Environment risk early warning method and system based on attention mechanism

Abstract

This application relates to the field of artificial intelligence early warning technology, and provides an environmental risk early warning method and system based on an attention mechanism to improve the accuracy and targeting of environmental risk early warning. The method includes: dividing labeled multi-dimensional environmental risk features into a hierarchical structure to obtain multi-level risk features; determining and updating weights and adjusting the initial pyramid network hierarchy parameters to construct an environmental risk early warning pyramid network; embedding attention gates and feature cross-models into each level of the network; matching the environmental monitoring data to be processed to the corresponding level of the network according to type, and extracting key risk features through attention gates; obtaining a risk feature matrix that integrates multi-dimensional correlation information by mining nonlinear correlations through the feature cross-model; inputting the data into a time-series cascaded prediction model for progressive inference to generate an environmental risk transmission inference graph; and combining hierarchical weights with risk node correlation assignments to generate hierarchical risk early warning results.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence early warning technology, specifically relating to an environmental risk early warning method and system based on an attention mechanism. Background Technology

[0002] With the acceleration of industrialization and urbanization, environmental risk factors are becoming increasingly complex, making environmental risk early warning a crucial link in ecological environmental protection and public safety. Existing environmental risk early warning technologies are mostly based on single-dimensional or limited-dimensional feature data. Common examples include using only air quality or water quality monitoring data for risk assessment. Some technologies attempt to integrate multi-dimensional data, but they often employ a flat data processing approach, failing to hierarchically classify environmental risk features and making it difficult to distinguish the different contributions of various features to risk early warning. Furthermore, existing technologies rely heavily on traditional statistical analysis methods or simple machine learning models, such as linear regression and decision trees, in the feature processing stage. These methods struggle to effectively uncover non-linear relationships between data, resulting in the underutilization of the potential value of risk features.

[0003] In terms of risk transmission analysis, existing technologies mostly focus on static risk status assessment, lacking time-series tracking and progressive reasoning of the risk transmission process, and cannot dynamically present the evolution path of risks. It is evident that existing technologies are insufficient to guarantee the accuracy and relevance of environmental risk early warning. Summary of the Invention

[0004] This application provides an environmental risk early warning method and system based on an attention mechanism to improve the accuracy and relevance of environmental risk early warning.

[0005] This application provides an environmental risk early warning method based on an attention mechanism, applied to an environmental risk early warning system. The method includes: Multi-level risk features are obtained by dividing labeled multi-dimensional environmental risk features into hierarchical structures. Update weights are determined according to the hierarchical order of each level of risk features. The hierarchical parameters of the initial pyramid network are adjusted using the update weights to obtain an environmental risk early warning pyramid network. Attention gates and feature cross-models are embedded in each level of the environmental risk early warning pyramid network; wherein, the attention gate is connected to the input end of the corresponding level network, and the feature cross-model is connected to the output end of the attention gate. The acquired environmental monitoring data to be processed is matched to each level of the environmental risk early warning pyramid network according to the data type. The key risk features of the environmental monitoring data to be processed are extracted by performing feature response analysis through the attention gates of each level of the network. The key risk features are received by the feature cross-model of each level network and the corresponding attention gate output. The nonlinear correlation between the key risk features is mined in the preset feature interaction space, and a risk feature matrix that integrates multi-dimensional correlation information is output. The risk feature matrix is ​​input into the time-series cascaded prediction model connected to the environmental risk early warning pyramid network to perform progressive reasoning in the time-series dimension, thereby obtaining the environmental risk transmission reasoning graph. The hierarchical weights of each network level are associated with the risk nodes in the environmental risk transmission inference graph and assigned values. Based on the assigned risk nodes and the risk transmission path corresponding to the environmental risk transmission inference graph, a hierarchical risk warning result is generated.

[0006] This application provides an environmental risk early warning system, which includes a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the above-described method.

[0007] This application provides a computer-readable storage medium including a computer program. When the computer program is run on an environmental risk early warning system, the computer program is used to cause the environmental risk early warning system to perform the steps of the above-described method.

[0008] This application's embodiments construct an environmental risk early warning pyramid network by dividing labeled multi-dimensional environmental risk features into a hierarchical structure and determining update weights. Attention gates and feature cross-models are embedded in each level of the network. Environmental monitoring data to be processed is matched to the corresponding level of the network according to type, and key risk features are extracted through attention gates. Then, nonlinear correlations are mined using the feature cross-model to obtain a risk feature matrix that integrates multi-dimensional correlation information. This matrix is ​​input into a time-series cascaded prediction model for progressive inference to generate an environmental risk transmission inference graph. Finally, hierarchical risk early warning results are generated by combining hierarchical weights and risk node correlation assignments. In this way, hierarchical and accurate modeling of environmental risk features is achieved. The synergistic effect of attention mechanisms and feature cross-models effectively captures potential data correlations. The time-series cascaded inference framework ensures dynamic tracking of risk transmission, and the hierarchical early warning results improve the targeting and effectiveness of risk response. This enhances the accuracy, comprehensiveness, and timeliness of environmental risk early warning, solving the problems of insufficient feature utilization, inadequate correlation mining, missing time-series tracking, and ambiguous result presentation in traditional early warning methods, thus improving the accuracy and targeting of environmental risk management and early warning. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating an environmental risk early warning method based on an attention mechanism, provided in an embodiment of this application.

[0010] Figure 2This is a schematic diagram of the structure of an environmental risk early warning system provided in an embodiment of this application.

[0011] Figure 3 This is a functional block diagram of an environmental risk early warning system provided in an embodiment of this application. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this application. Obviously, the described embodiments are only some embodiments of the technical solutions of this application, and not all embodiments. Based on the embodiments recorded in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the technical solutions of this application.

[0013] See Figure 1 This is an environmental risk early warning method based on an attention mechanism provided in the embodiments of this application. This method can be applied to an environmental risk early warning system. The specific process is as follows: steps 110-160.

[0014] Step 110: Divide the labeled multi-dimensional environmental risk features into a hierarchical structure to obtain multi-level risk features. Determine the update weights according to the hierarchical ranking of each level of risk features. Use the update weights to adjust the hierarchical parameters of the initial pyramid network to obtain the environmental risk early warning pyramid network.

[0015] In this embodiment, the environmental risk early warning system first acquires a multi-dimensional set of environmental risk features, including air quality characteristics, water quality characteristics, soil characteristics, meteorological characteristics, and pollution source characteristics. Air quality features include PM2.5 concentration, PM10 concentration, and sulfur dioxide concentration, while water quality features include pH value, chemical oxygen demand, and ammonia nitrogen concentration. All features are labeled with corresponding risk levels, such as low risk, medium risk, and high risk. The environmental risk early warning system hierarchically classifies features based on their direct impact on environmental risk and the complexity of their correlation. Features directly reflecting the environmental risk state, such as PM2.5 concentration and pH value, are classified as first-level risk features. Derivative features related to direct features, such as the correlation between PM2.5 and humidity, are classified as second-level risk features. Potential features affecting derivative features, such as pollution source emission patterns, are classified as third-level risk features, thus obtaining a multi-level risk feature set. Next, the environmental risk early warning system determines the update weights based on the hierarchical ranking of risk features at each level. Higher-level risk features correspond to larger update weights. For example, the update weight for the first-level risk features is set as the base weight value, the update weight for the second-level risk features is the base weight value multiplied by 1.2, and the update weight for the third-level risk features is the base weight value multiplied by 0.8. Subsequently, the environmental risk early warning system obtains an initial pyramid network, which includes an input layer, hidden layers, and an output layer. The hidden layers are divided according to different risk feature levels. The environmental risk early warning system uses the determined update weights to adjust the hierarchical parameters of the initial pyramid network, such as adjusting the number of neurons and connection weights at each level, ultimately obtaining the environmental risk early warning pyramid network.

[0016] Step 120: Embed attention gates and feature cross-models in each level of the environmental risk early warning pyramid network; wherein, the attention gate is connected to the input end of the corresponding level network, and the feature cross-model is connected to the output end of the attention gate.

[0017] In this embodiment, the environmental risk early warning system embeds an attention gate at the input of each level of the environmental risk early warning pyramid network. This attention gate employs a multi-head attention mechanism, containing multiple attention heads, each corresponding to a different feature dimension. Simultaneously, a feature cross-model is embedded at the output of the attention gate. This feature cross-model is a deep cross-model, comprising a feature cross-layer and a deep network layer. For example, in the first level of the network, the attention gate is embedded between the input layer and the hidden layer, receiving the input environmental risk features. After attention calculation, it outputs weighted features. The feature cross-model receives these weighted features, performs cross-combination of the features through the feature cross-layer, and then performs feature extraction through the deep network layer, finally outputting the processed features.

[0018] Step 130: Match the acquired environmental monitoring data to be processed according to data type to each level of the environmental risk early warning pyramid network, and perform feature response analysis through the attention gates of each level network to extract the key risk features of the environmental monitoring data to be processed.

[0019] In this embodiment, the environmental risk early warning system first acquires environmental monitoring data to be processed, such as real-time PM2.5 concentration data, pH value data, and soil heavy metal content data. Next, the system categorizes this data by type, for example, classifying PM2.5 and PM10 concentration data as air quality data, and pH and chemical oxygen demand data as water quality data. Then, based on the risk characteristic types corresponding to each level of the environmental risk early warning pyramid network, the system matches different types of environmental monitoring data to the corresponding level network. For example, air quality data is matched to the first level network, water quality data to the first level network, and pollution source emission data to the third level network. Subsequently, the attention gates of each level network perform feature response analysis on the environmental monitoring data to be processed, calculating the attention weight of each feature. For example, the attention weight for PM2.5 concentration is calculated to be 0.6, the attention weight for PM10 concentration is 0.4, and the attention weight for pH is 0.7. The environmental risk early warning system extracts key risk features based on attention weights, such as features with an attention weight greater than 0.5.

[0020] Step 131: Obtain multi-source environmental monitoring data to be processed, classify the multi-source environmental monitoring data to be processed according to the data representation dimension and data acquisition scenario, and obtain the environmental monitoring data to be processed after data type labeling.

[0021] In this embodiment, the environmental risk early warning system acquires multi-source environmental monitoring data from different monitoring stations and devices, such as air quality data from urban monitoring stations, water quality data from river monitoring stations, and pollution source data from industrial park monitoring stations. The system categorizes the data by data representation dimension into concentration-based data, ratio-based data, and count-based data, and by data collection scenario into urban environmental scenario data, industrial environmental scenario data, and natural environmental scenario data. Then, the system labels each piece of environmental monitoring data with a data type, for example, labeling PM2.5 concentration data collected from urban monitoring stations as "Air Quality - Concentration - Urban Environmental Scenario" data type, and labeling chemical oxygen demand data collected from industrial park monitoring stations as "Water Quality - Concentration - Industrial Environmental Scenario" data type, thus obtaining the labeled environmental monitoring data.

[0022] Step 132: Based on the risk characteristic hierarchical attributes corresponding to each level of the environmental risk early warning pyramid network, establish a matching mapping relationship between data types and hierarchical networks, and allocate the environmental monitoring data to be processed after data type labeling to the corresponding hierarchical networks according to the matching mapping relationship.

