Dynamic risk management and control method and system based on papermaking production safety management

Abstract

This application relates to a dynamic risk control method and system based on paper production safety management, including the following steps: acquiring multi-source datasets and constructing a feature vector sequence; inputting the feature vector sequence into a risk perception model to obtain preliminary judgment results; performing logical consistency verification on the preliminary judgment results and generating intermediate judgment results; determining the final judgment result, and executing risk processing and learning optimization processes based on the final judgment result. In summary, this application solves the problems of isolated multi-source information, black box decision-making, lack of confidence assessment, and closed-loop optimization in existing technologies by acquiring multi-source data, constructing feature vector sequences, inputting them into a risk perception model to output preliminary judgment results associated with uncertainty quantification scores, performing logical consistency verification to generate intermediate judgment results, triggering a collaborative judgment mechanism based on uncertainty score comparison to determine the final judgment result, and executing risk processing and learning optimization processes.

Description

Technical Field

[0001] This application relates to the technical field of safety production management, and in particular to a dynamic risk control method and system based on paper production safety management. Background Technology

[0002] In the field of safety production management in the paper industry, real-time monitoring and early warning of potential risks during production are core aspects of ensuring the safety of personnel and equipment. Currently, the risk control technologies commonly used in the industry are mainly divided into two categories. Specifically, the first type relies on discrete sensors and monitoring systems. This involves deploying vibration and temperature sensors to collect equipment operating parameters, using video cameras to monitor the work site, and combining this with environmental sensors to obtain information such as temperature, humidity, and gas concentration. However, this type of technology isolates multi-source information such as equipment status data, video image data, work environment data, and historical work data. Its risk assessment is based solely on a static threshold triggering alarm mechanism from a single data source, failing to effectively correlate multi-dimensional information such as equipment anomalies, personnel behavior, and environmental changes. This makes it difficult to identify systemic risks coupled by multiple factors or dynamically evolving, and its early warning results lack accuracy and foresight. The second type of technology, on the other hand, utilizes data... While analytical models learn from historical data to achieve intelligent early warning, existing data analysis models are often limited to training on single data types and lack the ability to deeply integrate and analyze contextual relationships with multi-source heterogeneous data. Furthermore, the decision-making process of existing data analysis models often exhibits "black box" characteristics, and their outputs may violate basic physical principles or industry safety regulations, making reliability difficult to guarantee. In complex and ambiguous scenarios, existing data analysis models lack a quantitative assessment mechanism for the confidence level of their own judgments, failing to automatically initiate more reliable verification processes when uncertainty is high, thus easily leading to false alarms or missed alarms. In addition, most existing technical solutions stop at the risk alarm stage, failing to establish a closed-loop mechanism that effectively feeds risk handling experience and results back to the model.

[0003] The aforementioned deficiencies, combined, have led to problems in current paper production safety risk management, including widespread monitoring blind spots, high false alarm rates, low levels of intelligence, and difficulty in adapting to dynamic changes in the production line. Summary of the Invention

[0004] To address the aforementioned shortcomings, this application provides a dynamic risk management method and system based on paper production safety management.

[0005] The above-mentioned objective of this application is achieved through the following technical solution:

[0006] A dynamic risk management method based on paper production safety management includes the following steps:

[0007] Obtain multi-source datasets of the target production area and construct a sequence of feature vectors based on the multi-source datasets, which include equipment status data, video image data, operating environment data, and historical operation data;

[0008] The feature vector sequence is input into a pre-trained risk perception model to obtain preliminary judgment results including risk type, risk location, risk evolution trend and dynamic risk level, and the preliminary judgment results are associated with uncertainty quantification scores;

[0009] The preliminary judgment results are logically consistent based on a pre-set security knowledge graph, and intermediate judgment results are generated based on the results of the logical consistency check.

[0010] The uncertainty quantification score is compared with a preset score threshold. If the score threshold is not exceeded, the intermediate judgment result is determined as the final judgment result. If the score threshold is exceeded, a preset collaborative judgment mechanism is triggered, and the final judgment result is determined.

[0011] Based on the final judgment result, the corresponding preset risk handling process and learning optimization process will be executed.

[0012] The second objective of this invention is achieved through the following technical solution:

[0013] A dynamic risk management system based on paper production safety management includes:

[0014] The data acquisition module is used to acquire multi-source datasets of the target production area and construct a feature vector sequence based on the multi-source datasets. The multi-source datasets include equipment status data, video image data, operating environment data, and historical operation data.

[0015] The first determination module is used to input the feature vector sequence into the pre-trained risk perception model to obtain a preliminary determination result including risk type, risk location, risk evolution trend and dynamic risk level, and the preliminary determination result is associated with an uncertainty quantification score.

[0016] The second judgment module is used to perform logical consistency verification on the preliminary judgment result based on the preset security knowledge graph, and generate intermediate judgment result based on the result of the logical consistency verification.

[0017] The third determination module is used to compare the uncertainty quantification score with a preset score threshold. If the score threshold is not exceeded, the intermediate determination result is determined as the final determination result. If the score threshold is exceeded, a preset collaborative determination mechanism is triggered, and the final determination result is determined.

[0018] The process execution module is used to execute the corresponding preset risk handling process and learning optimization process based on the final judgment result.

[0019] This application also relates to a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described dynamic risk control method based on paper production safety management.

[0020] This application also relates to a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the aforementioned dynamic risk control method based on paper production safety management.

[0021] In summary, the dynamic risk control method and system based on paper production safety management provided in this application constructs a feature vector sequence by acquiring multi-source data, inputs it into a risk perception model to output preliminary judgment results and associate them with uncertainty quantification scores, performs logical consistency verification to generate intermediate judgment results, triggers a collaborative judgment mechanism based on uncertainty score comparison to determine the final judgment result, and executes risk processing and learning optimization processes. This solves the problems of isolated multi-source information, black box decision-making, lack of confidence assessment and closed-loop optimization in existing technologies, realizes the integration of dynamic risk identification from multi-source data, provides uncertainty quantification to support intelligent review, ensures logical consistency and provides closed-loop learning optimization, thereby improving the accuracy, reliability and intelligence level of paper production safety management. Attached Figure Description

[0022] Figure 1 This is a flowchart of an embodiment of a dynamic risk control method based on paper production safety management according to this application;

[0023] Figure 2 This is a flowchart of step S20 in an embodiment of a dynamic risk control method based on paper production safety management in this application;

[0024] Figure 3 This is a flowchart of step S22 in an embodiment of a dynamic risk control method based on paper production safety management in this application;

[0025] Figure 4 This is a flowchart of step S23 in an embodiment of a dynamic risk control method based on paper production safety management in this application. Detailed Implementation

[0026] The following is in conjunction with the appendix Figures 1-4 This application will be described in further detail.

[0027] In paper production safety risk management, the processing of multi-source heterogeneous data has many shortcomings. Specifically, data from different modalities are usually collected and analyzed independently, and the risk assessment based on this mainly relies on triggering static thresholds set for a single data source, failing to achieve effective correlation of multi-dimensional information. As a result, existing risk management systems struggle to identify systemic risks formed by the dynamic coupling of multiple factors such as equipment malfunctions, personnel behavior, and environmental changes, making it difficult to provide highly accurate and forward-looking risk predictions. Furthermore, existing data analysis models are limited to a single data type during training, lacking the ability to deeply integrate multi-source heterogeneous data. Their decision-making process is opaque, and the output results may violate physical common sense or safety regulations, and there is a lack of a quantitative assessment mechanism for the confidence level of the judgment results. In addition, risk handling experience and results are not effectively fed back to the data analysis model, lacking a closed-loop optimization mechanism, causing the risk management system to be unable to continuously adapt to dynamic changes in the production line.

[0028] For example, in the drying section of a paper machine in a paper production line, vibration sensors monitor the operating status of the drying cylinder, temperature sensors detect bearing temperature, video cameras monitor operator behavior, and environmental sensors collect temperature, humidity, and gas concentration data. When the drying cylinder experiences slight vibration abnormalities, the vibration sensor triggers an alarm because it exceeds a preset threshold. However, the video image shows that the operator is working normally, and the environmental data is normal. The risk management system may mistakenly interpret this as an equipment malfunction and shut down the machine for inspection. At the same time, if the bearing temperature rises slowly, the operator approaches a high-temperature area, and the gas concentration increases locally, the coupling risks of these multi-dimensional signals are not identified, and the risk management system fails to warn of potential fire hazards. As a result, false alarms lead to unplanned shutdowns, while missed alarms may cause safety accidents. The coverage and reliability of risk monitoring cannot meet the needs of safe production.

[0029] If the above problems are not addressed, the blind spots in risk monitoring will continue to expand, and false alarms and missed alarms may continue to occur. The risk management system will be unable to accurately predict and prevent dynamically evolving risks, thereby increasing the possibility of personnel injury and equipment damage. More seriously, due to the lack of confidence assessment and closed-loop optimization mechanisms, the reliability of the risk management system's decision-making in complex scenarios is difficult to guarantee. Its low level of intelligence makes it difficult to adapt to frequent adjustments in papermaking production processes, ultimately threatening the safe and stable operation of the entire production process.

[0030] In one embodiment, this application discloses a dynamic risk management method based on paper production safety management, such as... Figure 1 As shown, the specific steps include the following:

[0031] S10: Obtain a multi-source dataset of the target production area and construct a feature vector sequence based on the multi-source dataset. The multi-source dataset includes equipment status data, video image data, operating environment data, and historical operation data.

