Intelligent assessment method and system for environmental pollution risk of chemical industry park

Abstract

The invention discloses a chemical industry park environmental pollution risk intelligent assessment method and system, and relates to the technical field of environmental risk management, and the method comprises the steps: extracting an association rule of a risk event based on a real-time data stream, coding the association rule into a symbolic logic expression, constructing a mixed neural symbol architecture, and carrying out the construction of a hybrid neural symbol architecture; optimizing the symbolic logic expression and the mixed neural symbolic architecture parameters through joint training, and outputting a preliminary risk level assessment result; according to the real-time data stream, extracting a'risk event-cause-consequence 'triple, generating a risk knowledge graph, and performing graph reasoning in combination with a preliminary risk level assessment result to obtain a risk event causal chain and a diffusion path; according to the method, efficient integration and dynamic optimization of the multi-source heterogeneous data are achieved by constructing the mixed neural symbol architecture, and the problem that space-time relevance is weakened due to data splitting type processing in an existing method is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of environmental risk management, in particular to an intelligent environmental pollution risk assessment method and system for a chemical industrial park. BACKGROUND

[0002] In the field of environmental pollution risk assessment of a chemical industrial park, existing methods mainly rely on static data collection and single model analysis. In early research, environmental monitoring data is obtained through a fixed sensor network, equipment operating state data is recorded through a supervisory control and data acquisition (SCADA) unit, and text log data is mostly archived in the form of manual reporting. Risk identification usually adopts process safety assessment technologies such as HAZOP (hazard and operability analysis) or LOPA (layer of protection analysis), and combines geographic information units to spatially mark high-risk devices to form a risk visualization hierarchical management.

[0003] Existing technologies have limitations in multi-source heterogeneous data fusion and dynamic causal chain modeling. Environmental risk of a chemical industrial park involves multi-source heterogeneous data such as environmental monitoring, equipment state, and text logs. However, existing methods mostly process different data sources with independent modules, which weakens the spatio-temporal correlation between data and makes it difficult to form unified risk event correlation rules. Existing risk assessment models have limited modeling capabilities for causal relationships, usually relying only on expert experience to build static risk atlases, or identifying correlations between variables through statistical methods, but cannot dynamically track the cause-effect chain of risk events. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides an intelligent environmental pollution risk assessment method for a chemical industrial park to solve the problems of low multi-source heterogeneous data fusion efficiency and insufficient dynamic causal chain modeling capability.

[0006] To solve the above technical problems, the present application provides the following technical solutions: In a first aspect, the present application provides an intelligent environmental pollution risk assessment method for a chemical industrial park, which comprises, collecting multi-source heterogeneous data of the chemical industrial park, performing spatio-temporal alignment on the multi-source heterogeneous data, and generating real-time data streams, wherein the multi-source heterogeneous data includes environmental monitoring data, equipment operating state data, and text log data; based on the real-time data streams, extracting correlation rules of risk events, encoding the correlation rules into symbolic logic expressions, constructing a hybrid neural-symbolic architecture, optimizing the symbolic logic expressions and the hybrid neural-symbolic architecture parameters through joint training, and outputting preliminary risk level assessment results; Based on real-time data streams, extract the "risk event-cause-effect" triplet to generate a risk knowledge graph. Combine the graph with the preliminary risk level assessment results to perform graph reasoning and obtain the causal chain and diffusion path of the risk event. By comparing the preliminary risk level assessment results with the actual handling records, misjudgment cases are identified, triggering dynamic optimization of the hybrid neural symbolic architecture; Based on risk assessment models and risk knowledge graphs, the risk knowledge graph, causal chains of risk events, and diffusion paths are overlaid on a digital twin map, and tiered early warning signals are pushed to managers through an early warning feedback interface.

[0007] As a preferred embodiment of the intelligent assessment method for environmental pollution risks in chemical industrial parks described in this invention, the step of extracting association rules for risk events based on real-time data streams and encoding the association rules into symbolic logic expressions refers to extracting association rules for risk events from real-time data streams and encoding the association rules into symbolic logic expressions containing trainable threshold parameters.