[0023] In this embodiment, the environmental risk early warning system analyzes the risk feature hierarchy attributes corresponding to each level of the environmental risk early warning pyramid network. The risk feature hierarchy attributes corresponding to the first level network are features that directly reflect the environmental risk status, and the corresponding data types are "air quality-concentration type," "water quality-concentration type," etc. The risk feature hierarchy attributes corresponding to the second level network are derived features associated with the direct features, and the corresponding data types are "air quality-related type," "water quality-related type," etc. The risk feature hierarchy attributes corresponding to the third level network are potential impact features, and the corresponding data types are "pollution source-pattern type," etc. Based on this, the environmental risk early warning system establishes a matching mapping relationship table between data types and hierarchical networks. For example, the "air quality-concentration type" data type is mapped to the first level network, and the "pollution source-pattern type" data type is mapped to the third level network. Subsequently, the environmental risk early warning system uses the matching mapping table to assign the environmental monitoring data to be processed after data type labeling to the corresponding hierarchical network. For example, PM2.5 concentration data labeled as "air quality - concentration category - urban environment scenario" is assigned to the first hierarchical network, and pollution source emission data labeled as "pollution source - regularity category - industrial environment scenario" is assigned to the third hierarchical network.

[0024] Step 133: Initialize the intra-level feature response of the environmental monitoring data received by each level network, load the level parameters of the corresponding level network through the attention gate, perform feature sensitivity analysis on the initialized environmental monitoring data, and screen out the initial response features that have a response correlation with risk warning.

[0025] In this embodiment, the environmental risk early warning system initializes the intra-level feature responses of the environmental monitoring data received by each layer of the network. For example, it converts PM2.5 concentration data and pH value data received by the first-level network into standard feature vector form and sets an initial response threshold. Next, the environmental risk early warning system loads the layer parameters of the corresponding layer network through an attention gate. These parameters include feature weight matrices and bias vectors. Then, the environmental risk early warning system uses the attention gate to perform feature sensitivity analysis on the initialized environmental monitoring data, calculating the sensitivity value of each feature. The sensitivity value is determined by the degree of correlation between the feature and the risk early warning result. For example, the sensitivity value of the PM2.5 concentration feature is calculated as the degree of influence of the feature change on the risk level change. The environmental risk early warning system selects features with sensitivity values ​​greater than the initial response threshold as initial response features, such as PM2.5 concentration features and pH value features.

[0026] Step 134: Perform cross-feature dimension correlation comparison between the initial response features and the multi-level risk features of the corresponding hierarchical network, identify the overlapping and differential feature dimensions between the initial response features and the multi-level risk features, and perform dimensionless feature enhancement processing on the differential feature dimensions based on the response intensity of the overlapping feature dimensions after standardization.

[0027] In this embodiment, the environmental risk early warning system performs a cross-feature dimension correlation comparison between the initial response features of the first-level hierarchical network, such as PM2.5 concentration and pH value, and the multi-level risk features of the same hierarchical network. The multi-level risk features include PM2.5 concentration and pH value features from the first-level risk features, and PM2.5-humidity correlation features from the second-level risk features. The environmental risk early warning system identifies overlapping feature dimensions as PM2.5 concentration and pH value features, and differential feature dimensions as PM2.5-humidity correlation features. Next, the environmental risk early warning system standardizes the response intensity of the overlapping feature dimensions, converting the response intensity values ​​to the range of 0-1. Subsequently, based on the standardized response intensity of the overlapping feature dimensions, the environmental risk early warning system performs dimensionless feature enhancement processing on the differential feature dimensions. For example, if the standardized response intensity of the overlapping feature dimension is 0.8, the enhancement coefficient of the differential feature dimension is set to 0.8, and the enhanced differential feature dimension is obtained by multiplying the feature value of the differential feature dimension by this enhancement coefficient.

[0028] Step 135: The enhanced differential feature dimension and the overlapping feature dimension are fused through the attention gate to generate a hierarchical response feature set, and the feature components used for risk indication are extracted from the hierarchical response feature set.

[0029] In this embodiment, the environmental risk early warning system fuses the enhanced differential and overlapping feature dimensions using an attention gate, converting them into feature vectors of the same dimension. Then, a vector concatenation operation is performed to generate a hierarchical response feature set. This set contains response features at different levels, such as first-level response features, second-level response features, etc. Next, the environmental risk early warning system extracts feature components from the hierarchical response feature set used for risk indication. These feature components can directly or indirectly reflect the changing trend of environmental risks, such as PM2.5 concentration feature components, pH value feature components, and enhanced PM2.5-humidity correlation feature components.

[0030] Step 136: Perform cross-level feature calibration and dimensionality bias correction on the feature components output by each level of the network to obtain key risk features of the environmental monitoring data to be processed, covering various data types and network levels.

[0031] In this embodiment, the environmental risk early warning system acquires feature components output from each level of the network, such as PM2.5 concentration and pH value feature components output from the first level network, PM2.5 and humidity correlation feature components output from the second level network, and pollution source emission pattern feature components output from the third level network. The environmental risk early warning system performs cross-level feature calibration on these feature components, adjusting the scale of feature components at different levels to bring them to the same order of magnitude. Then, the environmental risk early warning system corrects for dimensionality bias in the feature components, eliminating biases caused by different feature dimensions. Finally, the environmental risk early warning system obtains key risk features of the environmental monitoring data to be processed, covering various data types and network levels, such as calibrated PM2.5 concentration features and corrected pH value features.

[0032] Step 140: Receive key risk features output by the corresponding attention gates through the feature cross-model of each network level, mine the nonlinear correlation between the key risk features in the preset feature interaction space, and output a risk feature matrix that integrates multi-dimensional correlation information.

[0033] In this embodiment, the feature cross-model of each layer of the environmental risk early warning system receives key risk features output by the corresponding attention gate. For example, the feature cross-model of the first layer receives key risk features such as PM2.5 concentration and pH value. The environmental risk early warning system mines the nonlinear correlation between key risk features within a preset feature interaction space, which is set as a three-dimensional feature space, with each dimension corresponding to a different feature attribute. For example, in the first layer of the network, the feature cross-model mines the nonlinear correlation between PM2.5 concentration and pH value, finding that when PM2.5 concentration increases, pH value shows a decreasing trend, and this trend is not a linear relationship. Then, the environmental risk early warning system fuses the mined nonlinear correlation information with the key risk features to generate a risk feature matrix that integrates multi-dimensional correlation information. The rows of this matrix represent different key risk features, the columns represent different correlation dimensions, and the matrix elements represent the correlation strength between features.

[0034] Step 141: Receive the key risk features output by the corresponding attention gate through the feature cross-model of each layer network, initialize the preset feature interaction space based on the layer attributes of the corresponding layer network, and make the dimension of the preset feature interaction space match the feature dimension of the key risk features.

[0035] In this embodiment, the feature cross-model of each layer of the environmental risk early warning system receives key risk features output by the corresponding attention gate. For example, the feature cross-model of the second-layer network receives key risk features such as the correlation between PM2.5 and humidity, and the correlation between chemical oxygen demand and temperature. The environmental risk early warning system initializes a preset feature interaction space based on the hierarchical attributes of the corresponding layer network. The hierarchical attribute of the second-layer network is to process derived correlation features; therefore, the preset feature interaction space is set as a four-dimensional feature space, with each dimension corresponding to different attributes of the derived correlation features. The environmental risk early warning system adjusts the dimensions of the preset feature interaction space to match the feature dimensions of the key risk features. For example, if the feature dimension of the key risk features is 4, then the dimension of the preset feature interaction space is set to 4.

[0036] Step 142: Map the standardized key risk features to the preset feature interaction space to construct a spatial distribution representation of the key risk features, and determine the initial correlation strength between key risk features through spatial distance calculation.

[0037] In this embodiment, the environmental risk early warning system standardizes key risk features, converting their values ​​to the range of 0-1. Next, the system maps these standardized key risk features to a preset feature interaction space, with each key risk feature corresponding to a point in the space, thus constructing a spatial distribution representation of the key risk features. Then, the system determines the initial correlation strength between key risk features through spatial distance calculation; the smaller the spatial distance, the stronger the initial correlation. For example, the Euclidean distance between corresponding points of two key risk features is calculated, and the reciprocal of this distance is used as the initial correlation strength.

[0038] Step 143: Group key risk features based on initial association strength to form multiple initial association feature groups. Perform nonlinear association mining on key risk features within each initial association feature group and generate in-group association features through feature interaction operations.

[0039] In this embodiment, the environmental risk early warning system groups key risk features based on initial correlation strength. A correlation strength threshold is set, and key risk features with initial correlation strength greater than this threshold are grouped together, forming multiple initial correlation feature groups. Examples include correlation feature groups containing PM2.5 concentration and humidity features, and correlation feature groups containing pH and temperature features. Next, the environmental risk early warning system performs nonlinear correlation mining on the key risk features within each initial correlation feature group, using a multinomial regression algorithm to analyze the nonlinear relationships between features. Then, the environmental risk early warning system generates intra-group correlation features through feature interaction operations, such as multiplying PM2.5 concentration and humidity features to generate PM2.5-humidity interaction features.

[0040] Step 144: Perform cross-group interaction on the intra-group correlation features of each initial correlation feature group to explore the potential non-linear correlations between different correlation feature groups, generate cross-group correlation features, and calculate the correlation confidence between cross-group correlation features and intra-group correlation features.

[0041] In this embodiment, the environmental risk early warning system performs cross-group interaction between intra-group correlation features of a feature group containing PM2.5 concentration and humidity characteristics, and intra-group correlation features of a feature group containing pH and temperature characteristics. An association rule mining algorithm is used to uncover potential nonlinear correlations between different feature groups, revealing a potential negative correlation between the PM2.5 humidity interaction feature and the pH and temperature interaction feature. Next, the environmental risk early warning system generates cross-group correlation features, such as a comprehensive interaction feature between PM2.5 humidity and pH and temperature. Simultaneously, the system calculates the correlation confidence between cross-group and intra-group correlation features, determining the correlation confidence by calculating the correlation coefficient between features; the larger the absolute value of the correlation coefficient, the higher the correlation confidence.

[0042] Step 145: Filter intra-group and cross-group association features based on association credibility, extract feature components that meet the association credibility requirements, and sort and organize the filtered feature components according to feature dimensions.

[0043] In this embodiment, the environmental risk early warning system sets a correlation confidence threshold and filters out intra-group and cross-group correlation features with a correlation confidence greater than the threshold, such as PM2.5 humidity interaction features and pH temperature interaction features. Next, the environmental risk early warning system extracts feature components from these features, such as concentration and temperature components in the interaction features. Then, the environmental risk early warning system sorts and organizes the filtered feature components according to feature dimensions, such as concentration, temperature, and humidity.

[0044] Step 146: When the generated associated features no longer add valid associated information, combine all the obtained associated features according to the preset matrix arrangement rules and output a risk feature matrix that integrates multi-dimensional associated information.

[0045] In this embodiment, the environmental risk early warning system continuously generates associated features. When no new valid associated information is added after multiple consecutive generation of associated features, generation stops. Valid associated information refers to information that can improve the accuracy of risk early warning, and is determined by comparing the risk early warning accuracy before and after adding new associated features. Next, the environmental risk early warning system combines all obtained associated features according to a preset matrix arrangement rule. The preset matrix arrangement rule is to arrange the features in descending order of feature dimension, combining the associated features into a risk feature matrix. Rows in the matrix represent different associated features, columns represent different feature dimensions, and matrix elements represent the values ​​of the feature dimensions.

[0046] Step 150: Input the risk feature matrix into the time-series cascaded prediction model connected to the environmental risk early warning pyramid network to perform progressive reasoning in the time-series dimension, and obtain the environmental risk transmission reasoning graph.

