A chemical safety dynamic risk management and control method, device and medium

Abstract

This application discloses a method, equipment, and medium for dynamic risk management and control in chemical safety, relating to the field of chemical production monitoring technology. The method includes: fusing multi-source data, performing lightweight time-series anomaly detection on real-time monitoring data at an edge computing node, identifying potential abnormal data and uploading it to the cloud; performing risk evolution analysis on the cloud based on a physical mechanism model combined with a graph neural network model to predict potential runaway paths and corresponding dynamic risk indices; performing dual verification between the dynamic risk index and expert rules in a knowledge graph to obtain a final risk assessment result; generating control recommendations based on the final risk assessment result, converting the control recommendations into executable operation instructions for a distributed control system and pushing them to the distributed control system for execution; receiving the execution feedback results of the executable operation instructions, and performing online incremental learning and updating of the graph neural network model and the physical mechanism model based on the execution feedback results and real-time monitoring data.

Description

Technical Field

[0001] This application relates to the field of chemical production monitoring technology, and in particular to a method, equipment and medium for dynamic risk management of chemical safety. Background Technology

[0002] Chemical production processes are characterized by high risk, strong coupling, and strong correlation among multiple variables. Abnormal fluctuations in key process parameters such as temperature, pressure, and flow rate in production units can easily trigger chain reactions leading to major safety accidents. To ensure safety in chemical production, existing technologies generally employ distributed control systems combined with fixed-threshold alarm mechanisms. When monitored data exceeds a preset threshold, an audible and visual alarm is triggered, and operators rely on experience to determine the cause of the anomaly and take appropriate measures. However, exceeding the threshold for a single parameter often lags behind the occurrence of the anomaly, making it difficult to capture the coupling relationship between parameters and early signs of anomaly evolution. Furthermore, after an alarm is triggered, there is a lack of ability to diagnose and handle the cause, relying excessively on the operator's experience and response speed, making accurate and timely judgments difficult under complex operating conditions.

[0003] In recent years, some studies have attempted to use recurrent neural networks to detect anomalies in sensor time-series data, improving the sensitivity of identifying early abnormal fluctuations. However, the judgment results of these data-driven methods lack physical interpretability and are difficult to corroborate with the mechanisms of actual chemical processes, easily leading to false alarms or missed alarms. Furthermore, even when existing technologies can identify anomalies, they mostly stop at informing operators of the anomaly, failing to perform tracing analysis of the anomaly's cause, predict risk evolution trends, or automatically generate operating instructions for distributed control systems based on the analysis results to achieve closed-loop intervention. This results in time delays and missed golden windows for risk control. Summary of the Invention

[0004] This application provides a method, equipment, and medium for dynamic risk management of chemical safety to solve the above-mentioned technical problems.

[0005] On the one hand, embodiments of this application provide a method for dynamic risk management and control of chemical safety, including: Data fusion is performed on multi-source data collected from multiple heterogeneous systems. Lightweight time-series anomaly detection is performed on the fused real-time monitoring data at the edge computing node to identify potential abnormal data that deviates from the normal pattern, and the potential abnormal data is uploaded to the cloud. Based on a physical mechanism model combined with a graph neural network model, the potential abnormal data is analyzed for risk evolution in the cloud to predict and generate potential runaway paths and calculate the corresponding dynamic risk index. The dynamic risk index is double-verified with expert rules retrieved from the knowledge graph of chemical process safety to obtain the final risk assessment result; Based on the final risk assessment result, control recommendations are generated, the control recommendations are converted into executable operation instructions for the distributed control system, and the executable operation instructions are pushed to the distributed control system for execution. The system receives the execution feedback results of the executable operation instructions and performs online incremental learning updates on the graph neural network model and the physical mechanism model based on the execution feedback results and real-time monitoring data.

[0006] In one implementation of this application, a lightweight time-series anomaly detection is performed on the fused real-time monitoring data at the edge computing node to identify potential abnormal data deviating from the normal pattern, and the potential abnormal data is uploaded to the cloud, specifically including: Deploy a lightweight temporal neural network model on an edge computing node and acquire single-point time series data of at least one key process parameter from the fused real-time monitoring data; The time series data is input into the lightweight temporal neural network model to calculate the degree of deviation of the current data point from the historical normal pattern, and the degree of deviation is compared with the preset anomaly threshold at the edge. When the deviation exceeds the abnormal threshold, the current data point and the data segments in the neighborhood before and after the current data point are marked as potential abnormal data, triggering a local alarm. The potential abnormal data and the corresponding timestamp are then encapsulated and uploaded to the cloud server through a message queue. If the deviation does not exceed the abnormal threshold, then sliding window detection continues on subsequent real-time monitoring data.

[0007] In one implementation of this application, risk evolution analysis is performed on the potential anomaly data in the cloud based on a physical mechanism model combined with a graph neural network model to predict and generate potential runaway paths and calculate the corresponding dynamic risk index, specifically including: Based on a preset mapping table, the identification of the equipment, section or process parameter involved in the potential abnormal data is obtained, and the physical mechanism model associated with the relevant identification is called. The potential abnormal data is input as a boundary condition into the physical mechanism model to determine the suspected matching abnormal events corresponding to the potential abnormal data, and to calculate the deviation trend of the suspected matching abnormal events. The equipment nodes, material nodes, and process parameter nodes in the chemical production process are constructed as a graph structure, and the graph structure is input into a graph neural network model to analyze the coupling relationship between the nodes. By combining the output of the physical mechanism model with the calculation results of the graph neural network model, the potential runaway path of the suspected matching anomaly event propagating along the graph structure is predicted, and the dynamic risk index corresponding to each potential runaway path is output.

