Chemical industry park safety management system and method based on internet of things

Abstract

The application provides a chemical industry park safety management system and method based on the Internet of Things, generates a dynamic monitoring data set by collecting multi-source monitoring data in the chemical industry park; generates a path confidence distribution by imposing a space-time causal constraint on a multi-hop connection path in a chemical industry park equipment topology graph; constructs entity feature embeddings of each device in the chemical industry park, determines a time-varying causal correlation matrix through all entity feature embeddings and the path confidence distribution; divides risk device groups of different risk levels and constructs a risk propagation tensor of a device risk zoning topology network in the chemical industry park through each risk device group; generates a dynamic risk vector through the time-varying causal correlation matrix and the risk propagation tensor, and performs early warning on safety risks in the chemical industry park based on the dynamic risk vector. The scheme of the application can realize space-time registration of multi-source monitoring data of Internet of Things sensing nodes and mining of time-varying causal relationships between devices to realize dynamic safety risk early warning.

Description

Technical Field

[0001] This application relates to the field of Internet of Things (IoT) technology, and more specifically, to an IoT-based safety management system and method for chemical industrial parks. Background Technology

[0002] Chemical industrial parks, as core sites for chemical production, are characterized by dense equipment, complex processes, and the presence of flammable, explosive, toxic, and hazardous media, making safety management paramount. With the development of IoT technology, real-time data collection from distributed sensing nodes has become a crucial means of safety monitoring in these parks. However, the core challenge lies in transforming multi-source, heterogeneous monitoring data into precise risk warning information to address the complex relationships between equipment and their dynamically changing operational states.

[0003] In existing technologies, safety management in chemical industrial parks often relies on single sensor data or static topology analysis. For example, it involves monitoring single-point parameters and triggering threshold alarms using independent sensors, or performing simple risk path analysis based on fixed equipment connection diagrams. These methods struggle to achieve time synchronization and spatial correlation calibration of data collected by distributed IoT sensing nodes, leading to spatiotemporal deviations in equipment status representation and affecting the reliability of the data foundation for risk assessment. Furthermore, traditional methods fail to mine the dynamic causal relationships of multi-hop connection paths in the equipment topology diagram, making it impossible to quantify the causal strength of multi-hop paths over time to identify time-varying dependencies between equipment and adapt to the characteristics of safety risks evolving over time in chemical industrial parks. This can easily result in delayed or misjudged warnings. Therefore, how to achieve spatiotemporal registration of multi-source monitoring data from IoT sensing nodes and mine time-varying causal relationships between equipment to achieve dynamic safety risk warnings has become a challenge for the industry. Summary of the Invention

[0004] This application provides a safety management system and method for chemical industrial parks based on the Internet of Things (IoT), which can realize spatiotemporal registration of multi-source monitoring data from IoT sensing nodes and mine time-varying causal relationships between devices to achieve dynamic safety risk early warning.

[0005] Firstly, this application provides an IoT-based dynamic risk early warning method for chemical industrial parks, used by a safety management system to dynamically identify and issue early warnings of safety risks in chemical industrial parks. The distributed IoT is supported by narrowband IoT for communication. The method includes:

[0006] Multi-source monitoring data is collected in the chemical industrial park through distributed Internet of Things sensing nodes, thereby generating a dynamic monitoring dataset that is synchronized in time and space.

[0007] Obtain the equipment topology map of the chemical industrial park, apply spatiotemporal causal constraints to the multi-hop connection paths in the equipment topology map, and generate the path confidence distribution between equipment in the chemical industrial park.

[0008] Based on the dynamic monitoring dataset, entity feature embeddings of each device in the chemical industrial park are constructed. The time-varying causal association matrix between devices in the chemical industrial park is determined by all entity feature embeddings and the path confidence distribution.

[0009] Based on the multidimensional state vectors of the equipment in the dynamic monitoring dataset, the equipment in the chemical industrial park is divided into risk groups to obtain risk equipment groups with different risk levels. The entity feature embeddings of the equipment in each risk equipment group are used to propagate cross-group risk features, thereby constructing a risk propagation tensor of the risk partitioning topology network of equipment in the chemical industrial park.

[0010] A dynamic risk vector for the chemical industrial park is generated using the time-varying causal correlation matrix and the risk propagation tensor, and an early warning of safety risks within the chemical industrial park is provided based on the dynamic risk vector.

[0011] In some embodiments, applying spatiotemporal causal constraints to the multi-hop connection paths in the device topology graph to generate a path confidence distribution among devices within the chemical industrial park specifically includes:

[0012] Multi-scale path mining is performed on the device topology graph to determine the causal frame chain of each multi-hop connection path in the device topology graph;

[0013] The time-varying causality value of each causal frame chain is calculated using the time-varying Granger causality test algorithm;

[0014] The path confidence distribution between equipment in the chemical industrial park is generated based on the time-varying causal values ​​of all causal frame chains.

[0015] In some embodiments, constructing entity feature embeddings for various devices within the chemical industrial park based on the dynamic monitoring dataset specifically includes:

[0016] The dynamic monitoring dataset is subjected to feature engineering processing to construct a multi-dimensional initial feature set for each device in the chemical industrial park.

[0017] For each piece of equipment in the chemical industrial park, feature fusion is performed on the multidimensional initial feature set of the equipment to generate a graph convolution feature representation of the equipment during operation;

[0018] The initial feature embedding of the equipment is determined by the graph convolution feature representation, thereby obtaining the initial feature embedding of each piece of equipment in the chemical industrial park.

[0019] The initial feature embeddings of all equipment are compared and enhanced to generate the entity feature embeddings of each piece of equipment in the chemical industrial park.

[0020] In some embodiments, determining the time-varying causal correlation matrix between equipment within a chemical industrial park through all entity feature embeddings and the path confidence distribution specifically includes:

[0021] Construct a feature association matrix for equipment within the chemical industrial park by embedding all entity features;

[0022] Extract the path confidence scores between each device from the path confidence distribution;

[0023] All path confidences are convolved and fused with the feature association matrix to generate a time-varying causal association matrix between equipment in the chemical industrial park.

[0024] In some embodiments, risk groups are formed for equipment in the chemical industrial park based on the multidimensional state vectors of the equipment in the dynamic monitoring dataset, resulting in risk equipment groups with different risk levels, specifically including:

[0025] Extract the multidimensional state vectors of each piece of equipment in the chemical industrial park from the dynamic monitoring dataset;

[0026] Spatiotemporal feature decomposition is performed on all multidimensional state vectors to construct the equipment state feature space of the chemical industrial park;

[0027] In the device state feature space, all devices are clustered according to the risk propagation index of the devices to obtain risk device groups with different risk levels.

[0028] In some embodiments, the risk propagation tensor for constructing the risk partitioning topology network of equipment within the chemical industrial park by propagating the entity feature embeddings of equipment in each risk equipment group across groups specifically includes:

[0029] For each risk equipment group, a risk association matrix for risk propagation between equipment is constructed based on the entity feature embedding of the equipment in the risk equipment group;

[0030] The risk correlation matrix is ​​modeled in a time series to generate a spatiotemporal risk propagation map of risk equipment groups, thereby obtaining the spatiotemporal risk propagation map of each risk equipment group;

[0031] The spatiotemporal risk propagation maps of all risky equipment groups are fused to generate a risk zoning topology network for the chemical industrial park.

