A method and system for intelligent hazard assessment based on the fusion of multi-source perception and safety specification knowledge base

By integrating multi-source perception with a safety standard knowledge base, intelligent identification of safety hazards is achieved, solving the problems of fragmented data sources and uninterpretable judgments, improving the accuracy of judgments and the reliability of the system, and supporting expansion into multiple industries.

CN123115733APending Publication Date: 2026-11-17黄宇杰
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611444239.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-17
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing safety inspection systems suffer from problems such as fragmented data sources, disconnect between identification results and regulatory provisions, lack of scenario constraints in judgment, untraceable results, separation of closed-loop and model learning, and limited applicability, leading to inaccurate and unexplainable hazard identification.

Method used

By collecting multi-source heterogeneous data, performing spatiotemporal correlation and preprocessing, extracting multi-dimensional candidate hazard features, combining them with a safety standard knowledge base for fusion judgment, generating a traceable evidence chain, and updating and optimizing the model through manual review and learning, intelligent hazard judgment is achieved.

Benefits of technology

It improves the accuracy and interpretability of hazard identification, reduces the false alarm rate, enhances the system's reliability and cross-scenario applicability, and supports expansion into multiple industries.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115733A_ABST
    Figure CN123115733A_ABST
Patent Text Reader

Abstract

This invention discloses an intelligent hazard assessment method and system based on the fusion of multi-source perception and a safety specification knowledge base, belonging to the field of digital management technology for safe production. Addressing the problems of fragmented data sources, disconnect between identification and specifications, lack of scenario constraints in assessment, untraceable results, and separation of closed-loop learning in existing technologies, this invention constructs a complete technical chain from multi-source acquisition to controlled updates: multi-source heterogeneous safety data is spatiotemporally aggregated and structured scenario vectors are constructed; safety specifications are parsed into structured clause units containing multiple types of constraints; candidate rules are searched and matched in a two-stage process, and constraints are verified item by item. The fusion confidence score is calculated and output in a tiered manner by combining perception confidence, rule weight, data quality, and multi-source consistency; a traceable evidence chain and rectification tasks are generated, and the rules and models are iteratively updated using a controlled sedimentation mechanism. This invention significantly improves the accuracy and interpretability of hazard assessment, reduces false alarms and missed alarms, and supports cross-industry scenario expansion.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of digital management of safe production, specifically to a method and system for intelligent hazard judgment based on the integration of multi-source perception and safety standard knowledge base. Background Technology

[0002] Current security inspections primarily employ methods such as manual inspections, paper or mobile checklists, video surveillance, single-type sensor alarms, and image recognition based on general visual models. While manual inspections can rely on experience to make judgments, they suffer from limited coverage, inconsistent standard implementation, time-consuming evidence processing, and difficulties in closed-loop tracking. Video surveillance and sensors such as smoke detectors, gas detectors, temperature sensors, and vibration sensors can continuously collect data, but most systems only trigger alarms based on preset thresholds or single events.

[0003] Some intelligent safety systems can identify targets or behaviors such as not wearing safety helmets, personnel crossing boundaries, smoke, and vehicles approaching. Other systems associate the identification results with fixed rules. However, safety hazards are usually not determined by a single target, but rather by a combination of factors including the type of work, personnel identity, equipment status, area attributes, environmental parameters, time conditions, approval status, and exceptions to regulations. For example, the risk of the same person at a height differs depending on whether adequate platform protection is in place or not; the same temperature value also corresponds to different handling thresholds in a general warehouse and a flammable materials area.

[0004] Safety regulations typically exist as unstructured texts such as laws and regulations, national standards, industry standards, corporate policies, job operation procedures, and temporary control requirements. These regulations contain variations in version, scope of application, preconditions, thresholds, exceptions, and hierarchical conflicts. Relying solely on large language models to generate judgments can easily lead to unclear basis, distorted citations, or non-repeatable conclusions; relying solely on fixed rules makes it difficult to handle image semantics, complex scenarios, and updated regulations.

[0005] Specifically, the existing technology has the following core flaws: First, the data sources are fragmented. Images, videos, sensor data, equipment operating parameters, work permits, personnel qualifications, and location data are stored separately, making it difficult to form complete evidence from the same time, space, and work object.

[0006] Second, the identification results are disconnected from the regulatory provisions. Existing visual recognition systems mostly output "target detected" or "target not detected," but cannot reliably explain under what applicable conditions what kind of hazard it constitutes, which regulation it corresponds to, or why similar provisions are excluded.

[0007] Third, the judgment lacks scenario constraints. Relying solely on a single image or a single threshold is prone to false alarms, such as identifying a non-operational state as an illegal operation, or ignoring key conditions such as work permits, personnel qualifications, equipment interlocks, and area classification.

[0008] Fourth, the results are not traceable. The model input, specification version, rule parameters, evidence fragments, confidence levels, and manual review records are not linked, making it difficult to reproduce the judgment process when a dispute arises.

[0009] Fifth, the closed-loop process is separated from model learning. Hazard discovery, task assignment, rectification, review, and knowledge updating are all independent processes, and manual correction cannot optimize rule weights, scene templates, and recognition models in reverse.

