A risk source identification method based on ResGCN-GRU

Abstract

The present application relates to a kind of risk source identification methods based on ResGCN-GRU, comprising the following steps: obtaining the monitoring data of each monitoring station in detection area;The monitoring data obtained is cleaned, and data is divided into training set and test set;According to the propagation trend of monitoring data and the geographical location relationship of each monitoring station, an adjacency matrix is constructed;According to the adjacency matrix constructed and air pollution index standard, mark whether it is risk source time at each moment, the training set data of the cleaned monitoring data and its corresponding label constitute input vector;Input vector and adjacency matrix are input into ResGCN-GRU model for training;Test set is input into the ResGCN-GRU model trained to identify risk source, to determine whether there is risk source in detection area at a moment.The present application provides a kind of effective method for identifying risk source, can be more accurately identified risk source, provides effective technical support for atmospheric environmental management.

Description

Technical Field

[0001] This invention belongs to the field of deep learning technology, specifically a risk source identification method based on ResGCN-GRU. Background Technology

[0002] Air pollution is a key and difficult-to-solve environmental pollution problem, as pollutants in the atmosphere have a significant impact on human health. When the concentration of pollutants in urban air reaches certain levels, it not only negatively affects the entire ecological environment but also harms normal human life. Currently, the overall air quality in Chinese cities still largely depends on meteorological factors, and air pollution control remains a long and arduous task. Identifying and classifying risk sources during periods of heavy pollution and issuing timely warnings can enable staff to implement environmental control measures and prevent the further spread of these risk sources.

[0003] Risk source identification is mostly based on the analysis of image data from infrared and satellite remote sensing. Satellite remote sensing technology can monitor pollutants, but it is prone to confusion between high-pollution areas and cloud cover, resulting in low accuracy in image recognition and an inability to precisely identify risk sources. Furthermore, compared to grid-based monitoring, infrared and satellite remote sensing have drawbacks such as higher cost and more inconvenient operation and maintenance. Methods for extracting spatiotemporal features from monitoring data are relatively mature, but they are mostly used for air pollution prediction and are currently less common in risk source identification. Therefore, this paper proposes a risk source identification method based on ResGCN-GRU. Summary of the Invention

[0004] To enhance the air pollution monitoring capabilities of grid-based monitoring systems and fully leverage the role of deep learning in air pollution monitoring, this invention proposes a risk source identification method based on ResGCN-GRU. This invention uses monitoring data from 72 monitoring stations and their adjacency matrix as input to the model. ResGCN-GRU is used to extract the spatiotemporal features of the monitoring data, which are then input into a fully connected layer and a softmax function to obtain the classification results.

[0005] The technical solution adopted by this invention to achieve the above objectives is: a risk source identification method based on ResGCN-GRU, comprising the following steps:

[0006] 1) Obtain historical monitoring data from each monitoring station within the monitoring area;

[0007] 2) Clean the acquired historical monitoring data and divide the data into training and testing sets;

[0008] 3) Construct an adjacency matrix based on the propagation trend of historical monitoring data and the geographical relationships of each monitoring station;

[0009] 4) Based on the constructed adjacency matrix and air pollution index standards, mark each time point as a risk source time, and construct an input vector from the training set data and its corresponding labels;

[0010] 5) Input the input vector and adjacency matrix into the ResGCN-GRU model for training, and use the test set data for validation;

[0011] 6) Obtain monitoring data from a set time before a certain time at the monitoring station, input it into the trained ResGCN-GRU model to identify risk sources, and determine whether there are risk sources in the monitoring area at the current time.

[0012] The monitoring data includes PM2.5 pollution levels, geographical location, identification number, and station name.

[0013] Step 5) involves inputting the input vector and adjacency matrix into the ResGCN-GRU model for training and validating it using a test set. This includes the following steps:

[0014] Input the input vector and adjacency matrix into the ResGCN model to extract the spatial features of the data;

[0015] The temporal vector with spatial features output by the ResGCN model is input into the GRU network structure for model training to extract the spatiotemporal features of the data.

[0016] The spatiotemporal features are sequentially input into a fully connected layer and a softmax activation function, and the classification results are output to represent the risk source time and the non-risk source time.

