Hidden danger reporting and checking system

Abstract

The invention provides a hidden danger reporting and checking system, and relates to the technical field of safety management, and the system comprises a hidden danger reporting module which is used for receiving hidden danger information submitted by a user, and the hidden danger information comprises a hidden danger picture, text description, hidden danger classification, a place, reporting time and reporter information; the multi-hidden-danger association module is used for automatically associating related hidden danger records based on space-time overlap analysis, generating a composite risk report and triggering upgrading early warning to the early warning module; the early warning module is used for carrying out graded early warning according to hidden danger classification, relevance and historical data; the closed-loop regulation module is used for rectification task assignment, process feedback, safety department acceptance and result archiving; the knowledge base module is used for storing the mapping relation between the hidden danger type and the pre-control countermeasure and supporting fuzzy search and dynamic update of the countermeasure; the hidden danger management efficiency is improved, the hidden danger identification accuracy is improved, and resource waste is reduced.

Description

Technical Field

[0001] This invention relates to the field of safety management technology, and in particular to a hazard reporting and investigation system. Background Technology

[0002] In fields such as workplace safety, engineering management, and public facility maintenance, the timely detection, accurate assessment, and efficient remediation of potential hazards are core aspects of ensuring operational safety. Traditional hazard identification methods mainly rely on manual inspection records, paper form submissions, or simple information registration, which presents numerous insurmountable technical challenges.

[0003] First, the hazard reporting process lacks standardized constraints, making it difficult to guarantee the completeness and accuracy of reported information. Manually taken photos of hazards often suffer from insufficient clarity, obscured key features, or improper shooting angles, leading to difficulties in subsequent review and analysis. Text descriptions, due to differences in individual expression habits, often result in vague descriptions, missing key features, or contradictions with the actual hazards. Furthermore, hazard classification relies on manual selection, which is prone to errors due to insufficient professional knowledge, affecting subsequent statistics and the matching of countermeasures.

[0004] Secondly, the ability to analyze the correlation of hidden dangers is weak, making it difficult to identify complex risks. A single hidden danger may be a manifestation of a systemic problem, but the traditional approach of treating them in isolation cannot uncover the inherent connections between hidden dangers through dimensions such as spatiotemporal correlation and causal relationships. This leads to the neglect of potential complex risks, inaccurate determination of early warning levels, and missed opportunities for optimal remediation.

[0005] Furthermore, the review mechanism lacks intelligent support, resulting in low efficiency and a high risk of oversight. Manual review is not only time-consuming and labor-intensive, but also struggles to ensure consistency between textual descriptions and image information through cross-validation. The lack of unified standards for handling industry terminology and vague expressions leads to frequent false alarms and omissions. In addition, it lacks self-learning and optimization capabilities, making it impossible to continuously improve review accuracy based on historical review data.

[0006] Furthermore, the linkage between the closed-loop management of rectification and the knowledge base is insufficient. The level of informatization in the assignment of rectification tasks, process tracking and acceptance archiving is low, making it difficult to achieve full-process traceability. The matching of hazard types and pre-control countermeasures relies on manual experience, and the lack of a structured knowledge base to support rapid retrieval and dynamic updates makes it impossible to call effective countermeasures in a timely manner when similar hazards recur, affecting the efficiency and effectiveness of rectification.

[0007] Meanwhile, the mechanism for handling duplicate reports is imperfect. When the same hidden danger is reported multiple times, there is a lack of intelligent identification and hierarchical processing capabilities, which not only causes information redundancy but also increases the review and management costs. Furthermore, it is impossible to conduct time-series analysis on the duplicate report data to help assess the development trend of hidden dangers. Summary of the Invention

[0008] Based on the above problems, this invention proposes a hidden danger reporting and investigation system, which improves the efficiency of hidden danger management, enhances the accuracy of hidden danger identification, reduces resource waste, and provides an efficient risk prevention and control solution for fields such as safe production and engineering maintenance.

[0009] The objective of this invention is achieved through the following technical solution:

[0010] This invention provides a hazard reporting and investigation system, including:

[0011] The hazard reporting module is used to receive hazard information submitted by users, including hazard photos, text descriptions, hazard categories, locations, reporting times, and reporter information;

[0012] The multi-hazard association module is used to automatically associate related hazard records based on spatiotemporal overlap analysis, generate a composite risk report, and trigger an escalation warning to the warning module;

[0013] The early warning module is used to issue tiered early warnings based on hazard classification, correlation, and historical data.

[0014] The closed-loop remediation module is used for remediation task assignment, process feedback, safety department acceptance, and result archiving.

[0015] The knowledge base module stores the mapping relationship between hazard types and prevention and control measures, and supports fuzzy search and dynamic updates of the measures.

[0016] Preferably, the hazard reporting module includes:

[0017] The basic identification unit is used to receive at least one photo of the hazard, which includes the hazard subject and the surrounding environment, and automatically record the shooting time and geographical location;

[0018] The feature description unit is used to collect text descriptions containing specific manifestations and potential risks of hidden dangers, and associate them with keyword suggestions from the knowledge base to intelligently recommend hidden danger classifications based on the text description content;

[0019] The classification and positioning unit is used to provide standardized multi-level hazard classification options and multi-level positioning associated with electronic maps;

[0020] The intelligent review unit connects with the information collected by the above units, detects and identifies the compliance, clarity, and obstruction issues of potential problems in photos, and triggers a re-upload request for photos that do not meet the requirements; at the same time, it parses the text description content, matches it with a preset keyword library, and generates manual review prompts for submissions with contradictory descriptions or incorrect classifications.

[0021] Preferably, the intelligent review unit includes:

[0022] The cross-validation subunit is used to identify the type of objects and abnormal states in the photos of potential hazards; it compares the hazard characteristics described in the text with the image recognition results, and generates conflict markers and targeted review prompts for contradictory features;

[0023] The self-learning optimization subunit is used to automatically expand the thesaurus based on false positive samples reviewed by humans, record user descriptions with a deviation frequency greater than the frequency threshold, push customized filling guidance to users with consecutive classification errors exceeding the number threshold, and generate semantic completion prompts for fuzzy descriptions.

[0024] Preferably, the execution steps of the cross-validation subunit include:

[0025] Extract the core entities and core anomalies from the text description and perform rigid matching with the image recognition results. When the core entity is missing or the anomaly direction is completely contradictory, it is marked as a serious conflict.

[0026] By leveraging the semantic association network of the associated knowledge base, we can expand the matching of fuzzy descriptions or industry terms.

[0027] By combining the contextual information of the location of the hazard and the type of equipment, the scenario logic is verified, and features that violate the scenario constraints are marked as scenario logic conflicts.

[0028] Preferably, the step of extracting the core entities and core anomalies from the text description and rigidly matching them with the image recognition results includes:

[0029] Natural language processing is used to extract locative words related to the core anomaly from the text description, and spatial parameters are dynamically corrected by combining the electronic map coordinates and location type of the potential hazard location.

[0030] Heatmaps of target objects are generated based on object detection algorithms, and the heatmap regions are bound to absolute spatial coordinates. For irregular objects, the fan-shaped detection area is optimized into a rectangular detection band along the object's axis, and the angle range adapts to the object's orientation.

[0031] Calculate the spatial overlap rate between the pixel region of the target object in the photo and the heatmap, and introduce key feature weights;

[0032] When the overlap rate is less than the first overlap rate threshold, it is judged as an abnormal direction partial contradiction, and is not directly marked as a serious conflict, but a warning of insufficient view coverage and a reshoot guide are generated; if the overlap rate is less than the second overlap rate threshold, a serious conflict mark of complete abnormal direction contradiction is triggered; the first overlap rate threshold and the second overlap rate threshold are determined according to scene classification, risk coefficient and misjudgment rate.

[0033] Preferably, the step of performing scene logic verification by combining contextual information about the location of the potential hazard and the type of equipment, and marking features that violate scene constraints as scene logic conflicts, includes:

[0034] To address progressive hazards, a historical hazard photo library of the same location and equipment is accessed, and historical damage features are extracted using image segmentation algorithms to establish a historical feature sequence for the hazard.

[0035] Based on the historical damage feature sequence, the historical rate of change of the damage area is calculated to preliminarily determine the expected value of the damage area at the current time point.

[0036] The system automatically retrieves historical environmental data from the same location and calculates the impact weights of each environmental factor on the damage using a multiple linear regression model. The weights are then used to correct the damage area change curve to obtain the current expected value for environmental adaptation.

[0037] The corrected historical rate of change is input into the LSTM time series prediction model to generate a predicted trend of damage development; at the same time, damage features in the current photo are extracted and their deviation from the expected value after environmental correction is calculated.

