Coal enterprise personalized risk warning generation method, system, device and medium

Abstract

The application discloses a coal enterprise individualized risk warning generation method, system, device and medium, relates to the coal mine safety risk early warning technical field, solves the technical problems that various data are not deeply fused, individual data is difficult to combine and warning is strengthened, and the technical scheme points are a warning information generation mechanism based on a three-dimensional feature vector and personnel portrait fusion method, customized warning of 'one person one strategy' is realized, information overload and low efficiency in execution caused by traditional 'one size fits all' warning are effectively solved, through 'advance accurate early warning + targeted disposal guidance', production interruption, equipment damage, personnel casualties and other losses caused by risk expansion are avoided, through dynamic threshold adjustment, historical data correlation analysis and personnel behavior prediction, advance identification and early warning of risks are realized, and operable disposal guidance is provided, so that coal mine safety management is changed from 'after-the-fact remediation' to 'prevention', and the upgrading of the industry safety management mode is promoted.

Description

Technical Field

[0001] This application relates to the field of coal mine safety risk early warning technology, and in particular to a method, system, equipment and medium for generating personalized risk warnings for coal enterprises. Background Technology

[0002] Coal is the core of my country's energy structure, and its safe production is crucial to human lives, enterprise stability, and energy security. However, traditional coal mine risk early warning systems rely on fixed threshold alarms and general broadcast prompts, failing to combine real-time employee location, current work content, and historical violation records to provide differentiated early warning / warning messages for employees in different positions, with varying experience and behavioral habits. According to industry statistics, most coal mine safety accidents stem from improper operation or untimely warnings; the traditional "one-size-fits-all" approach is no longer adequate for the "precise warnings" required by intelligent mines.

[0003] Existing technologies have three core limitations:

[0004] (1) Standardized templates: Ignoring individual differences among employees. For example, new inspectors need detailed operating instructions when faced with "abnormal gas concentration", while experienced employees with 5 years of experience who have violated gas detection regulations need to be warned of the consequences of violations. However, traditional templates cannot differentiate between them, resulting in weak targeting and low rectification efficiency.

[0005] (2) Data fragmentation: Employees’ real-time location, work information and historical violation records are stored separately in the Internet of Things platform, production management system and safety assessment system, without deep integration. For example, when an employee is working in a high-risk work area, the system cannot simultaneously retrieve his “current location risk level + work type + similar violation records in the past 3 months”, making it difficult to generate accurate warnings.

[0006] (3) Logic static: There is no iterative mechanism that incorporates employee feedback. For example, if an employee repeatedly violates the rules by "not recording work data", they will still receive the same prompt. There is no reinforcement of warnings for their past violations, and the effect will continue to diminish.

[0007] The above problems can be categorized into three types: First, fixed templates are difficult to adapt to the dynamic location and work scenarios of employees; second, fragmented data makes it impossible to accurately locate the root cause of risks by combining historical violation records; and third, low intelligence means that only general notifications are provided, without differentiated warning statements based on "location-work-violation history".

[0008] Therefore, how to integrate various types of data to generate personalized risk warnings is a problem that this application aims to solve. Summary of the Invention

[0009] This application provides a method, system, equipment, and medium for generating personalized risk warnings for coal enterprises. Its technical purpose is to deeply integrate various types of data, thereby combining individual data to strengthen risk warnings and complete adaptive optimization of risk warnings, realizing a comprehensive intelligent upgrade of coal mine risk warnings from "general broadcasts" to "person-specific warning statements".

[0010] The above-mentioned technical objective of this application is achieved through the following technical solution:

[0011] A method for generating personalized risk warnings for coal enterprises includes:

[0012] Step S1: Calculate the feature weights of the input data; whereby the input data includes real-time scene data, personnel feature data, and legal and case data;

[0013] Step S2: Construct the job feature vector based on the input data and activate the role; wherein, the role activation includes role definition and loading the corresponding prompt knowledge base;

[0014] Step S3: Calculate the risk score based on the feature weights of the input data, and then classify the risk level based on the risk score;

[0015] Step S4: Generate risk warning prompts for different risk levels based on the role definition and the prompt knowledge base;

[0016] Step S5: Determine whether the risk warning message matches the focus of personnel's attention. If yes, proceed to step S6; otherwise, proceed to step S4.

[0017] Step S6: Input the risk warning prompt into the large language model to generate a draft prompt, then calculate the semantic fit of the draft prompt, and determine whether the semantic fit reaches a preset threshold. If it does, output the draft prompt; otherwise, supplement the draft prompt with details through the large language model until the semantic fit reaches the preset threshold.