[0008] In one implementation of this application, the dynamic risk index is double-verified with expert rules retrieved from a knowledge graph of chemical process safety to obtain a final risk assessment result, specifically including: Construct a knowledge graph for chemical process safety; the knowledge graph includes rule nodes and associated edges of pre-converted structured expert rules, and the expert rules contain triggering conditions and judgment conclusions; Obtain the key process parameters in the dynamic risk index and the corresponding potential runaway paths, and match the key process parameters with the triggering conditions of expert rules in the knowledge graph to obtain the judgment conclusion of the matched expert rules; If both the dynamic risk index and the key process parameters meet the triggering conditions of the expert rules, the final risk assessment result after double verification is output. If the dynamic risk index and the expert rule assessment result are inconsistent, the manual review process is triggered, and the corresponding conflict event is recorded to iteratively optimize the model.

[0009] In one implementation of this application, a control suggestion is generated based on the final risk assessment result, the control suggestion is converted into executable operation instructions for the distributed control system, and the executable operation instructions are pushed to the distributed control system for execution. Specifically, this includes: Obtain the risk level, risk type, and potential runaway path information from the final risk assessment result, and call the preset expert rule base to match the corresponding control suggestions according to the risk level and risk type; the control suggestions shall at least include the name of the process parameter to be adjusted, the adjustment direction, the adjustment range, or the safety interlock action to be triggered; The adjustment instructions in the control recommendations are encoded and converted according to the industrial standard communication protocol to generate one or more sets of executable operation instructions that can be parsed by the distributed control system. Through a pre-established application programming interface (API) channel, the executable operation instructions are pushed to the instruction queue of the target controller in a standardized message format, and the system waits for an execution confirmation signal, as well as receives the execution confirmation signal and real-time status data returned by the target controller.

[0010] In one implementation of this application, the graph neural network model and the physical mechanism model are updated online incrementally based on the execution feedback results and real-time monitoring data, specifically including: Collect actual process parameter change data within a preset time window after the execution of the executable operation command, and use the actual process parameter change data as the execution feedback result; the execution feedback result includes the actual process parameter change curve, equipment status response and safety index change; The execution feedback result is compared with the expected target effect in the control suggestion, the prediction deviation value is calculated, and the prediction deviation value, together with the real-time monitoring data before execution and the final risk judgment result, form an incremental training sample. Based on the incremental training samples, the graph neural network model and the physical mechanism model are trained online incrementally, the model parameters are updated to adapt to the new working conditions, and the updated model parameters are deployed to the cloud server.

[0011] In one implementation of this application, after performing online incremental learning updates on the graph neural network model and the physical mechanism model based on the execution feedback results and real-time monitoring data, the method further includes: When the executable operation instructions are pushed to the distributed operating system, a cross-system linkage strategy is generated based on the final risk assessment result; The cross-system linkage strategy includes instructions to retrieve real-time images of abnormal areas from the video surveillance system, to send graded evacuation instructions to the emergency broadcast system, and to send instructions to the personnel positioning system to query the distribution of personnel in high-risk areas and provide evacuation guidance. The unified API gateway simultaneously pushes corresponding linkage instructions to each target heterogeneous system and aggregates the execution status data returned by each target heterogeneous system. The intermediate data of anomaly identification, the executable operation instructions, the cross-system linkage strategy, and the execution status data are encapsulated into data blocks, hash values ​​are calculated, and they are distributed and stored in the blockchain.

[0012] In one implementation of this application, it further includes: A digital twin platform corresponding to a chemical production unit is constructed, and historical accident cases and typical disturbance scenarios are injected into the digital twin platform to form a simulated test environment; The entire process of anomaly identification, risk prediction, instruction generation, and multi-system linkage of the digital twin system was run in the simulated test environment. The risk warning time, anomaly identification accuracy, instruction execution success rate, and multi-system linkage time were collected during each run as evaluation indicator data. The evaluation index data is input into a reinforcement learning-based policy optimization neural network to iteratively search and generate optimization suggestions. The optimization suggestions are then pushed to the corresponding edge computing nodes to complete the iterative optimization of the system policy. The optimization suggestions include threshold adjustment, model parameter correction, or expert rule priority reordering suggestions.

[0013] On the other hand, this application also provides a chemical safety dynamic risk management and control device, the device comprising: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which are then executed by the at least one processor to enable the at least one processor to perform a dynamic risk management method for chemical safety as described above.

[0014] On the other hand, this application also provides a non-volatile computer storage medium storing computer-executable instructions, which, when executed, implement the above-described method for dynamic risk management of chemical safety.

[0015] This application provides a method, equipment, and medium for dynamic risk management of chemical safety, which has at least the following beneficial effects: Lightweight time-series anomaly detection is performed on the fused real-time monitoring data through edge computing nodes. Initial data screening and local alarms are completed at the edge, with only potentially abnormal data uploaded to the cloud. This leverages the low-latency response advantage of the edge while reducing the data processing pressure and network transmission burden on the cloud. In the cloud, risk evolution analysis is conducted based on a physical mechanism model combined with a graph neural network model. The physical mechanism model ensures that risk prediction conforms to the basic physical laws of chemical processes, while the graph neural network captures the complex coupling relationships between various equipment, materials, and process parameters in the production system, accurately predicting potential runaway paths and their propagation range. By double-verifying the dynamic risk index with expert rules retrieved from the chemical process safety knowledge graph, false alarms caused by data noise are reduced, while false negatives due to model limitations are avoided. Control recommendations are automatically generated based on the final risk assessment results and converted into executable operation instructions recognizable by the distributed control system for direct push execution. This establishes an instruction channel between the intelligent management and control system and the underlying control system, providing automatically executable control capabilities for safe chemical production. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart illustrating a dynamic risk management method for chemical safety provided in this application embodiment; Figure 2 This is a schematic diagram of the internal structure of a chemical safety dynamic risk management and control device provided in an embodiment of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0019] Figure 1 This is a flowchart illustrating a dynamic risk management method for chemical safety provided in an embodiment of this application.