[0010] Sixth, the scope of application is limited. The system is often pre-defined for a certain industry or a certain type of hidden danger, and it is difficult to expand to new industries and enterprise systems by adding standards, scenario templates and sensor access methods.

[0011] In view of this, in-depth research was conducted on the above issues, which led to the creation of this case. Summary of the Invention

[0012] The purpose of this invention is to provide a method and system for intelligent hazard judgment based on the fusion of multi-source perception and safety specification knowledge base, so as to solve the problems mentioned in the background art, such as fragmented data sources, disconnect between identification results and specification clauses, lack of scenario constraints in judgment, untraceable results, separation of closed loop and model learning, and limited scope of application.

[0013] To achieve the above objectives, the present invention provides the following technical solution: On the one hand, this invention provides an intelligent hazard assessment method based on the fusion of multi-source perception and a safety specification knowledge base, comprising the following steps: Step S1. Collect multi-source safety data: Collect multi-source heterogeneous on-site safety data. Each data point includes a data source identifier, collection time, spatial location, object identifier, and data quality parameters. Step S2. Perform data preprocessing and scene modeling: Perform clock calibration, spatial coordinate transformation and event aggregation on multi-source security data, and construct a structured scene vector based on the related data within the same time window and spatial neighborhood. The scene vector includes object attributes, behavioral attributes, environmental attributes and management attributes. Step S3. Extract candidate hazard features: Extract multi-dimensional candidate hazard features from the aggregated security event data using artificial intelligence models and deterministic algorithms; Step S4. Retrieve safety specifications and form a candidate rule set: Based on the scenario vector and the characteristics of the candidate hidden dangers, retrieve and match from the pre-built safety specification knowledge base to obtain a candidate rule set; the safety specification knowledge base is composed of unstructured safety specifications parsed into structured specification clause units; Step S5. Perform fusion judgment and risk classification: Perform constraint condition verification on each normative clause unit in the candidate rule set, calculate the fusion confidence based on the perception model confidence, rule reliability, data quality and multi-source consistency, classify the risk based on the fusion confidence and risk assessment results, and output the corresponding hidden danger judgment results; Step S6. Generate a traceable chain of evidence and rectification tasks: Associate and bind the hazard assessment results, triggering conditions, standard basis, original evidence, model and standard version information to generate a complete chain of evidence, and match the responsible party to generate rectification tasks; Step S7. Manual review, rectification closed loop and controlled learning: Receive feedback results of manual review and rectification tasks and store them in the sample pool. After meeting the preset sample quantity, permission approval and verification performance conditions, update the scene template, rule weight or model parameters, and retain the version before and after the update and rollback information.

[0014] Furthermore, in step S2, the construction of the scene vector specifically includes: Based on a preset time window and spatial neighborhood range, multi-source data of the same personnel, equipment, area or task are aggregated into a security event data packet; Extract attributes from the security event data packet, including region type, operation type, personnel role, equipment type, environment status, license status, time status, and organizational policy version, construct a scenario vector, and retain the data source and credibility mark of each attribute field.

[0015] Furthermore, in step S3, the extraction of candidate hazard features also includes multi-source feature fusion and conflict resolution: For cases where the same hazard feature is repeatedly observed from multiple data sources, feature fusion is performed based on data source reliability, temporal proximity, and spatial consistency. When feature recognition results from different data sources contradict each other, a conflict marker is generated and the multi-source consistency coefficient of the corresponding feature is reduced, without directly overwriting the recognition results of any data source.

[0016] Further, in step S4, the structure of the standard clause unit is R={id,v,t,a,o,j,p,n,f,e,h,m}, where: id is the standard and clause identifier, v is the version number, t is the effective time range, a is the applicable region or industry, o is the applicable object, j is the operation type, p is the set of necessary conditions, n is the set of prohibited conditions, f is the threshold function, e is the set of exception conditions, h is the mapping of hazard category and level, and m is the set of rectification measures; The retrieval and matching adopts a two-stage retrieval: the first stage recalls relevant regulatory clauses through semantic vectors or keywords, and the second stage filters fields according to version validity, applicable objects, applicable regions, job types and enterprise levels to obtain a candidate rule set.

