An AI security large model-based hazardous chemical operation risk prevention and control device and method

Abstract

The application discloses a dangerous chemical product operation risk prevention and control device and method based on an AI security large model, and particularly relates to the technical field of dangerous chemical product operation safety risk prevention and control, and is used for solving the problems that in the existing dangerous chemical product loading and unloading, tank area reverse loading and limited space operation, risk identification relies on artificial experience, monitoring data is scattered, and it is difficult to establish a unified model and a safety baseline on the operation unit granularity. Through taking the dangerous chemical product operation ticket as a starting point, an operation unit identifier is generated at the operation unit granularity, a digital twin model covering equipment, pipelines, operation steps and safety anchor points is constructed, under the condition of unified time service, multi-source monitoring records are converged to extract safety events to form a safety evidence graph, and the safety evidence graph is handed over to a safety large model to output risk assessment results and on-site control instructions under the constraint of safety anchor points, so that traceable, quantifiable and regulation-constrained risk identification and intervention decision-making are realized in the whole process of a single dangerous chemical product operation, and the risk of missed judgment and over-regulation execution is reduced.

Description

Technical Field

[0001] This invention relates to the field of safety risk management and control technology for hazardous chemical operations, specifically to a device and method for risk prevention and control of hazardous chemical operations based on an AI-based safety big data model. Background Technology

[0002] Currently, hazardous chemical loading and unloading, tank farm transfer, and confined space cleaning operations are generally managed through work permit management systems, process monitoring systems such as DCS / ESD, portable / fixed gas detectors, and video surveillance systems. Work permits are typically only used for manual approval and archiving, while on-site monitoring data is recorded sporadically in their respective systems by unit or area, lacking a systematic correlation at the level of a single operation. Monitoring personnel rely more on experience to interpret alarm information such as combustible gas concentration, oxygen content, pressure, and temperature, making it difficult to conduct timely and comprehensive risk assessment and review by combining specific work content, work steps, and personnel behavior.

[0003] With the advancement of information technology construction, some enterprises have introduced technologies such as work ticket systems, intelligent video analysis, and rule engines to automatically warn of typical violations and over-limit working conditions. However, existing technologies mostly remain at the level of "point monitoring + threshold alarm": on the one hand, monitoring data is usually stored by device, unit, or region, failing to model and collect data around "a complete work process corresponding to a certain work ticket," and unable to reconstruct equipment connectivity, work step sequence, and corresponding monitoring records at the work unit level; on the other hand, regulations and clauses often exist in the form of text or simple configuration items, making it difficult to solidify safety conditions into quantifiable and traceable constraint units in the system, and the correspondence between them and specific measurement points and work steps is unclear, making it difficult to form an interpretable risk evidence chain from the level of "which condition was violated, in which step, and by which monitoring signals."

[0004] Furthermore, existing risk warning attempts based on big data statistics or general algorithm models mainly rely on macro-level indicators such as historical alarm counts and accident statistics. They lack a structured representation of the risk evolution process across different time segments under the same work order, making it difficult to form a closed loop with pre-operation access conditions, in-operation intervention measures, and post-operation hazard rectification records. When monitoring signals are missing, timestamps drift, or networks are interrupted, existing systems generally only issue brief alarms and have not established a unified strategy for risk level adjustment, manual inspection prompts, and subsequent audit evidence collection. In summary, the core problem that existing technologies cannot solve is that they cannot, in hazardous chemical operation scenarios, integrate multi-source monitoring data, work process information, and safety procedure constraints into a unified model and correlation analysis at the granularity of a single operation corresponding to a work order, under unified time synchronization conditions. This would prevent the achievement of traceable, quantifiable, and compliant risk identification and intervention support throughout the entire process of work preparation, execution, and closure. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a risk control device and method for hazardous chemical operations based on an AI-based safety model, in order to solve the problems mentioned in the background section.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for risk prevention and control in hazardous chemical operations based on an AI-based safety big data model, comprising: S1. Generate a work unit identifier based on the hazardous chemicals work order, construct a digital twin model of the work unit according to the field structure, parse the safety anchor points from the safety regulations and bind them to the digital twin model; S2. Collect data from related work units under unified time synchronization, collect and extract safety events according to work unit identifiers, and associate safety event nodes and safety anchor point nodes to construct a safety evidence graph. S3. Input the operation digital twin model and safety evidence map into the safety big model, encode the safety evidence map, decode and output the risk assessment results, intervention suggestions and evidence node index under the constraint of safety anchor points; S4. Before the start of the operation, construct the disturbance conditions based on the digital twin model and input them into the safety model. Compare the risk assessment of the disturbance conditions to identify risk paths and generate access instructions and safety measure instructions accordingly. S5. Update the safety evidence diagram during operation, input the evidence image segments into the safety big model to obtain risk assessment, and convert intervention suggestions that meet the intervention trigger conditions into on-site control instructions through safety anchor point verification. S6. After the operation is completed, collect event records, organize the safety evidence diagram and risk assessment results to form a risk evidence chain, align the risk evidence chain with the records to generate training samples, which are used for incremental training of the safety large model and to update the safety anchor points.

[0007] Furthermore, S1 includes: A work unit identifier is generated from the hazardous chemicals work order. The work unit identifier consists of the work order number, the equipment area code, and the planned start and end times. The system identifies the equipment and pipelines involved in the operation from the equipment ledger and process flow card based on the operation unit identifier, extracts material connection relationships and process step sequences, and organizes them according to unified names and units of measurement to build a digital twin model containing equipment nodes, pipeline connection relationships, operation step nodes and rated operating parameters. On-site monitoring and control records carry work unit identifiers when they are generated.

[0008] Furthermore, the system retrieves clauses from the safety regulations database based on the operation type, medium hazard description, and equipment characteristics in the hazardous chemical operation ticket. Conditions with hard constraints are broken down into safety anchors. Each safety anchor records the anchor name, applicable operation type, applicable medium range, measurement object identifier, measurement point description, unit of measurement, allowable value range, required continuous compliance time, and subsequent measure level. These records are then associated with equipment nodes, area nodes, and operation step nodes in the digital twin model, forming a safety baseline that is stored with operation unit identifiers and safety anchor set corresponding to each other. The safety anchor version is then linked to the configuration record.

[0009] Furthermore, S2 includes: Under unified time synchronization, records of work unit identification and time stamps are collected from process monitoring devices, environmental monitoring devices, personnel wearing identification devices, and video intelligent identification devices; The system identifies safety events corresponding to changes in safety anchor point boundaries and process constraints based on the observation window. It registers these safety events as safety event nodes and establishes directed associations with equipment nodes, area nodes, work step nodes, and safety anchor point nodes in the digital twin model according to their temporal order, spatial adjacency, and safety procedure dependencies, thus forming a safety evidence graph.

[0010] Furthermore, S3 includes: The risk prevention and control service selects digital twin models and safety evidence diagrams based on the work unit identifiers, and submits the equipment, pipelines, areas, work steps, safety anchor points, safety events and directed relationships into a large safety model code; Under the constraints of safety anchor points, the safety big model generates risk assessment results, intervention suggestions, and evidence node indexes, and writes the risk assessment results, intervention suggestions, evidence node indexes, work unit identifiers, safety anchor point version tags, digital twin model version tags, and safety evidence graph version tags into the risk assessment log.

[0011] Furthermore, S4 includes: Before the start of operations, the risk control service generates a candidate set of original and disturbed operating conditions based on the digital twin model of the work unit and safety anchor points. The candidate set of operating conditions is input into the safety model to obtain risk assessment results under the constraint of safety anchor points; Based on the risk assessment results, risk paths and sensitive risk channels are identified, and access instructions and security measure instructions are generated. The enterprise operations management system displays access instructions and safety measure instructions and associates them with work unit identifiers, safety anchor version tags, and digital twin model version tags.

[0012] Furthermore, S5 includes: During the operation, the risk control service divides the time into observation windows and uses the safety evidence diagram and digital twin model of the operation unit to call the large safety model to obtain the risk assessment results and intervention suggestions for the time slice; The risk prevention and control service selects intervention suggestions based on preset intervention trigger conditions and safety anchor point rules, generates on-site control instructions containing work unit identifier, time slice number, target object identifier, action type and action parameters, issues them through the control channel, and records the execution results of the on-site control instructions to the control log based on idempotent flags and status flags.

[0013] Furthermore, S6 includes: After a single hazardous chemical operation is completed, the risk control service obtains event records from the operation management system and the on-site monitoring system based on the operation unit identifier, and arranges the event records, safety evidence map evolution trajectory and risk assessment results of each time segment to form a risk evidence chain. The risk prevention and control service selects training samples from multiple risk evidence chains based on the sample selection strategy, performs incremental training on the security big model in an offline environment, and generates security anchor point update suggestions. After confirmation by the safety management department, the security anchor version mark is updated and associated with the relevant risk evidence chain record.

