Coal mine safety knowledge vulnerability positioning system and method based on large model

Abstract

The application discloses a coal mine safety knowledge vulnerability positioning system and method based on a large model, wherein the method for positioning the coal mine safety knowledge vulnerability based on the large model comprises the following steps: acquiring a set of conditional knowledge points related to a post; performing an adversarial interrogation process and generating a cognitive state migration trajectory sequence; calculating a linkage influence matrix between dimensions and identifying abnormal linkage dimension pairs; generating a root misunderstanding hypothesis and outputting a knowledge vulnerability positioning result; and further uploading and clustering the desensitization misunderstanding structure characteristics of each mine through a federal aggregation mechanism, discovering a cross-mine common deviation mode, confirming an industry common root cognitive bias after distributed verification, and fusing and generating an enhanced knowledge vulnerability positioning result.

Description

Technical Field

[0001] This invention relates to the fields of coal mine safety management and artificial intelligence technology, and more specifically, to a coal mine safety knowledge vulnerability location system and method based on a large model. Background Technology

[0002] In coal mine safety production management, identifying knowledge gaps in employees is a crucial step in preventing accidents and ensuring safety. Existing large-scale model-based methods for identifying coal mine safety knowledge gaps typically operate within a single mine. They iteratively evaluate employees' cognitive states through an adversarial questioning process, where cognitive states are represented by multi-dimensional vectors. Each dimension corresponds to different conditional knowledge points. The evaluation terminates when the score change of any dimension of the cognitive state vector falls below a convergence threshold, and the lowest-scoring dimension in the final cognitive state vector is used to identify knowledge gaps.

[0003] The existing methods described above have the following technical problems: First, each dimension of the cognitive state vector is updated independently during the questioning process, and the convergence judgment is based solely on the magnitude of change of each dimension itself, without monitoring the linkage changes between dimensions. When practitioners have root-cause conceptual confusion, probing one conditional dimension may cause the score of another converged dimension to regress. This cross-dimensional abnormal migration phenomenon cannot be detected, resulting in the deep cognitive bias caused by conceptual confusion being masked. The cognitive state may stabilize at an erroneous value and be misjudged as converged, thus missing the corresponding knowledge gaps. Second, the sample size of single-mine assessments is limited, making it impossible to distinguish between individual-specific cognitive biases and industry-wide common cognitive biases. Furthermore, the cognitive assessment data of personnel from various mines are sensitive data for safety production and should not be directly shared centrally, resulting in the inability to discover and locate common root-cause cognitive biases across mines. Summary of the Invention

[0004] This invention provides a coal mine safety knowledge vulnerability localization system and method based on a large model, which solves the technical problems in related technologies that cannot effectively identify deep conceptual confusion and root cognitive bias in the safety knowledge system of coal mine practitioners, and that vulnerability localization results lack the ability to analyze the linkage relationship between dimensions.

[0005] This invention discloses a method for locating knowledge vulnerabilities in coal mine safety based on a large model, comprising the following steps: obtaining a set of job-related conditional knowledge points, wherein each knowledge point is associated with a set of applicable conditions, and each applicable condition corresponds to a dimension of the cognitive state vector; executing an adversarial questioning process for each knowledge point in the set of job-related conditional knowledge points, selecting a conditional dimension as the probed dimension in each round of questioning and updating the cognitive state vector, recording snapshots of the cognitive state vector in each round, and generating a cognitive state migration trajectory sequence; calculating the change matrix of all dimensions between adjacent trajectory points in the cognitive state migration trajectory sequence, accumulating and summing the change matrix to generate an inter-dimensional linkage influence matrix, identifying abnormal linkage dimension pairs corresponding to off-diagonal elements in the inter-dimensional linkage influence matrix whose absolute values ​​exceed the linkage threshold; based on the abnormal linkage dimension pairs and their associated question and answer content, generating root cause misunderstanding hypotheses through a large language model, and fusing the root cause misunderstanding hypotheses with conventional vulnerability analysis results to generate coal mine safety knowledge vulnerability location results based on a large model.

[0006] Furthermore, the acquisition of the set of job-related conditional knowledge points includes: inputting the text of the coal mine safety knowledge system into a large language model to extract knowledge points and their applicable conditions; based on the job description of the target practitioners, semantically matching the key functions in the job description with the applicable scenarios of the knowledge points; when the matching degree exceeds a preset relevance threshold, retaining the knowledge point, and generating the set of job-related conditional knowledge points.

[0007] Furthermore, each trajectory point in the cognitive state transition trajectory sequence includes the cognitive state vector of the current round, the condition dimension identifier being probed in this round, and the follow-up question and answer content of this round; the scores of each dimension of the cognitive state vector are processed by the mean normalization method based on the range, and the value range is a closed interval from zero to one; the termination condition of the adversarial follow-up questioning process is: the change amplitude of the scores of all dimensions of the cognitive state vector is lower than the convergence threshold within a consecutive preset sliding window length of rounds, or the number of follow-up questioning rounds reaches a preset maximum number of rounds; wherein, the convergence threshold constrains the upper limit of the absolute value of the difference between scores of two adjacent rounds, and when the absolute value of the difference between scores of two adjacent rounds of any dimension is less than the convergence threshold within a consecutive sliding window length of rounds, it is determined that the cognitive state vector has converged.

[0008] Furthermore, the step of accumulating and summing the change matrix to generate the inter-dimensional linkage influence matrix includes: for each round of follow-up questioning, the change matrix records the score change of each dimension only in the row corresponding to the currently probed dimension, and the elements in the remaining rows are set to zero; each element of the inter-dimensional linkage influence matrix is ​​the cumulative summation of the change matrix elements at the corresponding position over all follow-up questioning rounds, where the element in the i-th row and j-th column represents the cumulative score change of dimension j when dimension i is probed throughout the entire follow-up questioning process; since each round of follow-up questioning only probes one conditional dimension, the off-diagonal elements of the inter-dimensional linkage influence matrix specifically reflect the directional linkage relationship between dimensions; the linkage threshold is determined based on the mean of the absolute values ​​of all off-diagonal elements in the inter-dimensional linkage influence matrix plus a preset multiple of the standard deviation.

[0009] Furthermore, when identifying abnormal linkage dimension pairs, the linkage influence direction information is also recorded. When an element in the linkage influence matrix between dimensions is greater than the linkage threshold, it is marked as a positive abnormal linkage; when an element in the linkage influence matrix between dimensions is less than the negative value of the linkage threshold, it is marked as a negative abnormal linkage. The generation of root cause misunderstanding hypothesis through the large language model includes: inputting the dimension identifier, linkage influence direction, linkage influence feature vector, and related follow-up question and answer text and knowledge point content of the abnormal linkage dimension pair into the large language model. The large language model uses the dimension identifier to determine the specific conditional concepts involved in the confusion, uses the linkage influence direction to distinguish between positive and negative confusion, uses the value of the linkage influence feature vector to reflect the degree of confusion, and combines the specific expressions of practitioners in the follow-up question and answer text to infer the specific way of concept confusion, generating a root cause misunderstanding hypothesis description containing a description of the confusion relationship and the basis for inference.

