An environmental pollution inspection method and system

Abstract

The application discloses an environmental pollution inspection method and system, and belongs to the technical field of environmental monitoring. The technical scheme points of the application comprise the following steps: obtaining initial features corresponding to each node, and correcting the initial features according to a decay coefficient to obtain corrected features; correcting the weights of edges in a graph neural network according to the corrected features; calculating the branch degree and neighbor distance entropy corresponding to each node; updating the order of a convolution kernel in the graph neural network according to the branch degree and the neighbor distance entropy to obtain the order of the convolution kernel corresponding to each node; determining waste gas abnormal coordinates according to the order of the convolution kernel corresponding to each node; and formulating an inspection scheme according to the waste gas abnormal coordinates. The application dynamically updates the convolution kernel corresponding to each node in the graph neural network model through the node branch degree and the neighbor distance entropy, so that the convolution kernel can adapt to the topological structure corresponding to each node, and the model weight is subjected to a probability distribution according to Bayesian inference, thereby enhancing the generalization ability of the graph neural network model under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring technology, and more specifically to an environmental pollution inspection method and system. Background Technology

[0002] Industrial parks contain waste gas-related equipment (such as waste gas treatment towers, fans, various waste gas pipelines, valves, and storage tanks) that are responsible for the transmission and treatment of industrial waste gases (such as sulfur-containing and nitrogen-containing waste gases). Leaks or malfunctions in these facilities can easily lead to air pollution and impact the surrounding production and living environment. With increasing environmental protection requirements and demands for refined management, accurately locating waste gas anomalies (leaks, equipment malfunctions, etc.) has become a core requirement for waste gas anomaly inspections. However, industrial parks' waste gas-related equipment and pipeline networks are characterized by complex topologies, diverse waste gas media, and strong environmental interference. Traditional inspection methods struggle to balance location accuracy and reliability, posing challenges to the timeliness and accuracy of waste gas anomaly management. Therefore, existing technologies have shortcomings. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the present invention aims to provide an environmental pollution inspection method and system. By updating the convolution kernel corresponding to each node in the graph neural network model through the branch degree and neighbor distance entropy of each node in the exhaust gas system, the convolution kernel can be adapted to the topology of each node, thereby accurately determining the exhaust gas anomaly points and executing anomaly handling steps to avoid further environmental pollution.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] This invention provides an environmental pollution inspection method, comprising:

[0006] Obtain the initial features corresponding to each node, and correct the initial features according to the attenuation coefficient to obtain the corrected features;

[0007] The weights of edges in the graph neural network are corrected based on the corrected features;

[0008] Calculate the branching degree and neighbor distance entropy for each node;

[0009] The order of the convolution kernel in the graph neural network is updated based on the branch degree and the neighbor distance entropy to obtain the convolution kernel order corresponding to each node;

[0010] The coordinates of the exhaust gas anomaly are determined based on the convolution kernel order corresponding to each node.

[0011] An inspection plan is formulated based on the coordinates of the abnormal exhaust gas.

[0012] As a further improvement of the present invention, the initial feature includes concentration and pressure, the correction feature includes correction concentration and correction rate, and the step of correcting the initial feature according to the attenuation coefficient to obtain the correction feature includes:

[0013] The attenuation coefficient is obtained based on the digital twin model;

[0014] The corrected concentration is obtained based on the attenuation coefficient and the exhaust gas concentration;

[0015] The corrected speed is obtained based on the pressure and the medium coefficient.

[0016] As a further improvement of the present invention, the calculation of the branch degree and neighbor distance entropy corresponding to each node includes:

[0017] For each node, the branch degree is obtained based on the number of its directly adjacent nodes;

[0018] For each node, determine its distance to each neighbor node and the set of neighbor nodes;

[0019] The neighbor distance entropy is obtained based on the distance to each neighbor node and the set of neighbor nodes.

[0020] As a further improvement of the present invention, the step of updating the order of the convolution kernel in the graph neural network based on the branch degree and the neighbor distance entropy to obtain the convolution kernel order corresponding to each node includes:

[0021] For each node, if the corresponding branch degree is less than the preset value, the corresponding convolution kernel order is obtained based on the corresponding neighbor distance entropy and activation function.

[0022] If the corresponding branch degree is greater than or equal to the preset value, the corresponding convolution kernel order is obtained based on the corresponding branch degree and logarithmic function.

[0023] As a further improvement of the present invention, the step of determining the abnormal coordinates of exhaust gas based on the convolution kernel order corresponding to each node includes:

[0024] For each node, perform an iterative operation, which includes generating the current weight based on the weight probability distribution, obtaining dynamic aggregation features based on the current weight and the convolution kernel order, obtaining attention weights based on the dynamic aggregation features, obtaining the coordinates corresponding to the current weight based on the attention weights, until a preset termination condition is met, and outputting the coordinates corresponding to each weight.

[0025] The coordinates of the exhaust gas anomaly are obtained based on the coordinates corresponding to each weight.

[0026] As a further improvement of the present invention, the step of obtaining the dynamic aggregation feature based on the current weight and the convolution kernel order includes:

[0027] Based on the convolution kernel order, determine the multi-level neighbor nodes corresponding to the node;

[0028] For each neighbor node, the corresponding comprehensive weight is obtained based on the distance decay weight and the weight of the edge in the graph neural network;

[0029] The dynamic aggregation feature is obtained based on the current weight, the comprehensive weight, and the modified feature.