[0032] Extract the risk propagation tensor from the risk partitioning topology network.

[0033] In some embodiments, generating a dynamic risk vector for the chemical industrial park using the time-varying causal correlation matrix and the risk propagation tensor specifically includes:

[0034] Multimodal embedding fusion is performed on the time-varying causal correlation matrix and the risk propagation tensor to generate a risk state transition matrix for equipment within the chemical industrial park;

[0035] The risk state transition matrix is ​​used to predict the safety risks of the chemical industrial park, resulting in a dynamic risk vector for the chemical industrial park.

[0036] Secondly, this application provides an Internet of Things (IoT)-based safety management system for chemical industrial parks. This system includes a dynamic risk early warning unit for dynamically identifying and issuing early warnings of safety risks in the chemical industrial park. The distributed IoT is supported by narrowband IoT communication. The dynamic risk early warning unit includes:

[0037] The data acquisition module is used to collect multi-source monitoring data in the chemical industrial park through distributed Internet of Things sensing nodes, and then generate a dynamic monitoring dataset that is synchronized in time and space.

[0038] The processing module is used to obtain the equipment topology map of the chemical industrial park, apply spatiotemporal causal constraints to the multi-hop connection paths in the equipment topology map, and generate the path confidence distribution between equipment in the chemical industrial park.

[0039] The processing module is used to construct entity feature embeddings of each device in the chemical industrial park based on the dynamic monitoring dataset, and determine the time-varying causal association matrix between devices in the chemical industrial park through all entity feature embeddings and the path confidence distribution.

[0040] The processing module is used to divide the equipment in the chemical industrial park into risk groups based on the multidimensional state vectors of the equipment in the dynamic monitoring dataset, obtain risk equipment groups with different risk levels, perform cross-group risk feature propagation on the entity feature embedding of the equipment in each risk equipment group, and then construct the risk propagation tensor of the risk partitioning topology network of equipment in the chemical industrial park.

[0041] The execution module is used to generate a dynamic risk vector for the chemical industrial park through the time-varying causal correlation matrix and the risk propagation tensor, and to provide early warning of safety risks within the chemical industrial park based on the dynamic risk vector.

[0042] Thirdly, this application provides a computer device, which includes a memory and a processor. The memory stores code, and the processor is configured to acquire the code and execute the above-described IoT-based dynamic early warning method for park risks.

[0043] Fourthly, this application provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described IoT-based dynamic early warning method for park risks.

[0044] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects:

[0045] The IoT-based safety management system and method for chemical industrial parks provided in this application first collects multi-source monitoring data within the chemical industrial park through distributed IoT sensing nodes, thereby generating a spatiotemporally synchronized dynamic monitoring dataset; it then acquires a topology map of the chemical industrial park, applies spatiotemporal causal constraints to the multi-hop connection paths in the topology map, and generates a path confidence distribution among the devices within the chemical industrial park; based on the dynamic monitoring dataset, it constructs entity feature embeddings for each device within the chemical industrial park, and determines a time-varying causal correlation matrix among the devices through all entity feature embeddings and the path confidence distribution; it divides the devices within the chemical industrial park into risk groups according to the multidimensional state vectors of the devices in the dynamic monitoring dataset, obtaining risk device groups with different risk levels, and performs cross-group risk feature propagation on the entity feature embeddings of the devices in each risk device group, thereby constructing a risk propagation tensor for the risk partitioning topology network of the devices within the chemical industrial park; it generates a dynamic risk vector for the chemical industrial park through the time-varying causal correlation matrix and the risk propagation tensor, and provides early warning of safety risks within the chemical industrial park based on the dynamic risk vector.

[0046] Therefore, this application provides early warning of safety risks within chemical industrial parks based on the aforementioned dynamic risk vector. First, determining the path confidence distribution yields a path probability distribution characterizing the reliability of all equipment on multi-hop connection paths during risk propagation. This determination quantifies the causal strength of multi-hop paths dynamically changing over time, solving the problem that static analysis in existing technologies cannot characterize time-varying dependencies between equipment. It provides a probabilistic basis for constructing a time-varying causal correlation matrix, making the subsequent risk assessment model more closely aligned with the uncertainty characteristics of complex systems. Second, determining the time-varying causal correlation matrix yields a matrix quantifying the time-varying nature of causal relationships between equipment within chemical industrial parks. This determination quantifies the time-varying nature of causal relationships between equipment by capturing implicit correlations that change over time. This not only solves the deficiency in existing technologies where static topology maps cannot reflect the dynamic changes in equipment correlations over time, but also eliminates the correlation analysis caused by spatiotemporal biases in multi-source data through feature fusion in the spatiotemporal dimension. Distortion provides a structured carrier for dynamic modeling of implicit dependencies between devices, enabling more accurate identification of abnormal propagation paths and key risk nodes, thus improving the real-time performance and early warning accuracy of safety monitoring in chemical industrial parks. Finally, determining the risk propagation tensor yields a trend matrix characterizing the time-varying intensity of risk propagation among different risk groups in the chemical industrial park. This determination integrates the spatial correlation between risky device groups and captures the dynamic evolution of risk propagation over time, overcoming the limitation of static analysis in existing technologies that cannot characterize the spatiotemporal coupling characteristics of risk propagation. Furthermore, by extracting propagation intensity from time-series data through a sliding time window, dynamic quantification of risk diffusion patterns is achieved, providing a structured data carrier for modeling the spatiotemporal evolution of risk propagation. This upgrades risk analysis from "static single-point assessment" to "dynamic full-domain tracking." In summary, based on the above scheme, spatiotemporal registration of multi-source monitoring data from IoT sensing nodes can be achieved, and time-varying causal relationships between devices can be mined to realize dynamic safety risk early warning. Attached Figure Description

[0047] Figure 1 This is an exemplary flowchart of an IoT-based dynamic early warning method for park risks, as shown in some embodiments of this application.

[0048] Figure 2 This is an exemplary flowchart illustrating the determination of entity feature embedding according to some embodiments of this application;

[0049] Figure 3 This is an operational flowchart illustrating the determination of risky device groups according to some embodiments of this application;

[0050] Figure 4 This is a schematic diagram of the structure of a dynamic risk warning unit according to some embodiments of this application;

[0051] Figure 5 This is an internal structural diagram of a computer device that implements a dynamic early warning method for park risks based on the Internet of Things, according to some embodiments of this application. Detailed Implementation

[0052] To better understand the technical solution of this application, the technical solution of this application will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0053] refer to Figure 1 The figure is an exemplary flowchart of an IoT-based dynamic risk early warning method for industrial parks, according to some embodiments of this application. This IoT-based dynamic risk early warning method for industrial parks mainly includes the following steps:

[0054] In step 101, multi-source monitoring data within the chemical industrial park is collected through distributed IoT sensing nodes, thereby generating a spatiotemporally synchronized dynamic monitoring dataset.