[0014] On the other hand, the present invention provides a risk prevention and control device for hazardous chemical operations based on an AI-based safety big data model, comprising: The operation modeling module is used to generate operation unit identifiers based on hazardous chemical operation tickets, and to construct digital twin models of operation units corresponding to the operation unit identifiers according to the field structure. It also parses safety anchor points from safety regulations and binds them to the digital twin models. The data acquisition and security event extraction module is used to collect data carrying work unit identifiers under unified time synchronization conditions, aggregate the collected data according to the work unit identifiers and extract security events, and associate security event nodes with security anchor point nodes to construct a security evidence graph. The large-scale model analysis module is used to input the digital twin model of the work unit and the safety evidence map into the large-scale safety model, encode the safety evidence map, and decode it under the constraint of safety anchor points to obtain risk assessment results, intervention suggestions and evidence node indexes; The disturbance condition assessment module is used to construct disturbance conditions based on the digital twin model of the work unit before the start of the operation and input the disturbance conditions into the safety model. It compares the risk assessment results corresponding to the disturbance conditions to identify risk paths and generates access instructions and safety measure instructions based on the risk paths. The online intervention control module is used to update the safety evidence map corresponding to the work unit during the operation, input the safety evidence image segment into the safety big model to obtain the risk assessment result, and verify the intervention suggestions that meet the intervention trigger conditions under the safety anchor point constraint and convert them into on-site control instructions. The evidence chain management and incremental training module is used to collect event records related to the work unit after the operation is completed, organize the safety evidence graph and risk assessment results to form a risk evidence chain, align the risk evidence chain with the event records to generate training samples, and perform incremental training on the safety big model and update the safety anchor points based on the training samples.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. By generating work unit identifiers at the work unit level starting from the hazardous chemical work permit, a digital twin model covering equipment, pipelines, work steps, and safety anchor points is constructed. Under unified time synchronization, multi-source monitoring records are aggregated to extract safety events and form a safety evidence map, which is then handed over to the large safety model to output risk assessment results and verified on-site control instructions under the constraints of safety anchor points. This achieves traceable, quantifiable, and safety-compliant risk identification and intervention decisions throughout the entire process of preparation, execution, and completion of a single hazardous chemical operation, reducing the risks of missed judgments, misjudgments, and executions outside of regulations.

[0016] 2. By constructing disturbance conditions and identifying risk paths and sensitive risk channels based on a digital twin model before the start of the operation, and generating access instructions and safety measure instructions with constraints, and aligning the safety evidence map evolution trajectory, risk assessment results, on-site control behaviors and event records after the operation, a risk evidence chain is formed to drive incremental training of the safety big model and generate safety anchor point update suggestions. This achieves a closed-loop optimization mechanism that uniformly realizes operation access assessment, online risk control and post-event experience accumulation under different equipment and operation types, and continuously improves the adaptability and overall level of risk prevention and control in hazardous chemical operations. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a method for risk prevention and control in hazardous chemical operations based on an AI-powered safety model, as described in this invention. Figure 2 This is a schematic diagram of the structure of a hazardous chemical operation risk prevention and control device based on an AI safety big data model according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1: Figure 1 A flowchart illustrating a method for risk prevention and control in hazardous chemical operations based on an AI-powered safety model is provided. This method includes: S1. Generate a work unit identifier based on the hazardous chemicals work order, construct a digital twin model of the work unit according to the field structure, parse safety anchor points from safety regulations and bind them to the digital twin model. The specific implementation is as follows: In actual hazardous chemical loading and unloading, tank farm transfer, and confined space cleaning operations, to achieve traceable and reproducible risk control at the individual operation level, this method uses the hazardous chemical work permit as a starting point to uniquely identify and structurally describe each operation. The hazardous chemical work permit is a permit document filled out and approved by the company in accordance with national and industry standards before starting on-site operations. It typically includes information such as the type of operation, the name and hazard description of the involved medium, the operating equipment and area, the planned start and end times, a list of involved equipment, pipelines and valves requiring isolation, mandatory testing items, and requirements for monitoring personnel. The hazard description of the medium can be set to reference fields such as flash point, explosion limits, and toxicity level from the safety data sheet, used to subsequently determine monitoring targets and safety boundaries.

[0020] In this method, a work unit refers to a complete work process corresponding to a hazardous chemical work permit, which is completed within a specified time and area. The work unit identifier is a unique code string generated for the work unit, which is used for information association and repeated call control throughout the entire work cycle in the system. It is preferably formed by a combination of work permit number, device area code and planned start and end time. When there are multiple work permits in the same device and time period, a sequential number can be added to maintain uniqueness. All kinds of on-site monitoring records and control records carry the work unit identifier when they are generated, which is used for subsequent collection by work unit.

[0021] After obtaining the work unit identifier, the system retrieves basic information related to the work from the equipment ledger and process flow card. The equipment ledger is usually maintained by the equipment management department and records fields such as equipment name, equipment number, installation location, rated pressure, rated temperature, volume, material, and maintenance status for each container, pump, valve, heat exchanger, and auxiliary pipeline. The process flow card is maintained by the process department and describes the flow of each material, the opening and closing sequence of the main valves, normal operating conditions, and start-up and shutdown procedures. If necessary fields are missing from the equipment ledger or process flow card, the system can prompt on-site engineering technicians to supplement the data by referring to existing drawings and manuals to ensure the completeness of subsequent descriptions.

[0022] Based on the equipment, area, and equipment numbers listed in the hazardous chemicals work permit, the system filters out the equipment and pipelines involved in the operation from the equipment ledger, extracts the material connectivity relationships and planned process step sequences related to the operation from the process flow card, where material connectivity relationships refer to the physical relationship between the outlet of one piece of equipment and the inlet of another piece of equipment through pipelines, and process step sequences refer to the order of each step under normal operating conditions and the correspondence between steps and equipment; the system organizes the above information according to unified names and units of measurement, for example, converting different pressure units used on site into unified units, mapping the textual description of the installation location to a standardized area code, and mapping the textual description of the process steps to step identifiers with sequence numbers.

[0023] Based on this, a digital twin model of the work unit is constructed. In this method, the digital twin model is a set of records that structurally represent the equipment, pipelines, containers, work steps and expected operating conditions involved in this work. It at least indicates the name, number, location and rated operating parameters of each equipment node, the upstream and downstream equipment of each connection relationship and the requirement that it should be in an isolated or connected state in this work. It also sets the association between each work step and the equipment, medium and expected state, and can also set fields to indicate the progress and status of the current step, so that the field monitoring results and behavior records can be linked to specific equipment and steps in the future.

[0024] To ensure that the risk assessment process has a clear benchmark, this method relies on safety regulations to summarize and solidify the key conditions that need to be followed in this operation. Safety regulations include clauses in national and local laws and regulations on work safety, industry standards, and internal operating procedures of enterprises that are directly related to hot work, confined space operations, tank loading and unloading, such as requirements for combustible gas concentration, oxygen content, temperature, pressure, ventilation time, operation duration, number of personnel, and protective equipment.

[0025] Based on the operation type, medium hazard description, and equipment characteristics in the hazardous materials operation permit, the system retrieves the corresponding clauses from the safety regulations database. Each condition that has a hard constraint on on-site safety is broken down into a safety anchor point. In this method, a safety anchor point is a record used to describe a single safety constraint, which includes at least the anchor point name, applicable operation type and medium range, associated measurement object identifier, measurement point or area description, unit of measurement used, allowable upper and lower limits or range values, minimum time required to continuously meet the constraint, and the level of subsequent measures to be triggered if the constraint is violated. Measurement objects include quantifiable objects such as gas detection points in the work area, pressure or temperature measuring points on equipment, and the number of personnel and the status of protective equipment wearing.

[0026] The system establishes a connection between safety anchor points and specific equipment nodes, area nodes, or work step nodes in the digital twin model, forming a safety baseline covering the entire work unit. The safety baseline can be understood as the set of safety conditions that the work must meet in each stage of preparation, execution, and completion. It is stored in the system as a correspondence between work unit identifiers and the set of safety anchor points.

