Chemical production equipment safety emergency response system based on real-time data

Abstract

This invention belongs to the field of next-generation information technology and discloses a safety emergency response system for chemical production equipment based on real-time data. The system includes: an anomaly analysis module: storing current and historical dynamic propagation directed graphs into a first-in-first-out (FIFO) cache queue to obtain a time-series graph sequence; performing dynamic feature calculation based on the time-series graph sequence to output the dynamic features of the edges; and performing edge-level time-series association mining based on the dynamic features of the edges to obtain a risk dynamic propagation model; and an emergency response module: extracting critical vulnerable paths based on the risk dynamic propagation model to obtain a set of equipment anomaly propagation chains; and capturing abnormal equipment based on the set of equipment anomaly propagation chains to achieve safety emergency response control for chemical production equipment. This system can intervene before a fault causes widespread impact, significantly improving the foresight of the response.

Description

Technical Field

[0001] This invention relates to the field of next-generation information technology, and in particular to a safety emergency response system for chemical production equipment based on real-time data. Background Technology

[0002] In the field of chemical production, equipment safety is the lifeline for ensuring continuous and stable operation and preventing major accidents. Traditional safety emergency response systems mainly rely on single-point alarm mechanisms based on fixed thresholds. That is, a safety upper limit and lower limit are set for each key sensor (such as temperature, pressure, and flow). When the monitored data exceeds the threshold, an alarm is triggered. The core logic of this method has not changed fundamentally. Although it can alert to obvious and isolated equipment anomalies, it is essentially a passive response.

[0003] Existing emergency response systems for chemical production equipment lack systemic risk insight when facing complex chemical processes. Because chemical production processes involve dense equipment and highly coupled processes, an anomaly in one piece of equipment often triggers a chain reaction through material and energy flow. Existing single-point alarm systems cannot reveal the path and scope of this anomaly propagation, making it difficult for maintenance personnel to quickly locate the root cause and predict the impact chain of the fault from a massive number of alarms, easily missing the optimal intervention opportunity. Secondly, the response strategies of emergency response systems are static and rigid, usually based on historical case studies, making it difficult to adapt to real-time changing operating conditions. When the anomaly occurs with subtle differences from historical cases, static strategies may be ineffective or even misleading. Therefore, this paper proposes a real-time data-based emergency response system for chemical production equipment to address these problems. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art and to achieve the above objectives, the present invention proposes the following technical solution: A real-time data-based emergency response system for chemical production equipment includes: Status acquisition module: Collects initial status data of chemical equipment, and obtains equipment time-synchronized data through timestamp alignment and data stream synchronization. State synchronization module: Based on the device time-space data, the time window is dynamically divided according to the data fluctuation frequency to obtain spatiotemporal synchronization data. Based on the spatiotemporal synchronization data, the correlation coefficient matrix between sensors is obtained. Based on the correlation coefficient matrix, Granger causality test is performed to construct a dynamic propagation directed graph. Anomaly Analysis Module: Stores the current and historical dynamic propagation directed graphs into a first-in-first-out (FIFO) cache queue to obtain a time-series graph sequence. Based on the time-series graph sequence, it performs dynamic feature calculation to output the dynamic features of the edges. Based on the dynamic features of the edges, it performs edge-level time-series association mining to obtain a risk dynamic propagation model. Emergency Response Module: Based on the risk dynamic propagation model, critical vulnerability paths are extracted to obtain a set of equipment anomaly propagation chains. Based on the set of equipment anomaly propagation chains, abnormal equipment is captured to achieve safety emergency response control for chemical production equipment.

[0005] The process of obtaining device time and space data is as follows: Sensors are deployed to collect initial state data of chemical equipment. This initial state data is then timestamped, and a unified analysis time base is established. and reference sampling interval For any sensor i, at each reference time point The above method uses linear interpolation to calculate the time of the i-th sensor at the reference time point using two neighboring data points in its original time series. Approximate data values Then approximate data values Organized into a multidimensional time series matrix, where rows represent time points. The column represents sensor i, and the matrix elements are the corresponding approximate data values, thus obtaining the device time and space data X.