[0018] Step S7: Verify whether the initial draft of the prompt is compliant. If it is, generate the final draft of the risk warning. Otherwise, proceed to step S6 to revise the initial draft of the prompt using a large language model until the initial draft of the prompt passes the compliance verification.

[0019] Step S8: Collect feedback data, calculate feedback effect based on feedback data, optimize model parameters based on feedback effect, and update the association logic of "job position - real-time location - work information - historical violations - warning statement" based on the optimized parameters.

[0020] Preferably, step S1 includes:

[0021] Step S11: Standardize the input data, representing it as follows:

[0022] ;

[0023] in, Indicates the first The first sample Standardized feature values ​​of each feature Indicates the first The first sample 1 eigenvalue, This represents the minimum value among all eigenvalues. This represents the maximum value among all eigenvalues.

[0024] Step S12: Construct the probability matrix based on the standardized eigenvalues, then each probability value in the probability matrix is ​​represented as:

[0025] ;

[0026] in, Indicates the first The first sample The probability values ​​of each feature in different samples Indicates the total number of samples;

[0027] Step S13: Calculate the entropy value based on the probability value, and then calculate the redundancy based on the entropy value, expressed as:

[0028] - ;

[0029] ;

[0030] in, Indicates the first The entropy value of each feature, Indicates the first Redundancy of each feature;

[0031] Step S14: Calculate the feature weights based on the redundancy, as follows:

[0032] ;

[0033] in, Indicates the first Feature weights of each feature; Indicates the first of all features Redundancy of each feature ; This indicates the total number of features.

[0034] Preferably, step S2 includes:

[0035] Step S21: Extract job responsibility characteristics, operational procedure characteristics, and common risk characteristics from the coal industry knowledge base using an AI large model;

[0036] Step S22: Generate a job feature vector containing the core attributes of the job based on the job responsibility characteristics, the operation standard characteristics, and the common risk characteristics;

[0037] Step S23: Activate the corresponding role based on the job feature vector, that is, define the role based on the personnel feature data on the basis of the job feature vector, and simultaneously load the prompt knowledge base of the corresponding job of the role; wherein, the prompt knowledge base includes job-specific operating procedures, a list of common hidden dangers and emergency response procedures.

[0038] Preferably, step S3 includes:

[0039] Step S31: Obtain risk elements from the input data, including real-time monitoring elements, historical correlation elements, and personnel behavior elements;

[0040] Step S32: Calculate the risk score based on the feature weights and risk factors, as follows:

[0041] ;

[0042] in, Indicates the risk score; Indicates the time decay coefficient; Indicates the first The characteristic weights of each risk element Indicates the first Risk severity rating for each risk element;

[0043] Step S33: Classify the job risk level based on the risk score.

[0044] Preferably, step S5 includes:

[0045] Step S51: Calculate the correlation between the risk warning and the focus of the personnel, expressed as:

[0046] ;

[0047] in, Indicates the degree of relevance. Indicates length of employment. This indicates the average length of time an employee has been employed. Indicates the number of violations. Indicates the average number of violations. Indicates skill level. Indicates average skill level. This indicates the risk level corresponding to the job position indicated by the risk warning. This indicates the average risk level of the position; This indicates the degree of match between risk warnings and job responsibilities. This indicates the degree of matching between risk warnings and operational guidelines. This indicates the degree of matching between risk warnings and common risk characteristics. , , , Both represent weighting coefficients, and ;

[0048] Step S52: Match whether the correlation degree reaches a preset threshold. If yes, proceed to step S6; otherwise, proceed to step S4.

[0049] Preferably, in step S6, the semantic fit is expressed as:

[0050] ;

[0051] in, Indicates semantic fit. This indicates the degree of matching between the initial draft and the real-time scene data. This indicates the degree of matching between the initial draft and the personnel characteristic data. Indicates the weight.

[0052] Preferably, step S7 includes:

[0053] Step S71: Calculate the similarity between the initial draft of the prompt and the legal provisions, expressed as:

[0054] ;

[0055] in, Indicates similarity. A set of keywords representing legal provisions. This indicates the frequency of keywords in the query. This indicates the frequency of keywords in the initial draft of the suggestion. , , All of these represent adjustment parameters. This indicates the length of the initial draft. Indicates the average prompt length;

[0056] Step S72: Similarity If a preset threshold is reached, a final risk warning is generated; otherwise, it is determined to be a compliance conflict, triggering automatic correction and proceeding to step S6 to revise the initial warning draft using a large language model until the initial warning draft passes compliance verification.