[0055] It should be noted that in this application, the dynamic monitoring dataset is a spatiotemporally synchronized dataset composed of multiple monitoring data within the chemical industrial park. This dynamic monitoring dataset can provide standardized data support with a unified spatiotemporal benchmark for the extraction and relationship modeling of equipment risk characteristics within the chemical industrial park, ensuring the consistency of different sensor data in time series and spatial coordinates, and avoiding misjudgment of risks due to spatiotemporal deviations. In specific implementation, the generation of a spatiotemporally synchronized dynamic monitoring dataset by collecting multi-source monitoring data within the chemical industrial park through distributed IoT sensing nodes can be achieved in the following way: First, for each piece of equipment within the chemical industrial park, sensors (such as catalytic combustion gas sensors, thermocouple temperature sensors, piezoelectric pressure sensors, and electromagnetic flowmeters) deployed in the distributed IoT sensing nodes within the equipment can collect the combustible gas concentration, operating temperature, operating pressure, and medium flow rate of the equipment at preset collection time intervals (such as half an hour). The collection of all collected combustible gas concentrations, operating temperatures, operating pressures, and medium flow rates is used as the multi-source monitoring data of the equipment. Through the above steps, the multi-source monitoring data of each piece of equipment within the chemical industrial park can be obtained. The data is then processed using an improved Kalman filter algorithm (such as an adaptive extended Kalman filter) to time-calibrate the multi-source monitoring data of all devices. This is achieved by setting a sliding time window (e.g., a window size of 5 minutes and a sliding step size of 1 minute) to obtain time-synchronized multi-source monitoring data. A graph neural network based on a spatiotemporal attention mechanism is then used to spatially calibrate all the time-synchronized multi-source monitoring data to obtain a spatiotemporally synchronized dynamic monitoring dataset. The graph neural network based on the spatiotemporal attention mechanism constructs a graph structure based on the spatial distribution of distributed IoT sensing nodes. It automatically learns the spatial correlation of multi-source monitoring data from different devices through an attention weight matrix, thereby outputting a spatiotemporally synchronized dynamic monitoring dataset. The multi-source monitoring data is a collection of data covering combustible gas concentration, operating temperature, operating pressure, and medium flow rate during equipment operation within the chemical industrial park. This multi-source monitoring data provides a multi-dimensional information input basis for risk analysis in the chemical industrial park, comprehensively covering dimensions such as equipment status and environmental parameters. It can eliminate monitoring blind spots from a single data perspective and improve the completeness of risk identification. The distributed IoT is supported by narrowband IoT communication.

[0056] In step 102, the equipment topology map of the chemical industrial park is obtained, and spatiotemporal causal constraints are applied to the multi-hop connection paths in the equipment topology map to generate the path confidence distribution between equipment in the chemical industrial park.

[0057] In practice, obtaining the equipment topology map of a chemical industrial park can be achieved in the following way: The physical layout information of the equipment within the chemical industrial park can be obtained by reading the computer-aided design drawings of the park. A graph database (such as Neo4j) is then used to construct a topology map based on this information. Equipment (such as reactors, pipelines, valves, etc.) is defined as nodes, and physical connections (such as material flow direction, control signal transmission paths, etc.) are defined as edges. This results in the equipment topology map of the chemical industrial park. The equipment topology map is a directed graph representing the spatial layout and connection relationships of the equipment in the chemical industrial park. By integrating physical layout and industrial control connection data, this topology map accurately presents the material and energy flow paths between equipment, providing a structured model basis for causal constraint path analysis and making risk propagation path identification more consistent with actual process logic.

[0058] In some embodiments, applying spatiotemporal causal constraints to multi-hop connection paths in the device topology graph to generate path confidence distributions among devices within a chemical industrial park can be achieved through the following steps:

[0059] Multi-scale path mining is performed on the device topology graph to determine the causal frame chain of each multi-hop connection path in the device topology graph;

[0060] The time-varying causality value of each causal frame chain is calculated using the time-varying Granger causality test algorithm;

[0061] The path confidence distribution between equipment in the chemical industrial park is generated based on the time-varying causal values ​​of all causal frame chains.

[0062] In specific implementation, multi-scale path mining is performed on the equipment topology graph to determine the causal frame chain of each multi-hop connection path in the equipment topology graph. This can be achieved in the following way: First, the equipment topology graph can be layered into three abstract levels: physical layer, control layer, and information layer. In the physical layer, the physical connection relationships between devices are extracted based on computer-aided design drawings to construct an undirected basic graph. In the control layer, the control signal flow direction between devices is obtained through industrial protocols (such as Modbus protocol) to construct a directed control graph. In the information layer, the semantic associations between devices are defined and a semantic network graph is constructed. Then, an improved search algorithm (such as depth-first search algorithm) combined with knowledge rules in the chemical industry (such as material flow constraints) is used to mine multi-hop connection paths in parallel in the graph structure of the three abstract levels to obtain all multi-hop connection paths in the equipment topology graph. Finally, for each multi-hop connection path... A time window mechanism (e.g., a window size of 5 minutes and a sliding step of 1 minute) can be introduced to divide multi-hop connection paths into continuous time segments. The path state within each time segment constitutes a causal frame, and adjacent causal frames are connected in chronological order to form a causal frame chain. Through the above steps, the causal frame chain of each multi-hop connection path in the equipment topology diagram can be obtained. The causal frame chain is a temporal sequence of causal relationships characterizing the temporal evolution of risk propagation between devices on the multi-hop connection path in the equipment topology diagram. This causal frame chain accurately captures the temporal dependence of risk propagation between devices through coupled modeling of time slices and path states, avoiding the distortion of causal relationships caused by static analysis, and providing structured data support for dynamic risk assessment in the temporal dimension. The multi-hop connection path refers to an indirect connection path in the equipment topology diagram of a chemical industrial park that requires two or more intermediate device nodes to reach the target device node.

[0063] In specific implementation, the time-varying causality value of each causal frame chain can be calculated using the time-varying Granger causality test algorithm in the following way: For each causal frame chain, firstly, a sliding window mechanism is used to process the causal frame chain. The window size can be set to 20 time points and the sliding step size to 5 time points. Based on the causal frame chain within each window, a vector autoregressive model is constructed (i.e., the current device state vector is equal to the sum of the products of the state vectors at multiple previous time points and the coefficient matrix plus the error term), and the optimal lag order of the vector autoregressive model is automatically determined by the Bayesian information criterion; then, a Granger causality test is performed based on all the vector autoregressive models. The null hypothesis is "the historical value of device A cannot cause the current value of device B". The probability of the hypothesis being true is assessed using the F-test statistic, and the obtained F-test statistic is converted into a causal strength value (i.e., the logarithm of the negative F-test statistic). This causal strength value is then used as the time-varying causal value of the causal frame chain. Through the above steps, the time-varying causal value of each causal frame chain can be obtained. The time-varying causal value is a dynamic indicator characterizing the strength of causal influence between devices on a multi-hop connection path. This time-varying causal value can reflect the degree of causal correlation of device state changes at different time points. By quantifying the time-varying characteristics of causal relationships in real time, it adapts to the dynamic fluctuations of equipment operating parameters in chemical industrial parks, solves the problem that traditional fixed-parameter causal analysis cannot capture instantaneous causal changes, and improves the real-time performance and accuracy of risk propagation path identification.