[0047] In this embodiment, the environmental risk early warning system inputs a risk feature matrix that integrates multi-dimensional related information into a time-series cascaded prediction model. This model includes an input layer, a time-series processing layer, and an output layer. The time-series cascaded prediction model performs progressive time-series reasoning on the risk feature matrix. First, it analyzes the feature changes of the risk feature matrix at different time points, such as the PM2.5 concentration characteristics in the first hour and the PM2.5 concentration characteristics in the second hour. Then, based on the feature changes, the time-series cascaded prediction model infers the transmission path of environmental risks. For example, increased PM2.5 concentration leads to decreased visibility, which in turn affects traffic. Finally, an environmental risk transmission inference graph is formed, which includes risk nodes and the transmission paths between nodes.

[0048] Step 151: Input the risk feature matrix into the time series cascaded prediction model, and perform time series feature hierarchical decomposition on the risk feature matrix through the feature decomposition sub-model of the time series cascaded prediction model to obtain the feature component set of different time series stages.

[0049] In this embodiment, the environmental risk early warning system inputs the risk feature matrix into the feature decomposition sub-model of the time-series cascaded prediction model. This sub-model includes a feature extraction unit and a time-series stratification unit. The feature decomposition sub-model performs time-series feature stratification on the risk feature matrix. First, it extracts time-series features from the risk feature matrix, such as PM2.5 concentration features and pH value features at different time points. Then, the time-series stratification unit stratifies the time-series features according to time stages, such as classifying the features of the first 1-2 hours as the first time-series stage, the features of the third 3-4 hours as the second time-series stage, etc., to obtain feature component sets for different time-series stages.

[0050] Step 1511: Receive the risk feature matrix using the feature decomposition sub-model of the time series cascaded prediction model, and set time series stage division rules based on the time series evolution law of risk transmission. The time series stage division rules are adapted to the natural evolution cycle of environmental risks.

[0051] In this embodiment, the environmental risk early warning system utilizes the eigenvalue decomposition sub-model of the time-series cascaded prediction model to receive a risk feature matrix, which contains environmental risk features at multiple time points. The environmental risk early warning system sets time-series stage division rules based on the temporal evolution law of risk transmission. The temporal evolution law of risk transmission is the process of environmental risk from occurrence to diffusion and then to mitigation. Therefore, the time-series stage division rules are set according to different stages of risk evolution, such as the risk occurrence stage, risk diffusion stage, and risk mitigation stage. The duration of each stage is adapted to the natural evolution cycle of environmental risk; for example, the duration of the risk occurrence stage is set to 2 hours, the duration of the risk diffusion stage is set to 4 hours, and the duration of the risk mitigation stage is set to 6 hours.

[0052] Step 1512: According to the time series segmentation rules, the time dimension corresponding to the risk feature matrix is ​​segmented to obtain multiple continuous and non-overlapping time series segments. Each time series segment corresponds to a time interval and a subset of risk features within that time interval.

[0053] In this embodiment, the environmental risk early warning system segments the time dimension corresponding to the risk feature matrix according to a set time-series stage division rule, with the time dimension ranging from hour 0 to hour 12. The system divides hour 0-2 into the risk occurrence stage, corresponding to a time interval of 0-2 hours, within which a subset of risk features includes environmental risk features from hour 0, hour 1, and hour 2; it divides hour 3-6 into the risk diffusion stage, corresponding to a time interval of 3-6 hours, within which a subset of risk features includes environmental risk features from hour 3, hour 4, hour 5, and hour 6; and it divides hour 7-12 into the risk mitigation stage, corresponding to a time interval of 7-12 hours, within which a subset of risk features includes environmental risk features from hour 7 to hour 12, resulting in multiple continuous and non-overlapping time-series stages.

[0054] Step 1513: Normalize the feature dimensions of the risk feature subset corresponding to each time series stage. Through the time series feature extraction unit of the feature decomposition sub-model, perform time series correlation feature mining on each normalized risk feature subset to identify the time evolution trend of each feature component in the corresponding time series stage and its time series correlation with other feature components.

[0055] In this embodiment, the environmental risk early warning system normalizes the feature dimensions of the risk feature subset corresponding to each time series stage, converting the value of the feature dimension to the range of 0-1. Next, the environmental risk early warning system uses the time series feature extraction unit of the feature decomposition sub-model to perform time series correlation feature mining on the normalized risk feature subset, employing time series analysis algorithms to analyze the temporal evolution trend of each feature component within the corresponding time series stage, such as the upward trend of PM2.5 concentration during the risk occurrence stage and the downward trend during the risk mitigation stage. Simultaneously, the time series feature extraction unit identifies the time series correlation between each feature component and other feature components, such as the negative correlation between PM2.5 concentration and humidity during the risk diffusion stage.

[0056] Step 1514: Based on the temporal evolution trend and temporal correlation, perform feature stratification on each risk feature subset, and group feature components with the same evolution trend and correlation into the same feature layer to form a multi-layer feature set for each temporal stage.

[0057] In this embodiment of the application, the environmental risk early warning system performs feature layering on each risk feature subset based on the temporal evolution trend and temporal correlation. For example, in the risk occurrence stage, the feature components with an upward trend and negative correlation with humidity are classified into the first feature layer, and the feature components with a downward trend and positive correlation with temperature are classified into the second feature layer, thereby forming a multi-layer feature set for each temporal stage.

[0058] Step 1515: Extract the effective feature layer reflecting the temporal pattern of risk transmission from the multi-layer feature set of each time series stage, integrate the effective feature layers of all time series stages into feature component sets of different time series stages, and iteratively verify the temporal continuity of the feature component sets of each time series stage. If there is a temporal correlation break in the feature component sets of adjacent time series stages, readjust the time series stage division rules and feature layering standards until the feature component sets of all time series stages form a continuous temporal feature chain.

[0059] In this embodiment, the environmental risk early warning system extracts effective feature layers reflecting the temporal pattern of risk transmission from the multi-layered feature sets of each time series stage. Effective feature layers refer to feature layers that clearly reflect the risk transmission process, such as the first feature layer of the risk occurrence stage and the second feature layer of the risk diffusion stage. Next, the environmental risk early warning system integrates the effective feature layers of all time series stages into feature component sets for different time series stages. Then, the environmental risk early warning system iteratively verifies the temporal continuity of the feature component sets of each time series stage, checking for any breaks in the correlation between feature component sets of adjacent time series stages, such as whether there is a mismatch in feature dimensions between the feature component set of the risk occurrence stage and the feature component set of the risk diffusion stage. If a temporal correlation break exists, the environmental risk early warning system readjusts the time series stage division rules and feature layering standards, such as adjusting the time length of the time series stages and re-dividing the feature layers, until the feature component sets of all time series stages form a continuous temporal feature chain.

[0060] Step 152: According to the chronological order of the time series stages, the feature components of the previous time series stage are used as inputs, and the first-level progressive reasoning is performed through the reasoning sub-model of the time series cascaded prediction model to generate the initial risk transmission path and corresponding risk nodes.

[0061] In this embodiment, the environmental risk early warning system inputs the characteristic components of the risk occurrence stage into the inference sub-model of the time-series cascaded prediction model, according to the chronological order of the risk occurrence stages. The inference sub-model includes a path generation unit and a node generation unit. The path generation unit generates an initial risk transmission path based on the changes in the characteristic components, such as PM2.5 concentration increase → visibility decrease → traffic congestion. The node generation unit generates corresponding risk nodes based on the path, such as PM2.5 concentration increase node, visibility decrease node, traffic congestion node, etc.

[0062] Step 153: Integrate the initial risk transmission path and corresponding risk nodes output by the first-level progressive inference with the feature components of the current time series stage, and perform the second-level progressive inference through the inference sub-model to update the extension direction of the risk transmission path and the attribute information of the risk nodes.

[0063] In this embodiment, the environmental risk early warning system fuses the initial risk transmission path and corresponding risk nodes output by the first-level progressive inference with the feature components of the risk diffusion stage. The fusion method involves concatenating the features of the initial risk transmission path with the feature components of the risk diffusion stage. Next, the environmental risk early warning system performs a second-level progressive inference through an inference sub-model. The inference sub-model updates the extension direction of the risk transmission path based on the fused features, such as extending traffic congestion to traffic accidents, and simultaneously updates the attribute information of the traffic congestion node, such as updating the congestion level from mild to severe.

[0064] Step 154: Complete the progressive reasoning of all time-series stages in sequence to obtain a preliminary risk transmission reasoning graph containing the risk transmission path and risk nodes of the entire time-series stages. Perform path detection on the preliminary risk transmission reasoning graph to identify abnormal transmission paths that do not conform to the time-series logic.

[0065] In this embodiment, the environmental risk early warning system sequentially reasons through the risk occurrence stage, risk diffusion stage, and risk mitigation stage to obtain a preliminary risk transmission reasoning diagram containing the risk transmission paths and risk nodes across all time-series stages. Next, the environmental risk early warning system performs path detection on the preliminary risk transmission reasoning diagram, using a temporal logic checking algorithm to verify whether the paths conform to the chronological order, such as whether a path in the risk mitigation stage precedes a path in the risk occurrence stage. The environmental risk early warning system identifies abnormal transmission paths that do not conform to temporal logic, such as a traffic congestion path in the risk mitigation stage preceding a PM2.5 concentration increase path in the risk occurrence stage.

[0066] Step 155: Based on the feature components of the current time series stage and the effective inference results of the previous time series stage, the abnormal propagation path is corrected and adjusted, and the risk node attributes and path extension direction corresponding to the abnormal propagation path are re-determined.

[0067] In this embodiment, the environmental risk early warning system corrects and adjusts abnormal transmission paths based on the characteristic components of the risk mitigation phase and the effective reasoning results of the risk occurrence and risk diffusion phases. For example, if the abnormal transmission path is a traffic congestion path in the risk mitigation phase that precedes the PM2.5 concentration increase path in the risk occurrence phase, the environmental risk early warning system adjusts the traffic congestion path to the risk diffusion phase based on the effective reasoning results of the preceding time-series phases, and re-determines the risk node attributes corresponding to the abnormal transmission path, such as adjusting the time attribute of the traffic congestion node from the risk mitigation phase to the risk diffusion phase. Simultaneously, it adjusts the path extension direction, such as extending the traffic congestion path to traffic congestion → accident handling.

[0068] Step 156: Integrate the corrected transmission path with the effective transmission path to construct an environmental risk transmission reasoning graph that contains complete temporal correlations and consistent path logic. The environmental risk transmission reasoning graph is used to characterize the temporal transmission relationship and path correlation characteristics of each risk node.

[0069] In this embodiment, the environmental risk early warning system integrates the corrected transmission path with the effective transmission path, where the effective transmission path refers to a transmission path that conforms to temporal logic. The integration method involves concatenating the corrected transmission path and the effective transmission path in chronological order to construct an environmental risk transmission reasoning graph containing complete temporal relationships and consistent path logic. This graph is used to characterize the temporal transmission relationship and path association characteristics of each risk node, such as the temporal transmission relationship of PM2.5 concentration increase node → visibility decrease node → traffic congestion node → traffic accident node, as well as the path association strength between nodes.

[0070] Step 160: Associate the hierarchical weights of each network level with the risk nodes in the environmental risk transmission inference graph, and generate hierarchical risk warning results based on the assigned risk nodes and the risk transmission paths corresponding to the environmental risk transmission inference graph.