[0057] Preferably, in step S8, the feedback effect is expressed as:

[0058] ;

[0059] in, Indicates feedback effect, Indicates the rectification period. This indicates the average rectification time. Indicates the number of repeated violations. Indicates the total number of prompts. It has been stated that the rectification has been completed. This indicates that the prompt is unclear;

[0060] The optimized parameters are expressed as follows:

[0061] ;

[0062] in, This represents the optimized parameters. Indicates the learning rate. This represents the parameters before optimization. Represents the loss function. , This represents the threshold for the target effect.

[0063] A personalized risk warning generation system for coal enterprises, the system being used to implement the personalized risk warning generation method for coal enterprises, the system comprising:

[0064] The first calculation module calculates the feature weights of the input data; the input data includes real-time scene data, personnel feature data, and dynamic data.

[0065] The role activation module constructs a job feature vector based on the input data and activates the role; wherein, the role activation includes role definition and loading the corresponding prompt knowledge base;

[0066] The risk level classification module calculates the risk score based on the feature weights of the input data, and then classifies the risk level based on the risk score.

[0067] The prompt generation module generates risk warning prompts for different risk levels based on the role definition and the prompt knowledge base;

[0068] The first judgment module determines whether the risk warning prompt matches the focus of personnel's attention. If it does, it proceeds to the second judgment module; otherwise, it proceeds to the prompt generation module.

[0069] The second judgment module inputs the risk warning prompt into the large language model to generate a draft prompt, then calculates the semantic fit of the draft prompt, and judges whether the semantic fit reaches a preset threshold. If it does, the draft prompt is output; otherwise, the draft prompt is supplemented with details through the large language model until the semantic fit reaches the preset threshold.

[0070] The verification module verifies whether the initial draft of the prompt is compliant. If it is, a final draft of the risk warning is generated; otherwise, it is transferred to the second judgment module to revise the initial draft of the prompt through a large language model until the initial draft of the prompt passes the compliance verification.

[0071] The optimization module collects feedback data, calculates the feedback effect based on the feedback data, optimizes the model parameters based on the feedback effect, and updates the association logic of "job position - real-time location - work information - historical violations - warning statement" based on the optimized parameters.

[0072] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the method for generating personalized risk warnings for coal enterprises.

[0073] A computer storage medium storing a computer program, which, when executed by a processor, implements the steps of the method for generating personalized risk warnings for coal enterprises.

[0074] The above technical solution can achieve at least some of the following technical effects:

[0075] (1) By “multi-dimensional data fusion + entropy weight method weight optimization”, the problem of early warning information deviation caused by traditional technology relying on single-dimensional data is solved.

[0076] (2) Based on the three-dimensional feature vector (real-time monitoring elements, historical correlation elements and personnel behavior elements) and personnel profile fusion method, the warning information generation mechanism realizes the customized warning of "one person, one policy", which effectively solves the problem of information overload and inefficient execution caused by the traditional "one-size-fits-all" warning.

[0077] (3) By “early and accurate early warning + targeted handling guidance”, losses such as production interruption, equipment damage and casualties caused by the expansion of risks were avoided.

[0078] (4) Traditional technologies are mostly “threshold-triggered alarms”, which are passive responses; this application achieves early identification and warning of risks through dynamic threshold adjustment, historical data correlation analysis and personnel behavior prediction, and provides operable disposal guidance, so that coal mine safety management shifts from “post-event remediation” to “pre-event prevention”, and promotes the upgrading of the industry’s safety management model. Attached Figure Description

[0079] Figure 1 This is a flowchart of the method for generating personalized risk warnings for coal enterprises as described in the embodiments of this application. Detailed Implementation

[0080] The technical solution of this application will be described in detail below with reference to the accompanying drawings.

[0081] like Figure 1 As shown, the method for generating personalized risk warnings for coal enterprises described in this application includes:

[0082] Step S1: Calculate the feature weights of the input data; whereby the input data includes real-time scene data, personnel feature data, and regulations and case data.

[0083] In this embodiment of the application, the real-time scene data includes: obtaining gas concentration through a coal mine IoT interface. ( %), Equipment power-on status ( =0 indicates that the device is off. =1 indicates that the device is powered on), and the personnel's UWB positioning coordinates ( Electronic fence intrusion status ( =0 indicates no intrusion. =1 indicates an intrusion) and the current job type and job progress, etc., with a sampling frequency of 15 seconds / time.

[0084] Personnel characteristic data includes static data retrieved from the coal mine safety management system (such as length of employment). Month, Skill Level ) and dynamic data (number of violations in the past 30 days) Post-training violation rate , The number of violations after training. (Total number of training sessions).

[0085] The regulations and case data include the text of the "Coal Mine Safety Regulations", enterprise job operation specifications, hazard handling process database, and historical accident cases.

