Combustible toxic gas monitoring control method and system
By constructing dynamic process maps and risk assessment models, and combining them with real-time process data to identify operating conditions, the shortcomings of fixed threshold monitoring and control have been addressed, enabling precise monitoring and control of gas leaks and ensuring production safety.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- FOSHAN SHUNDE MANLING ELECTRICAL APPLIANCE CO LTD
- Filing Date
- 2026-03-10
- Publication Date
- 2026-06-09
AI Technical Summary
Existing gas monitoring and control technologies rely on fixed thresholds, which are difficult to adapt to dynamically changing production process conditions. This results in delayed early warnings under high-risk conditions and frequent false alarms under low-risk conditions, making it impossible to make accurate responses in the early stages of an accident.
A dynamic process map of the production system is constructed. Combined with real-time process equipment status data, the current operating mode is determined by a time-series pattern recognition model. A risk assessment model is used to simulate gas diffusion, calculate the risk probability value and leakage diffusion trend, match the hierarchical control strategy, and execute control commands.
It enables precise detection of potential gas leaks and diffusion patterns under different operating conditions, allowing for early risk assessment, prevention of safety accidents, and ensuring the safety of personnel and equipment during the production process.
Smart Images
Figure CN122176873A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial production safety monitoring and control technology, specifically to a method and system for monitoring and controlling combustible and toxic gases. Background Technology
[0002] In industrial production processes, especially in industries such as chemical, metallurgical, oil and gas processing, and coal chemical industries, the generation, storage, and transportation of flammable and toxic gases are common production processes. Once such gases leak, they may not only cause serious safety accidents such as combustion and explosion, but also seriously harm the health of on-site workers and have adverse effects on the surrounding ecological environment.
[0003] To ensure production safety, relevant companies usually install gas monitoring equipment at the production site to capture gas concentration information in real time and take corresponding control measures based on the characteristics of the production process.
[0004] Currently, with the increasing complexity and automation of industrial production processes, the operating conditions of production systems are showing dynamic changes. The diffusion patterns of gas leaks are closely related to factors such as production process conditions, equipment operating status, and material flow. How to combine the actual operating logic of the production system to achieve accurate monitoring and scientific control of flammable and toxic gases has become a key focus in the field of industrial production safety. Related monitoring and control technologies are also being continuously optimized and improved to adapt to the needs of complex and ever-changing production scenarios.
[0005] The limitations of existing technologies include at least the following problems: existing gas monitoring and control technologies rely on preset fixed concentration thresholds for alarm and linkage control, and their fundamental limitation is that they apply static threshold judgment systems to dynamically changing production processes.
[0006] The actual operation of the factory includes a variety of main operating conditions such as normal production, equipment start-up and shutdown, and pressurized maintenance. The inherent risk level of the system, the typical pattern of gas leakage, and the acceptable intervention threshold are significantly different under different operating conditions. Using a fixed threshold can easily cause false alarms due to normal fluctuations under low-risk operating conditions (such as normal and stable operation), which can interfere with production.
[0007] In high-risk operating conditions (such as start-up, shutdown, and maintenance), the failure to tighten thresholds in a targeted manner leads to delayed warnings, making it difficult for the control system to make accurate responses that are adapted to the current production status in the early stages of an accident. Summary of the Invention
[0008] To address the shortcomings of existing technologies, this invention provides a method and system for monitoring and controlling combustible and toxic gases. This solves the problem that existing technologies use fixed threshold gas monitoring, which makes it difficult to adapt to dynamically changing factory production processes and conditions, leading to delayed warnings under high-risk conditions and frequent false alarms under low-risk conditions.
[0009] To achieve the above objectives, the present invention provides the following technical solution: a method for monitoring and controlling combustible and toxic gases, comprising the following steps: constructing a dynamic process map of a production system, wherein the map uses production equipment, material pipelines, and monitoring sensors as nodes, and material flow direction, energy transfer, and signal connection relationships as edges; acquiring real-time process equipment status data and production parameters, and determining the current dominant process operating mode by combining the dynamic process map with a pre-trained time-series pattern recognition model; inputting the gas concentration, environmental parameters, and process equipment status data corresponding to the current dominant process operating mode into a pre-trained risk assessment model to execute a gas diffusion model. The proposed method involves calculating and generating a preliminary leak diffusion trend prediction, and combining this with process equipment status data to calculate risk probability values and leak diffusion trend predictions. Specifically, this includes: identifying spatial regions in the simulated concentration field that exceed a preset threshold as potential hazardous areas; calculating the concentration gradient vector along the normal direction of the boundary of this region as the dominant diffusion direction and velocity; determining the relative positions of potential hazardous areas, key equipment, and ventilation boundaries based on process equipment status data, and calculating risk probability values based on the dominant diffusion direction and velocity; and matching a hierarchical control strategy adapted to the current dominant process operating mode based on the risk probability values and leak diffusion trend predictions, and executing control commands.
[0010] Furthermore, the specific steps for constructing a dynamic process map are as follows: Based on the production process flow diagram and equipment layout diagram, determine all nodes and their connection relationships; define production equipment, material pipelines and monitoring sensors as nodes and store them in the map; define the material flow direction, energy transfer and signal connection relationship between nodes as edges and store them in the map.