[0038] If the deviation is less than the deviation threshold, it is determined that the time-dimensional scenario constraints are met.

[0039] If the deviation is greater than or equal to the deviation threshold, it is marked as a scene logic conflict, triggering a status anomaly review prompt, and automatically marking the areas in the photo that differ from historical features with bounding boxes.

[0040] Preferably, the multi-hazard association module includes a multi-dimensional association engine, which, based on spatiotemporal overlap analysis, achieves hazard association through the following extended dimensions:

[0041] The causal relationship between newly reported hazards and historical hazard records is determined by searching a hazard knowledge graph. When the relationship strength is greater than or equal to the relationship threshold, a relationship is established between the newly reported hazard and the historical hazard record. The hazard knowledge graph is pre-configured with a causal relationship network between hazard types.

[0042] Based on the type, severity, and surrounding environmental parameters of the hazard, the potential impact range of a single hazard is automatically calculated. When the spatial range of a newly reported hazard overlaps with or contains the impact range of a historical hazard, the association is triggered and the overlapping area is marked.

[0043] Preferably, the multi-dimensional association engine employs dynamic weighted association logic, including:

[0044] Adjustable weight parameters are configured for the dimensions of spatiotemporal overlap, causal correlation, and scope of influence, and the weight values ​​are dynamically adapted according to the type of hazard.

[0045] The association level is determined by the comprehensive association score (the sum of the products of the weights of each dimension and the association strength). When the comprehensive association score is greater than or equal to the score threshold, it is determined to be a strong association and included in the composite risk analysis.

[0046] Preferably, the multi-hazard association module further includes a duplicate reporting identification engine for determining and classifying duplicate reports, including:

[0047] The time threshold is dynamically set according to the urgency of the hidden danger. When the newly reported record and the historical record meet the preset spatial range and time difference requirements, the initial association is triggered.

[0048] For records that are initially associated, the consistency between core damage features and textual descriptions is compared through image feature extraction and semantic analysis.

[0049] When the comprehensive matching degree of space, features and semantics reaches the preset threshold, it is judged as a completely duplicate report; when the core features are the same but there are non-critical differences, it is judged as an approximately duplicate report.

[0050] For completely duplicate reports, the review process for new reports will be automatically terminated, a related prompt will be sent to the reporter, and the new report information will be merged into the history record and supplemented as time-series analysis data.

[0051] For near-duplicate reports that trigger the intelligent audit unit's difference verification, the feature difference area is marked and a confirmation guide is generated; based on the reporter's feedback, if it is confirmed to be a duplicate description, it is merged; if it is confirmed to be a new feature, the original record is upgraded to a composite risk and the association is retained.

[0052] Record the false positive rate of repeated judgments, and dynamically optimize the spatiotemporal judgment parameters through machine learning models for the types of hidden dangers with excessive false positive rates;

[0053] We analyzed the adoption rate of user feedback on the difference confirmation guidelines, and optimized the wording and supplementary information for guidelines with adoption rates below the threshold.

[0054] Preferably, the early warning module includes:

[0055] A multi-level early warning unit is used to establish a three-dimensional early warning scoring model based on the inherent risk level of the hidden danger classification, the comprehensive correlation score output by the multi-hidden danger correlation module, the rectification rate of similar hidden dangers in historical data, and the accident occurrence rate. In the three-dimensional early warning scoring model, the inherent risk level, comprehensive correlation score, historical rectification rate, and accident occurrence rate are each configured with dynamic weights, and the weight values ​​are dynamically calibrated according to the industry type. The early warning is divided into multiple levels according to the three-dimensional scoring results, and different levels of early warning correspond to different response time limits and processing priorities.

[0056] The early warning update unit is used to synchronize the rectification progress data of the closed-loop rectification module in real time. When a hidden danger enters the rectification process, the early warning level is automatically reduced. When the rectification is accepted and archived, the early warning is terminated. If the rectification is not completed within the time limit or a new hidden danger is added to the multi-hidden danger association module, the early warning level upgrade mechanism is triggered. The upgrade amount is determined according to the time limit or the risk level of the newly added associated hidden danger. It is connected to the external environmental monitoring system. When the environmental parameters of the hidden danger location exceed the safety threshold, the early warning level is dynamically upgraded based on the environmental impact factor model of the knowledge base module.

[0057] The beneficial effects of this invention include: the hazard reporting module automatically records spatiotemporal information through the basic identification unit, and combines this with the intelligent review unit to perform dual verification of photo compliance and text description accuracy, effectively solving problems such as blurry photos, missing information, and contradictory descriptions in traditional reporting. The keyword suggestion and classification recommendation functions of the feature description unit reduce the error rate of manual classification, and the standardized multi-level classification and positioning options provide a unified benchmark for subsequent data analysis, reducing information silos.

[0058] The multi-hazard correlation module, based on multi-dimensional analysis including spatiotemporal overlap, causal relationships, and impact scope, can automatically identify the inherent connections between scattered hazards, generate composite risk reports, and trigger escalation warnings. This correlation analysis capability breaks through the limitations of traditional isolated processing modes, enabling the early detection of systemic risks (such as cascading hazards caused by aging equipment in the same area). The warning module's three-dimensional scoring model combines inherent risks, correlation strength, and historical data to make warning levels more accurate, providing a scientific basis for prioritization decisions.

[0059] The cross-validation subunit of the intelligent review unit significantly reduces the workload of manual review through rigid matching of images and text, semantic expansion matching, and scene logic verification. Targeted review prompts for conflicting items improve the effectiveness of problem handling. The self-learning optimization subunit expands the vocabulary based on manual review samples and pushes customized guidance, enabling the system to dynamically evolve, continuously reducing false positive and classification error rates, and adapting to the expression habits of users in different industries.

[0060] The closed-loop rectification module covers the entire process of task assignment, process feedback, and acceptance archiving, ensuring that the responsibility for hazard rectification is assigned to specific individuals and is traceable, avoiding the loophole of "no follow-up after reporting" in traditional management. The knowledge base module stores the mapping relationship between hazard types and pre-control countermeasures, supports fuzzy search and dynamic updates, enabling newly reported hazards to be quickly matched with mature countermeasures, while rectification experience is fed back into the knowledge base, forming a virtuous cycle of "reporting - rectification - accumulation - reuse".

[0061] The duplicate reporting identification engine accurately distinguishes between completely duplicate and nearly duplicate reports through spatiotemporal thresholds, feature matching, and semantic analysis, avoiding information redundancy and resource waste. The merging processing of completely duplicate reports and the difference verification of nearly duplicate reports not only ensure data uniqueness but also track the changing trends of hidden dangers (such as the expansion of the damaged area) through time series analysis, providing data support for the prediction of hidden danger development. At the same time, the engine continuously improves the identification accuracy by optimizing the judgment parameters through machine learning.

[0062] Through multi-module collaboration, the system achieves intelligent management across the entire chain, from hazard discovery, analysis, and early warning to rectification. Dynamic updates to the knowledge base and countermeasure recommendations help frontline personnel quickly implement preventative measures. Closed-loop management data and historical hazard analysis provide management with decision-making support for equipment maintenance and process optimization (such as adjusting inspection frequency when a certain type of hazard occurs frequently), shifting from reactive response to proactive prevention and significantly reducing the accident rate. Attached Figure Description

[0063] Figure 1 This is a schematic diagram of the hidden danger reporting and investigation system according to an embodiment of the present invention.

[0064] Figure 2 This is a schematic diagram of the hidden danger reporting interface in an embodiment of the present invention. Detailed Implementation

[0065] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0066] See attached document Figure 1 This application provides a hazard reporting and investigation system, including:

[0067] The hazard reporting module is used to receive hazard information submitted by users, including hazard photos, text descriptions, hazard categories, locations, reporting times, and reporter information;

[0068] The multi-hazard association module is used to automatically associate related hazard records based on spatiotemporal overlap analysis (multiple hazards at the same location / equipment within a preset time period), generate a composite risk report, and trigger an escalation warning to the warning module;

[0069] The early warning module is used to issue tiered early warnings based on hazard classification, correlation, and historical data.

[0070] When the same type of hazard reaches a dynamic threshold (such as the average of the past 7 days ± 2 standard deviations) within a set period, a recurring alarm will be sent to the safety department.

[0071] If the rectification is not completed by the deadline, a reminder will be sent to the responsible person and the safety department at each level (e.g., a text message 3 days in advance, or automatic escalation to the superior department if the deadline is exceeded).

[0072] The closed-loop remediation module is used for remediation task assignment, process feedback, safety department acceptance, and result archiving.

[0073] The knowledge base module stores the mapping relationship between hazard types and prevention and control measures, and supports fuzzy search and dynamic updates of the measures.