[0032] In this embodiment, the multi-source dataset refers to a collection of various data obtained from different types of sensors, monitoring equipment, information systems, etc. within the paper production area. The multi-source dataset includes equipment operating parameters, video surveillance footage, environmental monitoring indicators, and historical accident records. Equipment status data refers to time-series signal data that reflects the health status of the equipment, collected in real time by various sensors installed on the paper production equipment. Video image data refers to image information collected by visual devices such as monitoring cameras and thermal imagers deployed in the production area. Work environment data refers to structured data reflecting the real-time physical and chemical environmental status of the production site and current work activity management information. Historical work data refers to textual knowledge and experience records related to safe production accumulated by the enterprise in the past. The feature vector sequence refers to a series of numerical representations arranged in chronological or logical order after preprocessing and feature extraction of the multi-source data. Each feature vector in the feature vector sequence represents the production status at a specific point in time or under a specific event.

[0033] Specifically, in practical applications, various sensors and monitoring devices can be deployed to continuously collect data. For example, equipment status data can be obtained through vibration sensors, temperature sensors, and pressure sensors to monitor the operational health of the equipment; video image data can be acquired through high-definition cameras to monitor personnel behavior and material stacking at the work site; work environment data can be obtained through temperature and humidity sensors, gas concentration sensors, etc., to assess the impact of environmental factors on safety; historical work data can be extracted from production management systems or accident reports to provide background knowledge and experience. After acquiring the above multimodal data, it can be initially organized and formatted, for example, by unifying the storage format of different types of data and performing preliminary noise reduction processing; subsequently, the organized data can be converted into a series of numerical representations, i.e., feature vectors. The obtained feature vectors are arranged in chronological order to form a feature vector sequence, which serves as input for subsequent risk analysis.

[0034] S20: Input the feature vector sequence into the pre-trained risk perception model to obtain preliminary judgment results including risk type, risk location, risk evolution trend and dynamic risk level, and the preliminary judgment results are associated with uncertainty quantification scores;

[0035] In this embodiment, the risk perception model refers to a trained computational model that can receive a sequence of feature vectors as input and output a risk assessment of the current production state. Specifically, the risk perception model is used to identify potential risk events, assess their severity, and predict their development trends. The preliminary judgment result refers to the initial risk assessment generated by the risk perception model after analyzing the current production state. The preliminary judgment result includes the risk category, the location of the risk occurrence, the possible evolution direction, and the quantified risk level, and is associated with an uncertainty quantification score. The uncertainty quantification score is a numerical value associated with the preliminary judgment result, used to characterize the confidence or reliability of the preliminary judgment result. The higher the uncertainty quantification score, the lower the confidence in the preliminary judgment result, and vice versa.

[0036] Specifically, the risk perception model can be a machine learning model such as a deep neural network or an ensemble learning model. After receiving a sequence of feature vectors, it analyzes the input data to identify possible abnormal patterns or potential risk signals in the data. For example, the risk perception model may comprehensively judge whether there is a risk by analyzing abnormal fluctuations in equipment vibration data, violations of personnel operation in video images, increases in ambient gas concentration, and the frequency of similar events in historical data. While outputting the risk type (e.g., equipment failure risk, personnel operation risk, environmental pollution risk), risk location (e.g., specific equipment number, work area coordinates), risk evolution trend (e.g., the risk is intensifying or spreading), and dynamic risk level (e.g., low, medium, high risk), the risk perception model also outputs an uncertainty quantification score. This uncertainty quantification score can be a probability value or a confidence index, used to represent the degree of confidence of the risk perception model in the preliminary judgment results. For example, the uncertainty quantification score can be calculated by the prediction variance within the risk perception model or the voting results of different sub-models or module layers in the ensemble risk perception model.

[0037] S30: Perform logical consistency verification on the preliminary judgment result based on the preset security knowledge graph, and generate intermediate judgment result based on the result of the logical consistency verification;

[0038] In this embodiment, the safety knowledge graph refers to a pre-set structured knowledge base that contains information such as safety regulations, equipment failure modes, risk propagation paths, and emergency response measures in the papermaking industry. The safety knowledge graph is organized in the form of nodes and edges to support the logical verification and reasoning of risk assessment. Logical consistency verification refers to comparing the preliminary assessment result with the known rules and facts in the safety knowledge graph to check whether the assessment result is reasonable, violates safety regulations, or has any contradictions. The intermediate assessment result refers to the result generated after the preliminary assessment result has undergone logical consistency verification. If the preliminary assessment result has logical anomalies, it will be corrected and accompanied by verification information to form the intermediate assessment result.

[0039] Furthermore, the construction of the security knowledge graph is based on diverse and heterogeneous data sources, specifically including:

[0040] Structured knowledge injection: The industry's papermaking safety regulations, enterprise equipment operating procedures, emergency plans and other documents are transformed into triples of entities, relations and attributes through natural language processing and information extraction technology, forming the rule backbone of the safety knowledge graph;

[0041] Unstructured experience transformation: Extract events and mine causal relationships from texts such as historical accident reports, hazard logs, and expert experience summaries to identify typical risk chains, such as "poor equipment lubrication, bearing temperature rise, abnormal equipment vibration, and mechanical failure," and convert them into risk propagation paths in a safety knowledge graph;

[0042] Production system data association: Information such as equipment ledgers, piping and instrumentation diagrams in the manufacturing execution system and chemical safety data tables in the ERP system are associated with equipment entities, spatial nodes, and chemical entities in the safety knowledge graph, and the safety knowledge graph is assigned corresponding physical attributes and spatial topological relationships.

[0043] Furthermore, the maintenance of the security knowledge graph can adopt a dynamic update strategy that combines automatic discovery with manual confirmation. Specifically, during operation, new abnormal patterns or amendments that are discovered or frequently occur during the logical consistency verification process but are not covered by the existing security knowledge graph will be continuously recorded. After the recorded new abnormal patterns or amendments are confirmed by security engineers in the human-computer interaction and display module, security knowledge graph update suggestions can be automatically generated, such as adding a rule or correcting an entity attribute. After authorization, these suggestions can be imported into the security knowledge graph, thereby enabling the security knowledge graph to evolve autonomously along with production practices and risk management processes.

[0044] Specifically, a safety knowledge graph can be a structured database containing information such as industry safety standards, equipment operating procedures, and causal chains of risk events. When the risk perception model outputs a preliminary judgment result, this preliminary judgment result is compared with the known safety rules and logic in the safety knowledge graph. For example, if the preliminary judgment result indicates that a certain device has a high temperature risk, but the safety knowledge graph clearly stipulates that the device should not have a high temperature under the current operating mode, then the preliminary judgment result may have a logical anomaly. When checking whether the preliminary judgment result conforms to the normal state evolution logic defined in the graph, or whether any prohibitive rules have been triggered, if a logical anomaly is detected, the preliminary judgment result will be corrected according to the relevant rules in the knowledge graph. For example, the risk level may be adjusted or the risk type may be reclassified, and a verification mark will be added to the corrected result to generate an intermediate judgment result.

[0045] S40: Compare the uncertainty quantification score with a preset score threshold. If the score threshold is not exceeded, the intermediate judgment result is determined as the final judgment result. If the score threshold is exceeded, a preset collaborative judgment mechanism is triggered, and the final judgment result is determined.

[0046] In this embodiment, the score threshold refers to a preset value used for comparison with the uncertainty quantification score. When the uncertainty quantification score exceeds the score threshold, it indicates that the reliability of the preliminary judgment result is insufficient and further verification is required. The collaborative judgment mechanism refers to a strategy of introducing multiple parties or multiple models for joint evaluation when the risk perception model has low confidence in the preliminary judgment result. The collaborative judgment mechanism aims to improve the accuracy and reliability of the final risk judgment by integrating opinions or analysis results from multiple parties. The final judgment result refers to the risk assessment conclusion finally determined after the preliminary evaluation by the risk perception model, the logical verification of the security knowledge graph, and the possible collaborative judgment, serving as the basis for guiding subsequent risk handling and decision-making.

[0047] Specifically, the score threshold can be set according to the actual application scenario and the requirements for risk tolerance. For example, a low threshold can be set, indicating that the judgment result will only be directly adopted when the risk perception model is very confident in it. When the uncertainty quantification score is lower than or equal to the score threshold, it indicates that the results of the risk perception model and knowledge graph verification have high reliability. At this time, the intermediate judgment result can be directly adopted as the final judgment result. However, if the uncertainty quantification score exceeds the preset score threshold, it indicates that the confidence of the current risk judgment is insufficient and an additional review mechanism needs to be introduced. At this time, a collaborative judgment mechanism is triggered. For example, the current feature vector sequence and intermediate judgment results are submitted to multiple independent analysis units or human experts for review to obtain a more comprehensive evaluation and a more reliable conclusion. By integrating the review opinions, the final judgment result is determined.

[0048] S50: Execute the corresponding preset risk handling process and learning optimization process based on the final judgment result.

[0049] In this embodiment, the risk handling process refers to the preset response measures taken for the identified risks based on the final judgment result. The risk handling process may include issuing alarms, activating emergency plans, adjusting production operations, or notifying relevant personnel to intervene. The learning optimization process refers to re-inputting the results of risk handling, feedback information, and new data into the entire system process or risk perception model to update and improve the risk perception model, safety knowledge graph, or collaborative judgment mechanism, so as to achieve continuous improvement in the performance of the system process.

