A method for intelligent examination of hidden disaster-causing reports in coal mines

Abstract

The present application relates to the technical field of intelligent examination of hidden disaster-causing reports in coal mines, in particular to a method, device, equipment and computer storage medium for intelligent examination of hidden disaster-causing reports in coal mines.The present application constructs a domain knowledge graph containing hidden disaster-causing factors and regulatory standards, and configures a hybrid reasoning engine;the report is subjected to multi-level preprocessing and semantic understanding, key entities are extracted to form structured information;the structured information is compared with the knowledge graph in multiple dimensions, the risk level is verified by using a probabilistic graph model, and the matching degree of prevention and control measures and the compliance with regulations are checked;examination opinions are generated and closed-loop optimization is carried out based on feedback data.The present application can improve the examination efficiency and accuracy, reduce the factor omission rate and risk misjudgment rate, and realize the reuse of expert experience and the unification of examination standards.

Description

Technical Field

[0001] This invention relates to the field of intelligent review technology for hidden disaster reports in coal mines, and in particular to an intelligent review method, device, equipment, and computer storage medium for hidden disaster reports in coal mines. Background Technology

[0002] In the field of reviewing reports on hidden disaster-causing factors in coal mines, existing technologies primarily rely on expert experience for judgment. Reviewers need to verify the completeness of geological data, the accuracy of risk assessment, and the rationality of proposed measures in the report, based on regulations and standards such as the "Coal Mine Safety Regulations" and their own professional knowledge. However, this traditional review model has significant drawbacks: First, the experience and knowledge of senior experts exist implicitly, lacking a systematic knowledge representation and quantitative reasoning mechanism, making it difficult to achieve large-scale reuse and inheritance of experience; second, different technical personnel have significant differences in their understanding of the same geological conditions, lacking unified quantitative assessment standards, resulting in poor consistency in review results, a high omission rate of 8% to 15% in identifying hidden disaster-causing factors, and a high misjudgment rate of 10% to 20% in risk levels; third, the cycle from data collection and comprehensive analysis to conclusion formation is long, with manual review of a single report taking 4 to 8 hours, resulting in low work efficiency and difficulty in meeting the needs of dynamic production decision-making; furthermore, once risk perception lags behind the progress of mining and production, the best prevention and control opportunity is easily missed, seriously restricting the improvement of coal mine safety governance effectiveness.

[0003] Therefore, there is an urgent need for an intelligent review method that can automate the review process, structure knowledge representation, and quantify the reasoning process, in order to overcome the technical problems of low efficiency, poor consistency, and difficulty in knowledge reuse in existing technologies. Summary of the Invention

[0004] Therefore, the technical problem to be solved by the present invention is to overcome the problems of low efficiency, poor consistency, difficulty in knowledge reuse, and difficulty in achieving automation and quantitative reasoning in the review of coal mine hidden disaster reports due to reliance on human experience.

[0005] To address the aforementioned technical problems, this invention provides an intelligent review method for hidden disaster reports in coal mines, comprising:

[0006] A domain knowledge graph containing hidden disaster-causing factors, causal conditions, prevention and control measures, and regulatory standards is constructed based on multi-source heterogeneous data, and a hybrid reasoning engine supporting rule-based reasoning and similarity-based reasoning is configured. Multi-level preprocessing and domain semantic understanding are performed on the text of the coal mine hidden disaster-causing factor survey report to be reviewed, and key entities, risk factors, assessment conclusions and numerical parameters in the report are extracted to form structured information; The structured information is compared with the domain knowledge graph in multiple dimensions. The risk level judgment in the report is inferred and verified using a probabilistic graphical model. The matching degree between the prevention and control measures and risk factors and the compliance with regulations are also checked. Based on the comparison and reasoning verification results, review opinions are generated, including problem location, supporting explanations, and modification suggestions. The domain knowledge graph and semantic understanding model are then optimized and updated in a closed loop based on the review feedback data.

[0007] Preferably, the construction of a domain knowledge graph based on multi-source heterogeneous data, including hidden disaster-causing factors, causative conditions, prevention and control measures, and regulatory standards, includes: Collect legal and standard knowledge, domain expert knowledge, and historical case knowledge, and extract mandatory requirements, review points, and case tags to form a multi-source knowledge base; Define the core concept classes of hidden disaster-causing factors, causative conditions, disaster-prone environment, disaster-bearing bodies, prevention and control measures, risk assessment, regulations and standards, and review points, and construct a multi-dimensional relationship network connecting the above concept classes; A sequence labeling model based on a pre-trained language model is used to perform entity recognition on the multi-source knowledge base, an attention mechanism neural network model is used to extract relations, and a domain knowledge graph containing node attributes and relation attributes is generated through knowledge fusion and conflict resolution.

[0008] Preferably, the multi-level preprocessing and domain semantic understanding of the coal mine hidden disaster-causing factor survey report text to be reviewed includes: It unifies the processing of reports awaiting review in various formats, and uses parsing and recognition technologies to extract plain text while retaining chapter structure information; Accurate word segmentation and part-of-speech tagging are performed based on a domain dictionary, and words are filtered using a domain-specific stop word list; Restore the report's chapter hierarchy and establish a tree-like paragraph system that reflects the report's logical structure; The text in the tree-like paragraph system is encoded into semantic vectors using a fine-tuned semantic understanding model to support vector similarity-based analysis. Key entities are identified from the semantic vectors, factor types and governance measures are extracted to form a risk factor set, and risk level determination and assessment method descriptions are extracted to form an assessment conclusion set.

