A volatile organic pollutant online monitoring and early warning method and monitoring system

Abstract

The present application provides an online monitoring and early warning method and monitoring system for volatile organic pollutants, the method comprising: obtaining target VOCs monitoring data to be detected, extracting features from the target VOCs monitoring data, obtaining an implicit representation of the target VOCs monitoring data, obtaining a target monitoring object in the target VOCs monitoring data, determining a first involved monitoring object corresponding to the target monitoring object, determining a target focused implicit representation corresponding to the target monitoring object based on the implicit representation of the monitoring object of the first involved monitoring object and the corresponding focusing influence coefficient, implicitly integrating the implicit representation of the target VOCs monitoring data with the target focused implicit representation corresponding to the target monitoring object to obtain a target integrated implicit representation, and determining a VOCs detection result corresponding to the target VOCs monitoring data based on the target integrated implicit representation. The present application can improve the accuracy of VOCs detection results.

Description

Technical Field

[0001] The present application relates to the field of data processing, and in particular to an online monitoring and early warning method and a monitoring system for volatile organic pollutants. Background Art

[0002] Volatile organic pollutants (VOCs) are organic compounds with a saturated vapor pressure greater than 70 Pa at room temperature and a boiling point between 50-260°C at normal pressure. VOCs come from a wide range of sources, including industrial production processes (such as chemicals, coatings, printing, etc.), transportation, building decoration, agricultural activities, and the use of solvents in daily life.

[0003] VOCs have many serious hazards to the environment and human health. In view of the hazards of VOCs, accurate monitoring and early warning are of great significance. Effective monitoring and early warning can help relevant departments to timely grasp the emission of VOCs and take targeted pollution control measures to meet the increasingly stringent environmental quality standards and protect public health and ecological environment safety.

[0004] Traditional VOCs monitoring methods mainly include offline detection and online detection. They usually only focus on the VOCs concentration of a single monitoring point or a single pollution source, and lack comprehensive consideration of the relationship between VOCs emission sources in the entire region. In the actual environment, there may be complex interactions between VOCs emission sources. For example, factors such as wind direction and geographical layout will cause VOCs between different emission sources to transmit, diffuse and react with each other, thereby affecting the VOCs concentration distribution in the entire region. Traditional technology is difficult to accurately capture this complex relationship, which in turn affects the accurate judgment and early warning of VOCs pollution conditions. Summary of the invention

[0005] The purpose of this application is to provide an online monitoring and early warning method and monitoring system for volatile organic pollutants. This application is implemented as follows:

[0006] In a first aspect, the present application provides an online monitoring and early warning method for volatile organic pollutants, the method comprising: obtaining target VOCs monitoring data to be detected, performing feature extraction on the target VOCs monitoring data, and obtaining an implicit representation of the target VOCs monitoring data, wherein the implicit representation of the target VOCs monitoring data includes an implicit representation of a target substance index corresponding to the target VOCs monitoring data; obtaining a target monitoring object in the target VOCs monitoring data, and determining a first involved monitoring object corresponding to the target monitoring object; determining a target focused implicit representation corresponding to the target monitoring object through an implicit representation of the monitoring object of the first involved monitoring object and a corresponding focusing influence coefficient, wherein the focusing influence coefficient is obtained through an involvement coefficient between an implicit representation of the target substance index and an implicit representation of an involvement relationship, and the implicit representation of the involvement relationship is an implicit representation characterizing an involvement relationship of the monitoring object; implicitly integrating the implicit representation of the target VOCs monitoring data with the target focused implicit representation corresponding to the target monitoring object to obtain a target integrated implicit representation; and determining a VOCs detection result corresponding to the target VOCs monitoring data through the target integrated implicit representation.

[0007] On the other hand, the present application provides a monitoring system, comprising: one or more processors; a memory; one or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method as described above is implemented.