[0017] Further, in step S5, the constraint verification specifically includes: For each specification clause unit in the candidate rule set, the necessary conditions, prohibited conditions, threshold conditions, and exception conditions are checked in sequence. When all necessary conditions are met, prohibited conditions are triggered, and exception conditions are not met, a potential hazard is formed. Threshold conditions are judged based on the collected values, unit conversion, duration, and measurement uncertainty. The formula for calculating the fusion confidence level C is: C = Pm × Wr × Dq × Sc, where: Pm represents the confidence level of the perception model, which comes from the confidence level output by the artificial intelligence model when identifying or classifying the features of candidate hazards. For example, the probability value output by the model is the target detection, image classification, action recognition, time-series anomaly detection, or visual language model, and can be normalized to 0-1. Wr represents the reliability weight of the clause rules, derived from the historical accuracy and hierarchical weight of the normative clauses. The hierarchical weight is determined based on the normative levels such as laws, administrative regulations, departmental rules, national standards, industry standards, and corporate systems, and can optionally be adjusted by combining the clause's effective status, manual review conclusions, and historical review results. Dq is the data quality coefficient, which comes from data quality parameters. These data quality parameters include one or more of signal strength, measurement accuracy, historical reliability, data integrity, and timeliness, and are normalized to 0-1 according to a preset mapping function. Sc is the multi-source consistency coefficient, which comes from the multi-source data consistency verification results. The multi-source data consistency verification results include one or more of the following: the degree of consistency of observations of the same hidden danger feature or fact by multiple data sources, conflict markers, temporal proximity, and spatial consistency. When the identification results of different data sources are contradictory, the coefficient is reduced, and the value ranges from 0 to 1. The above formula is used to solve the specific technical problem of how to quantify reliability when judging multi-source heterogeneous data fusion, rather than abstract mathematical calculation; by uniformly mapping the uncertainty of the perception model, the reliability of the normative clauses, data quality and multi-source consistency to the fusion confidence, the result of the hidden danger judgment is calculable, interpretable and traceable. The formula for calculating the risk value Q is: Q = C × S × E × K, where S is the severity of the consequences, E is the exposure level, and K is the scenario correction factor; The system sets three-level decision intervals based on the fusion confidence level: when the automatic confirmation threshold is reached, the system outputs a decision; when the threshold is within the verification range, a decision awaiting manual verification is output; and when the threshold is below the suppression threshold, only candidate records are retained. The three-level decision intervals are a specific technical application of the fusion confidence level formula, achieving a hierarchical processing effect of automatic confirmation with high confidence, manual verification with medium confidence, and suppression with low confidence, thereby reducing false alarms and missed alarms and minimizing unnecessary manual verification.

[0018] Furthermore, in step S6, the traceable evidence chain specifically includes: hazard type, risk level, occurrence time and location, involved objects, triggering conditions, unmet safety requirements, standard name and specific clause number, evidence image or data fragment, identification model version, standard library version, fusion confidence level and suggested rectification measures; All evidence documents and rectification tasks are associated and bound with a unique task identifier, and a summary value is calculated for each evidence document to prevent mismatch.

[0019] Furthermore, in step S7, the controlled update specifically includes: The results of manual review and rectification acceptance are first entered into the feedback sample pool and do not directly cover the modification of formal rules or online models. When the sample pool meets the preset sample quantity requirement, is approved by the designated permission role, and reaches the preset performance index on the validation set, the scene template, rule weight, or recognition model parameters are officially updated. Each update retains the complete version and rollback path before and after the update, and supports version switching and historical judgment reproduction.

[0020] On the other hand, the present invention also provides a hidden danger intelligent judgment system based on the fusion of multi-source perception and safety standard knowledge base, including: The data access module is used to collect and access multi-source heterogeneous field safety data, assign a unified data identifier to each data point and attach collection time, space, object and quality attributes; The data governance module is used to complete the temporal calibration, spatial correlation, format conversion, quality assessment, event aggregation and scene label generation of multi-source data, and output structured scene vectors; The feature recognition module is used to extract multi-dimensional candidate hazard features from security incident data through artificial intelligence models and deterministic algorithms; A scenario model is used to store the relationships between regions, tasks, objects, organizational systems, and dynamic states, and to define the scope of application of rules. The security specification knowledge base is used to store structured specification clause units, specification original text location information, version relationships, hierarchical relationships, and conflict rules; The fusion judgment module is used to retrieve and match candidate rule sets based on scene vectors and candidate hazard features, perform constraint condition verification item by item, calculate fusion confidence and complete risk classification, and output hazard judgment results. The closed-loop management module is used to generate traceable evidence chains, rectification tasks, notifications and review records to complete the entire process of hazard rectification. The learning and updating module is used to manage the review feedback sample pool, perform model evaluation, rule revision, version approval and release, and version rollback, so as to achieve controlled rule and model iteration.

[0021] Furthermore, the fusion judgment module has a built-in rule verification engine and a confidence calculation unit; The rule verification engine is used to perform necessary condition satisfaction verification, prohibited condition triggering verification, threshold condition matching verification and exception condition exclusion verification for each candidate specification clause unit in sequence. The confidence calculation unit is used to calculate the fusion confidence based on the confidence of the perception model, the reliability weight of the rule, the data quality coefficient and the multi-source consistency coefficient, and output the corresponding judgment conclusion according to the preset three-level threshold.

[0022] Furthermore, the security specification knowledge base supports extended access to specifications across multiple industries and at multiple levels. By adding data adapters, scenario templates, and specification clause units, it can adapt to security supervision scenarios in different industries. The learning and update module has a built-in version management unit that uniformly freezes and releases the specification library version, model version, and rule version, and supports historical version tracing and judgment process reproduction.

[0023] Compared with existing technologies, the beneficial effects of this intelligent hazard assessment method and system based on the fusion of multi-source perception and safety specification knowledge base include: 1. Improve the accuracy of hazard identification. Utilize spatiotemporal correlation, multi-source consistency, and exception condition verification to reduce false alarms and missed alarms caused by relying solely on images or single thresholds.

