecting the emergency rescue intelligent command and decision-making, including limiting conditions such as personnel, environment, and equipment. The potential evaluation indicator set includes indicators for evaluating the pros and cons of the emergency rescue intelligent command and decision-making system; Among them, select an intelligent decision-making model with an equivalent internally interpretable structure as the proxy model of the target emergency rescue intelligent command and decision-making system, and use it as the emergency rescue intelligent command and decision-making system after training. Use the trained intelligent decision-making model to perform reasoning through interpretability analysis to generate intelligent decision-making results; For the data features in a sample of the emergency rescue intelligent command and decision-making system The calculation formula for the SHAP value of the decision-making result is as follows: ; Wherein: N is the set of all data features; S is N a subset of, excluding the data feature i ; represents using only the data feature vectors in the data feature subset S ; ; is the prediction function; represents the model prediction value using only the data feature vectors in the data feature subset S ; ; represents the contribution degree SHAP value of the data feature in a sample of the emergency rescue intelligent command and decision-making system to the decision-making result; Average the SHAP values of all samples of the emergency rescue intelligent command and decision-making system to obtain the contribution degree of each data feature i ; , and the calculation formula is: ; where is the total number of samples; is the j th sample, the feature i SHAP value.

[0059] Step S02, by analyzing and screening the evaluation indicators in the candidate evaluation indicator set, determine the evaluation indicators with causal relationships as potential key evaluation indicators, use the potential key evaluation indicators as nodes and the causal relationships as edges, construct a Bayesian causal relationship graph model using the structure learning algorithm, and determine the strength and direction of each causal relationship between the potential key evaluation indicators by parameter estimating each node and edge in the Bayesian causal relationship graph model; Use the prior distribution for Bayesian estimation, and use the following formula for posterior estimation of parameters:

[0060] where is the prior count, usually taken as 1.

[0061] Step S03, perform cross-validation on the Bayesian causal relationship graph model to verify the accuracy of the causal relationship strength and direction between the evaluation indicators in the candidate evaluation indicator set, optimize the causal relationship strength and direction between the evaluation indicators in the candidate evaluation indicator set, adjust and optimize the Bayesian causal relationship graph model, and screen out the key evaluation indicators; Step S04, intervene in and analyze the causal relationships among the key evaluation indicators, and determine the causal relationship effect values of the key evaluation indicators; Specifically, the causal relationship effect value is calculated by the following method: Set two key evaluation indicators as the cause evaluation indicator X and the result evaluation indicator Y respectively. ACE is defined as the expected difference of the result evaluation indicator Y when and implementing the intervention measures, ; wherein, and respectively represent the expected values of Y when the intervention cause evaluation indicator X is at specific values and ; By calculating the expectations of the result evaluation indicator Y under different intervention measures on the cause evaluation indicator X, the causal relationship effect value of the cause evaluation indicator X on the result evaluation indicator Y is obtained; Step S05, fuse the contribution degrees and causal relationship effect values of the key evaluation indicators, calculate the weights of each key evaluation indicator, and generate an evaluation indicator system.

[0062] Furthermore, the following formula is specifically used to calculate the weights of each key evaluation indicator:

[0063] where: is the comprehensive weight of the key evaluation indicator i; is the causal relationship effect value of the key evaluation indicator i after standardized processing; is the importance degree of the key evaluation indicator i after standardized processing; γ and δ are parameters for adjusting the influence of the causal relationship effect value and contribution degree; n is the total number of key evaluation indicators.