[0009] Preferably, the step of performing a multi-dimensional comparison between the structured information and the domain knowledge graph includes: Verify the completeness of the chapters and the coverage of the content in the report to be reviewed against the standards; By combining semantic similarity calculation with knowledge graph matching, the extracted set of risk factors is compared with the hidden disaster-causing factor entities in the domain knowledge graph, indicating missing factors and marking non-standard expressions. Verify the sufficiency of the basic data supporting the analysis conclusions and generate the results of the element integrity review.

[0010] Preferably, the step of using a probabilistic graphical model to infer and verify the risk level determination in the report includes: Construct a probabilistic graph network structure that includes geological condition nodes, engineering factor nodes, and risk level nodes, and input the reported observation evidence as evidence nodes into the probabilistic graph network; Calculate the posterior probability distribution of the risk level nodes, and compare the calculated posterior probability with the risk level determined in the report; When the probability deviation exceeds a preset threshold, a review prompt is triggered, and a risk assessment accuracy verification report containing the comprehensive reasoning results of multi-source evidence is output.

[0011] Preferably, the verification of the matching degree between the prevention and control measures and risk factors, as well as regulatory compliance, includes: By querying the applicable relationships in the knowledge graph, we can verify the matching relationship between the types of governance measures and risk factors, as well as the commensurability between the strength of the measures and the risk level. The rules engine is used to automatically check whether the content of the report to be reviewed complies with mandatory requirements and with prohibitions. Based on historical implementation effects analyzed through case library data mining, the effectiveness of governance measures is assessed, and the results of compliance and effectiveness review of the proposed measures are generated.

[0012] Preferably, the step of generating review opinions that include problem location, supporting explanations, and modification suggestions based on the comparison results and reasoning verification results includes: Define the critical issues that affect the core conclusions of the report, the general issues that affect the quality of the report, and the optimization issues that are improvement suggestions, and classify and prioritize the identified issues; It accurately pinpoints the location of the problem and automatically links it to relevant legal and standard provisions, knowledge graph nodes, and historical case references as supporting evidence. The recommendations are presented in layers: principle-based suggestions, methodological suggestions, and reference suggestions, forming a complete description of modification suggestions for each problem.

[0013] This invention also provides an intelligent review device for hidden disaster reports in coal mines, comprising: The knowledge graph construction module is used to build a domain knowledge graph based on multi-source heterogeneous data, which includes hidden disaster-causing factors, causative conditions, prevention and control measures and legal standards, and is configured with a hybrid reasoning engine that supports rule-based reasoning and similarity-based reasoning. The semantic understanding and extraction module is used to perform multi-level preprocessing and domain semantic understanding on the text of the coal mine hidden disaster-causing factor survey report to be reviewed, and to extract key entities, risk factors, assessment conclusions and numerical parameters in the report to form structured information. The comparison and reasoning verification module is used to perform multi-dimensional comparison between the structured information and the domain knowledge graph, use a probabilistic graphical model to reason and verify the risk level judgment in the report, and check the matching degree between the prevention and control measures and risk factors and the compliance with regulations. The feedback generation and optimization module is used to generate review opinions that include problem location, supporting explanations, and modification suggestions based on the comparison results and reasoning verification results, and to perform closed-loop optimization and updates on the domain knowledge graph and semantic understanding model based on the review feedback data.

[0014] This invention also provides an intelligent review device for hidden disaster reports in coal mines, comprising: Memory, used to store computer programs; A processor is used to implement the steps of the above-mentioned intelligent review method for reporting hidden disasters in coal mines when executing the computer program.

[0015] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described intelligent review method for reporting hidden disasters in coal mines.

[0016] The technical solution of the present invention has the following advantages compared with the prior art: The intelligent review method for coal mine hidden disaster reports described in this invention achieves the systematic solidification and large-scale reuse of expert experience by constructing a domain knowledge graph and configuring a hybrid inference engine. Through multi-level preprocessing and domain semantic understanding, unstructured reports are transformed into structured information, and Bayesian networks are used to verify risk level determination through probabilistic reasoning. This effectively overcomes the subjective differences and cognitive limitations of manual review, significantly reducing the omission rate of hidden disaster-causing factors and the misjudgment rate of risk levels. Simultaneously, this method shortens the review time for a single report from several hours to 10-20 minutes, greatly improving review efficiency. Furthermore, a closed-loop optimization mechanism enables continuous system evolution and unified review standards, thereby effectively meeting the demands of dynamic production decision-making for rapid response and high accuracy while ensuring the objectivity and consistency of review results. Attached Figure Description

[0017] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1 This is a flowchart illustrating the implementation of an intelligent review method for concealed disaster reports in coal mines provided by this invention. Figure 2 This is a structural block diagram of an intelligent review device for hidden disaster reports in coal mines, provided in an embodiment of the present invention. Detailed Implementation

[0018] The core of this invention is to provide an intelligent review method, device, equipment, and computer storage medium for coal mine hidden disaster reports, which effectively solves the problems of low review efficiency, poor consistency of results, difficulty in reusing expert knowledge, and lack of quantitative reasoning mechanism in the prior art, and realizes the automation, intelligence, and continuous evolution of the review process.