[0024] 2. Improve the interpretability of judgments. Each conclusion corresponds to specific facts, triggering conditions, valid regulatory provisions, and original evidence, facilitating review and accountability.

[0025] 3. Improve the efficiency of standard updates. Standards are managed in versioned clause units, so there is no need to redevelop all recognition functions when updating clauses.

[0026] 4. Improve the efficiency of hazard closure. Automatically generate responsibilities, deadlines, measures, and review conditions, reducing repetitive data entry in inspection record compilation, dispatching, and review.

[0027] 5. Enhance system reliability. By using conflict markers, data quality coefficients, confidence partitioning, manual review, and rollbackable updates, uncertain results are prevented from being directly treated as definitive conclusions.

[0028] 6. Capable of cross-scenario reuse. The system can be configured to adapt to scenarios such as mining, construction, manufacturing, energy, transportation, warehousing and logistics, and camp management. Attached Figure Description

[0029] Figure 1 This is a flowchart of the intelligent safety hazard determination method in an embodiment of the present invention; Figure 2 This is a schematic diagram of the intelligent safety hazard judgment system in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the structure and judgment logic of the specification clauses in an embodiment of the present invention; Figure 4 This is a flowchart illustrating the integration of confidence calculation and three-level judgment in an embodiment of the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only 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.

[0031] Please see Figures 1-4 The present invention provides the following technical solutions: I. Method Implementation Examples like Figure 1 As shown, this invention provides an intelligent hazard assessment method based on the fusion of multi-source perception and a safety specification knowledge base, comprising the following steps: Step S1: Collect multi-source security data.

[0032] The system acquires images, video frames, audio, sensor values, equipment status, personnel identity and qualifications, geographical location, time, work permits, inspection records, and corporate policy data through one or more of the following: mobile terminals, fixed cameras, drones, smoke detectors, gas detectors, temperature sensors, vibration sensors, vehicle or equipment controllers, positioning devices, and business system interfaces.

[0033] Each data entry includes a data source identifier, acquisition time, spatial location, object identifier, data quality parameters, and integrity check value. For offline scenarios, a local task identifier is generated and cached on the terminal, and then uploaded after deduplication based on the task identifier and timestamp upon network connection. Data quality parameters may include signal strength, measurement accuracy, and historical reliability.

[0034] Step S2: Perform data preprocessing and scenario modeling.

[0035] The system performs unified clock calibration, spatial coordinate transformation, noise reduction, missing value marking, and outlier identification on different data sources. Based on preset time windows and spatial neighborhoods, data from the same personnel, equipment, area, or task are aggregated into a safety event data package.

[0036] A scenario vector X is constructed based on region type, job type, personnel role, equipment type, environmental status, permit status, time status, and organizational policy version. The scenario vector includes at least object attributes, behavioral attributes, environmental attributes, and management attributes, and retains the source and reliability of the fields.

[0037] Specifically, based on a preset time window and spatial neighborhood range, multi-source data of the same personnel, equipment, area or task are aggregated into a security event data packet; attributes such as area type, task type, personnel role, equipment type, environmental status, permit status, time status and organizational system version are extracted from the security event data packet to construct a scenario vector, and the data source and credibility mark of each attribute field are retained.

[0038] For example, in a mining vehicle maintenance area, the scenario vector may include: area type "tire repair station", operation type "tire inflation", personnel role "maintenance worker", equipment type "large vehicle tire", environmental status "normal temperature", permit status "approved", time status "in operation", and organizational system version "Safety Operation Procedures of a Certain Enterprise 2026 Edition".

[0039] Step S3: Extract candidate hazard features.

[0040] By utilizing object detection, image segmentation, pose recognition, action recognition, OCR, temporal anomaly detection, voice event recognition, or threshold detection models, candidate features such as personal protective equipment, dangerous behavior, equipment status, area protection, material stacking, smoke and fire, gas concentration, temperature, speed, distance, and ticket fields are extracted from security event data packets.

[0041] For the same feature repeatedly observed from multiple data sources, fusion is performed according to data source reliability, temporal proximity, and spatial consistency. When different data sources contradict each other, conflict markers are generated instead of direct overwriting, and the multi-source consistency coefficient Sc of the corresponding feature is reduced.

[0042] For example, the image identifies "personnel located on the side of the tire," the pressure sensor identifies "the tire is inflated," the area label indicates "tire repair station," and the personnel location shows that the worker is within the possible trajectory of the tire bursting. If the image and location data do not agree on the personnel's location, a conflict marker is generated, and the multi-source consistency coefficient is reduced in the subsequent fusion determination.

[0043] Step S4: Retrieve security specifications and form a candidate rule set.

[0044] Laws, regulations, standards, procedures, and corporate policies are analyzed into normative clause units R. Each normative clause unit includes: normative identifier, version, effective date, expiration date, applicable industry, applicable region, applicable object, operation type, necessary conditions, prohibited conditions, threshold conditions, exception conditions, hazard category, risk level recommendation, rectification measures, and original text location.