[0017] The test set from the cleaned historical monitoring data is input into the trained ResGCN-GRU model for model validation.

[0018] The data cleaning process specifically involves: performing outlier detection on the acquired historical monitoring data, treating the detected outliers as missing values, and then using the K-nearest neighbor method to estimate the missing values ​​of a station by weighted averaging the values ​​of nearby associated monitoring stations.

[0019] The construction of the adjacency matrix includes the following steps:

[0020] Based on the monitoring data corresponding to all time periods of each monitoring station, the correlation between monitoring stations when risk sources appear and when risk sources disappear is obtained, and an adjacency matrix between monitoring stations is constructed by combining the geographical location of the monitoring stations.

[0021] The rows and columns of the adjacency matrix represent N stations respectively. According to the association relationship, the values ​​corresponding to two stations in the adjacency matrix are set to 0 or 1. 0 indicates that there is no correlation between the two stations, that is, there is no propagation trend; 1 indicates that there is a correlation between the two stations, that is, there is a propagation trend.

[0022] The specific method for marking risk sources is as follows: based on the set air pollution index standard, if the monitored value of a monitoring station at a certain moment is greater than the threshold A, it is determined to be a polluted station; if the number of associated pollution monitoring stations at a certain moment is greater than the threshold B, this moment is marked as a risk source moment, otherwise it is marked as a non-risk source moment.

[0023] A risk source identification device based on ResGCN-GRU, comprising:

[0024] The data acquisition module is used to acquire historical monitoring data from each monitoring station within the monitoring area;

[0025] The data cleaning module is used to clean the acquired historical monitoring data and divide the data into training and testing sets.

[0026] The model training module is used to construct an adjacency matrix based on the propagation trend of historical monitoring data and the geographical relationships of each monitoring station; based on the constructed adjacency matrix and the air pollution index standard, it marks each moment as a risk source moment; it uses the cleaned historical monitoring data and its corresponding labels to form an input vector; and it inputs the input vector and the adjacency matrix into the ResGCN-GRU model for training.

[0027] The risk source identification module is used to acquire monitoring data from a certain time before a certain time to a certain time from the monitoring station, input it into the trained ResGCN-GRU model to identify risk sources, and determine whether there are risk sources in the monitoring area at the current time.

[0028] The ResGCN model is as follows:

[0029] Mapping relationship X′ t =F(X) t )+X t X t It is the input vector, F(X) t Let X' be the residual function. t It is the output of the residual block, which can effectively fuse the spatial features extracted by the two layers of GCN and improve the accuracy of the model.

[0030] GCN uses convolution to extract features from graph-structured data. The (l+1)th layer in the convolutional network has the following non-linear activation function for forward propagation:

[0031]

[0032] In the above formula, σ(·) is the activation function; A is the adjacency matrix, I N It is the identity matrix. It is the adjacency matrix after adding self-connections; It is a degree matrix. express The degree of the i-th station, i.e. The sum of the i-th row; H (l) Let H be the activation matrix of the l-th layer, and H 0 =X t X t It is the input vector; W (l) Let be the weight matrix of the l-th layer.

[0033] A risk source identification device based on ResGCN-GRU includes a memory and a processor; the memory is used to store a computer program; the processor is used to implement the risk source identification method based on ResGCN-GRU when the computer program is executed.

[0034] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned risk source identification method based on ResGCN-GRU.

[0035] The present invention has the following beneficial effects and advantages:

[0036] 1. This invention provides a practical and effective method for identifying risk sources, which can accurately identify risk sources and provide effective technical support for atmospheric environmental governance.

