A VOCs pollution source tracing method and device based on a convolutional neural network and a space-time-component coupled image, and a storage medium

Abstract

The application provides a VOCs pollution source tracing method and device based on a convolutional neural network and a space-time-component coupled image and a storage medium. It relates to the technical field of atmospheric pollution. The method comprises a data set of sailing position coordinates, VOC species and concentrations; a multi-channel three-dimensional image tensor is constructed; a CNN model with global average pooling is used to extract feature information; an image data information is obtained by using a class activation map technology; the image data information is preprocessed to obtain activation intensity, which is used as a prior constraint to construct a constrained objective function model; the objective function is solved to obtain a source component spectrum matrix and a source contribution matrix. Through the application, the chemical fingerprint extracted by deep learning is used as a physical constraint, effectively solving the rotation ambiguity problem of the traditional PMF model solution, enabling dynamic inversion of the pollution source component spectrum changing over time, and realizing high-precision analysis and spatial positioning of the pollution source.

Description

Technical Field

[0001] This invention relates to the field of air pollution technology, and in particular to a method, apparatus and storage medium for tracing VOCs pollution sources based on convolutional neural networks and spatiotemporal-component (species) coupled images. Background Technology

[0002] Volatile organic compounds (VOCs) are a collective term for organic compounds that have high vapor pressure and are easily volatile at room temperature. They possess diverse chemical structures and mainly include 12 categories: alkanes, alkenes, alkynes, benzene compounds, alcohols, aldehydes, ketones, and esters. Source tracing of VOC pollution relies on various technologies, among which mobile monitoring is the most widely used. This technology uses mobile monitoring vehicles (such as mobile monitoring vehicles) to continuously collect atmospheric pollutant data while in motion, and uses mass spectrometry analysis to quickly identify VOC types and concentrations, creating real-time pollutant distribution maps, thereby accurately pinpointing areas of abnormally high values ​​and potential pollution sources. Positive definite matrix factorization (PMF) is a widely used statistical method for atmospheric pollutant source apportionment. By analyzing the pollutant concentration data matrix, it identifies and quantifies the contributions of different pollution sources. Its core idea is to decompose observed data into the product of unknown source components and contributing factors, combined with uncertainty analysis, to provide a quantitative assessment of source contributions.

[0003] However, methods relying on traditional PMF models for tracing origins generally suffer from the following problems:

[0004] 1. Source spectrum libraries suffer from spatiotemporal lag and model dependence on static source spectra and assumptions: Existing source spectrum libraries are typically constructed based on laboratory measurements or historical literature data, and are therefore "static libraries." The core assumption of traditional PMF models is also based on the premise that the chemical composition spectrum of pollution sources remains unchanged throughout the monitoring period. However, in actual industrial parks, the actual emission characteristics of pollution sources change in real time due to adjustments in production processes, changes in raw materials, or abnormal operating conditions; static spectrum libraries cannot characterize the current true emission situation, leading to biased analysis results.

[0005] 2. Source composition spectrum construction is difficult and time-consuming: Traditional methods for constructing localized source spectrum libraries require manual entry into the pollution source for sampling and analysis, which is not only costly and dangerous, but also cannot be updated frequently and is difficult to capture transient emission characteristics.

[0006] 3. The model solution suffers from rotational ambiguity and subjectivity: The mathematical solution of the PMF model is not unique and suffers from "rotational ambiguity," meaning that multiple mathematically equally valid solutions exist. The final selection of the number of factors, their physical interpretation, and naming largely depend on the professional experience and subjective judgment of the analysts, leading to insufficient objectivity and reproducibility of the results.

[0007] 4. Insufficient structured utilization of high-dimensional data: Mobile monitoring data has high spatiotemporal resolution, but traditional methods treat it as a simple time series, ignoring the covariant structure of different chemical components in space (i.e., “visual” features), resulting in the loss of key spatial correlation information when constructing source spectra.

[0008] In view of the above problems, it is necessary to provide a method for tracing VOC pollution sources based on convolutional neural networks and spatiotemporal-component coupled images. Summary of the Invention

[0009] To overcome the problems of traditional PMF model source tracing methods in the prior art, this invention provides a VOCs pollution source tracing method based on convolutional neural networks and spatiotemporal-component coupled images. By using chemical fingerprints extracted by deep learning as physical constraints, it effectively solves the rotational ambiguity problem of traditional PMF model solutions, and can dynamically invert the pollution source composition spectrum that changes over time, thus achieving high-precision analysis and spatial positioning of pollution sources.

[0010] To achieve the above objectives, according to one aspect of this application, a method for tracing VOC pollution sources based on convolutional neural networks and spatiotemporal-component coupled images is provided, the method comprising:

[0011] S201: Mobile data collection dataset, including: mobile location coordinates data, VOC species name data and corresponding species concentration data;

[0012] S202: Convert the dataset into an image dataset and construct a multi-channel three-dimensional image tensor of H × W × N; where H is the longitude data of the navigation position coordinates, W is the latitude data of the navigation position coordinates, and N is the number of VOC species.

[0013] S203: Construct a global average pooling convolutional neural network model to extract feature information of key VOC species spectrum related to pollution source categories from the image dataset; then use class activation map technology to obtain image data information containing the activation intensity of key VOC species spectrum.