[0045] The structure of the standard clause unit is R={id,v,t,a,o,j,p,n,f,e,h,m}, where: id is the standard and clause identifier, v is the version number, t is the effective time range, a is the applicable region or industry, o is the applicable object, j is the operation type, p is the set of necessary conditions, n is the set of prohibited conditions, f is the threshold function, e is the set of exception conditions, h is the mapping of hazard category and level, and m is the set of rectification measures.

[0046] A two-stage retrieval process is performed based on scenario vectors and candidate hazard features: the first stage recalls relevant regulatory clauses through semantic vectors or keywords; the second stage filters fields according to version validity, applicable objects, applicable regions, operation types, and enterprise levels to obtain a candidate rule set.

[0047] When multiple clauses make different provisions on the same matter, the system determines the priority based on legal level, professional nature, publication time, stricter requirements of the enterprise, and explicit scope of application; if the conflicting clauses cannot be automatically resolved, they will be presented to the reviewers simultaneously. The machine parsing results of the standard text must retain the original page numbers, paragraph numbers, or clause numbers for manual verification.

[0048] Step S5: Perform fusion determination and risk classification.

[0049] Constraint satisfaction is verified for each normative clause unit in the candidate rule set. A potential hazard is formed when all necessary conditions are met, prohibited conditions are triggered, and exceptional conditions are not met. Threshold clauses are judged based on collected values, unit conversions, duration, and measurement uncertainty.

[0050] The formula for calculating the fusion confidence level C is: C = Pm × Wr × Dq × Sc, where: Pm represents the confidence level of the perception model, which comes from the output of the AI ​​model, such as the probability value output by the visual language model for the candidate hidden danger features in object detection, image classification, action recognition, temporal anomaly detection, or visual language model. It can be normalized to 0-1. Wr represents the reliability weight of the clause / rule, derived from the historical accuracy and hierarchical weight of the regulatory clause. For example, the basic weight is set according to the hierarchy of laws, administrative regulations, departmental rules, national standards, industry standards, and corporate systems, and then adjusted based on the historical accuracy of the clause and the pass rate of manual review. Dq is the data quality coefficient, which comes from data quality parameters, including signal strength, measurement accuracy, historical reliability, integrity and timeliness, and is normalized to 0-1 through a preset mapping function; Sc is the multi-source consistency coefficient, which comes from the consistency verification results of multi-source data, such as images, sensors, positioning, tickets, etc. It is the degree of consistency of observations of the same fact by multi-source data. It decreases when there are conflict markers, time deviations, or spatial deviations, and takes a value of 0-1. The above formula addresses the difficulty in quantifying reliability when determining the fusion of multi-source heterogeneous data. It unifies model confidence, rule reliability, data quality, and multi-source consistency into a calculable fusion confidence, rather than abstract mathematical calculation. The formula for calculating the risk value Q is: Q = C × S × E × K, where S is the severity of the consequences, E is the exposure level, and K is the scenario correction factor; A three-tiered decision interval is set based on the fusion confidence level: when the automatic confirmation threshold is reached, the system outputs a decision; when the confidence level is within the review interval, a decision awaiting manual review is output; and when the confidence level is below the suppression threshold, only candidate records are retained. This three-tiered decision interval is a specific technical application of the above formula, achieving a hierarchical processing effect of automatic confirmation for high-confidence events, manual review for medium-confidence events, and suppression for low-confidence events. For high-consequence events, even if the confidence level does not reach the automatic confirmation threshold, an alert can be triggered and a review request can be made. The specific functions, weights, and hierarchical thresholds can be configured by the industry or enterprise.

[0051] For example, in a tire inflation scenario, if the image location confidence level Pm is 0.92, the rule weight Wr is 0.95, the data quality coefficient Dq is 0.90, and the multi-source consistency coefficient Sc is 0.88, then the fusion confidence level C = 0.92 × 0.95 × 0.90 × 0.88 ≈ 0.692. If the automatic confirmation threshold is 0.85 and the review interval is 0.60 to 0.85, then the system will output "Pending manual review". Simultaneously, since tire inflation and bursting is a high-consequence event, the system will directly trigger an alert and require review.

[0052] Step S6: Generate a traceable chain of evidence and rectification tasks.

[0053] Output the hazard type, risk level, time and location of occurrence, affected objects, triggering conditions, unmet safety requirements, standard name and specific clauses, evidence images or data fragments, model version, standard version, confidence level, and recommended measures.

[0054] The responsible parties for rectification are determined based on a mapping table of organizational, regional, equipment, or job responsibilities, generating rectification tasks, deadlines, review requirements, and notification information. All evidence documents are associated with and bound to rectification tasks through a unique task identifier, and a summary value is calculated for each evidence document to prevent mismatches.

[0055] For example, the system generates a chain of evidence including: a site annotation diagram (relative positions of personnel and tires), pressure curve segments, area labels, personnel positioning trajectories, the original text and clause number of the specification clauses, model version number, specification library version number, a fusion confidence level of 0.692, and recommended measures such as "stop inflation, evacuate personnel from the hazardous trajectory, use restraint devices, and remote inflation." Simultaneously, a rectification task is generated, with the maintenance team as the responsible party, a deadline of 2 hours, and a review requirement of "re-identifying personnel positions and the status of protective devices."

[0056] Step S7: Manual review, rectification loop, and controlled learning.