[0064] Embodiment 3 As Figure 1-5 shown, the present invention provides a system for constructing an evaluation indicator system of an emergency rescue intelligent command and decision-making system, which is used to implement the above-mentioned method for constructing an evaluation indicator system of an emergency rescue intelligent command and decision-making system, including: A candidate evaluation indicator screening module, which adjusts the input data characteristics, and uses model interpretability to preliminarily screen out the indicators with contribution degrees higher than a preset contribution degree threshold from the potential evaluation indicator set of the emergency rescue intelligent command and decision-making system through the SHAP algorithm, and forms a candidate evaluation indicator set. The data characteristics are the limiting conditions affecting the emergency rescue intelligent command and decision-making, including: limiting conditions such as personnel, environment, equipment, etc. The potential evaluation indicator set includes the indicators for evaluating the advantages and disadvantages of the emergency rescue intelligent command and decision-making system; Furthermore, the candidate evaluation index screening module includes an intelligent decision-making agent model training and inference sub-module and a SHAP index contribution calculation sub-module; The intelligent decision-making agent model training and inference sub-module includes an intelligent decision-making agent model training unit and an intelligent decision-making agent model inference unit; the intelligent decision-making agent model training unit is responsible for selecting an intelligent decision-making model with an equivalent and internally interpretable structure as the intelligent decision-making agent model for the intelligent decision-making target model, obtaining a training data set, and training the intelligent decision-making agent model; the intelligent decision-making agent model inference unit uses the trained intelligent decision-making model to perform inference through interpretability analysis, generates an intelligent decision-making result, and selects all possible calculated evaluation indexes from the set of potential evaluation indexes; The SHAP index contribution calculation sub-module: includes a data feature SHAP calculation unit, an evaluation index contribution calculation and sorting unit, and an index contribution screening unit; the data feature SHAP calculation unit is used to calculate the contribution SHAP values of all data features to the decision result, and the SHAP values of all data features are used as the contributions; the evaluation index contribution calculation and sorting unit is responsible for calculating the contribution SHAP values of the data features to the evaluation indexes through model ablation analysis, and sorting the evaluation indexes according to the SHAP values of the data features to the evaluation indexes. The index contribution screening unit preliminarily screens out the evaluation indexes with contributions higher than the preset contribution threshold to form a set of candidate evaluation indexes.

[0065] Furthermore, the data feature SHAP calculation unit adjusts the input data features, uses the SHAP value algorithm to interpret the intelligent decision-making result of the emergency rescue intelligent command decision-making system, calculates the SHAP values of the data features to the decision result and to the evaluation indexes respectively, and takes them as the contributions of the data features to the decision result and to the evaluation indexes. For the data features in a sample of the emergency rescue intelligent command decision-making system The formula for calculating the SHAP value of the data features to the decision result is as follows: ; Where: N is the set of all data features; S is N a subset of i ; represents using only the data feature vector in the data feature subset S ; ; is the prediction function; represents the model prediction value using only the data feature vector in the data feature subset S ; ; Represents the data characteristics in a sample of the emergency rescue intelligent command and decision-making system The SHAP value of the contribution degree to the decision result; Average the SHAP values of all samples of the emergency rescue intelligent command and decision-making system to obtain the contribution degree of each data characteristic i of , and the calculation formula is: ; Among them, is the total number of samples; is the j th sample, and the feature i is the SHAP value.

[0066] According to the linear relationship between the decision result and the evaluation index, based on the SHAP value of the data feature to the decision result, the SHARP value of the data feature to the evaluation index can be obtained.

[0067] For the calculation of the contribution degree of the evaluation index, first perform model ablation analysis. For each evaluation index, sort according to the importance of its SHAP value. For features with higher global SHAP values, it is considered that they have a significant impact on the model output. Then, remove the features and retrain the model. To verify the impact of each feature on the model performance, after removing the features one by one, retrain the model and observe the change of the evaluation index. Next, calculate the importance of the evaluation index, and the importance of the evaluation index can be measured by the change of the model performance.

[0068] Set a certain performance index E, and the performance after removing feature i is , and the performance of the original model is , then the importance of feature i can be defined as:

[0069] If is positive, it means that the model performance decreases after removing feature i , and feature i has a positive contribution to the model.

[0070] Sort all features according to the calculated importance of the evaluation index . The features with higher importance are ranked in the front. According to the business requirements and model complexity, set a threshold for the importance of the evaluation index. Select all evaluation indexes with importance higher than the threshold to form a candidate evaluation index set. These indexes are considered to have a significant contribution to the model performance and should be focused on in the subsequent model development and optimization.