[0006] The process of obtaining spatiotemporal synchronization data is as follows: By sliding a small segment containing the 10 most recent data points across the time series, the variance of the device time-space data X within that sliding segment is calculated. And take the average of the variances of all sensors as the overall fluctuation index U of the current segment; Preset a high volatility threshold and a low volatility threshold ,when When the analysis time window is shortened to 5 data points, a short-window spatiotemporal data block is generated. The time window for extended analysis is 30 data points, generating a long-window spatiotemporal data block. Spatiotemporal synchronization data D is obtained based on short-window and long-window spatiotemporal data blocks.

[0007] The process of obtaining the correlation coefficient matrix is ​​as follows: For any two data sequences corresponding to sensors i and j in the spatiotemporal synchronization data D, obtain the Pearson correlation coefficient of the data sequences. Then, a fixed window is preset, and the Pearson correlation coefficients of all two data sequences within the fixed window are used to form a correlation coefficient matrix R.

[0008] The process of obtaining a dynamically propagating directed graph is as follows: Set a correlation coefficient threshold Only for those who meet the requirements Granger causality test is performed on sensor pair (i,j), and a restricted model is constructed. With the full model ; Compare the sum of squared residuals of the two models. When the prediction accuracy of the full model is higher than that of the restricted model, sensor j is taken as the Granger cause of sensor i. Represent all causal relationships that pass the Granger causality test as directed edges, and construct a directed causal relationship network. According to directed causal networks The correlation coefficient matrix R assigns causal relationships to the association strength, forming a complete weighted directed graph, i.e., a dynamic propagation directed graph. Where W is the set of edge weights, V is the set of nodes corresponding to all sensors, and each edge in the set of directed edges E represents... .

[0009] Constrained Model With the full model Represented as: Constrained model: ; Full model: ; Where m is the lag order, These are the coefficients of the constrained model. It is the error term of the constrained model. For the full model coefficients, It is the error term of the entire model. Indicates a restricted model index. Represents the full model index. This represents the spatiotemporal synchronization data corresponding to sensor i. This represents the spatiotemporal synchronization data corresponding to sensor j.

[0010] The time sequence is propagated through a directed graph that is dynamically fed into the current window. The image in the current window Store the data in a first-in-first-out (FIFO) cache queue in chronological order, save the graphs of the K most recent windows, and obtain the time sequence graph sequence.

[0011] The process of outputting the dynamic features of the edges is as follows: The dynamic characteristics of the edge include moving average and relative rate of change; The moving average is for each edge The weights are the moving averages over the past K windows, and the relative rate of change is the ratio of the current relative rate of change to the baseline weights, where the baseline weights are... , This represents the weight of the current k-th window.

[0012] The process of obtaining the risk dynamic propagation model is as follows: Calculate each edge The Pearson correlation coefficient of the weight of the edge with all other edge weights is used to screen out the edges with a correlation coefficient ≥ 0.6 to form a group of associated edges, and the frequency f of high-risk associated edges in the historical K windows is counted. When f≥0.2, it is considered that a strong change occurs on average once every 5 windows; When ≥60% of the edges in an edge group undergo strong changes simultaneously, the edge group is marked as a high-risk associated edge group. And construct a propagation potential index for each node. ,in, The average acceleration of all incoming edges of a node. Let α be the average relative rate of change of all edges, and β be the weighting coefficients, where α + β = 1. Node potential energy is classified according to propagation potential energy index. It is a high potential energy node; When the proportion of high-potential nodes is ≥30%, the next window length is shortened to a short window of 5 data points. When the proportion of high-potential nodes is ≤10%, the next window length is extended to a long window of 30 data points to complete edge-level temporal correlation mining. Dynamically propagating the directed graph The generation process serves as the basic architecture of the model, and the process of edge-level temporal correlation mining serves as the model update process, thus obtaining the risk dynamic propagation model M.