[0050] Specifically, once the final judgment is determined, corresponding risk management measures are implemented based on that judgment. For example, if the final judgment is "equipment overheating, posing a fire risk," an audible and visual alarm may be issued immediately to notify on-site operators via the corresponding notification terminal, and the cooling system may be automatically activated or the relevant equipment may be shut down. Simultaneously, this includes learning and optimizing processes. Specifically, the actual effectiveness of risk management, operator feedback, and new data are collected to continuously optimize the risk perception model, safety knowledge graph, and collaborative judgment mechanism. For example, if a risk management process is found to be inefficient, this information is recorded, and the management strategy is adjusted in subsequent iterations. Through this closed-loop learning mechanism, the risk management system processes can continuously adapt to changes in the production environment, thereby improving its risk identification and handling capabilities.

[0051] For example, suppose a pulp pump 100 is running in the pulping workshop of a paper mill. First, a multi-source dataset is continuously acquired through a data acquisition terminal. Specifically, equipment status data includes sensor data such as the vibration frequency, motor current, and bearing temperature of the pulp pump 100; video image data includes real-time footage captured by a camera installed near the pulp pump 100 to monitor whether there is a leak in the pump or whether personnel are approaching; operating environment data includes the temperature and humidity of the workshop, the concentration of harmful gases, etc.; and historical operating data includes historical fault records, maintenance records, and operating data of similar pumps under similar operating conditions for the pulp pump 100. After collecting the above multimodal data, it is initially sorted and converted into a feature vector sequence, which is then input into the risk perception model.

[0052] Furthermore, after receiving the feature vector sequence, the risk perception model analyzes it. For example, the risk perception model may detect abnormal fluctuations in the vibration frequency and motor current of the pulp pump 100 within a short period of time, while the bearing temperature continues to rise. Through the identification of the above patterns, the risk perception model initially determines that there is a "mechanical failure risk of pulp pump 100", the risk location is "pulp pump 100 in the pulping workshop", the risk evolution trend is "the failure is intensifying and may lead to shutdown", and the dynamic risk level is "medium risk". At the same time, the risk perception model outputs an uncertainty quantification score, such as 0.65, indicating its confidence in the determination.

[0053] Furthermore, the preliminary judgment result is input into the safety knowledge graph for logical consistency verification. The safety knowledge graph stores the normal operating parameter range of pulp pump 100, common failure modes and their causes, and related safety operating procedures. The "mechanical failure risk of pulp pump 100" in the preliminary judgment result is compared with the knowledge graph. For example, the safety knowledge graph may record that "when the temperature of the pulp pump bearing exceeds 80°C and the vibration frequency is abnormal, it usually indicates bearing wear or poor lubrication." If the preliminary judgment result is consistent with the rules in the safety knowledge graph, it is considered logically consistent. However, if the preliminary judgment result contradicts a prohibitive rule in the safety knowledge graph, for example, the risk perception model determines "pump body leakage," but the video image data does not show any signs of leakage, and there is no other evidence in the safety knowledge graph to support this judgment, then a logical anomaly will be identified, and the preliminary judgment result will be corrected according to the rules in the safety knowledge graph to generate an intermediate judgment result. In this example, it is assumed that the preliminary judgment result is logically consistent with the safety knowledge graph, and the intermediate judgment result is the same as the preliminary judgment result.

[0054] Furthermore, the uncertainty quantification score of 0.65 is compared with a preset score threshold, such as 0.5. Since 0.65 exceeds 0.5, it indicates that the risk perception model has a relatively low confidence level in this judgment, thus triggering a collaborative judgment mechanism. At this point, the real-time feature vector sequence of pulp pump 100 and intermediate judgment results can be distributed to multiple independent analysis units, such as expert models for analyzing vibration data, expert models for analyzing motor status, and prediction models based on historical fault modes. Each of these expert models performs a sub-risk judgment and outputs its local confidence level. For example, the vibration expert model may determine "high risk of bearing wear" with a confidence level of 0.8; the motor expert model may determine "medium risk of motor overload" with a confidence level of 0.7, etc. After collecting the above sub-risk judgment results and their local confidence levels, they are integrated through a preset fusion model, such as through weighted average or Bayesian fusion, to finally determine "severe bearing wear of pulp pump 100, with a risk of shutdown," and the dynamic risk level is adjusted to "high risk," which is then determined as the final judgment result.

[0055] Finally, based on the final judgment, the corresponding risk handling process is executed. For example, the highest level alarm is issued to the workshop control room, and relevant maintenance personnel and production supervisors are notified via SMS. At the same time, it is recommended to immediately stop the operation of pulp pump 100 and arrange for the maintenance team to inspect and replace the bearings. During the risk handling process, information such as the feedback from maintenance personnel, downtime, and maintenance costs will be recorded and input into the learning and optimization process. Furthermore, the above information will be used to update the risk perception model. For example, the model's recognition weight for specific vibration modes will be adjusted, or the description of pulp pump 100 failure modes in the safety knowledge graph will be updated, thereby achieving more accurate and timely identification and handling of similar risks.

[0056] In summary, the above technical solutions, by integrating multi-source heterogeneous data, intelligent risk perception, knowledge graph logic verification, and collaborative judgment mechanisms, can significantly improve the dynamic risk control capabilities of paper production safety management.

[0057] Compared with existing technologies that rely on a single data source or static threshold for alarms, this application acquires multi-source datasets such as equipment status data, video image data, operating environment data, and historical operation data, and constructs feature vector sequences to achieve multi-dimensional perception of information about the production process. This enables the identification of systemic risks that are coupled with multiple factors and dynamically evolve, rather than just isolated abnormal signals, thus improving the accuracy and foresight of production safety early warnings. For example, in the pulp pump failure example mentioned above, existing technologies may only issue an alarm based on a single bearing temperature exceeding the limit, while this application can simultaneously consider vibration, current, video footage, and even historical data to conduct a more comprehensive risk assessment.

[0058] Furthermore, this application inputs the feature vector sequence into a pre-trained risk perception model to obtain preliminary judgment results and associates them with uncertainty quantification scores. This addresses the problem of existing computational models being "black boxes" in their decision-making process and lacking confidence assessment. By quantifying uncertainty, the reliability of risk judgments can be identified, and a collaborative judgment mechanism can be triggered when the confidence level is insufficient. This contrasts with existing computational models, which may produce false alarms, missed alarms, and fail to initiate a review process. In the pulp pump failure example, when the risk perception model has low confidence in the preliminary judgment results, multiple expert models are introduced for review, thereby effectively reducing the false alarm rate and improving the reliability of the final judgment.

[0059] Furthermore, this application performs logical consistency verification on the preliminary judgment results based on a pre-set safety knowledge graph, and generates intermediate judgment results based on the verification results. This makes the decision-making process of the risk perception model more transparent and interpretable, thereby avoiding situations where the output results violate basic physical common sense or industry safety regulations. Compared with the risk judgments that may be inconsistent with reality by existing calculation models, this application can ensure the rationality and accuracy of risk judgments through the constraints and corrections of the knowledge graph. For example, in the pulp pump failure example mentioned above, the safety knowledge graph can ensure that the failure type determined by the risk perception model is logically consistent with the actual observed data pattern.

[0060] Finally, this application implements the corresponding risk handling process and learning optimization process based on the final judgment result, constructing a closed-loop mechanism for risk management. This addresses the problem that most existing solutions stop at risk alarms and lack effective feedback of handling experience and results to the calculation model to achieve self-iterative optimization of the system process. Through continuous learning and optimization, the risk management system can continuously adapt to changes in the production line, improve its intelligence level and long-term operational stability. In the pulp pump failure example, the feedback from maintenance personnel and the actual handling results will be used to optimize the risk perception model, enabling more accurate judgments and more effective responses when facing similar failures in the future.

[0061] In one embodiment, step S10 includes:

[0062] S11: Determine the target time window, obtain the multi-source dataset of the target production area based on the target time window, and perform timestamp alignment and spatial coordinate mapping on the multi-source dataset to generate a spatiotemporally aligned dataset;

[0063] In this embodiment, defining a target time window aims to clarify the time range for data collection and analysis, ensuring that subsequent processing is conducted on a unified time benchmark, avoiding data redundancy or omissions, and focusing on the risk situation within a specific time period. The target time window can be determined manually by inputting the start and end times through the user interface, or automatically set according to preset rules such as production shifts and equipment operating cycles. Furthermore, a sliding time window can be dynamically set according to the needs of real-time risk monitoring, for example, updating the data from the most recent hour every 5 minutes. Based on the target time window, multi-source datasets of the target production area are acquired, ensuring that all required raw data is collected from various data sources within the specified time range. Data can be pulled in real-time or near real-time from SCADA systems, DCS systems, video management systems, environmental monitoring stations, MES systems, etc., through data acquisition agents or API interfaces, or by subscribing to relevant data streams through message queues and caching and filtering within the target time window. Timestamp alignment of the multi-source datasets aims to resolve time inconsistencies caused by different data sources due to collection frequency, transmission delays, etc., ensuring that events occurring at the same time are aligned. To ensure accurate correlation across different data modalities, interpolation can be used to unify all data to the smallest time granularity or a preset sampling frequency. Alternatively, a time synchronization protocol can be used to ensure consistent clocks across data source devices, recording precise timestamps during data acquisition and subsequently merging data based on those timestamps. Spatial coordinate mapping aims to unify data from different data sources into a unified spatial coordinate system for the target production area, enabling spatial correlation analysis of data from different modalities. This can be achieved by mapping the physical locations of various sensors, cameras, and work areas to a unified coordinate system using a pre-established 3D model or 2D plan of the factory, or by using Geographic Information System (GIS) technology to divide the production area into grids and classify data points into corresponding grid cells. Generating a spatiotemporally aligned dataset aims to integrate multi-source data after timestamp alignment and spatial coordinate mapping, forming a dataset that is consistent in both time and space, providing input for subsequent feature extraction. Specifically, the aligned data can be stored in a unified tabular format, including timestamps, spatial IDs, and the original or preprocessed data of each modality, or by constructing multidimensional tensors or graph structures to better represent the spatiotemporal relationships between data.