[0074] The working principle of the above technical solution is as follows:

[0075] The hazard reporting module serves as the system's information entry point, receiving hazard photos, text descriptions, classifications, locations, and other information submitted by users, providing basic data for subsequent processing.

[0076] The multi-hazard association module performs in-depth analysis of this information, automatically associates related hazard records based on spatiotemporal overlap analysis, generates composite risk reports, and triggers escalation warnings to the warning module; this allows managers to more clearly grasp the inherent connections between hazards, making it easier to formulate response strategies from a holistic perspective and reducing the limitations of handling hazards in isolation.

[0077] The early warning module implements tiered early warnings based on hazard classification, correlation, and historical data. When similar hazards reach a dynamic threshold (e.g., the average of the past 7 days ± 2 standard deviations) within a set period, a recurring alarm is sent to the safety department to ensure timely monitoring of high-frequency hazards. Simultaneously, for cases where rectification is not completed by the deadline, tiered reminders are sent to the responsible person and the safety department, such as a text message reminder 3 days in advance; if the deadline is exceeded, the reminder is automatically escalated to the superior department to ensure timely rectification.

[0078] The closed-loop rectification module receives early warning information and is responsible for assigning rectification tasks, collecting process feedback, conducting safety department acceptance, and archiving results, forming a complete rectification closed loop. This module clearly defines the responsibilities of each stage of rectification, ensuring reasonable task assignment, traceable processes, standardized acceptance, and archiving results, thus enhancing the standardization and enforceability of rectification work. The tiered reminder mechanism also effectively urges responsible parties to complete rectification on time.

[0079] The knowledge base module stores the mapping relationship between hazard types and pre-control countermeasures, supports fuzzy search and dynamic updates of countermeasures, provides knowledge support for hazard handling, and can continuously optimize its content based on actual applications. The mapping relationship between hazard types and pre-control countermeasures stored in the knowledge base module supports fuzzy search and dynamic updates, enabling relevant personnel to quickly obtain effective countermeasures, improve the efficiency of hazard handling, and promote the accumulation and sharing of knowledge, avoiding duplication of work.

[0080] In one possible implementation, the hazard reporting module includes:

[0081] The basic identification unit is used to receive at least one (1-5 photos) of the hazard subject and its surrounding environment (e.g., equipment hazards must show the equipment nameplate and the area markings), and automatically record the shooting time and geographical location. The basic identification unit avoids unclear rectification goals due to ambiguous information. The feature description unit has mandatory requirements for "specific manifestations + potential risks", combined with keyword prompts, to make the text description more accurate and reduce ambiguity.

[0082] The user information unit is used to automatically retrieve the reporter's information (without requiring manual filling) through the associated user account, including name, department, position, and contact information (for subsequent verification and supplementation of information); by automatically retrieving the reporter's information, the manual filling step is eliminated; the classification and positioning unit supports intelligent recommendation of categories and map point selection positioning, which lowers the professional threshold for users; the status mark unit automatically generates the reporting time, reduces human intervention, shortens the overall operation time of the reporting process, and improves the user experience;

[0083] The feature description unit is used to collect textual descriptions containing specific manifestations of hazards and potential risks, and associates them with keyword suggestions from the knowledge base. Based on the textual description content, it intelligently recommends hazard categories. The textual description must include specific manifestations of hazards (such as missing distribution box cabinet doors, or debris piled up in fire escape routes on the ground) and potential risks (such as potential electric shock or obstruction of emergency evacuation). By associating with keyword suggestions from the knowledge base and using the intelligent recommendation category function, it not only assists users in reporting in a more standardized manner, but also enriches the keyword database and category mapping relationship of the knowledge base through the effective data reported by users, thereby enhancing the practicality and dynamic adaptability of the knowledge base.

[0084] The classification and positioning unit provides standardized multi-level hazard classification options (such as primary classification: equipment and facilities / working environment / management deficiencies; secondary classification: electrical equipment / fire protection facilities / chemical storage, etc.) and multi-level precise positioning (down to the area that can be rectified) linked to an electronic map; classification options support manual selection or intelligent recommendation (based on text description matching); it supports multi-level positioning (such as Building A of the factory area / second-floor workshop / north side assembly line), and can be linked to electronic map points or input specific coordinates to ensure positioning accuracy down to the area that can be rectified; hazard photos are accompanied by the shooting time and geographical location information automatically recorded by the terminal device, which is matched and verified with the hazard location information; the multi-level precision of classification and positioning (down to the area that can be rectified) avoids the shirking of rectification responsibility due to ambiguous positioning;

[0085] The status marker unit automatically generates an unmodifiable reporting time and provides urgency options (general / urgent / critical). When the urgency level is "critical," the highest priority assignment of rectification tasks is automatically triggered. The "critical" option of the status marker unit can automatically trigger the highest priority assignment, enabling urgent hazards to quickly enter the processing flow, shortening response time, and reducing the risk of accidents. The standardized multi-level classification and positioning provide a unified and accurate data foundation for the spatiotemporal analysis of subsequent multi-hazard association modules and the hierarchical early warning of the early warning module, improving the overall data analysis efficiency of the system.

[0086] The intelligent review unit connects with the information collected by the aforementioned units, detects and identifies the compliance (must include equipment nameplates or area identifiers, and the geographical location must be bound to the factory map coordinates or latitude and longitude) and clarity and obstruction issues of hazard photos, and triggers a re-upload request for photos that do not meet the requirements; at the same time, it parses the text description content, matches it with a preset keyword library, and generates manual review prompts for submissions with contradictory descriptions or incorrect classifications; through photo compliance detection (such as geographical location bound to factory coordinates) and text description and classification matching verification, it effectively filters out non-compliant, erroneous or contradictory reported information, reducing the ineffective costs of subsequent review and rectification.

[0087] In one possible implementation, the intelligent auditing unit includes:

[0088] The spatiotemporal verification subunit is used to extract the shooting timestamp (accurate to the second) of hazard photos, calculate the difference between the shooting timestamp and the reporting time, and determine that the photo is not real-time when the difference exceeds a preset threshold (e.g., configurable to ≤30 minutes). It also calculates the matching degree between the photo's latitude and longitude information and the electronic map coordinates of the reporting location. When the matching degree is <90%, a targeted retransmission request for "supplementing real-time photos" is triggered. By accurately comparing the difference between the photo's shooting timestamp and the reporting time, and the matching degree between the photo's latitude and longitude and the reporting location coordinates, it strictly filters out non-real-time photos and reporting information that do not match the location, ensuring the on-site timeliness and spatial accuracy of hazard information. The configurability of the preset thresholds (e.g., 30-minute time difference, 90% matching degree) can adapt to the management needs of different scenarios, reducing problems such as "off-site reporting" and "false positioning," and providing reliable basic data for subsequent rectification.

[0089] The cross-verification subunit is used to identify object types (such as electrical boxes and fire hydrants) and abnormal states (such as damage and obstruction) in photos using the YOLO+ classification network. It compares the hazard characteristics of the text description with the image recognition results. When the core object mentioned in the text is not identified in the image (e.g., a report of "fire exit" but no corresponding scene), or the spatial overlap rate of abnormal types is <50% (e.g., the text "stacked miscellaneous items" does not match the image "shelf"), a conflict marker and targeted review prompt are generated. Leveraging the image recognition capabilities of the YOLO+ classification network, a rigid match between the text description and the photo content is achieved. Through core object identification, abnormal state comparison, and spatial overlap rate calculation, issues such as "text-image contradictions" (e.g., mentioning a fire exit but no corresponding scene) are accurately marked. This cross-dimensional verification overcomes the subjective limitations of manual review, significantly reduces misjudgments caused by inconsistent information, provides clear directional guidance for manual review, and improves the efficiency of review decisions.

[0090] The semantic learning subunit constructs a two-tiered architecture: a basic keyword library (associated with standard terminology from the knowledge base) and an industry dynamic thesaurus (automatically updated by crawling annual safety bulletins). Based on manually reviewed false alarm samples, it automatically expands the thesaurus (e.g., associating "exposed wires" with "damaged wire insulation"), records high-frequency user description deviations (deviation frequencies exceeding a frequency threshold) (e.g., confusing "mechanical injury" with "pinch injury"), and pushes customized filling guidance (e.g., a classification decision tree interface) to users with consecutive classification errors exceeding a threshold. It generates semantic completion prompts (e.g., "Does this refer to abnormal noise / oil leak / overheating") for vague descriptions (e.g., "the equipment has a problem"). It also records deviation frequencies. The semantic learning subunit's two-tiered thesaurus architecture (basic keyword library + industry dynamic thesaurus) ensures the standardized application of standard terminology while also allowing for real-time updates of new industry terms through crawling annual safety bulletins, avoiding misjudgments caused by terminology differences. The expanded thesaurus (e.g., associating "exposed wires" with "damaged wire insulation") and the user high-frequency deviation recording function can adapt to the different expression habits of different users and reduce the review obstacles caused by differences in description; the semantic completion prompts for fuzzy descriptions lower the threshold for users to fill in the information and improve the accuracy of text descriptions.