[0013] The process of obtaining the set of device anomaly propagation chains is as follows: The nodes whose last three data points in the current sensor window consistently exceed the normal range are designated as initial abnormal nodes. And on the risk dynamic propagation model M, find the initial abnormal node. Cumulative weight of paths from the origin to all other reachable nodes : , Indicates the weight of the edge; Dijkstra's algorithm is introduced to transform edge weights into distances and find the shortest distance, especially for nodes from outliers. arrive path , For each abnormal node Find a critical path to different target nodes, collect all critical paths found starting from all initial abnormal nodes, remove completely duplicate paths, and accumulate weights according to the paths. Sort the results from largest to smallest to obtain the set of anomaly propagation chains. , Indicates the number of abnormal propagation chains.

[0014] The present invention has the following beneficial effects: 1. By constructing a dynamic risk propagation model, this system not only identifies anomalies in individual devices but also reveals in real time the potential propagation paths and risk levels of these anomalies within the device network. This transforms security response from passively dealing with individual fault points to proactively predicting and cutting off the entire fault propagation chain, achieving true systemic security protection.

[0015] 2. Improved accuracy and foresight in emergency response: By extracting critical vulnerability paths and capturing abnormal equipment, the system can accurately locate the source of the anomaly and the affected critical equipment, and determine the priority of handling based on the cumulative weight of the propagation path. This changes the traditional experience-based response model, making emergency measures more precise and enabling intervention before the failure causes widespread impact, thus significantly improving the foresight of the response.

[0016] 3. Enhanced the adaptive and dynamic optimization capabilities of emergency strategies. By introducing a strategy decay correction mechanism based on real-time data and historical case matching, the generated emergency instructions can be adaptively adjusted according to the differences between the current working conditions and historical cases, avoiding the limitations of static strategies and ensuring that the system can dynamically optimize subsequent responses based on the effects of control actions, forming a continuously improving intelligent response closed loop.

[0017] 4. Reduced false alarm rate and improved alarm timeliness: Based on the construction of a dynamic propagation graph of Granger causality, it can filter out accidental and non-causal data fluctuations, thereby effectively reducing false alarms. Combined with adaptive window partitioning technology, it can reduce the computational load when the chemical equipment is stable and quickly capture transient anomalies when the data fluctuates drastically, thereby improving the timeliness of alarms for real anomalies while ensuring system reliability. Attached Figure Description

[0018] Figure 1 This is a system block diagram of a chemical production equipment safety emergency response system based on real-time data proposed in this invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1 like Figure 1As shown, the present invention proposes a real-time data-based emergency response system for chemical production equipment, comprising: Status acquisition module: Collects initial status data of chemical equipment, and obtains equipment time-space data through timestamp alignment and data stream synchronization. A sensor network deployed on chemical equipment such as reactors, distillation columns, and pipelines collects raw data once per second at a fixed frequency. The sensors include pressure sensors, temperature sensors, flow sensors, and concentration sensors. Each sensor independently generates a signal sequence that changes over time, outputting the initial state data of the chemical equipment. The initial state data of chemical equipment is a multidimensional, discrete time series data set, represented as follows: ; Because there are slight differences in the sampling clock and communication delay of different sensors, direct use will lead to analysis bias, and preprocessing is required. First, timestamp alignment and data stream synchronization are performed. Timestamp alignment adds a high-precision, uniform timestamp to each sensor data stream, using the Network Time Protocol (NTP) or Precision Clock Synchronization Protocol to give all data packets a uniform time tag. Data stream synchronization processing is achieved by setting a unified analysis time base. and reference sampling interval For any sensor i, at each reference time point The above method uses linear interpolation to utilize two neighboring data points in the original time series. and Calculate approximate data values The formula is: ; in, It refers to the data in the sensor's raw data sequence that is earlier than or equal to the reference time point. Distance from the reference time point The timestamp of the most recent actual sampling point It is in the sensor's raw data sequence that the data point is later than the reference time point. Distance from the reference time point The timestamp of the most recent actual sampling point For the i-th sensor at the timestamp The actual measured value at that location For the i-th sensor at the timestamp The actual measured value at that location This represents the value of the i-th sensor at the reference time point, estimated using a linear interpolation method. Approximate data values; Furthermore, This ensures that all sensors have corresponding data points at all times; Then approximate data values Organized into a multidimensional time series matrix, where rows represent time points. The column represents sensor i, and the matrix elements are the corresponding approximate data values, thus obtaining the device time and space data X.