[0008] As a preferred embodiment of the intelligent risk assessment method for environmental pollution in chemical industrial parks described in this invention, the step of jointly training and optimizing symbolic logic expressions and hybrid neural symbolic architecture parameters to output preliminary risk level assessment results includes the following specific steps. Symbolic logic expressions are executed through a symbolic logic layer of a hybrid neural symbolic architecture to generate confidence vectors for association rules; A neural network layer with a hybrid neural symbolic architecture is used to process real-time data streams and generate spatiotemporal feature vectors. The fusion layer of the hybrid neural symbolic architecture is used to fuse the rule confidence vector and the spatiotemporal feature vector to obtain the fused feature vector; Map the fused feature vectors to a risk probability distribution; The difference between the risk probability distribution and the actual risk label in historical accident reports is calculated using the cross-entropy loss function, and the cross-entropy loss function value is obtained. Backpropagation is used to generate the gradient matrix of the threshold parameter and the hybrid neural symbolic architecture parameter in the symbolic logic expression, and an adaptive learning rate optimizer is used to update the threshold parameter and the hybrid neural symbolic architecture parameter in the symbolic logic expression. The real-time data stream is input into the updated hybrid neural symbolic architecture to generate a risk probability distribution. The risk level is then classified according to the risk probability distribution to obtain a preliminary risk level assessment result.

[0009] As a preferred embodiment of the intelligent environmental pollution risk assessment method for chemical industrial parks described in this invention, the step of extracting "risk event-cause-effect" triples based on real-time data streams to generate a risk knowledge graph includes the following specific steps. Use a named entity recognition model to extract the "risk event-cause-effect" triple; Construct an initial causal graph based on the "risk event-cause-effect" triple; Transform the directed acyclic graph into a graph structure, define risk event nodes, cause nodes and consequence nodes, directed edges and edge weights, and form a risk knowledge graph.

[0010] As a preferred embodiment of the intelligent environmental pollution risk assessment method for chemical industrial parks described in this invention, the step of combining preliminary risk level assessment results with graph reasoning to obtain the causal chain and diffusion path of risk events includes the following specific steps. Starting from a high-risk event node, trace the causal chain backwards along the directed edges; Predict the consequence chain along the directed edge from the nodes of high-risk events to form a causal chain of risk events; Starting from a high-risk event node, traverse all paths along directed edges to generate a diffusion path.

[0011] As a preferred embodiment of the intelligent assessment method for environmental pollution risks in chemical industrial parks described in this invention, the specific steps for triggering the dynamic optimization of the hybrid neural symbolic architecture are as follows: The error rate of the preliminary risk level assessment results was statistically analyzed; The neural network layer structure of the hybrid neural symbolic architecture is adjusted based on the misclassification rate, and batch normalization processing units are inserted into the neural network layer of the hybrid neural symbolic architecture. Extract the real-time data stream of misjudged cases and adjust the threshold parameters of the symbolic logic layer and the fusion layer parameters of the hybrid neural symbolic architecture; The threshold parameters of the symbolic logic layer and the parameters of the fusion layer of the adjusted hybrid neural symbolic architecture are loaded into the hybrid neural symbolic architecture to obtain the risk assessment model.

[0012] As a preferred embodiment of the intelligent assessment method for environmental pollution risks in chemical industrial parks described in this invention, the specific steps for overlaying the risk knowledge graph, risk event causal chain, and diffusion path onto the digital twin map are as follows. Extract equipment coordinates from the digital twin map of the chemical industrial park and associate risk event nodes in the risk knowledge graph with the equipment coordinates; Transform the causal chain of risk events into spatial path segments; Based on environmental monitoring data, the pollutant diffusion trajectory was simulated using a Gaussian plume model to calculate the pollutant concentration distribution; The diffusion radius is determined according to the "Occupational Exposure Limits for Hazardous Factors in the Workplace".

[0013] Secondly, this invention provides an intelligent assessment system for environmental pollution risks in chemical industrial parks, including, The data acquisition module collects multi-source heterogeneous data from the chemical industrial park, performs spatiotemporal alignment on the multi-source heterogeneous data, and generates a real-time data stream. The multi-source heterogeneous data includes environmental monitoring data, equipment operating status data, and text log data. The assessment module extracts association rules of risk events based on real-time data streams, encodes the association rules into symbolic logic expressions, constructs a hybrid neural symbolic architecture, optimizes the parameters of the symbolic logic expressions and the hybrid neural symbolic architecture through joint training, and outputs preliminary risk level assessment results. The reasoning module extracts the "risk event-cause-effect" triples based on real-time data streams, generates a risk knowledge graph, and performs graph reasoning based on the preliminary risk level assessment results to obtain the causal chain and diffusion path of the risk event. The update module identifies misjudgment cases by comparing the preliminary risk level assessment results with the actual handling records, triggering dynamic optimization of the hybrid neural symbolic architecture; The early warning module, based on the risk assessment model and risk knowledge graph, overlays the risk knowledge graph, the causal chain of risk events, and the diffusion path onto the digital twin map, and pushes tiered early warning signals to managers through the early warning feedback interface.