[0091] Priority scheduling subunit: Combining the urgency of the hidden danger (urgent / critical) and the error type (e.g., irrelevant photos are serious errors, redundant descriptions are minor errors), the abnormal prompts are divided into three levels: P1, P2, and P3. Level P1 (serious errors of urgent hidden dangers) are automatically pushed to senior reviewers at the top, and level P3 are merged into a daily summary list; improving the scheduling efficiency and response time of review resources.

[0092] The self-learning optimization subunit records the differences between manual review and initial system judgment (such as misjudging occlusion as valid photos), and iteratively optimizes the model through random forest + BERT fine-tuning: dynamically adjusting the photo clarity threshold and strengthening the identification of high-frequency error descriptions; the self-learning mechanism continuously optimizes the review standards through data accumulation, reduces the frequency of manual intervention, lowers training and management costs, and forms a virtuous cycle of "reporting-reviewing-learning-optimization", thereby improving the standardization level of the entire hidden danger reporting system.

[0093] In one possible implementation, the execution steps of the cross-validation subunit include:

[0094] Extract the core entities and core anomalies from the text description and perform rigid matching with the image recognition results. When the core entity is missing or the anomaly direction is completely contradictory, it is marked as a serious conflict.

[0095] By leveraging the semantic association network of the associated knowledge base, we can expand the matching of fuzzy descriptions or industry terms.

[0096] By combining the contextual information of the location of the hazard and the type of equipment, the scenario logic is verified, and features that violate the scenario constraints are marked as scenario logic conflicts.

[0097] The working principle of the above technical solution is as follows:

[0098] Extracting "core entities" (such as "fire hydrant" and "electrical distribution box") and "core anomalies" (such as "missing" or "oil leak") from text descriptions, and rigidly comparing them with the object types and abnormal states identified in image recognition, ensures "consistency of subject" (e.g., if the text mentions "fire hydrant," the image must identify a fire hydrant) and "consistency of anomaly direction" (e.g., if the text mentions "liquid leakage," the image must identify "liquid stains / dripping features"). By extracting core entities and core anomalies from text descriptions and rigidly matching them with image recognition results, key links in information discrepancies can be accurately identified. When a core entity is missing or the anomaly direction is completely contradictory, it is marked as a serious conflict, avoiding oversights in the review process due to information misalignment, providing clear direction for subsequent manual review, and ensuring the consistency of potential hazard information from the source.

[0099] By leveraging the semantic association network (including synonyms, near-synonyms, and scenario-derived terms) of the knowledge base, ambiguous descriptions or industry terms can be expanded and matched (e.g., "running, leaking, dripping" can be associated with "liquid leakage"). This semantic association network effectively solves the information matching barriers caused by differences in expression habits and professional terminology. For example, for ambiguous descriptions like "the outer sheath of the line is damaged," the semantic network can be used to associate them with standard expressions such as "damaged wire insulation" and "damaged cable sheath." For industry terms like "running, leaking, dripping," the matching can be expanded to specific abnormal states such as "pipeline leakage" and "valve dripping," ensuring effective semantic connection between text and image information and reducing misjudgments caused by non-standard expressions.

[0100] By combining contextual information about the location of potential hazards and the type of equipment, scenario logic verification is performed. Features that violate scenario constraints (such as contradictions between descriptions and images related to open flames in a dust workshop) are marked as "scenario logic conflict." This contextual verification effectively identifies abnormal features that violate real-world scenario constraints. For example, in a "flammable and explosive warehouse" scenario, if the text describes "using open flame equipment" but the image does not show fire prevention measures, the system can mark it as a scenario logic conflict. For equipment like "high-voltage distribution cabinets," if the text mentions "hands-free operation" but the image shows no insulated protective tools, a logic verification alarm will also be triggered. This process ensures that hazard information conforms to the physical laws, safety regulations, and equipment characteristics of the specific scenario, improving the rationality and credibility of the information.

[0101] When the cross-validation subunit marks these two types of conflicts, it will directly trigger a targeted review prompt, clearly marking the conflict points and pushing it to the manual review stage or initiating a more stringent intelligent verification process. The conflict mark will be accompanied by specific conflict content (such as "the text description is 'pipeline leak' but the image does not identify the pipeline entity" or "the electrical equipment hazard appears in a restricted area"), providing reviewers with targeted review directions and providing the reporter with correction guidance (such as supplementing photos and correcting descriptions).

[0102] In one possible implementation, the cross-validation subunit uses a vision-language pre-trained model (such as CLIP) to achieve cross-modal matching, including: extracting deep semantic features from the image, encoding semantic vectors for the text, judging consistency by calculating the semantic vector similarity between the two (rather than keyword matching); and dynamically adjusting the similarity threshold based on the risk level of the hidden danger (high-risk hidden danger ≥90%, low-risk hidden danger ≥60%), the threshold can be optimized through self-learning from historical conflict cases.

[0103] In one possible implementation, the extraction of core entities and core anomalies from the text description and the rigid matching with the image recognition results include:

[0104] Natural Language Processing (NLP) is used to extract locative terms (such as "left / above / adjacent equipment") related to the core anomaly from the text description. These are then combined with the electronic map coordinates and location type of the potential hazard location to dynamically adjust spatial parameters. For example:

[0105] In indoor scenarios (such as workshops), the default distance threshold d between adjacent devices is 1-3 meters.

[0106] In outdoor scenarios (such as factory areas), the default value of d is 3-8 meters, and users can manually fine-tune it.

[0107] The relative orientations mentioned above are converted into absolute spatial coordinates (x, y, z axis positioning); directional words in the text description are extracted through natural language processing, and spatial parameters are dynamically corrected by combining the electronic map coordinates of the potential hazard location and the scene type (indoor / outdoor) so that relative orientations (such as "adjacent equipment") can be accurately converted into absolute spatial coordinates.

[0108] Based on the YOLOv8 object detection algorithm, a heatmap of the target object is generated, and the heatmap region is bound to the absolute spatial coordinates. For irregular objects (such as pipes and walls), the fan-shaped detection area is optimized into a "rectangular detection band along the object's axis", and the angle range adapts to the object's orientation (such as the detection band of a horizontal pipe extending ±20° in the horizontal direction). This more accurately defines the possible abnormal areas of the target object and reduces missed detections and false detections caused by the special shape of the object.

[0109] Calculate the spatial overlap rate between the pixel region of the target object in the photo and the heatmap, and introduce key feature weights;

[0110] Spatial overlap rate (SC) is calculated using a weighted pixel ratio method, as shown in the following formula:

[0111] SC=(Σ(q_i×O_i)) / (Σ(q_i×M_i))×100%

[0112] Where: O_i is the number of overlapping pixels of the i-th key feature region in the heatmap (key features include core anomalies such as crack endpoints, corrosion centers, and equipment nameplates, etc., which are mandatory); M_i is the total number of pixels in the i-th key feature region;

[0113] q_i is the weight value of the i-th key feature (for example, the weight of core anomaly points is set to 0.6-0.8, and the weight of equipment nameplates and other markings is set to 0.2-0.4. It can be dynamically adjusted according to the type of hidden danger. For example, in electrical hidden dangers, the weight of "abnormal wiring terminals" is higher than that of "scratches on the casing").

[0114] When the overlap rate is less than the first overlap rate threshold, it is judged as a partial contradiction in the abnormal direction. It is not directly marked as a serious conflict, but a warning for insufficient view coverage and a reshoot guide are generated. If the overlap rate is less than the second overlap rate threshold, a serious conflict mark for complete contradiction in the abnormal direction is triggered. For example, if the overlap rate between the core abnormal area of ​​the target object in the photo (such as the crack endpoint or corrosion center) and the detection area is ≥50%, the overall overlap rate threshold is relaxed to 60%. In other cases, 70% is used as the threshold. When the overlap rate is lower than the corresponding threshold, it is judged as a partial contradiction in the abnormal direction. It is not directly marked as a serious conflict, but a warning for insufficient view coverage and a reshoot guide are generated (such as asking for a close-up shot within 1 meter to the left of the device). If the overlap rate is <30% (complete deviation), a serious conflict mark for "complete contradiction in the abnormal direction" is triggered. It strictly controls the matching accuracy of key features while avoiding over-judgment due to deviations in non-core areas, making the conflict classification more in line with actual risks and reducing the situation of misjudging serious conflicts or missing minor deviations. It generates targeted reshoot guidance for different overlap rates (such as clearly indicating the shooting range and angle) to help users quickly understand the problem and accurately supplement information, avoiding repeated uploads caused by vague reshoot requirements.