[0021] State synchronization module: Based on the device time-space data, the time window is dynamically divided according to the data fluctuation frequency to obtain spatiotemporal synchronization data. Based on the spatiotemporal synchronization data, the correlation coefficient matrix between sensors is obtained. Based on the correlation coefficient matrix, Granger causality test is performed to construct a dynamic propagation directed graph. First, acquire spatiotemporal synchronization data: First, the variance of the time-space data X within the sliding time segment is calculated. The adaptive window partitioning unit takes the device time-space data X as input and slides across the time series through a small sliding segment containing the 10 most recent data points. For each sensor i, the variance of its corresponding device time-space data X within that sliding segment is calculated. The average variance of all sensors is taken as the overall fluctuation index U of the current segment; Then preset a high volatility threshold. and a low volatility threshold ; when When this occurs, it indicates that the process is in a rapidly changing or unstable state, requiring more detailed analysis. In this case, the analysis time window length is shortened to 5 data points to generate a short-window spatiotemporal data block. The short window can capture transient anomalies more quickly. when When the time window is 30 data points, it indicates that the process is stable. At this point, the analysis time window is extended to 30 data points to generate a long-window spatiotemporal data block. The long window can provide more stable statistical characteristics and reduce the computational load. The final output spatiotemporal synchronization data D is a three-dimensional data tensor with dimensions of [number of windows, window length, number of sensors]. Each window is an L×n matrix, where L represents the window length and n is the number of sensors. Then calculate the correlation coefficient matrix between the sensor pairs: For any two data sequences corresponding to sensors i and j in the spatiotemporal synchronization data D, obtain the Pearson correlation coefficient of the data sequences. Then, a fixed window is preset, and the Pearson correlation coefficients of all two data sequences within the fixed window are used to form a correlation coefficient matrix R. Granger causality inference is performed based on the correlation coefficient matrix to obtain a directed causal network: Set a correlation coefficient threshold Only for those who meet the requirements The sensor pairs (i,j) are then used for subsequent causality testing. For the selected sensor pairs, perform Granger causality tests; Specifically, the Granger causality test is a statistical hypothesis test used to determine whether the historical values ​​of one variable can help predict the current value of another variable; Establish two autoregressive models: Constrained model: ; Full model: ; Where m is the lag order, These are the coefficients of the constrained model. It is the error term of the constrained model. For the full model coefficients, It is the error term of the entire model. Indicates a restricted model index. Represents the full model index. This represents the spatiotemporal synchronization data corresponding to sensor i. This represents the spatiotemporal synchronization data corresponding to sensor j; By comparing the sum of squared residuals of the two models, if the prediction accuracy of the full model is higher than that of the restricted model, then sensor j is said to be a Granger cause of sensor i, denoted as . ; Represent all valid causal relationships as directed edges, construct a directed graph structure, and output the directed causal relationship network. Here, the node set V corresponds to all sensors, and each edge in the directed edge set E represents... That is, sensor j has a statistically significant Granger causal effect on sensor i; Then construct the dynamic propagation directed graph: First, input the directed causal network. And the correlation coefficient matrix R, assigning the correlation strength to the causal relationship, forming a complete weighted directed graph, for each directed edge in the directed causal relationship network. Extract the absolute value of the corresponding correlation coefficient from the correlation coefficient matrix R. This is used as the initial weight of the directed edge. Weight This reflects the strength of causal influence from node j to node i, thus yielding a dynamically propagating directed graph. Where W is the set of edge weights, ∈W; Specifically, the dynamic propagation directed graph describes the possible propagation directions and intensity of anomalies between sensor nodes within the current analysis window.