[0014] Thirdly, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the intelligent assessment method for environmental pollution risks in chemical industrial parks as described in the first aspect of the present invention.

[0015] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the intelligent assessment method for environmental pollution risks in chemical industrial parks as described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: By constructing a hybrid neural symbolic architecture, efficient integration and dynamic optimization of multi-source heterogeneous data are achieved, solving the problem of weakened spatiotemporal correlation caused by fragmented data processing in existing methods. This enables the hybrid neural symbolic architecture to adapt to risk characteristics in different scenarios, avoiding the fragmented data processing problem of independent modules in existing methods and reducing information loss caused by data silos in existing methods. By generating a risk knowledge graph, dynamic modeling and accurate reasoning of causal chains of risk events are achieved, solving the problem of insufficient static causal chain modeling in existing methods. This quantifies the propagation path and weight of risk events, enabling the knowledge graph to continuously adapt to real-world scenarios and predict nonlinear risk diffusion. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Fig. 1 This is a flowchart of an intelligent assessment method for environmental pollution risks in chemical industrial parks.

[0019] Fig. 2 This is a schematic diagram of an intelligent environmental pollution risk assessment system for chemical industrial parks.

[0020] Fig. 3 This is a schematic diagram of a hybrid neural symbolic architecture.

[0021] Fig. 4 A schematic diagram of risk knowledge graph overlaid on a digital twin map. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figs. 1-4 As an embodiment of the present invention, an intelligent assessment method for environmental pollution risks in chemical industrial parks is provided, comprising the following steps: S1: Collect multi-source heterogeneous data from the chemical industrial park, perform spatiotemporal alignment on the multi-source heterogeneous data, and generate a real-time data stream. The multi-source heterogeneous data includes environmental monitoring data, equipment operating status data, and text log data.

[0026] The specific steps are as follows: Gas sensors, water quality monitors, and weather stations are installed at storage tank areas, pipeline interfaces, and discharge outlets within the chemical industrial park to collect environmental monitoring data. The gas sensors monitor VOCs concentration, PM2.5, and sulfur dioxide, the water quality monitors collect pH, COD, and heavy metal concentrations, and the weather stations collect wind speed, temperature, and humidity.

[0027] The system connects to the programmable logic controller and supervisory control unit (SCADA) within the chemical industrial park via the OPC-UA protocol to read equipment operating status data, including tank pressure, valve opening, pump speed, and pipeline corrosion level.

[0028] This process involves collecting chemical accident reports, equipment manuals, and safety regulations, manually annotating risk events, causal entities, and consequences to create a chemical industry-specific annotated corpus. A BERT model is loaded as the encoder, with a BiLSTM layer added to capture contextual dependencies, and a CRF layer to constrain entity label transfer logic. The word segmentation sequences from the chemical industry-specific annotated corpus are input into the BERT model for word vector encoding, resulting in text segmentation vectors. These vectors are then input into a BiLSTM layer, where a bidirectional long short-term memory network captures the contextual dependencies of the segmented sequences, yielding contextual feature vectors. These contextual feature vectors are then input into a CRF layer, where a linear transformation layer maps the feature dimensions of the contextual feature vectors to the label count dimension, resulting in an emission score tensor. This emission score tensor is used as an element of the state score matrix to form the state score matrix, which is then processed using nn.Parameter. The transition score matrix is ​​initialized, and the optimal entity label sequence is calculated based on the transition score matrix and the state score matrix. The predicted labels of risk events, causes, and consequences are output. The negative log-likelihood loss function of CRF is used to compare the difference between the predicted labels and the manually labeled ones, and the prediction error of the predicted labels is calculated. The partial derivatives of the loss function with respect to the BERT model parameters, BiLSTM layer parameters, and CRF layer parameters are calculated through Adam optimizer. The BERT model parameters, BiLSTM layer parameters, and CRF layer parameters are updated by backpropagation layer by layer to obtain the named entity recognition model.

[0029] Text log data, including equipment operation and maintenance records, alarm logs, and operator shift handover reports, is periodically retrieved from the enterprise production log database via a RESTful API interface. Regular expressions and named entity recognition models are used to extract event types from the text log data: regular expressions are used to match event types in the text log data; if a null value is returned after iterating through all rules, a named entity recognition model is used to perform deep semantic analysis on the text log data to obtain predicted labels, and the predicted labels of risk events are used as event types. The event types of discrete text log data are mapped to numerical codes to obtain text log event type encoded values.