[0071] In this embodiment, the environmental risk early warning system first obtains the hierarchical weights of each level of the environmental risk early warning pyramid network. The hierarchical weights are set according to the importance of the hierarchical network; for example, the hierarchical weight of the first level network is 0.6, the second level network is 0.3, and the third level network is 0.1. Next, the environmental risk early warning system associates the hierarchical weights of each level network with risk nodes in the environmental risk transmission inference graph. For example, the hierarchical weight of 0.6 for the first level network is assigned to nodes indicating increased PM2.5 concentration or decreased pH value. Then, based on the assigned risk nodes and risk transmission paths, the environmental risk early warning system generates hierarchical risk early warning results. These results contain risk early warning information at different levels; for example, the first level risk warning information is the risk of increased PM2.5 concentration, the second level is the risk of decreased visibility, and the third level is the risk of traffic congestion.

[0072] Step 161: Extract the hierarchical weights of each level of the environmental risk early warning pyramid network. The hierarchical weights reflect the importance and characteristic contribution of the corresponding level network in risk early warning.

[0073] In this embodiment, the environmental risk early warning system extracts the hierarchical weights of each level of the environmental risk early warning pyramid network. The hierarchical weights are calculated using the feature contribution and importance of the hierarchical network. The feature contribution refers to the degree of influence of the hierarchical network's features on the risk early warning result, while the importance refers to the hierarchical network's position in the entire early warning system. For example, if the feature contribution of the first-level network is 0.7 and its importance is 0.8, the hierarchical weight is calculated as the feature contribution multiplied by the importance, i.e., 0.7 × 0.8 = 0.56.

[0074] Step 162: Traverse all risk nodes in the environmental risk transmission reasoning graph, identify the risk feature source hierarchy network corresponding to each risk node, and establish the source association relationship between the risk node and the hierarchy network.

[0075] In this embodiment, the environmental risk early warning system traverses all risk nodes in the environmental risk transmission inference graph, such as nodes indicating increased PM2.5 concentration, decreased visibility, and traffic congestion. The system identifies the risk feature source hierarchy network corresponding to each risk node; for example, the risk feature source hierarchy network corresponding to the PM2.5 concentration increase node is the first-level hierarchy network, and the risk feature source hierarchy network corresponding to the visibility decrease node is the second-level hierarchy network. Then, the system establishes a source association table between risk nodes and hierarchy networks, containing information such as the risk node name and the name of the source hierarchy network.

[0076] Step 1621: In response to the risk node tracing analysis task, load the topology data of the environmental risk transmission reasoning graph, parse the unique identification information and feature description field of each risk node in the topology data, and convert the feature description field of the risk node into a standardized risk node feature vector.

[0077] In this embodiment, the environmental risk early warning system responds to the risk node source tracing analysis task by loading the topological structure data of the environmental risk transmission inference graph. This data includes the location information and connection relationship information of the risk nodes. The environmental risk early warning system parses the unique identification information of each risk node in the topological structure data, such as the node ID, and the feature description fields, such as increased PM2.5 concentration and decreased visibility. Then, the environmental risk early warning system transforms the feature description fields of the risk nodes into standardized risk node feature vectors. The standardization method is to convert the feature description fields into numerical form, such as converting increased PM2.5 concentration to 1 and decreased visibility to 2.

[0078] Step 1622: Based on the hierarchical division log of the environmental risk early warning pyramid network, extract the characteristic output port information and characteristic transmission link identifier of each level network and generate a correspondence index table between the hierarchical network and the characteristic transmission link.

[0079] In this embodiment, the environmental risk early warning system is based on the hierarchical division log of the environmental risk early warning pyramid network. This log includes information such as the division time and basis for each network level. The environmental risk early warning system extracts the characteristic output port information of each network level, such as port 1 for the first-level network and port 2 for the second-level network, as well as characteristic transmission link identifiers, such as link 1 and link 2. Then, the environmental risk early warning system generates a mapping index table between the hierarchical networks and the characteristic transmission links, which includes information such as the hierarchical network name, characteristic output port information, and characteristic transmission link identifier.

[0080] Step 1623: Using the standardized risk node feature vector as the retrieval basis, perform multi-dimensional fuzzy matching operation in the corresponding relationship index table to determine the candidate layer network set that has potential association with the risk node feature vector, and record the matching similarity value between each candidate layer network and the risk node.

[0081] In this embodiment, the environmental risk early warning system uses standardized risk node feature vectors as the retrieval basis and performs multi-dimensional fuzzy matching operations in the corresponding relationship index table. These multi-dimensional operations include feature dimensions, port dimensions, and link dimensions. The environmental risk early warning system identifies a set of candidate hierarchical networks that have potential associations with the risk node feature vectors. For example, the candidate hierarchical network set for nodes with increased PM2.5 concentration is the first-level network and the second-level network. Simultaneously, the environmental risk early warning system records the matching similarity value between each candidate hierarchical network and the risk node. The matching similarity value is obtained by calculating the similarity between the risk node feature vector and the candidate hierarchical network feature vector. For example, the matching similarity value between the first-level network and the node with increased PM2.5 concentration is 0.8, and the matching similarity value between the second-level network and the node with increased PM2.5 concentration is 0.5.

[0082] Step 1624: For each layer network in the candidate layer network set, call the historical feature output record of the layer network, extract the key risk feature matrix fragment output by the layer network in the corresponding time series stage, and perform feature dimension alignment processing on the key risk feature matrix fragment and the risk node feature vector.

[0083] In this embodiment, the environmental risk early warning system, for each layer of the candidate layer network set, such as the first layer network and the second layer network, calls the historical feature output records of that layer network. These historical feature output records contain feature output data from different time points. The environmental risk early warning system extracts key risk feature matrix fragments output by that layer network within the corresponding time series stage, such as the key risk feature matrix fragment output by the first layer network during the risk occurrence stage. Next, the environmental risk early warning system aligns the key risk feature matrix fragments with the feature vectors of the risk nodes, adjusting the feature dimensions of the key risk feature matrix fragments to be consistent with the feature dimensions of the risk node feature vectors.

[0084] Step 1625: Perform deep correlation mining on the aligned feature data using the feature attribution analysis algorithm, identify the source correspondence of each feature component in the risk node feature vector in the candidate level network feature matrix segment, and mark feature component pairs with direct inheritance relationship and indirect correlation feature component pairs.

[0085] In this embodiment, the environmental risk early warning system uses a feature attribution analysis algorithm to perform deep correlation mining on the aligned feature data. This algorithm analyzes the inheritance relationships between features. The system identifies the source-to-source correspondence between each feature component in the risk node feature vector and a segment of the candidate-level network feature matrix. For example, the PM2.5 concentration component in the risk node feature vector corresponds to the PM2.5 concentration feature component in the candidate-level network feature matrix segment. Then, the system identifies feature component pairs with direct inheritance relationships, such as PM2.5 concentration component and PM2.5 concentration feature component, as well as indirectly related feature component pairs, such as PM2.5 concentration component and humidity feature component.

[0086] Step 1626: Calculate the association confidence between the candidate hierarchical network and the risk node based on the inheritance relationship strength of the feature components, and obtain the quantitative association confidence evaluation results.

[0087] In this embodiment, the environmental risk early warning system calculates the association confidence between the candidate hierarchical network and the risk node based on the inheritance relationship strength of the feature components. The inheritance relationship strength refers to the degree of inheritance between feature components, which is determined by calculating the correlation coefficient between feature components. The association confidence is calculated as the average value of the inheritance relationship strength. For example, if the inheritance relationship strength between the candidate hierarchical network and the risk node is 0.7 and 0.8 respectively, the association confidence is calculated as (0.7+0.8) / 2=0.75.

[0088] Step 1627: Based on the accuracy label of the environmental risk warning and the feature contribution benchmark value of the hierarchical network, set the correlation confidence threshold, determine the candidate hierarchical network with the correlation confidence higher than the correlation confidence threshold as the source hierarchical network of the risk node, and remove the hierarchical network with the correlation confidence not meeting the standard and record the reason for removal.

[0089] In this embodiment, the environmental risk early warning system sets a correlation confidence threshold based on the accuracy label of the environmental risk early warning and the feature contribution benchmark value of the hierarchical network. The accuracy label refers to the accuracy of the risk early warning result, and the feature contribution benchmark value refers to the standard value of the feature contribution of the hierarchical network. For example, the accuracy label is 0.9, the feature contribution benchmark value is 0.6, and the correlation confidence threshold is set to 0.7. The environmental risk early warning system determines candidate hierarchical networks with a correlation confidence score higher than 0.7 as the source hierarchical networks of the risk nodes. For example, if the correlation confidence score of the first hierarchical network is 0.75, it is determined as the source hierarchical network; if the correlation confidence score of the second hierarchical network is 0.6, the hierarchical network is removed and the reason for removal is recorded. If the correlation confidence score does not meet the standard...

[0090] Step 1628: For risk nodes with multiple source hierarchical networks, obtain the feature output time series of each source hierarchical network, determine whether there is a causal relationship between the feature data of different source hierarchical networks and the formation of the risk node, and construct the multi-source hierarchical network contribution sequence of the risk node.

[0091] In this embodiment, the environmental risk early warning system targets risk nodes with multiple source hierarchical networks, such as traffic congestion nodes. It acquires the feature output time series of each source hierarchical network, for example, the feature output time series of the first hierarchical network is hour 1, and the feature output time series of the second hierarchical network is hour 2. The environmental risk early warning system determines whether there is a causal relationship between the feature data of different source hierarchical networks in the formation of the risk node. For example, if the feature data of the first hierarchical network is output first and plays a major role in the formation of the risk node, and the feature data of the second hierarchical network is output later and plays a secondary role in the formation of the risk node, then the environmental risk early warning system constructs a multi-source hierarchical network contribution sequence for the risk node. This sequence includes information such as the source hierarchical network name, feature output time series, and degree of contribution.

[0092] Step 1629: Based on the contribution sequence of the multi-source hierarchical network, with risk nodes as core nodes and source hierarchical networks as associated nodes, construct a source association graph between risk nodes and hierarchical networks by using weighted edges to represent the association strength and time sequence relationship between the two.

[0093] In this embodiment, the environmental risk early warning system is based on a multi-source hierarchical network contribution sequence. It constructs a source association graph between risk nodes and the hierarchical networks, with risk nodes as core nodes and source hierarchical networks as associated nodes. In the graph, weighted edges represent the association strength and temporal relationship between the two; a larger weight indicates a stronger association. The direction of the edge indicates the temporal relationship. For example, an edge with a weight of 0.8 pointing from the first hierarchical network to a risk node indicates an association strength of 0.8, and the temporal relationship is that the first hierarchical network outputs feature data first.

[0094] Step 16210: Perform loop detection on the source association graph. If an abnormal circular association is found between hierarchical networks, the association is corrected by feature transmission timing priority.

[0095] In this embodiment, the environmental risk early warning system performs loop detection on the source association graph. The loop detection algorithm is used to check whether there are abnormal circular associations in the graph, such as hierarchical network A pointing to hierarchical network B, hierarchical network B pointing to hierarchical network C, and hierarchical network C pointing to hierarchical network A. If an abnormal circular association is found, the environmental risk early warning system corrects the association relationship by using feature transmission timing priority. Feature transmission timing priority refers to the order in which feature data is transmitted, with higher priority feature data being transmitted first. For example, if the feature transmission timing priority of hierarchical network A is higher than that of hierarchical network B, the edge from hierarchical network B to hierarchical network C is adjusted to point from hierarchical network A to hierarchical network C.

[0096] Step 163: Based on the source association relationship, assign the hierarchical weight of the corresponding hierarchical network to the risk node to form a risk node with hierarchical weight attribute. At the same time, for risk nodes associated with multiple source hierarchical networks, perform weighted aggregation on the hierarchical weights based on the identified and quantified association confidence of each source hierarchical network, and calculate the aggregated weight value of each risk node.

