A system for tracing the source of atmospheric pollutants based on big data
By combining the meteorological feature fusion, spatiotemporal feature extraction and Gaussian diffusion model of convolutional neural network and GRU, and combined with drone monitoring, efficient positioning of traceability atmospheric pollutants is achieved, solving the problem of inefficient traceability in traditional methods.
Patent Information
- Application Number
- CN202411128739.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-16
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2044-08-16
AI Technical Summary
Traditional atmospheric pollutant traceability methods cannot effectively track and locate pollution sources, especially under the influence of diffusion and cross-stations between multiple sites, resulting in inefficient traceability.
The meteorological feature fusion unit combined with convolutional neural network and GRU is used to extract multi-dimensional meteorological data, and the spatiotemporal feature extraction unit constructs a dynamic map, combining Gaussian diffusion model and drone monitoring to achieve dynamic positioning and mobile monitoring of pollution sources.
It improves the accuracy of detecting abnormal pollutant concentration, can quickly screen sites that are greatly affected by pollution, adjust the progress length and route in real time through drones, accurately locate pollution sources, and make up for the scope limitations of traditional methods.
Smart Images

Figure CN118643456B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of pollutant source tracing, and in particular to an atmospheric pollutant source tracing system based on big data. Background Art
[0002] Severe air pollution can cause ecological problems. To this end, the country has built an air quality monitoring network covering all cities at or above the prefecture level to conduct real-time monitoring of pollutants in the atmosphere and accumulate a large amount of air quality monitoring data. Air pollutant source tracing is one of the core tasks of air pollution prevention and control. The Chinese patent application number 202210496415.7 discloses "an air pollutant source tracing system and method based on big data technology, specifically involving the application field of big data technology environment, including data acquisition module, data aggregation module, pollution source fingerprint module, data processing module, data analysis module, source tracing warning module, and source tracing display module."
[0003] The above document combines big data real-time streaming technology to analyze the basic pollutant data, meteorological data, and geospatial data in the traceability area to track and locate pollution sources. However, the scale of air quality monitoring data is large and the correlation between sites changes dynamically. Atmospheric pollutants mainly rely on air flow for diffusion and transmission. The concentration of atmospheric pollutants will change dynamically with changes in meteorological factors, and will also migrate and exchange between different regions. Pollutants between multiple sites will cross-influence each other due to diffusion. The traditional method of tracing the source of atmospheric pollutants determines the source of pollution by tracking the gas transmission path, but this method can only locate the pollution source near the sensor and cannot track it in breadth, reducing the efficiency of tracing. Summary of the invention
[0004] The purpose of the present invention is to provide an atmospheric pollutant source tracing system based on big data to solve the problems raised in the above-mentioned background technology.
[0005] To achieve the above object, the present invention provides the following technical solutions: a big data-based atmospheric pollutant tracing system, comprising a meteorological feature fusion unit, wherein the meteorological feature fusion unit extracts meteorological features of n stations from meteorological data of multiple dimensions of temperature, humidity, wind direction and wind speed by combining a convolutional neural network with a GRU, and divides the meteorological features into multiple segments by a sliding window, wherein each sliding window contains n stations, and is used to dynamically learn the correlation between the n stations;
[0006] A spatiotemporal feature extraction unit, which extracts time features from the time series data of each station in each window to capture time correlation, obtains spatial features from the meteorological feature fusion unit, adaptively learns the correlation between each station, and constructs a dynamic graph by combining the time features and the spatial features;
[0007] An abnormality locking unit, which obtains the spatiotemporal information composed of time features and space features from the spatiotemporal feature extraction unit, and then calculates the probability density of the spatiotemporal information of each sliding window, so as to detect the abnormality of the concentration of atmospheric pollutants in the city;
[0008] A site concentration acquisition unit, wherein the site concentration acquisition unit establishes a Gaussian diffusion model for the suspected pollution source in the traceability area during the abnormal period, first performs coordinate transformation and then uses the Gaussian diffusion model to calculate the concentration value of the pollutants released by the pollution source diffused to each site position in the traceability area;
[0009] The pollution tracing unit combines the concentration value of each site location, eliminates the sites that are not greatly affected by the pollution, and uses the drone to continuously adjust the travel length to search for the pollution source in real time according to the changes in the site concentration value. It is used to realize mobile monitoring of the tracing area and locate the pollution source.