[0019] To enable those skilled in the art to better understand the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please refer to Figure 1. Figure 1 The flowchart illustrates the implementation of an intelligent review method for concealed disaster reports in coal mines provided by this invention; the specific operation steps are as follows: S101: Construct a domain knowledge graph based on multi-source heterogeneous data, which includes hidden disaster-causing factors, causative conditions, prevention and control measures, and regulatory standards, and configure a hybrid reasoning engine that supports rule-based reasoning and similarity-based reasoning. S102: Perform multi-level preprocessing and domain semantic understanding on the text of the coal mine hidden disaster-causing factor survey report to be reviewed, and extract key entities, risk factors, assessment conclusions and numerical parameters in the report to form structured information; S103: Perform multi-dimensional comparison between the structured information and the domain knowledge graph, use a probabilistic graphical model to infer and verify the risk level judgment in the report, and check the matching degree between the prevention and control measures and risk factors and the compliance with regulations. S104: Generate review comments that include problem location, supporting explanations, and modification suggestions based on the comparison results and reasoning verification results, and perform closed-loop optimization and updates on the domain knowledge graph and semantic understanding model based on the review feedback data.

[0021] In some embodiments, the probabilistic graphical model employs a Bayesian network. When using a Bayesian network to infer and verify risk level determination, the process of calculating the posterior probability distribution of the risk level nodes is as follows: Let the joint probability distribution of each variable node in the Bayesian network be... ,in This represents the variable of the i-th node. express The set of parent nodes of E; given the observation value e of the evidence node set E, the posterior probability distribution of the target variable Q is: , where other represents other latent variables besides Q and E, and the risk level is quantitatively verified by comparing the posterior probability with the report judgment result.

[0022] In other embodiments, the hybrid reasoning engine includes a deterministic rule-based logical deduction component and a vector-representation-based similarity analogy component. The rule-based reasoning component employs the SWRL rule engine and uses the Cypher query language to perform hard compliance checks on regulatory and standard clauses in the graph database. The similarity reasoning component uses the Node2Vec algorithm to learn the vector representations of knowledge graph nodes and calculates the cosine similarity between nodes to achieve case recommendation and analogy analysis.

[0023] In one specific embodiment, Neo4j graph database is used as the core storage engine when constructing the domain knowledge graph. During the knowledge fusion process, low-confidence results are filtered through a multi-model voting decision mechanism, where the confidence score is defined as the maximum value of the probability distribution of the predicted entity category by the model, with a value range of [0,1]. When the score is lower than a preset threshold, it is submitted for expert verification.

[0024] Specifically, during the multi-level preprocessing, for scanned PDF documents, OCR recognition technology is used to extract image text, and layout analysis algorithms are used to identify headers, footers, and column information. The domain semantic understanding employs a fine-tuned BERT model, encoding text fragments into 768-dimensional or 1024-dimensional semantic vectors. The vector dimension corresponds to the model's hidden layer size and is used to represent the deep semantic features of technical terms.

[0025] It should be noted that this embodiment, by constructing a domain knowledge graph and configuring a hybrid inference engine, achieves the explicit solidification and large-scale reuse of implicit expert experience in the field of coal mine safety. By integrating deterministic rule reasoning and uncertain probabilistic reasoning, it effectively overcomes the subjectivity and inconsistency problems caused by experience differences in traditional manual review. This embodiment reduces the review time of a single hidden disaster report from several hours in the traditional manual mode to 10 to 20 minutes, reduces the omission rate of hidden disaster factor identification to below 2%, and controls the misjudgment rate of risk level determination to within 5%, significantly improving review efficiency and accuracy.

[0026] Based on the above embodiments, in some embodiments, the construction of a domain knowledge graph based on multi-source heterogeneous data, including hidden disaster-causing factors, causative conditions, prevention and control measures, and regulatory standards, includes: Collect legal and standard knowledge, domain expert knowledge, and historical case knowledge, and extract mandatory requirements, review points, and case tags to form a multi-source knowledge base; Define the core concept classes of hidden disaster-causing factors, causative conditions, disaster-prone environment, disaster-bearing bodies, prevention and control measures, risk assessment, regulations and standards, and review points, and construct a multi-dimensional relationship network connecting the above concept classes; A sequence labeling model based on a pre-trained language model is used to perform entity recognition on the multi-source knowledge base, an attention mechanism neural network model is used to extract relations, and a domain knowledge graph containing node attributes and relation attributes is generated through knowledge fusion and conflict resolution.

[0027] In some embodiments, the multi-relationship network includes causal relationships, correlational relationships, hierarchical relationships, temporal relationships, spatial relationships, and applicability relationships. Causal relationships connect causal conditions and hidden disaster-causing factors to express risk transmission paths; spatial relationships express the spatial configuration of geological bodies; and applicability relationships connect prevention and control measures and hidden disaster-causing factors to guide the matching and review of measures.

[0028] In other embodiments, the sequence labeling model based on the pre-trained language model adopts the BIO labeling scheme. During entity recognition, a remote supervision mechanism is introduced to expand the training data using existing relation instances in the knowledge base, and an active learning strategy is used to prioritize labeling samples with high model uncertainty. The model uncertainty is quantified by prediction probability entropy, with higher entropy values ​​indicating greater uncertainty.