[0010] Furthermore, the generation of the conventional vulnerability analysis results includes: for knowledge points where no abnormal linkages were detected, determining the final cognitive state vector according to the cognitive state vector convergence judgment logic, and arranging the final cognitive state vectors of each knowledge point according to the knowledge point dimension to form a global cognitive state matrix; performing column-by-column analysis on the global cognitive state matrix to identify conditional dimensions with low scores in multiple knowledge points and marking them as systemic weak conditional dimensions; performing row-by-row analysis on the global cognitive state matrix to identify cases where a low-scoring dimension exists in a certain knowledge point but the dimension scores normally in other knowledge points and marking them as local vulnerability knowledge points; the vulnerability types in the coal mine safety knowledge vulnerability location results based on the large model include mine-specific root cause deviation type, systemic weakness type, and local vulnerability type, and are sorted according to the rule that root cause deviation takes precedence over conventional vulnerabilities.

[0011] Furthermore, the anonymized misunderstanding structure feature vector is extracted from the root cause misunderstanding hypothesis descriptions generated locally in each mine. This misunderstanding structure feature vector contains four components: knowledge domain encoding, type encoding of confused concept pairs, direction encoding of linkage influence, and linkage intensity quantification value. Specific personnel information and mine sensitive parameters are replaced with typified labels, generating an anonymized misunderstanding structure feature set for each mine and uploading it to the federated aggregation node. The federated aggregation node performs cluster analysis on the anonymized misunderstanding structure feature vectors uploaded by all mines using a clustering algorithm. The similarity measure is based on weighted distance. The weighted distance uses encoding consistency judgment for knowledge domain encoding, type encoding of confused concept pairs, and direction encoding of linkage influence, and uses numerical difference measurement for linkage intensity quantification value. The sum of the preset weights corresponding to each component is one. The weighted distance is defined as the sum of the absolute value of the difference between the contribution weight value corresponding to the inconsistency of the discrete component encoding and the continuous component multiplied by the corresponding weight value. Deviation pattern clusters that occur more frequently across mines than a preset commonality threshold are marked as suspected industry common deviation patterns, generating a suspected industry common deviation pattern set and distributing it to each mine.

[0012] Furthermore, each mine matches the suspected industry-wide common deviation patterns with the root cause misunderstanding hypothesis descriptions already generated in its own mine. For patterns not found in the mine but belonging to the suspected industry-wide common deviation patterns, the type encoding and linkage influence direction encoding based on the confusion concept of the pattern are used to locate the corresponding knowledge points and conditional dimensions in the mine's safety knowledge system. The located knowledge point content, conditional dimension information, and deviation pattern structural description are input into a large language model to generate a localized verification and detection question set. The answer texts of practitioners to the localized verification and detection question set are obtained, and the semantic matching degree of the answer texts is calculated with the expected answer of the deviation hypothesis and the correct expected answer, respectively. When the semantic matching degree of the practitioner's answer with the expected answer of the deviation hypothesis is higher than the semantic matching degree of the practitioner's answer with the correct expected answer, it is determined that the practitioner has verified and confirmed the deviation pattern, and the verification and confirmation rate of each deviation pattern is calculated.

[0013] Furthermore, each mine uploads the verification confirmation rate of each deviation pattern to the federated aggregation node. The federated aggregation node summarizes and calculates the cross-mine weighted average confirmation rate. The cross-mine weighted average confirmation rate is the sum of the products of the sample size of practitioners participating in the verification of each mine and the corresponding verification confirmation rate, divided by the sum of the sample sizes of practitioners participating in the verification of each mine. Deviation patterns with a cross-mine weighted average confirmation rate exceeding a preset industry confirmation threshold are marked as confirmed industry common root cause cognitive deviations and distributed to each mine. Each mine integrates and updates the confirmed industry common root cause cognitive deviations with the coal mine safety knowledge vulnerability location results based on the large model. An industry common root cause deviation type is added to the vulnerability type. The vulnerability types are comprehensively sorted according to the rules of prioritizing industry common root cause deviations over mine-specific root cause deviations, root cause deviations over regular vulnerabilities, and high cross-mine confirmation rates over low confirmation rates, generating enhanced coal mine safety knowledge vulnerability location results based on the large model.

[0014] This invention provides a coal mine safety knowledge vulnerability localization system based on a large model, comprising: a knowledge point acquisition module, used to acquire the text of the coal mine safety knowledge system and generate a set of job-related conditional knowledge points based on job responsibility descriptions; a follow-up questioning execution module, used to execute an adversarial follow-up questioning process for each knowledge point in the set of job-related conditional knowledge points, selecting a conditional dimension as the probed dimension in each round of follow-up questioning and updating the cognitive state vector, recording snapshots of the cognitive state vector in each round and generating a cognitive state migration trajectory sequence; a linkage analysis module, used to calculate the change matrix of the cognitive state migration trajectory sequence, accumulate and sum the change matrix to generate a linkage influence matrix between dimensions, and identify abnormal linkage dimension pairs corresponding to off-diagonal elements whose absolute values ​​exceed the linkage threshold; and a vulnerability localization module, used to generate root cause misunderstanding hypotheses based on the abnormal linkage dimension pairs and their associated follow-up question and answer content through a large language model, and to fuse the root cause misunderstanding hypotheses with conventional vulnerability analysis results to generate coal mine safety knowledge vulnerability localization results based on a large model.

[0015] This invention addresses the technical problems of existing methods failing to detect cross-dimensional abnormal migrations and missing deep conceptual confusion-type knowledge gaps by recording cognitive state migration trajectory sequences and constructing inter-dimensional linkage and influence matrices during adversarial questioning processes. It achieves the technical effect of identifying root-cause cognitive biases caused by conceptual confusion stabilizing at erroneous values. Simultaneously, this invention solves the technical problems of insufficient sample size in single mines and the inability to directly share sensitive data, which prevent the discovery of common root-cause cognitive biases across mines, by desensitizing and structurally abstracting the descriptions of root-cause misunderstandings and performing cross-mine clustering analysis and distributed verification through federated aggregation nodes. It achieves the technical effect of distinguishing between individual-specific biases and industry-wide common biases without directly sharing sensitive assessment data from each mine. Attached Figure Description

[0016] Figure 1 This is a flowchart of the coal mine safety knowledge vulnerability location method based on a large model provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the cognitive state migration trajectory (changes in scores of each dimension) of EMP-047 provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the fourth round of Dimension 5 score regression anomaly (Dimension 4 detection triggers Dimension 5 regression) provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the heatmap of the interdimensional linkage influence matrix provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the comparison of the final cognitive state vector scores of EMP-047 provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the linkage intensity distribution of the desensitization misunderstanding feature vectors of three mines (cluster analysis input data) provided in an embodiment of the present invention; Figure 7 This is a schematic diagram comparing the verification confirmation rates of the P-02 deviation mode in three mines provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the semantic similarity distribution (partial employees) of the verification detection problem in Mine A provided by an embodiment of the present invention. Detailed Implementation

[0017] According to an embodiment of this implementation, a method for locating knowledge vulnerabilities in coal mine safety based on a large model is provided. It should be understood that the main implementers of this method are assessment terminals deployed locally in each mine and federated aggregation nodes. Each mine's assessment terminal possesses the ability to invoke large language models and local data processing capabilities, while the federated aggregation nodes are used to receive anonymized feature data uploaded by each mine and perform aggregation calculations.