[0008] Beneficial effects of the present application: In an embodiment of the present application, the pre-trained network includes a feature extraction network, a focused implicit representation extraction network and an integrated neural network. The focused implicit representation extraction network can determine the sample focused implicit representation corresponding to the sample monitoring object based on the implicit representation of the monitoring object and the corresponding focused influence coefficient. The focused influence coefficient is obtained based on the involvement coefficient between the implicit representation of the sample material index and the implicit representation of the involvement relationship. The implicit representation of the involvement relationship is an implicit representation that characterizes the involvement relationship of the monitoring object. Because the focused influence coefficient is obtained according to the involvement coefficient between the implicit representation of the sample material index and the implicit representation of the involvement relationship, the influence of the implicit representation of the monitoring object involving the monitoring object on the implicit representation of the sample monitoring object can be determined based on the characteristics of the VOCs monitoring data sample, so as to determine the focused influence coefficient. In this way, the obtained features can improve the judgment of the data pattern of the VOCs monitoring data sample, increase the data understanding effect of the obtained pre-trained network and the application layer neural network, and increase the accuracy of the VOCs detection results. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in describing the embodiments of the present application are briefly introduced below.

[0010] Figure 1 It is a flow chart of a method for online monitoring and early warning of volatile organic pollutants provided in an embodiment of the present application.

[0011] Figure 2 It is a schematic diagram of the composition of a monitoring system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0012] The following describes the embodiments of the present application in conjunction with the drawings in the embodiments of the present application. The terms used in the implementation method part of the embodiments of the present application are only used to explain the specific embodiments of the present application, and are not intended to limit the present application.

[0013] The executor of the online monitoring and early warning method for volatile organic pollutants provided in the embodiment of the present application is a monitoring system, including but not limited to a server, a personal computer, a laptop computer, a tablet computer, a smart phone, etc. The monitoring system includes a user device and a network device. Among them, the monitoring system can be operated alone to implement this application, or it can be connected to the network and implemented through interactive operations with other monitoring systems in the network. Among them, the network where the monitoring system is located includes but is not limited to the Internet, a wide area network, a metropolitan area network, a local area network, a VPN network, etc.

[0014] like Figure 1 As shown, the volatile organic pollutant online monitoring and early warning method comprises the following steps:

[0015] Step S100: Acquire target VOCs monitoring data to be detected, perform feature extraction on the target VOCs monitoring data, and obtain an implicit representation of the target VOCs monitoring data, wherein the implicit representation of the target VOCs monitoring data includes an implicit representation of a target substance index corresponding to the target VOCs monitoring data.

[0016] In the embodiment of the present application, the target VOCs monitoring data is a series of data obtained by monitoring volatile organic pollutants in the environment. For example, multiple monitoring points are set up around an industrial plant, and each monitoring point is equipped with corresponding monitoring equipment. These equipment can detect a variety of parameters related to VOCs, such as the concentration of different types of volatile organic compounds, meteorological conditions (temperature, humidity, wind speed, etc.). The data collected by these monitoring equipment within a certain period of time constitutes the target VOCs monitoring data.

[0017] After the monitoring system obtains these target VOCs monitoring data, it performs feature extraction. Feature extraction is a data processing technology that aims to mine representative and distinguishing features from raw data. Assume that the target VOCs monitoring data can be represented as a matrix X, where each row represents a sample (such as monitoring data at a moment) and each column represents a feature (such as the concentration of a certain VOCs substance). One way of feature extraction is principal component analysis (PCA), and its basic formula is: ; where n is the sample size, X T is the transposed matrix of X, and C is the covariance matrix. By performing eigenvalue decomposition on the covariance matrix C, eigenvectors and eigenvalues ​​can be obtained. By selecting the eigenvector corresponding to the larger eigenvalue, the original data can be projected into a low-dimensional space, and the data in this low-dimensional space is the extracted features.

[0018] Another way is Linear Discriminant Analysis (LDA), which finds a projection direction so that data of different categories are separated as much as possible in the projected space. Suppose there are k categories, and the i-th category has samples, and the sample mean is , the overall mean is , the intra-class scatter matrix and the between-class scatter matrix The calculation formula is as follows:

[0019] ;

[0020] ;

[0021] Then by solving the generalized eigenvalue problem The eigenvector corresponding to the maximum eigenvalue of is used to obtain the projection direction, thereby realizing feature extraction.

[0022] Through the above-mentioned technical means or other appropriate feature extraction methods, the monitoring system obtains the implicit representation of the target VOCs monitoring data. This implicit representation is a processed data form that better reflects the essential characteristics of the data. For example, after PCA processing, some redundant information in the original data is removed, and only those features that can explain the data variance to the greatest extent are retained. These features combined together constitute the implicit representation of the target VOCs monitoring data. The implicit representation of the target substance index is part of the implicit representation of the target VOCs monitoring data. The target substance index can be understood as some key indicators related to specific volatile organic pollutants. For example, for a possible volatile organic compound such as benzene, its concentration value, volatilization rate under specific environmental conditions and other indicators are all target substance indicators. In the feature extraction process, these features related to the target substance index are further abstracted and combined to form an implicit representation of the target substance index. It describes the relevant characteristics of the target substance in a more compact and meaningful way, which is helpful for the analysis and judgment of the target substance in subsequent steps.