[0071] The Bayesian causal relationship graph construction module constructs a Bayesian causal relationship graph model, and uses the Bayesian causal relationship graph model to determine the causal relationship strength and direction among the evaluation indicators in the candidate evaluation indicator set; verifies the accuracy of the causal relationship strength and direction among the evaluation indicators in the candidate evaluation indicator set, optimizes the causal relationship strength and direction among the evaluation indicators in the candidate evaluation indicator set, and screens out the key evaluation indicators; Further, the Bayesian causal relationship graph construction module includes: a key indicator determination unit, a Bayesian causal relationship graph model construction unit, a Bayesian model parameter estimation unit, and a Bayesian model verification unit; The key indicator determination unit is used to screen and analyze the candidate evaluation indicators in the candidate evaluation indicator set, and determine the candidate evaluation indicators with causal relationships as the key evaluation indicators for constructing the Bayesian causal relationship graph; The Bayesian causal relationship graph model construction unit: uses the determined key evaluation indicators as nodes and the causal relationships as edges, and uses a structure learning algorithm to establish the network structure of the evaluation indicators, forming a Bayesian causal relationship graph model; by analyzing the causal relationships between the nodes, constructs a directed acyclic graph to represent the causal relationships among the key evaluation indicators; For the screened key indicators, a Bayesian causal relationship graph model is constructed. The core of this step is to use a structure learning algorithm to establish the network structure of the evaluation indicators, thereby forming a Bayesian causal relationship graph. The Bayesian causal relationship graph is a directed acyclic graph (DAG), as Figure 5 shown, where the nodes represent random variables (i.e., key evaluation indicators), and the edges represent the causal relationships between the variables.

[0072] Structure learning algorithms are mainly divided into three categories: constraint-based algorithms, score-based algorithms, and hybrid algorithms.

[0073] (1) Constraint-based algorithms: This method uses statistical independence tests to determine the conditional independence relationships between variables, thereby constructing the graph structure. Typical algorithms include the PC algorithm and the Grow-Shrink algorithm. These algorithms gradually determine whether there are direct causal connections between variables through a series of conditional independence tests.

[0074] (2) Score-based algorithms: This method defines a scoring function (such as BIC, AIC, MDL, etc.), scores the possible network structures, and then selects the network with the highest score as the optimal structure. Common search strategies include the greedy method, simulated annealing, and genetic algorithms, etc.

[0075] (3) Hybrid algorithms: Combine the advantages of the above two methods, first use the constraint method to narrow the structure search space, and then use the scoring method to find the optimal structure within the narrowed space.

[0076] After completing the structure learning, a Bayesian causal relationship graph representing the causal relationships between key indicators is obtained. This graph clarifies which indicators have direct causal effects on other indicators and provides a basis for subsequent parameter estimation.

[0077] The Bayesian model parameter estimation unit determines the strength and direction of each causal relationship by performing parameter estimation on each node and edge in the Bayesian causal relationship graph model; The Bayesian model verification unit is used to verify the accuracy of the Bayesian model; by verifying the Bayesian causal relationship graph model and adjusting and optimizing the model. Furthermore, the goal of parameter estimation is to determine the probability distribution of each node given its parent nodes, that is, to calculate the conditional probability table (CPT); to avoid the zero-probability problem caused by sparse samples, the Bayesian model parameter estimation unit uses the prior distribution for Bayesian estimation. In this embodiment, the Dirichlet prior distribution is adopted, and the following formula is used for the posterior estimation of parameters:

[0078] where is the prior count, usually taken as 1.

[0079] For each node in the network, its CPT is calculated according to the above method, and the complete Bayesian network model is jointly determined by the network structure and the conditional probability table.

[0080] Furthermore, by verifying the Bayesian causal relationship graph model, the purpose is to evaluate the accuracy and reliability of the Bayesian causal relationship graph model to ensure that it can effectively reflect the causal relationships between key indicators. The following process is adopted for verification in this embodiment.

[0081] (1) Cross-validation: Divide the dataset into a training set and a test set, use the training set for the structure learning and parameter estimation of the model, and then evaluate the performance of the model on the test set. Commonly used evaluation indicators include log-likelihood, prediction accuracy, etc.

[0082] (2) Model comparison: Try different model structures or parameter estimation methods, compare their performances on the test set, and select the optimal model.

[0083] (3) Sensitivity analysis: Analyze the sensitivity of the model to changes in input data and evaluate the robustness of the model.

[0084] Based on the verification results, it may be necessary to adjust the structure and parameters of the model. For example, if it is found that some causal relationships do not match the actual situation, the structure learning process can be reexamined, and the method or threshold of the independence test can be adjusted; if the performance of the model on the test set is not good, it may be necessary to introduce more data, adopt a more complex model or incorporate prior knowledge.