[0010] Preferably, the meteorological feature fusion unit divides the meteorological features into multiple segments by means of a sliding window, and inputs the meteorological data of multiple dimensions of temperature, humidity, wind direction and wind speed into a convolutional layer, thereby extracting meteorological features, and fuses the meteorological features of the n stations with the monitoring data of the n stations in the corresponding sliding window, and inputs them into the spatiotemporal feature extraction unit to dynamically learn the correlation between the n stations.
[0011] Preferably, the spatiotemporal feature extraction unit includes a time feature extraction module and a space feature extraction module. The time feature extraction module uses a recurrent neural network to extract historical information of the t time period before the current site and historical information of the site related to the current site in the t-1 time period before, and normalizes the extracted historical information. According to the historical information, time features are extracted from the time series data of each site in each sliding window. The space feature extraction module obtains spatial features from the meteorological feature fusion unit, and combines meteorological features to model the dynamic interdependence between n sites using a graph attention mechanism. Then, a dynamic graph is constructed by combining time features and spatial features. Each node in the dynamic graph represents a monitoring point. The attention score of each node is obtained through the graph attention mechanism. Each edge represents the relationship between two sites. The weight of the edge is used to quantify the correlation. The attention score of each node constitutes an attention matrix and is trained.
[0012] Preferably, the abnormal locking unit includes a graph convolution module and a normalized flow module. The graph convolution module aggregates the trained attention matrix and time features, and captures the spatial information and time information of the neighborhood of each site, so as to obtain the spatiotemporal information of each sliding window. The graph convolution layer is used to aggregate the hidden states of n sites in a sliding window, and the spatiotemporal information within the sliding window period is feature-marked. The normalized flow module obtains the feature mark of the input spatiotemporal information from the graph convolution module, so as to perform a more detailed density estimation on the spatiotemporal information of each window, calculate the probability density mapped to the target distribution function, and use maximum likelihood estimation to optimize the parameters of the entire network.
[0013] Preferably, the abnormal locking unit also includes an abnormal score module, which is responsible for calculating the abnormal score according to the probability density of each window. The size of the abnormal score is used to measure the degree of abnormality of each window. The abnormal field score is calculated using the abnormal judgment formula, and the threshold is set using the interquartile range. The window with an abnormal score exceeding the threshold is judged as an abnormal situation.
[0014] Preferably, in the site concentration acquisition unit, the Gaussian diffusion model is specifically:
[0015] ;
[0016] in, Indicates the pollution source in the plane space Spread to coordinates The concentration value at Indicates wind speed, Indicates source strength, express The diffusion coefficient on the axis, express The diffusion coefficient on the axis, and They represent the plane coordinate axes, represents the natural exponential function, Represents a natural constant.
[0017] Preferably, the pollution tracing unit includes a pollution source search module, a concentration mean acquisition module and a behavior adjustment module. The pollution source search module is equipped with atmospheric pollutant sensors in four directions of the UAV in sequence, front, back, left and right, and uses the sensors to timely obtain the concentration values of atmospheric pollution in four directions. The concentration mean acquisition module enables the UAV to adjust its step length in real time. The farther the distance between the UAV and the pollution source is, the larger the step length is, and vice versa. The concentration values of the four sensors are averaged in combination with the UAV's step length to obtain the concentration mean. The behavior adjustment module compares the concentration mean with a preset concentration critical value to determine the UAV's route, trace the pollution location, and hover over the source.