[0064] In specific implementation, the path confidence distribution between equipment in the chemical industrial park can be generated based on the time-varying causal values ​​of all causal frame chains in the following way: a Dirichlet process mixture model can be used to model the time-varying causal values ​​of all causal frame chains to obtain the path confidence distribution between equipment in the chemical industrial park. The Dirichlet process mixture model can convert all time-varying causal values ​​into high-dimensional feature vectors (i.e., vectors containing statistical features such as the mean, variance, and trend slope of all time-varying causal values). It is assumed that the high-dimensional feature vectors follow a Gaussian mixture distribution and the number of mixture components is automatically determined through the Dirichlet process. Then, the variational inference algorithm is used to estimate the model parameters, and the variational parameters are iteratively updated until convergence. After convergence, each causal frame chain is assigned to a different cluster, and the cluster center represents a typical causal pattern. Then, the posterior probability of each causal frame chain belonging to each cluster is calculated, and the obtained posterior probability is used as the parameter of the path confidence distribution. The path confidence distribution of probabilistic causality between equipment is generated through Monte Carlo sampling.

[0065] It should be noted that in this application, the path confidence distribution is a causal path probability distribution that characterizes the reliability of all devices on a multi-hop connection path in risk propagation. This path confidence distribution automatically determines the confidence level of the causal path in a data-driven manner, which can avoid the subjectivity of manual experience assignment and provide probabilistic risk propagation weights for the construction of time-varying causal correlation matrices, making the subsequent risk assessment model more in line with the uncertainty characteristics of complex systems.

[0066] In step 103, entity feature embeddings of each piece of equipment in the chemical industrial park are constructed based on the dynamic monitoring dataset. The time-varying causal association matrix between the equipment in the chemical industrial park is determined by all entity feature embeddings and the path confidence distribution.

[0067] In some embodiments, reference Figure 2 The figure is an exemplary flowchart illustrating the determination of entity feature embedding according to some embodiments of this application. The entity feature embedding of each piece of equipment within the chemical industrial park based on the dynamic monitoring dataset in this application can be achieved using the following steps:

[0068] In step 1031, feature engineering is performed on the dynamic monitoring dataset to construct a multidimensional initial feature set for each piece of equipment in the chemical industrial park.

[0069] In step 1032, for each piece of equipment in the chemical industrial park, feature fusion is performed on the multidimensional initial feature set of the equipment to generate a graph convolution feature representation of the equipment during operation;

[0070] In step 1033, the initial feature embedding of the equipment is determined by the graph convolution feature representation, thereby obtaining the initial feature embedding of each piece of equipment in the chemical industrial park;

[0071] In step 1034, the initial feature embeddings of all equipment are compared and enhanced to generate the entity feature embeddings of each piece of equipment in the chemical industrial park.

[0072] In specific implementation, feature engineering processing is performed on the dynamic monitoring dataset to construct a multi-dimensional initial feature set for each device in the chemical industrial park. This can be achieved in the following way: First, for each device in the chemical industrial park, multi-source monitoring data of the device is extracted from the dynamic monitoring dataset, and multi-dimensional feature extraction is performed on the multi-source monitoring data. That is, the time-domain features (including mean, standard deviation, and peak factor) of combustible gas concentration, operating temperature, operating pressure, and medium flow rate in the multi-source monitoring data are calculated to characterize the parameter fluctuation characteristics. Then, the combustible gas concentration, operating temperature, operating pressure, and medium flow rate in the multi-source monitoring data are extracted through Fast Fourier Transform. The frequency domain characteristics of the flow (such as the amplitude of the dominant frequency and the proportion of harmonic energy) are used to capture periodic abnormal signals. Then, the set of all obtained time domain characteristics and all frequency domain characteristics is used as the multidimensional initial feature set of the equipment. Through the above steps, the multidimensional initial feature set of each piece of equipment in the chemical industrial park can be obtained. The multidimensional initial feature set is the basic feature set that characterizes the multidimensional indicators of the equipment in the time domain and frequency domain of the chemical industrial park. This multidimensional initial feature set can provide multi-perspective basic data input for equipment status analysis, covering the temporal fluctuations and frequency characteristics of equipment operation, avoiding the one-sidedness of single-dimensional features, and providing comprehensive original feature materials for subsequent spatiotemporal feature fusion.

[0073] In specific implementation, feature fusion of the multidimensional initial feature set of the equipment to generate the graph convolutional feature representation of the equipment during operation can be achieved in the following way: An existing graph convolutional network model (such as a spatiotemporal graph convolutional network model) can be loaded, and this model can be used to fuse the multidimensional initial feature sets of all equipment to generate the graph convolutional feature representation of the equipment during operation. The graph convolutional network model defines a graph structure based on the equipment topology graph of the chemical industrial park, with nodes representing equipment and edges representing physical connections. A graph Laplacian matrix is ​​used to represent the spatial correlation of equipment. In the time dimension, a one-dimensional convolutional kernel (such as a kernel kernel) is applied to the multidimensional initial feature set of the equipment. The size can be set to 10 time steps to capture the evolution of parameters over time. Spatial neighborhood features and time series features are processed alternately through spatiotemporal convolutional layers to achieve feature fusion. The output feature matrix containing spatiotemporal coupling information is used as a graph convolutional feature representation. The graph convolutional feature representation is a feature matrix that characterizes the spatiotemporal correlation characteristics of the device's operating state. Each element in the graph convolutional feature representation represents the spatiotemporal correlation state of the device at the corresponding time step. By modeling the spatial connection relationship of the device through graph structure and capturing the dynamics of parameters through temporal convolution, the spatiotemporal propagation law of device state changes can be effectively extracted, solving the problem that traditional methods are difficult to handle complex spatial correlations.

[0074] In specific implementation, determining the initial feature embedding of the device through the graph convolutional feature representation can be achieved in the following way: the graph convolutional feature representation can be input into the fully connected layer of the graph convolutional network model for nonlinear mapping to generate the initial feature embedding vector of the device; wherein, the fully connected layer automatically focuses on risk-sensitive feature dimensions such as over-temperature and over-pressure by performing soft maximization normalization on the graph convolutional feature representation and weighting the aggregated value vector, and concatenates and linearly transforms the output results to generate an initial feature embedding containing the device's individual state and neighborhood interaction information (e.g., the dimension can be set to 128 dimensions); wherein, the initial feature embedding refers to a low-dimensional vector containing the device's individual state and neighborhood interaction information. This initial feature embedding focuses on risk-sensitive feature dimensions such as over-temperature and over-pressure, and by strengthening the expression of key parameters for risk assessment and suppressing irrelevant noise interference, the device features are made more in line with the needs of safety risk analysis.