[0115] Among them, the first overlap rate threshold and the second overlap rate threshold are determined based on scenario classification, risk coefficient, and misjudgment rate; the risk coefficient is determined by the accident occurrence rate of similar hidden dangers;

[0116] The first overlap rate threshold T1 = T1_base × (1 + α × S_type + β × F_risk); where: T1_base (base value): when there is a core abnormal region, T1_base = 60%; when it contains only device identifiers, T1_base = 70%;

[0117] S_type (Scene Correction Coefficient): Small objects (valves, etc.): S_type = -0.05 (corresponding to a 5% decrease); High-risk scenarios (flammable and explosive areas): S_type = +0.1 (corresponding to a 10% increase); Other scenarios: S_type = 0;

[0118] F_risk (risk coefficient): Calculates the accident incidence rate P for similar hazards based on historical rectification data.

[0119] When P < 1%, F_risk = -0.03 (appropriately relaxed); when 1% ≤ P ≤ 5%, F_risk = 0 (maintain the baseline); when P > 5%, F_risk = +0.03 (strictly tightened); α and β are weighting coefficients: α = 0.7, β = 0.3 (prioritize scenario adaptability);

[0120] The first overlap rate threshold T2 algorithm includes:

[0121] T2 = T2_base × (1 + γ × C_error)

[0122] Where: T2_base is the baseline value, verified through historical data as the critical value for complete deviation, for example, 30%; C_error (misjudgment correction coefficient): Statistical analysis of the T2 misjudgment rate Er (number of misjudged cases / total number of judgments) in manual review over the past 3 months.

[0123] When Er < 5%, C_error = 0 (maintain baseline);

[0124] When 5% ≤ Er ≤ 10%, C_error = +0.05 (increased to 31.5% to reduce missed detections);

[0125] When Er > 10%, C_error = -0.05 (reduced to 28.5% to reduce false positives);

[0126] γ is the adjustment weight: γ = 1 (adjustment based on direct correlation with misjudgment rate);

[0127] The algorithm constraints include:

[0128] The T1 calculation result should be limited to the range of 50%-80% (to avoid extreme values);

[0129] The T2 calculation result should be limited to the range of 25%-35% (to ensure the stability of the judgment);

[0130] The system automatically incorporates the latest data (scenario classification, risk coefficient, and misjudgment rate) for iterative calculations each quarter.

[0131] In one possible implementation, the scenario logic verification, which combines contextual information about the location of the potential hazard and the type of equipment, and marks features that violate scenario constraints as scenario logic conflicts, includes:

[0132] For progressive damage, a historical hazard photo library of the same location and equipment is accessed, and historical damage features (such as corrosion area, crack length and other quantitative parameters) are extracted using image segmentation algorithms (such as U-Net) to establish a historical feature sequence for the hazard.

[0133] Based on historical damage feature sequences, calculate the historical rate of change (e.g., weekly average growth rate) of damage area (or length) to preliminarily determine the expected value of damage area at the current time point; obtain the single-time rate of change by using the difference between the damage area or length recorded in the i-th and i-1-th timestamps, and the timestamps of the i-th and i-1-th timestamps; determine the historical average rate of change by using multiple single-time rates of change obtained from the feature sequence; and obtain the preliminary expected value based on the time interval between the current time and the most recent record, the damage area of ​​the most recent record, and the average rate of change.

[0134] The system automatically retrieves historical environmental data from the same location (such as temperature, humidity, and concentration of corrosive media, and links it to the environmental monitoring system), and calculates the weight of each environmental factor on the damage through a multiple linear regression model (e.g., for every 10% increase in humidity, the rate of increase in corrosion area increases by 15%).

[0135] By using the aforementioned weighted correction of the damage area change curve, the current expected value for environmental adaptation is obtained; let the set of environmental factors (such as temperature and humidity) be E = [e1, e2, ..., e m The corresponding historical influence weights are calculated using a multiple linear regression model as W = [w1, w2, ..., w]. m ];w1+w2+...+w m =1; the deviation of the current environmental factor from the historical mean is ΔE. j ;

[0136]

[0137] The environmental correction factor K is:

[0138]

[0139] Historical average environmental factors; Current environmental factors;

[0140] The current expected value is determined based on the preliminary expected value and the environmental adjustment factor. For example, the current expected value is the product of the preliminary expected value and the environmental adjustment factor.

[0141] The corrected historical rate of change is input into the LSTM time series prediction model to generate a predicted trend of damage development; at the same time, damage features in the current photo are extracted and their deviation δ from the "expected value after environmental correction" is calculated.

[0142] If δ < deviation threshold (e.g., 30%), the current damage state is determined to meet the time dimension scenario constraints;

[0143] If δ ≥ the deviation threshold (e.g., the crack actually shrinks), it is determined to be a scene logic conflict, triggering a "state abnormality" review prompt, and automatically marking the areas in the photo that differ from historical features (e.g., the crack shortens unexpectedly, the corrosion area shrinks abnormally).

[0144] The above technical solution achieves the following results: For progressive damage, by accessing a historical database of potential hazards and extracting quantitative features, combined with environmental data to correct expected values, it is possible to scientifically determine whether the current damage conforms to the development pattern over time. When the actual damage deviates significantly from the expected value adapted to the environment (such as unexpected shrinkage of cracks or abnormal reduction in corrosion area), it is marked as a scene logic conflict. This effectively identifies abnormal situations such as data tampering and false reports of damage recovery, avoiding misjudgments of potential hazards due to subjective judgment biases and ensuring the objectivity of damage status assessment.

[0145] By generating damage development trends through an LSTM time series prediction model and combining historical change rates with the weights of environmental factors, the evolution path of damage can be predicted in advance (e.g., the corrosion area may increase by 20% in the next two weeks). This function breaks through the limitations of traditional judgments based solely on the current state, enabling managers to proactively formulate prevention and control measures, avoid accidents caused by sudden deterioration of damage, and improve the initiative in risk prevention and control.

[0146] When an anomaly verification prompt is triggered, the system automatically marks the areas in the photo that differ from historical features (such as shortened crack sections or reduced corrosion areas) with bounding boxes. This eliminates the need for manual verification to compare historical data one by one, allowing the system to directly focus on the abnormal areas and significantly reduce verification time and workload. Simultaneously, the quantified deviation index provides a clear basis for verification judgment, avoiding subjective differences in verification results and improving process efficiency.

[0147] By calculating the influence weights of environmental factors on damage (such as the specific impact ratios of humidity and corrosive medium concentration) using a multiple linear regression model, key driving factors of damage development can be revealed. For example, once it is determined that "for every 10% increase in humidity, the corrosion rate increases by 15%", environmental control measures in the area can be adjusted accordingly (such as adding dehumidification equipment) to slow down damage development at the source and improve the precision of prevention and control measures.

[0148] Information such as historical damage feature sequences, environmental data, and deviation judgment results will be continuously accumulated into system data assets, providing samples for model iteration. By analyzing high-frequency scene logical conflict cases, the feature extraction accuracy of image segmentation algorithms, the prediction accuracy of LSTM models, and the calculation method of environmental factor weights can be continuously optimized, enabling the system to continuously improve its performance as the application deepens, forming a virtuous cycle of "data accumulation - model optimization - accurate judgment".

[0149] In one possible implementation, the multi-hazard association module includes a multi-dimensional association engine, which, based on spatiotemporal overlap analysis, achieves hazard association through the following extended dimensions:

[0150] Causal association dimension: Configure a hidden danger knowledge graph, which has a pre-set causal relationship network between hidden danger types (such as the mapping relationship between equipment aging and functional failure). By searching the graph, we can mine the causal association strength between newly reported hidden dangers and historical hidden danger records. When the association strength is greater than or equal to the association threshold, even if the spatiotemporal overlap is less than the threshold, the association between the newly reported hidden danger and the historical hidden danger record is still established.

[0151] Impact Range Association Dimension: Based on the type, severity, and surrounding environmental parameters of the hazard, the potential impact range of a single hazard is automatically calculated. When the spatial range of a newly reported hazard overlaps with or contains the impact range of a historical hazard, the association is triggered and the overlapping area is marked. Specifically, based on the hazard type (e.g., fire, leak), severity (e.g., emergency, urgent), and surrounding environmental parameters (e.g., wind speed, population density), the potential impact range of both is calculated using a preset model (e.g., the impact range of a fire hazard is calculated based on "burning radius + smoke diffusion distance").