[0011] Furthermore, the specific steps for determining the current dominant process operating mode are as follows: associate the real-time acquired process equipment status data and production parameters with the corresponding nodes of the dynamic process graph; convert the graph after data association into the input format required by the time-series pattern recognition model; input the converted graph into the model and receive its output of the current dominant process operating mode.
[0012] Furthermore, the specific steps for matching the hierarchical control strategy are as follows: Based on the risk probability value, leakage diffusion trend prediction, and the current dominant process operating mode, query the pre-stored strategy mapping table, which records the correspondence between the operating mode, risk value range, diffusion trend, and hierarchical control strategy; extract the appropriate hierarchical control strategy from the strategy mapping table.
[0013] Furthermore, the steps for dynamic adjustment during the query process include: real-time monitoring of whether the risk probability value and leakage spread trend prediction change suddenly during the query; if the change causes the original query result to become invalid, the strategy mapping table is queried again immediately based on the result after the change; and the strategy obtained from the re-query is used as the final hierarchical control strategy to be executed.
[0014] Furthermore, it also includes the step of optimizing the policy mapping table: recording the policy executed each time and its corresponding risk outcome and operating mode; periodically iterating and optimizing the correspondence in the policy mapping table based on reinforcement learning algorithm with the goal of maximizing the treatment effect.
[0015] Furthermore, the specific steps for identifying potential hazardous areas are as follows: spatial grid traversal and threshold comparison are performed on the concentration field generated by gas diffusion simulation calculation; connected component analysis is performed on adjacent grid nodes with concentrations exceeding the threshold, and they are marked as independent potential hazardous areas; the spatial boundary coordinate set of all potential hazardous areas is output.
[0016] Furthermore, the specific steps for calculating the dominant diffusion direction and velocity are as follows: For a point on the boundary of a potential hazardous area, calculate the spatial gradient vector of its concentration field and the unit normal vector of the boundary; project the spatial gradient vector onto the direction of the unit normal vector to obtain the dominant diffusion direction and velocity scalar of that point; aggregate the scalars of all points on the boundary to generate the dominant diffusion direction and velocity of the region.
[0017] Further, the specific steps for determining the relative position and calculating the risk probability value are as follows: extract the spatial coordinates of the key equipment and the ventilation boundary from the process equipment status data; calculate the shortest spatial distance between the potential hazardous area and the above coordinates, and if it is less than the safety buffer distance, it is determined to be a proximity; calculate the risk arrival time and comprehensively calculate the risk probability value based on the proximity relationship, the dominant diffusion direction and speed.
[0018] A combustible and toxic gas monitoring and control system includes: a graph construction module for constructing a dynamic process graph of the production system, wherein the graph uses production equipment, material pipelines, and monitoring sensors as nodes, and material flow direction, energy transfer, and signal connection relationships as edges; a working condition mode recognition module for acquiring process equipment status data and production parameters in real time, and determining the current dominant process working condition mode by combining the dynamic process graph with a pre-trained time-series pattern recognition model; a dynamic risk assessment module for inputting gas concentration, environmental parameters, and process equipment status data corresponding to the current dominant process working condition mode into a pre-trained risk assessment model to perform gas diffusion simulation calculations to generate preliminary leakage diffusion trend predictions, and calculating risk probability values and leakage diffusion trend predictions by combining process equipment status data; and a control strategy execution module for matching a hierarchical control strategy adapted to the current dominant process working condition mode based on the risk probability value and leakage diffusion trend predictions, and executing control commands.
[0019] The present invention has the following beneficial effects:
[0020] (1) This method for monitoring and controlling combustible and toxic gases constructs a dynamic process map with production equipment, material pipelines and monitoring sensors as nodes and material flow direction, energy transfer and signal connection relationship as edges. It combines real-time acquired process equipment status data and production parameters, uses a pre-trained time sequence pattern recognition model to determine the current dominant process operating mode, performs gas diffusion simulation calculation through a risk assessment model, identifies potential dangerous areas, determines the dominant diffusion direction and speed, and calculates the risk probability value in combination with process equipment status data. This breaks the limitation of data isolation in traditional monitoring, deeply integrates process operation logic with gas monitoring, can accurately capture potential hidden dangers and diffusion patterns of gas leakage under different operating conditions, predict risks in advance, avoid safety accidents caused by incomplete monitoring and inaccurate judgment, and effectively protect the safety of personnel and equipment in the production process.
[0021] (2) The combustible and toxic gas monitoring and control method constructs a dynamic process map. First, based on the production process flow diagram and equipment layout diagram, all nodes and their connection relationships are determined. Then, production equipment, material pipelines and monitoring sensors are defined as nodes, and the material flow direction, energy transfer and signal connection relationship between nodes are defined as edges. This ensures that the map can truly and comprehensively reflect the overall operating status of the production system. The real-time acquired process equipment status data and production parameters are associated with the corresponding nodes of the map, converted into the format required by the time sequence pattern recognition model and input into the model to achieve accurate identification of the dominant process condition mode. This map construction and condition recognition method fully fits the actual layout and process logic of the production site, can adapt to process adjustments in the production process, ensure the continuity and accuracy of condition recognition, avoid blind monitoring that is detached from the actual production scenario, and reduce safety risks and control deviations caused by errors in condition judgment.