[0085] A causal inference module intervenes in and analyzes the causal relationships between key evaluation indicators to determine the causal relationship effect values of the key evaluation indicators; Furthermore, the causal inference module includes a causal intervention analysis sub-module, which specifically includes a causal intervention data generation unit, a causal intervention execution unit, and a causal effect evaluation unit; The causal intervention data generation unit generates causal intervention data to simulate different intervention measures; The causal intervention execution unit executes the intervention measures simulated by the causal intervention data generation unit and records the effects on the results in the causal relationship; The causal effect evaluation unit analyzes the data generated by the causal intervention execution unit to evaluate the effects of different intervention measures and determine the causal relationship effect values of the key evaluation indicators.

[0086] Furthermore, in the causal effect evaluation unit, the causal relationship effect value is calculated by the following method: Two key evaluation indicators are respectively set as the cause evaluation indicator X and the result evaluation indicator Y. ACE (Average Causal Effect) is defined as the difference in the expectations of the result evaluation indicator Y when the intervention measures and are carried out; ; where and respectively represent the expected values of Y when the intervention cause evaluation indicator X is at specific values and ; For the convenience of calculation, the following indirect estimation is made, that is:

[0087] where Z is all confounding variables. By adjusting Z, the influence of confounding factors can be eliminated to accurately estimate the causal effect.

[0088] By calculating the expectations of the result evaluation indicator Y under different intervention measures on the cause evaluation indicator X, the causal relationship effect value of the cause evaluation indicator X on the result evaluation indicator Y is obtained.

[0089] An index system generation module integrates the contribution degrees and causal relationship effect values of the key evaluation indicators, calculates the weights of each key evaluation indicator, and generates an evaluation index system.

[0090] Furthermore, the index system generation module includes an index weight calculation unit and an index hierarchy division unit; the index weight calculation unit fuses the contribution degree and causal relationship effect value of key evaluation indexes to calculate the weights of each key evaluation index; the index hierarchy division unit divides the evaluation indexes according to the Bayesian network hierarchy to generate an evaluation index system.

[0091] Even further, the index weight calculation unit fuses the contribution degree and causal relationship effect value of key evaluation indexes to calculate the weights of each key evaluation index, and specifically uses the following formula for calculation:

[0092] Where: is the comprehensive weight of key evaluation index i; is the causal relationship effect value after standardization of key evaluation index i; is the importance degree after standardization of key evaluation index i; γ and δ are parameters that adjust the influence of the causal relationship effect value and contribution degree; n is the total number of key evaluation indexes.

[0093] Adopt the standardization method of min-max normalization to map the causal effect and importance degree to the interval [0,1]. By adjusting the values of γ and δ, the influence degree of the causal effect and index importance degree on the comprehensive weight can be controlled. If the causal effect is considered more important, the value of γ can be increased.

[0094] According to the structure of the Bayesian network, divide the evaluation indexes into different levels to construct a hierarchical index system. For example, the top layer is the target variable, followed by the first-level factors, and then the second-level factors, etc. The hierarchical index system helps to clearly show the causal relationship and influence path between each index.

[0095] Embodiment 4 According to a specific implementation scheme of the present invention, the description of the evaluation index system construction system of the emergency rescue intelligent command and decision-making system of the present invention is carried out below, and the specific calculation process refers to Embodiment 3.

[0096] For the emergency rescue intelligent command and decision-making system, constructing a scientific and reasonable evaluation index system is crucial for improving the emergency response efficiency and ensuring the rescue effect. The evaluation index system construction system of the emergency rescue intelligent command and decision-making system of the present invention, based on the interpretability of the SHAP model and Bayesian causal reasoning, can effectively screen out key evaluation indexes, construct a comprehensive and accurate evaluation index system, so as to provide strong support for intelligent decision-making.

[0097] (1) The candidate evaluation index screening module conducts screening of the candidate evaluation index set The generation of emergency rescue plans involves various complex factors, namely data characteristics, such as the type and scale of natural disasters, resource allocation, personnel scheduling, environmental changes, personnel training levels, equipment availability, communication network coverage, rainfall, affected area, casualties, rescue team size, equipment investment, and so on.

[0098] According to expert experience and peer review, collect candidate evaluation indicators for the emergency rescue intelligent command and decision-making system, such as: a) Response time: The time from the occurrence of a disaster to the initiation of an emergency response.

[0099] b) Rescue efficiency: The number of people rescued or the number of affairs processed per unit time.