[0023] Step S200: Acquire a target monitoring object in the target VOCs monitoring data, and determine a first involved monitoring object corresponding to the target monitoring object.

[0024] In the embodiment of the present application, as mentioned above, the target VOCs monitoring data is a collection of various data related to volatile organic pollutants obtained by monitoring equipment set up in a specific area. These data contain a variety of information, such as data on VOCs concentration, meteorological conditions, etc. collected by monitoring stations at different locations.

[0025] The target monitoring object is an object that is specifically selected from the target VOCs monitoring data and needs to be focused on or analyzed. Taking the VOCs monitoring of a chemical park as an example, suppose there are multiple production workshops in the park, and each workshop emits VOCs of different types and concentrations. If one of the workshops has recently adjusted its production process, which may cause a significant change in a certain VOCs substance it emits, then the target monitoring object can be a specific pollution source (such as a specific workshop, equipment, etc.), or a specific monitoring point or a certain type of VOCs substance, etc. Determine the first involved monitoring object corresponding to the target monitoring object, that is, to find other monitoring objects that have a correlation with the target monitoring object. This correlation may be based on a variety of factors.

[0026] One possible correlation factor is the physical location relationship. For example, in terms of wind direction, other workshops or monitoring points that are downwind of the target monitoring object may be affected by the wind direction. VOCs emitted by the target monitoring object may be transmitted to these areas, so these downwind workshops or monitoring points are the first involved monitoring objects. From a technical implementation perspective, this positional relationship can be determined through geographic information system (GIS) technology. Assume that the coordinates of the target monitoring object are (x1, y1), and the coordinates of other monitoring points in the monitoring area are (x i ,y i ), the wind direction vector is , then we can calculate the vector and Angle ,when Within a certain range (such as When , it can be determined that the monitoring point is the first involved monitoring object. The specific calculation formula can be .

[0027] Another correlation factor is the reaction relationship of chemical substances. In VOCs monitoring, chemical reactions may occur between certain substances. For example, a VOCs substance A may react with another substance B to generate a new substance under specific temperature, humidity and light conditions. If the target monitoring object is the source of the emission substance A, then the source of the emission substance B may be the first involved monitoring object. The determination of this chemical substance reaction relationship can be assisted by chemical databases and reaction kinetic models. Suppose there is a chemical substance reaction database that records the reaction equations between various VOCs substances and the conditions under which the reactions occur. The monitoring system can query this database, compare the relevant information of the substances emitted by the target monitoring object, and find out the emission sources of other substances that may react with these substances as the first involved monitoring objects.

[0028] Another correlation factor is the production process correlation. In the industrial production process, different production links may have upstream and downstream relationships. For example, in a plastic production enterprise, certain VOCs intermediates produced in the raw material processing workshop (target monitoring object) may be further reacted or released in the subsequent processing workshop (first involved monitoring object). To determine this production process correlation, the company's production process management system (PMS) can be used. The PMS records in detail the input and output materials, operating conditions and other information of each production link. The monitoring system can analyze the data in the PMS and find out the monitoring objects corresponding to other production links that have upstream and downstream relationships with the target monitoring object in the production process as the first involved monitoring objects.

[0029] Determining the first involved monitoring object helps to fully understand the position and influence of the target monitoring object in the entire VOCs monitoring system. It not only takes into account the characteristics of the target monitoring object itself, but also links it with the surrounding related monitoring objects. This lays the foundation for a more in-depth analysis of the status of the target monitoring object, prediction of its changing trend, and accurate VOCs monitoring and early warning in the subsequent steps. For example, if the VOCs emissions of the target monitoring object suddenly increase, through the relevant data of the first involved monitoring object (such as the production status of the associated workshop, the change in VOCs concentration of the associated monitoring point, etc.), it is possible to more comprehensively analyze whether this change is due to its own production process problems or is affected by surrounding related factors, so as to make early warnings and decision-making more accurately.