[0037] 2. This invention uses PM2.5 values ​​monitored by 72 small-scale monitoring stations in Shenyang City and an adjacency matrix constructed based on the geographical location information of the monitoring stations and the risk source propagation trend as input vectors. It fully leverages the spatial feature extraction capabilities of the GCN and the temporal feature extraction capabilities of the GRU. The ResGCN, constructed by adding a residual network to the GCN layer, improves the accuracy of the final classification results. Attached Figure Description

[0038] Figure 1 This is a basic flowchart of a risk source identification method based on ResGCN-GRU according to the present invention;

[0039] Figure 2 Here is a structural diagram of the ResGCN-GRU;

[0040] Figure 3 This is a structural diagram of ResGCN. Detailed Implementation

[0041] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0042] Currently, grid-based monitoring technology for air pollution prevention and control has been gradually established in various cities. The application of grid-based air pollution prevention and control systems can provide real-time monitoring data and information on the atmospheric environment. Furthermore, grid-based monitoring equipment for air pollution prevention and control has many advantages, such as low cost and low operation and maintenance requirements. By selecting an appropriate deep learning model to train the data, the distribution characteristics and patterns of the data can be learned, thereby solving the problems encountered by traditional methods in pollution source identification research. Therefore, this invention proposes a risk source identification method based on ResGCN-GRU.

[0043] ResGCN combines residual networks with GCNs, enabling more effective learning of spatial features. The residual network effectively addresses the vanishing gradient problem caused by over-smoothing in multi-layer GCNs, and it also effectively integrates the spatial features learned by multiple GCNs, resulting in better overall model classification performance. The GRU layer effectively extracts temporal features from the dataset and, compared to LSTM, reduces model complexity, converges more easily, and accelerates training.

[0044] like Figure 1 As shown, the risk source identification method based on ResGCN-GRU includes the following steps:

[0045] Step 1: The grid-based monitoring network in Shenyang City comprises 72 micro-monitoring stations. The pollutant concentrations monitored by these stations are uploaded to a host computer. The host computer then retrieves the PM2.5 concentration data and latitude / longitude information of each monitoring station. The grid-based monitoring network divides the urban area into grids based on the monitoring stations, representing the air quality data within each grid area.

[0046] Step 2: Clean the acquired monitoring data, mainly by handling outliers and missing values, and divide the data into training and testing sets;

[0047] Step 3: Construct an adjacency matrix based on the propagation trend of historical pollution data among monitoring stations and the geographical relationships of each monitoring station;

[0048] Step 4: Based on the constructed adjacency matrix and the air pollution index standard in the host computer, mark each time moment as a risk source moment; then, construct the input vector from the cleaned monitoring data and the labels just marked for each time moment.

[0049] Step 5: Input the input vector and adjacency matrix into the ResGCN model to extract the spatial features of the data;

[0050] Step 6: Input the temporal vector with spatial features output by ResGCN into the GRU network structure for model training, and then extract the spatiotemporal features of the data;

[0051] Step 7: Input the spatiotemporal features into the fully connected layer and the softmax activation function to output the classification results and identify whether the moment is a risk source moment;

[0052] Step 8: Identify risk sources using the trained ResGCN-GRU model to determine whether a risk source exists in Shenyang at a given moment, issue timely warnings, and enable staff to implement environmental remediation measures to prevent further spread of the risk source. Here, "a given moment" refers to the model's input, which can represent a future moment. In other words, the model uses the input vector composed of this future moment and previous moments to identify whether a risk source exists at the current moment. For example, to identify whether November 12, 2021, at 15:00 is a risk source, the monitoring data from all stations between 10:00 and 15:00 on November 12, 2021, are input into the trained model to obtain the identification result for that moment.

[0053] The data cleaning process specifically involves: performing outlier detection on the data obtained from the host computer, treating the detected outliers as missing values, and then using the K-nearest neighbor method to estimate the missing value of a station by weighted averaging the values ​​of nearby associated stations;

[0054] The construction of the adjacency matrix specifically involves observing the monitoring data of each station every hour in October and November 2021, observing the correlation between stations when risk sources appear and disappear, and constructing an adjacency matrix between stations based on the geographical location of the monitoring stations. The adjacency matrix has 72 rows and 72 columns. Based on the above correlation, the values ​​corresponding to two stations in the adjacency matrix are set to 0 or 1, where 0 indicates no correlation and 1 indicates correlation. The correlation indicates the propagation trend between stations observed based on historical data, and stations with propagation trends are correlated.