[0086] Preferably, step S1 includes:

[0087] Step S11: Standardize the input data (gas concentration, number of violations, etc.), and represent it as follows:

[0088] ;

[0089] in, Indicates the first The first sample Standardized feature values ​​of each feature Indicates the first The first sample 1 eigenvalue, This represents the minimum value among all eigenvalues. This represents the maximum value among all eigenvalues.

[0090] Step S12: Construct the probability matrix based on the standardized eigenvalues, then each probability value in the probability matrix is ​​represented as:

[0091] ;

[0092] in, Indicates the first The first sample The probability values ​​of each feature in different samples This represents the total number of samples.

[0093] Step S13: Calculate the entropy value based on the probability value, and then calculate the redundancy based on the entropy value, expressed as:

[0094] - ;

[0095] ;

[0096] in, Indicates the first The entropy value of each feature, Indicates the first Redundancy of each feature.

[0097] Step S14: Calculate the feature weights based on the redundancy, as follows:

[0098] ;

[0099] in, Indicates the first Feature weights of each feature; Indicates the first of all features Redundancy of each feature ; This indicates the total number of features.

[0100] Step S2: Construct the job feature vector based on the input data and activate the role; wherein, the role activation includes role definition and loading the corresponding prompt knowledge base.

[0101] Preferably, step S2 includes:

[0102] Step S21: Extract job responsibility characteristics, operation standard characteristics, and common risk characteristics from the coal industry knowledge base using an AI big data model.

[0103] Step S22: Generate a job feature vector containing the core attributes of the job based on the job responsibility characteristics, the operation standard characteristics, and the common risk characteristics.

[0104] Step S23: Activate the corresponding role based on the job feature vector, that is, define the role based on the personnel feature data on the basis of the job feature vector, and load the prompt knowledge base of the corresponding job of the role simultaneously; wherein, the prompt knowledge base includes job-specific operating procedures (such as "gas detection frequency standard" for gas inspectors), a list of common hidden dangers (such as "equipment overload warning threshold" for coal mining machine operators) and emergency response procedures (such as "ventilation failure coordination steps" for dispatchers).

[0105] In this embodiment of the application, the roles are defined by making them concrete based on personnel data. For example, a gas inspector is matched with the role of "safety supervision engineer with 8 years of underground experience", and a dispatcher is matched with the role of "production scheduling specialist familiar with the mine's ventilation system".

[0106] Step S3: Calculate the risk score based on the feature weights of the input data, and then classify the risk level based on the risk score.

[0107] Preferably, step S3 includes:

[0108] Step S31: Obtain risk elements from the input data, including real-time monitoring elements, historical correlation elements, and personnel behavior elements.

[0109] In this embodiment of the application, historical related elements are obtained by calling a large language model to calculate the semantic similarity between "current location + work scenario" and "historical accident cases + the employee's historical violation records".

[0110] Risk factors related to personnel behavior include the number of violations in the past 30 days and the frequency of violations in similar scenarios.

[0111] Step S32: Calculate the risk score based on the feature weights and risk factors, as follows:

[0112] ;

[0113] in, Indicates the risk score; Indicates the time decay coefficient; Indicates the first The characteristic weights of each risk element Indicates the first Risk severity score for each risk element =1,2,3. Risk factors are extracted from the input data, therefore the feature weights of risk factors are also obtained through steps S11 to S14.

[0114] In the embodiments of this application, The score ranges from 0 to 5, with higher scores indicating higher risk and higher priority. When the risk factor is a real-time monitoring factor, =1.0; Time decay coefficient of historical related elements It decreases by 0.1 every hour, with a minimum of 0.5.

[0115] In the embodiments of this application, This represents the feature weights corresponding to the real-time monitored elements. This indicates the severity score of the risk corresponding to the real-time monitoring element (e.g., 5 points for exceeding the gas concentration limit).

[0116] This represents the feature weights corresponding to historically related elements. The score indicates the severity of risk corresponding to historical related elements (e.g., a score of 4 corresponds to a high degree of similarity in historical accidents).

[0117] This represents the feature weights corresponding to personnel behavior elements. The risk severity score indicates the level of risk associated with the individual's behavioral elements (e.g., 3 violations in the past 30 days correspond to 3 points).

[0118] , , All are calculated using the feature weight calculation method in step S1.

[0119] By integrating three core data categories—real-time monitoring elements, historical correlation elements, and personnel behavior elements—and quantifying the feature weights of each dimension using the entropy weight method, the shortcomings of traditional technologies that "emphasize hardware monitoring and neglect human-scenario collaboration" are addressed. This enables risk assessment dimensions to cover the entire chain of "environment-history-personnel," effectively improving the accuracy of early warning.

[0120] Step S33: Classify the job risk level based on the risk score.