[0030] A dynamic time warping algorithm is used to align environmental monitoring data, equipment operating status data, and text log data.

[0031] The locations and equipment coordinates of gas sensors, water quality monitors, and weather stations are mapped to a GIS grid. Geohashing encoding is used to associate the grid cells of the GIS grid with environmental monitoring data, equipment operation status data, and text log data.

[0032] Environmental monitoring data, equipment operating status data, and text log data are integrated into a real-time data stream according to time windows.

[0033] It should also be noted that by acquiring and aligning multi-source heterogeneous data in real time, a unified time reference and spatial mapping were achieved for environmental monitoring data, equipment operating status data, and text log data, resolving the issues of inconsistent timestamps and spatial location separation in existing methods. The dynamic time warping algorithm effectively eliminated differences in sensor data acquisition frequencies, and geohashing encoding associated discrete spatial locations with a unified GIS grid, improving the spatiotemporal correlation of the data stream. This provides a highly consistent data foundation for subsequent risk event association rule extraction, avoiding assessment bias caused by data fragmentation and reducing the impact of data silos on risk assessment.

[0034] S2: Based on real-time data streams, extract association rules for risk events, encode association rules into symbolic logic expressions, construct a hybrid neural symbolic architecture, optimize the parameters of the symbolic logic expressions and the hybrid neural symbolic architecture through joint training, and output preliminary risk level assessment results.

[0035] The specific steps are as follows: The PyPDF2 library in Python was used to parse historical accident reports and equipment operation and maintenance manuals of the chemical industrial park. A rule extraction method based on text templates was adopted to match the "condition-event" pattern and extract the association rules of risk events.

[0036] The association rules of risk events are encoded into symbolic logic expressions using fuzzy logic encoding: the condition combinations corresponding to the logical operator "AND" in the association rules of risk events are converted into multiplication operations, and the joint probability of multiple conditions being satisfied is represented by continuous multiplication; the condition combinations corresponding to the logical operator "OR" in the association rules of risk events are converted into maximum value operations, and the maximum probability among multiple conditions is taken as the degree of satisfaction; the threshold comparisons in the association rules of risk events are converted into Sigmoid functions, so that hard threshold boundaries are transformed into continuously differentiable soft probability outputs; the initial values ​​of the thresholds in the association rules of risk events are set as trainable parameters, and the thresholds are dynamically adjusted through backpropagation to adapt to the characteristics of real-time data streams.

[0037] The hybrid neurosymbolic architecture includes a symbolic logic layer, a neural network layer, and a fusion layer; The symbolic logic layer is used to load symbolic logic expressions as computing nodes, calculate soft probability outputs based on real-time data streams, and perform continuous multiplication or maximum value operations on the soft probability outputs according to the symbolic logic expressions to obtain the association rule confidence vector. The neural network layer uses a 3-layer fully connected network to process real-time data streams. Z-score standardization is used to standardize environmental monitoring data and equipment operating status data. Min-max normalization is used to normalize the text log event type encoding values. The normalized text log event type encoding values ​​are then linearly transformed, and the ReLU function is used to zero out negative values ​​and retain positive values ​​to generate a spatiotemporal feature vector. The fusion layer is used to concatenate all association rule confidences and feature vectors along the feature dimension to generate a fused feature vector. The fused feature vector is then mapped to the number of risk level categories, and the risk probability distribution is output through the Softmax function.

[0038] The difference between the risk probability distribution and the actual risk labels in historical accident reports is calculated using the cross-entropy loss function, yielding the cross-entropy loss function value, expressed as: ; in, The cross-entropy loss function value represents the difference between the risk probability distribution and the true risk labels in historical accident reports. For the number of risk categories, Indexed by the number of risk categories, For the first True risk label for risk class (unique hot coding format). For the first Risk probability of class risk; The threshold parameter in the symbolic logic expression is optimized using gradient descent: the partial derivative of the cross-entropy loss function with respect to the threshold parameter is calculated, and the threshold parameter is updated. The expression is as follows: ; in, For the updated threshold parameters, For threshold parameters, This is the learning rate.

[0039] Real-time data streams are input into the trained hybrid neural symbolic architecture to generate a risk probability distribution. Risk levels are then classified according to this distribution: based on the risk probability distribution, a percentile method is used to set medium-risk and high-risk thresholds. For example, the 30th percentile risk probability is used as the medium-risk threshold, and the 70th percentile risk probability is used as the high-risk threshold. Risk probabilities below the medium-risk threshold are classified as low-risk, risk probabilities not lower than the medium-risk threshold but lower than the high-risk threshold are classified as medium-risk, and risk probabilities not lower than the high-risk threshold are classified as high-risk. The risk level, association rules, and confidence levels of association rules are combined into a preliminary risk level assessment result.