[0014] S204: Preprocess the image data information to obtain the activation intensity in the spatiotemporal and species dimensions; use the activation intensity as a priori constraint, and solve the positive matrix factorization objective function of the convolutional neural network-class activation graph constraint by applying regularization to the source component spectrum matrix in the positive matrix factorization objective function to obtain the source component spectrum matrix and the source contribution matrix.

[0015] Furthermore, in S201, a chemical ionization source mass spectrometer is used for navigation; the time interval is 4-5 seconds.

[0016] Further, in S202, the dataset is converted into an image dataset, the method including:

[0017] S202-1 combines longitude and latitude data into H × W spatial grid cells;

[0018] S202-2, calculate the ratio of the concentration of each VOC species in each spatial grid cell to the total VOC concentration in that spatial grid cell, and obtain an image dataset containing VOC weight percentages;

[0019] The formula for calculating the VOC weight percentage is as follows:

[0020] (1);

[0021] In equation (1), j represents the index of the VOC species, corresponding to a specific channel in the image; u represents the index of the spatial grid cell, corresponding to a specific pixel position in the image; It is the first VOC-like substances in the first Concentration values ​​within a spatiotemporal grid unit; yes The total concentration of all VOCs within the corresponding spatiotemporal grid unit; It is the first VOC-like substances in the first Weight percentage within a spatiotemporal grid unit.

[0022] Furthermore, in S203, the global average pooling convolutional neural network model uses the multi-channel three-dimensional image tensor described in S202 as input data, pollution source categories, and corresponding key VOC species spectra as outputs to construct the model.

[0023] Furthermore, in S203, the calculation formula for the class activation graph is as follows:

[0024] (2);

[0025] In equation (2), An index representing the category of pollution sources; Indicates the corresponding to the first The spatial coordinates of each grid cell; Indicates the first Class activation graph of pollution source categories; This represents the feature map channel index of the last convolutional layer in a convolutional neural network; : indicates the first The feature map is used to identify the first feature map. Weighted contribution of different types of pollution sources; : indicates the first Each feature map in spatial location The activation value at that location.

[0026] Furthermore, in S203, the calculation formula for the activation intensity of key VOC species is as follows:

[0027] (3);

[0028] In equation (3), An index representing the category of pollution sources; : Indicates an index of VOC species; : indicates the calculated result regarding the first The first category of pollution sources The activation intensity of a chemical species; : Represents an aggregation function that simplifies a two-dimensional matrix into a single scalar value; The formula (2) calculates the first... Two-dimensional activation graph of a pollution source; This represents the element-wise multiplication operator; : indicates that the first digit obtained by formula (1) A two-dimensional image matrix for each species channel.

[0029] Furthermore, in S204, the objective function of positive matrix factorization under convolutional neural network-class activation graph constraints, i.e., the PMF function under CNN-CAM constraints, is as follows:

[0030] (4)

[0031] In equation (4), This represents the index of the sample collected during mobile data collection, i=1...n; This represents the VOC species index, j=1...m; This represents the index of VOCs pollution source factors, k=1...p; This indicates that the PMF algorithm seeks to minimize the constraint objective function; This represents the observed concentration matrix of the j-th VOC species in the i-th sample. This represents the source contribution matrix element of the k-th VOCs pollution source in the i-th sample being solved; This represents the source component spectral matrix element of the j-th VOC species from the k-th VOC pollution source being solved; :and The corresponding uncertainty matrix; These represent the total number of samples, VOC species, and VOC pollution sources, respectively. This represents a hyperparameter used to control the constraint strength of the activation intensity; The first expression obtained by formula (3) The first VOCs pollution source The activation intensity of VOC species; : Indicates the activation intensity after normalization.

[0032] Furthermore, the aforementioned It is between 0.1 and 0.75.

[0033] Furthermore, this application also includes:

[0034] S205: Calculate the source component spectrum matrix and source contribution matrix for different monitoring periods from the data collected by mobile survey according to steps S202~S204, generate a series of source component spectrum matrices and source contribution matrices that evolve over time, and form a dynamic source component spectrum matrix and dynamic source contribution matrix.

[0035] Preferably, the monitoring period is in units of seconds, minutes, days, months, quarters, or years.

[0036] According to another aspect of this application, a VOCs pollution source tracing device based on convolutional neural networks and spatiotemporal-species coupled images is provided, the device comprising:

[0037] The data acquisition unit collects datasets, including: navigation location coordinates, VOC species names, and corresponding species concentrations.

[0038] The first computing unit converts the dataset from the data acquisition unit into an image dataset and constructs a multi-channel three-dimensional image tensor of size H × W × N; where H is the longitude data of the navigation position coordinates, W is the latitude data of the navigation position coordinates, and N is the number of VOC species.

[0039] The second computing unit constructs a global average pooling convolutional neural network model and extracts feature information of key VOC species spectrum related to pollution source categories from the image dataset of the first computing unit; then, it uses class activation map technology to obtain image data information containing the activation intensity of key VOC species spectrum.

[0040] The output unit preprocesses the image data information from the second calculation unit to obtain activation intensity weights in the species and spatiotemporal dimensions. Using these activation intensity weights as prior constraints, the source component spectrum matrix in the positive matrix factorization objective function is regularized to solve the positive matrix factorization objective function of the convolutional neural network-class activation graph constraint, thereby obtaining the source component spectrum matrix and the source contribution matrix.