[0022] Anomaly Analysis Module: Stores the current and historical dynamic propagation directed graphs into a first-in-first-out (FIFO) cache queue to obtain a time-series graph sequence. Based on the time-series graph sequence, it performs dynamic feature calculation to output the dynamic features of the edges. Based on the dynamic features of the edges, it performs edge-level time-series association mining to obtain a risk dynamic propagation model. Input the dynamic propagation directed graph of the current window The image in the current window Store the graphs in a first-in, first-out (FIFO) buffer queue in chronological order, saving the graphs of the K most recent windows (e.g., K=20), to obtain the time sequence graph sequence. This sequence constitutes the core data structure of the risk dynamic propagation model, where k is the window index and K is the total index; The process of calculating the dynamic features of the output edge is as follows: The dynamic characteristics of an edge include moving average and relative rate of change; First, a moving average is calculated for each edge. The weights are the moving averages over the past K windows, expressed by the formula:

[0023] Where λ is the smoothing factor, It is the average weight of the current k-th window. , It is the moving average of the previous time step. It is the moving average at the current moment, reflecting the baseline level of the edge weight; Furthermore, the ratio of the current moving average to the baseline weight is calculated to obtain the relative rate of change, expressed by the formula: ,in, Indicates the baseline weight. The larger the edge, the more drastic the change in the propagation relationship is. Then, edge-level temporal association mining is performed based on the dynamic features of the edges: First, edge-level temporal association mining is performed to calculate the time-series associations for each edge. The Pearson correlation coefficients of the weight sequence with all other edge weight sequences are used to select edges with a correlation coefficient ≥ 0.6 to form associated edge groups, for example: side and The correlation coefficient is 0.75, so the two belong to the same associated edge group; Then, the dynamic characteristics of the associated edge groups are jointly analyzed. When ≥60% of the edges in the group show strong changes and accelerated changes at the same time, the edge group is marked as a high-risk associated edge, indicating that the equipment abnormality may spread rapidly along the associated chain. Among them, the frequency f of high-risk associated edges within K historical windows is statistically analyzed. When f≥0.2, it means that a strong change occurs on average once every 5 windows. Then, a propagation potential index is constructed for each node, taking into account the dynamic characteristics of the node's incoming and outgoing edges: ; in, To propagate the potential energy index, The average relative rate of change for all edges reflects the ability of a node to propagate as a source of anomalies. The average acceleration of all edges of a node reflects the sensitivity of the node as an affected object. α and β are weight coefficients, and α+β=1. Node potential energy is classified according to propagation potential energy index. These are high-potential nodes, meaning they are likely to become key nodes for the propagation of anomalies. It is a mid-potential node. It is a low potential node; Based on the real-time calculation of node potential energy distribution by the model, when the proportion of high potential energy nodes is ≥30%, the next window length is automatically shortened to a short window of 5 data points to improve the anomaly response speed. When the proportion of high-potential nodes is ≤10%, the window is automatically extended to 30 data points to reduce the computational load and complete the edge-level temporal correlation mining. Dynamically propagating the directed graph The generation process serves as the basic architecture of the model, the process of edge-level temporal association mining serves as the model update process, and the risk dynamic propagation model M is trained. The training process is as follows: For every 100 windows accumulated, normal and abnormal samples from historical data are calibrated through manual annotation or equipment failure records. The relative change rate threshold (30%) and the correlation coefficient threshold (0.6) of the associated edge group are re-optimized using the logistic regression algorithm. Fluctuations exceeding 30% are judged as increased risk. Only correlation coefficients > 0.6 are included in the dynamic propagation directed graph to ensure the model's adaptability to different working conditions. Finally, a complete risk dynamic propagation model M is obtained. Among them, the frequency f of high-risk associated edges within K historical windows can be counted. When f≥0.2, that is, a strong change occurs once every 5 windows on average, K is increased by 20% to retain more historical data for trend analysis. When f≤0.05, K is decreased by 20%, with a minimum of 10, to improve the efficiency of model updates. Furthermore, the risk dynamic propagation model M comprises three parts: the current graph structure, the dynamic feature mapping of the historical graph sequence, and so on. The current graph structure is the latest dynamically propagating directed graph. That is, the current state of the model; Historical graph sequences are cached graph sequences. That is, the historical process of model evolution; The dynamic feature mapping is a mapping table structure, where the key is each edge in the dynamically propagating directed graph. Its value is the moving average of the dynamic characteristics of that edge and its relative rate of change; Ultimately, the risk dynamic propagation model M includes not only the graph structure at the current moment. It also implicitly contains the historical evolution information of the weight of each edge, making it a composite model that includes temporal context information.