[0064] S12: Perform feature extraction and encoding on the spatiotemporal aligned dataset to obtain a feature vector set, and generate a feature vector sequence based on the feature vector set. The feature vector set includes state feature vectors, visual feature vectors, working environment feature vectors, and background knowledge feature vectors.

[0065] In this embodiment, feature extraction and encoding are performed on the spatiotemporally aligned dataset to extract information meaningful for risk perception from the original spatiotemporally aligned data and convert it into a numerical representation that can be processed by a machine learning model. For example, for equipment status data, statistical features of operating parameters, outliers, and fault codes can be extracted; for video image data, target detection results, behavior recognition results, smoke, and flame recognition can be extracted; for work environment data, temperature, humidity, and concentration of harmful gases can be extracted; for historical work data, historical accident types, frequency of occurrence, and handling measures can be extracted. Encoding can be performed using one-hot encoding, embedded vectors, or by using a deep learning model to automatically learn and extract high-dimensional features from the original data and encode them into fixed-length vectors. The resulting feature vector set aims to organize the features extracted from different modalities into a vector set. This can be achieved by storing the extracted status feature vectors, visual feature vectors, work environment feature vectors, and background knowledge feature vectors as independent vectors and concatenating or fusing them in subsequent steps, or by encapsulating these vectors in a data structure, such as a dictionary or a custom structure. The system defines objects for easy management and access; it generates feature vector sequences based on feature vector sets, aiming to arrange feature vector sets from different time steps in chronological order to form a sequence, thereby capturing dynamic information about risk evolution and providing temporal input for the risk perception model. It can use the feature vector set generated within each time window as input for one time step, and then concatenate these feature vector sets in chronological order to form a two-dimensional or three-dimensional tensor as a sequence; or it can use sliding window technology to generate a sequence containing feature vector sets from the past N time steps at each time point. The feature vector sets include: state feature vectors, which describe equipment operating status, personnel physiological status, etc., such as equipment temperature, pressure, current, personnel heart rate, body temperature, etc.; visual feature vectors, which describe visual information obtained through video image analysis, such as personnel location, behavior, equipment appearance, environmental anomalies, etc.; work environment feature vectors, which describe environmental parameters of the production site, such as temperature, humidity, harmful gas concentration, noise, etc.; and background knowledge feature vectors, which describe static or semi-static information related to risk, such as equipment model, maintenance records, operating procedures, historical accident patterns, expert experience, etc.

[0066] Specifically, by introducing a target time window, timestamp alignment, spatial coordinate mapping, and multimodal feature extraction and encoding, the original multi-source dataset is preprocessed to construct a feature vector sequence. First, by determining the target time window, the scope of data collection is limited, ensuring the temporal relevance of the data. Then, timestamp alignment and spatial coordinate mapping are performed on the multi-source datasets acquired within the target time window to address the inherent temporal asynchrony and spatial heterogeneity issues between different data sources, integrating scattered data into a spatiotemporally aligned dataset and improving the inherent consistency and correlation of the data. Based on this, feature extraction and encoding are performed on the spatiotemporally aligned dataset. Representative state feature vectors, visual feature vectors, operational environment feature vectors, and background knowledge feature vectors are extracted from multiple dimensions such as device status, video images, operating environment, and background knowledge, forming a feature vector set. This feature vector set is then organized into a feature vector sequence according to chronological order.

[0067] Through the above technical solutions, this application can effectively solve the problem of temporal and spatial inconsistency of multi-source heterogeneous data, ensuring that the data input into the risk perception model has high accuracy and relevance. By defining the target time window, performing timestamp alignment and spatial coordinate mapping, data from different modalities can be accurately integrated and understood, avoiding risk misjudgment or omission due to data misalignment or missing data. In addition, by performing multi-dimensional feature extraction and encoding on the spatiotemporally aligned dataset, potential risk factors in the production site are characterized from multiple levels such as equipment status, visual information, working environment and background knowledge, enriching the input information for risk perception. This provides a more comprehensive, accurate and spatiotemporally consistent data foundation for the risk perception model, which can significantly improve the accuracy and real-time performance of dynamic risk control in paper production safety management.

[0068] In one embodiment, the risk perception model includes a scene understanding layer, a feature fusion layer, and a decision output layer, such as... Figure 2 As shown, step S20 includes:

[0069] S21: The scenario understanding layer identifies the operation scenario type information corresponding to the target production area based on the received feature vector sequence;

[0070] In this embodiment, the risk perception model has a hierarchical structure, including a scenario understanding layer, a feature fusion layer, and a decision output layer. The scenario understanding layer primarily performs preliminary analysis of the input data to identify the type or state of the current working environment, helping subsequent processing layers adjust strategies according to the specific scenario. For example, the scenario understanding layer could be a deep learning-based classifier to identify whether it is an equipment maintenance scenario, a normal production scenario, or an emergency shutdown scenario; or it could be a rule engine to determine whether a specific operating mode is currently in use based on sensor data. The feature fusion layer integrates features from different data sources to discover features that are difficult to reveal with a single modality. Deeply correlated or coupled risks can be addressed through techniques such as attention mechanisms, graph neural networks, or multimodal fusion networks in the feature fusion layer. These techniques can be used to evaluate and weight the importance of different features. Alternatively, collaborative filtering or tensor decomposition can be employed to uncover the intrinsic connections between different modalities. The decision output layer generates specific risk judgment results based on the fused risk signals and assesses their confidence level. For example, the decision output layer could be a multi-task learning model that simultaneously outputs the risk type, location, evolution trend, and level. Alternatively, it could be a Bayesian network that combines prior knowledge and evidence for probabilistic reasoning and outputs an uncertainty quantification score.

[0071] In this embodiment, step S21 aims to provide contextual information for subsequent risk analysis, making the risk perception process more targeted. For example, in paper production, different work scenarios have different risk patterns and concerns, such as paper machine operation, equipment maintenance, and material handling. The implementation of step S21 may include: training the feature vector sequence using a classification algorithm so that it can map the sequence to a predefined work scenario type; or, using a clustering algorithm to perform unsupervised learning on historical data to automatically discover and define different work scenario types.

[0072] S22: The feature fusion layer performs cross-modal weighted correlation analysis on the feature vector sequence based on the job scenario type information to identify potential coupling risk signals;

[0073] In this embodiment, step S22 aims to fully utilize the complementarity of multi-source data. By performing weighted fusion and correlation analysis on different modal features, it can discover complex risk patterns that are difficult to reveal with single modal data, especially coupled risks caused by multiple factors. The implementation of step S22 may include: using an attention-based fusion method to dynamically adjust the weights of different modal features according to the type of work scenario. For example, in a device failure scenario, the weight of device status data may be higher; or, using a graph neural network to construct a multimodal feature map and capturing the correlation between different feature nodes through graph convolution operations, thereby identifying potential coupled risk signals.

[0074] S23: The output layer determines the risk type, risk location, risk evolution trend, and dynamic risk level based on the potential coupled risk signal, and simultaneously generates an uncertainty quantification score that represents the confidence level of the determination.

[0075] In this embodiment, step S23 aims to transform the abstract risk signal into specific and actionable risk information and provide an assessment of the reliability of the judgment result. The implementation of step S23 may include: designing a multi-head output neural network model, wherein each head is responsible for predicting a risk attribute, and an additional head is set up to predict the uncertainty quantification score; or, using a probabilistic graphical model to model the potential coupled risk signal, thereby simultaneously outputting each risk attribute and its corresponding confidence level.

[0076] Specifically, the risk perception model, with its layered architecture, more effectively identifies dynamic risks in the paper production process from heterogeneous, multi-source data. Specifically, the scenario understanding layer performs preliminary analysis on the received feature vector sequences to identify the specific operational scenario type of the current target production area, providing crucial contextual information for subsequent risk analysis. Furthermore, the feature fusion layer, based on the operational scenario type information, performs cross-modal weighted correlation analysis on feature vectors from different modalities, including equipment status data, video image data, operational environment data, and historical operational data. Through this weighted correlation analysis, the risk perception model can deeply explore the intrinsic connections and complementary information between different data modalities, thereby identifying potential coupled risk signals that are difficult to detect from a single data source. For example, abnormal equipment vibration and signs of operator fatigue may jointly indicate a higher risk of operational errors in a specific operational scenario. Finally, the judgment output layer, based on the identified potential coupled risk signals, comprehensively generates preliminary judgment results including risk type, risk location, risk evolution trend, and dynamic risk level, and simultaneously outputs an uncertainty quantification score representing the confidence level of the judgment.

[0077] As a specific implementation method, the risk perception model can be constructed as a deep learning framework. The scene understanding layer can employ a sequence classifier based on the Transformer architecture. This classifier receives a sequence of feature vectors as input and outputs the current work scene type, such as "high-speed paper machine operation," "pulp preparation," or "equipment maintenance." The feature fusion layer can employ a multimodal attention network, which dynamically assigns attention weights to features of different modalities in the feature vector sequence based on the work scene type information output by the scene understanding layer. For example, in an equipment maintenance scenario, the weight of equipment status features may be increased, while the weight of personnel behavior features in video images may also increase accordingly. This multimodal attention... The force network integrates weighted features into a unified risk scenario representation through weighted summation or more complex fusion mechanisms, and identifies potential coupled risk signals from it. The decision output layer can adopt a multi-task prediction network, which receives the fused risk signals and outputs risk type (such as "equipment overheating" and "personnel violation"), risk location (such as "drying section" and "grinding area"), risk evolution trend (such as "rapid escalation" and "slow diffusion"), and dynamic risk level (such as "low", "medium", and "high") in parallel. At the same time, the multi-task prediction network can also generate uncertainty quantification scores associated with each decision result through Monte Carlo sampling or Bayesian neural network methods to quantify the confidence level of the risk perception model in its own decisions.