[0075] It should be noted that in this application, entity feature embedding is a highly discriminative feature vector that represents the individual state of a device and its interaction with its neighborhood. This entity feature embedding improves feature discrimination by maximizing the similarity of similar devices and minimizing the differences between dissimilar devices, thereby solving the problem of insufficient discriminative power of traditional feature embedding. It enables the features of devices with different risk levels to be significantly separated in space, providing a more reliable feature basis for subsequent risk group division and propagation analysis.

[0076] In practice, the initial feature embeddings of all equipment are compared and enhanced to generate entity feature embeddings for each piece of equipment within the chemical industrial park. This can be achieved in the following way: First, positive sample pairs of similar equipment can be generated based on the process unit (such as reaction unit and distillation unit) and spatial location (such as within a radius of 50 meters) to which the equipment belongs. Negative sample pairs are generated by adding Gaussian noise with a mean of 0 and a standard of 0.1 to the entity feature embeddings. Then, the cosine similarity between the entity feature embedding of each equipment and the positive and negative sample embeddings is calculated using existing contrast loss functions (such as information noise contrast loss function). The optimization objective is set to maximize the similarity of positive sample pairs to be close to 1 and minimize the similarity of negative sample pairs to be close to -1. After 100 rounds of iterative training, the entity feature embeddings corresponding to each piece of equipment are generated.

[0077] In some embodiments, determining the time-varying causal correlation matrix between equipment within a chemical industrial park by embedding all entity features and the path confidence distribution can be achieved using the following steps:

[0078] Construct a feature association matrix for equipment within the chemical industrial park by embedding all entity features;

[0079] Extract the path confidence scores between each device from the path confidence distribution;

[0080] All path confidences are convolved and fused with the feature association matrix to generate a time-varying causal association matrix between equipment in the chemical industrial park.

[0081] In specific implementation, the feature association matrix of equipment within the chemical industrial park can be constructed by embedding all entity features in the following way: First, a multi-head attention mechanism can be used to perform multi-dimensional association calculations on the entity feature embeddings of all equipment. That is, the entity feature embedding of each equipment is mapped to a query vector, a key vector, and a value vector, respectively. In multiple (e.g., 8) parallel attention heads, the attention scores between different equipment are calculated by dot product operation, and the attention scores of all attention heads are weighted and summed to obtain the multi-dimensional association scores between different equipment. Then, a blank matrix is ​​constructed by using the sequence of all equipment as the rows and columns of the matrix, and all the obtained multi-dimensional association scores are filled into the corresponding positions in the matrix to obtain the feature association matrix of equipment within the chemical industrial park. The feature association matrix is ​​a matrix that characterizes the degree of correlation between the operating states of equipment in the chemical industrial park. This feature association matrix models the association relationship based on the differences between equipment features and states, providing a similarity measurement basis for the individual states of equipment for the time-varying causal association matrix.

[0082] In specific implementation, the path confidence between each device can be extracted from the path confidence distribution in the following way: For each pair of devices in the chemical industrial park, all possible multi-hop connection paths between the two devices can be traversed in the path confidence distribution, and the posterior probability of each multi-hop connection path can be obtained as the confidence of each corresponding multi-hop connection path. Then, the average of the confidence of all multi-hop connection paths can be used as the path confidence between the two devices. Through the above steps, the path confidence between each pair of devices in the chemical industrial park can be obtained, thereby obtaining the path confidence between each device. The path confidence is an indicator that characterizes the reliability of the risk propagation path between two devices in the time series dynamics. This path confidence can reflect the uncertainty and time-varying characteristics of the causal relationship between devices in the complex pipeline network of the chemical industrial park, and improve the credibility of risk propagation path analysis.

[0083] In specific implementation, the time-varying causal relationship matrix between equipment in the chemical industrial park can be generated by convolving and fusing all path confidences with the feature association matrix. This can be achieved in the following way: an existing convolutional neural network (such as a three-dimensional convolutional neural network) can be used to convolve and fuse the feature association matrix and all path confidences to generate the time-varying causal relationship matrix between equipment in the chemical industrial park. Specifically, the feature association matrix and the path confidence matrix can be convolved using a two-dimensional separable convolution kernel in the convolutional neural network. First, a 1×3 convolution kernel is used to perform row convolution on each matrix to extract the row direction dependency, and then a 3×1 convolution kernel is used to perform column convolution to extract the column direction dependency. By replacing the standard two-dimensional convolution with two one-dimensional convolutions, the computational complexity is reduced. Finally, the convolution result is normalized using a soft maximization normalization function to generate a time-varying causal relationship matrix with a dimension of equipment number × equipment number. Each element in the time-varying causal relationship matrix represents the time-varying nature of the causal relationship between the corresponding equipment pairs.

[0084] It should be noted that in this application, the time-varying causal correlation matrix is ​​a matrix that quantifies the time-varying nature of causal relationships between equipment in a chemical industrial park. By capturing the implicit correlations between equipment over time, this time-varying causal correlation matrix can more accurately identify abnormal propagation paths and key risk nodes, thereby improving the real-time performance and early warning accuracy of safety monitoring in chemical industrial parks.

[0085] In step 104, the equipment in the chemical industrial park is divided into risk groups based on the multidimensional state vectors of the equipment in the dynamic monitoring data set, resulting in risk equipment groups with different risk levels. The entity feature embeddings of the equipment in each risk equipment group are used to propagate cross-group risk features, thereby constructing a risk propagation tensor of the risk partitioning topology network of equipment in the chemical industrial park.

[0086] In some embodiments, reference Figure 3 The figure is a flowchart illustrating the operation of determining risk equipment groups according to some embodiments of this application. In this application, risk groups are divided into equipment in the chemical industrial park based on the multidimensional state vectors of the equipment in the dynamic monitoring data to obtain risk equipment groups with different risk levels. This can be achieved by the following steps:

[0087] Extract the multidimensional state vectors of each piece of equipment in the chemical industrial park from the dynamic monitoring dataset;

[0088] Spatiotemporal feature decomposition is performed on all multidimensional state vectors to construct the equipment state feature space of the chemical industrial park;

[0089] In the device state feature space, all devices are clustered according to the risk propagation index of the devices to obtain risk device groups with different risk levels.

[0090] In specific implementation, the extraction of multidimensional state vectors for each piece of equipment in the chemical industrial park from the dynamic monitoring dataset can be achieved in the following way: For each piece of equipment in the chemical industrial park, multi-source monitoring data of the equipment can be extracted from the dynamic monitoring dataset, and the vector composed of data within each time window (such as combustible gas concentration, operating temperature, operating pressure, and medium flow rate) can be taken as a time slice. Then, the multidimensional vector composed of time slices under all time windows can be taken as the multidimensional state vector of the equipment. Through the above steps, the multidimensional state vectors of each piece of equipment in the chemical industrial park can be obtained. The multidimensional state vector is a three-dimensional data structure that organizes the multi-source monitoring data of the equipment in the dynamic monitoring data according to the time dimension. This multidimensional state vector can provide standardized input for the spatiotemporal feature decomposition of high-dimensional complex data and solve the problem that traditional tabular data is difficult to capture multidimensional coupling relationships.