[0024] Emergency Response Module: Based on the risk dynamic propagation model, critical vulnerability paths are extracted to obtain a set of equipment anomaly propagation chains. Based on the set of equipment anomaly propagation chains, abnormal equipment is captured to achieve safety emergency response control for chemical production equipment. Set a normal operating threshold range for each sensor. Iterate through all sensor nodes. If the values ​​of the last 3 data points in the current window continuously exceed the threshold range, mark that sensor node as an initial abnormal node. All initial abnormal nodes form a set. ;

[0025] for Each initial abnormal node in , In the risk dynamic propagation model M, we find the cumulative weight of the path from v to all other reachable nodes in the graph. The weight of the edge is regarded as the risk transmission intensity. The path with the largest cumulative weight indicates that the anomaly is most likely to spread along this path and cause the greatest impact. Furthermore, Dijkstra's algorithm is introduced to convert the reciprocal (or negative logarithm) of the edge weights into distances. Finding the shortest distance is equivalent to finding the maximum cumulative weight. For edge weights from... arrive path Its cumulative weight , for each Find one or more critical paths to different target nodes; Specifically, the critical path refers to the cumulative weight of the path calculated in a risk dynamic propagation model, starting from a device or sensor node that has been identified as the initial anomaly. The largest (or first few) propagation path, that is, "the abnormal propagation route that is most likely to trigger a chain of failures and cause the greatest damage in the current state." Collect all critical paths found starting from all initial anomalous nodes, remove completely duplicate paths, and then assign these paths according to their cumulative weights. Sort the results from largest to smallest, and finally output the set of anomaly propagation chains. , Indicates the number of abnormal propagation chains; Each of these transmission chains It is an ordered list of sensor nodes, representing a potential risk path starting from the current anomaly point that is most likely to cause the fault to escalate, with the path's cumulative weight attached. ; Then, based on the high-risk correlation propagation chain output by the risk dynamic propagation model M, all involved device nodes on the chain are located. Combining the node propagation potential energy classification, high potential energy nodes are prioritized as suspected abnormal source devices. Subsequently, by collecting sensor data from these nodes in real time and comparing it with preset safety thresholds, the system verifies whether the device is in an abnormal state and accurately captures abnormal devices. Next, based on the direction and intensity of the abnormal propagation chain, determine the spread range and trend of the abnormality, and simultaneously trigger the corresponding level of emergency response. For a single abnormal device, immediately cut off its associated feed and discharge valves and activate the device's built-in safety protection device. For multiple equipment malfunctions that have formed a transmission chain, the equipment in the chain is isolated in sequence according to the transmission order. The dynamic changes of the malfunction transmission chain are continuously tracked. Combined with the real-time feedback of equipment status data, emergency control measures are dynamically adjusted until all malfunctioning equipment is restored to a safe state, ensuring the stable operation of chemical production.