[0018] Preferably, the steps of determining the route of the drone of the behavior adjustment module are as follows:
[0019] S1. When the mean concentration is less than or equal to the critical concentration value, the UAV updates its position in a Z-shaped variable step length and performs a global search;
[0020] The real-time position update formula is:
[0021] ;
[0022] in, express Point The direction of time, express Point The direction of time, Indicates a The constant of the interval, Indicates the highest concentration critical value, represents the mean concentration, is the minimum step length of the drone;
[0023] S2. When the mean concentration is greater than the critical concentration value, the drone continues to move in a straight line toward the maximum concentration. The maximum concentration calculation formula is:
[0024] ;
[0025] in, represents the maximum pollutant concentration, Respectively represent the pollution concentration values in the front, back, left, and right directions of the drone. Indicates the source location of air pollution sources, express Point The direction of time, Indicates the highest concentration critical value, Indicates a The constant of the interval, is the minimum step length of the drone.
[0026] Compared with the prior art, the present invention has the following beneficial effects:
[0027] 1. The present invention uses a meteorological feature fusion unit to fuse external meteorological factors and then uses a graph attention mechanism to dynamically capture the dynamic spatial correlation between sites, and captures the time correlation of each site based on the historical information of the site, thereby improving the detection accuracy. At the same time, the abnormal concentration of atmospheric pollutants in the city is detected through an abnormal locking unit, and the time period when the abnormality occurs is obtained, so that the pollution tracing unit can perform source tracing and positioning;
[0028] 2. The present invention considers the atmospheric diffusion process in combination with the Gaussian diffusion model, screens sites that are greatly affected by pollution, avoids interference from redundant sites, and calculates the concentration value of pollutants released by the pollution source diffused to each site location in the tracing area through the site concentration acquisition unit. It can provide information to the drone, enabling the drone to quickly search and locate the pollution source, making up for the shortcomings of the sensor's small-range search, and well realizing mobile monitoring of the tracing area. The drone can adjust the travel length and route in real time according to the distance from the pollution source. After finding the pollution location, it will inform by circling at the source. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] Figure 1 A schematic diagram of the overall system flow provided by an embodiment of the present invention;
[0030] Figure 2 An internal module block diagram of a spatiotemporal feature extraction unit provided in an embodiment of the present invention;
[0031] Figure 3 A block diagram of the internal modules of the abnormal locking unit provided by an embodiment of the present invention;
[0032] Figure 4 A block diagram of the internal modules of the pollution tracing unit provided in an embodiment of the present invention;
[0033] Figure 5 A flow chart of a method for determining a route taken by a drone provided by an embodiment of the present invention.
[0034] In the figure: 1. Meteorological feature fusion unit; 2. Spatiotemporal feature extraction unit; 201. Time feature extraction module; 202. Spatial feature extraction module; 3. Anomaly locking unit; 301. Graph convolution module; 302. Normalized flow module; 303. Anomaly score module; 4. Site concentration acquisition unit; 5. Pollution tracing unit; 501. Pollution source search module; 502. Concentration mean acquisition module; 503. Behavior regulation module. DETAILED DESCRIPTION
[0035] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0036] See also Figure 1-Figure 5, the present invention provides a technical solution: a big data-based atmospheric pollutant tracing system, comprising a meteorological feature fusion unit 1, the meteorological feature fusion unit 1 extracts meteorological features of n stations from meteorological data of multiple dimensions of temperature, humidity, wind direction and wind speed by combining a convolutional neural network with a GRU, and divides the meteorological features into multiple segments by a sliding window, each sliding window containing n stations, for dynamically learning the correlation between the n stations;
[0037] The spatiotemporal feature extraction unit 2 extracts time features from the time series data of each station in each window to capture time correlation, obtains spatial features from the meteorological feature fusion unit 1, adaptively learns the correlation between each station, and constructs a dynamic graph by combining time features and spatial features;
[0038] The abnormality locking unit 3 obtains the spatiotemporal information composed of time features and space features from the spatiotemporal feature extraction unit 2, and then calculates the probability density of the spatiotemporal information of each sliding window to detect the abnormal concentration of atmospheric pollutants in the city;
[0039] The site concentration acquisition unit 4 establishes a Gaussian diffusion model for the suspected pollution source in the traceability area during the abnormal period. First, coordinate transformation is performed and then the Gaussian diffusion model is used to calculate the concentration value of the pollutants released by the pollution source diffused to each site position in the traceability area. The coordinate change is used to adapt to the change of wind speed and wind direction. The Gaussian diffusion model can be used to more easily approach the pollution source.