[0030] Step S300: Determine the target focus implicit representation corresponding to the target monitoring object through the implicit representation of the monitoring object of the first involved monitoring object and the corresponding focusing influence coefficient, wherein the focusing influence coefficient is obtained through the involvement coefficient between the implicit representation of the target material indicator and the implicit representation of the involvement relationship, and the implicit representation of the involvement relationship is an implicit representation that characterizes the involvement relationship of the monitoring object.

[0031] In step S300 of the embodiment of the present application, the monitoring system determines the target focused implicit representation corresponding to the target monitoring object through the implicit representation of the monitoring object of the first involved monitoring object and the corresponding focusing influence coefficient. In step S100, the monitoring system performs feature extraction on the target VOCs monitoring data to obtain an implicit representation of the target VOCs monitoring data, which includes implicit representations of each monitoring object. The implicit representation of the monitoring object of the first involved monitoring object here is a feature representation form obtained for the first involved monitoring object on this basis. This implicit representation is a data form that has been abstracted and transformed, and it can more effectively reflect the essential characteristics of the first involved monitoring object in the entire VOCs monitoring system.

[0032] For example, suppose there are multiple monitoring sites in a city's VOCs monitoring network, and each site can be regarded as a monitoring object. After feature extraction, the implicit representation of the monitoring object of a first-involved monitoring site may be a multidimensional vector. Each dimension of this vector may represent different features, such as the concentration characteristics of specific VOCs substances around the site (represented after some mathematical transformation), the meteorological conditions in the area where the site is located (such as the processed representation of factors that affect the diffusion of VOCs, such as temperature and humidity), and the characteristics related to the operating status of the site's own equipment (such as the comprehensive representation of relevant factors such as the detection accuracy of the equipment).

[0033] The focus influence coefficient is obtained by the correlation coefficient between the implicit representation of the target substance index and the implicit representation of the involved relationship. The implicit representation of the target substance index is also obtained in step S100, and it is an implicit representation of the characteristics related to the specific substance involved in the target monitoring object. The implicit representation of the involved relationship is an implicit representation that characterizes the involved relationship of the monitoring object, and this involved relationship can be based on a variety of factors, such as physical, chemical or production process connections.

[0034] Assuming that the target monitoring object is a workshop that produces specific chemicals, the implicit representation of the target substance index includes the characteristics of a certain key VOCs substance emitted by the workshop. The first involved monitoring object is other workshops or monitoring points located downwind of the workshop. The implicit representation of the involvement relationship may include a comprehensive representation of factors such as wind direction, distance, and possible chemical substance transmission and reaction relationships. The involvement coefficient between the two is calculated through a specific algorithm. A possible simple algorithm is a calculation based on distance and wind direction weights. Assuming that the distance between the target monitoring object A and the first involved monitoring object B is d, and the weight of the wind direction on VOCs transmission is w (this weight can be determined based on meteorological data and experience), then the involvement coefficient k can be expressed as k=w / d (this is just a simple example, the actual situation may be more complicated).

[0035] After obtaining the involvement coefficient, the focus influence coefficient can be determined. The focus influence coefficient reflects the quantitative value of the influence of the first involved monitoring object on the target monitoring object. For example, if the involvement coefficient between the first involved monitoring object and the target monitoring object is high, indicating that the association between them is strong, then the corresponding focus influence coefficient will also be large, indicating that the first involved monitoring object has a greater influence in determining the target focus implicit representation of the target monitoring object.

[0036] After determining the implicit representation of the monitoring object and the focus influence coefficient of the first involved monitoring object, the monitoring system can determine the target focus implicit representation corresponding to the target monitoring object. This process can be analogized to a weighted summation process. Assume that the implicit representation of the monitoring object of the first involved monitoring object i is I i , and its corresponding focusing influence coefficient is f i , the target focus implicit representation F corresponding to the target monitoring object can be obtained by a similar formula (The actual situation may be more complicated, this is just a conceptual example formula) to calculate.

[0037] In a holistic sense, determining the target-focused implicit representation helps to focus more on the key information related to the target monitoring object in a complex VOCs monitoring system. For example, in a large industrial area with multiple pollution sources and monitoring points, the target monitoring object may be affected by multiple first-involved monitoring objects. By determining the target-focused implicit representation, the monitoring system can integrate these scattered information that are related to the target monitoring object to varying degrees, thereby more accurately characterizing the state and characteristics of the target monitoring object in the entire VOCs monitoring scenario. This is of great significance for the subsequent accurate judgment of the status of the target VOCs monitoring data and effective monitoring and early warning. It can help to extract the most relevant and influential information about the target monitoring object from a large number of monitoring data and complex monitoring object relationships, thereby providing more accurate data support for subsequent analysis and decision-making.