[0041] Furthermore, it also includes a display unit that receives the source component spectrum matrix and source contribution matrix from the output unit for different monitoring periods, generates a series of source component spectrum matrices and source contribution matrices that evolve over time, and forms a dynamic source component spectrum matrix and a dynamic source contribution matrix.

[0042] According to another aspect of this application, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored executable program, wherein, when the executable program is executed, it controls the device where the computer-readable storage medium is located to perform the above-described VOCs pollution source tracing method.

[0043] According to another aspect of this application, an electronic device is also provided, comprising:

[0044] Memory, which stores executable programs;

[0045] A processor is used to run the program, wherein the program executes the VOCs pollution source tracing method described above when it runs.

[0046] According to another aspect of this application, a computer program product is also provided, including computer instructions, characterized in that the computer instructions, when executed by a processor, implement the above-described VOCs pollution source tracing method.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] (1) This application utilizes an innovative data reconstruction method to “translate” high spatiotemporal resolution mobile monitoring data into a “chemical image” that can be understood by computer vision algorithms; then, a convolutional neural network (CNN) specially designed for scientific interpretability is used to “read” these images and automatically extract objective “chemical fingerprints” representing different pollution sources; finally, this data-driven “chemical fingerprint” is used as a powerful physical prior knowledge to inject into and constrain the classic PMF source apportionment model, thereby fundamentally solving its inherent mathematical fuzziness problem and ultimately generating pollution source apportionment results that are dynamically changing over time, have clear physical meaning, and can be accurately traced.

[0049] (2) This application breaks through the limitations of traditional manual sampling to construct spectral libraries. It uses mobile monitoring data and deep learning algorithms to reverse deduce and reconstruct the real-time chemical composition spectrum of pollution sources from environmental receptor data, and dynamically constructs a localized source spectrum library.

[0050] (3) This application proposes a data construction method of “spatiotemporal-component coupling image”, which transforms discrete monitoring data into chemical images containing spatial texture, enabling the model to accurately extract source component spectra by recognizing “texture features”.

[0051] (4) This application solves the rotational ambiguity problem of the PMF model by constructing a spatiotemporal-component coupled image and using a source spectrum library constructed in real time as a physical constraint, thereby achieving high-precision analysis and spatial positioning of pollution sources. Attached Figure Description

[0052] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0053] Figure 1 A hardware block diagram of a computer terminal for implementing a VOCs pollution source tracing method based on convolutional neural networks and spatiotemporal-component coupled images is shown.

[0054] Figure 2 This is a flowchart of an optional VOCs pollution source tracing method based on convolutional neural networks and spatiotemporal-component coupled images, according to Embodiment 1 of this application.

[0055] Figure 3 This is a schematic diagram of an optional VOCs pollution source tracing device based on a convolutional neural network and a spatiotemporal-component coupled image, according to Embodiment 1 of this application.

[0056] Figure 4 This is an application of the regression analysis of the prediction accuracy of the convolutional neural network-class activation graph constrained positive matrix factorization objective function model involved in Example 1.

[0057] Figure 5 This is a graph showing the comparison of source contribution rates for different methods involved in Application Example 1. Detailed Implementation

[0058] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0059] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0060] Example 1

[0061] According to an embodiment of this application, an embodiment of a VOCs pollution source tracing method based on a convolutional neural network and a spatiotemporal-component coupled image is also provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0062] The method embodiment provided in Embodiment 1 of this application can be executed on a mobile terminal, computer terminal, or similar computing device. Figure 1 This diagram illustrates the hardware structure of a computer terminal (or mobile device) for a VOCs pollution source tracing method based on convolutional neural networks and spatiotemporal-species coupled images. Figure 1 As shown, the computer terminal 10 (or mobile device) may include one or more processors 102 (shown as 102a, 102b, ..., 102n in the figure) 102 (processor 102 may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.), a memory 104 for storing data, and a transmission device 106 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of a BUS bus), a network interface, a power supply, and / or a camera. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, computer terminal 10 may also include... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0063] It should be noted that the aforementioned one or more processors 102 and / or other data processing circuits are generally referred to herein as "data processing circuits". These data processing circuits may be embodied, in whole or in part, in software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuits may be a single, independent processing module, or may be integrated, in whole or in part, into any other element within the computer terminal 10 (or mobile device). As involved in the embodiments of this application, the data processing circuits serve as a processor control mechanism (e.g., selection of a variable resistor termination path connected to an interface).

[0064] The memory 104 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the VOCs pollution source tracing method in this embodiment. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, thereby realizing the aforementioned VOCs pollution source tracing method. The memory 104 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include memory remotely located relative to the processor 102, and these remote memories can be connected to the computer terminal 10 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0065] The transmission device 106 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of the computer terminal 10. In one example, the transmission device 106 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 106 may be a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0066] The display may be, for example, a touchscreen liquid crystal display (LCD) that allows the user to interact with the user interface of the computer terminal 10 (or mobile device).

[0067] Under the aforementioned operating environment, this application provides the following: Figure 2 The method shown is a VOCs pollution source tracing method based on convolutional neural networks and spatiotemporal-component coupled images. Figure 2 This is a flowchart of a VOCs pollution source tracing method based on convolutional neural networks and spatiotemporal-component coupled images, according to Embodiment 1 of this application.

[0068] S201: Mobile data collection dataset, including: mobile location coordinates, VOC species names, and corresponding species concentrations.

[0069] A chemical ionization source mass spectrometer can be used for the cruise test; the time interval is 4 to 5 seconds.