[0022] (3) This method for monitoring and controlling combustible and toxic gases, by performing spatial grid traversal and threshold comparison on the concentration field generated by gas diffusion simulation calculation, marks independent potential hazardous areas and outputs spatial boundary coordinates, calculates the concentration gradient vector and boundary unit normal vector for the boundary points of hazardous areas, aggregates them to obtain the dominant diffusion direction and velocity of the area, and calculates the shortest spatial distance and risk probability value by combining the spatial coordinates of key equipment and ventilation boundaries. Thus, it can accurately grasp the diffusion law after gas leakage, clarify the range, diffusion direction and velocity of potential hazardous areas, and judge the degree of risk by combining the positional relationship of key equipment, avoiding a general judgment of risk. Based on the risk probability value and leakage diffusion trend prediction, it matches a graded control strategy that is compatible with the current dominant process operating mode, and can accurately take control measures according to the degree of risk, effectively curb the expansion of hazardous areas and reduce the probability of safety accidents.
[0023] (4) The combustible and toxic gas monitoring and control system realizes the coordinated operation of the entire process of combustible and toxic gas monitoring and control by setting up a map construction module, a working condition mode recognition module, a dynamic risk assessment module and a control strategy execution module. The map construction module is responsible for building a dynamic process map that fits the actual production, providing basic data support for the entire system. The working condition mode recognition module obtains the status data of process equipment and production parameters in real time, and determines the current dominant process working condition mode in combination with the map to ensure that the system can adapt to the production operation status. The dynamic risk assessment module completes gas diffusion simulation, potential hazard area identification and risk probability value calculation through a pre-trained risk assessment model, and accurately predicts hidden dangers. The control strategy execution module matches the hierarchical control strategy according to the assessment results and executes the control command to form a closed-loop control, thereby realizing the automated and coherent operation from hidden danger prediction to risk control, reducing the error of manual intervention, improving the efficiency of monitoring and control, and ensuring that the control strategy is accurately matched with the production working condition and risk level, so as to fully guarantee the safe and stable operation of the production system.
[0024] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description
[0025] Figure 1 This is a flowchart of a method for monitoring and controlling combustible and toxic gases according to the present invention.
[0026] Figure 2 This is a flowchart illustrating the specific steps involved in determining the current dominant process operating mode in a method for monitoring and controlling combustible and toxic gases according to the present invention.
[0027] Figure 3 This is a block diagram of a combustible and toxic gas monitoring and control system according to the present invention. Detailed Implementation
[0028] Please see Figure 1 This invention provides a technical solution: a method for monitoring and controlling combustible and toxic gases, comprising the following steps: constructing a dynamic process map of a production system, with production equipment, material pipelines, and monitoring sensors as nodes, and material flow direction, energy transfer, and signal connection relationships as edges; acquiring real-time process equipment status data and production parameters, and determining the current dominant process operating mode by combining the dynamic process map and a pre-trained time-series pattern recognition model; inputting the gas concentration, environmental parameters, and process equipment status data corresponding to the current dominant process operating mode into a pre-trained risk assessment model to perform gas diffusion simulation calculations to generate preliminary leakage diffusion trend predictions, and calculating risk probability values and leakage diffusion trend predictions by combining the process equipment status data, specifically: identifying spatial regions in the simulated concentration field that exceed a preset threshold as potential hazardous areas; calculating the concentration gradient vector in the normal direction of the boundary of the region as the dominant diffusion direction and velocity; determining the relative positions of potential hazardous areas, key equipment, and ventilation boundaries based on the process equipment status data, and calculating risk probability values by combining the dominant diffusion direction and velocity; matching a graded control strategy adapted to the current dominant process operating mode based on the risk probability value and leakage diffusion trend predictions, and executing control commands.
[0029] Among them, the risk probability value The calculation expression is:
[0030] ;
[0031] in, The weighting coefficients are dynamically adjusted based on the current dominant process operating mode, and satisfy the following conditions: The dynamic adjustment rules for the weighting coefficients are preset in the system;
[0032] For example, in the "normal and stable production" mode, it might be set to... It focuses on diffusion dynamics;
[0033] In special operating conditions with lower risk tolerance, such as "start-up and shutdown" or "equipment maintenance," adjustments may be made to... This significantly improves sensitivity to proximity and time urgency;
[0034] The diffusion dynamics factor is based on the concentration gradient vector, and its value is positively correlated with the average concentration gradient magnitude at the boundary of the potential hazard area. It is obtained by normalizing the gradient magnitude to the [0, 1] interval.
[0035] For example, ;
[0036] in, It is a reference gradient value;
[0037] The device proximity factor is based on relative location. Its value depends on the shortest distance between the potentially hazardous area and all critical targets (equipment / boundaries). The closer the distance, the higher the factor value. When the distance is less than the minimum safe distance, the factor value reaches the upper limit of 1.
[0038] For example, ;
[0039] in, It is the distance to the t-th critical target;
[0040] This is a time urgency factor based on the estimated arrival time. Its value is negatively correlated with the shortest estimated arrival time; the shorter the time, the higher the factor value.
[0041] For example, ;
[0042] in, The preset time constant (e.g., 300 seconds);
[0043] Hierarchical control strategy based on The threshold range that the user is in, which is bound to the operating mode, is used for tiered triggering;
[0044] For example, for the "normal production" mode, the threshold range might be set as follows:
[0045] [0, 0.25) represents Level 1 warning, [0.25, 0.6) represents Level 2 intervention, and [0.6, 1.0] represents Level 3 interlocking.