[0078] For example, the risk background is that the high temperature on the surface of the drying cylinder in the papermaking drying section can easily cause a fire due to the long-term accumulation of lubricating oil or paper fibers.

[0079] The system simultaneously analyzes the temperature and vibration data of the drying cylinder bearing (equipment status), infrared thermal imaging video (video images, monitoring surface temperature distribution hotspots), and surrounding VOC concentration (operating environment, monitoring combustible volatiles). When it identifies "continuous abnormal increase in local temperature" (time-series characteristic) and "simultaneous increase in VOC concentration," the feature fusion layer associates it with a coupled signal of "lubricating oil evaporation due to heat." The output layer combines the inherent development pattern of the "fire" risk type to generate a dynamic risk level of "high fire hazard." If the uncertainty is low, it directly triggers a fire protection system warning and notifies the inspection team. If the uncertainty is high, such as blurred smoke recognition, it triggers collaborative judgment, scheduling the infrared image analysis model and the gas leak model for consultation, thereby improving the reliability of the judgment.

[0080] Through the above technical solution, this application can effectively distinguish risk characteristics under different operating scenarios when multi-source heterogeneous data is directly input into the risk perception model in the complex environment of papermaking production, thereby improving the accuracy and efficiency of risk perception. Specifically, by introducing a scenario understanding layer, the current operating scenario type information can be identified first, providing accurate contextual information for subsequent risk analysis, making the risk perception process more targeted. The feature fusion layer performs cross-modal weighted correlation analysis on multimodal features based on scenario information, which can more deeply explore the potential coupled risk signals between different data modalities and avoid the limitations of a single data source. The judgment output layer can comprehensively and accurately generate risk type, risk location, risk evolution trend and dynamic risk level, and provide uncertainty quantification score, thereby improving the accuracy, robustness and interpretability of risk perception.

[0081] In one embodiment, such as Figure 3 As shown, step S22 includes:

[0082] S221: Based on the pre-set attention mechanism and job scenario type information, calculate the correlation weights between each feature vector in the feature vector set;

[0083] In this embodiment, the pre-defined attention mechanism is a computational model used to simulate the attention allocation process in human cognition, thereby dynamically allocating different levels of attention or weights based on different parts of the input data. This model can automatically focus on the information most relevant to the current task when processing large amounts of heterogeneous data, thus improving the efficiency and accuracy of information processing. For example, a self-attention mechanism based on the Transformer architecture can be used to dynamically generate weights by calculating the similarity between queries, keys, and values; or a channel attention mechanism can be used to allocate weights by learning the importance of different feature channels. Job scenario type information refers to descriptive data about the specific operating environment or activity state of the current target production area, which provides information for risk perception. This provides important context, enabling the adjustment of feature analysis and risk reasoning strategies based on different scenario characteristics. For example, job scenario type information can include "equipment operation," "equipment maintenance," "material conveying," and "personnel inspection." The correlation weights between feature vectors in the feature vector set are calculated to quantify the degree of mutual influence or correlation between different modal features in the feature vector set. By calculating the correlation weights, it is possible to identify which feature modalities or feature combinations are more critical for risk identification in a specific job scenario. For example, the output of an attention mechanism can be used as the correlation weight, or a dedicated neural network layer can be used to learn and predict the correlation weights. This neural network layer can receive feature vectors and job scenario type information as input.

[0084] S222: Based on the calculated relevance weights, the feature vectors of different modalities in the feature vector set are weighted and fused to generate a risk scenario representation;

[0085] In this embodiment, weighted fusion of feature vectors from different modalities in the feature vector set refers to combining feature vectors from different data sources according to their calculated relevance weights to form a more informative and representative comprehensive feature representation. Weighted fusion effectively integrates multi-source information while highlighting important features and suppressing irrelevant or noisy features. For example, a weighted summation method can be used, multiplying each feature vector by its corresponding weight and then accumulating the sums; or a gating mechanism can be used, dynamically controlling the contribution ratio of different modal features by learning a corresponding gating function. Risk scenario representation refers to the abstract representation generated after weighted fusion that comprehensively and compactly describes the potential risk state in the current work scenario. It integrates multimodal information and highlights key features related to risk. For example, risk scenario representation can be a high-dimensional vector whose dimension and value encode the risk features of the current scenario.

[0086] S223: Based on risk scenario representation, perform risk correlation reasoning that matches the job scenario type information to identify potential risk signals.

[0087] In this embodiment, performing risk correlation reasoning based on the matching of work scenario type information refers to calling or activating a predefined reasoning logic or model that matches the current work scenario type to identify specific potential risk signals from the risk scenario representation. This scenario-matching reasoning method can ensure the targeting and accuracy of risk identification. For example, a rule-based reasoning engine can be preset to activate different rule sets for different work scenario types; or, multiple scenario-specific machine learning models can be trained, and the appropriate model can be selected for reasoning based on the work scenario type information. Potential risk signals refer to specific signs or patterns discovered through risk correlation reasoning that indicate the possibility of safety accidents or abnormal situations. Among them, potential risk signals can be early signs of equipment failure, patterns of personnel violations, or indicators of environmental anomalies, etc.

[0088] Specifically, the solution in this application enhances the risk perception model's ability to identify potential coupled risk signals by introducing a pre-defined attention mechanism and work scenario type information. Specifically, when the feature vector sequence enters the feature fusion layer, based on the current work scenario type information and combined with the pre-defined attention mechanism, the correlation weights between each feature vector in the feature vector set are calculated. This allows for the determination of the importance and interrelationship of different modal features in the current scenario based on the specific work context. For example, in the "equipment operation" scenario, equipment vibration data may be given higher weight, while in the "personnel inspection" scenario, personnel behavior features in video images may receive more attention. Furthermore, based on the calculated correlation weights, the feature vectors of different modalities in the feature vector set are weighted and fused to generate a risk scenario representation that more accurately and comprehensively reflects the current risk state. This representation not only integrates multi-source information but also highlights key features closely related to the current scenario risk through weight allocation. Finally, based on the risk scenario representation, risk association reasoning that matches the work scenario type information is executed. This enables the dynamic activation or selection of the most suitable reasoning model or rule set for the current work scenario, thereby specifically identifying potential coupled risk signals.

[0089] By introducing a pre-set attention mechanism and job scenario type information through the above technical solution, it is possible to calculate the correlation weights between different modal features in a targeted manner and perform weighted fusion to generate a more representative risk scenario representation. Through this context-based feature weighting and fusion, the deep coupling relationship between different features can be captured more accurately, avoiding information redundancy or loss of key information that may result from simple fusion. In addition, by performing risk association reasoning that matches job scenario type information, the most appropriate reasoning logic can be dynamically activated according to the specific production scenario, thereby improving the accuracy and robustness of potential coupling risk signal identification and effectively reducing the risk of false alarms and false negatives.

[0090] In one embodiment, such as Figure 4 As shown, step S23 includes:

[0091] S231: Determine multimodal evidence and risk type based on potential coupling risk signals, and identify the support score and reliability weight of multimodal evidence;

[0092] In this embodiment, the potential coupled risk signal refers to a comprehensive signal identified through cross-modal weighted correlation analysis of feature vector sequences from different modalities, such as equipment status data, video image data, operating environment data, and historical operating data. This signal indicates the existence of potential risks. The potential coupled risk signal is the result of the interaction and mutual corroboration of multiple data modalities, reflecting deep-seated risk correlations that are difficult to reveal with single-modal data. Multimodal evidence refers to specific data fragments or information units extracted from the potential coupled risk signal, representing support for or rebuttal of specific risk determinations. Multimodal evidence originates from different data modalities, such as equipment sensor readings, abnormal behavior in video footage, environmental monitoring data, or historical fault records. Risk type refers to the label or description used to classify the identified potential risks, such as equipment failure, personnel violations, environmental pollution, and fire hazards. Support score is a quantitative indicator of the degree to which each piece of multimodal evidence supports a specific risk type or risk attribute, reflecting the strength or relevance of the evidence. Reliability weight is a quantitative indicator of the reliability of each source of multimodal evidence or the evidence itself. For example, calibrated, high-precision sensor data may have a higher reliability weight, while blurry video images or unverified human recordings may have a lower reliability weight.

[0093] Furthermore, the risk location can be obtained by identifying the evidence among the multimodal evidence that provides the most direct spatial information, and by determining the specific physical location or regional extent of the risk in the target production area based on the identified evidence.

[0094] Furthermore, based on the inherent development pattern of the temporal characteristics representing state changes in multimodal evidence and the pre-defined association between risk type, the evolution trend of risk can be inferred and output. The risk evolution trend includes at least one of diffusion direction and escalation speed.

[0095] Among them, the specific physical location or area of ​​the risk in the target production area refers to the precise location or scope of the risk event, which can be determined by modal evidence with spatial positioning capabilities; the evolution trend of the risk refers to the direction and speed of the risk event in the time dimension, which can be inferred based on the temporal characteristics reflecting state changes in multimodal evidence, combined with the inherent development pattern of the pre-set risk type.

[0096] Furthermore, in the field of industrial safety, the inherent development patterns of pre-defined risk types include:

[0097] Linear growth / diffusion pattern: Risk parameters (such as temperature, pressure, concentration) change linearly over time. For example, a small leak may cause the concentration of harmful gases to diffuse linearly in space at a relatively constant rate.