[0097] In this embodiment, the environmental risk early warning system assigns hierarchical weights to risk nodes based on source association relationships. For example, the hierarchical weight of 0.56 for the first-level hierarchical network is assigned to the node indicating an increase in PM2.5 concentration. For risk nodes associated with multiple source hierarchical networks, such as a traffic congestion node associated with the first-level and second-level hierarchical networks, the environmental risk early warning system performs weighted aggregation of the hierarchical weights based on the identified and quantified association confidence of each source hierarchical network. The association confidences are 0.75 and 0.6, and the hierarchical weights are 0.56 and 0.48, respectively. The aggregated weight value is calculated as (0.75×0.56+0.6×0.48) / (0.75+0.6)=(0.42+0.288) / 1.35=0.708 / 1.35≈0.524.

[0098] Step 164: Analyze the path length and path association strength of the risk transmission paths in the environmental risk transmission reasoning graph. The path association strength is the average level of the aggregated weight values ​​of all risk nodes on the path. Determine the priority of each risk transmission path based on the path length and path association strength.

[0099] In this embodiment, the environmental risk early warning system analyzes the path length of risk transmission paths in the environmental risk transmission inference graph. Path length refers to the number of risk nodes on the path. For example, the path length for the path of increased PM2.5 concentration → decreased visibility → traffic congestion is 3. Simultaneously, the environmental risk early warning system analyzes the path association strength, which is the average level of the aggregated weight values ​​of all risk nodes on the path. For example, if the aggregated weight values ​​of risk nodes on the path are 0.56, 0.48, and 0.524, the path association strength is calculated as (0.56 + 0.48 + 0.524) / 3 = 1.564 / 3 ≈ 0.521. Next, the environmental risk early warning system determines the priority of each risk transmission path based on the path length and path association strength. The shorter the path length and the greater the path association strength, the higher the priority. For example, the priority calculation method is set to divide the path association strength by the path length; the larger the priority value, the higher the priority.

[0100] Step 165: Sort the risk transmission paths in descending order of priority, mark the core risk transmission paths that meet the preset priority requirements, and mark the risk nodes on the core risk transmission paths as key risk nodes.

[0101] In this embodiment, the environmental risk early warning system sorts the risk transmission paths in descending order of priority. For example, path 1 has a priority of 0.6, path 2 has a priority of 0.5, and path 3 has a priority of 0.4, resulting in the sorted path order: path 1, path 2, and path 3. Next, the environmental risk early warning system sets preset priority requirements. For example, if the priority is greater than 0.5, core risk transmission paths that meet the preset priority requirements are marked, such as path 1 and path 2. Then, the environmental risk early warning system marks the risk nodes on the core risk transmission paths as critical risk nodes. For example, nodes indicating increased PM2.5 concentration, decreased visibility, and traffic congestion on path 1 are all marked as critical risk nodes.

[0102] Step 166: Perform feature aggregation processing on key risk nodes by using their hierarchical weight attributes, temporal transmission characteristics, and path association information to generate key risk node aggregated features.

[0103] In this embodiment, the environmental risk early warning system performs feature aggregation processing on key risk nodes based on their hierarchical weight attributes, temporal transmission characteristics, and path association information. The hierarchical weight attribute refers to the hierarchical weight value of the key risk node, the temporal transmission characteristic refers to the transmission order of the key risk node in time, and the path association information refers to the association strength between the key risk node and other nodes. The feature aggregation processing method involves concatenating the above features to generate aggregated features for key risk nodes. For example, the aggregated features may include a hierarchical weight value of 0.56, a temporal transmission order of 1, and an association strength of 0.8.

[0104] Step 167: Based on the aggregation characteristics of key risk nodes and the priority ranking of core risk transmission paths, generate hierarchical risk warning sub-results corresponding to each level of the network according to the hierarchical division rules of the hierarchical network.

[0105] In this embodiment, the environmental risk early warning system generates hierarchical risk early warning sub-results corresponding to each level of the network based on the aggregation characteristics of key risk nodes and the priority ranking of core risk transmission paths, according to the hierarchical division rules of the network. The hierarchical division rule is to divide the network from high to low priority. For example, the first level of the network corresponds to high-priority risk early warning sub-results, the second level corresponds to medium-priority risk early warning sub-results, and the third level corresponds to low-priority risk early warning sub-results. For example, the priority of core risk transmission path 1 is 0.6, corresponding to the first level of the network, generating a high-priority risk early warning sub-result, which states that increased PM2.5 concentration leads to decreased visibility, thereby causing traffic congestion, and the risk level is high risk.

[0106] Step 168: Based on the redundancy and conflict of early warning information between levels, the hierarchical risk early warning sub-results of each level network are fused across levels to obtain hierarchical risk early warning results covering the hierarchical network and core risk transmission paths.

[0107] In this embodiment, the environmental risk early warning system performs cross-level fusion of hierarchical risk early warning sub-results from different network levels based on the redundancy and conflict of early warning information between levels. Redundancy refers to duplicate content in the early warning sub-results of different network levels, while conflict refers to contradictory content in the early warning sub-results of different network levels. The environmental risk early warning system deduplicates redundant information and resolves conflicting information. Then, it performs cross-level fusion of the processed early warning sub-results to obtain a hierarchical risk early warning result covering the network levels and core risk transmission paths. This result includes early warning information from each network level and early warning information from the core risk transmission paths.

[0108] Step 1681: Import the hierarchical risk warning sub-results output by each level of the network into the information input layer of the cross-level fusion processing model according to the level number, and generate a standardized warning information set by the information input layer.

[0109] In this embodiment, the environmental risk early warning system imports the hierarchical risk early warning sub-results output from each network level into the information input layer of the cross-level fusion processing model according to the level number, such as 1, 2, 3, etc. The information input layer performs standardization processing on the imported early warning sub-results by converting them into a unified format, such as including fields for early warning level, early warning content, and early warning time, thereby generating a standardized early warning information set.

[0110] Step 1682: Transmit the standardized early warning information set to the redundancy detection layer of the cross-level fusion processing model. After dimensionless processing of the indicators of each dimension in the early warning information, perform multi-dimensional redundancy identification based on similarity calculation. For any two early warning information, if the similarity in the preset number of dimensions exceeds the preset redundancy threshold, it is determined that there is a redundant relationship. Perform complementary deduplication processing on the early warning information with redundant relationship.

[0111] In this embodiment, the environmental risk early warning system transmits a standardized set of early warning information to the redundancy detection layer of the cross-level fusion processing model. The redundancy detection layer performs dimensionless processing on the indicators of each dimension in the early warning information, converting the indicator values ​​to the range of 0-1. Next, the redundancy detection layer performs multi-dimensional redundancy identification based on similarity calculation. These dimensions include early warning level, early warning content, and early warning time. For any two pieces of early warning information, if the similarity in a preset number of dimensions exceeds a preset redundancy threshold (e.g., the preset number of dimensions is 3, and the preset redundancy threshold is 0.8), a redundancy relationship is determined to exist. Then, the redundancy detection layer performs complementary deduplication processing on the early warning information with redundant relationships, retaining the early warning information containing more comprehensive information and deleting duplicate information.

[0112] Step 1683: Transmit the complementary and deduplicated early warning information to the conflict resolution layer of the cross-level fusion processing model and scan the core fields in the early warning information; if there are numerical or representational differences in the core fields of the same early warning object for different levels of early warning information, they are marked as conflict early warning information groups; for each conflict early warning information group, retrieve the level weights, feature contribution rates, and early warning model accuracy indicators corresponding to each level of the network, and construct a conflict evaluation matrix after normalizing the above indicators; use the normalized level weights as the main weight factor, and calculate the credibility score of each conflict early warning information in combination with the normalized feature contribution rate and model accuracy indicator; prioritize the conflict early warning information according to the credibility score, take the early warning information with the highest credibility score as the benchmark information and take the other conflict information as the adjustment object, and correct the core fields of the adjustment object through the feature calibration algorithm; if the difference in credibility scores of the conflict early warning information groups is less than a preset threshold, perform a cross-level feature negotiation task, and fuse the core features of each conflict information to form a comprehensive early warning conclusion that needs to be confirmed by the feature verification link corresponding to each level of the network.

[0113] In this embodiment, the environmental risk early warning system transmits complementary and deduplicated early warning information to the conflict resolution layer of the cross-level fusion processing model. The conflict resolution layer scans the core fields in the early warning information, including the early warning level and risk degree. If different levels of early warning information have numerical or descriptive differences in the core fields for the same early warning object, such as the first-level network having a high-risk early warning level and the second-level network having a medium-risk early warning level, it is marked as a conflict early warning information group. For each conflict early warning information group, the environmental risk early warning system retrieves the level weights, feature contributions, and early warning model accuracy indicators corresponding to each level of the network. The level weights are 0.56 and 0.48, the feature contributions are 0.7 and 0.6, and the early warning model accuracy indicators are 0.9 and 0.8, respectively. The environmental risk early warning system normalizes the above indicators by converting the indicator values ​​to the range of 0-1. For example, the hierarchical weight is normalized to 0.56 / (0.56+0.48) = 0.56 / 1.04 ≈ 0.538, 0.48 / 1.04 ≈ 0.462; the feature contribution is normalized to 0.7 / (0.7+0.6) = 0.7 / 1.3 ≈ 0.538, 0.6 / 1.3 ≈ 0.462; and the early warning model accuracy is normalized to 0.9 / (0.9+0.8) = 0.9 / 1.7 ≈ 0.529, 0.8 / 1.7 ≈ 0.471. Next, the environmental risk early warning system constructs a conflict assessment matrix, which includes the normalized hierarchical weights, feature contributions, and early warning model accuracy. Then, the environmental risk early warning system uses the normalized hierarchical weights as the primary weighting factor, combined with the normalized feature contribution and model accuracy indicators, to calculate the credibility score of each conflict early warning message. The calculation method is: hierarchical weight × 0.6 + feature contribution × 0.3 + model accuracy indicator × 0.1. The credibility score of the first-level network is calculated as 0.538 × 0.6 + 0.538 × 0.3 + 0.529 × 0.1, and the credibility score of the second-level network is calculated as 0.462 × 0.6 + 0.462 × 0.3 + 0.471 × 0.1. The environmental risk early warning system prioritizes conflict early warning messages based on their credibility scores, with early warning messages from the first-level network having higher priority than those from the second-level network. The early warning message with the highest credibility score is used as the baseline information, i.e., the early warning message from the first-level network, while other conflict information is used as the adjustment object, i.e., the early warning message from the second-level network. The environmental risk early warning system corrects the core fields of the object to be adjusted through a feature calibration algorithm. The calibration method is to adjust the core field values ​​of the object to be adjusted to the core field values ​​of the baseline information.If the difference in credibility scores among conflict warning information groups is less than a preset threshold (e.g., the preset threshold is 0.1, the credibility score of the first-level network is 0.5, and the credibility score of the second-level network is 0.45, with a difference of 0.05 which is less than 0.1), the environmental risk warning system performs a cross-level feature negotiation task, integrating the core features of each conflict information to form a comprehensive warning conclusion. The comprehensive warning conclusion needs to be confirmed through the feature verification link corresponding to each level of the network. The feature verification link refers to the process of verifying the accuracy of the features.