[0038] Step S400: implicitly integrate the target VOCs monitoring data implicit representation with the target focus implicit representation corresponding to the target monitoring object to obtain a target integrated implicit representation.

[0039] In the embodiment of the present application, as mentioned above, the implicit representation of the target VOCs monitoring data is obtained by the monitoring system in step S100 by extracting features from the target VOCs monitoring data. The target focus implicit representation corresponding to the target monitoring object is determined in step S300. This representation reflects the comprehensive impact of the first-involved monitoring object related to the target monitoring object on the target monitoring object. When the monitoring system integrates implicit representations, a possible technical means is vector addition or matrix multiplication, which depends on the specific mathematical form of the implicit representation. If the implicit representation of the target VOCs monitoring data and the implicit representation of the target focus are both in vector form, such as \vec{V} and \vec{F} assumed above, then vector addition can be used for integration. The integrated vector , this new vector It is a simple form of implicit representation of target integration.

[0040] Another possible situation is if the implicit representation is in matrix form. For example, the target VOCs monitoring data is implicitly represented as matrix M V , the target focus is implicitly represented by the matrix M F , then the integration can be achieved through matrix multiplication. Assuming that the matrix multiplication rule conforms to the conventional linear algebra definition, the integrated matrix M T =M V ×M F (Here we need to ensure the feasibility of multiplication according to the matrix dimensions and multiplication rules), M T That is the target integrated implicit representation.

[0041] This implicit representation integration is to comprehensively consider the monitoring data features directly related to the target monitoring object (implicit representation of target VOCs monitoring data) and indirectly related influencing factors (target-focused implicit representation). Taking the VOCs monitoring of a city as an example, the target monitoring object may be a large chemical enterprise located in the center of the city's industrial zone. The implicit representation of the target VOCs monitoring data includes the VOCs concentration of the enterprise's own emission outlet, the meteorological data collected by the small-scale meteorological stations around the enterprise, and other directly monitored information. The target-focused implicit representation may include the potential impact of the production activities of other surrounding enterprises (the first involved monitoring object) on the large chemical enterprise, such as the VOCs emission during the transportation of raw materials of adjacent enterprises may affect the VOCs concentration distribution around the enterprise, or the intermediate products produced by certain production processes of surrounding enterprises may react chemically with the VOCs emitted by the enterprise. By integrating these two implicit representations, the monitoring system can obtain a more comprehensive representation that reflects the state of the target monitoring object in the entire VOCs monitoring environment.

[0042] This integrated target integrated implicit representation contains more dimensional information, not only limited to the monitoring data of the target monitoring object itself, but also incorporates the influence of other related factors. This enables subsequent judgments based on richer and more comprehensive information when determining the VOCs test results, thereby improving the accuracy and reliability of the test results. For example, when judging whether the chemical company has abnormal VOCs emissions, the integrated implicit representation can comprehensively consider the internal factors of the company (its own production process, equipment operation status, etc. are reflected in the implicit representation of the target VOCs monitoring data) and external environmental factors (the influence of surrounding companies is reflected in the target focused implicit representation), avoiding inaccurate judgments based on a single data source.

[0043] Step S500: Determine the VOCs detection result corresponding to the target VOCs monitoring data through target integrated implicit representation.

[0044] In the embodiment of the present application, one way is to construct a classification model. For example, a support vector machine (SVM) algorithm can be used. The basic idea of ​​SVM is to find an optimal hyperplane in the feature space to separate data of different categories. For the classification of VOCs detection results, normal VOCs emissions and abnormal VOCs emissions can be regarded as two categories.

[0045] Assume that the target integration is implicitly represented as the vector , for the two-class SVM model, its decision function can be expressed as ,in is the normal vector of the hyperplane, and b is the bias term. It is determined by the training dataset (containing target integrated implicit representation samples corresponding to known normal and abnormal VOCs emissions) and b. >0, it can be judged as a certain type (such as abnormal emission). When <0, it is judged as another category (such as normal emission).