[0020] The analysis method involved in the embodiments of this application can be implemented by a terminal device or a server, and this application does not impose any special limitations on it. For ease of understanding and description, the following embodiments are all described in detail using a server as an example.

[0021] It should be noted that the server can be a single device or a system composed of multiple devices, i.e., a distributed server. This application does not make any specific limitations on this.

[0022] like Figure 1 As shown in the embodiment of this application, a dynamic risk management method for chemical safety includes: Step 101: Perform data fusion on the multi-source data collected by the multi-heterogeneous system, perform lightweight time-series anomaly detection on the fused real-time monitoring data at the edge computing node, identify potential abnormal data that deviates from the normal pattern, and upload the potential abnormal data to the cloud.

[0023] In this embodiment, chemical production sites typically deploy multiple heterogeneous systems from different manufacturers. These readily available systems include distributed control systems, safety instrumented systems, video surveillance systems, personnel positioning systems, and environmental monitoring systems. Each system operates independently, collecting different types of production process data. For example, the distributed control system provides real-time process parameters such as temperature, pressure, and flow rate of the reactor; the environmental monitoring system provides data on the concentration of toxic and harmful gases; and the personnel positioning system provides real-time location information of workers within the plant area.

[0024] To comprehensively acquire multi-dimensional information on the status of chemical production, this application first uses a unified data acquisition gateway or message middleware to simultaneously access multiple heterogeneous systems via multi-protocol adaptation, acquiring multi-source data from different data sources. It should be noted that the data collected includes equipment status, process parameters, safety data, environmental data, and personnel behavior. Equipment status data includes pressure, temperature, flow rate, vibration, and corrosion rate, sourced from DCS, SCADA, and IoT sensors. Process parameters include feed ratio, reaction temperature gradient, residence time, and catalyst activity. Safety data includes HAZOP analysis tables, SIL levels, alarm records, and SOP execution trajectories. Environmental data includes VOCs concentration, toxic gas diffusion simulation, and weather conditions such as wind direction and speed. Personnel behavior data includes location records from positioning systems, work duration, and whether personnel have entered high-risk areas. Simultaneously, accident cases, maintenance records, and change management (MOC) records are also collected to form a historical event database.

[0025] Understandably, these multi-source data differ significantly in timestamp alignment, sampling frequency, data format, and units of measurement, thus requiring data fusion processing. Specifically, data fusion includes operations such as timestamp synchronization, unit normalization, outlier removal, missing value imputation, and feature extraction. The aim is to transform heterogeneous, multimodal raw data into fused real-time monitoring data with a unified time reference and standard format, providing high-quality data input for subsequent anomaly detection.

[0026] It should be noted that the multi-source heterogeneous data originates from DCS systems, SCADA systems, video surveillance systems, personnel positioning systems, environmental monitoring systems, Safety Instrumented Systems (SIS), and historical databases. Specifically, the DCS system includes process parameters such as temperature, pressure, and flow rate, with a sampling frequency of 1Hz; the SCADA system includes equipment status, such as start / stop and fault codes; the video surveillance system includes abnormal behavior identification results, such as personnel approaching high-risk areas; the personnel positioning system includes personnel location and dwell time; and the environmental monitoring system includes… , Concentration, Safety Instrumented System (SIS) includes interlock action records, and historical database includes accident cases and maintenance records.

[0027] The data fusion process includes several steps: timestamp alignment, unit normalization, sampling frequency unification, missing value handling, feature extraction, and data fusion output. For timestamp alignment, all systems are calibrated using a unified time server based on the NTP time synchronization protocol, with an error of 10ms. For unit normalization, units and dimensions are standardized, such as converting ℃ to K and pressure from MPa to kPa. For sampling frequency unification, high-frequency data is downsampled from 1Hz to 10Hz using interpolation or downsampling methods, while low-frequency data is filled in using linear interpolation (1 time / minute). For missing value handling, LSTM models are used to predict missing values ​​based on time series models, with an error of within 5%. For feature extraction, 23 dimensions of features are extracted based on temporal and statistical features, including mean, variance, slope, number of inflection points, and periodicity index. During data fusion, a unified data frame is constructed, aligned by timestamps, and a fusion data packet (one packet per second) containing all dimensional information is generated.

[0028] After obtaining the fused real-time monitoring data, this application does not directly upload all the data to the cloud, but instead performs preliminary anomaly detection on edge computing nodes close to the data source. It should be noted that these edge computing nodes can be industrial edge gateways, embedded smart boxes, or programmable logic controllers with computing capabilities deployed near workshops or equipment. Lightweight temporal neural network models are pre-deployed on these edge computing nodes. It should be noted that "lightweight" refers to the model's parameter count and computational complexity being specifically optimized to run in real-time on resource-constrained edge devices. This lightweight temporal neural network model is trained offline, learning the time-series characteristics of various key process parameters of a chemical production unit under normal operating conditions, including the normal fluctuation range, rate of change, and temporal dependencies of the parameters.

[0029] Specifically, after deploying a lightweight time-series neural network model on an edge computing node, single-point time-series data of at least one key process parameter from the fused real-time monitoring data are acquired. The time-series data is then input into this lightweight time-series neural network model, and the deviation of the current data point from the historical normal pattern is calculated point by point. It should be noted that the deviation is a quantitative indicator that combines amplitude deviation, rate of change deviation, and time-series prediction error.