[0046] For the "live maintenance" mode, the threshold is tightened to:
[0047] [0, 0.15) is Level 1, [0.15, 0.4) is Level 2, and [0.4, 1.0] is Level 3.
[0048] Specifically, the steps for constructing a dynamic process map are as follows:
[0049] Based on the production process flow diagram and equipment layout diagram, all nodes and their connection relationships are determined, specifically as follows:
[0050] Each independent functional unit and physical monitoring point is abstracted as a unique node. The connection edges between nodes are determined based on the process piping and instrumentation diagram (P&ID) and energy flow diagram, and the edge direction is consistent with the main flow direction.
[0051] Production equipment, material pipelines, and monitoring sensors are defined as nodes and stored in the graph, specifically as follows:
[0052] Define a multidimensional feature vector for each type of node, which includes ID, type, key design parameters and location information, and store it in key-value pairs.
[0053] For example, for a "centrifugal pump" device node, its feature vector may include:
[0054] [Equipment tag number "P-101A", type code "centrifugal pump", rated flow "100m³ / h", rated head "50m", motor power "22kW", installation coordinates (x, y, z)];
[0055] For a "hydrogen concentration sensor" node, its feature vector may include:
[0056] [Sensor tag number "AT-102", sensor type "electrochemical", monitored gas "H2", range "0-100%LEL", alarm threshold "20%LEL", installation coordinates (x, y, z), pipeline "P-201"];
[0057] The material flow direction, energy transfer, and signal connection relationships between nodes are defined as edges and stored in the graph. Specifically:
[0058] Define a relation type vector and dynamic weight for each edge. The dynamic weight is updated based on real-time process data (such as flow rate) and stored in an adjacency list structure.
[0059] Relational type vectors can use one-hot encoding, for example:
[0060] [1, 0, 0] represents material flow, [0, 1, 0] represents energy flow, and [0, 0, 1] represents signal flow.
[0061] Dynamic weights Initial values can be derived from design parameters (such as rated flow rate). Determined, and updated with real-time data (such as actual traffic). Normalization update, for example .
[0062] In this implementation plan, the specific steps for constructing a dynamic process map enable the map to better reflect actual production and possess dynamic adaptability. By reasonably abstracting nodes and determining connection relationships based on relevant drawings, and combining node feature vectors containing various necessary information, different types of nodes can be clearly distinguished, accurately presenting the layout and attributes of equipment, pipelines, and sensors in the production system. This allows the map to truly reflect the overall production situation. At the same time, by clarifying the relationship types of edges and setting dynamic weights that update with real-time process data, combined with a scientific storage structure, the map can adapt to the dynamic changes in the production process in real time and respond promptly to adjustments in process parameters.
[0063] Specifically, such as Figure 2As shown, the specific steps to determine the current dominant process operating mode are as follows:
[0064] The real-time acquired process equipment status data and production parameters are associated with corresponding nodes in the dynamic process graph, specifically as follows:
[0065] Read the corresponding real-time time series data according to the node type, and concatenate it with the node static features to form an enhanced map with time series features;
[0066] For example, for a pump node, real-time data may include:
[0067] Start / stop status (1 / 0), operating frequency (Hz), outlet pressure (MPa), current (A);
[0068] The correlated data map is then converted into the input format required by the time-series pattern recognition model, specifically as follows:
[0069] The temporal pattern recognition model is a hybrid model of graph attention network (GAT) and gated recurrent unit (GRU);
[0070] First, the node features are mapped to a high-dimensional space, and then passed through multiple layers of graph attention convolution.
[0071] In the In a layer, the aggregation of node features is determined by the attention-weighted sum of the features of its neighboring nodes, where the attention coefficient is... The calculation is as follows:
[0072] ;
[0073] This formula can dynamically assign importance to the different neighbors of each node;
[0074] For example, during the "reaction heating" phase, the connection between temperature-related sensor nodes and reactor nodes receives higher attention weight.
[0075] Finally, the node features of the entire graph are pooled to obtain a graph-level feature sequence.
[0076] Pooling operations, for example, can employ global average pooling, which means averaging the high-dimensional features of all nodes at the same time.
[0077] Input the converted graph into the model and receive its output of the current dominant process condition mode, which is as follows:
[0078] Input the graph-level feature sequence into a bidirectional GRU and input its final hidden state into a classifier. The output is the probability distribution of the predefined working condition mode, and the mode with the highest probability is taken as the output.
[0079] Manual confirmation is triggered when the maximum probability is below the confidence threshold;
[0080] The predefined operating condition modes can be divided according to specific processes, for example:
[0081] "Normal production", "catalytic reaction", "product separation", "catalyst regeneration", "shutdown cleaning", "emergency release", etc.
[0082] The confidence threshold can be set to 0.7. If the maximum probability output by the model is less than 0.7, the model is considered to be uncertain and the operator needs to confirm the current working condition based on experience.
[0083] In this implementation scheme, by processing the temporal changes of the nodes themselves, the graph attention mechanism can be used to actively mine and quantify the dynamic interactions and influence between nodes such as equipment and sensors, thereby capturing system-level operation mode characteristics that are difficult to detect by traditional methods. The model finally outputs a probabilistic judgment on predefined operating conditions, enabling the system to clearly know whether the current production is in a normal and stable operation, a start-up and shutdown transition, or a high-risk maintenance stage. This accurate identification of operating conditions gives the entire monitoring and control process the key scene perception and understanding capabilities.