[0070] Chemical ionization source mass spectrometers offer millisecond-level response times, require no sample pretreatment, and can meet the real-time capture needs of mobile monitoring vehicles for transient pollution plumes during operation. Setting a 4-5 second time interval is to match the typical travel speed of the mobile monitoring vehicle (20-30 km / h), ensuring the spatial resolution of the data is controlled at approximately 20-30 meters, thus avoiding the omission of small-scale pollution source emission characteristics due to excessively large sampling intervals. For example, a mobile monitoring vehicle continuously monitored along the main roads of a chemical industrial park at a speed of 20 km / h, with a sampling interval of 5 seconds, collecting data continuously for 2 hours, obtaining a total of 1440 sets of raw datasets containing monitoring time, GPS latitude and longitude coordinates, and concentrations of 50 target VOC species.

[0071] S202: Convert the dataset into an image dataset and construct a multi-channel three-dimensional image tensor of H × W × N; where H is the longitude data of the navigation position coordinates, W is the latitude data of the navigation position coordinates, and N is the number of VOC species.

[0072] The method for converting the dataset into an image dataset includes:

[0073] S202-1 combines monitoring time data and mobile position coordinate data into H × W spatial grid cells;

[0074] S202-2, calculate the ratio of the concentration of each VOC species in each spatial grid cell to the total VOC concentration in that spatial grid cell, and obtain an image dataset containing VOC weight percentages;

[0075] The formula for calculating the VOC weight percentage is as follows:

[0076] (1);

[0077] In equation (1), j represents the index of the VOC species, corresponding to a specific channel in the image; u represents the index of the spatial grid cell, corresponding to a specific pixel position in the image; It is the first VOC-like substances in the first Concentration values ​​within a spatiotemporal grid unit; yes The total concentration of all VOCs within the corresponding spatiotemporal grid unit; It is the first VOC-like substances in the first Weight percentage within a spatiotemporal grid unit.

[0078] S203: Construct a global average pooling convolutional neural network model to extract feature information of key VOC species spectrum related to pollution source categories from the image dataset; then use class activation map technology to obtain image data information containing the activation intensity of key VOC species spectrum.

[0079] The global average pooling convolutional neural network model uses the S202 multi-channel 3D image tensor as input data and the corresponding key VOC species spectrum as output data to construct the model. The pollution source categories and corresponding key VOC species spectra required for model construction can be obtained by measuring data collected by instruments in polluted industrial parks, or by analyzing sample data collected on-site at typical pollution source emission outlets, or by simulation datasets generated based on publicly available standard source spectrum libraries (such as the SPECIATE library) combined with Gaussian diffusion models.

[0080] The model can use ResNet or VGG backbone networks. The number of output channels of the last convolutional layer is equal to the preset number of pollution source categories. The last convolutional layer is followed by a global average pooling (GAP) layer to preserve spatial structure information. The loss function can be the cross-entropy loss function. The optimizer can use the Adam algorithm and output classification results for different pollution source categories.

[0081] In S203, the calculation formula for the class activation graph is as follows:

[0082] (2);

[0083] In equation (2), An index representing the category of pollution sources; Indicates the corresponding to the first The spatial coordinates of each grid cell; Indicates the first Class activation graph of pollution source categories; This represents the feature map channel index of the last convolutional layer in a convolutional neural network; : indicates the first The feature map is used to identify the first feature map. Weighted contribution of different types of pollution sources; : indicates the first Each feature map in spatial location The activation value at that location.

[0084] The calculation formula for the activation intensity of key VOC species in S203 is as follows:

[0085] (3);

[0086] In equation (3), An index representing the category of pollution sources; : Indicates an index of VOC species; : indicates the calculated result regarding the first The first category of pollution sources The activation intensity of a chemical species; : Represents an aggregation function that simplifies a two-dimensional matrix into a single scalar value; The formula (2) calculates the first... Two-dimensional activation graph of a pollution source; This represents the element-wise multiplication operator; : indicates that the first digit obtained by formula (1) A two-dimensional image matrix for each species channel.

[0087] S204: Preprocess the image data information to obtain the activation intensity in the spatiotemporal and species dimensions; use the activation intensity as a priori constraint, and solve the positive matrix factorization objective function of the convolutional neural network-class activation graph constraint by applying regularization to the source component spectrum matrix in the positive matrix factorization objective function to obtain the source component spectrum matrix (referred to as F matrix) and the source contribution matrix (referred to as G matrix).

[0088] The objective function for positive matrix factorization under convolutional neural network-class activation map constraints, i.e., the PMF function under CNN-CAM constraints, is as follows:

[0089] (4)

[0090] In equation (4), This represents the index of the sample collected during mobile data collection, i=1...n; This represents the VOC species index, j=1...m; This represents the index of VOCs pollution source factors, k=1...p; This indicates that the PMF algorithm seeks to minimize the constraint objective function; This represents the observed concentration matrix of the j-th VOC species in the i-th sample. This represents the source contribution matrix element of the k-th VOCs pollution source in the i-th sample being solved; This represents the source component spectral matrix element of the j-th VOC species from the k-th VOC pollution source being solved; :and The corresponding uncertainty matrix; These represent the total number of samples, VOC species, and VOC pollution sources, respectively. This represents a hyperparameter used to control the constraint strength of the activation intensity; The first expression obtained by formula (3) The first VOCs pollution source The activation intensity of VOC species; : Represents the normalized activation strength. Hyperparameter It can be set between 0.1 and 0.75 to balance the residuals of the observed data with the prior fingerprint constraints and prevent the model from relying too much on prior information.