[0030] Then, the deviation is compared with a preset anomaly threshold at the edge. When the deviation exceeds the anomaly threshold, the system determines the current data point as potentially abnormal data, marks the data point and its surrounding data segments, and triggers a primary alarm locally, for example, by displaying a prompt message on the operator station's human-machine interface or illuminating an alarm light. Conversely, if the deviation does not exceed the anomaly threshold, the system continues to perform sliding window detection on subsequent real-time monitoring data.

[0031] It should be noted that the anomaly threshold is not a fixed value, but rather a dynamically set value based on the statistical distribution of key process parameters under historical normal operating conditions. Specifically, taking reactor temperature as an example, before system deployment, the average value is calculated by collecting at least 30 days of stable operating data. with standard deviation A dynamic threshold mechanism is adopted: set to ,in This is an adjustable coefficient, typically ranging from 2.5 to 3.0, used to balance sensitivity and false alarm rate; simultaneously, a time-adaptive mechanism is introduced: when the device is in transitional conditions such as start-up, shutdown, or batch switching, the sensitivity is automatically reduced. To prevent false alarms, the value can be reduced to 1.5.

[0032] The abnormal thresholds are set according to the recommended principles for setting thresholds for safety-related parameters in ISO 13849-1 "Safety of machinery—Safety-related components of control systems", with the normal fluctuation range identified in the HAZOP analysis report as the boundary reference.

[0033] In one embodiment, when a lightweight temporal neural network model is deployed on an edge computing node, the anomaly detection module obtains the time series feature distribution under normal operating conditions through offline training and generates a dynamic threshold accordingly.

[0034] After the edge computing node completes the initial screening of anomalies and marks potentially abnormal data, the data needs to be uploaded to the cloud server for in-depth analysis. Considering the actual situation of limited bandwidth and low stability in the network environment of chemical production sites, this embodiment has specifically optimized the data upload process. Specifically, the edge computing node and the cloud server communicate using the Message Queuing Telemetry Transport (MQTT) protocol. This protocol is a lightweight message transmission protocol based on the publish-subscribe model, which has the advantages of low protocol overhead, simple implementation, and reliable transmission in unstable network environments, making it very suitable for data upload in industrial IoT scenarios. To ensure the security of data transmission and prevent potentially sensitive process data from being intercepted or tampered with, this embodiment uses a transport layer security protocol to encrypt the communication link at the transport layer. All uploaded data packets are encrypted before being sent to the cloud. At the same time, to further reduce network bandwidth consumption and improve the real-time performance of data transmission under low bandwidth or high latency network conditions, the edge computing node uses a lossless compression algorithm to compress the data packet content before sending the data packet. After encryption and compression, the data packets are sent to a designated topic on the cloud server via a message queue telemetry transmission protocol. The cloud service program that subscribes to the topic receives, decompresses, and decrypts the data to restore the original data content for subsequent risk evolution analysis. Through the above design, this application can achieve efficient and reliable data collaboration between the edge and the cloud while ensuring data security, ensuring that the system can operate stably even in chemical production sites with limited network conditions.

[0035] In this way, the edge device completes the initial screening of massive real-time data, uploading only data related to anomalies. This ensures that abnormal states can receive local responses within seconds or sub-seconds, while significantly reducing the data processing pressure and network bandwidth consumption in the cloud, thus achieving synergy between rapid edge response and in-depth cloud analysis.

[0036] Step 102: On the cloud, based on the physical mechanism model and graph neural network model, perform risk evolution analysis on potential abnormal data, predict and generate potential runaway paths, and calculate the corresponding dynamic risk index.

[0037] In this embodiment, after receiving potentially abnormal data uploaded by edge nodes, the cloud server performs a more in-depth risk analysis. It should be noted that the cloud possesses powerful computing resources and storage capabilities, enabling the execution of complex model algorithms. This application employs a dual-drive analysis architecture in the cloud that integrates physical mechanism models and graph neural network models. The physical mechanism model is a set of mathematical equations based on fundamental principles of chemical thermodynamics, reaction kinetics, and fluid mechanics. For example, the physical mechanism model for a polymerization reactor includes reaction heat balance equations, material conservation equations, and heat and mass transfer equations. Graph neural networks are deep learning models specifically designed to process graph-structured data, effectively capturing the complex coupling relationships between entities in a chemical production system.

[0038] Specifically, firstly, based on a pre-defined mapping table, the identifiers of the equipment, sections, or process parameters involved in the potential anomaly data are obtained. This mapping table includes a pre-established correspondence between data tags and physical entities; for example, the identifier of a temperature measuring point corresponds to a specific reactor or distillation column. Based on the obtained identifiers, the associated physical mechanism model is retrieved from the model library. Then, the potential anomaly data is input as boundary conditions into the physical mechanism model. By solving the differential or algebraic equations within the model, the suspected matching anomaly event corresponding to the potential anomaly data is determined. It should be noted that a suspected matching anomaly event refers to a physical cause that can explain the current anomaly, such as a sudden increase in catalyst feed rate, a decrease in cooling water flow rate, or heat exchanger fouling. Simultaneously, the physical mechanism model also calculates the deviation trend of the suspected matching anomaly event, such as quantitative indicators like the rate of temperature rise and the rate of pressure accumulation.