[0057] Reviewers can confirm, reject, reclassify, adjust the level, supplement evidence, or require re-collection. After the person responsible for rectification uploads the rectification evidence, the system will re-collect or compare the characteristics before and after rectification according to the original judgment rules, and the reviewers can then close or return the task.

[0058] The results of manual review and rectification acceptance first enter the feedback sample pool, without directly overwriting or modifying the formal rules or online model; when the sample pool meets the preset sample quantity requirements, is approved by the designated permission role, and reaches the preset performance indicators on the validation set, the scene template, rule weight, or recognition model parameters are officially updated; each update retains the complete version before and after the update and the rollback path, supporting version switching and historical judgment reproduction.

[0059] For example, during a review, the system re-identifies personnel positions and the status of protective devices, confirming that personnel have evacuated from dangerous areas and that remote inflation devices have been used. Once these conditions are met, manual confirmation and closure are initiated. The review results are entered into a sample pool. If the cumulative number of similar samples reaches a preset number (e.g., 50), and is approved by the safety administrator, and the accuracy on the validation set improves beyond a preset threshold, then the rule weights or model parameters are updated.

[0060] II. System Implementation Examples like Figure 2 As shown, the present invention also provides a hazard intelligent judgment system based on the fusion of multi-source perception and safety standard knowledge base, comprising: Data Access Module: Used to collect and access multi-source heterogeneous on-site safety data, assigning a unified data identifier to each data entry and attaching spatiotemporal, object, and quality attributes. It receives standardized data including images, videos, audio, sensor data, equipment operation data, personnel data, location data, ticket data, and text data.

[0061] The data governance module is used to perform time calibration, spatial correlation, format conversion, quality assessment, event aggregation, and scene label generation for multi-source data, outputting structured scene vectors. It aggregates multi-source data into security event data packets based on preset time windows and spatial neighborhoods.

[0062] Feature recognition module: Used to extract multi-dimensional candidate hazard features from security incident data through artificial intelligence models and deterministic algorithms. It can employ object detection, image segmentation, pose recognition, action recognition, OCR, temporal anomaly detection, voice event recognition, or threshold detection models.

[0063] Scenario Model: Used to store the relationships between areas, operations, objects, organizational systems, and dynamic states, and to define the scope of rule application. The scenario model, alongside the safety specification knowledge base, is used to map on-site facts to the applicable conditions of the specifications.

[0064] Security Specification Knowledge Base: Used to store structured specification clause units, specification original text location information, version relationships, hierarchical relationships, and conflict rules. The specification clause unit structure is R={id,v,t,a,o,j,p,n,f,e,h,m}, supporting extended access to specifications from multiple industries and at multiple levels.

[0065] The fusion judgment module is used to retrieve and match candidate rule sets based on scene vectors and candidate hazard features, perform constraint condition verification item by item, calculate fusion confidence, complete risk classification, and output hazard judgment results. The fusion judgment module has a built-in rule verification engine and confidence calculation unit. The rule verification engine performs necessary condition fulfillment verification, prohibited condition triggering verification, threshold condition matching verification, and exception condition exclusion verification for each candidate specification clause unit in sequence. The confidence calculation unit calculates the fusion confidence based on the perception model confidence, rule reliability weight, data quality coefficient, and multi-source consistency coefficient, and outputs the corresponding judgment conclusion based on the preset three-level thresholds.

[0066] Closed-loop management module: used to generate traceable evidence chains, rectification tasks, notifications and review records, and complete the entire process of hazard rectification.

[0067] The Learning and Update Module manages the review feedback sample pool, performs model evaluation, rule revision, version approval and release, and version rollback, achieving controlled rule and model iteration. The Learning and Update Module includes a built-in version management unit for unified freezing and release management of specification library versions, model versions, and rule versions, supporting historical version tracing and judgment process reproduction.

[0068] The structured and integrated judgment logic of the normative clauses is as follows: Figure 3 As shown.

[0069] III. Specific Application Examples In the mine vehicle repair area, a worker is inflating a large tire after it has been disassembled. A mobile terminal uploads images of the scene, a pressure sensor uploads data on tire pressure changes, an area label indicates that the location is a tire repair station, and personnel location data shows that the worker is within the possible trajectory of a tire bursting.

[0070] The feature recognition module extracts candidate features such as "the tire is inflated," "personnel are positioned to the side of the tire or in a direction where the rim may fly off," and "no isolated inflation cage or remote clamping inflation device was detected." The standard retrieval retrieves relevant safety operating procedure clauses based on the scenario field of "tire repair—inflation—personnel positioning—large vehicle tires," and verifies that the clause is a valid version and does not contain exceptions such as "equipment is in a fully depressurized maintenance state."

[0071] The fusion judgment module combines image positioning reliability, consistency of continuous pressure increase, region label reliability, and rule weights to determine the existence of a safety hazard: "personnel are within the danger zone during inflation." It generates an evidence chain including a site-marked diagram, pressure curve segments, the relative positions of personnel and tires, regulatory basis, and model version, and proposes corrective measures such as stopping inflation, evacuating personnel from the danger zone, using restraint devices, and remote inflation. During review, the system re-identifies personnel positions and the status of protective devices; if conditions are met, manual confirmation and shutdown are initiated.