[0055] The specific steps for marking risk sources are as follows: According to the air pollution index standard, if the monitoring value of a monitoring station is greater than 75 at a certain moment, it is determined to be a polluted station. If the number of associated polluted stations is greater than 3 at a certain moment, this moment is marked as a risk source moment, which is recorded in Excel and then combined with the cleaned monitoring data to form an input vector and input into the model; otherwise, it is marked as a non-risk source moment.

[0056] The ResGCN layer is specifically described as follows: ResGCN combines a residual network with a GCN, and its structure is as follows. Figure 2 , Figure 3 As shown:

[0057] Its mapping relationship X′ t =F(X) t )+X t X t It is the input vector, F(X) t Let X' be the residual function. t It is the output of the residual block, which can effectively fuse the spatial features extracted by the two layers of GCN and improve the accuracy of the model.

[0058] GCN uses convolution to extract features from graph-structured data. The (l+1)th layer in the convolutional network has the following non-linear activation function for forward propagation:

[0059]

[0060] In the above formula, σ(·) is the activation function; A is the adjacency matrix, I N It is the identity matrix. It is the adjacency matrix after adding self-connections; It is a degree matrix. express The degree of the i-th station, i.e. The sum of the i-th row; H (l) Let H be the activation matrix of the l-th layer, and H 0 =X t X t It is the input vector; W (l) Here is the weight matrix for the l-th layer;

[0061] It is the adjacency matrix after adding a self-join. The adjacency matrix contains information about the associated points of a certain point. If there is an association, the corresponding position in the matrix is ​​1. In the original adjacency matrix, a certain point is considered to be unrelated to itself, and the corresponding position is 0. Adding a self-join means that the point is considered to be related to itself, and the corresponding position is set to 1.

[0062] The GRU layer specifically refers to:

[0063] 1) Update gate z t The expression used to control how candidate hidden states containing information about the current time step are updated to the hidden state of the current time step is:

[0064] z t =σ(W z ·[h t-1 ,x t ]+b z )

[0065] Where σ represents the Softmax function, W z It is the weight parameter, h t-1 It is the hidden state from the previous step, x t It is the input for the current step, b z It is a bias parameter.

[0066] 2) Reset the door r t The expression used to determine whether to retain the previous hidden state is:

[0067] r t =σ(W r ·[h t-1 ,x t ]+b r )

[0068] Among them, W r It is the weight parameter, b r This is the bias parameter; the meanings of the other parameters are the same as in the above formula.

[0069] 3) Candidate hidden state Used to assist in subsequent hidden state calculations, it is calculated by multiplying the output of the reset gate at the current time step element-wise with the hidden state at the previous time step, concatenating the result with the input at the current time step, and finally calculating it through the activation function. Its expression is:

[0070]

[0071] Where tanh represents the tanh function, W h It is the weight parameter, b h It is the bias parameter, r t This is the current step reset gate output; the meanings of other parameters are the same as in the above formula.

[0072] 4) Hidden state h t The calculation expression:

[0073]

[0074] Among them, z t This is the current step update gate output. This represents the candidate hidden state for the current step; the meanings of the other parameters are the same as in the above formula.

[0075] In conjunction with the application scenario, x t It is the input at each time step, that is, the input feature matrix corresponding to each time step, h t It is the output of each time step, that is, the matrix after fusing spatiotemporal features in the current step.

Claims

1. A risk source identification method based on ResGCN-GRU, characterized in that, Includes the following steps: 1) Obtain historical monitoring data from each monitoring station within the monitoring area; 2) Clean the acquired historical monitoring data and divide the data into training and testing sets; 3) Construct an adjacency matrix based on the propagation trend of historical monitoring data and the geographical relationships of each monitoring station; 4) Based on the constructed adjacency matrix and air pollution index standards, mark each time point as a risk source time, and construct an input vector from the training set data and its corresponding labels; 5) Input the input vector and adjacency matrix into the ResGCN-GRU model for training, and use the test set data for validation; 6) Obtain monitoring data from a set time before a certain time at the monitoring station, input it into the trained ResGCN-GRU model to identify risk sources, and determine whether there are risk sources in the monitoring area at the current time; Step 5) involves inputting the input vector and adjacency matrix into the ResGCN-GRU model for training, and then validating it using test set data. This includes the following steps: Input the input vector and adjacency matrix into the ResGCN model to extract the spatial features of the data; The temporal vector with spatial features output by the ResGCN model is input into the GRU network structure for model training to extract the spatiotemporal features of the data. The spatiotemporal features are sequentially input into a fully connected layer and a softmax activation function, and the classification results are output to represent the risk source time and the non-risk source time. The test set from the cleaned historical monitoring data is input into the trained ResGCN-GRU model for model validation.