[0039] Meanwhile, this application abstracts various entities in the chemical production process as nodes in a graph structure data. Specifically, nodes can include equipment nodes, material nodes, and process parameter nodes. Equipment nodes include reactors, distillation columns, pumps, and compressors; material nodes include raw materials, intermediate products, and finished products; and process parameter nodes include temperature measuring points, pressure measuring points, and flow measuring points. The edges between nodes represent the relationships between entities, such as material flow relationships, energy transfer relationships, control coupling relationships, or spatial proximity relationships. The fused real-time monitoring data is used as the initial features of each node and input into a pre-trained graph neural network model. Through a message passing mechanism, each node aggregates information from its neighboring nodes to update its own representation, thereby calculating the probability distribution of abnormal states propagating in the graph.

[0040] When fusing the output of the physical mechanism model with the computational results of the graph neural network model, the deviation trend calculated by the physical mechanism model is used as a weight correction factor for the node features in the graph neural network, or the suspected matching anomaly event identified by the physical mechanism model is used as the starting node for propagation in the graph neural network. Ultimately, the system predicts the potential runaway path of this suspected matching anomaly event propagating along the graph structure. For example, starting from an abnormal reactor temperature, it predicts that it may lead to a series of chain reactions such as overpressure in the subsequent separation unit and dry burning of the reboiler, and outputs a corresponding dynamic risk index for each potential runaway path. This dynamic risk index is a time-varying quantitative indicator that comprehensively assesses the probability of anomaly occurrence, the scope of propagation, the severity of consequences, and the effectiveness of the current safety protection layer.

[0041] Step 103: Perform a double verification between the dynamic risk index and the expert rules retrieved from the knowledge graph of chemical process safety to obtain the final risk assessment result.

[0042] In this embodiment, to further improve the reliability of risk assessment, this application introduces a dual verification mechanism based on expert rules using knowledge graphs. It should be noted that a knowledge graph is a structured semantic knowledge base that can transform dispersed, unstructured expert experience into computable and reasonable graph-structured data.

[0043] Specifically, the first step is to construct a knowledge graph for chemical process safety. This involves collecting a large amount of unstructured or semi-structured process safety documents accumulated by chemical companies over a long period. Examples include hazard and operability analysis reports, which detail potential deviations, causes, consequences, and recommended measures for each process node; clauses in chemical process safety management regulations, specifying safety limits and handling requirements under various operating conditions; and the temporal logic and operational steps in standard operating procedures. Then, using natural language processing technology combined with a rule extraction engine, the expert experience knowledge in these texts is automatically or semi-automatically transformed into structured rule nodes and associated edges. For example, a typical expert rule might be expressed as: if the reactor pressure exceeds a pressure threshold and the reaction temperature exceeds a temperature threshold, and this state lasts for more than a duration threshold, then an overpressure risk is identified, and the corresponding risk level and handling recommendations are output. These rules are stored in the knowledge graph, and the rule nodes are interconnected through causal, temporal, or hierarchical relationships, forming a vast expert knowledge network.

[0044] During the dual verification process, the system first obtains the current values ​​of the dynamic risk index and key process parameters in potential runaway paths. Then, the system queries the knowledge graph and retrieves one or more expert rules related to the current equipment or section identifier. The key process parameters are matched against the trigger conditions of the expert rules to obtain the judgment conclusion of the matched expert rules. When the dynamic risk index is high and the measured values ​​of the key process parameters simultaneously meet all the trigger conditions set in a certain expert rule, the system determines that the expert rule has been matched. At this point, the final risk judgment result after dual verification is output. This result adopts the conclusion of the expert rules, or a weighted sum of the conclusion of the expert rules and the dynamic risk index output by the model.

[0045] When the dynamic risk index and the expert rule's judgment result are inconsistent—for example, the dynamic risk index indicates a high risk, but the key process parameters do not trigger any expert rule conditions; or conversely, the expert rule triggers the warning condition, but the dynamic risk index is at a low level—this embodiment designs a special arbitration mechanism to handle such conflicting situations. Specifically, the system does not simply select one side as the final result, but triggers a manual review process. The system integrates the model prediction result, expert rule matching status, and relevant real-time monitoring data into a conflict event report, which is pushed to the safety management personnel's operating terminal or mobile device, prompting manual intervention for comprehensive judgment. Simultaneously, the system records complete information about the conflict event, including the event occurrence time, equipment identification, model prediction value, rule trigger condition matching status, and the final manual review judgment conclusion, forming a structured conflict event record. These records will serve as valuable feedback samples for subsequent iterative optimization and training of the model. For example, if manual review confirms the existence of a risk but the expert rule fails to trigger, it indicates that the rule's conditions may be too strict. The system can use this case as a positive sample to adjust the rule parameters or expand the rule base. If manual review confirms that there is no actual risk but the model falsely reports a risk, this case can be used as a negative sample to optimize the weight parameters of the graph neural network model and reduce the false alarm rate in similar scenarios.

[0046] Through the synergistic work of the aforementioned dual verification and arbitration mechanisms, this application can not only quickly output reliable risk assessment results when the model and rules reach an agreement, but also ensure the accuracy of the decision through manual review when the two disagree. At the same time, it uses conflict events to continuously drive the self-evolution of the model, so that the risk identification capability of the system can be continuously improved as the running time increases.

[0047] Step 104: Generate control recommendations based on the final risk assessment results, convert the control recommendations into executable operation instructions for the distributed control system, and push the executable operation instructions to the distributed control system for execution.

[0048] In this embodiment, after obtaining the final risk assessment result that has undergone double verification, this application achieves a complete closed loop from risk assessment to automatic control execution. It should be noted that in traditional systems, the risk assessment result is usually only displayed to the operator as alarm information, requiring the operator to manually determine whether and how to intervene. This process suffers from problems such as response delays, human error, and inconsistent operations. This application, however, automatically generates specific control suggestions based on the final risk assessment result and converts human-readable control suggestions into machine-executable operation instructions.