[0114] Step 1684: Transmit the redundancy and conflict-resolved early warning information to the feature fusion layer of the cross-level fusion processing model. Based on the hierarchical relationship of the environmental risk early warning pyramid network, construct an early warning information hierarchical fusion rule library containing same-level fusion rules, cross-level progressive fusion rules, and core path priority fusion rules. According to the priority order in the fusion rule library, feature aggregation is performed on early warning information within the same level to form hierarchical comprehensive early warning fragments. Using the core risk transmission path as a clue, early warning fragments of different levels are progressively fused according to the risk transmission time sequence to establish early warning information association links between levels. The weight of early warning information is adjusted based on the early warning requirements corresponding to the core risk transmission path. During the feature fusion process, the feature vector of the fused early warning information is extracted and compared with the feature vector of the original early warning sub-result of each level network for difference update. The fused early warning information is then structured and organized according to the hierarchical network division and core risk transmission path classification to generate hierarchical risk early warning results containing hierarchical early warning semantics, path early warning focus, risk transmission trend, and comprehensive early warning level.

[0115] In this embodiment, the environmental risk early warning system transmits the redundancy- and conflict-resolved early warning information to the feature fusion layer of the cross-level fusion processing model. Based on the hierarchical relationship of the environmental risk early warning pyramid network, the feature fusion layer constructs a hierarchical fusion rule base for early warning information. This rule base includes same-level fusion rules, cross-level progressive fusion rules, and core path priority fusion rules. Same-level fusion rules refer to fusion of early warning information within the same level; cross-level progressive fusion rules refer to fusion of early warning information from different levels in a progressive order; and core path priority fusion rules refer to prioritizing the fusion of early warning information along the core risk transmission path. The feature fusion layer, according to the priority order in the fusion rule base (from highest to lowest priority: core path priority fusion rules, cross-level progressive fusion rules, and same-level fusion rules), aggregates features of early warning information within the same level to form a hierarchical comprehensive early warning fragment. For example, it aggregates early warning information from the first-level network to form a high-priority hierarchical comprehensive early warning fragment. Next, the feature fusion layer progressively fuses early warning segments from different levels according to the risk transmission sequence, using the core risk transmission path as a guide. The fusion method involves concatenating high-priority and medium-priority comprehensive early warning segments according to the risk transmission sequence. Then, the feature fusion layer establishes inter-level early warning information linkages, referring to the connections between early warning information at different levels. The feature fusion layer adjusts the weights of early warning information based on the early warning requirements corresponding to the core risk transmission path; these requirements indicate the level of importance attached to the early warning information, with higher requirements resulting in higher weights. During feature fusion, the feature fusion layer extracts the feature vectors of the fused early warning information and compares them with the feature vectors of the original early warning sub-results from each level of the network. The comparison method involves calculating the similarity between feature vectors; if the similarity is less than a preset threshold, an update is performed. Finally, the feature fusion layer structures the fused early warning information according to the hierarchical network division and the classification of core risk transmission paths, generating a hierarchical risk early warning result that includes hierarchical early warning semantics, path early warning focus, risk transmission trend, and comprehensive early warning level. Hierarchical early warning semantics refers to the early warning descriptions of different levels, path early warning focus refers to the key content of the early warning of the core risk transmission path, risk transmission trend refers to the direction of risk development, and comprehensive early warning level refers to the overall risk level.

[0116] Optionally, the method further includes: Step 210: Extract the key risk features and risk node attributes corresponding to the core risk transmission path in the hierarchical risk warning results, and align the key risk features with the preset risk evolution benchmark feature library to obtain the feature alignment results.

[0117] In this embodiment, the environmental risk early warning system extracts key risk features and risk node attributes corresponding to the core risk transmission path from the hierarchical risk early warning results. Key risk features include PM2.5 concentration and pH value, while risk node attributes include risk level and risk time. Next, the environmental risk early warning system aligns the key risk features with a preset risk evolution benchmark feature library, which contains benchmark feature values ​​corresponding to different risk levels. The feature alignment method involves adjusting the feature dimensions of the key risk features to match those of the benchmark feature library, resulting in a feature alignment result that includes the matching status between the key risk features and the benchmark features.

[0118] Step 220: Based on the feature alignment results, identify the direction and relative degree of evolutionary deviation between key risk features and benchmark features in the standardized feature space, and determine the dominant deviation factors of risk evolution by combining the hierarchical weights in the normalized risk node attributes, thus obtaining a set of dominant deviation factors.

[0119] In this embodiment, the environmental risk early warning system, based on feature alignment results, identifies the direction and relative degree of evolutionary deviation between key risk features and benchmark features within a standardized feature space. The direction of evolutionary deviation refers to the direction of change of the key risk feature relative to the benchmark feature, such as an increase or decrease; the degree of relative deviation refers to the degree of difference between the key risk feature and the benchmark feature. Next, the environmental risk early warning system normalizes the hierarchical weights in the risk node attributes by converting them to a range of 0-1. Then, the environmental risk early warning system combines the normalized hierarchical weights to determine the dominant deviation factors in risk evolution. The dominant deviation factor refers to the factor with the greatest impact on risk evolution deviation. For example, the hierarchical weight of the PM2.5 concentration feature is 0.6, and the degree of evolutionary deviation is 0.8, thus it is determined as the dominant deviation factor, resulting in a set of dominant deviation factors.

[0120] Step 230 inputs the set of deviation-dominant factors into the risk evolution trend extrapolation model. Through the trend analysis units corresponding to each level of the network in the risk evolution trend extrapolation model, the future evolution state of the core risk transmission path is extrapolated in multiple scenarios, and risk evolution trend sequences under each scenario are generated.

[0121] In this embodiment, the environmental risk early warning system inputs the set of dominant deviation factors into a risk evolution trend projection model. This model includes trend analysis units corresponding to each network level. The trend analysis units perform multi-scenario projections of the future evolution of the core risk transmission path, including normal scenarios, extreme weather scenarios, and scenarios with increased pollution sources. For example, in a normal scenario, the projected future evolution of the core risk transmission path is a slow decrease in PM2.5 concentration; in an extreme weather scenario, the projected future evolution of the core risk transmission path is a rapid increase in PM2.5 concentration. Then, the environmental risk early warning system generates risk evolution trend sequences for each scenario, with the sequences containing the risk evolution states at different time points.

[0122] Step 240: Compare the feature similarity of the risk evolution trend sequences under each scenario, select the core trend sequences with the highest consistency in evolution trends, and generate a supplementary risk evolution warning report by combining the risk level in the hierarchical risk warning results.

[0123] In this embodiment, the environmental risk early warning system compares the feature similarity of risk evolution trend sequences under various scenarios. The feature similarity comparison method involves calculating the feature vector similarity between different sequences. The environmental risk early warning system selects the core trend sequence with the highest consistency in evolution trends, such as the risk evolution trend sequence under normal scenarios. Then, the environmental risk early warning system combines the risk levels in the hierarchical risk early warning results to generate a supplementary risk evolution early warning report. The report includes information such as the content of the core trend sequence and changes in the risk level.

[0124] Step 250: Integrate the supplementary risk evolution warning report with the hierarchical risk warning results to generate risk warning information that includes the current risk status and future evolution trend, and output it to the environmental risk management terminal.

[0125] In this embodiment, the environmental risk early warning system fuses the supplementary risk evolution early warning report with the hierarchical risk early warning results by combining the content of the report with the content of the early warning results. Then, the environmental risk early warning system generates risk early warning information that includes the current risk status and future evolution trend. The current risk status refers to the current environmental risk situation, and the future evolution trend refers to the direction in which the risk will develop. Finally, the environmental risk early warning system outputs the risk early warning information to the environmental risk management terminal, which displays the early warning information and performs corresponding management operations.

[0126] Optionally, the method further includes: Step 310: Collect all process data during the generation of the hierarchical risk warning results and establish a data chain for the warning process.

[0127] In this embodiment, the environmental risk early warning system collects data from the entire process of generating hierarchical risk early warning results. This data includes data from the data input stage, the feature processing stage, the model inference stage, and the result output stage. The environmental risk early warning system then establishes a data chain for the early warning process, connecting the data from each stage in chronological order.

[0128] Step 320: Perform feature analysis on the data chain of the early warning process, extract the quality features and correlation features of the data in each link and perform standardization processing, identify the feature loss nodes and deviation introduction nodes in the data transmission and processing process, and generate a process quality assessment report.

[0129] In this embodiment, the environmental risk early warning system performs feature analysis on the data chain of the early warning process. The analysis method involves analyzing the characteristics of data at each stage of the data chain. Next, the environmental risk early warning system extracts the quality characteristics and correlation characteristics of the data at each stage. Quality characteristics refer to the accuracy and completeness of the data, while correlation characteristics refer to the degree of correlation between data points. Then, the environmental risk early warning system standardizes these characteristics by converting the feature values ​​to the range of 0-1. The environmental risk early warning system identifies feature loss nodes and deviation introduction nodes during data transmission and processing. Feature loss nodes are nodes where features are lost during data transmission, and deviation introduction nodes are nodes where deviations are introduced during data processing. Finally, the environmental risk early warning system generates a process quality assessment report, which includes information on feature loss nodes, deviation introduction nodes, and data quality assessment results.

[0130] Step 330: Based on the standardized quality characteristics in the process quality assessment report, and combined with the normalized hierarchical weights of each level of the network, determine the key data links that have the greatest comprehensive impact on the early warning results, and obtain a list of key data links.

[0131] In this embodiment, the environmental risk early warning system, based on the standardized quality characteristics in the process quality assessment report and combined with the normalized hierarchical weights of each network level, determines the key data links with the greatest comprehensive impact on the early warning results. The comprehensive impact is calculated by multiplying the quality characteristic by the hierarchical weight; for example, if the quality characteristic is 0.8 and the hierarchical weight is 0.6, the comprehensive impact is 0.8 × 0.6 = 0.48. The environmental risk early warning system filters out the key data links with the greatest comprehensive impact, such as the feature extraction stage in the data input phase, to obtain a list of key data links.

[0132] Step 340: Generate a data quality optimization strategy for each link in the list of key data links, and apply the data quality optimization strategy to the parameter adjustment of the environmental risk early warning pyramid network.

[0133] In this embodiment, the environmental risk early warning system generates data quality optimization strategies for each stage in the list of key data stages. For example, the data quality optimization strategy for the feature extraction stage in the data input stage includes increasing the number of samples for feature extraction and optimizing the feature extraction algorithm. Next, the environmental risk early warning system applies the data quality optimization strategies to the parameter adjustment of the environmental risk early warning pyramid network. The parameter adjustment methods include adjusting the network's learning rate and the number of neurons in the hidden layer.

[0134] Step 350: Reprocess the environmental monitoring data to be processed using the optimized parameters to generate optimized hierarchical risk warning results. Compare the warning results before and after optimization using features and output the optimized warning results.

[0135] In this embodiment, the environmental risk early warning system reprocesses the environmental monitoring data to be processed using optimized parameters, and the processing procedure is consistent with the previous processing procedure. Then, the environmental risk early warning system generates optimized hierarchical risk early warning results. Next, the environmental risk early warning system compares the early warning results before and after optimization, comparing features such as the risk level and early warning content of the early warning results. Finally, the environmental risk early warning system outputs optimized early warning results, which include comparison information of the early warning results before and after optimization, and the optimized early warning content.

[0136] Optionally, the method further includes: Step 410: Obtain the real-time environmental perception data at the time the hierarchical risk warning result is generated, and perform feature standardization on the real-time environmental perception data through the input adaptation layer of the environmental risk warning pyramid network to obtain a real-time feature set that matches the feature dimensions of the hierarchical risk warning result.