[0030] As a further improvement of the present invention, the weighted probability distribution is obtained based on Bayesian inference.

[0031] As a further improvement of the present invention, the weighted probability distribution is obtained based on Bayesian inference, including:

[0032] Define prior distribution and variational distribution;

[0033] For each observation data point, the corresponding loss function value is obtained based on the current variational distribution;

[0034] The variational distribution is updated based on the loss function value until the variational distribution converges, thus obtaining the weight probability distribution.

[0035] As a further improvement of the present invention, the step of obtaining the corresponding loss function value according to the current variational distribution for each observation data includes:

[0036] Multiple resampling is performed based on the current variational distribution to obtain multiple weight values;

[0037] The probability of the observed data is obtained based on each weight value;

[0038] The expected value is obtained based on the probability.

[0039] Calculate the KL divergence between the current variational distribution and the prior distribution;

[0040] The loss function value is obtained based on the expected value and the KL divergence.

[0041] This invention provides an environmental pollution inspection system for implementing the aforementioned environmental pollution inspection method, comprising:

[0042] The correction module is used to obtain the initial features corresponding to each node, and correct the initial features according to the attenuation formula to obtain the corrected features;

[0043] The first update module is used to correct the weights of edges in the graph neural network based on the corrected features;

[0044] The calculation module is used to calculate the branching degree and neighbor distance entropy for each node;

[0045] The second update module is used to update the order of the convolution kernel in the graph neural network according to the branch degree and the neighbor distance entropy, so as to obtain the convolution kernel order corresponding to each node;

[0046] The positioning module is used to determine the coordinates of the exhaust gas anomaly based on the convolution kernel order corresponding to each node.

[0047] The strategy module is used to formulate inspection plans based on the coordinates of the abnormal exhaust gas.

[0048] This invention updates the convolution kernel of each node in the graph neural network model by using the branch degree and neighbor distance entropy of each node in the exhaust gas system. This allows the convolution kernel to adapt to the topology of each node, thereby accurately determining the location of exhaust gas anomalies and generating reasonable inspection plans to prevent further pollution. Furthermore, by using Bayesian inference to make the model weights follow a probability distribution, the generalization ability of the graph neural network model under complex working conditions is enhanced, solving the problem that existing technologies are difficult to adapt to complex environments and complex topologies. Attached Figure Description

[0049] Figure 1 This is a schematic diagram illustrating the steps of an environmental pollution inspection method according to the present invention;

[0050] Figure 2 This is a schematic diagram of the node topology;

[0051] Figure 3 A schematic diagram illustrating the steps for determining the coordinates of abnormal exhaust gas;

[0052] Figure 4 A schematic diagram illustrating the steps involved in determining the weighted probability distribution. Detailed Implementation

[0053] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof.

[0054] The term "and / or" in the following text is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0055] like Figure 1 As shown in the embodiment of this application, an environmental pollution inspection method is provided, including:

[0056] Obtain the initial features corresponding to each node, and correct the initial features according to the attenuation coefficient to obtain the corrected features;

[0057] The weights of edges in the graph neural network are corrected based on the corrected features;

[0058] Calculate the branching degree and neighbor distance entropy for each node;

[0059] The order of the convolution kernel in the graph neural network is updated based on the branch degree and the neighbor distance entropy to obtain the convolution kernel order corresponding to each node;

[0060] The coordinates of the exhaust gas anomaly are determined based on the convolution kernel order corresponding to each node.

[0061] Develop an inspection plan based on the coordinates of abnormal exhaust gas.

[0062] The waste gas system of the enterprise park is composed of multiple discrete functional units, such as fans, treatment towers, and pipelines. If the entire waste gas system is analyzed directly, the amount of data is large and it is difficult to accurately locate the abnormal area. Therefore, this embodiment divides the waste gas system into nodes, making each node an independent analysis unit, thereby reducing the difficulty of data processing and improving the accuracy of anomaly location. Nodes include treatment nodes (fans, treatment towers, etc.) and connection nodes (pipelines) in the waste gas system. Edges represent the physical connectivity between nodes. For each node, its corresponding initial features include concentration, pressure, and velocity, and the correction features include corrected concentration, corrected pressure, and corrected velocity. The initial features can be collected by sensors. Based on the nodes and edges, the initial topology of the waste gas system is obtained. Furthermore, in this embodiment, the abnormality of exhaust gas mainly refers to exhaust gas leakage and equipment failure. Exhaust gas leakage means that the exhaust gas is not transmitted according to the designed path, but leaks directly into the environment through gaps or damage in equipment or pipelines. Equipment failure means that the exhaust gas cannot be transmitted normally and there is a blockage. When there is an exhaust gas leakage, the concentration of exhaust gas around the leakage point increases abnormally, and the pressure will drop locally due to the leakage. For example, the pressure of the pipeline leakage section is lower than the normal transmission pressure, and the velocity downstream of the leakage point decreases. When there is an equipment failure, the concentration and pressure of the node increase, and the velocity drops sharply. Therefore, this embodiment selects concentration, pressure and velocity as initial features to accurately identify abnormal coordinates.