[0027] To facilitate subsequent traceability and iteration, each adjustment to safety procedure clauses, numerical limits, or scope of application will generate a new safety anchor version in the system and assign a version tag. The version tag and work unit identifier are stored together in the configuration record and recorded along with the generated risk evidence chain, enabling reviewers to clearly understand under what safety baseline constraints a particular operation was completed. Preferably, in the tank farm loading scenario, the work unit identifier can be set as a fixed-length coded string. The beginning corresponds to the plant and equipment area, the middle corresponds to the operation date and planned start time, and the end is used to distinguish multiple work tickets within the same time period. The rated pressure recorded for each container node in the digital twin model can be set to be measured in a unified unit and retain a certain number of digits; the rated temperature can be set to be measured in degrees Celsius and retain an integer; the volume can be set to be measured in cubic meters and retain a certain number of digits; and the conditions for combustible gas detection in the safety anchor can be set to... Before hot work, the volume fraction of combustible gas at the corresponding location shall not exceed the upper limit determined by the enterprise according to the safety technical specifications and regulations, and the detection results shall remain below the upper limit for several consecutive minutes. For confined space operations, the conditions can be set as follows: the oxygen content in the space is maintained within a certain range, the ventilation time is not less than a certain duration, and the detection interval during the operation is not higher than a certain rhythm. The above numerical ranges are given by the enterprise in combination with the characteristics of the equipment and regulatory requirements. Once determined, they are locked through the safety anchor version mark. Based on this, those skilled in the art can select specific values ​​on different devices and achieve the same modeling capability. As long as the digital twin model is used to fully reflect the equipment and steps involved in this operation, and the capability and scope of application of the safety anchor to solidify the safety baseline in a measurable form and bind it to the work unit identifier remains unchanged, it can be ensured that the subsequent risk assessment, intervention decision and evidence chain solidification links are all carried out with clear and consistent technical processing around the same work unit.

[0028] S2. Under unified time synchronization, collect data from associated work units, aggregate and extract safety events according to work unit identifiers, and construct a safety evidence graph by associating safety event nodes and safety anchor nodes. The specific implementation is as follows: In the preparation and execution phases of a task, to continuously monitor the site conditions at the individual task level and provide a reliable basis for subsequent risk assessment, this method collects information related to the task unit from process monitoring devices, environmental monitoring devices, personnel-worn identification devices, and video intelligent identification devices under unified time synchronization conditions. Unified time synchronization conditions mean that the aforementioned devices within the plant area are time-stamped under the same time source constraint. The time source is preferably a plant-level clock or uniformly distributed via a time synchronization gateway. After synchronization, each device adds time stamps to the collected records with second-level precision, ensuring the comparability of the chronological order of information from different sources on the timeline.

[0029] Process monitoring devices refer to measuring devices installed on containers, pipelines, and pumps / valvees involved in the digital twin model, used to acquire process monitoring values ​​such as temperature, pressure, liquid level, and flow rate. Temperature is preferably measured in degrees Celsius, pressure is preferably measured in a unified pressure unit, liquid level is preferably measured as a percentage or liquid column height, and flow rate is preferably measured as volumetric flow rate or mass flow rate. Environmental monitoring devices refer to combustible gas detectors, toxic gas detectors, and oxygen content detectors arranged in the work area, used to collect environmental monitoring values ​​such as combustible gas concentration, toxic gas concentration, and oxygen content, which are preferably measured as volume fractions. The system includes the detection point location and its area code; personnel identification devices refer to the identification tags and positioning devices worn by workers and supervisors, which are used to identify the wearing status of personal protective equipment, such as whether they are wearing air respirators, explosion-proof walkie-talkies, protective clothing, etc., and also to record the location information and duration of stay of personnel in the area; video intelligent identification devices refer to the camera devices and their associated image analysis functions deployed near the work area, which are used to identify behaviors and phenomena such as smoking, open flames, abnormal gathering of people, suspected leakage signs, and people falling in the video footage, and output them in the form of events.

[0030] When the above-mentioned devices are running on site, they will send records to the risk control service according to a preset rhythm. The records will include at least the work unit identifier, time stamp, device or personnel identifier, measurement value or event type and quality mark. The quality mark is used to indicate whether the record was obtained when the device passed the self-test and whether there were any situations such as over-range or signal interruption.

[0031] The risk control service merges all records into the corresponding work unit name based on the work unit identifier at the receiving end, and sorts them by time stamp. During the sorting process, records that are significantly outside the work plan time range can be removed or marked separately to ensure that subsequent analysis is carried out around the actual time window of this work.

[0032] In this method, to facilitate the identification of safety-related state changes within a limited time, the concept of an observation window is set. The observation window refers to a continuous time segment selected on the time axis. Within this segment, the system comprehensively considers various monitoring records and the progress of work steps to determine whether a safety event has occurred. The length and sliding step of the observation window are uniformly set according to the device characteristics, the hazard of the medium, and the type of work when the system is deployed.

[0033] In this method, a safety incident refers to a change in state or behavior that has a significant impact on on-site safety. This includes events where the value crosses a certain safety anchor limit, such as the concentration of combustible gas rising from below a certain control value to near or above the upper limit set by the safety anchor. It also includes events where the work process deviates from expectations, such as skipping to subsequent steps without confirming the work steps in the order recorded in the digital twin model. Furthermore, it includes personnel behavior events, such as personnel entering an area marked as unauthorized without permission, a supervisor leaving the supervised area for more than the permitted time, and workers not wearing necessary protective equipment as required by the safety anchor.

[0034] Within each observation window, the system compares and combines monitoring values ​​and behavior records belonging to that work unit and whose time markers fall within the observation window according to pre-set judgment rules. If the change satisfies the boundary conditions or process constraints in the safety anchor point, it is registered as a safety event. When registering a safety event, information such as the work unit identifier, time marker, identifier of the measurement object or personnel involved, associated safety anchor point markers, and event type are attached and added to the safety evidence set of this operation in the form of a safety event node.

[0035] The equipment nodes, area nodes, and work step nodes defined in the digital twin model, along with the aforementioned safety anchor nodes and safety event nodes, together constitute the node set of the safety evidence graph. To reflect the relationships between nodes, this method further establishes directed associations between nodes based on temporal sequence, spatial adjacency, and dependencies in safety procedures. Temporal sequence refers to the sequential relationship between safety event nodes and work step nodes that occur sequentially within the same work unit. Spatial adjacency refers to the adjacency relationship between nodes located in the same equipment, the same area, or directly affected by material connectivity. Dependencies in safety procedures refer to the dependency relationship where a safety anchor requires a certain condition to be met before subsequent steps can be performed, or where a certain anomaly must trigger specific subsequent measures. Through these directed associations, the safety evidence graph can reflect the subsequent risk chains that a safety event may trigger and which equipment, areas, or steps constrain it.

[0036] The resulting safety evidence map is a set of records containing nodes and their directed associations, with the work unit as the scope. It includes work unit identifiers and version tags. The version tags are consistent with the digital twin model and safety anchor version, and are used to indicate the equipment topology, work process and safety baseline version used at that time.

[0037] The security evidence graph is stored persistently in the risk control service, preferably in a storage system optimized for time-series and relational queries. It records the generation time, source device, and association rule number of security event nodes and relations, so that when the security big model is called for risk assessment later, the evidence context of a certain moment can be quickly reconstructed according to the work unit and time segment. It also facilitates the backtracking of the event evolution process during the entire operation period according to the work unit identifier after the operation is completed. In this method, the risk control service refers to the set of risk management service modules deployed on the enterprise side, which is used to receive on-site monitoring records and behavior records, maintain the digital twin model of the work unit, the security evidence graph, the risk assessment log and the control log, and interact with the security big model.

[0038] Preferably, in a petrochemical enterprise tank area loading and confined space cleaning scenario, the process monitoring device can be set to report monitoring values ​​at a rhythm of tens of seconds to several minutes, the environmental monitoring device can be set to report at a shorter rhythm to promptly capture rapid changes in combustible and toxic gases, the personnel-worn identification device can be set to report position and status at a near real-time rhythm, the video intelligent recognition device can be set to report identified behavioral events in a continuous event stream, the observation window length can be set to several minutes, the sliding step size can be set to the minute level, and the proximity to the safety anchor point appears multiple times within several consecutive observation windows. When the monitoring value of the upper limit is exceeded or when process violations or regional violations occur repeatedly, they are all registered as safety events. The corresponding nodes and directed relationships in the safety evidence map are fully recorded. Based on this, technicians familiar with the site conditions can select specific rhythms and observation window lengths for different devices and operation types. However, the capability and scope of application of unified time synchronization for data collection, aggregation by operation unit identification, extraction of safety events within the observation window based on safety anchors and process rules, and construction of safety evidence maps with time and space dependencies remain unchanged. This provides a public, sufficient, and traceable site basis for subsequent risk assessment based on the safety big model.

[0039] S3. Input the operational digital twin model and safety evidence map into the large safety model, encode the safety evidence map, and decode and output the risk assessment results, intervention suggestions, and evidence node index under the constraint of safety anchor points. The specific implementation is as follows: When it is necessary to conduct risk assessment for a single hazardous chemical operation, the risk prevention and control service, driven by the aforementioned operation unit identifier, selects the digital twin model of the operation unit corresponding to the current moment and the safety evidence map formed within the current moment or a designated observation window. The attribute information of the equipment, pipelines, areas, operation steps, safety anchors, and safety events described therein, as well as their directed correlations, are submitted to the safety big model for assessment.