[0084] Specifically, the steps for matching the hierarchical control strategy are as follows:
[0085] Based on the risk probability value, leakage diffusion trend prediction, and the current dominant process operating mode, the pre-stored strategy mapping table is queried. This table records the correspondence between operating mode, risk value range, diffusion trend, and hierarchical control strategy, specifically as follows:
[0086] The strategy mapping table has a multi-layer decision structure;
[0087] Using operating mode as the primary keyword;
[0088] The second key is the interval obtained by dynamically binning historical risk values under this model;
[0089] The third keyword is a combination of key attributes of the diffusion trend (such as whether it points to a key area and whether the arrival time exceeds the limit);
[0090] Through mapping function Unique retrieval strategy;
[0091] "Dynamic binning" refers to the process of dividing the bins based on historical data generated under this operating condition. Risk ranges are defined by the statistical distribution of values (such as decimals) rather than fixed thresholds, which allows risk classification to adapt to the risk characteristics of specific working conditions.
[0092] "Key attributes" can include Boolean conditions, such as:
[0093] C1: Does the potentially hazardous area include "open flame equipment" or "densely populated areas"?
[0094] C2: Does the dominant diffusion direction point to the "factory boundary" or "important facilities"?
[0095] C3: Is the estimated shortest arrival time less than the "emergency response time" (e.g., 3 minutes)?
[0096] The appropriate hierarchical control strategy is extracted from the strategy mapping table, specifically as follows:
[0097] The strategy is a structured instruction set, which includes a list of control actions, execution order, delay, and expected effect;
[0098] For example, a level 3 strategy targeting a risk value between 0.4 and 0.6 that satisfies condition C1 might have the following instruction set:
[0099] 1. Trigger on-site audible and visual alarms and central control alarm (execute immediately);
[0100] 2. Automatically turn on the ventilation system in the hazardous area and upwind direction to the maximum air volume (execution after a 2-second delay).
[0101] 3. Send a 50% load reduction instruction to potentially affected production units (execution after a 5-second delay);
[0102] 4. If the risk value does not decrease within 30 seconds, the upstream process valve closest to the leak source will be interlocked off (conditional execution).
[0103] The steps for dynamically adjusting during the query process include:
[0104] During the query, real-time monitoring is conducted to check for any sudden changes in the risk probability value and the predicted leakage spread trend. Specifically:
[0105] Monitor the instantaneous rate of change of risk probability value, diffusion direction and dangerous area area. If any rate of change exceeds the corresponding threshold, it is judged as a sudden change.
[0106] "Instantaneous rate of change" can be defined as the difference or ratio between the current value and the value of the previous sampling period (e.g., 1 second ago);
[0107] Thresholds are set based on parameter characteristics and safety requirements; for example, the threshold for the rate of change of risk probability values. It can be set to 0.2 / second, the threshold for the change in diffusion direction angle. It can be set to 45 degrees / second;
[0108] If a mutation causes the original query result to become invalid, the strategy mapping table will be queried again immediately based on the result after the mutation. Specifically:
[0109] Compare the severity levels of the strategies corresponding to the new and old query results. If the new strategy has a higher severity level, the original strategy is deemed invalid, the original command is interrupted, and the query is performed again.
[0110] The "severity level" of a strategy is predefined, for example;
[0111] Level 1 early warning is 1, Level 2 intervention is 2, and Level 3 interlocking is 3;
[0112] If the original strategy level is 2 (intervention) and the new query strategy level is 3 (interlocking), then it is determined to be invalid;
[0113] The strategy obtained from the re-query will be used as the final hierarchical control strategy to be executed, specifically as follows:
[0114] Set the new strategy as the highest priority and execute it, and record this adjustment event;
[0115] The system will generate a dynamic adjustment log containing a timestamp, the original policy ID, the new policy ID, and the reason for the change.
[0116] It also includes the step of optimizing the policy mapping table:
[0117] Record the strategy executed each time, along with its corresponding risk outcomes and operating conditions, specifically as follows:
[0118] Record the state when the policy is triggered, the policy identifier executed, and the comprehensive reward value for post-event evaluation. The reward value takes into account the security outcome, economic loss, and system stability.
[0119] Reward Value Specific structural examples:
[0120] ;
[0121] in, For safety rewards, +10 is given for no accidents, 0 for minor leaks, and -100 for fires and explosions.
[0122] The economic incentives are inversely proportional to the downtime and production losses caused by the implementation of the strategy;
[0123] To ensure stable rewards, the rewards are inversely proportional to the fluctuations in process parameters caused by the implementation of the strategy.
[0124] Weight Reflecting the management focus, usually maximum;
[0125] Periodically, based on reinforcement learning algorithms, the correspondence in the policy mapping table is iteratively optimized with the goal of maximizing the treatment effect. Specifically:
[0126] The Q-learning algorithm is used, which utilizes historical experience tuples. Update strategy value;
[0127] Value updates follow these rules:
[0128] ;
[0129] After learning is complete, for each state Choose the one with the highest The action (policy) of the value is used to update the policy mapping table. ;
[0130] The update process is offline; the system maintains an experience replay buffer, accumulating a sufficient amount of data. After recording, start batch learning;
[0131] state That is, the combination of conditions that trigger the strategy (operating mode, risk range, diffusion trend conditions), and the action. That is, the policy identifier and the next state. The new state observed after performing the action and waiting for a period of time;
[0132] Through continuous learning, the system can gradually associate higher-reward strategies with corresponding states, thereby optimizing the strategy mapping table.