[0091] S205: Calculate the source component spectrum matrix and source contribution matrix for different monitoring periods using the data collected from mobile surveys, following steps S202 to S204. Generate a series of source component spectrum matrices and source contribution matrices that evolve over time, forming a dynamic source component spectrum matrix and a dynamic source contribution matrix. The monitoring period can be in units of seconds, minutes, days, months, quarters, or years.

[0092] The PMF function constrained by CNN-CAM is embedded in the core of PMF solving algorithms (such as the least squares iterative algorithm). It obtains the source component spectrum matrix and source contribution matrix by solving for the minimum value. During the calculation, the computer continuously adjusts the values ​​of the source component spectrum matrix and source contribution matrix. Each time it is adjusted, it substitutes them into this formula again until the minimum value is obtained.

[0093] After repeated iterative optimization by the computer, the source component spectrum matrix and the source contribution matrix are finally output, which are the two core tables.

[0094] Among them, the source composition matrix is ​​a composition spectrum with clear chemical characteristics after physical constraint correction (for example, clearly indicating "this is a coating source, with a benzene content of 80%)); the source contribution matrix is ​​the concentration value of each pollution source on each spatiotemporal network unit.

[0095] Based on this, this embodiment can also divide the entire monitoring period into multiple sub-periods and apply the above process independently to the data of each sub-period to generate a series of source component spectra that evolve over time, thereby forming a "dynamic source component spectrum" that reflects the real-time changes in the emission characteristics of pollution sources.

[0096] Furthermore, by associating and matching the source contribution matrix (i.e., the G matrix, the source contribution values ​​of each spatiotemporal network unit) output by the PMF function constrained by CNN-CAM with the mobile location coordinate data (such as GPS coordinates) collected in the first step, a high-resolution spatial distribution map of the contribution of each pollution source can be drawn in the Geographic Information System (GIS), thereby intuitively locating pollution hotspots and tracing them back to specific emitting enterprises or facilities, thus achieving precise source tracing.

[0097] Sources 1-3 represent a pollution source (such as painting, transportation, chemical industry, etc.); The numerical value represents the contribution of that pollution source to the total VOCs concentration at that time and location (usually in ppb or...). ).

[0098] Taking the following source contribution matrix as an example, assuming that the park monitors 3 spatiotemporal network units, the model identifies 3 pollution sources (source 1: painting workshop, source 2: traffic exhaust, source 3: biomass combustion).

[0099] Calculated source contribution matrix as follows:

[0100] Spatiotemporal Network Unit ID Time GPS coordinates (Location) Source 1 contribution value (painting) Source 2 contribution value (transportation) Source 3 contribution value (combustion) 1 10:00:01 (116.35, 39.98) 85.4 12.1 5.3 2 10:00:05 (116.36, 39.98) 40.2 60.5 4.8 3 10:00:09 (116.37, 39.99) 10.1 15.3 2.1

[0101] The matrix is ​​interpreted as:

[0102] Spatiotemporal Network Unit 1: At 10:00:01, most of the VOCs in the air come from the painting workshop (85.4), indicating that the monitoring vehicle may be passing by the entrance of the painting plant.

[0103] Spatiotemporal Network Unit 2: After 4 seconds, the contribution of the paint source decreased, while the contribution of the traffic source surged (60.5), indicating that the car may have driven to the intersection of the main road.

[0104] Data application: By using the GPS coordinates in this table as points on the map and the contribution values ​​of source 1 (85.4, 40.2,...) as the shades of color on the map, we can obtain the "Pollution Distribution Map of the Painting Workshop".

[0105] Example 2

[0106] This application also provides a VOCs pollution source tracing device based on convolutional neural networks and spatiotemporal-species coupled images. It should be noted that the VOCs pollution source tracing device of this application can be used to execute the VOCs pollution source tracing device provided in this application. The following describes the VOCs pollution source tracing device provided in this application.

[0107] According to embodiments of this application, a VOCs pollution source tracing device based on convolutional neural networks and spatiotemporal-species coupled images is also provided, such as... Figure 3 As shown, the device includes: a data acquisition unit 301, a first calculation unit 302, a second calculation unit 303, an output unit 304, and a display unit 305.

[0108] Specifically, the data acquisition unit 301 collects a dataset, including: navigation position coordinate data, VOC species name data, and corresponding species concentration data.

[0109] A chemical ionization source mass spectrometer can be used for mobile monitoring, with a time interval of 4-5 seconds. Chemical ionization source mass spectrometers have millisecond-level response speeds, require no sample pretreatment, and can meet the real-time capture requirements of transient pollution plumes during mobile monitoring. Setting the time interval to 4-5 seconds is to match the typical travel speed of the mobile monitoring vehicle (20-30 km / h), ensuring the spatial resolution of the data is controlled at approximately 20-30 meters, thus avoiding the omission of small-scale pollution source emission characteristics due to excessively large sampling intervals. For example, a mobile monitoring vehicle continuously monitors along the main roads of a chemical industrial park at a speed of 20 km / h, with a sampling interval of 5 seconds, collecting data continuously for 2 hours, obtaining a total of 1440 sets of raw datasets containing monitoring time, GPS latitude and longitude coordinates, and concentrations of 50 target VOC species.