[0040] In this method, the safety big model refers to a type of multi-layer nonlinear mapping model obtained through offline training based on historical work cases and risk evidence chains. This model takes the work unit as the basic object and can determine the risk level and risk type of the current work status given a digital twin model and safety evidence map. It also provides a combination of on-site handling actions that match the risk and a set of key evidence node markers to support the determination. During the offline training phase, the enterprise collects risk evidence chains from multiple work units and uses the digital twin model, typical safety evidence image segments, and actual accident or hazard records as references. By repeatedly adjusting the internal parameters of the model, the risk judgment obtained by the safety big model under similar working conditions is made as close as possible to historical experience and safety regulations.

[0041] During the online assessment phase, when risk assessment conditions are triggered during the work preparation or execution phase, the risk control service sends the process structure, equipment connectivity, and work step sequence reflected in the digital twin model of the current work unit, along with the multi-source monitoring events, safety event nodes, and their temporal, spatial, and safety procedure dependencies reflected in the safety evidence diagram, to the safety big model. The safety big model first forms a comprehensive representation of the overall state of the current operation internally, and then, combined with the currently effective safety anchor constraints, determines the risk level and risk type of the current operation at the current moment, forming a risk assessment result. In this method, the risk assessment result refers to the record used to describe the current risk level and the main risk source categories. The risk level is preferably divided into several fixed levels, and the risk source categories preferably include categories that match the on-site working conditions, such as leakage, deflagration, asphyxiation, poisoning, and falls.

[0042] Based on the risk assessment results, the safety big model further deduces the combination of actions that should be implemented on-site to reduce the risk while adhering to the safety anchor points, forming intervention recommendations. In this method, intervention recommendations refer to the set of operational instructions for on-site control systems, monitoring personnel, and operators, including at least content that can be directly translated into on-site operations such as stopping or slowing down the handling of a certain material, adjusting ventilation measures, restricting access to a certain area, increasing monitoring efforts, or extending the frequency of inspections.

[0043] To ensure the interpretability and verifiability of risk assessments, the security big data model selects several nodes with the greatest impact on the current assessment from the security evidence graph as supporting evidence when generating risk assessment results and intervention recommendations. The markings of these nodes are compiled into an evidence node index. In this method, the evidence node index refers to a set of identifiers used to trace which device monitoring values, security events, and security anchors jointly supported the current risk assessment and intervention decision during subsequent presentation or auditing.

[0044] To prevent recommendations that violate procedures from entering the execution chain, the safety big data model compares them with the currently effective set of safety anchors before and after generating intervention recommendations. Actions that conflict with the provisions of the safety anchors are eliminated or replaced. For example, it prohibits continuing to recommend feeding operations when the concentration of combustible gas is close to the upper limit. For situations that are determined to be of a high-risk level but have insufficient or inconsistent evidence nodes within the current observation window, the safety big data model can mark the risk assessment results of this round as a risk uncertainty state and generate conservative intervention recommendations that are not directly converted into on-site control instructions. Conservative intervention recommendations include increasing the frequency of detection, keeping the current operation steps unchanged, suspending the execution of subsequent steps, increasing manual review, and increasing on-site monitoring.

[0045] After each risk assessment, the risk control service writes the risk assessment results, intervention suggestions, and evidence node indexes provided by the safety big data model, along with the corresponding work unit identifier, safety anchor version marker, digital twin model version marker, and safety evidence graph version marker, into the risk assessment log. In this method, the risk assessment log refers to a collection of logs recording the risk judgment trajectory at all assessment moments for a single operation. It includes at least the assessment time, model version marker, risk level, risk type, intervention suggestion summary, and evidence node index, and is persistently stored in a structured and searchable format. This provides a complete basis for subsequent intervention execution in later stages of the operation, the solidification of the risk evidence chain in the post-operation stage, and the incremental training of the safety big data model. Preferably, in a petrochemical tank loading scenario, the enterprise can set the risk level to four levels. When the safety big data model detects multiple consecutive observation windows in the safety evidence graph showing a combination of safety event nodes such as combustible gas concentration approaching the upper limit of the safety anchor point, personnel gathering on the loading platform, and work step jumps, it classifies the current operation as a third-level risk. The risk type is classified as deflagration risk, and intervention recommendations are generated. These recommendations include suspending the current loading operation, closing relevant feed valves, strengthening platform ventilation, and organizing the temporary evacuation of personnel. Simultaneously, the corresponding combustible gas detection nodes, operation step nodes, and personnel location nodes are marked in the evidence node index. The risk assessment log records the risk level, risk type, key points of the intervention recommendations, evidence node index, and the version of the safety anchor point used at that time. Those skilled in the art can adjust the risk level classification, risk source category, and intervention recommendations according to different devices and operation types. However, when risk assessment is required, the digital twin model and safety evidence map are integrated into a large safety model, using the operation unit as the basic object. Under the constraints of the safety anchor point, an assessment record containing risk assessment results, intervention recommendations, and evidence node indexes is formed. Version locking and traceability are maintained through the risk assessment log, ensuring that this process capability and scope of application remain unchanged. This guarantees that this method can achieve risk identification and decision support through a unified technical chain in different hazardous chemical operation scenarios.

[0046] S4. Before the operation begins, construct disturbance conditions based on the digital twin model and input them into the large safety model. Compare the risk assessment of the disturbance conditions to identify risk paths, and generate access instructions and safety measure instructions accordingly. The specific implementation is as follows: Before the actual hazardous chemical operation begins, in order to identify the risk-sensitive combination of operating conditions in advance without changing the actual operating status of the equipment on site, this method generates the original operating conditions and potential disturbance operating conditions based on the aforementioned constructed digital twin model. In this method, the original operating conditions refer to the planned operating conditions determined according to the hazardous chemical operation ticket, equipment ledger, process flow card, and current safety anchor point. These conditions include at least the planned operation time, operation duration, predetermined material temperature and pressure, ventilation capacity, number of personnel allowed to work simultaneously, number of related equipment allowed to be used simultaneously, and the equipment connectivity status and operation step sequence recorded in the digital twin model.

[0047] In this method, a disturbance condition refers to a combination of operating conditions obtained by making bounded adjustments to one or more key parameters of the original operating condition in a virtual space through interpolation or extrapolation, while keeping the actual state of the on-site equipment unchanged. The key parameters preferably include ambient temperature of the work area, ventilation capacity, planned operation duration, number of simultaneous workers, material loading rate, predetermined upper limit of liquid level, and evacuation time. The adjustment range can be set to shift a certain percentage or a certain absolute value to the higher and lower sides respectively within the allowable range specified by the safety anchor point, and under any disturbance condition, it shall not exceed the range that the safety anchor point has clearly prohibited.

[0048] When generating disturbance conditions, the system prioritizes the connection relationship of fixed equipment and the sequence of operation steps, and only adjusts the above-mentioned key parameters and parameters related to personnel organization. It also assigns a disturbance identifier and detailed parameter records to each group of disturbance conditions, and packages the original conditions and each disturbance condition into a candidate set of conditions for this operation.

[0049] When an operational access assessment is required, the risk control service submits each disturbing condition in the original working condition and the candidate working condition set to the aforementioned safety big model for evaluation. The method of invocation is consistent with that of online risk assessment, that is, based on the state of the digital twin model and the initial state of the safety evidence diagram under the corresponding working condition, the risk assessment results, intervention suggestions and evidence node indexes for each working condition are obtained under the constraint of safety anchor points. After obtaining the risk assessment results for a set of working conditions, the system compares the risk levels and risk types between different working conditions, identifies the trajectory of change in risk level as certain parameters change in a certain direction, or the risk type changes from general risk to severe categories such as deflagration risk and asphyxiation risk, and registers this trajectory as a risk path.

[0050] In this method, a risk path refers to a sequence of working conditions in the candidate set of working conditions where a change in a certain key parameter or combination of parameters is the dominant factor, leading to an adverse change in the risk assessment results. A sensitive risk channel in this method refers to a parameter dimension or combination of parameters that has a significant impact on risk, summarized from several risk paths, such as the combined effect of ambient temperature and ventilation capacity, or the combined effect of work duration and the number of people working at the same time.

[0051] When registering risk paths and sensitive risk channels, the system should at least record the names of the key parameters that cause the risk changes, the direction of the changes, the starting and ending points of the risk level jump, the types of risks involved, and the corresponding evidence node indexes, so that they can be referenced in subsequent access determination and security measure formulation.