[0121] In the embodiments of this application, This is considered a high-risk scenario, corresponding to situations that "may directly cause personal injury or major equipment accidents" (such as gas concentration ≥1.5% or roof pressure exceeding critical values).

[0122] The risk level is medium, corresponding to a scenario where there are potential safety hazards but no immediate emergency risks (such as equipment operating temperature being too high or ventilation volume being slightly below standard).

[0123] This is considered a low-risk scenario, corresponding to "non-standard operation but no direct safety threat" (such as failure to record test data as required or improper wearing of safety protective equipment).

[0124] Step S4: Generate risk warning prompts for different risk levels based on the role definition and the prompt knowledge base.

[0125] Specifically, warning statements are designed based on different risk levels, taking into account employees' real-time location, work information, and historical violation records, to ensure that the content is accurately matched to individual scenarios. These include:

[0126] (1) High-risk scenarios: Risk warning prompts strictly follow the three-stage structure of "problem location - basis reference - disposal suggestion", forcibly referencing specific clauses of the "Coal Mine Safety Regulations", generating multi-level response suggestions, and triggering the collaborative confirmation mechanism of superior management personnel to ensure closed-loop disposal actions;

[0127] (2) Medium-risk scenarios: Risk warning prompts focus on "early warning + behavioral guidance", and adjust the tone intensity according to the employee's past response behavior (such as using a warning tone for employees who have violated the rules multiple times, and a reminder tone for employees who have triggered the rules for the first time). For example, "The ventilation volume of the No. 3 working face you are responsible for is 8% lower than the standard value (background). Please check the ventilation fan filter within 1 hour (task) to avoid gas accumulation due to insufficient air volume (requirement)".

[0128] (3) Low-risk scenarios: Risk warning prompts are mainly based on "knowledge push + habit formation", combined with job safety knowledge to generate lightweight content, such as "Today's safety tips: Recording gas detection data according to the specifications can shorten the time for tracing hidden dangers (knowledge), and it is recommended that you synchronize to the system immediately after each test (guidance)", to enhance employees' safety awareness.

[0129] Step S5: Determine whether the risk warning message matches the focus of personnel's attention. If yes, proceed to step S6; otherwise, proceed to step S4.

[0130] Preferably, step S5 includes:

[0131] Step S51: Calculate the correlation between the risk warning and the focus of the personnel, expressed as:

[0132] ;

[0133] in, Indicates the degree of relevance. Indicates length of employment. This indicates the average length of time an employee has been employed. Indicates the number of violations. Indicates the average number of violations. Indicates skill level. Indicates average skill level. This indicates the risk level of the corresponding job (e.g., high risk / medium risk / low risk corresponds to 3 / 2 / 1). This indicates the average risk level of the position; This indicates the degree of matching between risk warnings and job responsibilities. For example, a gas inspector's risk warning for "gas detection" is 1, while the risk warning for "tunneling operation" is 0.3. This indicates the degree of alignment between risk warnings and operational procedures, such as operational procedures including "evacuate if gas levels exceed limits." If the value is 1, then the value is 0.5; This indicates the degree of matching between the risk warning and common risk characteristics. For example, if an individual has a large number of historical violations, the corresponding high-risk warning is 1. , , , Both represent weighting coefficients, and .

[0134] Traditional risk warnings often use standardized templates (e.g., "Gas exceeding limits, evacuate immediately"), failing to consider differences in job roles (e.g., the different operational authority of gas inspectors and tunneling workers) and personnel capabilities (e.g., the different emergency response levels of new and experienced employees). This application constructs a three-dimensional feature vector of "job responsibility characteristics - operational procedure characteristics - common risk characteristics," and calculates the correlation degree by combining it with personnel profiles (length of service, skill level, and historical violation records). This enables customized warning content for each individual, solving the problems of "information overload" or "insufficient targeting" in traditional warnings.

[0135] Step S52: Determine whether the correlation degree reaches a preset threshold. If it does, determine that it is a high match and proceed to step S6; otherwise, if it is not a match, proceed to step S4.

[0136] In this embodiment of the application, when A score ≥0.7 is considered a high match. If a mismatch is determined, proceed to step S4 to adjust the core elements of the risk warning message. Specifically, this includes: re-screening the core content of the risk warning message by identifying mismatches between the three-dimensional job feature vector and the personnel profile (such as low job suitability or mismatch in operating procedures), then recalculating the relevance in step S5 until the threshold is met, and then proceeding to step S6 to adjust the wording details using the Large Language Model (LLM).