[0152] Calculate the ratio of the overlapping area of ​​the affected area to the minimum affected area of ​​the two (e.g., the affected area of ​​the historical hazard is 80㎡, the new hazard is 60㎡, and the overlapping area is 40㎡, then the affected area correlation score is 40 / 60≈0.67).

[0153] If an inclusion relationship exists (such as the impact range of a new hidden danger completely encompassing the impact range of a historical hidden danger), the impact range correlation score is directly set to 1.0.

[0154] The multi-dimensional association engine employs dynamic weighted association logic, including:

[0155] Adjustable weight parameters are configured for the dimensions of spatiotemporal overlap, causal relationship, and scope of influence. The weight values ​​are dynamically adapted according to the type of hazard (e.g., the weight of the spatiotemporal overlap dimension is increased for emergency hazards, and the weight of the causal relationship dimension is increased for progressive hazards).

[0156] The association level is determined by a comprehensive association score (the sum of the products of the weights of each dimension and the association strength, i.e., the weighted average of the association strength of each association dimension). When the score is greater than or equal to the score threshold, it is determined to be a strong association and included in the composite risk analysis. The association strength of the association dimension ranges from [0,1].

[0157] The mechanism by which the multi-hazard association module generates a composite risk report includes:

[0158] Multi-dimensional risk assessment integration: The report integrates the risk level of related hidden dangers themselves, the risk level of diffusion based on the relationship, and the level of historical difficulty in handling them to form a comprehensive risk assessment matrix;

[0159] The interactive interface displays the spatial distribution of related hazards, the network of relationships, and the risk transmission path, allowing users to view the simulation effects of different treatment plans through visual operations.

[0160] The upgraded early warning mechanism triggered by the multi-hazard association module includes:

[0161] Predictive early warning: Configure a machine learning prediction model to generate a prediction of the development trend of complex risks based on the feature sequence of related hidden dangers and historical data. When the prediction result reaches the preset early warning threshold, trigger an upgraded early warning in advance.

[0162] Tiered and refined early warning: Based on the comprehensive risk level of related hidden dangers, early warnings are divided into multiple levels (such as extremely high / high / medium / low), and different levels correspond to different handling procedures (such as automatically activating the emergency response plan for extremely high-level early warnings, and limiting the time threshold for feedback on handling plans for high-level early warnings).

[0163] By analyzing the correlations across three dimensions—spatial-temporal, causal, and scope of impact—this approach overcomes the limitations of a single dimension and can capture complex relationships such as those that are spatially and temporally dispersed but have causal chains, or those that have overlapping scopes of impact but no direct spatiotemporal correlation. Dynamic weighting makes the analysis more closely aligned with the characteristics of potential hazards (e.g., prioritizing spatiotemporal correlation for urgent hazards), while the comprehensive correlation scoring formula quantifies the strength of the correlation, avoiding subjective judgment bias and improving the accuracy of identifying strongly correlated hazards.

[0164] The composite risk report integrates its own risks, diffusion risks, and the difficulty of handling them, forming a comprehensive risk assessment matrix that can fully reflect the overall risk level of related hidden dangers. The interactive interface visualizes the spatial distribution, related networks, and risk transmission paths, enabling users to intuitively understand the formation and diffusion logic of risks. The function of simulating the effects of handling plans provides a pre-decision tool for decision-making, reducing the probability of decision-making errors.

[0165] Predictive early warning, based on machine learning models, captures risk development trends in advance, allowing for more time to respond compared to traditional post-event warnings. Tiered and detailed early warnings match differentiated processes according to risk levels. For example, extremely high-level warnings can directly activate emergency plans, avoiding resource waste or response delays caused by a one-size-fits-all approach. This ensures that high-risk hazards are prioritized and handled efficiently, while low-risk hazards are handled according to reasonable procedures, thus improving the overall efficiency of risk management.

[0166] By uncovering the relationships between potential hazards, the multi-hazard correlation module transforms isolated hazard records into a systemic risk network, helping managers shift from addressing individual hazards to overall risk management. For example, after identifying the correlation chain of pipeline corrosion, abnormal pressure, and leakage risks, preventative measures covering the entire chain can be developed, rather than just targeting a single link for partial rectification, fundamentally reducing the probability of complex risks occurring.

[0167] In one possible implementation, the early warning module includes a multi-level early warning determination unit, specifically configured as follows:

[0168] A three-dimensional early warning scoring model is established based on the inherent risk level of hazard classification, the comprehensive correlation score output by the multi-hazard correlation module, the rectification rate and accident occurrence rate of similar hazards in historical data;

[0169] In the aforementioned three-dimensional early warning scoring model, inherent risk level, comprehensive correlation score, historical rectification rate, and accident occurrence rate are each assigned dynamic weights, and the weight values ​​are dynamically calibrated according to industry type; for example, the inherent risk level accounts for 40%-60% of the weight, the comprehensive correlation score accounts for 20%-30% of the weight, and the historical rectification rate and accident occurrence rate account for 20%-20% of the weight; the scores of each dimension are weighted and averaged to obtain the three-dimensional scoring result;

[0170] Based on the three-dimensional scoring results, the warnings are divided into multiple levels (such as urgent, important, general, and alert). For example, when the three-dimensional score is ≥80 points, it is judged as a Level 1 warning (urgent), 60-79 points as a Level 2 warning (important), 40-59 points as a Level 3 warning (general), and <40 points as a Level 4 warning (alert). Different levels of warnings correspond to different response time limits and processing priorities.

[0171] The early warning module also includes a dynamic early warning update unit, configured as follows:

[0172] The rectification progress data of the closed-loop rectification module is synchronized in real time. When the hidden danger enters the rectification process, the warning level is automatically reduced by one level. When the rectification is accepted and archived, the warning is terminated.

[0173] If rectification is not completed within the time limit or new associated hazards are added to the multi-hazard module, the early warning level upgrade mechanism will be triggered. The upgrade level will be determined based on the length of the overdue period or the risk level of the newly added associated hazards (up to the highest level of early warning).

[0174] By connecting to the external environmental monitoring system, when environmental parameters (such as rainfall, temperature, and wind force) at the potential hazard location exceed the safety threshold, the warning level is dynamically adjusted upward based on the environmental impact factor model in the knowledge base module.

[0175] In one possible implementation, the multi-hazard association module further includes a duplicate reporting identification engine for determining and classifying duplicate reports, including:

[0176] The time threshold is dynamically set according to the urgency of the hidden danger. When the newly reported record and the historical record meet the preset spatial range and time difference requirements, the initial association is triggered.

[0177] Specifically:

[0178] The spatial overlap between newly reported hazards and historical hazards is calculated to determine whether it meets a preset standard. The core indicator is spatial matching degree (calculated based on IoU intersection-union ratio).

[0179] When the spatial matching degree is greater than or equal to the spatial judgment threshold (default 0.5, which can be adjusted according to the type of hazard), it is judged as spatial overlap. For example, if the spatial overlap rate between a newly reported equipment corrosion hazard and a corrosion record at the same location 3 days ago reaches 60%, it meets the spatial range requirement.

[0180] For linear objects (such as pipes and cables), the spatial range determination is optimized to "axial distance + vertical coverage width". If the axial distance between the newly reported location and the historical record is ≤1 meter (indoor) or 3 meters (outdoor), and the vertical coverage width overlap rate is ≥50%, it is determined that the spatial range is in compliance.

[0181] The time interval between newly reported records and historical records must fall within the dynamically calculated time threshold range (i.e., 0 < time difference ≤ time threshold):

[0182] Time difference = New report time - Historical record creation time (take the most recent related historical record);

[0183] For example, the time threshold for emergency hazards is calculated to be 15 hours. If the time interval between a newly reported record and a historical record is 12 hours (≤15 hours), then the time difference requirement is met; if the interval is 20 hours (>15 hours), then it is not met.

[0184] When the spatial matching degree is greater than or equal to the spatial judgment threshold and the time difference is less than or equal to the time threshold, the system automatically triggers the initial association and proceeds to the next step of feature and semantic comparison.

[0185] Wherein, the time threshold T = T0 × (1 - a1 × Ep) × (1 + a2 × H)

[0186] T0 is the base time threshold (e.g., 24 hours, set according to industry default values ​​or the average of the time from the first report to the start of rectification for multiple such events);

[0187] Ep is the urgency coefficient of the hazard (value 0-1, urgent hazard E=0.8, general hazard E=0.3). The higher the urgency, the greater the compression ratio of the time threshold.

[0188] a1 represents the urgency level impact weight (a fixed value of 0.5, ensuring a significant reduction in T for urgent potential hazards);

[0189] H represents the historical frequency of repeated reporting (the ratio of the average repeated reporting interval of this type of hidden danger in the past 3 months to the standard handling cycle, with a value of 0-1). The more frequent the repeated reporting, the more relaxed the threshold should be.