[0029] In one specific embodiment, the pre-trained language model uses BERT-base-Chinese or RoBERTa-wwm-ext, and a coal mining corpus containing a professional terminology dictionary, regulatory standard texts, and historical report samples is constructed to fine-tune the basic model for domain adaptation. Each knowledge graph node is designed with a complete attribute system, including identification attributes, feature attributes, state attributes, relation attributes, source attributes, and application attributes. The feature attributes of the node cover fault drop, dip angle, and extension length; the state attributes cover the degree of investigation, risk level, and trend of change.

[0030] Specifically, the knowledge fusion and conflict resolution include: filtering low-confidence results through a multi-model voting decision-making mechanism, performing entity alignment to merge different representations of the same entity, performing relation disambiguation to resolve conflicts, storing high-confidence results in a graph database, and submitting low-confidence results for expert verification, thus forming a human-machine collaborative quality assurance mechanism.

[0031] It should be noted that this embodiment significantly improves the accuracy of entity recognition for coal mine professional terms by combining a domain-adapted pre-trained language model with an active learning strategy; and ensures the consistency of multi-source heterogeneous knowledge fusion through multi-model voting and entity alignment mechanisms, providing a high-quality, conflict-free knowledge foundation for subsequent intelligent review and reasoning.

[0032] Based on the above embodiments, in some embodiments, the multi-level preprocessing and domain semantic understanding of the coal mine hidden disaster-causing factor survey report text to be reviewed includes: It unifies the processing of reports awaiting review in various formats, and uses parsing and recognition technologies to extract plain text while retaining chapter structure information; Accurate word segmentation and part-of-speech tagging are performed based on a domain dictionary, and words are filtered using a domain-specific stop word list; Restore the report's chapter hierarchy and establish a tree-like paragraph system that reflects the report's logical structure; The text in the tree-like paragraph system is encoded into semantic vectors using a fine-tuned semantic understanding model to support vector similarity-based analysis. Key entities are identified from the semantic vectors, factor types and governance measures are extracted to form a risk factor set, and risk level determination and assessment method descriptions are extracted to form an assessment conclusion set.

[0033] In some embodiments, the unified processing of reports to be reviewed in multiple formats includes: extracting text streams using a PDF parsing engine, processing image text in scanned documents using OCR recognition technology, identifying header, footer, and column information using a layout analysis algorithm, and outputting plain text data that retains the original chapter structure information.

[0034] In other embodiments, the precise word segmentation and part-of-speech tagging based on the domain dictionary adopts a dictionary matching plus conditional random field sequence tagging method. Paragraph structure recognition adopts a combination of title keyword matching, chapter numbering pattern recognition and font format analysis. Chapter numbering pattern recognition adopts regular expression matching for numbering levels in the form of "1", "1.1", "1.1.1".

[0035] In one specific embodiment, the semantic similarity calculation uses the cosine similarity formula: Where A and B represent the semantic vectors of the text fragment of the report to be reviewed and the standard entity of the knowledge graph, respectively, and n represents the number of dimensions of the vector. and Let represent the component values ​​of vectors A and B in the ii-th dimension, respectively. Set a similarity threshold (e.g., 0.85). When the maximum similarity is lower than this threshold, it is judged as a non-standard representation and mapped to the closest standard entity.

[0036] Specifically, the extraction of the risk factor set adopts a sequence labeling plus multi-field extraction model, and the extracted content includes factor type, location description, degree of identification, characteristic parameters, risk analysis and governance measures; the extraction of the assessment conclusion set adopts template matching plus machine reading comprehension, and the extracted content includes risk level determination, assessment method description and uncertainty description.

[0037] It should be noted that this embodiment effectively solves the problems of diverse formats, complex professional terminology, and high degree of unstructured nature in coal mine hidden disaster reports by cascading multi-level preprocessing and deep semantic understanding technology. It achieves automated and accurate extraction of key review elements, providing a high-quality structured data foundation for subsequent comparison review and Bayesian network inference.

[0038] Based on the above embodiments, in some embodiments, the multi-dimensional comparison of the structured information with the domain knowledge graph includes: Verify the completeness of the chapters and the coverage of the content in the report to be reviewed against the standards; By combining semantic similarity calculation with knowledge graph matching, the extracted set of risk factors is compared with the hidden disaster-causing factor entities in the domain knowledge graph, indicating missing factors and marking non-standard expressions. Verify the sufficiency of the basic data supporting the analysis conclusions and generate the results of the element integrity review.

[0039] In some embodiments, the comparison with the standard to check the completeness of the chapters in the report to be reviewed includes: matching the sequence of chapter titles at all levels under the root node of the report with the list of required chapters in the "Coal Mine Safety Production Standardization Management System" specification template layer by layer, traversing the text paragraphs in each chapter, calculating the keyword coverage rate to determine the completeness of the specified content, and outputting a list of structural defects marked with missing chapter numbers and chapters with insufficient content coverage.

[0040] In other embodiments, the step of combining semantic similarity calculation with knowledge graph matching includes: calculating the cosine similarity between the extracted entity vector and the standard entity vector in the graph; traversing all disaster-causing factor nodes that should exist in the current geological scenario in the graph; if no corresponding entity is detected in the report, it is marked as an omitted factor; and outputting a risk factor comparison report containing the name of the omitted factor, the original text of the non-standard expression, and standardized suggestions.

[0041] In one specific embodiment, the verification of the sufficiency of the basic data supporting the analysis conclusion includes: grouping and statistically analyzing the number, type distribution, and time span of measured data points by disaster type, logically comparing them with the corresponding minimum sample size threshold in the professional standard, and outputting the element integrity review result indicating whether the amount of data meets the requirements of the professional standard.