[0137] Step S6: Input the risk warning prompt into the large language model to generate a draft prompt, then calculate the semantic fit of the draft prompt, and determine whether the semantic fit reaches a preset threshold. If it does, output the draft prompt; otherwise, supplement the draft prompt with details through the large language model until the semantic fit reaches the preset threshold.

[0138] In this embodiment, the Large Language Model (LLM) generates natural language warnings tailored to the scenario and the user based on its understanding of the intent behind the risk warning. The specific process includes:

[0139] (1) High-risk draft: LLM performs "dual-channel verification" - based on the risk warning prompts, it first matches the text in the regulatory library (such as the clauses of the "Coal Mine Safety Regulations"), and then associates the historical case handling logic to generate a prompt draft;

[0140] (2) Medium / low risk initial draft: LLM directly generates the initial draft prompt by calling the job knowledge base (such as operating specifications) according to the prompt.

[0141] Preferably, the semantic fit is expressed as:

[0142] ;

[0143] in, Indicates semantic fit. This indicates the degree of matching between the initial draft and the real-time scene data. This indicates the degree of matching between the initial draft and the personnel characteristic data. Indicates the weight.

[0144] In the embodiments of this application, =0.5.

[0145] The semantic similarity between "risk warning prompts" and "real-time scene data" is calculated using a Large Language Model (LLM). For example, the risk warning prompt text (such as "Gas concentration exceeds the standard and evacuation is required") and the real-time scene data text (such as "Current scene: tunnel face, gas concentration 1.6%)" are input into the LLM, and the semantic similarity is output (the value ranges from 0 to 1, and the closer it is to 1, the higher the matching degree).

[0146] The semantic similarity between "risk warning prompts" and "personnel characteristic data" is calculated using a Large Language Model (LLM). For example, the text of risk warning prompts (such as "new employees need to strengthen gas detection operations") and the text of personnel characteristic data (such as "personnel: 2 months on the job, basic skill level") are input into the LLM, and the semantic similarity is output (values ​​range from 0 to 1).

[0147] when If so, the initial draft of the prompt will be output. Then, targeted modifications are made through LLM, for example, if If the data is low, supplement the scene data. If the risk level is low, adjust the details (such as disassembly steps or reducing the complexity of terminology), but do not change the core content of the risk warning.

[0148] Step S7: Verify whether the initial draft of the prompt is compliant. If it is, generate the final draft of the risk warning. Otherwise, proceed to step S6 to revise the initial draft of the prompt using a large language model until the initial draft of the prompt passes the compliance verification.

[0149] Preferably, step S7 includes:

[0150] Step S71: Calculate the similarity between the initial draft of the prompt and the legal provisions, expressed as:

[0151] ;

[0152] in, Indicates similarity. A set of keywords representing legal provisions. This indicates the frequency of keywords in the query. This indicates the frequency of keywords in the initial draft of the suggestion. , , All of these represent adjustment parameters. This indicates the length of the initial draft. This indicates the average prompt length.

[0153] In the embodiments of this application, =1.2, =100, =0.75.

[0154] Step S72: Similarity If a preset threshold is reached, a final risk warning is generated; otherwise, it is determined to be a compliance conflict, triggering automatic correction and proceeding to step S6 to revise the initial warning draft using a large language model until the initial warning draft passes compliance verification.

[0155] In this embodiment of the application, when If a compliance conflict is detected, automatic correction will be triggered.

[0156] Step S8: Collect feedback data, calculate feedback effect based on feedback data, optimize model parameters based on feedback effect, and update the association logic of "job position - real-time location - work information - historical violations - warning statement" based on the optimized parameters.

[0157] In this embodiment of the application, the feedback data includes active feedback data and passive feedback data.

[0158] Proactive feedback data includes: employees clicking "rectified" ( =1), "Unclear prompt" =0).

[0159] Passive feedback data includes: rectification time. and number of repeated violations .

[0160] Preferably, the feedback effect is expressed as:

[0161] ;

[0162] in, Indicates feedback effect, Indicates the rectification period. This indicates the average rectification time. Indicates the number of repeated violations. Indicates the total number of prompts. It has been stated that the rectification has been completed. This indicates that the prompt is unclear.

[0163] The optimized parameters are expressed as follows:

[0164] ;

[0165] in, This represents the optimized parameters. Indicates the learning rate. , This represents the parameters before optimization. Represents the loss function. , This represents the threshold for the target effect.

[0166] Specifically, the parameters include job role weight, risk level determination threshold, and prompt length threshold.

[0167] The personalized risk warning generation system for coal enterprises described in this application includes a first calculation module, a role activation module, a risk level classification module, a prompt generation module, a first judgment module, a second judgment module, a verification module, and an optimization module.

[0168] The first calculation module is used to calculate the feature weights of the input data; the input data includes real-time scene data, personnel feature data, and dynamic data.