[0040] The pollution tracing unit 5 combines the concentration value of each site location, eliminates the sites that are not greatly affected by the pollution, and uses the drone to continuously adjust the travel length to search for the pollution source in real time according to the change of the site concentration value, so as to realize mobile monitoring of the tracing area and locate the pollution source.
[0041] The meteorological feature fusion unit 1 divides the meteorological features into multiple segments by means of a sliding window, and inputs the meteorological data of multiple dimensions of temperature, humidity, wind direction and wind speed into the convolution layer to extract the meteorological features, and fuses the meteorological features of the n stations with the monitoring data of the n stations in the corresponding sliding window, and inputs them into the spatiotemporal feature extraction unit 2 to dynamically learn the correlation between the n stations;
[0042] The spatiotemporal feature extraction unit 2 includes a time feature extraction module 201 and a space feature extraction module 202. The time feature extraction module 201 uses a recurrent neural network to extract historical information of a period t before the current site and historical information of a period t-1 before the site related to the current site, and normalizes the extracted historical information, and extracts time features from the time series data of each site in each sliding window according to the historical information;
[0043] The spatial feature extraction module 202 obtains spatial features from the meteorological feature fusion unit 1, and combines the meteorological features to model the dynamic interdependence between the n sites using the graph attention mechanism, and then constructs a dynamic graph by combining the time features and the spatial features. Each node in the dynamic graph represents a monitoring point, and the attention score of each node is obtained through the graph attention mechanism. Each edge represents the relationship between two sites, and the weight of the edge is used to quantify the correlation. The attention score of each node constitutes an attention matrix and is trained;
[0044] The calculation formula of the graph attention mechanism is:
[0045] ;
[0046] in, Representation Node The output representation of is still the feature vector of the current node. represents the sigmoid activation function, Representation and Node The number of adjacent nodes, Representation Node For Node The attention score of the node For Node The contribution of For Node The adjacent nodes of , Represents the nodes in the graph and nodes The characteristic vector of Indicates the parameters, represents the natural exponential function, represents a nonlinear activation function, represents the trainable column vector parameters, , Represents the dimension of each node feature vector;
[0047] The abnormal locking unit 3 includes a graph convolution module 301 and a normalized flow module 302. The graph convolution module 301 aggregates the trained attention matrix and time features, and captures the spatial information and time information of the neighborhood of each site, so as to obtain the spatiotemporal information of each sliding window, and uses the graph convolution layer to aggregate the hidden states of n sites in a sliding window, and features the spatiotemporal information within the sliding window period;
[0048] The feature labeling formula is:
[0049] ;
[0050] in, Before The time period is a feature mark of the spatiotemporal information of the sliding window. represents the attention matrix, Before All hidden features of the time period, Before All hidden features of the time period, Represents the parameters for converting each node information in the dynamic graph, Indicates the parameters for converting the historical information of each node. Indicates additional transformation, all three parameters are trainable. represents the activation function;
[0051] The normalization flow module 302 obtains the feature tags of the input spatiotemporal information from the graph convolution module 301, thereby performing a more detailed density estimation of the spatiotemporal information of each window, calculating the probability density mapped to the target distribution function, and optimizing the parameters of the entire network using maximum likelihood estimation;
[0052] The density estimation formula is:
[0053] ;
[0054] in, represents the input sequence, Indicates windows, Represents the characteristic mark of the input sequence, that is, the conditional information, represents the distribution parameter, represents the density estimation function, Indicates Density estimation of windows;
[0055] The target distribution function is specifically:
[0056] ;
[0057] in, represents the probability density, Indicates the number of sites, represents the density estimate, represents the mean of the density estimate, represents the mean parameter, and represents a natural constant;
[0058] The abnormal locking unit 3 also includes an abnormal score module 303, which is responsible for calculating the abnormal score according to the probability density of each window. The size of the abnormal score is used to measure the degree of abnormality of each window. The abnormal score is calculated using the abnormal judgment formula, and the threshold is set using the interquartile range. The window with an abnormal score exceeding the threshold is judged as an abnormal situation. The higher the abnormal score, the lower the density of the window, and the higher the possibility of abnormality.