[0110] The first computing unit 302 converts the dataset from the data acquisition unit into an image dataset and constructs a multi-channel three-dimensional image tensor of H x W x N; where H is the longitude data of the navigation position coordinates, W is the latitude data of the navigation position coordinates, and N is the number of VOC species.

[0111] The method for converting the dataset into an image dataset using the first computing unit 302 includes:

[0112] S202-1 combines monitoring time data and mobile position coordinate data into H × W spatial grid cells;

[0113] S202-2, calculate the ratio of the concentration of each VOC species in each spatial grid cell to the total VOC concentration in that spatial grid cell, and obtain an image dataset containing VOC weight percentages;

[0114] The formula for calculating the VOC weight percentage is as follows:

[0115] (1);

[0116] In equation (1), j represents the index of the VOC species, corresponding to a specific channel (range) in the image. ); u represents the index of the u-th spatial grid cell (corresponding to an image pixel, range) ), which corresponds to a specific pixel location in the image; It is the first VOC-like substances in the first Concentration values ​​within a spatiotemporal grid unit; yes The total concentration of all VOCs within the corresponding spatiotemporal grid unit; It is the first VOC-like substances in the first Weight percentage within a spatiotemporal grid unit.

[0117] The second computing unit 303 constructs a global average pooling convolutional neural network model and extracts feature information of key VOC species spectrum related to pollution source categories from the image dataset of the first computing unit; then, it uses class activation map technology to obtain image data information containing the activation intensity of key VOC species spectrum.

[0118] The global average pooling convolutional neural network model uses the multi-channel three-dimensional image tensor obtained by the first computing unit 302 as input, and the data pollution source category and the corresponding key VOC species spectrum as output to construct the model.

[0119] The formula for calculating the class activation graph is as follows:

[0120] (2);

[0121] In equation (2), express Indicates the corresponding to the first The spatial coordinates of each grid cell; Indicates the first Class activation graph of pollution source categories; This represents the feature map channel index of the last convolutional layer in a convolutional neural network; : indicates the first The feature map is used to identify the first feature map. Weighted contribution of different types of pollution sources; : indicates the first Each feature map in spatial location The activation value at that location.

[0122] The formula for calculating the activation intensity of key VOC species is as follows:

[0123] (3);

[0124] In equation (3), express :express : indicates the calculated result regarding the first The first category of pollution sources The activation intensity of a chemical species (i.e., prior fingerprint). : Represents an aggregation function (such as global average pooling, mean, median, variance) that simplifies a two-dimensional matrix into a single scalar value; The formula (2) calculates the first... Two-dimensional activation graph of a pollution source; This represents the element-wise product (Hadamard product) operator; : indicates that the first digit obtained by formula (1) A two-dimensional image matrix for each species channel.

[0125] The output unit 304 preprocesses the image data information of the second calculation unit to obtain the activation intensity weights in the species dimension and the spatiotemporal dimension; using the activation intensity weights as prior constraints, it solves the positive matrix factorization objective function of the convolutional neural network-class activation graph constraint by applying regularization to the source component spectrum matrix in the positive matrix factorization objective function, and obtains the source component spectrum matrix and the source contribution matrix.

[0126] The objective function for positive matrix factorization under convolutional neural network-class activation map constraints, i.e., the PMF function under CNN-CAM constraints, is as follows:

[0127] (4)

[0128] In equation (4), This represents the index of the sample collected during mobile data collection, i=1...n; This represents the VOC species index, j=1...m; This represents the index of VOCs pollution source factors, k=1...p; This indicates that the PMF algorithm seeks to minimize the constraint objective function; This represents the observed concentration matrix of the j-th VOC species in the i-th sample. This represents the source contribution matrix element of the k-th VOCs pollution source in the i-th sample being solved; This represents the source component spectral matrix element of the j-th VOC species from the k-th VOC pollution source being solved; :and The corresponding uncertainty matrix; These represent the total number of samples, VOC species, and VOC pollution sources, respectively. This represents a hyperparameter used to control the constraint strength of the activation intensity; The first expression obtained by formula (3) The first VOCs pollution source The activation intensity of VOC species; : Represents the normalized activation strength. Among them, the hyperparameters... It can be between 0.1 and 0.75.

[0129] The display unit 305 receives source component spectrum matrices and source contribution matrices from the output unit for different monitoring periods, and generates a series of source component spectrum matrices and source contribution matrices that evolve over time, forming dynamic source component spectrum matrices and dynamic source contribution matrices. The monitoring periods are measured in seconds, minutes, days, months, quarters, or years.

[0130] Application Example 1

[0131] This application example selects the area surrounding a typical chemical industrial park in eastern China as the research object to verify the effectiveness of the VOCs pollution source tracing method based on convolutional neural networks and spatiotemporal-component coupled images proposed in this application.

[0132] Data Acquisition and Preprocessing

[0133] The mobile monitoring vehicle equipped with a single-photon ionization time-of-flight mass spectrometer (SPI-MS) was used for regional mobile monitoring. The mobile speed was controlled at about 20 km / h and the sampling time interval was set to 5 seconds. A total of 3,600 valid observation data points were collected, covering 50 typical VOC species (including alkanes, alkenes, aromatic hydrocarbons and oxygen-containing volatile organic compounds).