[0190] a2 is the frequency adjustment coefficient (fixed value of 0.3, to avoid excessive relaxation);

[0191] For records that are initially associated, the consistency between core damage features and textual descriptions is compared through image feature extraction and semantic analysis.

[0192] When the combined matching degree of space, features and semantics reaches the matching degree threshold, it is judged as a completely duplicate report; when the core features are the same but there are non-critical differences, it is judged as an approximately duplicate report.

[0193] For completely duplicate reports, the review process for new reports will be automatically terminated, a related prompt will be sent to the reporter, and the new report information will be merged into the history record and supplemented as time-series analysis data.

[0194] For near-duplicate reports that trigger the intelligent audit unit's difference verification, the feature difference area is marked and a confirmation guide is generated; based on the reporter's feedback, if it is confirmed to be a duplicate description, it is merged; if it is confirmed to be a new feature, the original record is upgraded to a composite risk and the association is retained.

[0195] Record the false positive rate of repeated judgments, and dynamically optimize the spatiotemporal judgment parameters through machine learning models for the types of hidden dangers with excessive false positive rates;

[0196] We analyzed the adoption rate of user feedback on the difference confirmation guidelines, and optimized the wording and supplementary information for guidelines with adoption rates below the threshold.

[0197] The effects of the above technical solution are as follows:

[0198] By employing a dynamic time threshold algorithm and multi-dimensional spatial judgment rules, precise matching of spatiotemporal conditions is achieved. The time threshold is dynamically adjusted based on the urgency coefficient of the hazard and the frequency of historical repeated reporting, avoiding the accumulation of duplicate reports; the time window for general hazards is more flexible, reducing the probability of misjudging new reports.

[0199] Completely duplicate reports automatically terminate the review process and merge records, reducing the workload of invalid reviews; near-duplicate reports trigger intelligent difference verification, which marks the difference areas through image feature extraction and semantic analysis, guiding the reporter to confirm, thus avoiding the omission of new features and ensuring the effective integration of similar potential risks.

[0200] For potential risks with excessive false positive rates, the machine learning model automatically adjusts the spatiotemporal judgment parameters (such as spatial threshold and time threshold coefficient) to ensure that the judgment rules continuously adapt to the actual scenario. For confirmation guidelines with an adoption rate below the threshold, the expression format (such as adding graphic comparison examples) and auxiliary information (such as historical similar cases) are optimized to improve user confirmation efficiency.

[0201] Merged records of completely duplicate reports form a complete time-series data chain, which facilitates the analysis of the development trend of hidden dangers (such as the weekly average growth rate of corrosion area); new features confirmed in near-duplicate reports upgrade the original records to composite hidden dangers, providing a more comprehensive feature dimension for risk assessment.

[0202] Accurate duplicate identification reduces redundant information flow and saves reviewers time from repetitive work; the hierarchical processing mechanism ensures that duplicate reports of urgent hazards are quickly filtered out, avoiding the occupation of emergency resources; and the strict verification of near-duplicates prevents subtle changes from being overlooked (such as an abnormal increase in the corrosion depth of chemical pipelines), reducing the risk omission rate.

[0203] In one possible implementation, the early warning module includes an early warning information generation unit, configured as follows:

[0204] For Level 1 and Level 2 warnings, an early warning report is automatically generated, which includes the core features of the hidden danger (image annotation of the difference area, text description of key information), a list of related hidden dangers, historical similar cases, and pre-control countermeasures matched with the knowledge base.

[0205] The report includes an embedded electronic map location link and contact information for the responsible unit / personnel, along with a countdown reminder (based on response time limit);

[0206] For Level 3 and Level 4 warnings, a concise warning notification is generated, which includes a summary of the hazard, preliminary handling suggestions, and a link to view details.

[0207] The early warning module also includes a precise push unit, configured as follows:

[0208] Based on the multi-level positioning information of the classification and positioning units, and matched with the preset responsibility area division table, the early warning information is pushed to the directly responsible person in the corresponding area.

[0209] If no confirmation reply is received from the responsible person within the first preset time (within 2 hours) (Level 1 warning) or no confirmation is received within the second preset time (e.g., within 4 hours) (Level 2 warning), a supervision warning will be automatically sent to the superior management department and the safety department.

[0210] By combining the user's historical processing records with their expertise tags (such as electrical hazards, fire hazards), warning information for cross-regional or complex hazards is additionally pushed to personnel with corresponding processing experience.

[0211] The early warning module includes an early warning effectiveness evaluation unit, configured as follows:

[0212] Record the response time, rectification completion rate, and hazard recurrence rate for each warning, and conduct correlation analysis with the warning level and the matching degree of the target audience;

[0213] For entities that fail to respond to warnings in a timely manner for more than three consecutive times, a capability assessment report will be generated and sent to the management department.

[0214] Based on the evaluation data, the weight parameters and response time of the three-dimensional early warning scoring model are dynamically optimized. For hazard types with a historical false alarm rate (confirmed as non-hazard after early warning) exceeding 5%, the intelligent review unit is linked to strengthen the pre-verification.

[0215] The multi-level early warning judgment unit integrates the inherent risk level of the hidden danger classification, the comprehensive correlation score, the historical rectification rate and the accident occurrence rate through a three-dimensional early warning scoring model. It combines dynamic weights (calibrated according to industry type) for weighted calculation, so that the judgment of the early warning level no longer depends on a single indicator, avoids the bias of one-size-fits-all judgment, and makes the early warning level more in line with the actual risk situation, providing accurate guidance for subsequent handling.

[0216] The dynamic early warning update unit dynamically adjusts the early warning level by synchronizing rectification progress, associating with new hidden dangers, and monitoring changes in environmental parameters in real time. The warning level automatically downgrades when a hidden danger enters rectification phase and automatically upgrades when rectification exceeds the deadline or when new, more significant hidden dangers are associated with it, ensuring that the early warning status is synchronized with changes in risk. For example, if the rectification of a hidden danger exceeds the deadline by 3 days and is associated with a new high-risk hidden danger, the early warning level directly upgrades from "general" to "emergency," forcibly accelerating the handling process. This dynamic mechanism breaks the static state after the early warning is issued, allowing the response pace to closely follow the dynamics of risk and reducing the risk of loss of control due to information lag.

[0217] The person directly responsible receives the alert first, and if the alert is not confirmed in a timely manner, it is automatically forwarded to the superior for supervision. For cross-regional or complex hazards, the alert is also pushed to experienced personnel. For example, electrical hazard warnings are not only pushed to the person in charge of the area, but also simultaneously pushed to technical personnel with electrical expertise, ensuring that professional forces can intervene quickly, reducing information redundancy, and significantly improving the timeliness of the warning response (e.g., the confirmation rate of a level 1 warning increases to over 90% within 2 hours).

[0218] Level 1 and Level 2 warnings include core characteristics of potential hazards, a list of related issues, historical cases, and preventative measures. They also include location links, contact information, and countdown reminders, providing one-stop information support for emergency response. Level 3 and Level 4 warnings, with their simplified notifications, focus on core information, reducing the information load in non-emergency scenarios. This tiered reporting design ensures the completeness of information in high-risk warnings while improving the reading efficiency of low-risk warnings, helping relevant personnel quickly grasp the key points for handling.

[0219] Capability assessment reports are generated for responsible parties who fail to respond promptly. For hazard types with false alarm rates exceeding 5%, pre-emptive verification is strengthened, and the weights and response times of the 3D model are dynamically adjusted. For example, if a historical data indicator in a certain industry is found to have excessively high weights leading to misjudgments, its weight can be lowered to a reasonable range. Regarding the high false alarm rate for electrical hazards, the intelligent review unit is linked to add a photo compliance verification dimension. This continuous optimization capability enables the early warning system to constantly adapt to real-world scenarios, gradually reducing false alarm and missed alarm rates, and improving overall risk management effectiveness.

[0220] This invention is described from the perspectives of purpose, effectiveness, progress, and novelty, and meets the functional enhancement and use requirements emphasized by the Patent Law. The above description and accompanying drawings are only preferred embodiments of this invention and are not intended to limit the invention. Therefore, all structures, devices, features, etc., that are similar to or identical to those of this invention, i.e., all equivalent substitutions or modifications made in accordance with the scope of this patent application, should fall within the scope of protection of this patent application.

Claims

1. A hidden danger reporting and investigation system, characterized in that, The system comprises: a hidden danger reporting module for receiving hidden danger information submitted by a user, including hidden danger photos, text description, hidden danger classification, location, reporting time and reporter information; a multi-hidden danger correlation module for automatically correlating related hidden danger records based on time and space overlap analysis, generating a composite risk report, and triggering an upgrade warning to the warning module; a warning module for hierarchical warning according to hidden danger classification, correlation and historical data; a closed-loop rectification module for rectification task assignment, process feedback, security department acceptance and result archiving; a knowledge base module for storing the mapping relationship between hidden danger types and pre-control countermeasures, supporting fuzzy search and dynamic update of countermeasures.