[0026] In the application, several formulas are calculated by removing dimensions and taking their numerical values. The formulas are established by collecting a large amount of data and simulating the most recent real situation. Some coefficients or weights in the formulas are set by those skilled in the art according to the actual situation, so they will not be elaborated here.

[0027] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0028] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A real-time data-based emergency response system for chemical production equipment, characterized in that, include: Status acquisition module: Collects initial status data of chemical equipment, and obtains equipment time-synchronized data through timestamp alignment and data stream synchronization. State synchronization module: Based on the device time-space data, the time window is dynamically divided according to the data fluctuation frequency to obtain spatiotemporal synchronization data. Based on the spatiotemporal synchronization data, the correlation coefficient matrix between sensors is obtained. Based on the correlation coefficient matrix, Granger causality test is performed to construct a dynamic propagation directed graph. Anomaly Analysis Module: Stores the current and historical dynamic propagation directed graphs into a first-in-first-out (FIFO) cache queue to obtain a time-series graph sequence. Based on the time-series graph sequence, it performs dynamic feature calculation to output the dynamic features of the edges. Based on the dynamic features of the edges, it performs edge-level time-series association mining to obtain a risk dynamic propagation model. The process of obtaining the risk dynamic propagation model is as follows: Calculate each edge The Pearson correlation coefficient of the weight of the edge with all other edge weights is used to screen out edges with a correlation coefficient ≥ 0.6 to form a group of associated edges, and the frequency f of high-risk associated edges in the historical K windows is counted. When f≥0.2, it is considered that a strong change occurs on average once every 5 windows; When ≥60% of the edges in an edge group undergo strong changes simultaneously, the edge group is marked as a high-risk associated edge group. And construct a propagation potential index for each node. ,in, The average acceleration of all incoming edges of a node. Let α be the average relative rate of change of all edges, and β be the weighting coefficients, where α + β = 1. Node potential energy is classified according to propagation potential energy index. It is a high potential energy node; When the proportion of high-potential nodes is ≥30%, the next window length is shortened to a short window of 5 data points. When the proportion of high-potential nodes is ≤10%, the next window length is extended to a long window of 30 data points to complete edge-level temporal correlation mining. Dynamically propagating the directed graph The generation process serves as the basic architecture of the model, and the process of edge-level temporal correlation mining serves as the model update process, thus obtaining the risk dynamic propagation model M. The model update process includes: Every 100 windows, based on normal and abnormal samples in historical data, and calibrated through manual annotation or equipment failure records, the relative rate of change threshold and the correlation coefficient threshold of associated edge groups are re-optimized using a logistic regression algorithm. Fluctuations exceeding 30% are considered as increased risk, and only when the correlation coefficient is >0.6 is the directed graph dynamically propagated. Emergency Response Module: Based on the risk dynamic propagation model, critical vulnerability paths are extracted to obtain a set of equipment anomaly propagation chains. Based on the set of equipment anomaly propagation chains, abnormal equipment is captured to achieve safety emergency response control for chemical production equipment.

2. The emergency response system for chemical production equipment based on real-time data according to claim 1, characterized in that, The process of obtaining device time and space data is as follows: Sensors are deployed to collect initial state data of chemical equipment. This initial state data is then timestamped, and a unified analysis time base is established. and reference sampling interval For any sensor i, at each reference time point Above, where k represents the index and represents a constant, the ith sensor at the reference time point is calculated using linear interpolation, utilizing two neighboring data points in its original time series. Approximate data values Then approximate data values Organized into a multidimensional time series matrix, where rows represent time points. The column represents sensor i, and the matrix elements are the corresponding approximate data values, thus obtaining the device time and space data X.