[0040] It should also be noted that by encoding association rules into symbolic logic expressions and jointly training them with neural network layers, a deep integration of rule logic and data-driven approaches is achieved. The symbolic logic layer employs fuzzy logic encoding (e.g., multiplication represents "AND", maximum value represents "OR") to transform hard thresholds into differentiable soft probability outputs. The neural network layer learns the nonlinear characteristics of real-time data streams through a fully connected network. The fusion layer concatenates two types of features to generate a risk probability distribution. The design of the cross-entropy loss function and backpropagation to optimize threshold parameters allows the symbolic logic expressions to dynamically adapt to real-time data features. Percentile-based risk level classification (e.g., 30th and 70th percentile thresholds) further enhances the adaptability of risk level classification. This addresses the shortcomings of static rules in adapting to dynamic scenarios, improving the accuracy and interpretability of preliminary assessments.

[0041] S3: Based on the real-time data stream, extract the "risk event-cause-effect" triplet, generate a risk knowledge graph, and combine it with the preliminary risk level assessment results to perform graph reasoning to obtain the causal chain and diffusion path of the risk event.

[0042] The specific steps are as follows: The named entity recognition model is used to identify risk events, causes, and consequences in text log data. Dependency parsing is used to extract the relationship of "risk event-cause-effect" triples and match the "subject-verb-object" structure. Environmental monitoring data and equipment operation status data are associated with the corresponding "risk event-cause-effect" triples.

[0043] Risk events, causes, and consequences are treated as risk event nodes, cause nodes, and consequence nodes, respectively. There are undirected edges between all nodes, forming an initial causal graph. For each pair of adjacent nodes, the undirected edges are converted into directed edges according to the order of their occurrence. If the entities corresponding to each pair of adjacent nodes occur at the same time, the undirected edges are converted into bidirectional edges. The last directed edge established in all closed loop connections is deleted, generating a directed acyclic graph. Calculate the co-occurrence frequency and time difference of occurrence of all two entities in the "risk event-cause-effect" triple, and then calculate the directed edge weights, expressed as follows: ; in, For directed edge weights, The co-occurrence frequency of the two entities. The time difference between the appearance of the two entities; The directed acyclic graph is transformed into a graph structure. The nodes of the graph include risk event nodes, causal nodes, and consequence nodes. The directed edges of the graph are causal relationship types (such as "cause", "promote", "inhibit"), and the edge weights are the directed edge weights of the corresponding nodes. The preliminary risk level assessment results are embedded into the risk event nodes to form a risk knowledge graph.

[0044] Starting from a high-risk event node, recursively visit upstream nodes of the high-risk event node, and arrange all nodes in the recursive visit process in reverse order of visit to form a causal chain. Starting from a high-risk event node, traverse downstream nodes originating from the high-risk event node in a hierarchical manner, and arrange the nodes traversed in each layer of the hierarchical visit process in ascending order of visit to form a consequence chain. Connect the causal chain and the consequence chain with the high-risk event node as the hub to form a risk event causal chain. Starting from the high-risk event node, traverse all paths along the directed edges, sum the directed edge weights and risk level coefficients of each path to obtain the total weight of each path. The risk level coefficient is determined based on the proportion of actual loss amount for different risk levels in historical accident reports. Sort the total weights of each path in descending order and select the top K paths as diffusion paths, where K is usually 5.

[0045] It should also be noted that the named entity recognition model automatically extracts triples from text logs to construct a directed acyclic graph with a clear causal direction. It traces the causal chain backwards and predicts the consequence chain forwards to form a complete causal chain; it calculates the total weight of each path and selects the top K key diffusion paths. This overcomes the limitations of traditional static graphs, quantifies the strength of risk propagation paths, and enhances the predictive ability of nonlinear diffusion.

[0046] S4: By comparing the preliminary risk level assessment results with the actual handling records, misjudgment cases are identified, triggering dynamic optimization of the symbolic logic expression of the hybrid neural symbolic architecture.