[0052] After completing the risk path identification, the system generates access instructions and safety measure instructions by combining the original working conditions and sensitive risk channels. In this method, the access instruction refers to the record of the judgment on whether the operation is allowed to start and under what constraints it is allowed to start. It includes at least the following situations: allowed to start, not allowed to start, and not allowed to start until specific safety measures are added. It also includes the reason for the judgment and the risk path number cited. In this method, the safety measure instruction refers to a set of specific safety measures that need to be implemented before the operation starts in order to keep the original working conditions in a relatively safe working condition range. It includes at least the addition or activation of additional gas detection points in sensitive areas, adjustment of ventilation facility operating parameters, shortening of single continuous operation time or increasing rest intervals, limiting the number of people working at the same time or prohibiting the start of the operation during the time period when several high-risk operations overlap, and increasing the execution requirements of certain safety anchor points. The content of the measures, the effective stage and the responsible role are clearly defined in the form of fields.

[0053] After the access instructions and safety measure instructions are generated, they are presented to dispatchers, supervisors, and on-site managers through the enterprise operation management system. The operation management system can display "whether access is permitted," "access restrictions," and "list of safety measures to be implemented" in a structured manner on the work ticket issuance interface and work preparation list interface, and require relevant responsible persons to confirm each item. At the same time, the risk control service will register the version number and effective time of the access instructions and safety measure instructions generated this time together with the work unit identifier, safety anchor version mark, and digital twin model version mark, as the basis for condition verification and risk assessment comparison in the subsequent operation execution stage.

[0054] Preferably, in a tank loading scenario at a petrochemical enterprise, the enterprise can set up perturbation conditions under the premise that the outdoor ambient temperature is within a certain range, the ventilation capacity meets the design value, the planned loading time does not exceed a certain duration, and the number of workers at the same time does not exceed two. These perturbation conditions include a gradual increase in ambient temperature to near the local summer extreme high temperature, a percentage-based decrease in ventilation capacity, a percentage-based extension of the planned operation duration, and an increase in the number of workers at the same time to multiple. This creates several combinations of conditions without changing the valve isolation status or the sequence of operation steps. After analyzing these perturbation conditions, if the safety model identifies that when the ambient temperature increases while ventilation capacity is simultaneously reduced and operation time is extended, the deflagration risk level significantly increases from a lower level to a higher level, and the evidence nodes are concentrated at combustible gas detection nodes and personnel gathering nodes, the system can register this combination as a sensitive risk channel and generate an access instruction that allows only... Loading operations can only commence when ventilation capacity reaches design values, the estimated operation time does not exceed the originally planned duration, and the number of operators does not exceed two. Safety measures include adding combustible gas detection points around the platform, postponing operations during the high-temperature period of the day or conducting them at night, and requiring the addition of a dedicated monitoring personnel. While those skilled in the art can adjust the selection parameters, risk level classification, and specific safety measures for disturbance conditions based on different device types and media characteristics, the process capability and scope of application remain unchanged. This ensures that the method can conduct a fully transparent risk review of key conditions before commencement and support feasible safety decisions in different hazardous chemical operation scenarios.

[0055] S5. During operation, update the safety evidence diagram, input the evidence image segments into the safety big model to obtain a risk assessment, and convert intervention suggestions that meet the intervention trigger conditions into on-site control instructions through safety anchor point verification. The specific implementation is as follows: During the operation, in order to continuously perceive risks and trigger on-site control when necessary without interrupting normal on-site operations, this method relies on the aforementioned unified time synchronization mechanism. Process monitoring devices, environmental monitoring devices, and personnel wearing identification devices continuously collect process monitoring values, environmental monitoring values, and personnel status information according to a predetermined rhythm. Each record carries an operation unit identifier and a time stamp. After receiving the record, the risk control service adds the new record to the safety evidence diagram of the corresponding operation unit according to the operation unit identifier, so that the safety evidence diagram extends continuously on the timeline as the operation progresses.

[0056] To facilitate segmented analysis over time, this method adopts the aforementioned observation window setup. The safety evidence map related to a specific work unit is divided into several continuous time segments along the time axis according to a fixed observation window length and sliding step size. Each time segment corresponds to an observation interval. The system assigns a time slice number to each time segment to identify its sequence within the current operation. At the boundary of each time segment, the risk control service extracts a safety evidence image segment from the safety evidence map. This image segment preserves the equipment nodes, area nodes, work step nodes, safety anchor point nodes, and safety event nodes within that time segment, as well as their temporal sequence, spatial adjacency, and dependencies. This image, along with the digital twin model of the current work unit, is then submitted to the large-scale safety model for analysis. The large-scale safety model provides the risk assessment results and intervention recommendations corresponding to that time segment.

[0057] To ensure the controllability of intervention triggering, before deploying this method, the safety management department, in conjunction with the company's safety objectives and equipment characteristics, sets intervention triggering conditions for different risk types and risk levels. In this method, intervention triggering conditions refer to the judgment rules that determine when the risk assessment results meet certain combinations, which are considered necessary for on-site control. Preferably, these include triggering intervention when a certain risk level remains above a certain level for several consecutive observation windows, or immediately triggering intervention when the risk type belongs to high-risk categories such as deflagration or asphyxiation and the risk level reaches a preset level. Intervention triggering conditions can be stored in the configuration as rule records and version-marked together with the safety anchor point set and the safety big model version.

[0058] During operation, after each round of risk assessment, the risk control service compares the intervention suggestions output by the safety model with the currently effective intervention trigger conditions and safety anchor rules. For intervention suggestions that do not yet meet the intervention trigger conditions, they are only registered as reference information in the risk assessment log and are not converted into on-site control actions. For intervention suggestions that conflict with safety anchor rules, such as suggestions to continue loading under conditions that do not meet combustible gas detection requirements, they are directly rejected or replaced. Only when the risk assessment result corresponding to the intervention suggestion meets the intervention trigger conditions and the suggestion does not violate any safety anchors will the system organize the intervention suggestion into an on-site control instruction.

[0059] In this method, field control instructions refer to control records that can be directly executed by the field control system or operator terminal. They include at least the work unit identifier, time slice number, target object identifier, action type, and necessary action parameters. The target object identifier is used to identify the controlled equipment or personnel terminal and is consistent with the equipment node and personnel identifier in the digital twin model. The action type is used to describe the specific operation, such as closing a valve, starting or stopping a ventilation device, reducing the operating load of a pump, or pushing an evacuation prompt to the operator terminal.

[0060] On-site control commands are issued through a control channel. In this method, the control channel refers to the communication link between the risk control service and the on-site execution agency or terminal used to transmit control information, which can be implemented based on the plant's industrial network. To avoid the same control action being executed multiple times due to network retransmission or system repetition, when the control channel receives an on-site control command from the risk control service, it forms an idempotent marker based on the combination of the work unit identifier, time slice number, and action type. This idempotent marker is registered as the unique key for this control action. When the control channel receives a control command with the same idempotent marker again, it only records the duplicate arrival and does not send a new action request to the downstream execution agency, thereby ensuring that the same control action under the same time slice is executed only once.

[0061] To ensure that the sequence of control actions aligns with the time sequence of risk assessment, the control channel sorts and schedules field control commands according to their time slice numbers when forwarding them. If a control command is delayed due to communication or execution-side issues, the control channel can retry a limited number of times within a preset response time window. In this method, the response time window refers to the maximum permissible time interval from the issuance of the control command to confirmation of its completion. Control commands that do not receive valid execution feedback after this time interval are marked as timed out. The control channel will no longer send the command downstream but will record the timeout in the control log for subsequent manual verification. In this method, the control log refers to a collection of logs used to record the issuance and execution results of field control commands. It includes at least the control command content, idempotency flag, status flag, and time information, and can be associated with the risk assessment log and risk evidence chain.

[0062] To facilitate monitoring and auditing, after forwarding field control instructions to the executor or terminal, the control channel requires the executor to return a status flag when the action is completed or cannot be executed. In this method, the status flag refers to classification information used to summarize the result of this control execution, preferably including categories such as completion, time limit exceeded, version mismatch, and insufficient permissions. Among them, completion indicates that the control action has been implemented on-site as required by the instruction; time limit exceeded indicates that the control action failed to be completed within the preset response time window; version mismatch indicates that the executor detected that the digital twin model version or security anchor version carried by the control instruction is outdated compared to the version currently used on-site, and refused to execute the instruction to avoid performing operations that may no longer be compatible under inconsistent version conditions; insufficient permissions indicates that the current control channel or terminal does not have the authorization to perform this type of operation on the target object.

[0063] Upon receiving a status marker, the risk control service records it in the control log along with the corresponding on-site control instructions, work unit identifier, time slice number, and the risk assessment results at that time. The control log can be linked with the risk assessment log and the risk evidence chain for post-event review and model effectiveness evaluation.