[0072] IV. Alternative Implementation Methods The data collection method can use only images or videos, or a combination of images and any type of sensor; in the absence of a network, the edge computing terminal can complete the initial identification and rule verification, and synchronize the evidence and results after connecting to the network.

[0073] The standardized knowledge base can be stored using relational databases, document databases, or graph databases, or it can be a combination of vector databases and rule engines; semantic retrieval can be achieved through lexical retrieval, vector retrieval, knowledge graph path retrieval, or a combination thereof.

[0074] The perception model can be a locally deployed model, a cloud-based model, or a combination of both; it can use one or more of the following: object detection, visual language model, temporal model, expert rules, or statistical model.

[0075] Fusion determination is not limited to multiplication formulas; it can employ weighted summation, Bayesian inference, evidence theory, fuzzy comprehensive evaluation, decision trees, or validated learning ranking models, as long as the influence of data quality and rule constraints on the conclusion is preserved.

[0076] Risk classification can be achieved using enterprise-defined matrices, job condition hazard assessments, consequence-probability matrices, or levels directly stipulated by regulations.

[0077] The rectification loop can be integrated with existing enterprise office, equipment management, personnel management, SMS, instant messaging or work order systems, or it can run as an independent system.

[0078] The standard clauses can be manually structured, or they can be extracted by a natural language processing model and then manually reviewed and published; simple threshold clauses can be automatically converted, while complex exception clauses require manual confirmation.

[0079] Contents not described in detail in this specification are prior art known to those skilled in the art. Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A hidden danger intelligent judgment method based on multi-source perception and safety specification knowledge base fusion, characterized in that, Includes the following steps: Step S1. Collect multi-source safety data: Collect multi-source heterogeneous on-site safety data. Each data point includes a data source identifier, collection time, spatial location, object identifier, and data quality parameters. Step S2. Perform data preprocessing and scene modeling: Perform clock calibration, spatial coordinate transformation and event aggregation on multi-source security data, and construct a structured scene vector based on the related data within the same time window and spatial neighborhood. The scene vector includes object attributes, behavioral attributes, environmental attributes and management attributes. Step S3. Extract candidate hazard features: Extract multi-dimensional candidate hazard features from the aggregated security event data using artificial intelligence models and deterministic algorithms; Step S4. Retrieve safety specifications and form a candidate rule set: Based on the scenario vector and the characteristics of the candidate hidden dangers, retrieve and match from the pre-built safety specification knowledge base to obtain a candidate rule set; the safety specification knowledge base is composed of unstructured safety specifications parsed into structured specification clause units; Step S5. Perform fusion judgment and risk classification: Perform constraint condition verification on each normative clause unit in the candidate rule set, calculate the fusion confidence score by combining the perception model confidence score, rule reliability, data quality and multi-source consistency, classify the risk based on the fusion confidence score and risk assessment results, and output the corresponding hidden danger judgment results; where, the perception model confidence score Pm comes from the confidence score output by the artificial intelligence model, the rule reliability Wr comes from the historical accuracy and hierarchical weight of the normative clause, the data quality Dq comes from the data quality parameters and includes one or more of signal strength, measurement accuracy and historical reliability, and the multi-source consistency Sc comes from the multi-source data consistency verification results; Step S6. Generate a traceable chain of evidence and rectification tasks: Associate and bind the hazard assessment results, triggering conditions, standard basis, original evidence, model and standard version information to generate a complete chain of evidence, and match the responsible party to generate rectification tasks; Step S7. Manual review, rectification closed loop and controlled learning: Receive feedback results of manual review and rectification tasks and store them in the sample pool. After meeting the preset sample quantity, permission approval and verification performance conditions, update the scene template, rule weight or model parameters, and retain the version before and after the update and rollback information.

2. The hidden danger intelligent judgment method based on multi-source perception and safety specification knowledge base fusion according to claim 1, characterized in that: In step S2, the construction of the scene vector specifically includes: Based on a preset time window and spatial neighborhood range, multi-source data of the same personnel, equipment, area or task are aggregated into a security event data packet; Extract attributes from the security event data packet, including region type, operation type, personnel role, equipment type, environment status, license status, time status, and organizational policy version, construct a scenario vector, and retain the data source and credibility mark of each attribute field.

3. The hidden danger intelligent judgment method based on multi-source perception and safety specification knowledge base fusion according to claim 1, characterized in that: In step S3, the extraction of candidate hidden danger features also includes multi-source feature fusion and conflict resolution: For cases where the same hazard feature is repeatedly observed from multiple data sources, feature fusion is performed based on data source reliability, temporal proximity, and spatial consistency. When feature recognition results from different data sources contradict each other, a conflict marker is generated and the multi-source consistency coefficient of the corresponding feature is reduced, without directly overwriting the recognition results of any data source.