[0047] The specific steps are as follows: The preliminary risk level assessment results were compared with the actual risk level labels in the actual handling records one by one, and risk events with inconsistent levels were marked as misjudgments. The percentage of misjudged cases out of the total number of assessed cases is used to obtain the misjudgment rate of the preliminary risk level assessment. Based on the historical misjudgment rate, a misjudgment threshold is set using the percentile method. For example, the historical misjudgment rate at the 80th percentile is used as the misjudgment threshold. When the misjudgment rate exceeds the misjudgment threshold, the neural network layer structure of the hybrid neural symbolic architecture is adjusted. Batch normalization processing units are inserted between the original convolutional layers and activation layers in the neural network layer of the hybrid neural symbolic architecture.

[0048] Extract the real-time data stream of misjudged cases, recalculate the cross-entropy loss function value, use backpropagation to calculate the partial derivative of the new cross-entropy loss function value with respect to the threshold parameter, and use the Adam optimizer to adjust the optimized threshold parameter; The parameters of the symbolic logic layer and neural network layer of the hybrid neural symbolic architecture are fixed. The real-time data stream of misjudged cases is input into the hybrid neural symbolic architecture and executed to the fusion layer to output the risk probability distribution. The difference between the risk probability distribution and the actual risk label is compared as the risk probability loss value. The partial derivative of the risk probability loss value with respect to the attention weight parameters of the fusion layer is calculated, and the attention weight parameters are adjusted using the Adam optimizer. The adjusted threshold parameters and attention weight parameters are loaded into the hybrid neural symbolic architecture to obtain the risk assessment model.

[0049] It should also be noted that: by comparing the preliminary assessment results with the actual handling records, misjudgment cases are identified, triggering a closed-loop optimization mechanism. Backpropagation calculates the partial derivatives of the cross-entropy loss with respect to the threshold parameters and the partial derivatives of the risk probability loss value with respect to the attention weight parameters of the fusion layer. Combined with the Adam optimizer, the symbolic logic expression parameters and attention weight parameters are dynamically adjusted; this achieves the iteration of the hybrid neural symbolic architecture: threshold parameter optimization improves the generalization ability to complex scenarios, solves the problem of misjudgment accumulation caused by fixed parameters, enables the assessment to have self-optimization capabilities, and reduces the long-term misjudgment rate.

[0050] S5: Based on the risk assessment model and risk knowledge graph, the risk knowledge graph, the causal chain of risk events and the diffusion path are overlaid on the digital twin map, and the tiered early warning signals are pushed to the managers through the early warning feedback interface.

[0051] The specific steps are as follows: Extract equipment coordinates from the digital twin map of the chemical industrial park and associate risk event nodes in the risk knowledge graph with the equipment coordinates; transform the causal chain of risk events into spatial path segments: connect the causal nodes, risk event nodes, and consequence nodes sequentially with arrows, and distinguish the line segment colors according to the risk level, with red indicating high risk level, yellow indicating medium risk level, and blue indicating low risk level.

[0052] Based on environmental monitoring data, with the leak source as the origin, the prevailing wind direction as the positive x-axis, the horizontal diffusion along the vertical wind direction as the y-axis, and the vertical height as the z-axis, a Gaussian plume model is used to simulate the pollutant diffusion trajectory and calculate the pollutant concentration at point (x, y, z). The expression is as follows: ; in, Let (x,y,z) be the pollutant concentration at point (x,y,z). For the strength of the leakage source, For wind speed, The horizontal diffusion parameter, The vertical diffusion parameter, This represents the offset distance perpendicular to the wind direction. Vertical height The effective height of the leak source, This represents the distance downwind.

[0053] The safe concentration threshold of pollutants is determined according to the "Occupational Exposure Limits for Hazardous Factors in the Workplace". The distance at which the safe concentration threshold of the pollutant is first exceeded along the prevailing wind direction is defined as the diffusion radius.

[0054] The system pushes tiered early warning signals to managers through an early warning feedback interface, including risk level, causal chain of risk events, diffusion path, and diffusion radius.

[0055] It should also be noted that overlaying the risk knowledge graph with a digital twin map enables spatial mapping of risk elements. The Gaussian plume model simulates pollutant diffusion trajectories based on environmental monitoring data, and calculates the diffusion radius using the safe concentration thresholds from the "Occupational Exposure Limits for Hazardous Factors in the Workplace," quantifying the scope of environmental impact. Tiered early warning signals, through visualization of causal chains and diffusion paths, help managers quickly locate risk sources and their impact range. Transforming abstract risk assessments into intuitive spatial decision support improves emergency response efficiency.