[0100] c) Resource utilization rate: The ratio of resource input to actual use.

[0101] d) Personnel safety: The casualty rate of rescue personnel and assisted personnel.

[0102] e) Information transfer rate: The timeliness and accuracy of information transfer between departments.

[0103] f) Degree of collaborative cooperation: The level of collaboration between different departments and agencies.

[0104] g) Plan completeness: The comprehensiveness of scenarios covered by the plan and response measures.

[0105] h) Public satisfaction: The degree of satisfaction of affected groups with rescue operations.

[0106] i) Cost-benefit ratio: The ratio of the economic input of rescue operations to the results.

[0107] j) Secondary disaster prevention and control: The prevention and control effect on possible secondary disasters.

[0108] k) Scope of environmental damage: The scope of environmental damage caused by rescue operations.

[0109] Select the XGBoost machine learning algorithm as the surrogate model of the emergency rescue intelligent command and decision-making system. The intelligent decision-making agent model inference unit of the surrogate model of the emergency rescue intelligent command and decision-making system completes model training, inference, and calculation of candidate evaluation indicators.

[0110] Then, use the evaluation indicator contribution degree calculation and ranking unit to complete the calculation and ranking of the contribution degrees of the above candidate evaluation indicators, and then screen out the following relatively important evaluation indicators.

[0111] Candidate evaluation indicators = {Response time, Rescue efficiency, Personnel safety, Resource allocation, Resource utilization rate, Cost-benefit ratio} (2) The Bayesian causal relationship diagram construction module constructs a Bayesian causal relationship diagram, determines the causal relationship strength and direction among the evaluation indicators in the candidate evaluation indicator set using the Bayesian causal relationship diagram model; verifies the accuracy of the causal relationship strength and direction among the evaluation indicators in the candidate evaluation indicator set, optimizes the causal relationship strength and direction among the evaluation indicators in the candidate evaluation indicator set, and screens out the key evaluation indicators; (3) The causal inference module intervenes in and analyzes the causal relationship among the key evaluation indicators, determines the causal relationship effect value of the key evaluation indicators, and optimizes the causal relationship diagram; It can be seen from the Bayesian evaluation indicator causal relationship diagram of the emergency rescue intelligent command and decision-making system evaluation indicator system that: ① The type and scale of disasters directly affect the response time and resource allocation.

[0112] ② The response time affects the rescue efficiency, and a faster response helps to improve the efficiency.

[0113] ③ The resource allocation affects the rescue efficiency and resource utilization rate.

[0114] ④ The rescue efficiency affects the safety of personnel, and high efficiency may reduce casualties.

[0115] ⑤ The resource utilization rate affects the cost-benefit ratio, and the effective use of resources can reduce costs.

[0116] According to the above evaluation indicator causal diagram, the evaluation indicators and their mutual influence relationships can be clearly seen, and it can be analyzed which factors have direct or indirect effects on the final rescue effect.

[0117] Identify the intervention points to improve the effect of the emergency rescue plan.

[0118] (4) The indicator system generation module integrates the contribution degree and causal relationship effect value of the key evaluation indicators, calculates the weights of each key evaluation indicator, and generates an evaluation indicator system for causal effect analysis.

[0119] Finally, the following set of independent and orthogonal evaluation indicators is identified: The final evaluation indicator system = {personnel safety, cost-benefit ratio, resource allocation}.

[0120] The above is only the preferred embodiment of the present invention and is not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. A method for constructing an evaluation index system of an emergency rescue intelligent command and decision-making system, characterized in that It includes the following steps: Step S01: Adjust the input data features. Using model interpretability, initially screen out evaluation indicators with a contribution degree higher than a preset contribution degree threshold from the potential evaluation indicator set of the emergency rescue intelligent command decision-making system through the SHAP algorithm to form a candidate evaluation indicator set. The data features are the limiting conditions affecting the emergency rescue intelligent command decision-making, and the potential evaluation indicator set includes indicators for evaluating the pros and cons of the emergency rescue intelligent command decision-making system; Step S02: Construct a Bayesian causal relationship graph model, and use the Bayesian causal relationship graph model to determine the causal relationship strength and direction among the evaluation indicators in the candidate evaluation indicator set; Step S03: Verify the accuracy of the causal relationship strength and direction among the evaluation indicators in the candidate evaluation indicator set, optimize the causal relationship strength and direction among the evaluation indicators in the candidate evaluation indicator set, and screen out the key evaluation indicators; Step S04: Intervene in the causal relationship among the key evaluation indicators and conduct an analysis to determine the causal relationship effect value of the key evaluation indicators; Step S05: Integrate the contribution degree and causal relationship effect value of the key evaluation indicators, calculate the weights of each key evaluation indicator, and generate an evaluation indicator system.