[0133] In this implementation plan, based on a deep analysis of the current production status and risk situation, the system can quickly match the most suitable handling plan from the pre-set strategies, ensuring that the control measures are highly consistent with the specific scenario. When the risk situation changes abruptly during the response process, the system can perceive it in real time and decisively upgrade to a higher level of response strategy, effectively avoiding the expansion of consequences that may be caused by untimely or insufficient response. By recording the complete context and actual effect of each action and applying reinforcement learning algorithms to continuously train and iterate the strategy library, the system enables its decision-making logic to continuously absorb practical experience, ultimately forming an intelligent safety decision-making closed loop with experience accumulation and evolution capabilities.
[0134] Specifically, the steps for identifying potentially hazardous areas are as follows:
[0135] The concentration field generated by the gas diffusion simulation is subjected to spatial grid traversal and threshold comparison, specifically as follows:
[0136] The concentration value of each grid point in the simulated three-dimensional discrete concentration field is compared with a dynamic threshold to generate a binary label matrix;
[0137] "Dynamic threshold" can be a fixed alarm threshold (such as 20% of the lower explosion limit LEL) or a value that is finely adjusted according to the current dominant process operating mode (such as using a lower threshold at night or under special conditions).
[0138] Connectivity analysis is performed on adjacent grid nodes where the concentration exceeds the threshold, and these nodes are marked as independent potential hazard areas. Specifically:
[0139] Perform three-dimensional connected component analysis on the binary matrix to classify spatially adjacent out-of-range grids into the same region;
[0140] "Adjacent" usually refers to a 6-neighborhood (up / down, left / right, front / back) or 26-neighborhood in three-dimensional space;
[0141] Analysis algorithms can employ seed-filling or disjoint-set data structure methods;
[0142] Output the set of spatial boundary coordinates for all potentially hazardous areas, specifically:
[0143] Calculate the minimum and maximum indices of each connected component in the three coordinate directions, and convert them into physical space coordinates to form a bounding box;
[0144] For example, if a connected domain spans grid indices i=10 to 15 in the X direction with a grid spacing of 1 meter, then its physical X-direction boundary is [9.5, 15.5] meters (considering the grid center).
[0145] In this implementation scheme, by performing spatial grid traversal and threshold comparison on the concentration field, and combining it with a dynamic threshold that can be finely adjusted according to the current dominant process operating mode, the judgment of concentration exceeding the standard is made more in line with the actual production scenario, avoiding misjudgment or omission caused by fixed standards. Three-dimensional connected domain analysis of the exceeding grid can classify adjacent exceeding grids into independent danger areas, clearly distinguish different leakage ranges, avoid area confusion, and output a set of spatial boundary coordinates, which can clearly define the specific range of the danger area, making it easier to accurately grasp the area affected by the leakage.
[0146] Specifically, the steps for calculating the dominant diffusion direction and velocity are as follows:
[0147] For points on the boundary of a potentially hazardous area, the spatial gradient vector of the concentration field and the unit normal vector of the boundary are calculated as follows:
[0148] Extract the boundary points of the region, calculate the concentration gradient of each point by combining the central difference method with the concentration values of the surrounding grid, and obtain the unit normal vector pointing out of the region at that point.
[0149] Boundary points can be obtained by extracting concentration isosurfaces (e.g., concentration = threshold);
[0150] The normal vector can be obtained by calculating the direction of the perpendicular line to the local tangent plane of the isosurface at that point;
[0151] Projecting the spatial gradient vector onto the direction of the unit normal vector yields the dominant diffusion direction and velocity scalar at that point, specifically:
[0152] The concentration gradient vector is projected onto the unit normal direction, and the propulsion velocity of the diffusion front along the normal direction at that point is estimated based on Fick's law and the threshold concentration.
[0153] Fick's law states that diffusion flux is proportional to the concentration gradient. It assumes the concentration at the diffusion front is approximately a threshold value. The diffusion coefficient is Then the propulsion speed ,in The unit normal vector;
[0154] The scalar values at all points on the aggregation boundary generate the dominant diffusion direction and velocity in that region, specifically:
[0155] Using the area represented by each boundary point as the weight, the velocity and direction vectors are weighted and averaged to obtain the overall dominant diffusion direction and average velocity of the region.
[0156] The overall direction is the unit vector after weighted average, and the overall velocity is the scalar value after weighted average.
[0157] In this implementation scheme, by extracting the boundary points of potential hazardous areas and calculating the concentration gradient vector and the boundary unit normal vector, the accuracy of the basic data is ensured, which conforms to the actual physical laws of gas diffusion. The spatial gradient vector is projected onto the direction of the unit normal vector, and the diffusion velocity at a single point is estimated in combination with relevant laws, making the calculation of diffusion parameters at a single point more reasonable. The area represented by the boundary point is used as the weight to perform a weighted average of all single-point scalars to obtain the overall dominant diffusion direction and velocity of the region, avoiding the bias caused by single-node data and comprehensively reflecting the diffusion characteristics of the region.