[0134] Based on the data transformation method proposed in this application, the monitoring area is divided into 100×100 spatial grid units; for each grid, the concentration ratio of 50 VOC species is calculated, and a multi-channel three-dimensional image tensor (H×W×N) with a size of 100×100×50 is constructed; this tensor preserves the spatial texture features and correlations of VOC components.

[0135] CNN Feature Extraction and CAM Constraint Construction

[0136] The constructed 3D image tensor is input into a pre-built convolutional neural network model (see Example 1 for details). This model uses ResNet as the backbone network for feature extraction, and connects a global average pooling (GAP) layer after the last convolutional layer to preserve spatial structure information and output classification results for different pollution source categories. In this application example, five pollution sources are set: industrial coating source, chemical production process source, storage tank emission source, motor vehicle exhaust source, and biomass combustion source.

[0137] After the model training is completed, the spatial response for different pollution source categories is calculated according to formula (2) using the principle of Class Activation Map (CAM) technology, and the activation intensity of the key VOC species spectrum is obtained by aggregation using formula (3). The activation intensity is normalized and used as a physical prior constraint, which is then introduced into the objective function of the Positive Matrix Factorization (PMF) model to construct the PMF objective function with CNN-CAM constraint as shown in formula (4). In this embodiment, the hyperparameter α is set to 0.5 to balance the data residuals and fingerprint constraints.

[0138] 3. Results Analysis and Verification

[0139] The objective function is solved using the method described in this application, resulting in optimized source component spectra and source contribution matrices, which are then compared with those obtained using the traditional PMF method (i.e., the PMF objective function without CNN-CAM constraints). Thanks to the introduced chemical fingerprint spatial constraint, the method in this application effectively overcomes the rotational ambiguity of the traditional PMF solution, achieving accurate analysis of pollution sources under complex emission scenarios.

[0140] The following focuses on verifying the model from two aspects: the accuracy of its predictions and the quantitative analysis of the source contribution rate.

[0141] (1) Analysis of model prediction accuracy ( Figure 4 analyze)

[0142] Figure 4 The results of linear regression analysis between the model-reconstructed concentration and the actual observed concentration are presented to evaluate the model's interpretability of the original data. In the figure, the horizontal axis represents the total VOCs concentration calculated by the model in Example 1 (Predicted), and the vertical axis represents the actual observed total VOCs concentration (Observed).

[0143] The circular dots represent the traditional PMF function, with a linear fit R² of only 0.65. The data points are relatively scattered, indicating that the traditional method has a large error when reconstructing complex environmental data. The triangular dots represent the CNN-CAM-PMF function in Example 1, with a linear fit R² improved to 0.93, and the vast majority of data points are closely distributed near the 1:1 reference line (the dashed line in the figure). This proves that the results obtained by the method in Example 1 of this invention are more consistent with the actual environmental characteristics and significantly improve the mathematical fitting accuracy of the model.

[0144] (2) Source contribution rate analysis deviation analysis Figure 5 analyze)

[0145] To verify the quantitative accuracy of source apportionment, the method in Example 1 of this invention introduces the analysis results of the Chemical Mass Balance Model (CMB) as a "true value" reference (because CMB is based on known source spectra, its results are often regarded in the academic community as the benchmark closest to the real situation).

[0146] Figure 5 The results of calculating the contribution rate of major pollution sources in the study area using three different methods are presented and compared.

[0147] observe Figure 5 It can be seen that the traditional PMF calculation results deviate significantly from the CMB benchmark in the two key industrial emission categories of "solvent use source (i.e., industrial coating source)" and "process emission source (i.e. chemical production process source)". This shows the limitations of traditional unsupervised methods in the absence of prior information.

[0148] In contrast, the contribution rate distribution of the method in Embodiment 1 of the present invention is highly consistent with the CMB benchmark for each pollution source, with deviations all controlled within 5%.

[0149] This further verifies that the "chemical fingerprint" extracted by deep learning in this application, as a key physical constraint, can effectively guide the model to converge to a more physically realistic solution, significantly improving the quantitative accuracy of source analysis.

Claims

1. A method for tracing VOCs pollution sources based on convolutional neural networks and spatiotemporal-component coupled images, characterized in that, include: S201: Mobile data collection dataset, including: mobile location coordinates data, VOC species name data and corresponding species concentration data; S202: Convert the dataset into an image dataset and construct a multi-channel three-dimensional image tensor of H × W × N; where H is the longitude data of the navigation position coordinates, W is the latitude data of the navigation position coordinates, and N is the number of VOC species. S203: Construct a global average pooling convolutional neural network model to extract feature information of key VOC species spectrum related to pollution source categories from the image dataset; then use class activation map technology to obtain image data information containing the activation intensity of key VOC species spectrum. S204: Preprocess the image data information to obtain the activation intensity in the spatiotemporal dimension and species dimension; use the activation intensity as a prior constraint, and solve the positive matrix factorization objective function of convolutional neural network-class activation graph constraint by applying regularization to the source component spectral matrix in the positive matrix factorization objective function to obtain the source component spectral matrix and the source contribution matrix. The objective function for positive matrix factorization constrained by the convolutional neural network-like activation graph is as follows: （4） In equation (4), This represents the index of the sample collected during mobile data collection, i=1...n; This represents the VOC species index, j=1...m; This represents the index of VOCs pollution source factors, k=1...p; This indicates that the PMF algorithm seeks to minimize the constraint objective function; This represents the observed concentration matrix of the j-th VOC species in the i-th sample. This represents the source contribution matrix element of the k-th VOCs pollution source in the i-th sample being solved; This represents the source component spectral matrix element of the j-th VOC species from the k-th VOC pollution source being solved; :and The corresponding uncertainty matrix; These represent the total number of samples, VOC species, and VOC pollution sources, respectively. This represents a hyperparameter used to control the constraint strength of the activation intensity; The first expression obtained by formula (3) The first VOCs pollution source The activation intensity of VOC species; : Indicates the activation intensity after normalization.