[0064] Preferably, in a petrochemical enterprise tank area loading scenario, the enterprise can set the observation window length to several minutes, divide the risk level into several levels, and agree that intervention will be triggered when the deflagration risk level reaches a high level and lasts for more than two observation windows. After receiving the high-risk assessment result and intervention suggestion to stop loading from the safety big data model at the boundary of two consecutive time slots, the risk control service will organize the suggestion into a field control command that closes the loading feed valve, starts the ventilation equipment around the platform, and pushes an evacuation prompt to the operator's terminal, and issues it through the control channel. The control channel generates an idempotent flag by combining the work unit identifier, the current time slot number, and the action type of closing the feed valve, ensuring that the closing operation is only triggered once in the event of a re-attempt. The process is executed once, and the system waits for the actuator to return a status marker such as completion or time limit exceeding within a preset response time window. Those skilled in the art can adjust the observation window length, response time window, and intervention triggering conditions according to the response characteristics of different devices and corporate safety strategies. However, during the operation, continuous risk assessment is carried out on a time-slice basis based on the safety evidence map and the safety big model. Under the premise of meeting the preset intervention triggering conditions and complying with the safety anchor point constraints, on-site control instructions with idempotent markers and status marker mechanisms are generated and implemented to specific equipment and personnel terminals through the control channel. This process capability and scope of application remain unchanged, thereby ensuring that this method can achieve traceable and controllable real-time risk intervention in different hazardous chemical operation scenarios.

[0065] S6. After the task is completed, collect event records, organize the safety evidence graph and risk assessment results to form a risk evidence chain, align the risk evidence chain with the records to generate training samples, which are used for incremental training of the large safety model and to update safety anchors. The specific implementation is as follows: After a single hazardous chemical operation is completed, in order to systematically archive the safety performance of the entire operation and provide a basis for subsequent model updates, this method involves the risk control service reading event records related to the operation unit from the enterprise's operation management system and on-site monitoring system under the constraint of the operation unit identifier. The event records include at least the operation start time, operation pause and resumption time, operation termination time, on-site hazard registration, accident report, and rectification closure record. The rectification closure record refers to the measures taken for the aforementioned hazards and accidents, the completion time, and the review and confirmation status, and the original time stamp is kept unchanged during reading. In this method, the operation management system refers to the management system used to manage hazardous chemical operation tickets, operation plans, and operation status, and the on-site monitoring system refers to the monitoring system used to collect process monitoring, environmental monitoring, and video monitoring information.

[0066] The risk prevention and control service associates the aforementioned event records with the evolution trajectory of the safety evidence map formed throughout the entire operation, according to the work unit identifier and time sequence. In this method, the evolution trajectory of the safety evidence map refers to the sequence formed by the safety evidence map continuously accumulated from the operation preparation stage and the operation execution stage, with the time slice number increasing. At the same time, the risk assessment results, intervention suggestions, sorted on-site control instructions, and safety anchor point version marks given by the safety big model in each time slice during the entire operation are introduced. This information is linked and sorted on the time axis to form the risk evidence chain of the work unit.

[0067] In this method, the risk evidence chain refers to a set of records compiled in chronological order at the work unit level. These records include the results of disturbance condition analysis, access instructions, and safety measure instructions during the pre-operation access phase; safety evidence images, risk assessment results, intervention triggers, and on-site control execution during each time slice of the operation execution phase; and records of hidden dangers and accidents and rectification closure records after the operation is completed. This chain is used to fully reflect the risk evolution process of a single operation from planning and execution to completion, as well as the corresponding control behaviors.

[0068] Over long-term operation, the system accumulates multiple risk evidence chains from different work units. Based on a pre-defined sample selection strategy, a set of records for parameter updates is extracted from these risk evidence chains. The sample selection strategy can be set to prioritize work units that contain accidents, hazards, or multiple high-risk assessments, or it can select work units that cover typical safety conditions proportionally according to the type of work and the characteristics of the medium, so as to ensure that the training samples include both negative cases and safe execution cases.

[0069] The selected risk evidence chains are organized to form training samples. The training samples include at least the job type, key fields in the digital twin model, such as the main equipment type, connectivity features and typical operating conditions, and compressed markers of the safety evidence graph, such as the density of high-risk events and the occurrence of key safety event combinations obtained in each time slice. These are used to retain structural information that affects risk judgment while ensuring privacy and computational efficiency. The training samples also include the risk assessment results at each key moment, the corresponding on-site control execution results, and subsequent hazard or accident markers. In this method, hazard or accident markers refer to the classification information of whether a hazard, general accident or major accident occurs within a certain time range after a certain time slice.

[0070] In the offline environment, the risk prevention and control service uses the above training samples to perform incremental training on the safety big model. In this method, incremental training refers to adjusting the internal parameters of the model based on the new training samples while retaining the overall capabilities of the original model. This allows the safety big model to gradually adapt to new combinations of working conditions, changes in the range of working conditions caused by equipment modifications, and changes in personnel behavior patterns, while retaining the ability to identify existing typical scenarios.

[0071] To avoid model updates becoming disconnected from safety procedures, the system analyzes newly formed risk patterns during incremental training and after training is completed. When it is found that a certain type of behavior, parameter combination, or work arrangement is repeatedly and significantly correlated with risk level improvement or accident occurrence in multiple risk evidence chains, the system can generate safety anchor update suggestions. In this method, safety anchor update suggestions refer to suggestions to solidify the above-mentioned behavior restrictions, parameter limits, or work arrangement requirements into new safety anchor items, or suggestions to tighten the limits of existing safety anchors.

[0072] The safety anchor update recommendations are reviewed by the safety management department, which decides whether to adopt them based on regulatory requirements and the company's actual management practices. Once adopted, they are incorporated into the company's safety procedures, and the safety anchor version markers are updated simultaneously. The risk control service will link and archive the updated safety anchor version with the relevant work unit identifiers and the risk evidence chain records used as the basis, thereby establishing a locking relationship between the safety procedure configuration and the safety big model version, as well as leaving a trace of the evidence chain. This facilitates tracing back to which historical cases a particular version configuration was based on during post-event audits.

[0073] Throughout the entire process, for situations such as long-term missing monitoring signals, significant deviations in time stamps, and network transmission failures with exhausted retries, the system records this situation along with the current work unit identifier, safety anchor version marker, and known safety evidence image segments into the risk evidence chain. This situation is marked as a special node with insufficient information completeness. At the same time, the system raises the current risk level to no less than the pre-set level through pre-configured risk assessment rules and issues prompts to on-site monitoring personnel and safety management personnel to strengthen manual inspections or temporarily take conservative control measures. This ensures that risks are not underestimated under conditions of incomplete information from the perspective of safety and compliance.

[0074] Preferably, in a tank loading scenario at a petrochemical enterprise, the enterprise can select the risk evidence chain corresponding to all work units that have experienced gas alarms, emergency shutdowns, or on-site handling records as the source of training samples. Key fields of the digital twin model may include the number of loading arms, tank capacity, and medium type. The safety evidence map compression markers may include the proportion of time slices in which high-risk events occur during loading, the overlap between high-value flammable gas events and personnel gathering events, etc. After incremental training, the system finds a stable correlation between reducing monitoring personnel and extending continuous operation time during a certain period at night and the increased risk of deflagration. That is, it generates a safety anchor point update suggestion to limit continuous loading time and increase monitoring personnel during that period. After approval by the safety management department... This requirement is incorporated into the safety procedures, and the corresponding safety anchor version takes effect in all subsequent nighttime loading operations. Those skilled in the art can adjust the training sample selection strategy and incremental training rhythm according to the management level, monitoring methods, and model types of different enterprises. However, after the operation is completed, the risk evidence chain is formed by combining the records of the operation management system and monitoring system with the safety evidence map, risk assessment results, and on-site control behaviors. On this basis, the safety model is incrementally trained, safety anchor update suggestions are generated and implemented, and the process capability and scope of application of increasing the risk level and requiring manual inspection when information is incomplete remain unchanged. This ensures that this method can continuously accumulate experience, optimize models and procedures, and improve the overall safety level in multiple rounds of operations.

[0075] Example 2: Figure 2 A schematic diagram of a hazardous chemical operation risk control device based on an AI safety big data model is provided. The device includes: The operation modeling module is used to generate operation unit identifiers based on hazardous chemical operation tickets, and to construct digital twin models of operation units corresponding to the operation unit identifiers according to the field structure. It also parses safety anchor points from safety regulations and binds them to the digital twin models. The data acquisition and security event extraction module is used to collect data carrying work unit identifiers under unified time synchronization conditions, aggregate the collected data according to the work unit identifiers and extract security events, and associate security event nodes with security anchor point nodes to construct a security evidence graph. The large-scale model analysis module is used to input the digital twin model of the work unit and the safety evidence map into the large-scale safety model, encode the safety evidence map, and decode it under the constraint of safety anchor points to obtain risk assessment results, intervention suggestions and evidence node indexes; The disturbance condition assessment module is used to construct disturbance conditions based on the digital twin model of the work unit before the start of the operation and input the disturbance conditions into the safety model. It compares the risk assessment results corresponding to the disturbance conditions to identify risk paths and generates access instructions and safety measure instructions based on the risk paths. The online intervention control module is used to update the safety evidence map corresponding to the work unit during the operation, input the safety evidence image segment into the safety big model to obtain the risk assessment result, and verify the intervention suggestions that meet the intervention trigger conditions under the safety anchor point constraint and convert them into on-site control instructions. The evidence chain management and incremental training module is used to collect event records related to the work unit after the operation is completed, organize the safety evidence graph and risk assessment results to form a risk evidence chain, align the risk evidence chain with the event records to generate training samples, and perform incremental training on the safety big model and update the safety anchor points based on the training samples.