[0018] At least one embodiment of the present invention discloses a method for locating knowledge gaps in coal mine safety based on a large model, such as... Figure 1 As shown, it includes the following steps: Step 1: Obtain a set of job-related conditional knowledge points; The text of the coal mine safety knowledge system is obtained and input into a large language model to extract knowledge points and their applicable conditions. Each knowledge point is associated with a set of applicable conditions, and each applicable condition corresponds to a dimension of the cognitive state vector. Based on the job descriptions of the target practitioners, the extracted knowledge points are filtered for relevance, retaining those directly related to the job responsibilities, thus generating a set of job-related conditional knowledge points.

[0019] It should be noted that the aforementioned set of applicable conditions refers to the set of preconditions that make a certain knowledge point valid or applicable. For example, for knowledge points related to gas detection, the set of applicable conditions may include conditions such as the type of detection instrument, the detection location, the ventilation status, and the gas concentration threshold, with each condition constituting an independent dimension of the cognitive state vector.

[0020] It should be noted that the aforementioned relevance filtering refers to semantically matching the key functions in the job description with the applicable scenarios of the knowledge points. When the matching degree exceeds a preset relevance threshold, the knowledge point is retained. For example, for the gas detection position, the knowledge points retained after relevance filtering cover conditional knowledge directly related to the position, such as gas concentration determination, ventilation system linkage, and gas emission assessment.

[0021] Step 2: Execute the adversarial questioning process and generate a sequence of cognitive state transition trajectories; For each knowledge point in the set of conditional knowledge points related to the job, an adversarial probing process is executed locally in each mine. In each round of probing, a conditional dimension is selected as the currently probed dimension, and probing questions are posed to the employee regarding that conditional dimension. After obtaining the employee's response text, the cognitive state vector is updated. After each round of probing, a complete snapshot of the cognitive state vector is recorded, and the snapshots from each round are arranged chronologically to generate a cognitive state transition trajectory sequence for that knowledge point.

[0022] It should be understood that each trajectory point in the cognitive state transition trajectory sequence contains three pieces of data: the cognitive state vector of the current round, the condition dimension identifier being probed in this round, and the content of the follow-up questions and answers in this round.

[0023] Let the dimension of the cognitive state vector be . Execution of adversarial questioning process The sequence of cognitive state transition trajectories generated after the round is represented as follows: ,in For the first Cognitive state vector after rounds of follow-up questioning For the first The cognitive state vector after the first round of questioning The rating values ​​for each dimension, For the first The conditional dimension identifier for the wheel being detected. For the first The follow-up questions and answers regarding the wheel Let be the dimension number of the cognitive state vector. This represents the total number of rounds in the adversarial questioning process. Scores for each dimension of the cognitive state vector. The normalized value has a range of values. The original scores were processed using a mean normalization method based on the range.

[0024] It should be noted that the termination condition for the above adversarial questioning process is: when the rating changes of all dimensions of the cognitive state vector are within a continuous range. All rounds were below the convergence threshold When, or when asked if the number of rounds has reached the preset maximum number of rounds. The questioning should end here. The length of the sliding window used for convergence judgment. This is a convergence threshold used to determine whether the cognitive state vector tends to stabilize. This is the preset maximum number of follow-up questions.

[0025] Furthermore, the aforementioned convergence threshold The variation range of scores in each dimension of the cognitive state vector is constrained, since the scores in each dimension have been normalized to... interval, The range of values ​​is , representing the upper limit of the absolute value of the difference between two adjacent rounds of scores, that is, for any dimension ,when In continuous All within the wheel When all conditions are met, the cognitive state vector is considered to have converged.

[0026] Step 3: Calculate the inter-dimensional linkage influence matrix and identify abnormal linkage dimension pairs; For each knowledge point, calculate the change matrix across all dimensions between adjacent trajectory points in the cognitive state transition trajectory sequence. For the... Follow-up questioning, when the conditional dimension being probed is identified as At that time, the change matrix The Line 1 Column elements are defined as: in For the first The cognitive state vector after the first round of questioning The rating values ​​for each dimension, For the first The score after a round of follow-up questions For the first The conditional dimension identifier for the wheel being detected. This represents the dimension of the cognitive state vector. The change matrix. The remaining rows are set to zero, meaning only the influence of the currently probed dimension on each other is recorded. Since the scores for each dimension of the cognitive state vector have been normalized... Within the interval, the scoring dimensions of each dimension are consistent, and the calculation of each element of the change matrix is ​​not affected by the difference in dimensions.

[0027] By summing the change matrix, an inter-dimensional linkage and influence matrix is ​​generated. : in This indicates the probing dimension throughout the entire questioning process. Time dimension The cumulative change in ratings This represents the total number of rounds in the adversarial questioning process. For the first The first round of change matrix Line 1 The elements of the column. Since each round of follow-up questions only probes one conditional dimension, In the Wheel only when The time is non-zero, therefore Only dimensions were accumulated. Dimensions when being detected The changes in the scores excluded the impact on the dimensions when other dimensions were being probed. Interference from this can cause interdimensional linkages to affect the matrix. Off-diagonal elements can specifically reflect the directional linkage relationship between dimensions.

[0028] Identify the inter-dimensional linkage influence matrix The absolute value exceeds the preset linkage threshold. The dimension pairs corresponding to the off-diagonal elements, that is, satisfying and Dimensions These are marked as abnormal linkage dimension pairs. A set of abnormal linkage dimension pairs and their corresponding linkage influence feature vectors are generated for each mine, where the linkage influence feature vector is the inter-dimensional linkage influence matrix. A vector consisting of the off-diagonal elements of the corresponding row.

[0029] It should be noted that the linkage threshold The linkage threshold is determined based on the statistical distribution of all off-diagonal elements in the inter-dimensional linkage influence matrix. The mean of the absolute values ​​of all off-diagonal elements plus a preset multiple of the standard deviation is taken as the linkage threshold. Because... The range of values ​​is affected by the cognitive state vector scoring range. Constraints throughout the questioning process In the wheel, The absolute value is at most no more than Linkage threshold The reasonable range of values ​​is .

[0030] It should be noted that the above-mentioned abnormal linkage dimensions reflect cross-dimensional abnormal transfer phenomena during the follow-up questioning process. When the dimensions... Being marked as abnormal linkage means that in the detection condition dimension During the process, conditional dimension The ratings changed beyond the normal range, and this correlated change indicates that practitioners may be at risk in certain dimensions. and dimensions There is confusion between the corresponding concepts.

[0031] In this embodiment of the application, in order to distinguish the different indicative roles of positive and negative linkage in addressing root cause deviations, information on the direction of linkage influence is also recorded when identifying abnormal linkage dimension pairs. When The time marker is marked as a positive anomaly linkage, indicating the detection dimension. Time dimension The rating rose abnormally; when The time marker is a negative anomaly linkage, indicating the detection dimension. Time dimension The rating has regressed. Negative anomaly linkage exists, dimension... It had previously tended to converge, but due to dimensionality The fact that the score regresses during the detection process strongly indicates the existence of a fundamental conceptual confusion.