[0056] This embodiment also provides an intelligent assessment system for environmental pollution risks in chemical industrial parks, including: The data acquisition module collects multi-source heterogeneous data from the chemical industrial park, performs spatiotemporal alignment on the multi-source heterogeneous data, and generates a real-time data stream. The multi-source heterogeneous data includes environmental monitoring data, equipment operating status data, and text log data. The assessment module extracts association rules of risk events based on real-time data streams, encodes the association rules into symbolic logic expressions, constructs a hybrid neural symbolic architecture, optimizes the parameters of the symbolic logic expressions and the hybrid neural symbolic architecture through joint training, and outputs preliminary risk level assessment results. The reasoning module extracts the "risk event-cause-effect" triples based on real-time data streams, generates a risk knowledge graph, and performs graph reasoning based on the preliminary risk level assessment results to obtain the causal chain and diffusion path of the risk event. The update module identifies misjudgment cases by comparing the preliminary risk level assessment results with the actual handling records, triggering dynamic optimization of the hybrid neural symbolic architecture; The early warning module, based on the risk assessment model and risk knowledge graph, overlays the risk knowledge graph, the causal chain of risk events, and the diffusion path onto the digital twin map, and pushes tiered early warning signals to managers through the early warning feedback interface.

[0057] This embodiment also provides a computer device applicable to the intelligent assessment method for environmental pollution risks in chemical industrial parks, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the intelligent assessment method for environmental pollution risks in chemical industrial parks as proposed in the above embodiment.

[0058] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices 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 and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0059] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the intelligent assessment method for environmental pollution risks in chemical industrial parks as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0060] In summary, this invention achieves efficient integration and dynamic optimization of multi-source heterogeneous data by constructing a hybrid neural symbolic architecture. This solves the problem of weakened spatiotemporal correlation caused by fragmented data processing in existing methods, enabling the hybrid neural symbolic architecture to adapt to risk characteristics in different scenarios. It avoids the fragmented data processing problem of independent modules in existing methods and reduces information loss caused by data silos. Furthermore, by generating a risk knowledge graph, it achieves dynamic modeling and accurate reasoning of causal chains of risk events, solving the problem of insufficient static causal chain modeling in existing methods. It quantifies the propagation path and weight of risk events, enabling the knowledge graph to continuously adapt to real-world scenarios and predict nonlinear risk diffusion.

[0061] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligent assessment of environmental pollution risks in chemical industrial parks, characterized by: include, Collect multi-source heterogeneous data from the chemical industrial park, perform spatiotemporal alignment on the multi-source heterogeneous data, and generate a real-time data stream. The multi-source heterogeneous data includes environmental monitoring data, equipment operating status data, and text log data. Based on real-time data streams, association rules of risk events are extracted, and these association rules are encoded into symbolic logic expressions. A hybrid neural symbolic architecture is constructed, and the parameters of the symbolic logic expressions and the hybrid neural symbolic architecture are optimized through joint training to output preliminary risk level assessment results. Based on real-time data streams, extract the "risk event-cause-effect" triplet to generate a risk knowledge graph. Combine the graph reasoning with the preliminary risk level assessment results to obtain the causal chain and diffusion path of the risk event. By comparing the preliminary risk level assessment results with the actual handling records, misjudgment cases are identified, triggering dynamic optimization of the hybrid neural symbolic architecture; Based on risk assessment models and risk knowledge graphs, the risk knowledge graph, causal chains of risk events, and diffusion paths are overlaid on a digital twin map, and tiered early warning signals are pushed to managers through an early warning feedback interface.

2. The intelligent assessment method for environmental pollution risks in chemical industrial parks as described in claim 1, characterized in that: The step of extracting association rules for risk events based on real-time data streams and encoding these association rules into symbolic logic expressions refers to extracting association rules for risk events from real-time data streams and encoding these association rules into symbolic logic expressions containing trainable threshold parameters.

3. The intelligent assessment method for environmental pollution risks in chemical industrial parks as described in claim 1, characterized in that: The process of jointly training and optimizing symbolic logic expressions and hybrid neural symbolic architecture parameters to output preliminary risk level assessment results is as follows. Symbolic logic expressions are executed through a symbolic logic layer of a hybrid neural symbolic architecture to generate confidence vectors for association rules; A neural network layer with a hybrid neural symbolic architecture is used to process real-time data streams and generate spatiotemporal feature vectors. The fusion layer of the hybrid neural symbolic architecture is used to fuse the rule confidence vector and the spatiotemporal feature vector to obtain the fused feature vector; Map the fused feature vectors to a risk probability distribution; The difference between the risk probability distribution and the actual risk label in historical accident reports is calculated using the cross-entropy loss function, and the cross-entropy loss function value is obtained. Backpropagation is used to generate the gradient matrix of the threshold parameter and the hybrid neural symbolic architecture parameter in the symbolic logic expression, and an adaptive learning rate optimizer is used to update the threshold parameter and the hybrid neural symbolic architecture parameter in the symbolic logic expression. The real-time data stream is input into the updated hybrid neural symbolic architecture to generate a risk probability distribution. The risk level is then classified according to the risk probability distribution to obtain a preliminary risk level assessment result.