3. The emergency response system for chemical production equipment based on real-time data according to claim 2, characterized in that, The process of obtaining spatiotemporal synchronization data is as follows: By sliding a small segment containing the 10 most recent data points across the time series, the variance of the device time-space data X within that sliding segment is calculated. And take the average of the variances of all sensors as the overall fluctuation index U of the current segment; Preset a high volatility threshold and a low volatility threshold ,when At that time, the analysis time window length is shortened to 5 data points, and a short-window spatiotemporal data block is generated. The time window for extended analysis is 30 data points, generating a long-window spatiotemporal data block. Spatiotemporal synchronization data D is obtained based on short-window and long-window spatiotemporal data blocks.

4. The emergency response system for chemical production equipment based on real-time data according to claim 3, characterized in that, The process of obtaining the correlation coefficient matrix is ​​as follows: For any two data sequences corresponding to sensors i and j in the spatiotemporal synchronization data D, obtain the Pearson correlation coefficient of the data sequences. Then, a fixed window is preset, and the Pearson correlation coefficients of all two data sequences within the fixed window are used to form a correlation coefficient matrix R.

5. The emergency response system for chemical production equipment based on real-time data according to claim 4, characterized in that, The process of obtaining a dynamically propagating directed graph is as follows: Set a correlation coefficient threshold Only for those who meet the requirements Granger causality test is performed on sensor pair (i,j), and a restricted model is constructed. With the full model ; Compare the sum of squared residuals of the two models. When the prediction accuracy of the full model is higher than that of the restricted model, sensor j is taken as the Granger cause of sensor i. Represent all causal relationships that pass the Granger causality test as directed edges, and construct a directed causal relationship network. According to directed causal networks The correlation coefficient matrix R assigns causal relationships to the association strength, forming a complete weighted directed graph, i.e., a dynamic propagation directed graph. Where W is the set of edge weights, V is the set of nodes corresponding to all sensors, and each edge in the set of directed edges E represents... .

6. The emergency response system for chemical production equipment based on real-time data according to claim 5, characterized in that, Constrained models With the full model Represented as: Constrained model: ; Full model: ; Where t represents the timestamp and m is the lag order. These are the coefficients of the constrained model. It is the error term of the constrained model. For the full model coefficients, It is the error term of the entire model. Indicates a restricted model index. Represents the full model index. This represents the spatiotemporal synchronization data corresponding to sensor i. This represents the spatiotemporal synchronization data corresponding to sensor j.

7. The emergency response system for chemical production equipment based on real-time data according to claim 5, characterized in that, The time sequence is propagated through a directed graph that is dynamically fed into the current window. The image in the current window Store the data in a first-in-first-out (FIFO) cache queue in chronological order, save the graphs of the K most recent windows, and obtain the time sequence graph sequence.

8. The emergency response system for chemical production equipment based on real-time data according to claim 7, characterized in that, The process of outputting the dynamic features of the edges is as follows: The dynamic characteristics of the edge include moving average and relative rate of change; The moving average is for each edge The weights are the moving averages over the past K windows, and the relative rate of change is the ratio of the current relative rate of change to the baseline weights, where the baseline weights are... , This represents the weight of the current k-th window. It is the moving average at the current moment.

9. A real-time data-based emergency response system for chemical production equipment according to claim 8, characterized in that, The process of obtaining the set of device anomaly propagation chains is as follows: The nodes whose last three data points in the current sensor window consistently exceed the normal range are designated as initial abnormal nodes. And on the risk dynamic propagation model M, find the initial abnormal node. Cumulative weight of paths from the origin to all other reachable nodes : , Indicates the weight of the edge; Dijkstra's algorithm is introduced to transform edge weights into distances and find the shortest distance, especially for nodes from outliers. arrive path , For each abnormal node Find a critical path to different target nodes, collect all critical paths found starting from all initial abnormal nodes, remove completely duplicate paths, and accumulate weights according to the paths. Sort the results from largest to smallest to obtain the set of anomaly propagation chains. , Indicates the number of abnormal propagation chains.

Images