2. The hidden trouble reporting and checking system according to claim 1, characterized in that, The hidden danger reporting module comprises: a basic recognition unit for receiving at least one hidden danger photo containing a hidden danger subject and surrounding environment, and automatically recording the shooting time and geographical location; a feature description unit for collecting text description containing hidden danger specific performance and potential risk, and correlating knowledge base keyword prompts, and intelligently recommending hidden danger classification based on text description content; a classification positioning unit for providing standardized multi-level hidden danger classification options and multi-level positioning associated with an electronic map; an intelligent audit unit for interfacing information collected by the above units, detecting and identifying the compliance and clarity of hidden danger photos, and shielding problems, and triggering a re-upload request for photos that do not meet the requirements; while analyzing the content of the text description, matching the preset keyword library, and generating an artificial review prompt for the submitter of contradictory or misclassified descriptions. 3.The hidden trouble reporting and checking system according to claim 1, characterized in that, The intelligent audit unit comprises: a cross-validation sub-unit for identifying object types and abnormal states in hidden danger photos; comparing hidden danger features in text description with image recognition results, and generating conflict markers and directional review prompts for contradictory features; a self-learning optimization sub-unit for automatically expanding the synonym library based on false positive samples of artificial review, recording user descriptions with a deviation frequency greater than a frequency threshold, and pushing customized reporting guidelines to users with continuous misclassification exceeding a number threshold; generating semantic completion prompts for ambiguous descriptions.

4. The hidden trouble reporting and checking system according to claim 3, characterized in that, The execution steps of the cross-validation sub-unit include: extracting core entities and core abnormalities from text description, and rigidly matching them with image recognition results, and marking as a serious conflict when core entities are missing or abnormal directions are completely contradictory; expanding matching for ambiguous descriptions or industry terms through the semantic association network of the knowledge base; performing scene logic verification in combination with context information such as hidden danger location and equipment type, and marking features that violate scene constraints as scene logic conflicts.

5. The hidden trouble reporting and checking system according to claim 4, characterized in that, The extraction of core entities and core abnormalities from text description and the rigid matching with image recognition results include: using natural language processing to extract orientation words related to core abnormalities in text description, and dynamically correcting spatial parameters in combination with electronic map coordinates and location type of hidden danger location; generating a heat map of target objects based on target detection algorithms, and binding heat map regions with absolute spatial coordinates; for irregular objects, optimizing the fan-shaped detection area into a rectangular detection area along the object axis, and adapting the angle range to the object direction; calculating the spatial overlap rate of the pixel area of the target object in the photo and the heat map, and introducing key feature weights; When the overlap rate is less than the first overlap rate threshold, it is determined that the abnormal direction part is contradictory, not directly marked as a serious conflict, but a view coverage warning and a retake instruction are generated; if the overlap rate is less than the second overlap rate threshold, a serious conflict mark of abnormal direction complete contradiction is triggered; wherein the first overlap rate threshold and the second overlap rate threshold are determined according to scene classification, risk coefficient and misjudgment rate.

6. The hidden trouble reporting and checking system according to claim 4, characterized in that, The scene logic verification is performed according to the context information of the hidden danger site and the equipment type, and a feature that violates the scene constraint is marked as a scene logic conflict, including: For progressive hidden dangers, a historical hidden danger photo library of the same site and equipment is called, historical damage features are extracted through an image segmentation algorithm, and a historical feature sequence of the hidden danger is established; Based on the historical damage feature sequence, a historical change rate of the damage area is calculated, and an expected value of the damage area at the current time node is preliminarily determined; The historical environmental data of the same site are automatically called, the influence weight of each environmental factor on the damage is calculated through a multivariate linear regression model, the damage area change curve is corrected using the weight, and the current expected value adapted to the environment is obtained; The corrected historical change rate is input into an LSTM time sequence prediction model to generate a prediction trend of the damage development; meanwhile, damage features in the current photo are extracted, and a deviation degree of the damage features from the expected value corrected according to the environment is calculated; If the deviation degree is less than a deviation degree threshold, it is determined that the time dimension scene constraint is met; If the deviation degree is greater than or equal to the deviation degree threshold, it is marked as a scene logic conflict, a state abnormality review prompt is triggered, and a difference region between the historical features and the photo is automatically marked with a frame.

7. The hidden trouble reporting and checking system according to claim 1, characterized in that, The multi-hidden danger correlation module includes a multi-dimensional correlation engine, which, on the basis of spatio-temporal overlap analysis, realizes hidden danger correlation through the following extended dimensions: The causal correlation strength between a newly reported hidden danger and a historical hidden danger record is mined through hidden danger knowledge graph retrieval, and when the correlation strength is greater than or equal to a correlation threshold, an association between the newly reported hidden danger and the historical hidden danger record is established; the hidden danger knowledge graph is preconfigured with a causal relationship network between hidden danger types; Based on hidden danger types, severity and surrounding environmental parameters, the potential influence range of a single hidden danger is automatically calculated, and when the spatial range of a newly reported hidden danger overlaps or contains the influence range of a historical hidden danger, correlation is triggered and the overlapping area is marked.

8. The hidden trouble reporting and checking system according to claim 7, characterized in that, The multi-dimensional correlation engine adopts a dynamic weight correlation logic, including: Adjustable weight parameters are configured for spatio-temporal overlap, causal correlation and influence range correlation dimensions, and the weight values are dynamically adapted according to the hidden danger types; The correlation level is determined by the comprehensive correlation score (the sum of the products of the weight of each dimension and the correlation strength), and when the comprehensive correlation score is greater than or equal to a score threshold, it is determined as strong correlation and included in the compound risk analysis.

9. The hidden trouble reporting and checking system according to claim 1, characterized in that, The multi-hidden danger correlation module further includes a repeated reporting recognition engine for repeated reporting determination and hierarchical processing, including: A time threshold is dynamically set according to the hidden danger urgency, and when a newly reported record and a historical record meet the preconfigured spatial range and time difference requirements, preliminary correlation is triggered; For the records of preliminary correlation, the consistency of the core damage features and the textual description is compared through image feature extraction and semantic analysis. When the comprehensive matching degree of space, feature and semantics reaches the preset threshold, it is determined as complete duplicate reporting; when the core features are consistent but there are non-critical differences, it is determined as approximate duplicate reporting; For complete duplicate reporting, the review process of new reporting is automatically terminated, the associated prompt is pushed to the reporter, and the new reporting information is merged into the historical record to supplement the time sequence analysis data; For approximate duplicate reporting, the difference checking of the intelligent review unit is triggered, the feature difference area is marked and the confirmation guide is generated; according to the feedback of the reporter, if it is confirmed as duplicate description, the merging is performed, and if it is confirmed as new feature, the original record is upgraded to a composite hazard and the association is retained. The misjudgment rate of record duplicate determination is recorded, and for the hazard types with misjudgment rate exceeding the standard, the spatio-temporal determination parameters are dynamically optimized through the machine learning model; The adoption rate of the user to the difference confirmation guide is counted, and for the guide content with an adoption rate less than the adoption rate threshold, the expression form and auxiliary information are optimized.

10. The hidden trouble reporting and checking system according to claim 7, characterized in that, The early warning module comprises: A multi-level early warning unit is configured to establish a three-dimensional early warning scoring model based on inherent risk levels of hazard classification, comprehensive correlation scores output by the multi-hazard correlation module, historical data of the same type of hazard rectification rate and accident occurrence rate; in the three-dimensional early warning scoring model, the inherent risk level, the comprehensive correlation score, the historical rectification rate and the accident occurrence rate are respectively configured with dynamic weights, and the weight values are dynamically calibrated according to the industry type; according to the three-dimensional scoring results, the early warning is divided into multiple levels, and different levels of early warning correspond to different response time limits and processing priorities. An early warning updating unit is configured to synchronize the rectification progress data of the closed-loop rectification module in real time, automatically reduce the early warning level when the hazard enters the rectification state, terminate the early warning when the rectification acceptance is passed and the archiving is completed, trigger the early warning level promotion mechanism if the rectification is overdue or the multi-hazard correlation module adds a strongly correlated hazard, and determine the promotion range according to the overdue time or the risk level of the newly added correlated hazard; connect to an external environment monitoring system, and when the environmental parameters of the hazard location exceed the safety threshold, dynamically increase the early warning level based on the environmental influence factor model of the knowledge base module.

Images