[0063] In the initial topology, the weights and directions of edges are typically determined based on the time difference of exhaust gas transport or the degree of concentration decay. For example, when constructing the initial topology, if the nodes... The abnormal exhaust gas concentration was detected earlier than the node. And nodes The detected concentration is greater than the node This indicates that the direction of exhaust gas transmission is node-based. To node That is, the direction of the edge between these two nodes is the node. →Node Whether the exhaust gas concentration is abnormal can be determined according to a preset standard. This embodiment does not impose any restrictions on this. Based on the exhaust gas concentrations monitored by these two nodes and the distance between the two nodes, the weight of the edge between these two nodes in the initial topology can be determined. for:

[0064] ;

[0065] in, Represents a node The monitored concentration of exhaust gas, Represents a node The monitored concentration of exhaust gas, Represents a node and nodes The distance between them, and the weight used to measure the exhaust gas from the node Transmit to node The greater the probability, the higher the weight value, indicating a higher probability, as the distance between nodes is smaller. However, this embodiment found that determining the weight solely by concentration may be affected by environmental factors, such as if the nodes... The presence of other gaseous interference sources nearby can easily lead to node interference. The exhaust gas concentration was excessively weakened, leading to an incorrect estimation of the edge weights and directions. Therefore, the weights need further correction. for:

[0066] ;

[0067] ;

[0068] in, The reliability factor is determined by the dominant airflow direction between two nodes. For example, if the dominant airflow direction between two nodes is the node... To node And in the initial topology, the direction of the edges is the node. To node This indicates that the concentration direction is consistent with the physical direction. The corrected weights at this point It can be abbreviated as This indicates that the weights obtained from the signal are relatively reliable and do not weaken the weights. If the dominant airflow direction is inconsistent with the direction of the edges in the initial topology, it means that the concentration direction contradicts the physical direction. However, 0.5 is merely an example, used to mitigate the impact of erroneous weights and prevent the graph neural network model from learning incorrect patterns. Those skilled in the art can set other values. The dominant airflow direction refers to the airflow direction during pipeline design or the dominant direction of ambient wind, etc.

[0069] This is the airflow correlation coefficient between two nodes. Its specific value is equal to the component of the airflow velocity along the line connecting the nodes. For example, if the airflow velocity is 5 m / s and the angle between the line connecting the nodes and the airflow is 30°, then the component is 5 × cos30° ≈ 4.33 m / s. Its value is positive in the direction of gravity and negative in the direction of gravity. For nodes To node The slope support factor is used to quantify the degree to which gravity supports exhaust gas transport. Used to control slope The extent of the impact can be determined through historical data. For example, historical cases of exhaust gas leaks within the industrial park can be collected, and tests can be conducted at different levels. The next step is to determine whether the corrected edge weights better reflect the actual direction of exhaust gas transport, and finally select the one with the smallest error. , For nodes To node The slope support factor, under gravity conditions, such as at high nodes. Lowest node Because gravity promotes the transport of exhaust gases, Greater than 1, in the case of reverse gravity, such as at lower nodes. Highest node Gravity will hinder the transmission of exhaust gases, therefore If the value is less than 1, then when calculating the corrected weights, take... and The maximum value in the middle, under gravity and Increase edge weights when there is no gravity and When there is a contradiction, reduce the edge weight to make the edge weight more consistent with the transmission characteristics of exhaust gas. This represents the influence factor of the gas flow, determined by the gas transmission medium between two nodes. For example, if the gas is transported between these two nodes via a pipeline, the gas flow depends on the pressure difference. Therefore, a pressure difference can be set in this case. If the exhaust gas between two nodes is transmitted through the open air, such as the transmission of exhaust gas between a breathing valve and an intake hood, the transmission of exhaust gas is greatly affected by the ambient airflow. Therefore, it is possible to set... The value is a small value, such as 0.5, but this embodiment does not limit the specific value. This is a physical direction amplification factor used to enhance the side weights when the exhaust gas transmission direction is consistent with the physical direction. The value should be relatively large, and... Similarly, this can also be determined using historical data. In this embodiment, the reliability factor, medium airflow influence factor, and physical direction amplification factor range from 0 to 1, are dimensionless parameters, and do not require normalization. The airflow correlation coefficient, however, is measured in m / s, which differs significantly in dimension from other parameters and requires normalization.

[0070] This embodiment first distinguishes the role of airflow in different exhaust gas transport scenarios by using the medium airflow influence factor, ensuring that the effect of airflow on exhaust gas transport in different scenarios is accurately characterized. Then, it uses the airflow support factor based on the direction of gravity and the magnitude of airflow-related parameters, and compares it with the airflow sensitivity coefficient. The calculation quantifies the degree of support of gravity for exhaust gas transmission, conforming to the physical law of exhaust gas diffusion along the direction of gravity. The reliability factor and physical direction amplification factor determine whether the exhaust gas transmission direction is consistent with the physical direction. When there is a contradiction, the reliability of the weight is reduced, and when they are consistent, the weight ratio of the physical direction is amplified. This makes the corrected weight retain the data characteristics of exhaust gas transmission such as timing and concentration in the sensor signal, while also forcibly constraining the exhaust gas propagation to follow the physical laws such as gravity and equipment operation logic. This allows the graph neural network to improve the accuracy and physical interpretability of anomaly location in complex exhaust gas systems by using edge weights that are more consistent with the actual physical process of exhaust gas transmission when aggregating neighbor features.

[0071] After correcting the weights, the order of the convolution kernel in the graph neural network is updated according to the branch degree and neighbor distance entropy of each node, so as to obtain the convolution kernel order of each node. Finally, based on the convolution kernel order of each node and the corrected features, the location of the exhaust gas anomaly is determined and an inspection plan is formulated.