[0049] Specifically, the process begins by acquiring the risk level, risk type, and potential runaway path information from the final risk assessment results. Risk levels can be categorized into multiple levels, such as early warning, preparedness, emergency response, and recovery. Risk types can include overpressure risk, overtemperature risk, leakage risk, and reaction runaway risk. Then, a pre-defined expert rule base is invoked, which is either linked to or independently stored within the aforementioned knowledge graph. This expert rule base predefines corresponding response strategies for different risk levels, risk types, and potential runaway paths. For example, for the risk of a rapid temperature rise in a polymerization reactor due to excessive catalyst feed, control recommendations could include reducing the catalyst feed valve opening by a specified percentage, increasing the cooling water flow rate by a specified value, and lowering the reaction temperature setpoint by several units. Control recommendations can also include non-process parameter operations, such as triggering safety interlock logic, starting a backup pump, or closing an emergency shut-off valve. Control recommendations are represented in a structured data format, including at least the name of the process parameter to be adjusted, the adjustment direction, the adjustment range, or the safety interlock action to be triggered.

[0050] After generating control recommendations, these recommendations are further converted into executable operation instructions that the distributed control system can directly recognize and execute. Distributed control systems typically use specific industrial communication protocols for data exchange, such as open platform communication unified architecture protocols or industrial automation communication protocols. During the conversion, the corresponding protocol encoder is first invoked based on the communication protocol type of the target distributed control system. Then, each adjustment instruction in the control recommendations is encapsulated into a corresponding data structure according to the protocol specification. For analog quantity adjustments, a write instruction packet containing fields such as parameter address, target value, and change slope is generated; for digital quantity interlocking, an instruction packet containing fields such as trigger bit and hold time is generated. These encoded and converted data packets are the executable operation instructions.

[0051] Finally, through a pre-established application programming interface (API) channel—which can be implemented based on a representative state transfer architecture or a remote procedure call (RPC) protocol—executable commands are pushed to the command queue of the target distributed control system in a standardized message format. The system waits for the target controller to return an execution confirmation signal and receives real-time status data after execution to confirm that the command has been executed correctly.

[0052] In one embodiment of this application, while pushing executable operation instructions to the distributed control system for execution, the method also includes generating a cross-system linkage strategy based on the final risk assessment result. Specifically, the method includes: sending instructions to the video surveillance system to retrieve real-time footage of the abnormal area for remote confirmation of the situation; sending graded evacuation instructions to the emergency broadcast system to notify personnel in relevant areas to evacuate based on the scope of risk impact; and sending instructions to the personnel positioning system to query the distribution of personnel in high-risk areas and provide evacuation guidance, obtaining the list and location of personnel in the affected area, and further sending evacuation guidance instructions to the positioning tags or mobile terminals carried by these personnel. It should be noted that all the above-mentioned cross-system linkage instructions are simultaneously pushed to the interfaces of each target heterogeneous system through a unified API gateway using a standard Hypertext Transfer Protocol or Message Queuing Telemetry Transmission Protocol. After each system executes the linkage action, it returns execution status data, such as a video retrieval success flag, broadcast confirmation information, and personnel location list. The system encapsulates this execution status data, along with intermediate data from the anomaly identification process, executable operation instruction content, cross-system linkage strategy, and other key information, into data blocks, calculates hash values, and distributes them in a blockchain network to form an immutable, fully traceable record.

[0053] Step 105: Receive the execution feedback results of the executable operation instructions, and perform online incremental learning and updates on the graph neural network model and physical mechanism model based on the execution feedback results and real-time monitoring data.

[0054] In this embodiment, to ensure that the risk management model can continuously adapt to new operating conditions, equipment aging, raw material changes, and environmental drift in the chemical production process, this application introduces an online incremental learning mechanism. It should be noted that online incremental learning differs from traditional offline batch training; it can use newly generated data to fine-tune the model in real-time or near real-time during system operation, ensuring that the model always maintains optimal adaptability to the current operating conditions.

[0055] Specifically, after executable commands are pushed to the distributed control system for execution, the system continuously collects actual process parameter change data within a preset time window after command execution, such as temperature response curves, pressure recovery times, and flow rate trends, and uses this data as execution feedback results. These feedback results can specifically include actual process parameter change curves, equipment status responses, and changes in safety indicators. Then, the system compares these actual execution effects with the expected target effects in the control recommendations and calculates the prediction deviation value. For example, if the control recommendations expect that reducing the catalyst feed rate should cause the reaction temperature to decrease by a specified amount within a specified time, but the actual decrease is less than expected, a positive prediction deviation occurs. This deviation value reflects an error in the current model's understanding of the system's dynamic characteristics.

[0056] Subsequently, the system combines the predicted deviation value with real-time monitoring data before execution, the final risk assessment result, and feedback data after execution to form an incremental training sample. This sample contains complete sequence information from anomaly occurrence to control execution and effect evaluation. Using this incremental training sample, the system performs online incremental training on the adjustable empirical parameters in the cloud-based graph neural network model and the physical mechanism model. Incremental training employs methods such as mini-batch gradient descent or online stochastic gradient descent, which does not require retraining the entire model but only fine-tunes the model parameters. For example, the attention weights between nodes in the graph neural network model are adjusted based on the feedback results, making the prediction of the anomaly propagation path more consistent with the actual observed response process. Empirical parameters such as the heat transfer coefficient and reaction rate constant in the physical mechanism model are also corrected based on the actual response curve. The updated model parameters are saved and deployed to the cloud service for subsequent real-time risk analysis. Through this closed-loop feedback learning mechanism, this application achieves continuous model evolution and adaptive capabilities.