[0158] Specifically, the steps for determining the relative position and calculating the risk probability value are as follows:
[0159] The spatial coordinates of key equipment and ventilation boundaries are extracted from the process equipment status data, specifically as follows:
[0160] Obtain its real-time three-dimensional coordinates or spatial range representation from a predefined list of key equipment and ventilation boundary information;
[0161] The list of key equipment includes, but is not limited to:
[0162] Open flame equipment (heating furnace), high temperature equipment, power distribution room, control room, and personnel living area;
[0163] Ventilation boundaries include: doors and windows, ventilation openings, and factory boundary fences;
[0164] Coordinates can be imported from the factory's 3D digital model or design drawings;
[0165] Calculate the shortest spatial distance between the potential hazard area and the above coordinates. If it is less than the safety buffer distance, the area is considered to be adjacent. Specifically:
[0166] Calculate the spatial distance between each potential hazard area (represented by its center of gravity or boundary) and each critical target, and compare it with the corresponding safety buffer distance for that target;
[0167] Safe buffer distance It can be preset according to the properties of the gas (explosion limits, toxicity), the attributes of the target equipment (explosion protection level) and industry safety standards;
[0168] For example, the safe buffer distance between a hydrogen leak source and an open flame device may be set at 15 meters;
[0169] Based on location proximity, dominant diffusion direction and speed, the arrival time of the risk is calculated, and the risk probability value is calculated comprehensively, as follows:
[0170] For each "location-adjacent" pairing, the arrival time is estimated by combining the regional diffusion velocity, the angle between the diffusion direction and the target direction. The local risk probability value is a function of arrival time, concentration level, and target vulnerability.
[0171] The global risk probability value is taken as the maximum value or weighted sum of all local risks;
[0172] Arrival time estimation formula:
[0173] ;
[0174] in For distance, For diffusion rate, The angle between the diffusion direction and the target direction;
[0175] like If so, it is assumed that the destination will not be reached directly;
[0176] Local risks ,function It can be a monotonically decreasing function;
[0177] For example ,in The target vulnerability coefficient, is the time constant.
[0178] In this implementation plan, the spatial coordinates of key equipment and ventilation boundaries are extracted from the process equipment status data to match the actual layout of the plant, ensuring the accuracy of location judgment. The shortest spatial distance between potential hazardous areas and key targets is calculated, and the proximity relationship is determined by combining the safety buffer distance. The degree of risk association is clearly distinguished. The arrival time of risk is estimated by combining the proximity relationship, the dominant diffusion direction and speed. The global and local risk probability values are calculated comprehensively, and all kinds of influencing factors are fully considered to avoid the one-sidedness of risk judgment.
[0179] Please see Figure 3 This invention provides a technical solution: a combustible and toxic gas monitoring and control system, comprising: a graph construction module for constructing a dynamic process graph of a production system, wherein the graph uses production equipment, material pipelines, and monitoring sensors as nodes, and material flow direction, energy transfer, and signal connection relationships as edges; a working condition mode recognition module for acquiring process equipment status data and production parameters in real time, and determining the current dominant process working condition mode by combining the dynamic process graph with a pre-trained time-series pattern recognition model; a dynamic risk assessment module for inputting gas concentration, environmental parameters, and process equipment status data corresponding to the current dominant process working condition mode into a pre-trained risk assessment model to perform gas diffusion simulation calculations to generate preliminary leakage diffusion trend predictions, and calculating risk probability values and leakage diffusion trend predictions by combining process equipment status data; and a control strategy execution module for matching a hierarchical control strategy adapted to the current dominant process working condition mode based on the risk probability value and leakage diffusion trend predictions, and executing control commands.
[0180] The system is deployed using industrial edge computing gateways or industrial control computers as carriers, integrating real-time data buses and industrial communication protocols, and interconnecting with the factory information layer (MES) and control layer (PLC / DCS).
[0181] The topology construction module obtains real-time process data from DCS / SCADA via OPC UA / MODBUS TCP protocol and parses static equipment topology from plant information models (such as AVEVA AIM) or CAD drawings;
[0182] The working condition pattern recognition module and the dynamic risk assessment module, as the core analysis engines, are deployed on an edge server with GPU acceleration capabilities.
[0183] The system sets up independent safety relay circuits for the three-level control commands to ensure the reliability of the highest priority interlocking action;
[0184] This safety relay circuit (or safety PLC) is separated from the main control system, receives hard-wired signals or safety communication protocol (such as PROFIsafe) signals from the edge server of this system, and directly drives the final actuators such as emergency shut-off valves (ESD valves) to meet the SIL2 and above safety level requirements;
[0185] The human-machine interface (HMI) uses a 3D plant map as the base map and integrates dynamic process diagrams and risk fields for visualization;
[0186] The visualized content includes:
[0187] The nodes (equipment icons) and edges (flow lines) of the dynamic process diagram.
[0188] Real-time / predicted gas concentration fields are displayed as overlaid semi-transparent color patches, with colors mapping concentration levels.
[0189] Arrow symbols are used to dynamically indicate the dominant diffusion direction and speed of potentially hazardous areas;
[0190] Highlight devices in alarm status and control commands being executed;
[0191] The system also provides functions for historical data query, event retrieval, and report generation.
[0192] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.
[0193] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.