[0072] This embodiment updates the convolution kernel of each node in the graph neural network model by using the inspection data of the exhaust gas system and the branch degree and neighbor distance entropy of each node in the system. This allows the convolution kernel to adapt to the topology of each node, thereby accurately identifying abnormal points in the exhaust gas system and executing abnormal handling steps to avoid further air pollution. Furthermore, by using Bayesian inference to make the model weights follow a probability distribution, the generalization ability of the graph neural network model under complex working conditions is enhanced, solving the problem that existing technologies are difficult to adapt to complex environments and complex topologies.

[0073] Furthermore, this embodiment provides a step for correcting the initial characteristics based on the attenuation coefficient to obtain the corrected characteristics, including:

[0074] The attenuation coefficient is obtained based on the digital twin model;

[0075] The corrected concentration is obtained based on the attenuation coefficient and the exhaust gas concentration;

[0076] The corrected airflow velocity is obtained based on the pressure and medium coefficient.

[0077] For example, in a digital twin model-simulated exhaust gas system, virtual leak points and faulty devices can be set up at different locations and of different types of nodes, and virtual sensors can be deployed to monitor the transmission of exhaust gas. The attenuation coefficient of exhaust gas concentration in different environments (such as different temperatures, humidity, and airflow environments) can be obtained through data fitting.

[0078] Specifically, after obtaining the attenuation coefficients corresponding to different environments, each node selects an appropriate attenuation coefficient based on its own environment, and uses the ratio of the collected exhaust gas concentration to the attenuation coefficient as the correction concentration. Ideally, the diffusion rate of the medium corresponding to each node... for:

[0079] ;

[0080] in, This represents the difference between the corrected pressure and the average historical normal operating pressure at that node. The density of the waste gas medium is determined based on the medium transported by the equipment corresponding to that node in the waste gas system. However, different equipment (such as fans, heat exchangers, spray towers, etc.) have different internal structures (such as impeller shape, heat exchange tube arrangement, spray device design, etc.), resulting in varying resistance to waste gas flow. Furthermore, different waste gas media have different viscosities, all of which affect the velocity of the waste gas medium. Therefore, this embodiment condenses the combined influence of equipment structure and medium into a single medium coefficient. This makes the velocity formula more closely match the actual physical laws, resulting in a corrected airflow velocity. for:

[0081] ;

[0082] The medium coefficient varies depending on the combination of different exhaust gas media and corresponding equipment. The medium coefficient can be fitted by simulation experiments. For example, different types of media can be injected into equipment with different structures (simulating actual fans, heat exchangers, etc.) in the experimental field, leaks can be artificially created, and the pressure difference and measured diffusion rate can be recorded. By fitting the above formula through multiple sets of experiments, the medium coefficient under different equipment structures and different combinations of exhaust gas media can be obtained.

[0083] Specifically, pressure can be corrected based on the ideal gas law using temperature. The specific formula is:

[0084] ;

[0085] in, This represents the actual measured pressure. This indicates the actual measured temperature. This indicates the standard temperature (293K).

[0086] This embodiment eliminates environmental interference by modifying the initial features to obtain the true characteristics of the exhaust gas. This makes the features input into the graph neural network conform to the physical laws of exhaust gas transmission, thereby enabling the model to learn and reason based on the correct physical relationships and improving the accuracy of anomaly location in the park's exhaust gas system.

[0087] Furthermore, this embodiment provides a step for calculating the branching degree and neighbor distance entropy corresponding to each node, including:

[0088] For each node, the branch degree is obtained based on the number of its directly adjacent nodes;

[0089] For each node, determine its distance to each neighbor node and the set of neighbor nodes;

[0090] The neighbor distance entropy is obtained based on the distance to each neighbor node and the set of neighbor nodes.

[0091] Specifically, for each node, the number of its directly adjacent nodes is its branch degree, such as... Figure 2 In the middle node, node A has a branch degree of 2, and node B has a branch degree of 3. Directly adjacent nodes are connected by directed edges (e.g., pipe connections, equipment-to-pipe connections). For example, nodes... The corresponding branch degree is denoted as Its corresponding neighbor distance entropy for:

[0092] ;

[0093] ;

[0094] in, Represents nodes A set of directly adjacent nodes, where each node in the set is denoted as a node. For each corresponding neighbor node, Represents a node and The first in The distance between nodes, such as the connection length between equipment and pipeline, and the laying length between pipelines, etc. Represents a node and The first in The distance between nodes and Used as an index for traversal. Each neighbor node in the network, For nodes To the The distance between nodes accounts for the distance between nodes. The proportion of the total distance to all neighboring nodes.

[0095] This embodiment constructs the neighbor distance entropy based on the basic formula of entropy in information theory, and then assigns nodes... To the The distance distribution of each node can be compared to the distance concentration of the neighbors. The probability of this event at each node is determined by traversing... The neighbor distance entropy obtained for each neighbor node in the algorithm quantifies the uniformity of the distance distribution among each neighbor node. The higher the entropy, the better the distance between each neighbor node and the node. The more uneven the distance distribution between nodes, the easier it is to determine the convolution kernel corresponding to nodes with different topologies by calculating the neighbor distance entropy and branching degree. This allows the graph neural network model to more accurately aggregate effective features on nodes with different topologies, thereby improving the accuracy of exhaust gas anomaly localization.

[0096] Furthermore, this embodiment provides a step for updating the order of the convolution kernel in a graph neural network based on the branch degree and neighbor distance entropy to obtain the convolution kernel order corresponding to each node, including:

[0097] For each node, if the corresponding branch degree is less than the preset value, the corresponding convolution kernel order is obtained based on the corresponding neighbor distance entropy and activation function.