[0076] In the operational scenario shown in this embodiment: In the tank farm loading scenario of a petrochemical enterprise along the river, this method takes a single gasoline loading operation as the object, and operates step by step around a single operation unit according to the aforementioned steps to form a closed loop. Before the operation starts, the enterprise fills out and approves a gasoline loading hazardous chemical operation ticket according to the specifications. The operation management system generates an operation unit identifier based on the operation ticket, and combines the operation ticket number, the tank farm and the unit area code where the loading arm is located, and the planned start and end times into a unique code. Based on this, the storage tanks, loading arms, connecting pipelines and related valves participating in this loading are screened from the equipment ledger and process flow cards. Information such as material flow direction, valve opening and closing sequence and start-up and shutdown steps are extracted. After unifying the unit of measurement and area code, a digital twin model of this loading operation is constructed. Each storage tank and loading arm node is marked with name, number, tank farm, rated pressure, rated temperature and volume. Each connection relationship is marked with upstream and downstream equipment and the requirement that it should be in an isolated or connected state during this operation. The process step sequence is then linked with... The system establishes a connection between specific equipment and media. Based on the work type and media hazard description in the work order, the system retrieves clauses related to tank loading from the safety regulations library. Hard constraints such as the upper limit of combustible gas volume fraction, allowable oxygen content range, ventilation time, upper limit of work duration, upper limit of simultaneous workers, and protective equipment configuration requirements are broken down into a set of safety anchors. These anchors point to quantifiable objects such as gas detection points on the loading platform, pressure and temperature measurement points of storage tanks and pipelines, number of personnel in the loading area, and the wearing status of personal protective equipment. The safety anchors are then linked one by one with the equipment nodes, area nodes, and work step nodes in the digital twin model to form a safety baseline covering the preparation, execution, and closing stages of this loading. The safety anchors are assigned version tags and stored in the configuration record along with the work unit identifier.

[0077] During the preparation and execution phases, pressure, temperature, liquid level, and flow monitoring devices within the tank area report process monitoring values ​​according to a unified time synchronization rhythm. Combustible gas and oxygen content detection devices around the loading platform report environmental monitoring values. Identification devices worn by operators and supervisors report their location and the status of their protective equipment. Intelligent video recognition devices identify behaviors such as smoking, open flames, excessive personnel gathering, and suspected leaks in the monitoring footage and generate event records. All records carry an operation unit identifier and a timestamp upon generation. The risk control service at the receiving end merges these records according to the operation unit identifier, sorts them by timestamp, and then divides them according to the preset observation window length and sliding step size. For continuous time segments, within each observation window, a safety event is determined based on the boundary conditions and process constraints corresponding to the safety anchor point. Examples include a continuous increase in combustible gas concentration approaching the upper limit, skipping steps without confirmation in sequence during loading, and workers repeatedly entering and exiting restricted areas within a short period. Changes that meet the conditions are registered as safety event nodes, and together with equipment nodes, area nodes, operation step nodes, and safety anchor point nodes in the digital twin model, a safety evidence graph is constructed. Directed associations are established through temporal sequence, spatial adjacency, and procedural dependencies, so that the possible impact paths between a certain abnormal detection value and the risks of subsequent steps can be presented in the graph.

[0078] Before the operation begins, the risk control service first generates the original operating conditions based on the digital twin model of the current loading operation. This records parameters such as the planned loading time, the predetermined loading volume, the number of personnel allowed to work simultaneously on the platform, and the operating capacity of the ventilation equipment. Without altering the actual valve positions and equipment connectivity, bounded disturbances are made to key parameters such as ambient temperature, ventilation capacity, operation duration, number of simultaneous workers, and loading volume to create several disturbance operating conditions. The original operating conditions and the corresponding digital twin model states and initial states of the safety evidence diagrams are then submitted to the large-scale safety model for analysis. By comparing the risk assessment results under each operating condition, a jump in the deflagration risk level is detected when certain parameters change towards higher or lower values. The sequence of work conditions where the risk level changes from general operational risk to deflagration or suffocation risk is registered as a risk path, and sensitive risk channels such as ambient temperature and ventilation capacity, operation duration and number of workers simultaneously are summarized. Based on this, the system generates access instructions to clarify whether the current original working conditions allow loading to start and under what constraints. At the same time, it generates safety measure instructions, requiring the addition or activation of additional combustible gas detection points, adjustment of ventilation equipment operation levels, limitation of the number of workers simultaneously on the platform, and control of single loading time and rest intervals before the operation starts. These instructions are displayed to dispatchers, supervisors, and on-site supervisors through the operation management system. The relevant responsible persons must confirm and implement each item before the work permit can be issued.

[0079] After loading begins, the risk control service continuously receives newly generated monitoring and behavior records, updates the evolution trajectory of the safety evidence map according to time slices, and extracts the corresponding safety evidence image segment at the boundary of each time slice. This image is then submitted to the large-scale safety model for analysis along with the current digital twin model. Under the constraints of safety anchor points, the large-scale safety model provides the current risk level and type, along with intervention suggestions such as stopping loading, reducing loading flow, increasing ventilation, or temporarily evacuating personnel. The safety management department pre-configures intervention trigger conditions for high-risk categories such as deflagration risk and asphyxiation risk. For example, when the deflagration risk level reaches a high level and does not decrease within several consecutive observation windows, on-site control must be implemented. After each analysis, the risk control service compares the intervention suggestions with the intervention trigger conditions and safety anchor point rules. Suggestions that do not yet meet the trigger conditions are only... The risk assessment log records suggestions that conflict with safety anchors, eliminating those that meet the triggering conditions and do not violate safety anchors. These suggestions are then compiled into on-site control instructions, including the work unit identifier, time slice number, controlled object identifier, and action type. These instructions are issued to valve actuators, ventilation equipment control units, or operator terminals via control channels. The control channels use the work unit identifier, time slice number, and action type to form idempotent markers, executing control instructions in time slice order. Controls that are not completed within the preset response time window are marked as timed out and recorded in the control log. The execution side returns status markers such as completed, time exceeded, version mismatch, or insufficient permissions. The risk control service solidifies these status markers along with the corresponding control instructions and risk assessment results for post-event review of each risk assessment and intervention during the loading process.

[0080] After the operation is completed, the risk control service reads the start time, pause and resumption records, termination time, hazard registration, emergency shutdown records, and rectification closure status of this loading from the operation management system and monitoring system. These events are then linked in chronological order with the safety evidence map evolution trajectory of this operation, the risk assessment results of each time segment, intervention suggestions, on-site control instructions, and safety anchor point version markers to form a complete risk evidence chain, reflecting the risk evolution and control path of this loading from access assessment, operation execution to final rectification.

[0081] After accumulating a certain number of risk evidence chains for loading operations, the system selects work units containing gas alarms, emergency shutdowns, or on-site handling records according to a sample selection strategy. It then organizes key fields of the digital twin model, compressed safety evidence graph markers, risk assessment results at each critical moment, control execution results, and subsequent hazard or accident markers to form training samples for incremental training of the large-scale safety model. In an offline environment, the system adjusts the internal parameters of the large-scale safety model to better identify early risk signals under similar working condition combinations. Simultaneously, the system analyzes recurring high-risk patterns in multiple risk evidence chains, such as the stable correlation between reducing monitoring personnel and extending continuous loading time during a certain period at night and the increased risk of deflagration. It automatically generates safety anchor point update suggestions for limiting continuous loading time and increasing monitoring personnel during that period. After approval by the safety management department, these suggestions are incorporated into the safety procedures and the safety anchor point version markers are updated, enabling all subsequent nighttime loading operations to automatically operate under the new safety baseline.