Claims
1. A method for monitoring and controlling combustible and toxic gases, characterized in that, Includes the following steps: Construct a dynamic process map of the production system. The map uses production equipment, material pipelines and monitoring sensors as nodes, and material flow direction, energy transfer and signal connection relationship as edges. Real-time acquisition of process equipment status data and production parameters, combined with dynamic process graphs, and determination of the current dominant process operating mode through a pre-trained time-series pattern recognition model; The gas concentration, environmental parameters, and process equipment status data corresponding to the current dominant process operating mode are input into the pre-trained risk assessment model to perform gas diffusion simulation calculations to generate preliminary leakage diffusion trend predictions. The risk probability value and leakage diffusion trend predictions are then calculated in combination with the process equipment status data. Specifically, spatial areas in the simulated concentration field that exceed a preset threshold are identified as potential hazardous areas. Calculate the concentration gradient vector along the normal direction of the boundary of the region as the dominant diffusion direction and velocity; determine the relative positions of potential hazardous areas and key equipment with ventilation boundaries based on process equipment status data, and calculate the risk probability value in combination with the dominant diffusion direction and velocity; Based on the risk probability value and leakage diffusion trend prediction, a hierarchical control strategy adapted to the current dominant process operating mode is matched, and control commands are executed.
2. The method for monitoring and controlling combustible and toxic gases according to claim 1, characterized in that, The specific steps for constructing a dynamic process map are as follows: Based on the production process flow diagram and equipment layout diagram, determine all nodes and their connection relationships; Define production equipment, material pipelines, and monitoring sensors as nodes and store them in the graph; The material flow direction, energy transfer and signal connection relationship between nodes are defined as edges and stored in the graph.
3. The method for monitoring and controlling combustible and toxic gases according to claim 1, characterized in that, The specific steps to determine the current dominant process operating mode are as follows: The real-time acquired process equipment status data and production parameters are associated with the corresponding nodes in the dynamic process graph. The correlated data map is converted into the input format required by the time-series pattern recognition model; Input the converted graph into the model and receive its output of the current dominant process condition mode.
4. The method for monitoring and controlling combustible and toxic gases according to claim 1, characterized in that, The specific steps for matching hierarchical control strategies are as follows: Based on the risk probability value, leakage diffusion trend prediction and the current dominant process operating mode, the pre-stored strategy mapping table is queried, which records the correspondence between operating mode, risk value range, diffusion trend and graded control strategy. Extract the appropriate hierarchical control strategy from the strategy mapping table.
5. The method for monitoring and controlling combustible and toxic gases according to claim 4, characterized in that, The steps for dynamically adjusting during the query process include: During the query, monitor in real time whether there are any sudden changes in the risk probability value and leakage spread trend prediction; If a mutation causes the original query result to become invalid, the strategy mapping table will be queried again immediately based on the result after the mutation. The strategy obtained from the re-query will be used as the final hierarchical control strategy to be executed.
6. The method for monitoring and controlling combustible and toxic gases according to claim 4, characterized in that, It also includes optimizing the policy mapping table, the specific steps of which are as follows: Record the strategy executed each time, along with its corresponding risk outcomes and operating conditions; Periodically, based on reinforcement learning algorithms, the correspondence in the policy mapping table is iteratively optimized with the goal of maximizing the treatment effect.
7. The method for monitoring and controlling combustible and toxic gases according to claim 1, characterized in that, The specific steps for identifying potentially hazardous areas are as follows: Spatial grid traversal and threshold comparison are performed on the concentration field generated by gas diffusion simulation calculation; Connectivity analysis is performed on adjacent grid nodes where the concentration exceeds the threshold, and they are marked as independent potential hazard areas. Output the set of spatial boundary coordinates for all potentially hazardous areas.
8. The method for monitoring and controlling combustible and toxic gases according to claim 1, characterized in that, The specific steps for calculating the dominant diffusion direction and velocity are as follows: For points on the boundary of a potentially hazardous area, calculate the spatial gradient vector of the concentration field and the unit normal vector of the boundary. Projecting the spatial gradient vector onto the direction of the unit normal vector yields the dominant diffusion direction and velocity scalar at that point. The scalars of all points on the aggregation boundary generate the dominant diffusion direction and velocity in that region.
9. The method for monitoring and controlling combustible and toxic gases according to claim 1, characterized in that, The specific steps for determining the relative position and calculating the risk probability value are as follows: Extract the spatial coordinates of key equipment and ventilation boundaries from process equipment status data; Calculate the shortest spatial distance between the potential danger zone and the above coordinates. If it is less than the safety buffer distance, the location is considered to be close. Based on location proximity, dominant diffusion direction and speed, the arrival time of the risk is calculated and the risk probability value is calculated comprehensively.
10. A combustible and toxic gas monitoring and control system, employing the combustible and toxic gas monitoring and control method according to any one of claims 1-9, characterized in that, include: The graph construction module is used to build a dynamic process graph of the production system. The graph uses production equipment, material pipelines and monitoring sensors as nodes, and material flow direction, energy transfer and signal connection relationship as edges. The operating condition pattern recognition module is used to acquire real-time status data and production parameters of process equipment, and combine them with dynamic process graphs to determine the current dominant process operating condition pattern through a pre-trained time-series pattern recognition model. The dynamic risk assessment module is used to input the gas concentration, environmental parameters and process equipment status data corresponding to the current dominant process operating mode into the pre-trained risk assessment model to perform gas diffusion simulation calculations to generate preliminary leakage diffusion trend predictions, and combine the process equipment status data to calculate and generate risk probability values and leakage diffusion trend predictions. The control strategy execution module is used to match a hierarchical control strategy that is adapted to the current dominant process operating mode based on the risk probability value and leakage diffusion trend prediction, and execute control commands.