[0042] Specifically, the multidimensional comparison sequentially performs chapter structure and content coverage review, risk factor semantic comparison, and basic data sufficiency verification. The list of structural defects output in the previous step serves as the contextual constraint for subsequent content comparison, ensuring the rigor of the review logic.

[0043] It should be noted that this embodiment achieves automated quantitative assessment of the compliance of report format and the completeness of content through a multi-level comparison mechanism, effectively avoiding the problem of chapter omissions or missing key data due to negligence in manual review, and significantly improving the objectivity and consistency of review results.

[0044] Based on the above embodiments, in some embodiments, the step of using a probabilistic graphical model to infer and verify the risk level determination in the report includes: Construct a probabilistic graph network structure that includes geological condition nodes, engineering factor nodes, and risk level nodes, and input the reported observation evidence as evidence nodes into the probabilistic graph network; Calculate the posterior probability distribution of the risk level nodes, and compare the calculated posterior probability with the risk level determined in the report; When the probability deviation exceeds a preset threshold, a review prompt is triggered, and a risk assessment accuracy verification report containing the comprehensive reasoning results of multi-source evidence is output.

[0045] In some embodiments, the probabilistic graphical network employs a Bayesian network. When constructing the Bayesian network structure, geological condition nodes and engineering factor nodes serve as parent nodes pointing to risk level nodes, forming a causal dependency. The reported observational evidence includes measured values ​​of gas content and fault development degree, which are instantiated as evidence nodes in the Bayesian network.

[0046] In other embodiments, the posterior probability distribution of the risk level nodes is calculated using an exact inference algorithm or an approximate sampling algorithm. The risk level corresponding to the maximum probability in the posterior probability distribution is taken as the risk level estimated by the system, or the expected value of the probability of each level is calculated and its consistency is verified with the risk level determined in the report.

[0047] In one specific embodiment, the formula for calculating the posterior probability distribution of risk level nodes is: Where Risk represents the risk level node variable, r represents the specific risk level value (e.g., low risk, medium risk, high risk), Evidence represents the set of evidence nodes, and e represents the observed value of the evidence node. The prior probability representing the risk level. This represents the likelihood probability of evidence appearing under a given risk level, with the normalization factor in the denominator.

[0048] Specifically, after the review prompt is triggered, the system traces back the inference path of the Bayesian network to identify the key influencing factors that contribute the most to the posterior probability. The key influencing factors are determined by calculating the probability influence factor of each evidence node. The influence factor is defined as the difference in posterior probability when the evidence node takes the observed value and the default value.

[0049] It should be noted that this embodiment uses Bayesian network quantitative reasoning to transform the risk assessment of hidden disaster-causing factors from qualitative judgment to quantitative calculation, effectively overcoming the problem of misjudgment caused by subjective experience differences in manual review, significantly improving the accuracy and consistency of risk level determination, and enhancing the interpretability of review conclusions through the identification of key influencing factors.

[0050] Based on the above embodiments, in some embodiments, the verification of the matching degree between prevention and control measures and risk factors and regulatory compliance includes: By querying the applicable relationships in the knowledge graph, we can verify the matching relationship between the types of governance measures and risk factors, as well as the commensurability between the strength of the measures and the risk level. The rules engine is used to automatically check whether the content of the report to be reviewed complies with mandatory requirements and with prohibitions. Based on historical implementation effects analyzed through case library data mining, the effectiveness of governance measures is assessed, and the results of compliance and effectiveness review of the proposed measures are generated.

[0051] In some embodiments, the querying of applicable relationships in the knowledge graph includes: starting from the risk factor entity, retrieving associated prevention and control measure nodes along predefined applicable relationship edges, comparing whether the measure types extracted from the report are within the recommended measure set, and verifying whether the measure strength parameter meets the proportionality threshold requirement based on the risk level attribute value.

[0052] In other embodiments, the automatic check using a rule engine includes: comparing key statements in the report with IF-THEN logical expressions in the rule base one by one, focusing on verifying whether mandatory exploration quantities are omitted or whether there is a mining layout that violates prohibitions, and outputting regulatory compliance check results containing an index of specific violation clauses.

[0053] In one specific embodiment, the evaluation of the effectiveness of governance measures includes: executing data mining algorithms to statistically analyze the implementation effect records of similar governance measures in historical scenarios, and calculating the success rate index and cost-benefit ratio. If the historical success rate of the current measure is lower than a preset threshold or the cost-benefit ratio is significantly worse than the best case of its kind, then the measure is deemed insufficiently effective. The formula for calculating the success rate index is: ,in This indicates the number of times the measure has been successfully implemented in historical cases. This indicates the total number of times the measure has been implemented.

[0054] Specifically, the proposed measures' compliance and effectiveness review results integrate matching deviations, regulatory violations, and effectiveness assessment conclusions, thereby shifting the review of prevention and control measures from qualitative experience-based judgment to quantitative automated verification.

[0055] It should be noted that this embodiment combines semantic association queries of knowledge graphs with logical judgments of rule engines to achieve automated verification of prevention and control measures review, effectively avoiding omissions or misjudgments caused by differences in expert subjective cognition. At the same time, the effectiveness evaluation mechanism based on historical data ensures the economic rationality and technical feasibility of governance recommendations.