[0137] In this embodiment, the environmental risk early warning system acquires real-time environmental sensing data at the moment the hierarchical risk early warning results are generated. This real-time environmental sensing data includes real-time PM2.5 concentration data, real-time pH value data, etc. Next, the environmental risk early warning system performs feature standardization on the real-time environmental sensing data through the input adaptation layer of the environmental risk early warning pyramid network. The standardization method involves converting the data into a form that matches the feature dimensions of the hierarchical risk early warning results, thus obtaining a real-time feature set.

[0138] Step 420: Input the real-time feature set into the attention gate for dynamic response analysis, and extract the dynamic difference features in the real-time feature set that deviate from the hierarchical risk warning result in the standardized feature space.

[0139] In this embodiment, the environmental risk early warning system inputs a real-time feature set into an attention gate for dynamic response analysis. This dynamic response analysis is used to analyze the deviation between the real-time feature set and the hierarchical risk early warning results. Next, the environmental risk early warning system extracts dynamic difference features within a standardized feature space that deviate from the hierarchical risk early warning results. These dynamic difference features refer to the differences between the real-time features and the early warning result features.

[0140] Step 430: Integrate the dynamic difference features and the associated features mined by the feature cross model into the risk feature matrix, and use the time-series cascaded prediction model to perform supplementary reasoning to obtain real-time risk transmission adjustment information. Based on the real-time risk transmission adjustment information, correct the risk node attributes and path priorities in the hierarchical risk warning results.

[0141] In this embodiment, the environmental risk early warning system integrates dynamic difference features and their associated features mined through a feature cross-model into a risk feature matrix. This integration involves adding features to the risk feature matrix. Next, the environmental risk early warning system uses a time-series cascaded prediction model to perform supplementary reasoning to obtain real-time risk transmission adjustment information. This real-time risk transmission adjustment information includes adjustments to risk node attributes and path priorities. Then, based on the real-time risk transmission adjustment information, the environmental risk early warning system corrects the risk node attributes and path priorities in the hierarchical risk early warning results. This correction involves adjusting the risk node attributes to new attribute values ​​and adjusting the path priorities to a new priority order.

[0142] Step 440: Extract the feature changes of the corrected early warning results and the hierarchical risk early warning results, combine them with the confidence level of real-time environmental perception data, generate a dynamic update report of the early warning results, and output the corrected early warning results and the dynamic optimization report to the environmental risk management terminal simultaneously.

[0143] In this embodiment, the environmental risk early warning system extracts the characteristic changes of the corrected early warning results and the hierarchical risk early warning results, where the characteristic change refers to the degree of change in the characteristic value. Next, the environmental risk early warning system combines the confidence level of real-time environmental sensing data (where confidence level refers to the accuracy of the real-time data) to generate a dynamic update report of the early warning results. The report includes information on the characteristic change, confidence level information, and the corrected early warning results. Finally, the environmental risk early warning system synchronously outputs the corrected early warning results and the dynamic optimization report to the environmental risk management terminal, which displays the updated early warning information.

[0144] This application's embodiments construct an environmental risk early warning pyramid network by dividing labeled multi-dimensional environmental risk features into a hierarchical structure and determining update weights. Attention gates and feature cross-models are embedded in each level of the network. Environmental monitoring data to be processed is matched to the corresponding level of the network according to type, and key risk features are extracted through attention gates. Then, nonlinear correlations are mined using the feature cross-model to obtain a risk feature matrix that integrates multi-dimensional correlation information. This matrix is ​​input into a time-series cascaded prediction model for progressive inference to generate an environmental risk transmission inference graph. Finally, hierarchical risk early warning results are generated by combining hierarchical weights and risk node correlation assignments. In this way, hierarchical and accurate modeling of environmental risk features is achieved. The synergistic effect of attention mechanisms and feature cross-models effectively captures potential data correlations. The time-series cascaded inference framework ensures dynamic tracking of risk transmission, and the hierarchical early warning results improve the targeting and effectiveness of risk response. This enhances the accuracy, comprehensiveness, and timeliness of environmental risk early warning, solving the problems of insufficient feature utilization, inadequate correlation mining, missing time-series tracking, and ambiguous result presentation in traditional early warning methods, thus improving the accuracy and targeting of environmental risk management and early warning.

[0145] Based on the same inventive concept, embodiments of this application also provide an environmental risk early warning system. See also... Figure 2 As shown, this is a schematic diagram of a possible environmental risk early warning system provided in an embodiment of this application. Figure 2 In the environmental risk early warning system 200, there are a processor 210 and a memory 220. The processor 210 and the memory 220 are connected to each other via a communication bus. The memory 220 stores computer programs that can be executed by the processor 210. By executing the instructions stored in the memory 220, the processor 210 can perform the steps of the aforementioned environmental risk early warning method based on the attention mechanism.

[0146] Based on the same inventive concept, embodiments of this application provide a computer-readable storage medium including a computer program. When the computer program is run on an environmental risk early warning system, it causes the environmental risk early warning system to perform the steps of the aforementioned attention-based environmental risk early warning method. In some possible implementations, various aspects of the attention-based environmental risk early warning method provided in this application can also be implemented as a program product including a computer program. When the program product is run on an environmental risk early warning system, it causes the environmental risk early warning system to perform the steps of the aforementioned attention-based environmental risk early warning method. For example, the environmental risk early warning system can perform actions such as... Figure 1The steps are shown in the diagram. The computer-readable storage medium includes volatile or non-volatile or a combination thereof, and may be removable or non-removable. Examples of computer-readable storage media include, but are not limited to, phase-change random access memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random-access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), flash memory or other memory technologies, CD-ROM, Digital Video Disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium.

[0147] like Figure 3 The diagram shown is a functional block diagram of the environmental risk early warning system provided in this embodiment. The environmental risk early warning system includes an environmental risk early warning device, which includes: The network parameter update module is used to divide the labeled multi-dimensional environmental risk features into a hierarchical structure to obtain multi-level risk features, determine the update weight according to the hierarchical order of the risk features at each level, and adjust the hierarchical parameters of the initial pyramid network using the update weight to obtain the environmental risk early warning pyramid network. A hierarchical network embedding module is used to embed attention gates and feature cross-models into each level of the environmental risk early warning pyramid network; wherein, the attention gate is connected to the input end of the corresponding level network, and the feature cross-model is connected to the output end of the attention gate. The feature response analysis module is used to match the acquired environmental monitoring data to be processed to each level of the environmental risk early warning pyramid network according to the data type, and to perform feature response analysis through the attention gates of each level of the network to extract the key risk features of the environmental monitoring data to be processed. The feature matrix output module is used to receive key risk features output by corresponding attention gates through the feature cross-model of each level network, mine the nonlinear correlation between the key risk features in the preset feature interaction space, and output a risk feature matrix that integrates multi-dimensional correlation information. The risk transmission reasoning module is used to input the risk feature matrix into the time-series cascaded prediction model connected to the environmental risk early warning pyramid network to perform progressive reasoning in the time-series dimension, and obtain the environmental risk transmission reasoning graph. The hierarchical risk warning module is used to associate the hierarchical weights of each level of the network with the risk nodes in the environmental risk transmission inference graph, and generate hierarchical risk warning results based on the assigned risk nodes and the risk transmission paths corresponding to the environmental risk transmission inference graph.

[0148] Accordingly, this application also provides a computer program product, which includes a computer program or instructions that, when executed by a processor, cause the processor to implement the steps in the above method embodiments. It should be understood that each step or combination of steps in the above method flow can be implemented by the computer program or instructions. Furthermore, these computer programs or instructions can be applied to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device, enabling the processor of the general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to function as an apparatus for implementing the corresponding functions in the above method embodiments.

[0149] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0150] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0151] Finally, it should be noted that the above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. An environmental risk early warning method based on an attention mechanism, characterized in that, include: Multi-level risk features are obtained by dividing labeled multi-dimensional environmental risk features into hierarchical structures. Update weights are determined according to the hierarchical order of each level of risk features. The hierarchical parameters of the initial pyramid network are adjusted using the update weights to obtain an environmental risk early warning pyramid network. Attention gates and feature cross-models are embedded in each level of the environmental risk early warning pyramid network; wherein, the attention gate is connected to the input end of the corresponding level network, and the feature cross-model is connected to the output end of the attention gate. The acquired environmental monitoring data to be processed is matched to each level of the environmental risk early warning pyramid network according to the data type. The key risk features of the environmental monitoring data to be processed are extracted by performing feature response analysis through the attention gates of each level of the network. The key risk features are received by the feature cross-model of each level network and the corresponding attention gate output. The nonlinear correlation between the key risk features is mined in the preset feature interaction space, and a risk feature matrix that integrates multi-dimensional correlation information is output. The risk feature matrix is ​​input into the time-series cascaded prediction model connected to the environmental risk early warning pyramid network to perform progressive reasoning in the time-series dimension, thereby obtaining the environmental risk transmission reasoning graph. The hierarchical weights of each network level are associated with the risk nodes in the environmental risk transmission inference graph and assigned values. Based on the assigned risk nodes and the risk transmission path corresponding to the environmental risk transmission inference graph, a hierarchical risk warning result is generated.

2. The method as described in claim 1, characterized in that, The process involves matching the acquired environmental monitoring data to be processed according to data type and assigning it to each level of the environmental risk early warning pyramid network. Feature response analysis is then performed using attention gates at each level of the network to extract key risk features from the environmental monitoring data, including: Acquire multi-source environmental monitoring data to be processed, classify the multi-source environmental monitoring data to be processed according to data representation dimensions and data acquisition scenarios, and obtain the environmental monitoring data to be processed after data type labeling. Based on the risk characteristic hierarchical attributes corresponding to each level of the environmental risk early warning pyramid network, a matching mapping relationship between data types and hierarchical networks is established, and the environmental monitoring data to be processed after being labeled with data types is allocated to the corresponding hierarchical networks according to the matching mapping relationship. The environmental monitoring data to be processed received by each layer of the network is initialized with intra-layer feature response. The layer parameters of the corresponding layer of the network are loaded through the attention gate. Feature sensitivity analysis is performed on the initialized environmental monitoring data to be processed to screen out the initial response features that have a response correlation with risk warning. The initial response features are compared with the multi-level risk features of the corresponding hierarchical network across feature dimensions to identify the overlapping and different feature dimensions between the initial response features and the multi-level risk features. Based on the response intensity of the overlapping feature dimensions after standardization, the different feature dimensions are subjected to dimensionless feature enhancement processing. The enhanced differential feature dimension and the overlapping feature dimension are fused through the attention gate to generate a hierarchical response feature set, and the feature components used for risk indication are extracted from the hierarchical response feature set. By performing cross-level feature calibration and dimensionality bias correction on the feature components output by each level of the network, key risk features of the environmental monitoring data to be processed, covering various data types and network levels, are obtained.

3. The method as described in claim 1 or 2, characterized in that, The key risk features received by the feature cross-model of each layer of the network and the corresponding attention gate output are then mined within a preset feature interaction space to uncover the nonlinear correlations between the key risk features, and a risk feature matrix fused with multi-dimensional correlation information is output, including: The key risk features output by the corresponding attention gate are received through the feature cross-model of each level network. The preset feature interaction space is initialized based on the hierarchical attributes of the corresponding level network, so that the dimension of the preset feature interaction space is adapted to the feature dimension of the key risk features. The standardized key risk features are mapped to the preset feature interaction space to construct a spatial distribution representation of the key risk features, and the initial correlation strength between key risk features is determined by spatial distance calculation. Based on the initial association strength, key risk features are grouped into multiple initial association feature groups. Nonlinear association mining is performed on the key risk features within each initial association feature group, and in-group association features are generated through feature interaction operations. Cross-group interaction is performed on the intra-group association features of each initial association feature group to explore the potential non-linear associations between different association feature groups, generate cross-group association features, and calculate the association confidence between cross-group association features and intra-group association features. Based on the association credibility, the intra-group association features and cross-group association features are filtered, and the feature components that meet the association credibility requirements are extracted. The filtered feature components are then sorted and organized according to the feature dimension. When the generated associated features no longer add valid associated information, all the obtained associated features are combined according to the preset matrix arrangement rules to output a risk feature matrix that integrates multi-dimensional associated information.