[0082] Preferably, in the above-mentioned tank loading scenario, enterprises can set the observation window length to several minutes and the sliding step size to minute levels, divide the risk level into several grades, and specify the upper limit of the volume fraction of combustible gas before hot work and during loading to a certain percentage. In a nighttime summer high-temperature loading operation, the safety model identifies combustible gas detection nodes approaching the upper limit of the safety anchor point and personnel gathering nodes and operation step jump nodes appearing simultaneously in multiple consecutive time slices. It determines that the deflagration risk level of this operation has reached a high level and triggers intervention conditions. Based on this, the risk control service issues orders to close the feed valve and start the surrounding ventilation equipment. The on-site control instructions for personnel evacuation, along with all records of the control process and subsequent rectification and closure, are incorporated into the risk evidence chain and play a role in incremental training and safety anchor updates. Those skilled in the art can adjust specific parameter configurations in other hazardous chemical loading and unloading, tank area reloading, or confined space cleaning scenarios. However, the preparation, execution, and retrospective capabilities and their applicable scope surrounding the technical chain of work unit identification, digital twin model, safety anchors, safety evidence diagrams, large-scale safety model, access instructions, on-site control instructions, and risk evidence chain remain consistent, thereby reproducing the overall operational effect of this embodiment.

[0083] All calculations involved in the embodiments are dimensionless numerical calculations, and the preset parameters and thresholds in the calculations are set by those skilled in the art according to the actual situation.

[0084] It should be noted that this invention can be deployed on the device itself to realize embedded applications, or it can run on a PC or other terminal with a user interface, thereby meeting various hardware environments and usage requirements.

[0085] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wireless or wired transmission; wired transmission methods include optical fiber, twisted pair, coaxial cable, etc.; wireless transmission includes infrared, microwave, etc. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center containing one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0086] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0087] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0088] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0089] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0090] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0091] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0092] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for risk prevention and control in hazardous chemical operations based on an AI-powered safety model, characterized in that, include: S1. Generate a work unit identifier based on the hazardous chemicals work order, construct a digital twin model of the work unit according to the field structure, parse the safety anchor points from the safety regulations and bind them to the digital twin model; S2. Collect data from related work units under unified time synchronization, collect and extract safety events according to work unit identifiers, and associate safety event nodes and safety anchor point nodes to construct a safety evidence graph. S3. Input the operation digital twin model and safety evidence map into the safety big model, encode the safety evidence map, decode and output the risk assessment results, intervention suggestions and evidence node index under the constraint of safety anchor points; S4. Before the start of the operation, construct the disturbance conditions based on the digital twin model and input them into the safety model. Compare the risk assessment of the disturbance conditions to identify risk paths and generate access instructions and safety measure instructions accordingly. S5. Update the safety evidence diagram during operation, input the evidence image segments into the safety big model to obtain risk assessment, and convert intervention suggestions that meet the intervention trigger conditions into on-site control instructions through safety anchor point verification. S6. After the operation is completed, collect event records, organize the safety evidence diagram and risk assessment results to form a risk evidence chain, align the risk evidence chain with the records to generate training samples, which are used for incremental training of the safety large model and to update the safety anchor points.

2. The method for risk prevention and control of hazardous chemical operations based on an AI-powered safety model according to claim 1, characterized in that, S1 includes: A work unit identifier is generated from the hazardous chemicals work order. The work unit identifier consists of the work order number, the equipment area code, and the planned start and end times. The system identifies the equipment and pipelines involved in the operation from the equipment ledger and process flow card based on the operation unit identifier, extracts material connection relationships and process step sequences, and organizes them according to unified names and units of measurement to build a digital twin model containing equipment nodes, pipeline connection relationships, operation step nodes and rated operating parameters. On-site monitoring and control records carry work unit identifiers when they are generated.

3. The method for risk prevention and control of hazardous chemical operations based on an AI-powered safety model according to claim 2, characterized in that: The system retrieves clauses from the safety regulations database based on the operation type, medium hazard description, and equipment characteristics in the hazardous chemical operation ticket. Conditions with rigid constraints are broken down into safety anchors. Each safety anchor record includes the anchor name, applicable operation type, applicable medium range, measurement object identifier, measurement point description, unit of measurement, allowable value range, required continuous compliance time, and subsequent measure level. These records are then associated with equipment nodes, area nodes, and operation step nodes in the digital twin model, forming a safety baseline that is stored with corresponding operation unit identifiers and safety anchor set data. The safety anchor version is then linked to the configuration record.

4. The method for risk prevention and control of hazardous chemical operations based on an AI-powered safety model according to claim 1, characterized in that, S2 include: Under unified time synchronization, records of work unit identification and time stamps are collected from process monitoring devices, environmental monitoring devices, personnel wearing identification devices, and video intelligent identification devices; The system identifies safety events corresponding to changes in safety anchor point boundaries and process constraints based on the observation window. It registers these safety events as safety event nodes and establishes directed associations with equipment nodes, area nodes, work step nodes, and safety anchor point nodes in the digital twin model according to their temporal order, spatial adjacency, and safety procedure dependencies, thus forming a safety evidence graph.

5. The method for risk prevention and control of hazardous chemical operations based on an AI-powered safety model according to claim 1, characterized in that, S3 include: The risk prevention and control service selects digital twin models and safety evidence diagrams based on the work unit identifiers, and submits the equipment, pipelines, areas, work steps, safety anchor points, safety events and directed relationships into a large safety model code; Under the constraints of safety anchor points, the safety big model generates risk assessment results, intervention suggestions, and evidence node indexes, and writes the risk assessment results, intervention suggestions, evidence node indexes, work unit identifiers, safety anchor point version tags, digital twin model version tags, and safety evidence graph version tags into the risk assessment log.

6. The method for risk prevention and control of hazardous chemical operations based on an AI-powered safety model according to claim 1, characterized in that, S4 include: Before the start of operations, the risk control service generates a candidate set of original and disturbed operating conditions based on the digital twin model of the work unit and safety anchor points. The candidate set of operating conditions is input into the safety model to obtain risk assessment results under the constraint of safety anchor points; Based on the risk assessment results, risk paths and sensitive risk channels are identified, and access instructions and security measure instructions are generated. The enterprise operations management system displays access instructions and safety measure instructions and associates them with work unit identifiers, safety anchor version tags, and digital twin model version tags.

7. The method for risk prevention and control of hazardous chemical operations based on an AI-powered safety model according to claim 1, characterized in that, S5 include: During the operation, the risk control service divides the time into observation windows and uses the safety evidence diagram and digital twin model of the operation unit to call the large safety model to obtain the risk assessment results and intervention suggestions for the time slice; The risk prevention and control service selects intervention suggestions based on preset intervention trigger conditions and safety anchor point rules, generates on-site control instructions containing work unit identifier, time slice number, target object identifier, action type and action parameters, issues them through the control channel, and records the execution results of the on-site control instructions to the control log based on idempotent flags and status flags.

8. The method for risk prevention and control of hazardous chemical operations based on an AI-powered safety model according to claim 1, characterized in that, S6 include: After a single hazardous chemical operation is completed, the risk control service obtains event records from the operation management system and the on-site monitoring system based on the operation unit identifier, and arranges the event records, safety evidence map evolution trajectory and risk assessment results of each time segment to form a risk evidence chain. The risk prevention and control service selects training samples from multiple risk evidence chains based on the sample selection strategy, performs incremental training on the security big model in an offline environment, and generates security anchor point update suggestions. After confirmation by the safety management department, the security anchor version mark is updated and associated with the relevant risk evidence chain record.

9. A hazardous chemical operation risk control device based on an AI safety big data model, used to implement the hazardous chemical operation risk control method based on an AI safety big data model as described in any one of claims 1-8, characterized in that, include: The operation modeling module is used to generate operation unit identifiers based on hazardous chemical operation tickets, and to construct digital twin models of operation units corresponding to the operation unit identifiers according to the field structure. It also parses safety anchor points from safety regulations and binds them to the digital twin models. The data acquisition and security event extraction module is used to collect data carrying work unit identifiers under unified time synchronization conditions, aggregate the collected data according to the work unit identifiers and extract security events, and associate security event nodes with security anchor point nodes to construct a security evidence graph. The large-scale model analysis module is used to input the digital twin model of the work unit and the safety evidence map into the large-scale safety model, encode the safety evidence map, and decode it under the constraint of safety anchor points to obtain risk assessment results, intervention suggestions and evidence node indexes; The disturbance condition assessment module is used to construct disturbance conditions based on the digital twin model of the work unit before the start of the operation and input the disturbance conditions into the safety model. It compares the risk assessment results corresponding to the disturbance conditions to identify risk paths and generates access instructions and safety measure instructions based on the risk paths. The online intervention control module is used to update the safety evidence map corresponding to the work unit during the operation, input the safety evidence image segment into the safety big model to obtain the risk assessment result, and verify the intervention suggestions that meet the intervention trigger conditions under the safety anchor point constraint and convert them into on-site control instructions. The evidence chain management and incremental training module is used to collect event records related to the work unit after the operation is completed, organize the safety evidence graph and risk assessment results to form a risk evidence chain, align the risk evidence chain with the event records to generate training samples, and perform incremental training on the safety big model and update the safety anchor points based on the training samples.

Images