2. The method for constructing an evaluation index system of an emergency rescue intelligent command and decision-making system according to claim 1, wherein In step S01, by adjusting the input data features, the SHAP value algorithm is used to interpret the intelligent decision-making results of the emergency rescue intelligent command and decision-making system. The SHAP values of the data features for the decision-making results and their evaluation indicators are calculated respectively as the contribution degrees of the data features to the decision-making results and their evaluation indicators. The evaluation indicators are sorted according to the SHAP values of the data features for the evaluation indicators, and the evaluation indicators with contribution degrees higher than the preset contribution threshold are initially screened out to form a candidate evaluation indicator set for the data features in a sample of the emergency rescue intelligent command and decision-making system The calculation formula for the SHAP value of the decision-making result is as follows: ； Wherein: N is a set of all data features; S is N a subset of, and does not contain the data feature i ; Indicates using only the data feature subset S in the data feature vectors ; is a prediction function; Indicates using only a subset of data features S in the data feature vector of the model prediction value; Represent the data characteristics in a sample of the emergency rescue intelligent command and decision-making system The SHAP value of the contribution degree to the decision result; Average the SHAP values of all samples of the emergency rescue intelligent command and decision-making system to obtain the contribution degree of each data feature i of , and the calculation formula is: ； Among them, is the total number of samples; is the j th sample, and the SHAP value of the feature i is 3. The method for constructing an evaluation index system of an emergency rescue intelligent command and decision-making system according to claim 1, wherein In step S02, through the analysis and screening of the evaluation indicators in the candidate evaluation indicator set, determine the evaluation indicators with causal relationships as potential key evaluation indicators. Use the potential key evaluation indicators as nodes and the causal relationships as edges, and construct a Bayesian causal relationship graph model using the structure learning algorithm. By performing parameter estimation on each node and edge in the Bayesian causal relationship graph model, determine the strength and direction of each causal relationship among the potential key evaluation indicators.

4. The method for constructing an evaluation index system for an emergency rescue intelligent command and decision-making system according to claim 1, wherein, In step S04, simulate different intervention measures for the key evaluation indicators to generate an intervention dataset. Execute these intervention measures on the Bayesian causal relationship graph model, observe the impact on the target key evaluation indicator, and determine the impact degree of different intervention measures on the causal relationship among the key evaluation indicators by comparing the key evaluation indicators before and after the intervention, and determine the causal relationship effect value among the key evaluation indicators.

5. The method for constructing an evaluation index system for an emergency rescue intelligent command and decision-making system according to claim 1, wherein Integrate the contribution degree and causal relationship effect value of the key evaluation indicators, and calculate the weights of each key evaluation indicator. Specifically, use the following formula for calculation: Wherein: is the comprehensive weight of the key evaluation index i; is the causal relationship effect value after the standardized processing of the key evaluation index i; is the importance after the standardized processing of the key evaluation index i; γ and δ are parameters that adjust the influence of the causal relationship effect value and the contribution degree; n is the total number of key evaluation indexes.

6. An emergency rescue intelligent command and decision-making system evaluation index system construction system, characterized in that, For implementing the method for constructing an evaluation indicator system of the emergency rescue intelligent command decision-making system described in any one of claims 1-5, it includes: A candidate evaluation indicator screening module, which adjusts the input data features and initially screens out indicators with a contribution degree higher than a preset contribution degree threshold from the potential evaluation indicator set of the emergency rescue intelligent command decision-making system through the SHAP algorithm to form a candidate evaluation indicator set; A Bayesian causal relationship graph construction module, which constructs a Bayesian causal relationship graph model, and uses the Bayesian causal relationship graph model to determine the causal relationship strength and direction among the evaluation indicators in the candidate evaluation indicator set; verify the accuracy of the causal relationship strength and direction among the evaluation indicators in the candidate evaluation indicator set, optimize the causal relationship strength and direction among the evaluation indicators in the candidate evaluation indicator set, and screen out the key evaluation indicators; The causal inference module intervenes in and analyzes the causal relationships among key evaluation indicators to determine the causal relationship effect values of the key evaluation indicators; The indicator system generation module integrates the contribution degrees and causal relationship effect values of key evaluation indicators, calculates the weights of each key evaluation indicator, and generates an evaluation indicator system.