[0032] Step 4: Generate root cause misconception hypotheses and output the knowledge gap location results; Step 401: Generate the root cause misconception hypothesis; For the abnormal linkage dimension pairs identified locally in each mine, the relevant follow-up questions and answers and corresponding knowledge points are obtained. The dimension identifiers, linkage influence directions, linkage influence feature vectors, and relevant follow-up questions and answers and knowledge points of the abnormal linkage dimension pairs are input into a large language model to generate a root cause misunderstanding hypothesis description. The root cause misunderstanding hypothesis description includes the conceptual confusion relationships that practitioners may have and the basis for their inferences.

[0033] Furthermore, in the process of generating the aforementioned root cause misunderstanding hypothesis, the large language model uses the dimensional identifier of the abnormal linkage dimension pair to determine the specific conditional concepts involved in the confusion, distinguishes between positive confusion (probing one concept leads to an abnormal increase in the score of another concept) and negative confusion (probing one concept leads to a regression in the score of another converged concept) by the direction of linkage influence, reflects the degree of confusion by the magnitude of the linkage influence feature vector, and infers the specific way in which the practitioners confuse the two concepts by combining the specific statements of practitioners in the follow-up question and answer text, thereby generating a root cause misunderstanding hypothesis description that includes a description of the confusion relationship and the basis for inference.

[0034] Step 402: Generate routine vulnerability analysis results; For knowledge points in the cognitive state transition trajectory sequence where no abnormal linkages were detected, their final cognitive state vectors are determined according to the cognitive state vector convergence judgment logic. The final cognitive state vectors of each knowledge point are arranged according to the knowledge point dimension to form a global cognitive state matrix. A column-by-column analysis is performed on the global cognitive state matrix to identify conditional dimensions with consistently low scores across multiple knowledge points, marking them as systemically weak conditional dimensions. A row-by-row analysis is then performed on the global cognitive state matrix to identify cases where a certain knowledge point has a low-scoring dimension, but that dimension scores normally in other knowledge points, marking these as locally vulnerable knowledge points. Regular vulnerability analysis results are then generated.

[0035] Step 403: Integrate and generate knowledge vulnerability location results; This method integrates the set of knowledge points related to root cause misconceptions and their impact with the results of conventional vulnerability analysis, and sorts them according to the rule that root cause bias takes precedence over conventional vulnerabilities, generating a coal mine safety knowledge vulnerability localization result based on a large model. The coal mine safety knowledge vulnerability localization result based on the large model includes vulnerability knowledge point identifiers, vulnerability types, and descriptions of associated root cause misconceptions. Vulnerability types include mine-specific root cause bias, systemic weakness, and localized vulnerability.

[0036] It should be noted that the aforementioned mine-specific root cause deviation type refers to the vulnerabilities corresponding to deep cognitive biases caused by conceptual confusion, which are discovered through analysis of abnormal linkage dimensions; the systemic weakness type refers to the vulnerabilities corresponding to conditional dimensions with generally low scores across multiple knowledge points, which are discovered through column-by-column analysis of the global cognitive state matrix; and the local vulnerability type refers to vulnerabilities that only manifest as low scores on individual knowledge points and do not involve cross-dimensional abnormal linkages.

[0037] In this embodiment of the application, in order to discover common root cause cognitive biases across mines without directly sharing sensitive assessment data from each mine, the following steps 5 to 8 are also included, which enhance the knowledge vulnerability location results based on step 4.

[0038] Step 5: Generate a set of desensitized and misleading structural features and upload it to the federated aggregation node; For each mine, the root cause misunderstanding hypothesis description generated in step 401 is processed, and a desensitized misunderstanding structure feature vector is extracted. The misunderstanding structure feature vector contains four components: the knowledge domain code, the type code of the confused concept pair, the linkage influence direction code, and the linkage intensity quantification value. Specifically, the knowledge domain code is the code value determined according to the domain classification of the coal mine safety knowledge system; the type code of the confused concept pair is the type code value determined according to the type of concept relationship between the two confused concepts (such as causal relationship confusion, measurement relationship confusion, conditional relationship confusion, etc.); the linkage influence direction code is the code value for positive or negative abnormal linkage; and the linkage intensity quantification value is the absolute value of the corresponding element in the inter-dimensional linkage influence matrix. The knowledge domain code, the type code of the confused concept pair, and the linkage influence direction code are discrete categorical data, converted to numerical representation using integer encoding; the linkage intensity quantification value is a continuous numerical value, scaled to a mean normalization method based on the range. The interval is used to eliminate the impact of differences in the dimensions of the linkage intensity between different knowledge points on subsequent distance calculations.

[0039] The specific personnel information and mine-sensitive parameters in the root cause misconception hypothesis descriptions are replaced with typified labels to generate a set of de-identified misconception structural features for each mine. Each mine then uploads its set of de-identified misconception structural features to the federated aggregation node.

[0040] It should be noted that the above-mentioned typified label replacement refers to replacing the specific mine name with a mine type label (such as underground mine, open-pit mine, etc.) and the specific personnel identifier with a job type label (such as gas inspector, ventilation manager, etc.), thereby removing sensitive information that can be traced back to specific mines and personnel while retaining the characteristics of the misleading structure.

[0041] Step 6: Cluster analysis of the desensitization misunderstanding structure characteristics and discovery of common deviation patterns across mines; The federated aggregation node performs clustering analysis on the anonymized misunderstanding structure features uploaded by all mines using a clustering algorithm. The input is a set of anonymized misunderstanding structure feature vectors uploaded by each mine, and the output is several deviation pattern clusters and the cluster center structure feature vector of each cluster. The similarity measure is based on the weighted distance of each component in the misunderstanding structure feature vector: for discrete components such as knowledge domain encoding and type encoding of confused concept pairs, encoding consistency is used; for the encoding of linkage influence direction, directional consistency is used; and for continuous components such as the linkage intensity quantification value, numerical difference is used.

[0042] It should be noted that the weighted distance is calculated as follows: Let the two misunderstanding structure feature vectors be respectively... and The four components of the misunderstanding structure feature vector are, in order, the knowledge domain encoding, the type encoding of the confused concept pair, the direction encoding of the linkage influence, and the quantification value of the linkage strength, with corresponding weights as follows: , , , Then the weighted distance Defined as: in and They represent and The One portion, , This is an indicator function that takes the value when the condition within the parentheses is true. Otherwise, the value is ; , , , The preset weights for each component satisfy the following conditions: Since the linkage intensity quantization value has been normalized in step 5, The interval is consistent with the value range for discrete component coding consistency judgment, and the dimensions of each component are unified in the weighted distance calculation, and the weights are consistent. to It can directly reflect the relative importance of each component. For discrete components (knowledge domain encoding, confusion concept pair type encoding, linkage influence direction encoding), the contribution is zero when the corresponding components of the two misunderstanding structural feature vectors are inconsistent. For continuous components (linkage strength quantification value), the contribution is the product of its weight and the absolute value of the difference between the two misunderstanding structural feature vectors for that component.