[0098] Exponential growth / acceleration pattern: The rate of risk development accelerates dramatically over time or under certain conditions. For example, the reaction rate of some chemical reactions increases exponentially after reaching a critical temperature; or, the failure rate of equipment fatigue damage accelerates dramatically after reaching a critical point.

[0099] Phased or chain reaction pattern: The development of risk presents clear phases, with the result of the previous phase becoming the trigger for the next phase. For example, in a fire scenario, the typical inherent development pattern may be presupposed as follows: heat source ignites combustibles (initial stage), fire expands and produces a large amount of dense smoke (development stage), and may trigger an explosion or structural collapse (intense stage / derived disaster). Based on the detected evidence of "local high temperature" and "increased smoke concentration", it can be inferred that the evolution trend of "fire is in the development stage and may accelerate";

[0100] Diffusion / propagation model: focuses on the spatial spread of risk, which may include gas diffusion model, heat transfer model or pollutant migration model based on fluid dynamics and heat conduction model. For example, based on the nature of the leaked gas and the site layout, the diffusion model in the downwind direction or low-lying area is preset.

[0101] Attenuation / Dissipation Pattern: Some risks tend to decrease after intervention or changes in conditions. For example, the risk of electrical fires decreases rapidly after power is cut off, or the concentration of harmful gases decreases exponentially after ventilation.

[0102] S232: Calculate the overall evidence confidence level of the current risk assessment by weighting and summing the support scores of the multimodal evidence according to their reliability weights;

[0103] In this embodiment, the overall confidence level of evidence refers to a quantitative indicator of the overall reliability of the current risk assessment result. It can be calculated by weighting and summing the support scores of multimodal evidence, where the weights are the reliability weights of each piece of evidence.

[0104] S233: Based on the risk type, match its preset corresponding consequence severity coefficient and time urgency coefficient, and combine with the comprehensive evidence confidence level, calculate and generate a dynamic risk level through a preset risk quantification function;

[0105] In this embodiment, the consequence severity coefficient refers to a quantitative indicator of the potential harm caused by a specific type of risk once it occurs. This consequence severity coefficient is preset and can be defined based on historical accident data, industry standards, safety regulations, and expert experience. The time urgency coefficient refers to a quantitative indicator of the urgency to which a specific type of risk needs to be addressed. This time urgency coefficient is also preset and can be defined based on factors such as the speed of risk evolution, potential impact range, and controllability. The risk quantification function refers to a preset mathematical model or algorithm used to integrate multiple factors such as risk type, consequence severity coefficient, time urgency coefficient, and comprehensive evidence confidence level to calculate and generate a dynamic risk level. The dynamic risk level refers to the risk level obtained after real-time assessment of the current risk status, such as low, medium, high, and extremely high.

[0106] S234: Perform conflict identification and completeness assessment on multimodal evidence, and determine an uncertainty quantification score based on the results of conflict identification and completeness assessment.

[0107] In this embodiment, conflict identification refers to detecting whether there are contradictions or inconsistencies among multimodal evidence; completeness assessment refers to assessing whether the currently available multimodal evidence is sufficient to make a comprehensive and accurate judgment on the risk; uncertainty quantification score refers to a quantitative indicator of the degree of uncertainty in the risk judgment result, wherein the uncertainty quantification score is determined based on the results of conflict identification and completeness assessment.

[0108] Specifically, upon receiving a potential coupling risk signal identified by the feature fusion layer, the output layer does not directly output a preliminary judgment result. Instead, it analyzes the potential coupling risk signal to determine the multimodal evidence constituting the risk judgment and its corresponding risk type. During this process, it simultaneously identifies the support score of each multimodal evidence for the risk and its own reliability weight. For example, when a potential coupling risk signal indicates an overheating risk, temperature sensor data, infrared thermal imager images, and equipment operation logs are extracted from the potential coupling risk signal as multimodal evidence, and the support level and reliability of each piece of evidence for the overheating risk are evaluated. Further, based on the modal evidence providing the most direct spatial information, such as video image data or sensor data at specific locations, the specific physical location or area of ​​the risk in the target production area is determined. Simultaneously, based on the temporal characteristics representing state changes in the multimodal evidence, such as temperature change curves over time, and combined with the pre-defined inherent development pattern of the risk type, the evolution trend of the risk is inferred and output, including its possible diffusion direction or escalation speed. To ensure... The reliability of risk assessment is further enhanced by weighted summation of support scores based on the reliability weights of multimodal evidence. This calculation yields a comprehensive evidence confidence level for the current risk assessment, effectively integrating data from different sources and with varying degrees of credibility, resulting in a more comprehensive and accurate confidence evaluation. Building upon this, based on the identified risk type, a pre-defined severity coefficient and a time urgency coefficient for risk development are matched. These two coefficients, along with the comprehensive evidence confidence level, are input into a pre-defined risk quantification function to ultimately generate a dynamic risk level. Furthermore, to further improve the transparency and traceability of the assessment, conflict identification and completeness assessment are conducted on the collected multimodal evidence. Conflict identification aims to identify contradictions between different pieces of evidence, while completeness assessment determines whether the existing evidence is sufficient to support the current risk assessment. Finally, an uncertainty quantification score is determined based on the results of conflict identification and completeness assessment. This score directly reflects the degree of uncertainty in the current risk assessment, providing crucial input for subsequent logical consistency verification and collaborative assessment mechanisms.

[0109] By comprehensively considering the support score and reliability weight of multimodal evidence, and combining risk type, consequence severity coefficient, and time urgency coefficient, this application can generate a more accurate and persuasive dynamic risk level. At the same time, the conflict identification and completeness assessment of multimodal evidence quantifies the uncertainty of the judgment, providing more comprehensive information for subsequent decision-making, and has the effect of improving the transparency, reliability, and decision support capabilities of risk assessment.

[0110] In one embodiment, step S30 includes:

[0111] S31: Identify the risk entities in the preliminary judgment results, and perform semantic matching and association binding between the risk entities, risk locations and corresponding entity nodes, activity nodes and spatial nodes in the security knowledge graph, thereby activating the risk dynamic subgraph;

[0112] In this embodiment, a risk entity refers to a specific hazard source, abnormal event, or potential hazard identified in the preliminary judgment results, such as "equipment overheating," "personnel violation of operating procedures," or "chemical leak." A risk location refers to the specific physical space or area where the risk entity is located, such as "drying section," "paper machine operating table," or "chemical storage area." Identifying risk location information can typically be achieved by parsing the text description of the preliminary judgment results using natural language processing technology, or by extracting it from structured data using preset keyword matching, named entity recognition, or other methods. A safety knowledge graph is a structured knowledge base that represents entities, concepts, and their relationships in the field of paper production safety in the form of a graph. The safety knowledge graph contains various types of nodes, such as entity nodes (e.g., "drying section," "cooling system," "operator"), activity nodes (e.g., "equipment operation," "maintenance work," "chemical addition"), spatial nodes (e.g., "workshop A," "area B"), and edges connecting these nodes. Edges represent various relationships between entities (e.g., "located in," "contains," "causes," "requires"). Furthermore, the safety knowledge graph can also embed a large number of safety rules, operating procedures, and accident cases. Knowledge embedded in the knowledge graph typically exists in the form of logical rules or constraints. Furthermore, the construction of a security knowledge graph can be based on expert experience, historical accident reports, safety regulations, industry standards, etc., and can be achieved through manual construction, semi-automatic extraction, or machine learning. Semantic matching refers to semantically mapping the risk entities and risk locations identified in the preliminary judgment results to existing entity nodes, activity nodes, and spatial nodes in the security knowledge graph. Association binding refers to establishing or strengthening the connection relationships between these matched entities in the security knowledge graph after successful semantic matching, thus enabling the initial... The information in the judgment results can be effectively linked with relevant knowledge in the graph, which can be achieved through various technologies such as ontology-based matching algorithms, word vector similarity calculation, and rule reasoning. The risk dynamic subgraph refers to a local view of the security knowledge graph. Specifically, after semantic matching and association binding, it is a set of entities, relationships, and rules most relevant to the current risk situation dynamically extracted from the entire security knowledge graph, centered on the risk entities and risk locations in the preliminary judgment results. The purpose of activating the risk dynamic subgraph is to limit the scope of reasoning to the knowledge most relevant to the current risk, thereby improving reasoning efficiency and accuracy.

[0113] S32: Based on the risk dynamic subgraph, traverse and reason along the predefined rule path to detect whether there are logical anomalies in the preliminary judgment results;

[0114] In this embodiment, the predefined rule path refers to a knowledge chain or logical sequence pre-set in the safety knowledge graph for logical reasoning and verification. The rule path may include: prerequisite checks, prohibitive rules, and normal state evolution logic. The rule path is usually defined by domain experts based on safety production experience and standards, and stored in the safety knowledge graph in the form of logical expressions, production rules, or ontology constraints. Traversal refers to systematically checking every node and edge related to the preliminary judgment result along the predefined rule path in the risk dynamic subgraph. Reasoning calculation refers to using a logical reasoning engine during this traversal process to make logical judgments on the risk state or suggested actions described by the preliminary judgment result based on the rules and knowledge stored in the graph. Logical anomaly refers to the existence of contradictions, inconsistencies, or unreasonable aspects in the preliminary judgment result that are consistent with the domain knowledge, safety standards, or common sense represented by the safety knowledge graph.

[0115] S33: If a logical anomaly is detected, the corresponding rule chain and graph evidence node will be located.

[0116] In this embodiment, the rule chain refers to the specific reasoning path or a series of related rules that lead to the occurrence of logical anomalies; the graph evidence node refers to the entity node, relationship or rule in the knowledge graph that directly leads to or supports the judgment of logical anomalies in the rule chain.