[0059] The abnormality determination formula is:
[0060] ;
[0061] in, Indicates The anomaly score of the window, Indicates the total number of sites, Indicates The probability density of the window is Indicates The input sequence of windows;
[0062] Quartile is a kind of quantile in statistics. All numbers are arranged from small to large and divided into four equal parts according to quantity. The difference between the third quartile and the first quartile is called the interquartile range, which represents the degree of dispersion of the middle 50% of the data set. The formula is:
[0063] ;
[0064] in, represents the threshold value, represents the first quartile, represents the third quartile;
[0065] In the site concentration acquisition unit 4, the Gaussian diffusion model is specifically:
[0066] ;
[0067] in, Indicates the pollution source in the plane space Spread to coordinates The concentration value at Indicates wind speed, Indicates source strength, express The diffusion coefficient on the axis, express The diffusion coefficient on the axis, and They represent the plane coordinate axes, represents the natural exponential function, represents a natural constant;
[0068] The pollution source tracing unit 5 includes a pollution source search module 501, a concentration mean acquisition module 502 and a path adjustment module 503. The pollution source search module 501 is provided with air pollutant sensors in the four directions of the front, back, left and right of the UAV in sequence, and uses the sensors to timely obtain the concentration values of the air pollution in the four directions. The concentration mean acquisition module 502 enables the UAV to adjust the travel step length in real time. The farther the distance between the UAV and the pollution source is, the larger the step length is, otherwise the step length becomes smaller. The calculation formula for adjusting the travel step length is:
[0069] ;
[0070] in, Indicates a The constant of the interval, Indicates the highest concentration critical value, Indicates that the drone is in The concentration value obtained by position monitoring, is the minimum step length of the drone;
[0071] The concentration values of the four sensors are averaged according to the UAV's step length to obtain the average concentration. The calculation formula is:
[0072] ;
[0073] in, Respectively represent the pollution concentration values in the front, back, left, and right directions of the drone. Indicates that the current progress is increasing. represents the mean concentration;
[0074] The path adjustment module 503 compares the concentration mean with the preset concentration threshold to determine the route of the drone, trace the pollution source, and hover over the source;
[0075] The steps of determining the route of the drone in the behavior adjustment module 503 are as follows:
[0076] S1. When the mean concentration is less than or equal to the critical concentration value, the UAV updates its position in a Z-shaped variable step length and performs a global search;
[0077] The real-time position update formula is:
[0078] ;
[0079] in, express Point The direction of time, express Point The direction of time, Indicates a The constant of the interval, Indicates the highest concentration critical value, represents the mean concentration, is the minimum step length of the drone;
[0080] S2. When the mean concentration is greater than the critical concentration value, the drone continues to move in a straight line toward the maximum concentration. The maximum concentration calculation formula is:
[0081] ;
[0082] in, represents the maximum pollutant concentration, Respectively represent the pollution concentration values in the front, back, left, and right directions of the drone. Indicates the source location of air pollution sources, express Point The direction of time, Indicates the highest concentration critical value, Indicates a The constant of the interval, is the minimum step length of the drone.