[0043] Furthermore, since the consistency judgment result of the discrete component encoding takes the value... The quantization value of the continuous component linkage intensity has been normalized to interval, therefore The range of values ​​is The weighted distance is consistent with the contribution range of the discrete components. The range of values ​​is Each weight to All values ​​are positive and satisfy the normalization constraint. .

[0044] The frequency of cross-mine occurrence and the number of mines involved are calculated for each deviation pattern cluster. The frequency of cross-mine occurrence is the total number of misunderstanding structural features from different mines included in the deviation pattern cluster, and the number of mines involved is the number of different mine sources included in the deviation pattern cluster. Deviation pattern clusters with a cross-mine occurrence frequency exceeding a preset commonality threshold are marked as suspected industry common deviation patterns. A set of suspected industry common deviation patterns and its statistical characteristics are generated and distributed to each mine. Each pattern in the suspected industry common deviation pattern set includes the structural feature vector of the cluster center, the frequency of cross-mine occurrence, and the number of mines involved.

[0045] In this embodiment of the application, in order to verify suspected common deviation patterns in the industry and avoid directly determining common deviations based solely on clustering results without cross-validation, the following steps 7 and 8 are also included.

[0046] Step 7: Generate a localized set of verification probe questions and perform verification; Each mine locally matches suspected industry-wide common deviation patterns with the root cause misconception hypothesis descriptions generated in step 401. For matching patterns independently discovered by the mine, the confidence level of the matching patterns is increased based on their frequency of occurrence across mines.

[0047] For patterns not found in this mine but suspected to be common industry-wide deviations, the type and direction of their associated impact are encoded based on the concept of confusion within the pattern. This allows for the identification of specific knowledge points and conditional dimensions within the mine's safety knowledge system. The identified knowledge points, conditional dimensions, and the structural description of the deviation pattern are input into a large language model to generate a set of verification probe questions tailored to the mine's specific knowledge content. These verification probe questions, by setting specific combinations of conditions, ensure that correct and confused understandings correspond to different response paths, thereby enabling the discriminative verification of deviation hypotheses. This generates a localized set of verification probe questions for the deviations to be verified.

[0048] Obtain the text of the responses from mine workers to a localized verification and detection question set. Calculate the semantic matching degree between each response and the expected responses to the deviation hypothesis and the expected correct responses. Determine the verification result for each worker regarding the deviation pattern based on the semantic matching degree. Calculate the verification confirmation rate for each deviation pattern among the mine workers.

[0049] It should be noted that the semantic matching degree mentioned above uses cosine similarity calculation. This involves encoding the practitioner's response text, the expected response text based on the deviation hypothesis, and the correct expected response text into semantic vectors using a large language model. Then, the cosine similarity between the practitioner's response semantic vector and the expected response semantic vector, as well as the cosine similarity between the practitioner's response semantic vector and the correct expected response semantic vector, is calculated. When the cosine similarity between the practitioner's response and the expected response based on the deviation hypothesis is higher than the cosine similarity between the practitioner's response and the correct expected response, it is determined that the practitioner has verified and confirmed the deviation pattern.

[0050] It should be noted that the above-mentioned expected response based on the assumption of bias refers to the response that should be given if practitioners do indeed have this confusion bias, while the expected response based on the assumption of correct understanding of the relevant concepts refers to the response that practitioners should give.

[0051] Step 8: Summarize the cross-mine verification confirmation rate and identify common root causes of cognitive biases in the industry; Each mine uploads the validation confirmation rate for each deviation pattern to the federated aggregation node. The federated aggregation node aggregates and calculates the weighted average confirmation rate across mines, where the weight of each mine is determined based on the sample size of practitioners participating in the validation.

[0052] It should be noted that the above-mentioned cross-mine weighted average confirmation rate is calculated as follows: Let the total number of mines participating in the verification be... , No. The sample size of employees from each mine participating in the verification was [number missing]. , No. The validation confirmation rate of a certain deviation model for a single mine is [percentage missing]. The cross-mine weighted average confirmation rate of this deviation model Defined as: in This represents the total number of mines participating in the verification. For the first The total number of employees in each mine who participated in the verification of this deviation pattern. For the first The number of employees in each mine who verified and confirmed this deviation pattern The ratio, , , Number the mine shaft. .

[0053] Deviation patterns where the weighted average confirmation rate across mines exceeds the preset industry confirmation threshold are marked as confirmed common root cause cognitive biases in the industry, and then distributed to each mine after adding cross-mine confirmation statistics.

[0054] Furthermore, the aforementioned industry confirmation threshold is a fixed threshold preset in the federated aggregation node, with a value range of [value range missing]. This is used to distinguish between common biases prevalent among employees in multiple mines and localized biases occurring only in a few mines. When the cross-mine weighted average confirmation rate... When the industry confirmation threshold is exceeded, the corresponding deviation pattern is determined to be a confirmed common root cause cognitive bias in the industry; when the cross-mine weighted average confirmation rate If the deviation pattern does not exceed the industry confirmation threshold, it will only be retained as a suspected pattern and will not be included in the industry's common root cause cognitive deviation set.

[0055] After receiving the confirmed set of common industry-wide cognitive biases, each mine integrates and updates the coal mine safety knowledge vulnerability location results generated in step 403 with the set of common industry-wide cognitive biases based on the large model. Knowledge points affected by common root cause biases are labeled with industry-wide common bias tags, cross-mine confirmation statistics, and descriptions of confused concepts. Vulnerability types are updated, adding a category of industry-wide common root cause bias. Enhanced coal mine safety knowledge vulnerability location results are generated by comprehensively ranking vulnerabilities according to the rules of prioritizing industry-wide common root cause biases over mine-specific root cause biases, root cause biases over regular vulnerabilities, and high cross-mine confirmation rates over low confirmation rates. The enhanced coal mine safety knowledge vulnerability location results based on the large model include vulnerability knowledge point identifiers, vulnerability types, cross-mine bias confirmation rates, associated root cause misunderstanding descriptions, and comprehensive priority scores. Vulnerability types include industry-wide common root cause bias, mine-specific root cause bias, systemic weakness, and local vulnerability.

[0056] It should be noted that the calculation rules for the above comprehensive priority score are as follows: First, the basic priority level is determined according to the vulnerability type. The basic level of industry common root cause deviation type is higher than that of mine specific root cause deviation type, the basic level of mine specific root cause deviation type is higher than that of systemic weakness type, and the basic level of systemic weakness type is higher than that of local vulnerability type. Within the same basic level, the vulnerability is sorted from high to low according to the cross-mine confirmation rate or linkage intensity quantification value to generate the final comprehensive priority score.