4. The intelligent hazard assessment method based on the fusion of multi-source perception and safety standard knowledge base as described in claim 1, characterized in that: In step S4, the structure of the standard clause unit is R={id,v,t,a,o,j,p,n,f,e,h,m}, where: id is the standard and clause identifier, v is the version number, t is the effective time range, a is the applicable region or industry, o is the applicable object, j is the operation type, p is the set of necessary conditions, n is the set of prohibited conditions, f is the threshold function, e is the set of exception conditions, h is the mapping of hazard category and level, and m is the set of rectification measures; The retrieval and matching adopts a two-stage retrieval: the first stage recalls relevant regulatory clauses through semantic vectors or keywords, and the second stage filters fields according to version validity, applicable objects, applicable regions, job types and enterprise levels to obtain a candidate rule set.

5. The intelligent hazard judgment method based on the fusion of multi-source perception and safety standard knowledge base as described in claim 1, characterized in that: In step S5, the constraint verification specifically includes: For each specification clause unit in the candidate rule set, the necessary conditions, prohibited conditions, threshold conditions, and exception conditions are checked in sequence. When all necessary conditions are met, prohibited conditions are triggered, and exception conditions are not met, a potential hazard is formed. Threshold conditions are judged based on the collected values, unit conversion, duration, and measurement uncertainty. The formula for calculating the fusion confidence C is: C=Pm×Wr×Dq×Sc, where Pm is the confidence of the perception model, Wr is the reliability weight of the clause rule, Dq is the data quality coefficient, and Sc is the multi-source consistency coefficient. The formula for calculating the risk value Q is: Q = C × S × E × K, where S is the severity of the consequences, E is the exposure level, and K is the scenario correction factor; Three-level judgment intervals are set according to the fusion confidence level: when the automatic confirmation threshold is reached, the system judgment conclusion is output; when it is in the review interval, the conclusion to be manually reviewed is output; and when it is below the suppression threshold, only candidate records are retained.

6. The intelligent hazard judgment method based on the fusion of multi-source perception and safety standard knowledge base as described in claim 1, characterized in that: In step S6, the traceable evidence chain specifically includes: hazard type, risk level, occurrence time and location, involved objects, triggering conditions, unmet safety requirements, standard name and specific clause number, evidence image or data fragment, identification model version, standard library version, fusion confidence level and suggested rectification measures; All evidence documents and rectification tasks are associated and bound with a unique task identifier, and a summary value is calculated for each evidence document to prevent mismatch.

7. The intelligent hazard judgment method based on the fusion of multi-source perception and safety standard knowledge base as described in claim 1, characterized in that: In step S7, the controlled update specifically includes: The results of manual review and rectification acceptance are first entered into the feedback sample pool and do not directly cover the modification of formal rules or online models. When the sample pool meets the preset sample quantity requirement, is approved by the designated permission role, and reaches the preset performance index on the validation set, the scene template, rule weight, or recognition model parameters are officially updated. Each update retains the complete version and rollback path before and after the update, and supports version switching and historical judgment reproduction.

8. A hazard identification system based on the fusion of multi-source perception and a safety standard knowledge base, characterized in that, include: The data access module is used to collect and access multi-source heterogeneous field safety data, assign a unified data identifier to each data point and attach collection time, space, object and quality attributes; The data governance module is used to complete the temporal calibration, spatial correlation, format conversion, quality assessment, event aggregation and scene label generation of multi-source data, and output structured scene vectors; The feature recognition module is used to extract multi-dimensional candidate hazard features from security incident data through artificial intelligence models and deterministic algorithms; A scenario model is used to store the relationships between regions, tasks, objects, organizational systems, and dynamic states, and to define the scope of application of rules. The security specification knowledge base is used to store structured specification clause units, specification original text location information, version relationships, hierarchical relationships, and conflict rules; The fusion judgment module is used to retrieve and match candidate rule sets based on scene vectors and candidate hazard features, perform constraint condition verification item by item, calculate fusion confidence and complete risk classification, and output hazard judgment results. The closed-loop management module is used to generate traceable evidence chains, rectification tasks, notifications and review records to complete the entire process of hazard rectification. The learning and updating module is used to manage the review feedback sample pool, perform model evaluation, rule revision, version approval and release, and version rollback, so as to achieve controlled rule and model iteration.

9. The intelligent hazard judgment system based on the fusion of multi-source perception and safety standard knowledge base as described in claim 8, characterized in that: The fusion judgment module has a built-in rule verification engine and a confidence calculation unit; The rule verification engine is used to perform necessary condition satisfaction verification, prohibited condition triggering verification, threshold condition matching verification and exception condition exclusion verification for each candidate specification clause unit in sequence. The confidence calculation unit is used to calculate the fusion confidence based on the confidence of the perception model, the reliability weight of the rule, the data quality coefficient and the multi-source consistency coefficient, and output the corresponding judgment conclusion according to the preset three-level threshold.

10. The intelligent hazard judgment system based on the fusion of multi-source perception and safety standard knowledge base as described in claim 8, characterized in that: The security specification knowledge base supports extended access to specifications across multiple industries and at multiple levels. By adding data adapters, scenario templates, and specification clause units, it can adapt to security supervision scenarios in different industries. The learning and update module has a built-in version management unit that uniformly freezes and releases the specification library version, model version, and rule version, and supports historical version tracing and judgment process reproduction.