[0083] Working principle: The present invention divides meteorological features into multiple segments in a sliding window manner through the meteorological feature fusion unit 1, and inputs multiple dimensional meteorological data into the convolution layer, extracts meteorological features of n stations from multiple dimensional meteorological data of temperature, humidity, wind direction and wind speed, performs feature fusion on the meteorological features of the n stations with the monitoring data of the n stations in the corresponding sliding window, and inputs them into the spatiotemporal feature extraction unit 2 to dynamically learn the correlation between the n stations, extracts time features from the time series data of each station in each window through the spatiotemporal feature extraction unit 2 to capture time correlation, obtains spatial features from the meteorological feature fusion unit 1, combines meteorological features, uses the graph attention mechanism to model the dynamic interdependence between the n stations, and constructs a dynamic graph in combination with time features and spatial features, and constructs an attention matrix according to each node in the dynamic graph for training;
[0084] The spatiotemporal information composed of spatial features and temporal features is obtained through the abnormal locking unit 3, and the trained attention matrix and temporal features are aggregated through the graph convolution module 301. The graph convolution layer is used to aggregate the hidden states of n sites in a sliding window, and the spatiotemporal information in the sliding window period is marked with features to provide conditional information for the normalized flow module 302, so as to make a more detailed density estimation of the spatiotemporal information of each window and calculate the probability density mapped to the target distribution function. The abnormal score module 303 is responsible for calculating the abnormal score according to the probability density of each window, and measuring the abnormality of each window according to the size of the abnormal score. To determine the normal degree, during the abnormal period, the site concentration acquisition unit 4 establishes a Gaussian diffusion model for the suspected pollution source in the tracing area. First, coordinate transformation is performed and then the Gaussian diffusion model is used to calculate the concentration value of the pollutants released by the pollution source diffused to each site location in the tracing area. The coordinate change is used to adapt to the changes in wind speed and wind direction. The Gaussian diffusion model can more easily approach the pollution source. The pollution tracing unit 5 combines the concentration value of each site location to eliminate the sites that are not greatly affected by the pollution. According to the changes in the site concentration value, the drone is used to continuously adjust the travel length to search for the pollution source in real time, which is used to realize mobile monitoring of the tracing area and locate the pollution source.
[0085] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.
[0086] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A big data-based atmospheric pollutant tracing system, characterized in that: include: A meteorological feature fusion unit (1), wherein the meteorological feature fusion unit (1) extracts meteorological features of n stations from meteorological data of multiple dimensions of temperature, humidity, wind direction and wind speed by combining a convolutional neural network with a GRU, and divides the meteorological features into multiple segments by a sliding window, wherein each sliding window contains n stations, and is used to dynamically learn the correlation between the n stations; A spatiotemporal feature extraction unit (2), wherein the spatiotemporal feature extraction unit (2) extracts time features from the time series data of each station in each sliding window to capture time correlation, obtains spatial features from the meteorological feature fusion unit (1), adaptively learns the correlation between each station, and constructs a dynamic graph by combining the time features and the spatial features; An abnormality locking unit (3), wherein the abnormality locking unit (3) obtains the spatiotemporal information composed of time features and space features from the spatiotemporal feature extraction unit (2), and then calculates the probability density of the spatiotemporal information of each sliding window, so as to detect abnormal conditions of atmospheric pollutant concentrations in the city; The abnormal locking unit (3) includes a graph convolution module (301) and a normalized flow module (302). The graph convolution module (301) aggregates the trained attention matrix and time features, and captures the spatial information and time information of each site neighborhood, thereby obtaining the spatiotemporal information of each sliding window. The graph convolution layer is used to aggregate the hidden states of n sites in a sliding window, and the spatiotemporal information in the sliding window period is marked with features. The normalized flow module (302) obtains the feature marks of the input spatiotemporal information from the graph convolution module (301), thereby performing a more detailed density estimation on the spatiotemporal information of each window, calculating the probability density mapped to the target distribution function, and optimizing the parameters of the entire network using maximum likelihood estimation. The abnormality locking unit (3) further comprises an abnormality score module (303), wherein the abnormality score module (303) is responsible for calculating an abnormality score according to the probability density of each window, wherein the magnitude of the abnormality score is used to measure the degree of abnormality of each window, and an abnormality score is calculated using an abnormality determination formula, and a threshold is set using an interquartile range, and a window whose abnormality score exceeds the threshold is determined to be an abnormal situation; A site concentration acquisition unit (4), wherein the site concentration acquisition unit (4) establishes a Gaussian diffusion model for a suspected pollution source in the traceability area during an abnormal period, first performs coordinate transformation, and then uses the Gaussian diffusion model to calculate the concentration value of pollutants released by the pollution source diffused to each site position in the traceability area; The pollution tracing unit (5) combines the concentration value of each site location, eliminates sites that are not greatly affected by the pollution, and uses the drone to continuously adjust the travel length to search for pollution sources in real time according to the changes in the site concentration values, so as to realize mobile monitoring of the tracing area and locate the pollution source.