[0056] Based on the above embodiments, in some embodiments, the step of generating review opinions containing problem location, explanation of basis, and modification suggestions based on comparison results and inference verification results includes: Define the critical issues that affect the core conclusions of the report, the general issues that affect the quality of the report, and the optimization issues that are improvement suggestions, and classify and prioritize the identified issues; It accurately pinpoints the location of the problem and automatically links it to relevant legal and standard provisions, knowledge graph nodes, and historical case references as supporting evidence. The recommendations are presented in layers: principle-based suggestions, methodological suggestions, and reference suggestions, forming a complete description of modification suggestions for each problem.

[0057] In some embodiments, the serious problems include omission of major hidden disaster-causing factors, serious misjudgment of risk level, falsification of key data, or violation of mandatory regulations and standards, and the handling requirement is that the data must be revised and resubmitted; the general problems include incomplete data description, loopholes in analysis logic, or weak targeting of measures and suggestions, and the handling requirement is that the suggestions be revised and improved and then confirmed; the suggestions for optimization include insufficient standardization of expression, poor clarity of charts and graphs, or insufficient citation of case studies, and the handling requirement is that suggestions be adopted as appropriate and improvements encouraged.

[0058] In other embodiments, accurately pinpointing the location of the problem includes: using text anchoring technology to accurately extract the chapter number, paragraph number, and specific sentence fragment corresponding to each problem from the report tree structure as location information; and simultaneously calling the knowledge graph query interface to retrieve the original text of the legal and standard clauses, knowledge graph node attributes, and historical similar case records that match the problem entity.

[0059] In one specific embodiment, the generation of the hierarchical modification suggestions includes: for each problem instance, retrieving the corresponding solution strategy from the domain expert knowledge base, and combining the following to generate a structured modification guidance text: principle suggestions clarifying the direction of rectification, methodological suggestions providing specific technical parameters or operating procedures, and reference suggestions listing similar successful cases.

[0060] Specifically, the review comments support diversified outputs, including: a structured review report containing an overall evaluation, a list of issues, a summary of modification suggestions, and review conclusions; a visual display interface containing risk heatmaps, knowledge graph association diagrams, and statistical dashboards; and an interactive feedback function that includes online annotation, confirmation of corrections, supplementary questions, and communication and discussion.

[0061] It should be noted that this embodiment optimizes the allocation of review resources by defining and prioritizing review issues at a fine-grained hierarchical level; it significantly improves the interpretability and credibility of review conclusions by accurately locating issues down to the sentence level and automatically associating multiple sources of evidence; and it effectively reduces the difficulty of revisions for report preparers by presenting modification suggestions in a layered manner.

[0062] Based on the above embodiments, in some embodiments, the closed-loop optimization and update of the domain knowledge graph and semantic understanding model based on review feedback data includes: Establish an early warning result verification and annotation system to receive actual situation annotation data fed back by on-site personnel through mobile terminals. The annotation data includes confirmed problems, false alarms, and normal status. Newly labeled data is regularly added to the training set to learn and update the node attributes of the domain knowledge graph and the model parameters of the semantic understanding model online. Establish a false alarm analysis module to statistically analyze false alarm cases to identify model weaknesses that fail to recognize specific technical terms or special expression patterns, and collect targeted data for this scenario to fine-tune and optimize the model.

[0063] In some embodiments, the online learning update employs a mini-batch gradient descent algorithm to fine-tune the model parameters, focusing on adjusting the attention weights for specialized terminology in the domain-specific pre-trained language model. The model parameter update formula is as follows: ,in Let represent the model parameter vector at the t-th update, and η represent the learning rate (used to control the step size of parameter updates). The loss function L represents the parameters. In the current training mini-batch data gradient on, This represents the updated model parameters.

[0064] In other embodiments, the false alarm analysis module statistically analyzes frequently occurring variations of specific technical terms or unique expression patterns in false alarm cases to identify weaknesses in the model's semantic understanding. To address the data imbalance problem, Focal Loss and SMOTE oversampling techniques are employed during model training. The Focal Loss function is defined as: ,in This represents the probability that the model predicts the true class. γ represents the category balance weight factor (used to adjust the weight ratio of positive and negative samples), and γ represents the focusing parameter (used to control the degree of weight adjustment for easy and difficult samples). is the modulation factor.

[0065] In one specific embodiment, in the early warning result verification and annotation system, a confirmed problem refers to a report that does indeed have defects and needs rectification; a false alarm refers to a system error in judgment but a report that is actually compliant; and a normal state refers to a system warning of risk but which, after verification, is within the safety threshold. Confirmed problem samples are added as positive samples and false alarm samples as negative samples to the original training set. For the knowledge graph, the state attributes and relationship attributes of nodes are updated based on new facts in the confirmed problems.

[0066] Specifically, the closed-loop optimization update is performed periodically (e.g., weekly), and samples with high model uncertainty are prioritized for labeling through an active learning strategy. The model uncertainty is quantified using prediction probability entropy. ,in Entropy represents the probability that the model will predict the sample as class c. The higher the entropy value, the greater the uncertainty of the model's prediction for that sample.

[0067] It should be noted that this embodiment effectively utilizes on-site measured feedback data through a closed-loop optimization mechanism to continuously correct system deviations, significantly reducing the false alarm rate caused by diverse professional terminology or rich implicit semantics, enabling the intelligent review system to have adaptive evolution capabilities. As the usage time increases, the review accuracy and scenario adaptability are continuously improved.