[0169] The role activation module is used to construct a job feature vector based on the input data and activate the role; wherein, the role activation includes role definition and loading the corresponding prompt knowledge base;

[0170] The risk level classification module is used to calculate the risk score based on the feature weights of the input data, and then classify the risk level based on the risk score;

[0171] The prompt generation module is used to generate risk warning prompts for different risk levels based on the role definition and the prompt knowledge base;

[0172] The first judgment module is used to determine whether the risk warning prompt matches the focus of personnel's attention. If it does, it is transferred to the second judgment module; otherwise, it is transferred to the prompt generation module.

[0173] The second judgment module is used to input the risk warning prompt into the large language model to generate a prompt draft, then calculate the semantic fit of the prompt draft, and judge whether the semantic fit reaches a preset threshold. If it does, the prompt draft is output; otherwise, the prompt draft is supplemented with details through the large language model until the semantic fit reaches the preset threshold.

[0174] The verification module is used to verify whether the initial draft of the prompt is compliant. If it is, a final draft of the risk warning is generated; otherwise, it is transferred to the second judgment module to revise the initial draft of the prompt through a large language model until the initial draft of the prompt passes the compliance verification.

[0175] The optimization module is used to collect feedback data, calculate the feedback effect based on the feedback data, optimize the model parameters based on the feedback effect, and update the association logic of "job position - real-time location - work information - historical violations - warning statement" based on the optimized parameters.

[0176] The above are exemplary embodiments of this application, and the scope of protection of this application is defined by the claims and their equivalents.

Claims

1. A method for generating personalized risk warnings for coal enterprises, characterized in that, include: Step S1: Calculate the feature weights of the input data; whereby the input data includes real-time scene data, personnel feature data, and legal and case data; Step S2: Construct the job feature vector based on the input data and activate the role; wherein, the role activation includes role definition and loading the corresponding prompt knowledge base; Step S3: Calculate the risk score based on the feature weights of the input data, and then classify the risk level based on the risk score; Step S4: Generate risk warning prompts for different risk levels based on the role definition and the prompt knowledge base; Step S5: Determine whether the risk warning message matches the focus of personnel's attention. If yes, proceed to step S6; otherwise, proceed to step S4. Step S6: Input the risk warning prompt into the large language model to generate a draft prompt, then calculate the semantic fit of the draft prompt, and determine whether the semantic fit reaches a preset threshold. If it does, output the draft prompt; otherwise, supplement the draft prompt with details through the large language model until the semantic fit reaches the preset threshold. Step S7: Verify whether the initial draft of the prompt is compliant. If it is, generate the final draft of the risk warning. Otherwise, proceed to step S6 to revise the initial draft of the prompt using a large language model until the initial draft of the prompt passes the compliance verification. Step S8: Collect feedback data, calculate feedback effect based on feedback data, optimize model parameters based on feedback effect, and update the association logic of "position - real-time location - work information - historical violations - warning statement" based on the optimized parameters; Step S5 includes: Step S51: Calculate the correlation between the risk warning and the focus of the personnel, expressed as: ； in, Indicates the degree of relevance. Indicates length of employment. This indicates the average length of time an employee has been employed. Indicates the number of violations. Indicates the average number of violations. Indicates skill level. Indicates average skill level. This indicates the risk level corresponding to the job position indicated by the risk warning. This indicates the average risk level of the position; This indicates the degree of match between risk warnings and job responsibilities. This indicates the degree of matching between risk warnings and operational guidelines. This indicates the degree of matching between risk warnings and common risk characteristics. , , , , , Both represent weighting coefficients, and ; Step S52: Match whether the correlation degree reaches a preset threshold. If yes, proceed to step S6; otherwise, proceed to step S4.

2. The method for generating personalized risk warnings for coal enterprises as described in claim 1, characterized in that, Step S1 includes: Step S11: Standardize the input data, representing it as follows: ； in, Indicates the first The first sample Standardized feature values ​​of each feature Indicates the first The first sample 1 eigenvalue, This represents the minimum value among all eigenvalues. This represents the maximum value among all eigenvalues. Step S12: Construct the probability matrix based on the standardized eigenvalues, then each probability value in the probability matrix is ​​represented as: ； in, Indicates the first The first sample The probability values ​​of each feature in different samples Indicates the total number of samples; Step S13: Calculate the entropy value based on the probability value, and then calculate the redundancy based on the entropy value, expressed as: ； ； in, Indicates the first The entropy value of each feature, Indicates the first Redundancy of each feature; Step S14: Calculate the feature weights based on the redundancy, as follows: ； in, Indicates the first Feature weights of each feature; Indicates the first of all features Redundancy of each feature ; This indicates the total number of features.