[0057] In another preferred embodiment of this application, the system's control effectiveness is verified and optimized using a digital twin platform. Specifically, a digital twin platform is first constructed that corresponds one-to-one with a real chemical production unit in terms of geometry, material properties, control logic, and dynamic characteristics. Then, historical accident cases and typical disturbance scenarios are injected into the digital twin platform to form a simulated testing environment. The entire process of anomaly identification, risk prediction, instruction generation, and multi-system linkage described in this application is run in this simulated testing environment, and the risk warning time, anomaly identification accuracy, instruction execution success rate, and multi-system linkage time are collected during each run as evaluation index data. These evaluation index data are then input into a reinforcement learning-based policy optimization neural network to iteratively search and generate optimization suggestions, which are then pushed to the corresponding edge computing nodes to complete the iterative optimization of the system strategy. Optimization suggestions may include threshold adjustments, model parameter corrections, or expert rule priority reordering. In this way, this application can continuously optimize the system configuration in a safe simulation environment and then deploy the optimized strategy to the real production system, forming an outer self-optimizing closed loop.

[0058] The above are embodiments of the method proposed in this application. Based on the same inventive concept, embodiments of this application also provide a chemical safety dynamic risk management and control device, the structure of which is as follows: Figure 2 As shown.

[0059] Figure 2 This is a schematic diagram of the internal structure of a chemical safety dynamic risk management and control device provided in an embodiment of this application. Figure 2 As shown, the device includes: At least one processor; And, a memory that is communicatively connected to at least one processor; The memory stores instructions that can be executed by at least one processor, and the instructions, when executed by at least one processor, enable at least one processor to: Data fusion is performed on multi-source data collected from multiple heterogeneous systems. Lightweight time-series anomaly detection is performed on the fused real-time monitoring data at edge computing nodes to identify potential abnormal data that deviates from the normal pattern and upload the potential abnormal data to the cloud. Based on a physical mechanism model combined with a graph neural network model, risk evolution analysis is performed on potential abnormal data in the cloud to predict and generate potential runaway paths and calculate the corresponding dynamic risk index. The dynamic risk index is double-verified with expert rules retrieved from the knowledge graph of chemical process safety to obtain the final risk assessment result. Based on the final risk assessment results, control recommendations are generated, which are then converted into executable operation instructions for the distributed control system and pushed to the distributed control system for execution. It receives the execution feedback results of executable operation instructions and performs online incremental learning and updates of the graph neural network model and physical mechanism model based on the execution feedback results and real-time monitoring data.

[0060] This application also provides a non-volatile computer storage medium storing computer-executable instructions, which, when executed, can: Data fusion is performed on multi-source data collected from multiple heterogeneous systems. Lightweight time-series anomaly detection is performed on the fused real-time monitoring data at edge computing nodes to identify potential abnormal data that deviates from the normal pattern and upload the potential abnormal data to the cloud. Based on a physical mechanism model combined with a graph neural network model, risk evolution analysis is performed on potential abnormal data in the cloud to predict and generate potential runaway paths and calculate the corresponding dynamic risk index. The dynamic risk index is double-verified with expert rules retrieved from the knowledge graph of chemical process safety to obtain the final risk assessment result. Based on the final risk assessment results, control recommendations are generated, which are then converted into executable operation instructions for the distributed control system and pushed to the distributed control system for execution. It receives the execution feedback results of executable operation instructions and performs online incremental learning and updates of the graph neural network model and physical mechanism model based on the execution feedback results and real-time monitoring data.

[0061] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for IoT devices and media are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0062] The above description is merely an embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of this application should be included within the scope of the claims of this application.

Claims

1. A method for dynamic risk management and control of chemical safety, characterized in that, The method includes: Data fusion is performed on multi-source data collected from multiple heterogeneous systems. Lightweight time-series anomaly detection is performed on the fused real-time monitoring data at the edge computing node to identify potential abnormal data that deviates from the normal pattern, and the potential abnormal data is uploaded to the cloud. Based on a physical mechanism model combined with a graph neural network model, the potential abnormal data is analyzed for risk evolution in the cloud to predict and generate potential runaway paths and calculate the corresponding dynamic risk index. The dynamic risk index is double-verified with expert rules retrieved from the knowledge graph of chemical process safety to obtain the final risk assessment result; Based on the final risk assessment result, control recommendations are generated, the control recommendations are converted into executable operation instructions for the distributed control system, and the executable operation instructions are pushed to the distributed control system for execution. The system receives the execution feedback results of the executable operation instructions and performs online incremental learning updates on the graph neural network model and the physical mechanism model based on the execution feedback results and real-time monitoring data.

2. The method for dynamic risk management and control of chemical safety according to claim 1, characterized in that, Lightweight time-series anomaly detection is performed on the fused real-time monitoring data at the edge computing node to identify potential abnormal data that deviates from the normal pattern, and the potential abnormal data is uploaded to the cloud. Specifically, this includes: Deploy a lightweight temporal neural network model on an edge computing node and acquire single-point time series data of at least one key process parameter from the fused real-time monitoring data; The time series data is input into the lightweight temporal neural network model to calculate the degree of deviation of the current data point from the historical normal pattern, and the degree of deviation is compared with the preset anomaly threshold at the edge. When the deviation exceeds the abnormal threshold, the current data point and the data segments in the neighborhood before and after the current data point are marked as potential abnormal data, triggering a local alarm. The potential abnormal data and the corresponding timestamp are then encapsulated and uploaded to the cloud server through a message queue. If the deviation does not exceed the abnormal threshold, then sliding window detection continues on subsequent real-time monitoring data.