[0057] This implementation records a complete snapshot of the cognitive state vector and generates a sequence of cognitive state migration trajectories during the adversarial questioning process. It calculates a change matrix based on adjacent trajectory points and accumulates it to generate an inter-dimensional linkage influence matrix. Cross-dimensional abnormal linkage phenomena are detected by identifying off-diagonal elements in the inter-dimensional linkage influence matrix that exceed a linkage threshold. Since the off-diagonal elements of the inter-dimensional linkage influence matrix explicitly record the cumulative score changes in other dimensions when probing one conditional dimension, it can capture cross-dimensional abnormal migrations that are ignored in existing independent update methods, such as situations where probing one conditional dimension causes a regression in the score of another converged dimension. Based on the abnormal linkage dimension pairs and their associated follow-up questions and answers, a root cause misunderstanding hypothesis is generated. This allows for the identification and localization of deep cognitive biases caused by conceptual confusion leading to the cognitive state stabilizing at an erroneous value, rather than simply identifying superficial knowledge gaps based on low-scoring dimensions of the final cognitive state vector.

[0058] Furthermore, this implementation method desensitizes and structurally abstracts the description of root cause misunderstandings, replacing specific personnel information and mine sensitive parameters with typified labels. Only the misunderstanding structure feature vector is uploaded to the federated aggregation node, enabling cross-mine cluster analysis without directly sharing the sensitivity assessment data of each mine. Since the misunderstanding structure feature vector contains structured information such as knowledge domain encoding, confusion concept pair type encoding, linkage influence direction encoding, and linkage strength quantification, cluster analysis can group misunderstandings with similar structural features into the same deviation pattern cluster, thereby discovering common cross-mine deviation patterns caused by unified training materials or industry practices. By distributing the suspected industry-wide common deviation patterns discovered by federated aggregation to each mine and generating a localized verification and detection question set, and then performing cross-validation by summing the weighted average confirmation rate across mines after distributed verification, the identification of common industry-wide root cause cognitive biases does not rely on the limited sample size of a single mine, but is determined based on the distributed verification results of multiple mines, thus distinguishing between individual-specific biases and industry-wide common biases. Ultimately, this implementation method integrates and sorts the confirmed industry-wide common root cause deviations, mine-specific root cause deviations, and conventional vulnerability analysis results to achieve penetrating knowledge vulnerability localization, from superficial knowledge point deficiencies to deep-seated common concept confusion root causes.

[0059] The following is an example of an application of the present invention, such as Figure 2-8 As shown, the implementation process is as follows: Gas detectors at three mines (Mine A, Mine B, and Mine C) under a coal mining group recently underwent knowledge vulnerability assessments following safety training. All three mines used the same group-compiled training materials, with assessment terminals deployed locally at each mine and a federated aggregation node deployed at the group's safety management center. This study focuses on employee EMP-047 from Mine A as the primary target, demonstrating the complete data flow from knowledge point extraction to enhanced vulnerability localization results. Initial conditions: EMP-047 holds a gas detector certification, has 3 years of work experience, and this assessment targets the knowledge point of "handling excessive gas levels in mining faces."

[0060] The text of the group's unified safety training materials was input into a large language model to extract the set of applicable conditions for the knowledge point "Handling of Excessive Gas Levels in Mining Faces," identifying a total of 5 condition dimensions. Based on the job description of EMP-047 (responsible for gas detection and alarm response in mining faces), the extracted knowledge points were filtered for relevance. All 5 dimensions are directly related to the gas detector position and were retained.

[0061] Table 1. Definition of the dimensions of the cognitive state vector: Cognitive state vector dimension Initial cognitive state vector The initial score for each dimension was set to 0.50, representing a neutral state before the evaluation was initiated.

[0062] The mine assessment terminal A performs an adversarial interrogation process on EMP-047, setting a convergence sliding window. Convergence threshold Maximum number of follow-up questions The questioning process was executed for 6 rounds before triggering a termination condition, generating a cognitive state transition trajectory sequence, with a number of trajectory points. .

[0063] Table 2. Cognitive State Transfer Trajectory Sequence for EMP-047: Taking the fourth round as an example, the detected dimension was d4 (gas concentration threshold). After follow-up questioning in this round, the dimension 5 score was 0.58, a significant drop compared to the 0.72 score in the third round. The drop was [missing information]. Far exceeding the convergence threshold This anomaly will be captured in step 3.

[0064] The change matrix of the cognitive state transition trajectory sequence of EMP-047 is calculated round by round, and the inter-dimensional linkage influence matrix is ​​generated by accumulating the changes. Taking the 4th round (with a detected dimension of d4) as an example, the elements in the 4th row of the change matrix are calculated as follows: After accumulating six rounds of changes, the interdimensional linkage and influence matrix is ​​obtained. (Only rows with probed dimensions d4 and d5 are displayed; the values ​​in other rows are smaller): Key rows of the inter-dimensional linkage and influence matrix in Table 3 (dimensions 4 and 5): Linkage threshold Calculated based on the mean of the absolute values ​​of all off-diagonal elements plus 1.5 times the standard deviation, statistically, we obtain... .

[0065] Check off-diagonal elements: , All exceeded the linkage threshold, and the dimension pair (d4, d5) was marked as an abnormal linkage dimension pair. Marked as a negative abnormal linkage, it indicates that the score of the outflow assessment dimension has regressed when detecting the concentration threshold dimension. This is also a negative abnormal linkage.

[0066] The abnormal linkage dimension is compared with (d4, d5) (concentration threshold dimension and emission assessment dimension), the negative linkage direction, and the feature vector of non-diagonal linkage in row d4. In addition, the large language model for question-and-answer text input involved in rounds 1 and 4 generates the root cause misconception hypothesis: EMP-047 may confuse "gas concentration threshold" (the comparison between the concentration measurement value at a certain moment and the alarm / power outage standard) with "gas emission" (the volume of gas emitted per unit time, reflecting the gas occurrence state in the mine), mistakenly believing that the concentration threshold judgment depends on the calculation result of the emission, leading to a cognitively incorrect causal dependency between the two concepts.

[0067] Constructing a global cognitive state matrix, the final cognitive state vector of EMP-047 for the knowledge point of "handling excessive gas levels in mining faces" is: The scores for Dimension 1 (detection instrument type, 0.53) and Dimension 2 (detection location, 0.54) were too low, and were marked as systematic weaknesses.

[0068] Table 4. Initial knowledge vulnerability location results for Mine A EMP-047: The root cause of the misunderstanding of EMP-047 by Mine A is used to extract the desensitized misunderstanding structure feature vector, and the specific personnel identifier is replaced with "gas inspector" and the mine name is replaced with "underground mine".

[0069] Table 5. Desensitized Misunderstanding Structural Feature Vectors Uploaded by Mine A: Normalization of linkage intensity quantification value: Let the maximum absolute value of all off-diagonal elements in this assessment of the mine be 0.39, the minimum be 0.00, and the mean be 0.09, then: Mine B and Mine C simultaneously upload their respective sets of de-identified misunderstanding structural features.

[0070] The federated aggregation node summarizes the anonymized and misleading structural feature vectors uploaded from the three mines, receiving a total of 14 feature records, and performs cluster analysis. Taking the feature vector from mine A as an example... With a certain feature vector of mine B For example, let the weights be... , , , Calculate the weighted distance: The two are very close and are classified into the same deviation pattern cluster. After clustering, the deviation pattern cluster P-02 (confusion and emission measurement relationship, negative linkage) contains 9 records from three mines. The frequency of cross-mine occurrence is 9, and the number of mines involved is 3. It exceeds the preset commonality threshold (frequency ≥ 6), and is marked as a suspected industry common deviation pattern and distributed to each mine.