7. The system for constructing an evaluation index system of an emergency rescue intelligent command and decision-making system according to claim 6, characterized in that, The candidate evaluation indicator screening module includes an intelligent decision-making agent model training and inference sub-module and a SHAP indicator contribution degree calculation sub-module; The intelligent decision-making agent model training and inference sub-module includes an intelligent decision-making agent model training unit and an intelligent decision-making agent model inference unit; the intelligent decision-making agent model training unit is responsible for selecting an equivalent internally interpretable intelligent decision-making model as the proxy model for the intelligent decision-making target model, obtaining the training data set, and training the intelligent decision-making agent model; the intelligent decision-making agent model inference unit uses the trained intelligent decision-making model for inference, generates intelligent decision-making results, and selects all possible calculable evaluation indicators from the set of potential evaluation indicators; The SHAP indicator contribution degree calculation sub-module: includes a data feature SHAP calculation unit, an evaluation indicator contribution degree calculation and sorting unit, and an indicator contribution degree screening unit; The data feature SHAP calculation unit is used to calculate the contribution degree SHAP values of all data features to the decision result, and the SHAP values of all data features are used as the contribution degrees; the evaluation indicator contribution degree calculation and sorting unit is responsible for calculating the contribution degree SHAP values of data features to the evaluation indicators through model ablation analysis, and sorting the evaluation indicators according to the SHAP values of data features to the evaluation indicators. The indicator contribution degree screening unit preliminarily screens out the evaluation indicators with contribution degrees higher than the preset contribution threshold to form a set of candidate evaluation indicators.

8. The system for constructing an evaluation index system of an emergency rescue intelligent command and decision-making system according to claim 6, characterized in that, The Bayesian causal relationship diagram construction module includes: a key indicator determination unit, a Bayesian causal relationship diagram model construction unit, a Bayesian model parameter estimation unit, and a Bayesian model verification unit; The key indicator determination unit is used to determine the candidate evaluation indicators with causal relationships as the key evaluation indicators for constructing the Bayesian causal relationship diagram by screening and analyzing the candidate evaluation indicators in the set of candidate evaluation indicators; The Bayesian causal relationship diagram model construction unit: uses the determined key evaluation indicators as nodes and the causal relationships as edges, and uses a structure learning algorithm to establish the network structure of the evaluation indicators to form a Bayesian causal relationship diagram model; by analyzing the causal relationships between the nodes, a directed acyclic graph is constructed to represent the causal relationships among the key evaluation indicators; The Bayesian model parameter estimation unit determines the strength and direction of each causal relationship by parameter estimating each node and edge in the Bayesian causal relationship diagram model; The Bayesian model verification unit is used to verify the accuracy of the Bayesian model; by verifying the Bayesian causal relationship diagram model and adjusting and optimizing the model.

9. The system for constructing an evaluation index system of an emergency rescue intelligent command and decision-making system according to claim 6, characterized in that, The causal inference module includes a causal intervention analysis sub-module, specifically including a causal intervention data generation unit, a causal intervention execution unit, and a causal effect evaluation unit; The causal intervention data generation unit generates causal intervention data to simulate different intervention measures; The causal intervention execution unit executes the intervention measures simulated by the causal intervention data generation unit and records the effects on the outcomes in the causal relationship. The causal effect evaluation unit analyzes the data generated by the causal intervention execution unit to evaluate the effects of different intervention measures and determines the causal relationship effect values of the key evaluation indicators.

10. The system for constructing an evaluation index system of an emergency rescue intelligent command and decision-making system according to claim 6, wherein The index system generation module includes an index weight calculation unit and an index hierarchy division unit; the index weight calculation unit integrates the contribution degrees of the key evaluation indicators and the causal relationship effect values to calculate the weights of each key evaluation indicator. The index hierarchy division unit divides the evaluation indicators into hierarchies according to the Bayesian network hierarchy to generate an evaluation index system.

Images