[0068] Please refer to Figure 2 , Figure 2 A structural block diagram of an intelligent review device for hidden disaster reports in coal mines provided in this embodiment of the invention; the specific device may include: The knowledge graph construction module 100 is used to construct a domain knowledge graph based on multi-source heterogeneous data, which includes hidden disaster-causing factors, causative conditions, prevention and control measures and legal standards, and is configured with a hybrid reasoning engine that supports rule-based reasoning and similarity-based reasoning. The semantic understanding and extraction module 200 is used to perform multi-level preprocessing and domain semantic understanding on the text of the coal mine hidden disaster-causing factor survey report to be reviewed, and to extract key entities, risk factors, assessment conclusions and numerical parameters in the report to form structured information. The comparison and reasoning verification module 300 is used to perform multi-dimensional comparison between the structured information and the domain knowledge graph, use a probabilistic graphical model to reason and verify the risk level judgment in the report, and check the matching degree between the prevention and control measures and risk factors and the compliance with regulations. The opinion generation and optimization module 400 is used to generate review opinions that include problem location, supporting explanations and modification suggestions based on the comparison results and reasoning verification results, and to perform closed-loop optimization and update of the domain knowledge graph and semantic understanding model based on the review feedback data.

[0069] In some embodiments, the knowledge graph construction module 100 includes: a multi-source knowledge acquisition unit for acquiring legal and standard knowledge, domain expert knowledge, and historical case knowledge; an ontology modeling unit for defining core concept classes and constructing a multi-relationship network; a knowledge extraction unit for entity recognition using a sequence labeling model based on a pre-trained language model and relation extraction using an attention mechanism neural network model; and a knowledge fusion unit for achieving knowledge fusion and conflict resolution through multi-model voting decision-making, entity alignment, and relation resolution.

[0070] In other embodiments, the semantic understanding and extraction module 200 includes: a format standardization unit for uniformly processing reports to be reviewed in various formats; a word segmentation and tagging unit for performing accurate word segmentation and part-of-speech tagging based on a domain dictionary; a structure restoration unit for restoring the report chapter hierarchy and establishing a tree-like paragraph system; a semantic encoding unit for encoding the text into semantic vectors using a fine-tuned semantic understanding model; and an information extraction unit for identifying key entities, risk factor sets, and assessment conclusion sets from the semantic vectors.

[0071] In one specific embodiment, the comparison and reasoning verification module 300 includes: an element completeness review unit, used to check the completeness of chapters and the coverage of content against the standard; a risk factor comparison unit, used to prompt missing factors by combining semantic similarity calculation with knowledge graph matching; a Bayesian reasoning unit, used to construct a probability graph network structure and calculate the posterior probability distribution of risk level; and a measure compliance review unit, used to check the matching degree between prevention and control measures and risk factors and regulatory compliance.

[0072] Specifically, the opinion generation and optimization module 400 includes: an issue classification unit, used to classify and prioritize identified issues as serious issues, general issues, and issues requiring optimization; an opinion generation unit, used to accurately pinpoint the location of the issue and associate it with legal provisions, knowledge graph nodes, and historical cases as supporting evidence; a suggestion generation unit, used to present principled suggestions, methodological suggestions, and reference suggestions in a hierarchical manner; and a closed-loop optimization unit, used to perform online learning and updates of the knowledge graph and semantic understanding model based on review feedback data.

[0073] The intelligent review device for coal mine hidden disaster reports in this embodiment is used to implement the aforementioned intelligent review method for coal mine hidden disaster reports. Therefore, the specific implementation of the intelligent review device for coal mine hidden disaster reports can be found in the previous embodiment section of the intelligent review method for coal mine hidden disaster reports. For example, the knowledge graph construction module 100, the semantic understanding and extraction module 200, the comparison and reasoning verification module 300, and the opinion generation and optimization module 400 are respectively used to implement steps S101, S102, S103, and S104 in the aforementioned intelligent review method for coal mine hidden disaster reports. Therefore, its specific implementation can be referred to the description of the corresponding embodiments, which will not be repeated here.

[0074] This embodiment solidifies the method steps into functional modules, constructing a complete intelligent review device. It realizes intelligent processing of the entire process of coal mine hidden disaster-causing factor survey report from receiving, parsing, reviewing to feedback, significantly improving review efficiency and accuracy.

[0075] A specific embodiment of the present invention also provides an intelligent review device for coal mine hidden disaster reports, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the above-described intelligent review method for coal mine hidden disaster reports.

[0076] A specific embodiment of the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described intelligent review method for reporting hidden disasters in coal mines.

[0077] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0078] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0079] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0080] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0081] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for intelligent review of hidden disaster reports in coal mines, characterized in that, include: A domain knowledge graph containing hidden disaster-causing factors, causal conditions, prevention and control measures, and regulatory standards is constructed based on multi-source heterogeneous data, and a hybrid reasoning engine supporting rule-based reasoning and similarity-based reasoning is configured. Multi-level preprocessing and domain semantic understanding are performed on the text of the coal mine hidden disaster-causing factor survey report to be reviewed, and key entities, risk factors, assessment conclusions and numerical parameters in the report are extracted to form structured information; The structured information is compared with the domain knowledge graph in multiple dimensions. The risk level judgment in the report is inferred and verified using a probabilistic graphical model. The matching degree between the prevention and control measures and risk factors and the compliance with regulations are also checked. Based on the comparison and reasoning verification results, review opinions are generated, including problem location, supporting explanations, and modification suggestions. The domain knowledge graph and semantic understanding model are then optimized and updated in a closed loop based on the review feedback data.