[0117] S34: Correct the preliminary judgment result based on the located rule chain, and add verification marks to the corrected preliminary judgment result based on the rule chain and the graph evidence nodes to generate intermediate judgment result.

[0118] In this embodiment, revising the preliminary judgment result refers to adjusting or supplementing the risk type, risk location, risk evolution trend, or dynamic risk level in the preliminary judgment result based on the located rule chain and graph evidence node after detecting a logical anomaly; adding a verification mark refers to adding an identifier or metadata to the revised preliminary judgment result, indicating that the result has been verified by the logical consistency of the security knowledge graph and may have been revised based on the verification result; generating an intermediate judgment result refers to forming a risk judgment output that is more reliable and logically consistent than the preliminary judgment result after the logical consistency verification and possible revision of the security knowledge graph.

[0119] Specifically, the solution in this application improves the reliability and logical rationality of the preliminary judgment results output by the risk perception model by introducing a logical consistency verification mechanism based on a security knowledge graph. Specifically, after the risk perception model generates a preliminary judgment result, it first identifies key risk entities and risk location information from this result. Furthermore, it performs semantic matching and association binding between the identified information and entity nodes, activity nodes, and spatial nodes in the pre-constructed security knowledge graph, thereby dynamically extracting the local knowledge network most relevant to the current risk situation from the security knowledge graph, i.e., activating the risk dynamic subgraph. Once the risk dynamic subgraph is activated, it systematically traverses and performs reasoning calculations along predefined rule paths based on this risk dynamic subgraph. This reasoning process aims to comprehensively check... The risk status described in the preliminary judgment result or its implicit handling suggestions are compared with the prerequisites, prohibitive rules, and normal production state evolution logic stored in the knowledge graph. Through the above logical comparison, any logical anomalies or inconsistencies in the preliminary judgment result can be effectively detected. Once a logical anomaly is detected, the specific rule chain and related knowledge graph evidence nodes that caused the anomaly are further traced and located. The located rule chain and evidence nodes provide clear basis for correcting the preliminary judgment result. Based on the provided basis, the preliminary judgment result is corrected to eliminate its logical irrationality. Furthermore, the corrected preliminary judgment result is also attached with a verification mark to indicate that it has been logically verified and corrected by the knowledge graph, thereby generating a more reliable and accurate intermediate judgment result.

[0120] By introducing a safety knowledge graph for logical consistency verification, the above technical solutions enable deep semantic understanding and rule-based reasoning of preliminary judgment results. This allows for the timely detection and correction of potential logical anomalies, improving the accuracy, reliability, and rationality of risk assessment results and preventing erroneous decisions or delays in risk management due to model misjudgments. Intermediate judgment results verified and corrected by the knowledge graph provide a more reliable foundation for subsequent risk handling and learning optimization processes, thereby enhancing the dynamic risk control capabilities of the paper production safety management system.

[0121] In one embodiment, step S40 includes:

[0122] S41: If the uncertainty quantification score exceeds the score threshold, then based on the intermediate judgment result, the N expert models with the highest matching degree are dynamically scheduled from the preset expert model pool to form an expert model group.

[0123] In this embodiment, the pre-built expert model pool refers to a pre-constructed set of specialized risk assessment models optimized for specific risk types, data modalities, or scenarios. Each expert model in the pool can be based on different machine learning algorithms, such as a time-series analysis-based model for equipment failure prediction, a deep convolutional neural network model for visual anomaly detection, or a rule-based reasoning model for operational procedure violations. The purpose is to provide more detailed and professional auxiliary judgment capabilities when the uncertainty of the main risk perception model is high. Dynamically scheduling the N expert models with the highest matching degree refers to... The system quantifies the outcome and uncertainty scores, and selects N expert models from the expert model pool to best handle the current risk scenario. The matching degree can be evaluated based on factors such as the degree of fit between the expert model's preset expertise and the current risk type, the model's performance in similar historical cases, or the model's ability to process specific data modalities. N is a configurable integer representing the number of expert models to be invoked, and its value can be adjusted according to system resources, response time requirements, and risk complexity. An expert model group refers to a temporary collaborative unit composed of N expert models selected by dynamic scheduling, used to jointly conduct in-depth analysis of high uncertainty risks.

[0124] Furthermore, the expert models can be dynamically scheduled based on the intermediate judgment results. This can be achieved by using a preset model-scenario-capability mapping matrix, which defines the risk types that each expert model in the expert model pool is good at, the data modal it adapts to, and its historical performance scores in different work scenarios, such as "paper making", "coating preparation", and "equipment maintenance".

[0125] The scheduling process specifically includes:

[0126] Scene filtering: Based on the job scenario type information corresponding to the intermediate judgment results, a subset of pre-defined effective expert models under the job scenario type is initially selected as candidate models;

[0127] Risk matching: Based on the risk type corresponding to the intermediate judgment result, calculate the matching degree score of each candidate model in the mapping matrix for that risk type;

[0128] Evidence fit: Analyze the sources of uncertainty that trigger collaborative judgment. If the uncertainty mainly stems from a certain type of data, such as video blurring leading to low confidence in visual analysis, then prioritize scheduling expert models that are good at handling other complementary data modalities, such as device status and acoustics, to make up for the evidence shortcomings.

[0129] Performance weighting: The matching score is weighted by combining dynamic performance indicators such as the success rate of each candidate model in a recent preset time, the consistency between the judgment result and the subsequent processing feedback;

[0130] Ultimately, the N models with the highest overall scores were selected to form the expert model group for this collaborative judgment.

[0131] S42: Distribute the feature vector sequence in parallel to each expert model in the expert model group, so that each expert model can perform sub-risk determination;

[0132] In this embodiment, the parallel distribution of the feature vector sequence to each expert model in the expert model group means that the original feature vector sequence is sent to each expert model in the expert model group at the same time. This can improve processing efficiency and enable each expert model to analyze the same risk scenario independently and simultaneously, thereby speeding up the risk assessment. Sub-risk assessment refers to each expert model in the expert model group independently analyzing the received feature vector sequence based on its own professional knowledge and training data, and outputting its local judgment result on the risk. The local judgment result may include refinement or supplementation of the risk type, risk location, risk evolution trend or dynamic risk level, as well as the local confidence of the expert model in its own judgment.

[0133] S43: Collect the sub-risk assessment results and their local confidence scores of each expert model, and use a pre-set Bayesian fusion model to resolve conflicts and fuse evidence in the sub-risk assessment results to generate the final assessment result.

[0134] In this embodiment, local confidence refers to the quantitative assessment of the reliability or accuracy of each expert model's own judgment result when making sub-risk judgments. It reflects the degree of understanding and judgment of a single expert model regarding the current risk situation. The preset Bayesian fusion model is a decision fusion method based on Bayesian theory. It can integrate evidence from multiple information sources and weight them according to the reliability of these evidences to arrive at a more comprehensive and reliable overall judgment. The Bayesian fusion model is pre-trained and configured to adapt to the characteristics of paper production safety risks. Conflict resolution and evidence fusion refer to the Bayesian fusion model identifying and handling possible contradictions or inconsistencies between sub-risk judgment results from different expert models when it receives them. For example, when different expert models give different judgments on risk types or levels, the Bayesian fusion model will make a reasonable ruling on the conflict based on its preset fusion strategy and the local confidence of each expert model. Evidence fusion refers to the integration of all valid evidence after conflict resolution and the calculation of the final risk judgment result and its overall confidence through Bayesian inference.

[0135] Specifically, when the risk perception model has low confidence in the preliminary judgment result, i.e., the uncertainty quantification score exceeds a preset score threshold, a collaborative judgment mechanism is triggered. Specifically, firstly, based on the intermediate judgment result, the N expert models with the highest matching degree are dynamically selected from a preset expert model pool to form an expert model group. This ensures that only the most relevant and professional expert models are activated, thereby avoiding unnecessary computational resource consumption and improving the targeting of subsequent judgments. Secondly, the original feature vector sequence is distributed in parallel to each expert model in the expert model group. Each expert model, based on its specific professional domain and training data, independently performs sub-risk judgments on the feature vector sequence and outputs its local confidence score. This parallel processing method can shorten the judgment time, thus enabling rapid response to high uncertainty risks. Finally, a preset Bayesian fusion model collects the sub-risk judgment results and their local confidence scores from all expert models, and performs conflict resolution and evidence fusion on this information. The Bayesian fusion model can weight the evidence according to the local confidence scores of each expert model, effectively handling possible judgment differences between different expert models, thereby generating a more comprehensive, accurate, and more confident final judgment result.

[0136] As a specific implementation method, when the risk perception model lacks confidence in the preliminary judgment result, N expert models can be dynamically selected from the expert model pool based on the intermediate judgment result. For example, an infrared image analysis model specifically for "temperature anomaly," an equipment status monitoring model for "motor current / voltage anomaly," and a rule reasoning model for "cooling system failure" can be selected to form an expert model group. Furthermore, the feature vector sequence containing infrared image data, motor operation data, and cooling system sensor data is distributed in parallel to these three expert models. The infrared image analysis model can determine whether there is a local high temperature. The system analyzes the temperature range and provides its local confidence level; the equipment condition monitoring model can analyze whether the motor is overloaded or has abnormal vibration and provide its local confidence level; the rule reasoning model can check whether the coolant flow or pressure is below the safety threshold and provide its local confidence level; finally, the sub-risk judgment results and local confidence levels are integrated through a preset Bayesian fusion model. For example, if the infrared image model and the equipment condition model both indicate high temperature and overload, and the rule reasoning model also finds abnormal coolant flow, the Bayesian fusion model will fuse these pieces of evidence according to their weights and ultimately determine that "motor overheating leads to a high risk of shutdown" and provide a final judgment result with high confidence.