2. The big data-based atmospheric pollutant tracing system according to claim 1 is characterized by: The meteorological feature fusion unit (1) divides the meteorological features into multiple segments by means of a sliding window, and inputs the meteorological data of multiple dimensions of temperature, humidity, wind direction and wind speed into a convolution layer, thereby extracting meteorological features, and fuses the meteorological features of n stations with the monitoring data of the n stations in the corresponding sliding window, and inputs the features into the spatiotemporal feature extraction unit (2), so as to dynamically learn the correlation between the n stations.
3. The big data-based atmospheric pollutant tracing system according to claim 1 is characterized by: The spatiotemporal feature extraction unit (2) comprises a time feature extraction module (201) and a space feature extraction module (202). The time feature extraction module (201) uses a recurrent neural network to extract historical information of a period of t before the current station and historical information of a station related to the current station during a period of t-1 before the current station, and normalizes the extracted historical information. Time features are extracted from the time series data of each station in each sliding window according to the historical information. The space feature extraction module (202) obtains spatial features from the meteorological feature fusion unit (1), and combines the meteorological features to model the dynamic interdependence between the n stations using a graph attention mechanism. A dynamic graph is then constructed by combining the time features and the space features. Each node in the dynamic graph represents a monitoring point. The attention score of each node is obtained through the graph attention mechanism. Each edge represents the relationship between two stations. The weight of the edge is used to quantify the correlation. The attention score of each node forms an attention matrix, and is trained.
4. The big data-based atmospheric pollutant tracing system according to claim 1 is characterized by: In the site concentration acquisition unit (4), the Gaussian diffusion model is specifically: ; in, Indicates the pollution source in the plane space Spread to coordinates The concentration value at Indicates wind speed, Indicates source strength, express The diffusion coefficient on the axis, express The diffusion coefficient on the axis, and They represent the plane coordinate axes, represents the natural exponential function, Represents a natural constant.
5. The big data-based atmospheric pollutant tracing system according to claim 1 is characterized by: The pollution tracing unit (5) comprises a pollution source search module (501), a concentration mean acquisition module (502) and a path adjustment module (503). The pollution source search module (501) is provided with air pollutant sensors in four directions, namely, front, back, left, and right, of the UAV in sequence, and uses the sensors to timely obtain the concentration values of air pollution in the four directions. The concentration mean acquisition module (502) enables the UAV to adjust its step length in real time. When the distance between the UAV and the pollution source is far, the step length becomes larger, and vice versa. The concentration values of the four sensors are averaged in combination with the UAV's step length to obtain the concentration mean. The path adjustment module (503) compares the concentration mean with a preset concentration critical value to determine the UAV's route, trace the pollution source location, and hover over the source.
6. The big data-based atmospheric pollutant source tracing system according to claim 5 is characterized by: The steps of determining the route of the drone in the behavior adjustment module (503) are as follows: S1. When the mean concentration is less than or equal to the critical concentration value, the UAV updates its position in a Z-shaped variable step length and performs a global search; S2. When the mean concentration is greater than the critical concentration value, the UAV continues to move in a straight line toward the maximum concentration.