2. The method according to claim 1, characterized in that, The domain knowledge graph constructed based on multi-source heterogeneous data, which includes hidden disaster-causing factors, causative conditions, prevention and control measures, and regulatory standards, includes: Collect legal and standard knowledge, domain expert knowledge, and historical case knowledge, and extract mandatory requirements, review points, and case tags to form a multi-source knowledge base; Define the core concept classes of hidden disaster-causing factors, causative conditions, disaster-prone environment, disaster-bearing bodies, prevention and control measures, risk assessment, regulations and standards, and review points, and construct a multi-dimensional relationship network connecting the above concept classes; A sequence labeling model based on a pre-trained language model is used to perform entity recognition on the multi-source knowledge base, an attention mechanism neural network model is used to extract relations, and a domain knowledge graph containing node attributes and relation attributes is generated through knowledge fusion and conflict resolution.

3. The method according to claim 1, characterized in that, The multi-level preprocessing and domain semantic understanding of the coal mine hidden disaster-causing factor survey report text to be reviewed includes: It unifies the processing of reports awaiting review in various formats, and uses parsing and recognition technologies to extract plain text while retaining chapter structure information; Accurate word segmentation and part-of-speech tagging are performed based on a domain dictionary, and words are filtered using a domain-specific stop word list; Restore the report's chapter hierarchy and establish a tree-like paragraph system that reflects the report's logical structure; The text in the tree-like paragraph system is encoded into semantic vectors using a fine-tuned semantic understanding model to support vector similarity-based analysis. Key entities are identified from the semantic vectors, factor types and governance measures are extracted to form a risk factor set, and risk level determination and assessment method descriptions are extracted to form an assessment conclusion set.

4. The method according to claim 1, characterized in that, The step of performing a multi-dimensional comparison between the structured information and the domain knowledge graph includes: Verify the completeness of the chapters and the coverage of the content in the report to be reviewed against the standards; By combining semantic similarity calculation with knowledge graph matching, the extracted set of risk factors is compared with the hidden disaster-causing factor entities in the domain knowledge graph, indicating missing factors and marking non-standard expressions. Verify the sufficiency of the basic data supporting the analysis conclusions and generate the results of the element integrity review.

5. The method according to claim 1, characterized in that, The reasoning and verification of the risk level determination in the report using a probabilistic graphical model includes: Construct a probabilistic graph network structure that includes geological condition nodes, engineering factor nodes, and risk level nodes, and input the reported observation evidence as evidence nodes into the probabilistic graph network; Calculate the posterior probability distribution of the risk level nodes, and compare the calculated posterior probability with the risk level determined in the report; When the probability deviation exceeds a preset threshold, a review prompt is triggered, and a risk assessment accuracy verification report containing the comprehensive reasoning results of multi-source evidence is output.

6. The method according to claim 1, characterized in that, The verification and prevention measures' compatibility with risk factors and regulatory compliance include: By querying the applicable relationships in the knowledge graph, we can verify the matching relationship between the types of governance measures and risk factors, as well as the commensurability between the strength of the measures and the risk level. The rules engine is used to automatically check whether the content of the report to be reviewed complies with mandatory requirements and with prohibitions. Based on historical implementation effects analyzed through case library data mining, the effectiveness of governance measures is assessed, and the results of compliance and effectiveness review of the proposed measures are generated.

7. The method according to claim 1, characterized in that, The review comments generated based on the comparison results and reasoning verification results, including problem location, supporting explanations, and modification suggestions, include: Define the critical issues that affect the core conclusions of the report, the general issues that affect the quality of the report, and the optimization issues that are improvement suggestions, and classify and prioritize the identified issues; It accurately pinpoints the location of the problem and automatically links it to relevant legal and standard provisions, knowledge graph nodes, and historical case references as supporting evidence. The recommendations are presented in layers: principle-based suggestions, methodological suggestions, and reference suggestions, forming a complete description of modification suggestions for each problem.

8. An intelligent review device for reports of hidden disasters in coal mines, characterized in that, include: The knowledge graph construction module is used to build a domain knowledge graph based on multi-source heterogeneous data, which includes hidden disaster-causing factors, causative conditions, prevention and control measures and legal standards, and is configured with a hybrid reasoning engine that supports rule-based reasoning and similarity-based reasoning. The semantic understanding and extraction module is used to perform multi-level preprocessing and domain semantic understanding on the text of the coal mine hidden disaster-causing factor survey report to be reviewed, and to extract key entities, risk factors, assessment conclusions and numerical parameters in the report to form structured information. The comparison and reasoning verification module is used to perform multi-dimensional comparison between the structured information and the domain knowledge graph, use a probabilistic graphical model to reason and verify the risk level judgment in the report, and check the matching degree between the prevention and control measures and risk factors and the compliance with regulations. The feedback generation and optimization module is used to generate review feedback that includes problem location, supporting explanations, and modification suggestions based on the comparison results and reasoning verification results. It also performs closed-loop optimization and updates to the domain knowledge graph and semantic understanding model based on the review feedback data.

9. An intelligent review device for hidden disaster reports in coal mines, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the intelligent review method for reporting hidden disasters in coal mines as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the intelligent review method for reporting hidden disasters in coal mines as described in any one of claims 1 to 7.

Images