3. The method for generating personalized risk warnings for coal enterprises as described in claim 2, characterized in that, Step S2 includes: Step S21: Extract job responsibility characteristics, operational procedure characteristics, and common risk characteristics from the coal industry knowledge base using an AI large model; Step S22: Generate a job feature vector containing the core attributes of the job based on the job responsibility characteristics, the operation standard characteristics, and the common risk characteristics; Step S23: Activate the corresponding role based on the job feature vector, that is, define the role based on the personnel feature data on the basis of the job feature vector, and simultaneously load the prompt knowledge base of the corresponding job of the role; wherein, the prompt knowledge base includes job-specific operating procedures, a list of common hidden dangers and emergency response procedures.

4. The method for generating personalized risk warnings for coal enterprises as described in claim 3, characterized in that, Step S3 includes: Step S31: Obtain risk elements from the input data, including real-time monitoring elements, historical correlation elements, and personnel behavior elements; Step S32: Calculate the risk score based on the feature weights and risk factors, as follows: ； in, Indicates the risk score; Indicates the time decay coefficient; Indicates the first The characteristic weights of each risk element Indicates the first Risk severity rating for each risk element; Step S33: Classify the job risk level based on the risk score.

5. The method for generating personalized risk warnings for coal enterprises as described in claim 4, characterized in that, In step S6, the semantic fit is expressed as: ； in, Indicates semantic fit. This indicates the degree of matching between the initial draft and the real-time scene data. This indicates the degree of matching between the initial draft and the personnel characteristic data. Indicates the weight.

6. The method for generating personalized risk warnings for coal enterprises as described in claim 5, characterized in that, Step S7 includes: Step S71: Calculate the similarity between the initial draft of the prompt and the legal provisions, expressed as: ； in, Indicates similarity. A set of keywords representing legal provisions. This indicates the frequency of keywords in the query. This indicates the frequency of keywords in the initial draft of the suggestion. , , All of these represent adjustment parameters. This indicates the length of the initial draft. Indicates the average prompt length; Step S72: Similarity If a preset threshold is reached, a final risk warning is generated; otherwise, it is determined to be a compliance conflict, triggering automatic correction and proceeding to step S6 to revise the initial warning draft using a large language model until the initial warning draft passes compliance verification.

7. The method for generating personalized risk warnings for coal enterprises as described in claim 1, characterized in that, In step S8, the feedback effect is represented as follows: ； in, Indicates feedback effect, Indicates the duration of rectification. This indicates the average rectification time. Indicates the number of repeated violations. Indicates the total number of prompts. It has been stated that the rectification has been completed. This indicates that the prompt is unclear; The optimized parameters are expressed as follows: ； in, This represents the optimized parameters. Indicates the learning rate. This represents the parameters before optimization. Represents the loss function. , This represents the threshold for the target effect.

8. A personalized risk warning generation system for coal enterprises, the system being used to implement the personalized risk warning generation method for coal enterprises as described in any one of claims 1-7, characterized in that, The system includes: The first calculation module calculates the feature weights of the input data; the input data includes real-time scene data, personnel feature data, and dynamic data. The role activation module constructs a job feature vector based on the input data and activates the role; wherein, the role activation includes role definition and loading the corresponding prompt knowledge base; The risk level classification module calculates the risk score based on the feature weights of the input data, and then classifies the risk level based on the risk score. The prompt generation module generates risk warning prompts for different risk levels based on the role definition and the prompt knowledge base; The first judgment module determines whether the risk warning prompt matches the focus of personnel's attention. If it does, it proceeds to the second judgment module; otherwise, it proceeds to the prompt generation module. The second judgment module inputs the risk warning prompt into the large language model to generate a draft prompt, then calculates the semantic fit of the draft prompt, and judges whether the semantic fit reaches a preset threshold. If it does, the draft prompt is output; otherwise, the draft prompt is supplemented with details through the large language model until the semantic fit reaches the preset threshold. The verification module verifies whether the initial draft of the prompt is compliant. If it is, a final draft of the risk warning is generated; otherwise, it is transferred to the second judgment module to revise the initial draft of the prompt through a large language model until the initial draft of the prompt passes the compliance verification. The optimization module collects feedback data, calculates the feedback effect based on the feedback data, optimizes the model parameters based on the feedback effect, and updates the association logic of "job position - real-time location - work information - historical violations - warning statement" based on the optimized parameters.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the personalized risk warning generation method for coal enterprises as described in any one of claims 1 to 7.

10. A computer storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the personalized risk warning generation method for coal enterprises as described in any one of claims 1 to 7.

Images