2. The risk source identification method based on ResGCN-GRU according to claim 1, characterized in that, The monitoring data includes PM2.5 pollution levels, geographical location, identification number, and station name.

3. The risk source identification method based on ResGCN-GRU according to claim 1, characterized in that: The data cleaning process specifically involves: performing outlier detection on the acquired historical monitoring data, treating the detected outliers as missing values, and then using the K-nearest neighbor method to estimate the missing values ​​of a station by weighted averaging the values ​​of nearby associated monitoring stations.

4. The risk source identification method based on ResGCN-GRU according to claim 1, characterized in that: The construction of the adjacency matrix includes the following steps: Based on the monitoring data corresponding to all time periods of each monitoring station, the correlation between monitoring stations when risk sources appear and when risk sources disappear is obtained, and an adjacency matrix between monitoring stations is constructed by combining the geographical location of the monitoring stations. The rows and columns of the adjacency matrix represent N stations respectively. According to the association relationship, the values ​​corresponding to two stations in the adjacency matrix are set to 0 or 1. 0 indicates that there is no correlation between the two stations, that is, there is no propagation trend; 1 indicates that there is a correlation between the two stations, that is, there is a propagation trend.

5. The risk source identification method based on ResGCN-GRU according to claim 1, characterized in that: The specific method for marking risk sources is as follows: based on the set air pollution index standard, if the monitored value of a monitoring station at a certain moment is greater than the threshold A, it is determined to be a polluted station; if the number of associated pollution monitoring stations at a certain moment is greater than the threshold B, this moment is marked as a risk source moment, otherwise it is marked as a non-risk source moment.

6. A risk source identification device based on ResGCN-GRU, the device being used to implement the risk source identification method based on ResGCN-GRU as described in any one of claims 1-5, characterized in that, include: The data acquisition module is used to acquire historical monitoring data from each monitoring station within the monitoring area; The data cleaning module is used to clean the acquired historical monitoring data and divide the data into training and testing sets. The model training module is used to construct an adjacency matrix based on the propagation trend of historical monitoring data and the geographical relationships of each monitoring station; Based on the constructed adjacency matrix and air pollution index standards, each moment is marked as a risk source moment. The training set data from the cleaned historical monitoring data and its corresponding labels are used to construct the input vector. The input vector and the adjacency matrix are then input into the ResGCN-GRU model for training. The risk source identification module is used to acquire monitoring data from a certain time before a certain time to a certain time from the monitoring station, input it into the trained ResGCN-GRU model to identify risk sources, and determine whether there are risk sources in the monitoring area at the current time.

7. A risk source identification device based on ResGCN-GRU according to claim 6, characterized in that: The ResGCN model is as follows: Mapping relationship , It is the input vector. For the residual function, It is the output of the residual block. GCN uses convolution to extract features from graph-structured data, defining the first convolutional layer in a convolutional network. The layer, whose forward propagation nonlinear activation function is: ； In the above formula It is an activation function; Let A be the adjacency matrix. It is the identity matrix. It is the adjacency matrix after adding self-connections; It is a degree matrix. , express The degree of the i-th station, i.e. The sum of the i-th row; For the first The activation matrix of the layer, and , It is the input vector; For the first Layer weight matrix.

8. A risk source identification device based on ResGCN-GRU, characterized in that, It includes a memory and a processor; the memory is used to store a computer program; the processor is used to implement, when executing the computer program, a risk source identification method based on ResGCN-GRU as described in any one of claims 1-5.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements a risk source identification method based on ResGCN-GRU as described in any one of claims 1-5.

Images