[0071] Mine A matched the suspected industry-wide common deviation pattern P-02 with the root cause misunderstanding hypothesis already generated in its mine, confirming that the deviation of EMP-047 matched P-02, and increased the confidence of this hypothesis based on cross-mine frequency. For employees in Mine A who did not independently discover P-02 in step 401, based on the confusion concept of P-02, the type was coded (metric relationship confusion, code 2) and the negative linkage direction was coded (code 0), locating two specific knowledge points: "gas concentration threshold judgment" and "gas emission calculation". Verification detection question groups were generated (e.g., "When the CH4 concentration at the upper corner of the working face is 1.4%, is it necessary to first calculate the absolute emission of the shift before deciding whether to cut off power and evacuate personnel?"), setting two different answer paths corresponding to correct understanding and confused understanding. Mine A organized 12 gas detectors to participate in the verification.

[0072] Table 6. Semantic matching results for the verification detection problem in Mine A (partial): Of the 12 participants in mine A, 8 were confirmed, resulting in a confirmation rate of 8%. .

[0073] The three mines uploaded the verification confirmation rates of P-02 to the federated aggregation node. The number of participants in verification for each mine were 12 for Mine A, 15 for Mine B, and 10 for Mine C, with confirmation rates of 0.667, 0.733, and 0.700, respectively. Calculate the weighted average confirmation rate across the mines: The industry confirmation threshold is preset to 0.60. P-02 was confirmed as a common industry-wide cognitive bias. Cross-mine confirmation statistics (involving 3 mines, 37 participants, weighted confirmation rate of 0.703) were attached and distributed to all mines. Mine A integrated P-02 into the vulnerability location results of EMP-047, updated the vulnerability type to "Industry-wide Common Cognitive Bias Type," and generated an enhanced coal mine safety knowledge vulnerability location result based on a large model.

[0074] Table 7. Results of Enhanced Knowledge Vulnerability Location in Mine A EMP-047: The data flow logic of the entire implementation process is as follows: Step 1: Extract five conditional dimensions from the textbook text to provide a structural foundation for the subsequent cognitive state vector; Step 2: Generate a migration trajectory sequence containing rating snapshots through six rounds of follow-up questions, where the regression value of dimension 5 in the fourth round (0.72→0.58) is the key data source for subsequent anomaly detection; Step 3: Accumulate the round-by-round changes in the trajectory sequence into a linkage influence matrix. If the threshold of 0.12 is exceeded, the abnormal linkage dimension pair is marked; Step 4 generates a root cause misunderstanding hypothesis based on the abnormal linkage pair and merges it with the results of the regular low-score dimension pair to output the initial vulnerability location result; Step 5 desensitizes the root cause misunderstanding hypothesis into a four-component structural feature vector and uploads it to protect the sensitive data of the mine, where the linkage strength normalized value Step 6: Clustering at the federated nodes reveals a common cross-mine pattern for P-02. The weighted distance between the feature vector of Mine A and the corresponding feature vector of Mine B is 0.006, confirming that the two mines have a high degree of consistency in this type of deviation. Step 7: Verification questions are generated and semantic similarity data from the answers of 12 people is obtained, with a confirmation rate of 0.667%. Step 8: Data from 37 people across the three mines is aggregated to calculate a weighted confirmation rate of 0.703. After exceeding the threshold, P-02 is marked as a common industry-wide root cause cognitive bias and fed back to Mine A. Finally, the vulnerability type of EMP-047 is upgraded from "mine-specific root cause bias type" to "industry-wide common root cause bias type", achieving penetrating localization from single-mine surface scoring to deep cross-mine conceptual confusion root cause.

[0075] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for locating knowledge gaps in coal mine safety based on a large model, characterized in that, Includes the following steps: Obtain a set of conditional knowledge points related to the job, where each knowledge point is associated with a set of applicable conditions, and each applicable condition corresponds to a dimension of the cognitive state vector. For each knowledge point in the set of conditional knowledge points related to the job, an adversarial questioning process is executed. In each round of questioning, a conditional dimension is selected as the probed dimension and the cognitive state vector is updated. Snapshots of the cognitive state vectors in each round are recorded to generate a cognitive state transition trajectory sequence. Calculate the change matrix of all dimensions between adjacent trajectory points for the cognitive state migration trajectory sequence, accumulate and sum the change matrix to generate the inter-dimensional linkage influence matrix, and identify the abnormal linkage dimension pairs corresponding to the off-diagonal elements in the inter-dimensional linkage influence matrix whose absolute value exceeds the linkage threshold. Based on the aforementioned abnormal linkage dimensions and their associated follow-up questions and answers, a root cause misunderstanding hypothesis is generated through a large language model. This root cause misunderstanding hypothesis is then integrated with the results of conventional vulnerability analysis to generate a coal mine safety knowledge vulnerability location result based on the large model.

2. The method for locating knowledge gaps in coal mine safety based on a large model according to claim 1, characterized in that, The set of job-related conditional knowledge points includes: Input the text of the coal mine safety knowledge system into a large language model to extract knowledge points and their applicable conditions. Based on the job descriptions of target employees, the key functions in the job descriptions are semantically matched with the applicable scenarios of knowledge points. When the matching degree exceeds a preset relevance threshold, the knowledge point is retained, and a set of job-related conditional knowledge points is generated.

3. The method for locating knowledge gaps in coal mine safety based on a large model according to claim 1, characterized in that, Each trajectory point in the cognitive state transition trajectory sequence contains the cognitive state vector of the current round, the condition dimension identifier being probed in this round, and the follow-up questions and answers for this round; the scores of each dimension of the cognitive state vector are processed using a mean normalization method based on the range, and their values ​​range from zero to one in a closed interval; the termination condition of the adversarial follow-up questioning process is: The change in scores across all dimensions of the cognitive state vector is below the convergence threshold within a consecutive preset sliding window length, or the number of follow-up questions reaches the preset maximum number of rounds. The convergence threshold constrains the upper limit of the absolute value of the difference between two adjacent rounds of scores. When the absolute value of the difference between two adjacent rounds of scores in any dimension is less than the convergence threshold within a continuous sliding window length, the cognitive state vector is determined to have converged.

4. The method for locating coal mine safety knowledge gaps based on a large model according to claim 1, characterized in that, The step of accumulating and summing the change matrix to generate the inter-dimensional linkage influence matrix includes: For each round of follow-up questions, the change matrix records the change in scores for each dimension only in the row corresponding to the dimension currently being probed, and sets the elements of the remaining rows to zero; Each element of the interdimensional linkage influence matrix is ​​the cumulative sum of the change matrix elements at the corresponding position across all follow-up rounds, where the element in the i-th row and j-th column represents the cumulative change in the score of dimension j when probing dimension i throughout the entire follow-up process; Since each round of probing only probes one conditional dimension, the off-diagonal elements of the inter-dimensional linkage influence matrix specifically reflect the directional linkage relationship between dimensions; The linkage threshold is determined based on the mean of the absolute values ​​of all off-diagonal elements in the inter-dimensional linkage influence matrix plus a preset multiple of the standard deviation.