[0098] If the corresponding branch degree is greater than or equal to the preset value, the corresponding convolution kernel order is obtained based on the corresponding branch degree and logarithmic function.

[0099] Among them, with nodes For example, its corresponding convolution kernel order for:

[0100] ;

[0101] Specifically, This represents the Sigmoid activation function. This is used to control the sensitivity of the Sigmoid activation function to the neighbor distance entropy, i.e. The larger the value, the greater the entropy of the neighbor distance. When changes occur, The greater the magnitude of the numerical change, the more the convolution kernel order in the graph neural network model is used to determine the aggregation range. The higher the order, the larger the aggregation range. Corresponding branch degree When this occurs, it indicates that the node is directly connected to only two nodes; in this case, a threshold should be set. When the neighbor distance entropy is less than or equal to When the distance to neighboring nodes is uniform, it indicates that the propagation of leaked exhaust gas in the exhaust gas system follows the laws of fluid dynamics. The diffusion velocity after the exhaust gas leak is affected by factors such as pipe slope and exhaust gas density. The uniform distance between neighboring nodes indicates that the diffusion environment of the exhaust gas within the range included by the set of neighboring nodes is more consistent. At this time, the diffusion velocity and other characteristics carried by one layer of neighboring nodes can accurately reflect the diffusion situation of the leak. Therefore, only one layer of neighboring nodes is aggregated at this time. Based on this, this embodiment sets an activation function. When the neighbor distance entropy is less than or equal to hour, It is a negative number. The value of is closer to 0, that is Closer to 1, according to The rounded values ​​determine the number of neighbor nodes to be aggregated to level 1, meaning that only the features of two directly adjacent nodes are aggregated.

[0102] When the neighbor distance entropy is greater than When, it indicates that at this time, it is related to the node. Two directly connected nodes and nodes A large distance difference indicates an uneven distribution of neighboring nodes. In this case, a closer neighboring node has less noise and transmits information with higher accuracy, while a farther neighboring node has more noise and transmits information with lower accuracy. Aggregating only the features of these two neighboring nodes is insufficient for accurate localization; therefore, it's necessary to expand the aggregation range and increase the number of features to mitigate the impact of noise. Thus, when the neighbor distance entropy is greater than... hour Closer to 2, according to After rounding, the number of neighbor nodes in the aggregation is determined to be 2 levels. Therefore, it is necessary to obtain the set of neighbor nodes corresponding to these two neighbor nodes and use them as nodes. The second-level neighbor nodes have a larger distance, but the noise increases with the distance between them. Therefore, to avoid the negative impact of the noise outweighing the positive impact of increasing the number of features, this embodiment only adds one level of neighbor nodes. For example, such as... Figure 2 As shown, for node A, when the number of neighboring nodes to be aggregated is 2, the nodes that need to be aggregated include nodes B, C, D, and F.

[0103] when At this time, 3 is the preset value, indicating that the node... For multi-branch nodes (such as exhaust gas collection main pipes, exhaust gas treatment equipment with multiple interfaces, etc.), in this topology, leaked gas may propagate to the node from multiple directions. To ensure the comprehensiveness of aggregation characteristics, the aggregation range of such multi-branch nodes is larger than that of the aforementioned two-branch nodes. For the minimum number of aggregation layers, for example , This indicates rounding, but similarly, the more aggregation layers there are, the greater the noise. To avoid the number of aggregation layers growing indefinitely with the number of branches, this embodiment sets a logarithmic function so that the order increases with the number of branches, while limiting the growth rate of the number of aggregation layers. Furthermore, this embodiment also sets a minimum value function to prevent the order from growing indefinitely when the number of branches is too large. For example... hour, In this embodiment, the threshold is... , and minimum number of polymerization layers The specific value is not limited and can be determined by those skilled in the art based on historical data.

[0104] This embodiment adjusts the convolution kernel order based on the node topology. For nodes with a branch degree of 2, the neighbor distance entropy and activation function are used to smoothly and precisely control the convolution kernel order according to the uniformity of the neighbor node distribution. This allows for the aggregation of features from multiple layers of neighbor nodes when the neighbor distances are uneven to avoid noise interference. When the neighbor distances are uniform, aggregating features from only one layer of neighbor nodes is sufficient to cover the effective information. For multi-branch nodes, a logarithmic function is used to ensure that the order grows reasonably with the number of branches, guaranteeing the aggregation of leakage features from multiple directions. Furthermore, a minimum value operation effectively limits the unlimited growth of the order, preventing the aggregation of too many noisy features from distant neighbors. Overall, this embodiment achieves adaptive convolution kernel order to the local topology of different exhaust gas system nodes. This allows the graph neural network to fully utilize effective neighbor information and avoid noise interference during feature aggregation, greatly improving the accuracy and robustness of tasks such as locating exhaust gas leaks in industrial parks.

[0105] Furthermore, this embodiment provides a step for determining the coordinates of exhaust gas anomalies based on the convolution kernel order corresponding to each node, see [link to relevant documentation]. Figure 3 ,include:

[0106] For each node, perform an iterative operation, which includes generating the current weight based on the weight probability distribution, obtaining dynamic aggregated features based on the current weight and the convolution kernel order, obtaining attention weights based on the dynamic aggregated features, obtaining the coordinates corresponding to the current weight based on the attention weights, until a preset termination condition is met, and outputting the coordinates corresponding to each weight.