[0091] In specific implementation, the spatiotemporal feature decomposition of all multidimensional state vectors to construct the equipment state feature space of the chemical industrial park can be achieved in the following way: First, existing feature decomposition techniques (such as tensor train decomposition) can be used to reduce the dimensionality and extract features from the multidimensional state vectors of all equipment. The original multidimensional vectors are decomposed into the product of multiple low-dimensional vectors, and principal component analysis is performed on the decomposed low-dimensional vectors to extract principal components, thereby obtaining feature vectors describing the equipment state. Then, the equipment state feature space is constructed with the feature vectors of the equipment state as coordinate axes, based on the position of each equipment in the equipment topology diagram. Thus, the equipment state feature space of the chemical industrial park is obtained. The equipment state feature space is a low-dimensional feature vector space that maps the equipment operating state with key feature vectors as coordinate axes. This equipment state feature space removes redundant information and retains core features through dimensionality reduction techniques, compressing high-dimensional data to a computable dimension, providing efficient and representative feature input for subsequent cluster analysis, and improving the computational efficiency and accuracy of risk group classification.

[0092] In specific implementation, clustering all devices in the device state feature space according to the risk propagation index to obtain risk device groups of different risk levels can be achieved in the following way: First, the risk propagation index of each device can be extracted in the device state feature space, including the fluctuation range of device state (i.e., the standard deviation of operating temperature) and the trend of change (i.e., the slope of the operating pressure time series); then, using existing clustering algorithms (such as the improved density peak clustering algorithm), the weighted distance between devices is calculated with the risk propagation index as the weight, and the cluster center is automatically identified through the decision graph and all devices are clustered to obtain risk device groups of different risk levels; wherein, for each cluster center, the density peak clustering algorithm can calculate its risk propagation entropy based on the information entropy theory as a quantitative index of risk level, divide the devices into different risk levels, and perform secondary classification of boundary devices using semi-supervised learning, and adjust the classification boundary using a small number of samples labeled by domain experts to obtain risk device groups of different risk levels.

[0093] It should be noted that in this application, the risk equipment group is a cluster of equipment with similar risk propagation patterns. Through the risk equipment group, the hierarchical management of equipment in the chemical industrial park can be realized, so that safety resources can be focused on high-risk groups, and the problems of strong subjectivity and low efficiency of traditional manual classification can be solved.

[0094] In some embodiments, the risk propagation tensor for constructing the risk partitioning topology network of equipment within a chemical industrial park by propagating the entity feature embeddings of equipment in each risk equipment group across groups can be achieved through the following steps:

[0095] For each risk equipment group, a risk association matrix for risk propagation between equipment is constructed based on the entity feature embedding of the equipment in the risk equipment group;

[0096] The risk correlation matrix is ​​modeled in a time series to generate a spatiotemporal risk propagation map of risk equipment groups, thereby obtaining the spatiotemporal risk propagation map of each risk equipment group;

[0097] The spatiotemporal risk propagation maps of all risky equipment groups are fused to generate a risk zoning topology network for the chemical industrial park.

[0098] Extract the risk propagation tensor from the risk partitioning topology network.

[0099] In specific implementation, the risk association matrix for risk propagation between devices based on the entity feature embedding of devices in the risk device group can be constructed in the following way: First, the entity features of all devices in the risk device group are embedded into the input graph attention network. The multi-head attention mechanism of the graph attention network maps the entity feature embedding of each device into three vectors: query vector, key vector, and value vector. The risk propagation intensity between different devices is calculated by dot product. Then, a blank matrix is ​​constructed by using the sequence of all devices as the rows and columns of the matrix respectively. All the obtained risk propagation intensities are filled into the corresponding positions in the matrix to obtain the risk association matrix for risk propagation between devices in the risk device group. The risk association matrix is ​​a matrix that quantifies the risk propagation intensity between devices in the risk device group. This risk association matrix dynamically learns the risk propagation relationship based on the real-time operating characteristics of the devices. It can reflect the risk transmission association driven by the device state characteristics, accurately capture the non-linear coupling risk transmission path, and solve the problem that static weights cannot adapt to changes in device state.

[0100] In specific implementation, the risk correlation matrix is ​​modeled using a time-series evolution to generate a spatiotemporal risk propagation map of risky equipment groups. This can be achieved by introducing a time-series pulsation factor and using a gated loop unit to model the risk correlation matrix using a time-series evolution. The risk correlation matrix at each time step is used as the input sequence. The transmission and updating of risk information are controlled by the forget gate and update gate of the gated loop unit to generate a spatiotemporal risk propagation map containing time-series information. The spatiotemporal risk propagation map is a map that integrates the dynamic changes in the risk propagation intensity and time dimension between equipment. This spatiotemporal risk propagation map retains historical risk propagation information and captures the dynamic patterns of risk characteristics decay and enhancement over time through the forget gate and update gate mechanisms, providing a structured model support for the time-series dimension for real-time prediction and trend analysis of risks in chemical industrial parks.

[0101] In specific implementation, the risk partitioning topology network of the chemical industrial park can be generated by graph fusion of the spatiotemporal risk propagation maps of all risk equipment groups in the following way: Existing graph fusion algorithms (such as the graph community discovery algorithm based on modularity optimization) can be used to fuse all spatiotemporal risk propagation maps to generate the risk partitioning topology network of the chemical industrial park; wherein, the graph community discovery algorithm based on modularity optimization first constructs a global risk propagation graph, with nodes representing each risk equipment group and edge weights representing the risk propagation intensity between risk equipment groups (i.e., the similarity between the spatiotemporal risk propagation maps of risk equipment groups). Then, in each iteration of the graph community discovery algorithm, communities are divided by maximizing modularity, and spatial constraints are introduced to ensure that the partitioning conforms to the physical layout of the chemical industrial park. Through iterative optimization, the equipment is divided into different risk communities, generating the risk partitioning topology network.

[0102] It should be noted that in this application, the risk zoning topology network is a network composed of all equipment in the chemical industrial park divided according to the risk propagation pattern. The risk zoning topology network divides areas with similar risk propagation patterns by modular clustering, and combines the physical layout of the chemical industrial park to divide complex systems into hierarchical risk communities, providing a network structure basis for "zoning-based policy" for park safety management.

[0103] In specific implementation, the risk propagation tensor extracted from the risk partitioning topology network can be achieved in the following way: a sliding time window mechanism (e.g., the window size is set to 10 time steps and the sliding step size is set to 2 time steps) can be used to extract the time series data in the risk partitioning topology network, and the risk propagation intensity between any two risk groups within each time window can be extracted. Then, all the extracted risk propagation intensities are organized according to the three-dimensional dimensions of "risk group pair - time step - propagation intensity" to obtain the risk propagation tensor. Among them, the first dimension of the risk propagation tensor is the index of the risk group pair, the second dimension is the time step, and the third dimension is the risk propagation intensity within the corresponding window.