3. The method for dynamic risk management and control of chemical safety according to claim 1, characterized in that, Based on a physical mechanism model combined with a graph neural network model, risk evolution analysis is performed on the potential anomaly data in the cloud to predict and generate potential runaway paths and calculate the corresponding dynamic risk index, specifically including: Based on a preset mapping table, the identification of the equipment, section or process parameter involved in the potential abnormal data is obtained, and the physical mechanism model associated with the relevant identification is called. The potential abnormal data is input as a boundary condition into the physical mechanism model to determine the suspected matching abnormal events corresponding to the potential abnormal data, and to calculate the deviation trend of the suspected matching abnormal events. The equipment nodes, material nodes, and process parameter nodes in the chemical production process are constructed as a graph structure, and the graph structure is input into a graph neural network model to analyze the coupling relationship between the nodes. By combining the output of the physical mechanism model with the calculation results of the graph neural network model, the potential runaway path of the suspected matching anomaly event propagating along the graph structure is predicted, and the dynamic risk index corresponding to each potential runaway path is output.

4. The method for dynamic risk management and control of chemical safety according to claim 1, characterized in that, The dynamic risk index is double-verified with expert rules retrieved from the knowledge graph of chemical process safety to obtain the final risk assessment result, specifically including: Construct a knowledge graph for chemical process safety; the knowledge graph includes rule nodes and associated edges of pre-converted structured expert rules, and the expert rules contain triggering conditions and judgment conclusions; Obtain the key process parameters in the dynamic risk index and the corresponding potential runaway paths, and match the key process parameters with the triggering conditions of expert rules in the knowledge graph to obtain the judgment conclusion of the matched expert rules; If both the dynamic risk index and the key process parameters meet the triggering conditions of the expert rules, the final risk assessment result after double verification is output. If the dynamic risk index and the expert rule assessment result are inconsistent, the manual review process is triggered, and the corresponding conflict event is recorded to iteratively optimize the model.

5. The method for dynamic risk management and control of chemical safety according to claim 1, characterized in that, Based on the final risk assessment result, control recommendations are generated, the control recommendations are converted into executable operation instructions for the distributed control system, and the executable operation instructions are pushed to the distributed control system for execution. Specifically, this includes: Obtain the risk level, risk type, and potential runaway path information from the final risk assessment result, and call the preset expert rule base to match the corresponding control suggestions according to the risk level and risk type; the control suggestions shall at least include the name of the process parameter to be adjusted, the adjustment direction, the adjustment range, or the safety interlock action to be triggered; The adjustment instructions in the control recommendations are encoded and converted according to the industrial standard communication protocol to generate one or more sets of executable operation instructions that can be parsed by the distributed control system. Through a pre-established application programming interface (API) channel, the executable operation instructions are pushed to the instruction queue of the target controller in a standardized message format, and the system waits for an execution confirmation signal, as well as receives the execution confirmation signal and real-time status data returned by the target controller.

6. The method for dynamic risk management and control of chemical safety according to claim 1, characterized in that, Based on the execution feedback results and real-time monitoring data, the graph neural network model and physical mechanism model are updated online incrementally, specifically including: Collect actual process parameter change data within a preset time window after the execution of the executable operation command, and use the actual process parameter change data as the execution feedback result; the execution feedback result includes the actual process parameter change curve, equipment status response and safety index change; The execution feedback result is compared with the expected target effect in the control suggestion, the prediction deviation value is calculated, and the prediction deviation value, together with the real-time monitoring data before execution and the final risk judgment result, form an incremental training sample. Based on the incremental training samples, the graph neural network model and the physical mechanism model are trained online incrementally, the model parameters are updated to adapt to the new working conditions, and the updated model parameters are deployed to the cloud server.

7. The method for dynamic risk management and control of chemical safety according to claim 1, characterized in that, After performing online incremental learning updates on the graph neural network model and physical mechanism model based on the execution feedback results and real-time monitoring data, the method further includes: When the executable operation instructions are pushed to the distributed operating system, a cross-system linkage strategy is generated based on the final risk assessment result; The cross-system linkage strategy includes instructions to retrieve real-time images of abnormal areas from the video surveillance system, to send graded evacuation instructions to the emergency broadcast system, and to send instructions to the personnel positioning system to query the distribution of personnel in high-risk areas and provide evacuation guidance. The unified API gateway simultaneously pushes corresponding linkage instructions to each target heterogeneous system and aggregates the execution status data returned by each target heterogeneous system. The intermediate data of anomaly identification, the executable operation instructions, the cross-system linkage strategy, and the execution status data are encapsulated into data blocks, hash values ​​are calculated, and they are distributed and stored in the blockchain.

8. The method for dynamic risk management and control of chemical safety according to claim 1, characterized in that, The method further includes: A digital twin platform corresponding to a chemical production unit is constructed, and historical accident cases and typical disturbance scenarios are injected into the digital twin platform to form a simulated test environment; The entire process of anomaly identification, risk prediction, instruction generation, and multi-system linkage of the digital twin system was run in the simulated test environment. The risk warning time, anomaly identification accuracy, instruction execution success rate, and multi-system linkage time were collected during each run as evaluation indicator data. The evaluation index data is input into a reinforcement learning-based policy optimization neural network to iteratively search and generate optimization suggestions. The optimization suggestions are then pushed to the corresponding edge computing nodes to complete the iterative optimization of the system policy. The optimization suggestions include threshold adjustment, model parameter correction, or expert rule priority reordering suggestions.

9. A dynamic risk management and control device for chemical safety, characterized in that, The device includes: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform a dynamic risk management method for chemical safety as described in any one of claims 1-8.

10. A non-volatile computer storage medium storing computer-executable instructions, characterized in that, When the computer-executable instructions are executed, a dynamic risk management method for chemical safety as described in any one of claims 1-8 is implemented.

Images