[0107] Based on the coordinates corresponding to each weight, the coordinates of the exhaust gas anomaly are obtained.

[0108] For example, for nodes First, the current weights are generated. Since the graph neural network model includes multiple convolutional layers, the current weights include the weights of each current layer. The multiple weights in each layer follow a weight probability distribution. ,in This represents any trainable parameter in this layer. This represents the mean of the parameters. This represents the variance of the parameter. The mean is variance is It follows a normal distribution.

[0109] Then, nodes are generated based on the current weights. The corresponding dynamic aggregation features are:

[0110] ;

[0111] ;

[0112] in, This represents the dynamically aggregated features obtained after one convolution. This represents the set of all nodes within the aggregation range determined by the convolution kernel order. Used as an index for traversal. Each node in the middle, The normalization coefficient is... , Represents a node and nodes Distance decay weight, The rate coefficient representing distance decay is used to control the degree of influence of distance on weight decay. This embodiment... and The specific value is not limited. Represents a node and nodes distance, Represents a node in a graph neural network. and nodes The weight of the edge between nodes, if nodes Not with nodes For directly adjacent nodes, the connecting nodes need to be determined. and nodes For each edge, the product of the weights of each edge is ultimately used as... , Point and nodes The overall weight between them Represents a node The initial input features are defined as follows: for each node, the corresponding initial input features are feature vectors, where each element in the feature vector corresponds to the normalized values ​​of branch degree, correction concentration, correction rate, correction pressure, and neighbor distance entropy. This represents the trainable convolutional weight matrix in the first convolutional layer, obtained through the weight distribution. Similarly, the dynamically aggregated features obtained after the second convolution... for:

[0113] ;

[0114] This process continues until the last convolutional layer outputs dynamically aggregated features. , Represents a node The dynamic aggregation features obtained after one convolution are This represents the trainable convolutional weight matrix in the second convolutional layer. To further highlight nodes that contribute significantly to exhaust gas anomaly localization, an attention mechanism is needed to determine the nodes. Corresponding attention weights for:

[0115] ;

[0116] in, The scoring function assigns higher scores to more abnormal nodes, and the attention weight is greater. Indicates the total number of nodes. This is an exponential function. Repeating the above steps for each node yields the attention weights for that node. Then, the dynamically aggregated features of all nodes are weighted and fused to obtain the global features. for:

[0117] ;

[0118] in, Represents a node The process The dynamic aggregated features obtained after multiple convolutions The total number of all nodes. Represents a node The attention weights are then used to map the global features through a fully connected layer to obtain the coordinates of nodes with abnormal exhaust gas, and thus obtain the coordinates of the abnormal exhaust gas.

[0119] Then, the current weights of each layer are regenerated through the weight probability distribution, and this process is repeated multiple times until the preset number of iterations is reached. Multiple exhaust gas anomaly coordinates are obtained, and the variance of the multiple exhaust gas anomaly coordinates is calculated as an uncertainty index. If the uncertainty index is less than or equal to the threshold, the multiple exhaust gas anomaly coordinates and the corresponding uncertainty index are output. If it is greater than the threshold, the iteration steps are repeated.

[0120] Based on the above analysis, the steps for obtaining dynamically aggregated features based on the current weights and convolution kernel order can be summarized as follows:

[0121] The multi-level neighbor nodes corresponding to a node are determined based on the convolution kernel order.

[0122] For each neighbor node, the corresponding comprehensive weight is obtained based on the distance decay weight and the weight of the edge in the graph neural network;

[0123] Based on the current weight, the comprehensive weight, and the corrected features, the dynamic aggregated features are obtained.

[0124] This embodiment uses Bayesian inference to probabilistically model the weights and combines the adaptive adjustment of the dynamic convolution kernel with the branch degree and neighbor distance entropy, enabling the model to accurately capture the abnormal characteristics of exhaust gas under different topological structures of the park's exhaust gas system, effectively improving the accuracy of exhaust gas anomaly point location. On the other hand, uncertainty is quantified by calculating the variance of the location coordinates in multiple rounds, and a confidence decision mechanism based on a set threshold is used. When the uncertainty is low, the location result is directly output to guide operation and maintenance, and when the uncertainty is high, it is iterated again, ensuring the reliability of the location result and providing accurate and reliable anomaly location information support for subsequent work such as environmental pollution inspection.

[0125] Furthermore, this embodiment provides a step for obtaining a weight probability distribution based on Bayesian inference, including:

[0126] Define prior distribution and variational distribution;

[0127] For each observation data point, the corresponding loss function value is obtained based on the current variational distribution;

[0128] The variational distribution is updated based on the loss function value until the variational distribution converges, thus obtaining the weighted probability distribution.

[0129] Specifically, in order to obtain the weight probability distribution First, we define the prior distribution and the variational distribution. For example, we first assume that the model weights follow a normal distribution, i.e., the prior distribution. , The mean is variance is The normal distribution, whose mean and variance in the prior distribution can be determined from historical data, and the variational distribution is defined. , The mean is variance is The normal distribution, where and It will be continuously updated based on the loss function value. In the initial definition, the mean and variance of the variational distribution can be randomly initialized.