[0104] It should be noted that in this application, the risk propagation tensor is a trend matrix characterizing the change of risk propagation intensity among different risk groups in a chemical industrial park over time. By integrating the spatial correlation and temporal dynamic characteristics of risk propagation, this risk propagation tensor provides structured data support for accurately capturing the laws of risk diffusion and generating dynamic risk vectors, thereby improving the comprehensiveness and timeliness of risk early warning in the industrial park.

[0105] In step 105, a dynamic risk vector for the chemical industrial park is generated using the time-varying causal correlation matrix and the risk propagation tensor, and an early warning of safety risks within the chemical industrial park is issued based on the dynamic risk vector.

[0106] In some embodiments, generating a dynamic risk vector for a chemical industrial park using the time-varying causal correlation matrix and the risk propagation tensor can be achieved through the following steps:

[0107] Multimodal embedding fusion is performed on the time-varying causal correlation matrix and the risk propagation tensor to generate a risk state transition matrix for equipment within the chemical industrial park;

[0108] The risk state transition matrix is ​​used to predict the safety risks of the chemical industrial park, resulting in a dynamic risk vector for the chemical industrial park.

[0109] It should be noted that in this application, the dynamic risk vector is a multi-dimensional parameter vector that dynamically depicts the spatiotemporal evolution of risks in chemical industrial parks. It includes risk intensity and propagation speed. This dynamic risk vector not only reflects the current risk level of the chemical industrial park, but also predicts the propagation trend and impact range. It provides full-chain data support for park safety management, including "situational awareness, trend prediction, and emergency response," and solves the problem that traditional static risk assessment cannot capture dynamic changes.

[0110] In specific implementation, the risk state transition matrix of equipment within the chemical industrial park can be generated by multimodal embedding and fusion of the time-varying causal correlation matrix and the risk propagation tensor. This can be achieved as follows: First, the time-varying causal correlation matrix can be expanded into a vector sequence row-wise. The temporal dependencies between equipment are calculated using a multi-head self-attention mechanism, and the causal feature representations of all equipment are output. Then, a 3D convolutional kernel (such as a 3×3×3 convolutional kernel) is used to perform a 3D convolution operation on the risk propagation tensor to extract the spatiotemporal local features. This causal feature representation is then used as a query vector, and the spatiotemporal local features are used as key-value pairs. Attention weights are calculated and weighted summation is performed to achieve two modalities. The system involves the interactive fusion of features, followed by inputting the fused features into a fully connected 3D convolutional neural network. An activation function is used to generate a risk state transition matrix, where each element represents the probability of a device transitioning from its current state to another. Finally, a forget gate mechanism is introduced to exponentially decay and update the historical state transition matrix, ensuring that it reflects the latest risk propagation trends. This risk state transition matrix quantifies the probability of equipment transitioning between different risk states in a chemical industrial park. It provides a structured basis for dynamic risk prediction, accurately capturing risk transmission patterns and enhancing the scientific rigor and foresight of safety risk early warning.

[0111] In specific implementation, the dynamic risk vector of the chemical industrial park can be obtained by predicting the safety risks of the chemical industrial park through the risk state transition matrix. This can be achieved in the following way: a continuous-time Markov chain can be used to model the risk propagation process of the chemical industrial park based on the risk state transition matrix. The transition probability density function of the continuous-time Markov chain can be approximately solved by Chebyshev polynomial acceleration matrix exponential operation to generate a dynamic risk vector containing risk intensity value and propagation speed (i.e., the number of risk diffusion nodes per unit time).

[0112] In specific implementation, early warning of safety risks within the chemical industrial park based on the dynamic risk vector can be achieved in the following way: A multi-dimensional safety early warning system for the chemical industrial park can be constructed based on the dynamic risk vector to provide early warning of safety risks within the chemical industrial park. Specifically, the risk threshold determination module performs real-time evaluation of the indicators of each dimension in the dynamic risk vector. For the risk intensity dimension, the risk intensity value is compared with the preset three-level thresholds (e.g., safety threshold of 0.3, early warning threshold of 0.6, and high-risk threshold of 0.8) to trigger the corresponding level of early warning. For the propagation speed dimension, the rate of change of the number of risk diffusion nodes per unit time is calculated. When the rate of change exceeds 0.5 for three consecutive time steps, a trend early warning is triggered. Based on the dynamic risk vector, a long short-term memory network is used to predict the changing trend of risk indicators in the next 4 hours. When the predicted value exceeds 120% of the current threshold, an advanced early warning is triggered. Finally, a three-dimensional early warning information containing real-time risk level, evolution trend, and impact range is formed, and then visualized on the electronic map of the park through a geographic information system.

[0113] Furthermore, in another aspect of this application, in some embodiments, this application provides an Internet of Things-based safety management system for chemical industrial parks, which includes a dynamic risk early warning unit, as referenced. Figure 4 The figure is a structural schematic diagram of a risk dynamic early warning unit according to some embodiments of this application. The risk dynamic early warning unit 400 of the chemical industrial park safety management system based on the Internet of Things includes: a data acquisition module 401, a processing module 402, and an execution module 403, which are described below:

[0114] The acquisition module 401 in this application is mainly used to acquire multi-source monitoring data in the chemical industrial park through distributed Internet of Things sensing nodes, and then generate a dynamic monitoring dataset that is synchronized in time and space.

[0115] Processing module 402, in this application, is mainly used to obtain the equipment topology map of the chemical industrial park, apply spatiotemporal causal constraints to the multi-hop connection paths in the equipment topology map, and generate path confidence distributions between equipment in the chemical industrial park;

[0116] It should be noted that the processing module 402 in this application is also used to construct entity feature embeddings of each piece of equipment in the chemical industrial park based on the dynamic monitoring dataset, and to determine the time-varying causal association matrix between the equipment in the chemical industrial park through all entity feature embeddings and the path confidence distribution;

[0117] Additionally, it should be noted that the processing module 402 in this application is also used to divide the equipment in the chemical industrial park into risk groups based on the multidimensional state vector of the equipment in the dynamic monitoring data, obtain risk equipment groups with different risk levels, and propagate cross-group risk features by embedding the entity features of the equipment in each risk equipment group, thereby constructing a risk propagation tensor of the risk partitioning topology network of equipment in the chemical industrial park.

[0118] The execution module 403 in this application is mainly used to generate a dynamic risk vector of the chemical industrial park through the time-varying causal correlation matrix and the risk propagation tensor, and to provide early warning of safety risks in the chemical industrial park based on the dynamic risk vector.

[0119] The various modules in the aforementioned IoT-based chemical industrial park safety management system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, allowing the processor to invoke and execute the corresponding operations of each module.

[0120] In another embodiment, this application provides a computer device, which may be a server, and its internal structure diagram may be as follows. Figure 5 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores data for a dynamic early warning method for park risks based on the Internet of Things (IoT). The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements a dynamic early warning method for park risks based on the IoT.

[0121] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0122] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above embodiment of the IoT-based dynamic early warning method for park risks.