2. The VOCs pollution source tracing method based on convolutional neural networks and spatiotemporal-component coupled images as described in claim 1, characterized in that, In S201, a chemical ionization source mass spectrometer is used for mobile navigation; The time interval is 4-5 seconds; In S202, the dataset is converted into an image dataset, and the method includes: S202-1 combines longitude and latitude data into H × W spatial grid cells; S202-2, calculate the ratio of the concentration of each VOC species in each spatial grid cell to the total VOC concentration in that spatial grid cell, and obtain an image dataset containing VOC weight percentages; The formula for calculating the VOC weight percentage is as follows: （1）； In equation (1), j represents the index of the VOC species, corresponding to a specific channel in the image; u represents the index of the spatial grid cell, corresponding to a specific pixel position in the image; It is the first VOC-like substances in the first Concentration values ​​within a spatiotemporal grid unit; yes The total concentration of all VOCs within the corresponding spatiotemporal grid unit; It is the first VOC-like substances in the first Weight percentage within a spatiotemporal grid unit.

3. The VOCs pollution source tracing method based on convolutional neural networks and spatiotemporal-component coupled images as described in claim 2, characterized in that, In S203, the global average pooling convolutional neural network model uses the multi-channel three-dimensional image tensor described in S202 as input data, pollution source categories, and corresponding key VOC species spectra as outputs to construct the model. The formula for calculating the class activation graph is as follows: （2）； In equation (2), An index representing the category of pollution sources; Indicates the corresponding to the first The spatial coordinates of each grid cell; Indicates the first Class activation graph of pollution source categories; This represents the feature map channel index of the last convolutional layer in a convolutional neural network; : indicates the first The feature map is used to identify the first feature map. Weighted contribution of different types of pollution sources; : indicates the first Each feature map in spatial location The activation value at that location; The formula for calculating the activation intensity of key VOC species is as follows: （3）； In equation (3), An index representing the category of pollution sources; : Indicates an index of VOC species; : indicates the calculated result regarding the first The first category of pollution sources The activation intensity of a chemical species; : Represents an aggregation function that simplifies a two-dimensional matrix into a single scalar value; The formula (2) calculates the first... Two-dimensional activation graph of a pollution source; This represents the element-wise multiplication operator; : indicates that the first digit obtained by formula (1) A two-dimensional image matrix for each species channel.

4. The VOCs pollution source tracing method based on convolutional neural networks and spatiotemporal-component coupled images as described in claim 1, characterized in that, The It is between 0.1 and 0.

75.

5. The VOCs pollution source tracing method based on convolutional neural networks and spatiotemporal-component coupled images as described in claim 1, characterized in that, Also includes: S205: Calculate the source component spectrum matrix and source contribution matrix for different monitoring periods from the data collected by mobile survey according to steps S202~S204, generate a series of source component spectrum matrices and source contribution matrices that evolve over time, and form a dynamic source component spectrum matrix and dynamic source contribution matrix. The monitoring period is measured in seconds, minutes, days, months, quarters, or years.

6. A VOCs pollution source tracing device based on convolutional neural networks and spatiotemporal-component coupled images, employing the VOCs pollution source tracing method of claim 1, characterized in that, include: The data acquisition unit collects datasets, including: navigation location coordinates, VOC species names, and corresponding species concentrations. The first computing unit converts the dataset from the data acquisition unit into an image dataset and constructs a multi-channel three-dimensional image tensor of size H × W × N; where H is the longitude data of the navigation position coordinates, W is the latitude data of the navigation position coordinates, and N is the number of VOC species. The second computing unit constructs a global average pooling convolutional neural network model and extracts feature information of key VOC species spectrum related to pollution source categories from the image dataset of the first computing unit; then, it uses class activation map technology to obtain image data information containing the activation intensity of key VOC species spectrum. The output unit preprocesses the image data information from the second calculation unit to obtain the activation intensity in the species dimension and the spatiotemporal dimension; using this activation intensity as a prior constraint, the source component spectrum matrix and the source contribution matrix are obtained by applying regularization to the source component spectrum matrix in the positive matrix factorization objective function.

7. The VOCs pollution source tracing device as described in claim 6, characterized in that, It also includes a display unit that receives source component spectrum matrices and source contribution matrices from different monitoring periods from the output unit, generates a series of source component spectrum matrices and source contribution matrices that evolve over time, and forms dynamic source component spectrum matrices and dynamic source contribution matrices.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored executable program, wherein, when the executable program is executed, it controls the device on which the computer-readable storage medium is located to perform the VOCs pollution source tracing method according to any one of claims 1 to 5.

9. An electronic device, characterized in that, include: Memory, which stores executable programs; A processor for running the program, wherein the program executes the VOCs pollution source tracing method according to any one of claims 1 to 5 when it runs.

10. A computer program product comprising computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the VOCs pollution source tracing method according to any one of claims 1 to 5.

Images