[0130] Next, for each observation data in the historical data First, the weights are sampled multiple times from the variational distribution. For each sampled weight value... Calculate the probability of the observed data occurring under this weight. The mean of each probability obtained from multiple sampling weights is denoted as the expected value of the data likelihood. Next, the KL divergence between the variational distribution and the prior distribution is calculated, and the KL divergence is subtracted from the expected value of the data likelihood to obtain the loss function value. Finally, the loss function is updated using gradient descent. and The values ​​are then calculated, and the steps for calculating the expected value and KL divergence are repeated, continuously updating the values. and The value is obtained until the variational distribution converges. Variational distribution convergence means that the values ​​obtained in two consecutive iterations are... and If the numerical change is less than a certain threshold, this embodiment does not impose any restrictions on the value of the threshold, and the final output variational distribution is recorded as the weighted probability distribution.

[0131] Based on the above analysis, we can summarize as follows: Figure 4 As shown, the steps for obtaining the corresponding loss function value based on the current variational distribution for each observed data point include:

[0132] Multiple resampling is performed based on the current variational distribution to obtain multiple weight values;

[0133] The probability of the observed data is obtained based on each weight value;

[0134] The expected value is obtained based on the probability.

[0135] Calculate the KL divergence between the current variational distribution and the prior distribution;

[0136] The loss function value is obtained based on the expected value and the KL divergence.

[0137] This embodiment defines the weights of each layer in the network as random variables following a specific probability distribution, rather than the fixed values ​​of traditional neural networks. This allows the model to learn not only the mean of the weights but also their variance during training. This enables the model to naturally capture the error propagation process caused by input noise and data sparsity. For example, when the data collected by a node is severely affected by environmental interference (such as temperature and humidity), the variance of the corresponding input layer weights will increase. Through multi-layer probability propagation, this will ultimately manifest as higher location uncertainty in the output exhaust gas anomaly coordinates. This embodiment utilizes variational distribution to sample and optimize the weight distribution. Multiple location results can be obtained through multiple samplings during the inference phase, quantifying the uncertainty in the model's inference process and improving the reliability and interpretability of the exhaust gas anomaly location results.

[0138] This application provides an environmental pollution inspection system, including:

[0139] The correction module is used to obtain the initial features corresponding to each node, and correct the initial features according to the attenuation formula to obtain the corrected features;

[0140] The first update module is used to correct the weights of edges in the graph neural network based on the corrected features;

[0141] The calculation module is used to calculate the branching degree and neighbor distance entropy for each node;

[0142] The second update module is used to update the order of the convolution kernel in the graph neural network according to the branch degree and the neighbor distance entropy, so as to obtain the convolution kernel order corresponding to each node;

[0143] The positioning module is used to determine the coordinates of the exhaust gas anomaly based on the convolution kernel order corresponding to each node.

[0144] The strategy module is used to formulate inspection plans based on the coordinates of the abnormal exhaust gas.

[0145] Specifically, after obtaining the coordinates of the abnormal exhaust gas, the inspection range can be delineated based on the type of exhaust gas medium and the pressure of the exhaust gas system, with the coordinates as the center. For example, flammable and explosive exhaust gases (such as hydrocarbon exhaust gases) and high-pressure exhaust gases have a high risk of rapid leakage and diffusion, so a radius of 15 meters can be delineated. Ordinary exhaust gases at normal pressure diffuse more slowly, so a radius of 8 meters can be delineated. Furthermore, the neighboring nodes directly adjacent to the nodes corresponding to the abnormal exhaust gas coordinates should be identified to check for multiple leaks or pipeline cascading damage. Then, specific inspection items should be determined based on the type of exhaust gas medium. For example, for flammable and explosive exhaust gases, the focus should be on checking for odors and whether there is a risk of open flame on the equipment. For corrosive exhaust gases (such as sulfur-containing exhaust gases), the focus should be on checking the corrosion of the equipment and whether there are abnormal discoloration patches on the ground. If there are multiple abnormal exhaust gas coordinates and there are insufficient inspection personnel, the priority can be further determined based on uncertainty indicators, exhaust gas medium type, and distance from the surrounding environment (such as production workshops and residential areas). Inspections should then be carried out according to the priority based on the abnormal exhaust gas coordinates.

[0146] This application provides an environmental pollution inspection method and system. By updating the convolution kernel of each node in the graph neural network model through the branch degree and neighbor distance entropy of each node in the exhaust gas system, the convolution kernel can be adapted to the topology of each node, thereby accurately determining the location of exhaust gas anomalies and generating reasonable inspection plans to prevent further pollution expansion. Furthermore, by using Bayesian inference to make the model weights follow a probability distribution, the generalization ability of the graph neural network model under complex working conditions is enhanced, solving the problem that existing technologies are difficult to adapt to complex environments and complex topologies.