[0123] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps described in the above embodiment of the IoT-based dynamic early warning method for park risks.

[0124] In one embodiment, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps described in the embodiment of the IoT-based dynamic early warning method for campus risks.

[0125] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0126] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0127] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. An IoT-based dynamic risk early warning method for chemical industrial parks, used by the safety management system of chemical industrial parks to dynamically identify safety risks and issue early warnings, wherein... The distributed Internet of Things (IoT) is supported by narrowband IoT for communication, characterized by the following steps: Multi-source monitoring data is collected in the chemical industrial park through distributed Internet of Things sensing nodes, thereby generating a dynamic monitoring dataset that is synchronized in time and space. Obtain the equipment topology map of the chemical industrial park, apply spatiotemporal causal constraints to the multi-hop connection paths in the equipment topology map, and generate the path confidence distribution between equipment in the chemical industrial park. Based on the dynamic monitoring dataset, entity feature embeddings of each device in the chemical industrial park are constructed. The time-varying causal association matrix between devices in the chemical industrial park is determined by all entity feature embeddings and the path confidence distribution. Based on the multidimensional state vectors of the equipment in the dynamic monitoring dataset, the equipment in the chemical industrial park is divided into risk groups to obtain risk equipment groups with different risk levels. The entity feature embeddings of the equipment in each risk equipment group are used to propagate cross-group risk features, thereby constructing a risk propagation tensor of the risk partitioning topology network of equipment in the chemical industrial park. A dynamic risk vector for the chemical industrial park is generated using the time-varying causal correlation matrix and the risk propagation tensor, and an early warning of safety risks within the chemical industrial park is provided based on the dynamic risk vector. Specifically, applying spatiotemporal causal constraints to the multi-hop connection paths in the equipment topology graph to generate the path confidence distribution among equipment within the chemical industrial park includes: Multi-scale path mining is performed on the device topology graph to determine the causal frame chain of each multi-hop connection path in the device topology graph; The time-varying causality value of each causal frame chain is calculated using the time-varying Granger causality test algorithm; Generate path confidence distributions among equipment within the chemical industrial park based on time-varying causal values ​​of all causal frame chains; Specifically, determining the time-varying causal correlation matrix between equipment within the chemical industrial park through the embedding of all entity features and the path confidence distribution includes: Construct a feature association matrix for equipment within the chemical industrial park by embedding all entity features; Extract the path confidence scores between each device from the path confidence distribution; All path confidences are convolved and fused with the feature association matrix to generate a time-varying causal association matrix between equipment in the chemical industrial park.

2. The method of claim 1, wherein, The specific steps for constructing entity feature embeddings for various equipment within the chemical industrial park based on the aforementioned dynamic monitoring dataset include: The dynamic monitoring dataset is subjected to feature engineering processing to construct a multi-dimensional initial feature set for each device in the chemical industrial park. For each piece of equipment in the chemical industrial park, feature fusion is performed on the multidimensional initial feature set of the equipment to generate a graph convolution feature representation of the equipment during operation; The initial feature embedding of the equipment is determined by the graph convolution feature representation, thereby obtaining the initial feature embedding of each piece of equipment in the chemical industrial park. The initial feature embeddings of all equipment are compared and enhanced to generate the entity feature embeddings of each piece of equipment in the chemical industrial park.

3. The method of claim 1, wherein, Based on the multidimensional state vectors of the equipment in the dynamic monitoring dataset, the equipment in the chemical industrial park is divided into risk groups, resulting in risk equipment groups with different risk levels, specifically including: Extract the multidimensional state vectors of each piece of equipment in the chemical industrial park from the dynamic monitoring dataset; Spatiotemporal feature decomposition is performed on all multidimensional state vectors to construct the equipment state feature space of the chemical industrial park; In the device state feature space, all devices are clustered according to the risk propagation index of the devices to obtain risk device groups with different risk levels.

4. The method of claim 1, wherein, The risk propagation tensor for constructing a risk zoning topology network of equipment risk partitioning within a chemical industrial park involves embedding the entity features of equipment in each risk equipment group and propagating risk features across groups. Specifically, this includes: For each risk equipment group, a risk association matrix for risk propagation between equipment is constructed based on the entity feature embedding of the equipment in the risk equipment group; The risk correlation matrix is ​​modeled in a time series to generate a spatiotemporal risk propagation map of risk equipment groups, thereby obtaining the spatiotemporal risk propagation map of each risk equipment group; The spatiotemporal risk propagation maps of all risky equipment groups are fused to generate a risk zoning topology network for the chemical industrial park. Extract the risk propagation tensor from the risk partitioning topology network.

5. The method of claim 1, wherein, The generation of the dynamic risk vector for the chemical industrial park through the time-varying causal correlation matrix and the risk propagation tensor specifically includes: Multimodal embedding fusion is performed on the time-varying causal correlation matrix and the risk propagation tensor to generate a risk state transition matrix for equipment within the chemical industrial park; The risk state transition matrix is ​​used to predict the safety risks of the chemical industrial park, resulting in a dynamic risk vector for the chemical industrial park.

6. A safety management system for a chemical industrial park based on the Internet of Things (IoT), comprising a dynamic risk early warning unit for dynamically identifying safety risks in the chemical industrial park and issuing early warnings, wherein the system employs the method described in any one of claims 1 to 5 for dynamic risk early warning of the park, wherein... The distributed Internet of Things (IoT) is supported by narrowband IoT for communication, characterized in that the dynamic risk early warning unit includes: The data acquisition module is used to collect multi-source monitoring data in the chemical industrial park through distributed Internet of Things sensing nodes, and then generate a dynamic monitoring dataset that is synchronized in time and space. The processing module is used to obtain the equipment topology map of the chemical industrial park, apply spatiotemporal causal constraints to the multi-hop connection paths in the equipment topology map, and generate the path confidence distribution between equipment in the chemical industrial park. The processing module is used to construct entity feature embeddings of each device in the chemical industrial park based on the dynamic monitoring dataset, and determine the time-varying causal association matrix between devices in the chemical industrial park through all entity feature embeddings and the path confidence distribution. The processing module is used to divide the equipment in the chemical industrial park into risk groups based on the multidimensional state vectors of the equipment in the dynamic monitoring dataset, obtain risk equipment groups with different risk levels, perform cross-group risk feature propagation on the entity feature embedding of the equipment in each risk equipment group, and then construct the risk propagation tensor of the risk partitioning topology network of equipment in the chemical industrial park. The execution module is used to generate a dynamic risk vector for the chemical industrial park through the time-varying causal correlation matrix and the risk propagation tensor, and to provide early warning of safety risks within the chemical industrial park based on the dynamic risk vector. 7.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-6 when the computer program is executed by the processor. When the processor executes the computer program, it implements the steps of the IoT-based dynamic early warning method for park risks as described in any one of claims 1 to 5.

8. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 7. When the computer program is executed by the processor, it implements the steps of the IoT-based dynamic early warning method for park risks as described in any one of claims 1 to 5.

Images