4. The intelligent assessment method for environmental pollution risks in chemical industrial parks as described in claim 3, characterized in that: The process of extracting "risk event-cause-effect" triples from real-time data streams to generate a risk knowledge graph involves the following steps: Use a named entity recognition model to extract the "risk event-cause-effect" triple; Construct an initial causal graph based on the "risk event-cause-effect" triple; Transform the directed acyclic graph into a graph structure, define risk event nodes, cause nodes and consequence nodes, directed edges and edge weights, and form a risk knowledge graph.

5. The intelligent assessment method for environmental pollution risks in chemical industrial parks as described in claim 4, characterized in that: The process of combining the preliminary risk level assessment results with graph reasoning to obtain the causal chain and diffusion path of the risk event involves the following specific steps. Starting from a high-risk event node, trace the causal chain backwards along the directed edges; Predict the consequence chain along the directed edge from the nodes of high-risk events to form a causal chain of risk events; Starting from a high-risk event node, traverse all paths along directed edges to generate a diffusion path.

6. The intelligent assessment method for environmental pollution risks in chemical industrial parks as described in claim 1, characterized in that: The specific steps for triggering the dynamic optimization of the hybrid neural symbolic architecture are as follows. The error rate of the preliminary risk level assessment results was statistically analyzed; The neural network layer structure of the hybrid neural symbolic architecture is adjusted based on the misclassification rate, and batch normalization processing units are inserted into the neural network layer of the hybrid neural symbolic architecture. Extract the real-time data stream of misjudged cases and adjust the threshold parameters of the symbolic logic layer and the fusion layer parameters of the hybrid neural symbolic architecture; The threshold parameters of the symbolic logic layer and the parameters of the fusion layer of the adjusted hybrid neural symbolic architecture are loaded into the hybrid neural symbolic architecture to obtain the risk assessment model.

7. The intelligent assessment method for environmental pollution risks in chemical industrial parks as described in claim 1, characterized in that: The specific steps for overlaying the risk knowledge graph, risk event causal chain, and diffusion path onto the digital twin map are as follows. Extract equipment coordinates from the digital twin map of the chemical industrial park and associate risk event nodes in the risk knowledge graph with the equipment coordinates; Transform the causal chain of risk events into spatial path segments; Based on environmental monitoring data, the pollutant diffusion trajectory was simulated using a Gaussian plume model to calculate the pollutant concentration distribution; The diffusion radius is determined according to the "Occupational Exposure Limits for Hazardous Factors in the Workplace".

8. An intelligent assessment system for environmental pollution risks in chemical industrial parks, based on the intelligent assessment method for environmental pollution risks in chemical industrial parks as described in any one of claims 1 to 7, characterized in that: include, The data acquisition module collects multi-source heterogeneous data from the chemical industrial park, performs spatiotemporal alignment on the multi-source heterogeneous data, and generates a real-time data stream. The multi-source heterogeneous data includes environmental monitoring data, equipment operating status data, and text log data. The assessment module extracts association rules of risk events based on real-time data streams, encodes the association rules into symbolic logic expressions, constructs a hybrid neural symbolic architecture, optimizes the parameters of the symbolic logic expressions and the hybrid neural symbolic architecture through joint training, and outputs preliminary risk level assessment results. The reasoning module extracts the "risk event-cause-effect" triples from the real-time data stream, generates a risk knowledge graph, and performs graph reasoning based on the preliminary risk level assessment results to obtain the causal chain and diffusion path of the risk event. The update module identifies misjudgment cases by comparing the preliminary risk level assessment results with the actual handling records, triggering dynamic optimization of the hybrid neural symbolic architecture; The early warning module, based on the risk assessment model and risk knowledge graph, overlays the risk knowledge graph, the causal chain of risk events, and the diffusion path onto the digital twin map, and pushes tiered early warning signals to managers through the early warning feedback interface.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the intelligent assessment method for environmental pollution risk in chemical industrial parks as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the intelligent assessment method for environmental pollution risk in chemical industrial parks as described in any one of claims 1 to 7.

Images