[0147] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0148] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0149] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0150] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. An environmental pollution inspection method, characterized by, The method comprises: obtaining initial features corresponding to each node, and correcting the initial features according to a decay coefficient to obtain corrected features; correcting the weights of edges in the graph neural network according to the corrected features; calculating the branch degree and neighbor distance entropy corresponding to each node; updating the order of the convolution kernel in the graph neural network according to the branch degree and the neighbor distance entropy to obtain the convolution kernel order corresponding to each node; determining the exhaust abnormality coordinates according to the convolution kernel order corresponding to each node; formulating an inspection scheme according to the exhaust abnormality coordinates; wherein the updating of the order of the convolution kernel in the graph neural network according to the branch degree and the neighbor distance entropy to obtain the convolution kernel order corresponding to each node comprises: for each node, if the corresponding branch degree is less than a preset value, obtaining the corresponding convolution kernel order according to the corresponding neighbor distance entropy and an activation function; if the corresponding branch degree is greater than or equal to the preset value, obtaining the corresponding convolution kernel order according to the corresponding branch degree and a logarithmic function; wherein the convolution kernel order is: ； representative node corresponding convolution kernel order, representative node corresponding branch degree, representative node corresponding neighbor distance entropy, representative activation function, threshold used for controlling the sensitivity of the Sigmoid activation function to the neighbor distance entropy, threshold used for judging whether the distance distribution of the node is uniform, representative minimum aggregation layer number, 3 is a preset value; wherein the determination of the exhaust abnormality coordinates according to the convolution kernel order corresponding to each node comprises: for each node, performing an iteration operation, which comprises generating a current weight according to a weight probability distribution, obtaining a dynamic aggregation feature according to the current weight and the convolution kernel order, obtaining an attention weight according to the dynamic aggregation feature, obtaining a global feature by fusing the dynamic aggregation features corresponding to each node according to the attention weight, and mapping the global feature through a fully connected layer to obtain the coordinates corresponding to the current weight, until a preset termination condition is met, and outputting the coordinates corresponding to each weight; calculating a variance according to the coordinates corresponding to each weight, and obtaining the exhaust abnormality coordinates according to the variance; wherein the weight probability distribution is: ； denotes any one trainable parameter in any one convolutional layer in the graph neural network model, denotes the mean of the parameter, denotes the variance of the parameter, denotes a normal distribution with mean and variance . the dynamic aggregation feature is: ； Indicates the process The dynamic aggregated features obtained after multiple convolutions This represents the set of all nodes within the aggregation range determined by the convolution kernel order. Used as an index for traversal. Each node in The normalization coefficient is... , Represents a node and nodes Distance decay weight, , The rate coefficient representing distance decay, Represents a node and nodes distance, In a graph neural network, a node represents... and nodes The weight of the edges between them. Represents a node The initial input features, Indicates the first The trainable convolutional weight matrix in each convolutional layer; the attention weight is: ； representing nodes corresponding attention weights, is a scoring function, is an exponential function, representing the total number of nodes, representing nodes after the dynamic aggregation feature obtained by the convolution times the global feature is: ； representing nodes of the passage after the second convolution to obtain dynamic aggregation features, representing nodes attention weights, is an index used to traverse each node in 2. The method of claim 1, wherein, the initial features include concentration and pressure, and the corrected features include corrected concentration and corrected velocity, and the correction of the initial features according to the decay coefficient to obtain the corrected features comprises: obtaining a decay coefficient according to a digital twin model; obtaining the corrected concentration according to the decay coefficient and the exhaust concentration; obtaining the corrected velocity according to the pressure and a medium coefficient.

3. The method of claim 1, wherein, The calculation of the branch degree and the neighbor distance entropy corresponding to each node comprises: for each node, obtaining the corresponding branch degree according to the number of nodes directly adjacent thereto; for each node, determining the distance to each neighbor node and the neighbor node set; obtaining the neighbor distance entropy according to the distance to each neighbor node and the neighbor node set.

4. The method of claim 1, wherein, The obtaining of the dynamic aggregation feature according to the current weight and the convolution kernel order comprises: determining the multi-layer neighbor nodes corresponding to the node according to the convolution kernel order; for each layer of neighbor nodes, obtaining the corresponding comprehensive weight according to the distance decay weight and the weights of edges in the graph neural network; obtaining the dynamic aggregation feature according to the current weight, the comprehensive weight, and the corrected features.

5. The method of claim 1, wherein the step of detecting the environmental pollution comprises the steps of: detecting the environmental pollution by using a camera; and detecting the environmental pollution by using a sensor. The weight probability distribution is obtained according to Bayesian inference.

6. The method of claim 5, wherein the step of detecting the presence of the environmental pollutant comprises detecting the presence of the environmental pollutant by detecting the presence of a volatile organic compound. The weight probability distribution is obtained according to Bayesian inference, comprising: defining a prior distribution and a variational distribution; For each observation data, a corresponding loss function value is obtained according to a current variational distribution; The variational distribution is updated according to the loss function value until the variational distribution converges, and the weight probability distribution is obtained.

7. The method of claim 6, wherein the step of detecting the presence of the environmental pollutant comprises detecting the presence of the environmental pollutant by detecting the presence of a volatile organic compound. The method further includes: According to the current variational distribution, multiple weight values are obtained by resampling multiple times; According to each weight value, a probability of occurrence of the observation data is obtained; According to the probability, an expected value is obtained; A KL divergence between the current variational distribution and a prior distribution is calculated; According to the expected value and the KL divergence, the loss function value is obtained.

8. An environmental pollution inspection system for implementing the environmental pollution inspection method according to any one of claims 1 to 7, characterized by, The method further includes: A correction module configured to obtain an initial feature corresponding to each node, and correct the initial feature according to a decay formula to obtain a corrected feature; A first update module configured to correct a weight of an edge in the graph neural network according to the corrected feature; A calculation module configured to calculate a branch degree and a neighbor distance entropy corresponding to each node; A second update module configured to update an order of a convolution kernel in the graph neural network according to the branch degree and the neighbor distance entropy to obtain a convolution kernel order corresponding to each node; A positioning module configured to determine an exhaust gas abnormal coordinate according to the convolution kernel order corresponding to each node; A strategy module configured to formulate an inspection scheme according to the exhaust gas abnormal coordinate.

Images