5. The method for locating coal mine safety knowledge gaps based on a large model according to claim 1, characterized in that, When identifying abnormal linkage dimension pairs, the linkage influence direction information is also recorded. When the element in the linkage influence matrix between dimensions is greater than the linkage threshold, it is marked as a positive abnormal linkage. When the element in the linkage influence matrix between dimensions is less than the negative value of the linkage threshold, it is marked as a negative abnormal linkage. The process of generating root cause misunderstanding hypotheses through a large language model includes: inputting the dimension identifiers, direction of influence, feature vectors of influence, and related follow-up questions and answers, as well as the knowledge points, of the abnormal linkage dimension pairs into the large language model; using the dimension identifiers to determine the specific conditional concepts involved in the confusion; using the direction of influence to distinguish between positive and negative confusion; using the magnitude of the influence feature vectors to reflect the degree of confusion; and combining the specific expressions of practitioners in the follow-up questions and answers to infer the specific ways in which the concepts are confused, thereby generating a root cause misunderstanding hypothesis description that includes a description of the confusion relationship and the basis for inference.

6. The method for locating knowledge gaps in coal mine safety based on a large model according to claim 1, characterized in that, The generation of the conventional vulnerability analysis results includes: For knowledge points where no abnormal linkage was detected, the final cognitive state vector was determined according to the cognitive state vector convergence judgment logic, and the final cognitive state vectors of each knowledge point were arranged according to the knowledge point dimension to form a global cognitive state matrix. Perform column-by-column analysis on the global cognitive state matrix to identify conditional dimensions with low scores across multiple knowledge points and mark them as systematically weak conditional dimensions. Perform a row-by-row analysis on the global cognitive state matrix to identify cases where a certain knowledge point has a low-scoring dimension, but the same dimension scores normally in other knowledge points, and mark them as local vulnerability knowledge points. The vulnerability types in the coal mine safety knowledge vulnerability localization results based on the large model include mine-specific root cause deviation type, systemic weakness type, and local vulnerability type, and are sorted according to the rule that root cause deviation takes precedence over conventional vulnerability.

7. The method for locating coal mine safety knowledge gaps based on a large model according to claim 1, characterized in that, Also includes: For the root cause misunderstanding hypothesis descriptions generated locally in each mine, the desensitized misunderstanding structure feature vector is extracted. The misunderstanding structure feature vector contains four components: knowledge domain encoding, type encoding of confused concept pairs, direction encoding of linkage influence, and linkage intensity quantification value. Replace specific personnel information and mine sensitive parameters with typified labels, generate a set of desensitized misunderstanding structure features for each mine, and upload it to the federated aggregation node; The federated aggregation node performs cluster analysis on the desensitized misunderstanding structure feature vectors uploaded by all mines using a clustering algorithm. The similarity measure is based on weighted distance. The weighted distance uses coding consistency judgment for knowledge domain coding, type coding of confused concept pairs, and coding of linkage influence direction. The linkage intensity quantification value uses numerical difference measurement. The sum of the preset weights corresponding to each component is one. The weighted distance is defined as the sum of the absolute value of the difference between the contribution weights of the discrete components and the continuous components when the codes of the discrete components are inconsistent, multiplied by the corresponding weights. Deviation pattern clusters that occur more frequently across mines than a preset commonality threshold are marked as suspected industry common deviation patterns. A set of suspected industry common deviation patterns is generated and distributed to each mine.

8. The method for locating knowledge gaps in coal mine safety based on a large model according to claim 7, characterized in that, Also includes: Each mine matches the suspected common industry deviation patterns with the root cause misunderstanding hypothesis descriptions already generated in its own mine. For patterns that are not found in the mine but belong to the suspected common industry deviation patterns, the type encoding and linkage influence direction encoding based on the confusion concept of the pattern are used to locate the corresponding knowledge points and conditional dimensions in the mine's safety knowledge system. The content of the located knowledge points, conditional dimension information and structural description of the deviation pattern are input into the large language model to generate a localized verification and detection question set. Obtain the response text of practitioners to the localized verification probe question set, calculate the semantic matching degree between the response text and the expected answer of the deviation hypothesis and the expected answer of the correct answer, respectively. When the semantic matching degree between the practitioner's answer and the expected answer of the deviation hypothesis is higher than the semantic matching degree between the practitioner's answer and the expected answer of the correct answer, it is determined that the practitioner has verified and confirmed the deviation pattern, and the verification confirmation rate of each deviation pattern is calculated.

9. The method for locating knowledge gaps in coal mine safety based on a large model according to claim 8, characterized in that, Also includes: Each mine uploads the verification confirmation rate of each deviation mode to the federated aggregation node. The federated aggregation node summarizes and calculates the cross-mine weighted average confirmation rate. The cross-mine weighted average confirmation rate is the sum of the products of the sample size of practitioners participating in the verification of each mine and the corresponding verification confirmation rate, divided by the sum of the sample sizes of practitioners participating in the verification of each mine. Deviation patterns where the weighted average confirmation rate across mines exceeds the preset industry confirmation threshold are marked as confirmed common root cause cognitive biases in the industry and distributed to each mine. Each mine integrates and updates the confirmed industry-wide common root cause cognitive biases with the coal mine safety knowledge vulnerability location results based on the large model. An industry-wide common root cause bias category is added to the vulnerability types. The vulnerability types are then comprehensively sorted according to the rules that industry-wide common root cause bias takes precedence over mine-specific root cause bias, root cause bias takes precedence over conventional vulnerabilities, and high cross-mine confirmation rate takes precedence over low confirmation rate, thus generating an enhanced coal mine safety knowledge vulnerability location result based on the large model.

10. A coal mine safety knowledge vulnerability localization system based on a large model, used to execute the coal mine safety knowledge vulnerability localization method based on a large model as described in any one of claims 1 to 9, characterized in that, include: The knowledge point acquisition module is used to acquire the text of the coal mine safety knowledge system and generate a set of job-related conditional knowledge points based on the job description. The follow-up questioning module is used to execute an adversarial follow-up questioning process for each knowledge point in the set of conditional knowledge points related to the job. In each round of follow-up questioning, a conditional dimension is selected as the probed dimension and the cognitive state vector is updated. The module records snapshots of the cognitive state vector in each round and generates a sequence of cognitive state transition trajectories. The linkage analysis module is used to calculate the change matrix of the cognitive state migration trajectory sequence, accumulate and sum the change matrix to generate the inter-dimensional linkage influence matrix, and identify abnormal linkage dimension pairs corresponding to off-diagonal elements whose absolute values ​​exceed the linkage threshold. The vulnerability localization module is used to generate root cause misunderstanding hypotheses based on the abnormal linkage dimension and its associated follow-up questions and answers through a large language model, and to integrate the root cause misunderstanding hypotheses with the results of conventional vulnerability analysis to generate a coal mine safety knowledge vulnerability localization result based on a large model.

Images