[0137] Among them, the expert models in the expert model pool can be built, trained or acquired by technical personnel in the relevant technical field. Whether it is built or trained, they belong to conventional or public technologies in the relevant technical field, and their acquisition can be easily achieved by technical personnel in the relevant technical field. At the same time, the Bayesian fusion model is the same, and technical personnel in the relevant technical field can train it based on the characteristics of the paper production safety risk-related field.

[0138] Through the above technical solution, when the confidence level of the risk perception model in the preliminary judgment result is low, this application can conduct collaborative judgment by dynamically scheduling expert model groups, which can effectively avoid the limitations of single model judgment. Specifically, by utilizing the advantages of multiple specialized models, multi-dimensional and in-depth analysis of high uncertainty risks is carried out, and the judgments of various expert models are integrated through a Bayesian fusion model, thereby improving the accuracy, reliability and robustness of risk judgment. This enables the paper production safety management system to output more convincing final judgment results when facing complex, changeable or difficult-to-determine risk situations, effectively reducing the risk of misjudgment and omission, and ensuring production safety.

[0139] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0140] In one embodiment, a dynamic risk management system based on paper production safety management is provided. This dynamic risk management system corresponds one-to-one with the dynamic risk management method based on paper production safety management described in the previous embodiment. The dynamic risk management system based on paper production safety management includes:

[0141] The data acquisition module is used to acquire multi-source datasets of the target production area and construct a feature vector sequence based on the multi-source datasets. The multi-source datasets include equipment status data, video image data, operating environment data, and historical operation data.

[0142] The first determination module is used to input the feature vector sequence into the pre-trained risk perception model to obtain a preliminary determination result including risk type, risk location, risk evolution trend and dynamic risk level, and the preliminary determination result is associated with an uncertainty quantification score.

[0143] The second judgment module is used to perform logical consistency verification on the preliminary judgment result based on the preset security knowledge graph, and generate intermediate judgment result based on the result of the logical consistency verification.

[0144] The third determination module is used to compare the uncertainty quantification score with a preset score threshold. If the score threshold is not exceeded, the intermediate determination result is determined as the final determination result. If the score threshold is exceeded, a preset collaborative determination mechanism is triggered, and the final determination result is determined.

[0145] The process execution module is used to execute the corresponding preset risk handling process and learning optimization process based on the final judgment result.

[0146] Specific limitations regarding a dynamic risk control system based on paper production safety management can be found in the above description of a dynamic risk control method based on paper production safety management, and will not be repeated here. Each module in the aforementioned dynamic risk control system based on paper production safety management can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0147] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a dynamic risk control method based on paper production safety management.

[0148] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements a dynamic risk control method based on paper production safety management.

[0149] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A dynamic risk control method based on paper production safety management, characterized in that, Including the following steps: Obtain multi-source datasets of the target production area and construct a sequence of feature vectors based on the multi-source datasets, which include equipment status data, video image data, operating environment data, and historical operation data; The feature vector sequence is input into a pre-trained risk perception model to obtain preliminary judgment results including risk type, risk location, risk evolution trend and dynamic risk level, and the preliminary judgment results are associated with uncertainty quantification scores; The preliminary judgment results are logically consistent based on a pre-set security knowledge graph, and intermediate judgment results are generated based on the results of the logical consistency check. The uncertainty quantification score is compared with a preset score threshold. If the score threshold is not exceeded, the intermediate judgment result is determined as the final judgment result. If the score threshold is exceeded, a preset collaborative judgment mechanism is triggered, and the final judgment result is determined. Based on the final judgment result, the corresponding preset risk handling process and learning optimization process will be executed.

2. The dynamic risk control method based on paper production safety management according to claim 1, characterized in that: The step of acquiring a multi-source dataset of the target production area and constructing a feature vector sequence based on the multi-source dataset, wherein the multi-source dataset includes equipment status data, video image data, operating environment data, and historical operating data, includes the following steps: Determine the target time window, obtain multi-source datasets of the target production area based on the target time window, and perform timestamp alignment and spatial coordinate mapping on the multi-source datasets to generate a spatiotemporally aligned dataset; Feature extraction and encoding are performed on the spatiotemporally aligned dataset to obtain a feature vector set, and a feature vector sequence is generated based on the feature vector set. The feature vector set includes state feature vectors, visual feature vectors, working environment feature vectors, and background knowledge feature vectors.

3. The dynamic risk control method based on papermaking production safety management according to claim 1, characterized in that: The risk perception model includes a scene understanding layer, a feature fusion layer, and a judgment output layer. The step of inputting the feature vector sequence into the pre-trained risk perception model to obtain preliminary judgment results including risk type, risk location, risk evolution trend, and dynamic risk level, and whereby the preliminary judgment results are associated with an uncertainty quantification score, includes the following steps: The scenario understanding layer identifies the type of work scenario corresponding to the target production area based on the received feature vector sequence; Based on the job scenario type information, the feature fusion layer performs cross-modal weighted correlation analysis on the feature vector sequence to identify potential coupling risk signals; The output layer determines the risk type, risk location, risk evolution trend, and dynamic risk level based on potential coupled risk signals, and simultaneously generates an uncertainty quantification score that represents the confidence level of the determination.

4. The dynamic risk control method based on paper production safety management according to claim 3, characterized in that: The feature fusion layer, based on job scenario type information, performs cross-modal weighted correlation analysis on the feature vector sequence to identify potential coupling risk signals, including the following steps: Based on the pre-set attention mechanism and job scenario type information, the correlation weights between each feature vector in the feature vector set are calculated. Based on the calculated relevance weights, the feature vectors of different modalities in the feature vector set are weighted and fused to generate a risk scenario representation. Based on risk scenario representation, risk correlation reasoning is performed to match the information of the type of execution scenario in order to identify potential risk signals.

5. The dynamic risk control method based on paper production safety management according to claim 4, characterized in that: The step of generating risk type, risk location, risk evolution trend, and dynamic risk level based on potential coupled risk signals in the output layer, and simultaneously generating an uncertainty quantification score representing the confidence level of the judgment, includes the following steps: Based on potential coupling risk signals, multimodal evidence and risk types are determined, and the support score and reliability weight of multimodal evidence are identified. The overall evidence confidence level for the current risk assessment is calculated by weighting and summing the support scores of the multimodal evidence according to their reliability weights. Based on the risk type, a dynamic risk level is generated by matching its corresponding pre-set consequence severity coefficient and time urgency coefficient, and combining the comprehensive evidence confidence level through a pre-set risk quantification function. Conflict identification and completeness assessment are performed on multimodal evidence, and uncertainty quantification scores are determined based on the results of conflict identification and completeness assessment.

6. The dynamic risk control method based on paper production safety management according to claim 1, characterized in that: The step of performing logical consistency verification on the preliminary judgment result based on a preset security knowledge graph, and generating intermediate judgment results based on the results of the logical consistency verification, includes the following steps: Identify risk entities in the preliminary judgment results, and perform semantic matching and association binding between risk entities, risk locations and corresponding entity nodes, activity nodes and spatial nodes in the security knowledge graph, thereby activating the risk dynamic subgraph; Based on the risk dynamic subgraph, traversal and reasoning calculations are performed along predefined rule paths to detect whether there are logical anomalies in the preliminary judgment results; If a logical anomaly is detected, the corresponding rule chain and graph evidence node will be located. The preliminary judgment result is corrected based on the located rule chain, and verification marks are added to the corrected preliminary judgment result based on the rule chain and the graph evidence nodes to generate intermediate judgment result.

7. The dynamic risk control method based on paper production safety management according to claim 1, characterized in that: The step of comparing the uncertainty quantification score with a preset score threshold, and determining the intermediate judgment result as the final judgment result if the score does not exceed the threshold, and triggering a preset collaborative judgment mechanism and determining the final judgment result if the score exceeds the threshold, includes the following steps: If the uncertainty quantification score exceeds the score threshold, then based on the intermediate judgment result, the N expert models with the highest matching degree are dynamically scheduled from the preset expert model pool to form an expert model group. The feature vector sequence is distributed in parallel to each expert model in the expert model group, so that each expert model can perform sub-risk determination. Collect the sub-risk assessment results and their local confidence scores from each expert model, and use a pre-set Bayesian fusion model to resolve conflicts and fuse evidence in the sub-risk assessment results to generate the final assessment result.

8. A dynamic risk control system based on paper production safety management, characterized in that, include: The data acquisition module is used to acquire multi-source datasets of the target production area and construct a feature vector sequence based on the multi-source datasets. The multi-source datasets include equipment status data, video image data, operating environment data, and historical operation data. The first determination module is used to input the feature vector sequence into the pre-trained risk perception model to obtain a preliminary determination result including risk type, risk location, risk evolution trend and dynamic risk level, and the preliminary determination result is associated with an uncertainty quantification score. The second judgment module is used to perform logical consistency verification on the preliminary judgment results based on the preset security knowledge graph, and generate intermediate judgment results based on the results of the logical consistency verification. The third determination module is used to compare the uncertainty quantification score with a preset score threshold. If the score threshold is not exceeded, the intermediate determination result is determined as the final determination result. If the score threshold is exceeded, a preset collaborative determination mechanism is triggered, and the final determination result is determined. The process execution module is used to execute the corresponding preset risk handling process and learning optimization process based on the final judgment result.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the dynamic risk control method based on paper production safety management as described in any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the dynamic risk control method based on paper production safety management as described in any one of claims 1-7.

Images