4. The method as described in claim 1, characterized in that, The step of inputting the risk feature matrix into a time-series cascaded prediction model connected to the environmental risk early warning pyramid network for progressive reasoning along the time-series dimension yields an environmental risk transmission reasoning graph, including: The risk feature matrix is ​​input into the time series cascaded prediction model. The risk feature matrix is ​​then subjected to time series feature hierarchical decomposition through the feature decomposition sub-model of the time series cascaded prediction model to obtain a set of feature components for different time series stages. According to the chronological order of the time series stages, the feature components of the previous time series stage are used as input, and the first-level progressive reasoning is performed through the reasoning sub-model of the time series cascaded prediction model to generate the initial risk transmission path and corresponding risk nodes. The initial risk transmission path and corresponding risk nodes output by the first-level progressive inference are fused with the feature components of the current time series stage. The second-level progressive inference is then performed through the inference sub-model to update the extension direction of the risk transmission path and the attribute information of the risk nodes. By sequentially completing the progressive reasoning of all time-series stages, a preliminary risk transmission reasoning graph containing the risk transmission paths and risk nodes of the entire time-series stages is obtained. Path detection is performed on the preliminary risk transmission reasoning graph to identify abnormal transmission paths that do not conform to the time-series logic. Based on the characteristic components of the current time series stage and the effective inference results of the previous time series stage, the abnormal propagation path is corrected and adjusted, and the risk node attributes and path extension direction corresponding to the abnormal propagation path are re-determined. The modified transmission path is integrated with the effective transmission path to construct an environmental risk transmission reasoning graph that contains complete temporal correlations and consistent path logic. The environmental risk transmission reasoning graph is used to characterize the temporal transmission relationship and path correlation characteristics of each risk node.

5. The method as described in claim 4, characterized in that, The risk feature matrix is ​​input into a time-series cascaded prediction model, and the risk feature matrix is ​​subjected to time-series feature hierarchical decomposition through the feature decomposition sub-model of the time-series cascaded prediction model to obtain a set of feature components for different time-series stages, including: The risk feature matrix is ​​received using the feature decomposition sub-model of the time series cascaded prediction model. Based on the time series evolution law of risk transmission, a time series stage division rule is set, which is adapted to the natural evolution cycle of environmental risks. According to the time series segmentation rules, the time dimension corresponding to the risk feature matrix is ​​segmented to obtain multiple continuous and non-overlapping time series segments. Each time series segment corresponds to a time interval and a subset of risk features within that time interval. For each time series stage, the risk feature subset is normalized in terms of feature dimension. Then, through the time series feature extraction unit of the feature decomposition sub-model, time series correlation feature mining is performed on each normalized risk feature subset to identify the time evolution trend of each feature component in the corresponding time series stage and its time series correlation with other feature components. Based on the temporal evolution trend and temporal correlation, feature stratification is performed on each risk feature subset, and feature components with the same evolution trend and correlation are grouped into the same feature layer to form a multi-layer feature set for each temporal stage. Effective feature layers reflecting the temporal pattern of risk transmission are extracted from the multi-layer feature sets of each time series stage. The effective feature layers of all time series stages are integrated into feature component sets of different time series stages. The temporal continuity of the feature component sets of each time series stage is checked iteratively. If there is a temporal correlation break in the feature component sets of adjacent time series stages, the time series stage division rules and feature layering standards are readjusted until the feature component sets of all time series stages form a continuous temporal feature chain.

6. The method according to any one of claims 1-2 and 4-5, characterized in that, The step of associating and assigning values ​​to the hierarchical weights of each network level with the risk nodes in the environmental risk transmission inference graph, and generating hierarchical risk warning results based on the assigned risk nodes and the risk transmission paths corresponding to the environmental risk transmission inference graph, includes: Extract the hierarchical weights of each level of the environmental risk early warning pyramid network. The hierarchical weights reflect the importance and characteristic contribution of the corresponding level network in risk early warning. Traverse all risk nodes in the environmental risk transmission reasoning graph, identify the risk feature source hierarchy network corresponding to each risk node, and establish the source association relationship between the risk node and the hierarchy network; Based on the source association relationship, the hierarchical weights of the corresponding hierarchical network are assigned to the risk nodes to form risk nodes with hierarchical weight attributes. At the same time, for risk nodes that are associated with multiple source hierarchical networks, the hierarchical weights are weighted and aggregated based on the identified and quantified association confidence of each source hierarchical network to calculate the aggregated weight value of each risk node. The path length and path association strength of the risk transmission paths in the environmental risk transmission reasoning graph are analyzed. The path association strength is the average level of the aggregated weight values ​​of all risk nodes on the path. The priority of each risk transmission path is determined based on the path length and path association strength. The risk transmission paths are sorted in descending order of priority, and the core risk transmission paths that meet the preset priority requirements are marked. The risk nodes on the core risk transmission paths are marked as key risk nodes. The key risk nodes are aggregated by using their hierarchical weight attributes, temporal transmission characteristics, and path association information to generate aggregated features of key risk nodes. Based on the aggregation characteristics of key risk nodes and the priority ranking of core risk transmission paths, hierarchical risk warning sub-results are generated according to the hierarchical division rules of the hierarchical network. Based on the redundancy and conflict of early warning information between levels, the hierarchical risk early warning sub-results of each level network are fused across levels to obtain hierarchical risk early warning results that cover the hierarchical network and the core risk transmission path.

7. The method as described in claim 6, characterized in that, The process of traversing all risk nodes in the environmental risk transmission reasoning graph, identifying the risk feature source hierarchy network corresponding to each risk node, and establishing the source association relationship between risk nodes and the hierarchy network includes: In response to the risk node tracing and analysis task, the topology data of the environmental risk transmission reasoning graph is loaded, and the unique identification information and feature description fields of each risk node in the topology data are parsed. The feature description fields of the risk nodes are transformed into standardized risk node feature vectors. Based on the hierarchical division log of the environmental risk early warning pyramid network, the feature output port information and feature transmission link identifiers of each level network are extracted and a correspondence index table between the hierarchical network and the feature transmission link is generated. Using the standardized risk node feature vector as the retrieval basis, a multi-dimensional fuzzy matching operation is performed in the corresponding relationship index table to determine the candidate layer network set that has potential association with the risk node feature vector. At the same time, the matching similarity value between each candidate layer network and the risk node is recorded. For each layer network in the candidate layer network set, the historical feature output record of the layer network is called, and the key risk feature matrix fragment output by the layer network in the corresponding time series stage is extracted. The key risk feature matrix fragment is then aligned with the feature dimension of the risk node feature vector. The alignment of feature data is deeply mined using a feature attribution analysis algorithm. This identifies the source correspondence of each feature component in the feature vector of a risk node in the feature matrix fragment of the candidate hierarchical network. Feature component pairs with direct inheritance relationships and indirect associations are marked. Based on the inheritance relationship strength of the feature components, the association confidence between the candidate hierarchical network and the risk node is calculated, resulting in a quantitative association confidence assessment. Based on the accuracy label of environmental risk early warning and the feature contribution benchmark value of the hierarchical network, a correlation confidence threshold is set. Candidate hierarchical networks with a correlation confidence higher than the correlation confidence threshold are identified as the source hierarchical networks of the risk node, and hierarchical networks with a correlation confidence that does not meet the threshold are removed and the reasons for removal are recorded. For risk nodes with multiple source hierarchical networks, the feature output time sequence of each source hierarchical network is obtained, and it is determined whether there is a causal relationship between the feature data of different source hierarchical networks and the formation of the risk node. A multi-source hierarchical network contribution sequence of the risk node is constructed. Based on the contribution sequence of the multi-source hierarchical network, with risk nodes as core nodes and source hierarchical networks as associated nodes, a source association graph between risk nodes and hierarchical networks is constructed by using weighted edges to represent the association strength and time sequence relationship between the two. Loop detection is performed on the source association graph. If anomalies of cyclic association between hierarchical networks are found, the association relationship is corrected by feature transmission time sequence priority.

8. The method as described in claim 6, characterized in that, Based on the redundancy and conflict of early warning information between levels, the hierarchical risk early warning sub-results of each level network are fused across levels to obtain hierarchical risk early warning results covering the hierarchical network and core risk transmission paths, including: The hierarchical risk warning sub-results output by each level of the network are imported into the information input layer of the cross-level fusion processing model according to the level number, and the information input layer generates a standardized warning information set. The standardized early warning information set is transmitted to the redundancy detection layer of the cross-level fusion processing model. After the dimensionless processing of the indicators in each dimension of the early warning information, multi-dimensional redundancy identification based on similarity calculation is performed. For any two early warning information, if the similarity in the preset number of dimensions exceeds the preset redundancy threshold, it is determined that there is a redundant relationship. The early warning information with redundant relationship is subjected to complementary deduplication processing. The complementary and deduplicated early warning information is transmitted to the conflict resolution layer of the cross-level fusion processing model, and the core fields in the early warning information are scanned. If there are numerical or descriptive differences in the core fields of the same early warning object for different levels of early warning information, they are marked as conflict early warning information groups. For each conflict early warning information group, the level weights, feature contributions, and early warning model accuracy indicators of each level network are retrieved, and the above indicators are normalized to construct a conflict evaluation matrix. The credibility score of each conflict early warning information is calculated by using the normalized level weights as the main weight factor, combined with the normalized feature contributions and model accuracy indicators. The conflict early warning information is prioritized according to the credibility score, and the early warning information with the highest credibility score is used as the benchmark information and other conflict information is used as the adjustment object. The core fields of the adjustment object are corrected by the feature calibration algorithm. If the difference in credibility scores of the conflict early warning information groups is less than a preset threshold, a cross-level feature negotiation task is performed to fuse the core features of each conflict information to form a comprehensive early warning conclusion that needs to be confirmed by the feature verification link corresponding to each level network. After redundancy and conflict resolution, the early warning information is transmitted to the feature fusion layer of the cross-level fusion processing model. Based on the hierarchical relationship of the environmental risk early warning pyramid network, a hierarchical fusion rule library for early warning information is constructed, which includes same-level fusion rules, cross-level progressive fusion rules, and core path priority fusion rules. According to the priority order in the fusion rule library, early warning information within the same level is feature-aggregated to form hierarchical comprehensive early warning fragments. Using the core risk transmission path as a clue, early warning fragments from different levels are progressively fused according to the risk transmission sequence to establish a link between early warning information at different levels. The weight of early warning information is adjusted based on the early warning requirements corresponding to the core risk transmission path. During the feature fusion process, the feature vector of the fused early warning information is extracted and compared with the feature vector of the original early warning sub-results of each level network for difference updating. The fused early warning information is then structured and organized according to the hierarchical network division and core risk transmission path classification to generate a hierarchical risk early warning result that includes hierarchical early warning semantics, path early warning focus, risk transmission trend, and comprehensive early warning level.

9. An environmental risk early warning system, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program that, when executed by the processor, causes the processor to perform the steps of the environmental risk warning method based on the attention mechanism according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a computer program that, when run on the environmental risk early warning system, causes the environmental risk early warning system to perform the steps of the attention-based environmental risk early warning method according to any one of claims 1 to 8.

Images