[0046] Another technical approach is based on neural networks. For example, a multi-layer perceptron (MLP) neural network can be constructed. The target integrated implicit representation is used as the input of the input layer neurons of the neural network. Assume that there are n neurons in the input layer, corresponding to the n feature components of the target integrated implicit representation. The neural network can have multiple hidden layers, and the number of neurons in each hidden layer can be set according to actual conditions, such as using m neurons. The hidden layers and the hidden layers and the output layer are connected through activation functions (such as ReLU function y=max(0,x)) and weight matrices.

[0047] During the training process, the neural network is trained with a large number of target integrated implicit representation samples corresponding to known VOCs detection results, and the value of the weight matrix is ​​adjusted so that the neural network can accurately classify the input target integrated implicit representation. The number of neurons in the output layer is determined according to the number of categories of the detection results. For example, for a two-classification problem (normal and abnormal emissions), the output layer can have one neuron, and an output value close to 1 indicates abnormal emissions, and close to 0 indicates normal emissions.

[0048] Take the VOCs monitoring of an industrial park as an example. There are many factories in the industrial park, and different factories emit different types and concentrations of VOCs. The target integrated implicit representation contains the emission characteristics of each factory (target monitoring object) and the influencing factors between each other. The monitoring system analyzes the target integrated implicit representation through the above-mentioned technical means (such as SVM or neural network) to determine whether the VOCs emissions of the entire industrial park are within the normal range. If it is determined to be abnormal emission, this may mean that there is a problem with the production process of a factory, the environmental protection equipment fails, or the meteorological conditions cause the accumulation of pollutants. This test result is of great guiding significance for the environmental supervision department to take corresponding measures (such as requiring the factory to make rectifications, strengthening the supervision of the operation of environmental protection equipment, etc.).

[0049] In addition, when determining the VOCs detection results, the uncertainty and error of the data must also be considered. Due to the accuracy limitations of the monitoring equipment itself, the interference of environmental factors, and approximate calculations in the data processing process, the target integrated implicit representation may have certain errors. In order to deal with this situation, a probabilistic model, such as a Bayesian network, can be used. The Bayesian network can calculate the probabilities of different VOCs detection results based on prior knowledge (such as the statistical laws of historical VOCs monitoring data) and the current target integrated implicit representation. For example, given the target integrated implicit representation , calculate the probability of normal emission P(normal | ) and the probability of abnormal emission P(abnormal| ), and determine the final detection result based on the probability.

[0050] In summary, in step S500, the monitoring system uses appropriate technical means (such as classification models, neural networks, probability models, etc.) to integrate the implicit representation of the target, comprehensively considers factors such as data characteristics and uncertainty, and determines the VOCs detection results corresponding to the target VOCs monitoring data.

[0051] As an implementation manner, step S300, determining a target focus implicit representation corresponding to a target monitoring object through a first monitoring object implicit representation and a corresponding focus influence coefficient involving the monitoring object, includes:

[0052] Step S310: obtaining an object involvement map composed of a first involved monitoring object and a target monitoring object;

[0053] Step S320: for a graph monitoring object in the object involvement graph, obtaining an implicit representation of the involvement relationship that represents the involvement relationship between the graph monitoring object and adjacent monitoring objects;

[0054] Step S330: Obtain an implicit representation involvement coefficient through an implicit representation of the involvement relationship and an implicit representation of the target material index, and determine a focusing influence coefficient corresponding to an adjacent monitoring object through the implicit representation involvement coefficient;

[0055] Step S340: determining the target focus implicit representation corresponding to the atlas monitoring object through the focus influence coefficient corresponding to the adjacent monitoring object and the monitoring object implicit representation of the adjacent monitoring object;

[0056] Step S350: extracting the target focus implicit representation corresponding to the target monitoring object from the target focus implicit representations corresponding to each atlas monitoring object of the object involvement atlas.

[0057] In step S310, the object involvement graph is a data structure (either a directed graph or an undirected graph) used to represent the relationship between monitoring objects. For example, in a VOCs monitoring scenario in a chemical park, the target monitoring object is a major production workshop, and the first involved monitoring object may include other workshops around the workshop, storage facilities, logistics channels, and surrounding environmental monitoring points. There are many complex relationships between these objects, such as material transmission relationships, pollutant diffusion relationships caused by wind direction, upstream and downstream relationships in production processes, etc. The monitoring system represents these relationships in the form of an object involvement graph. This graph can be represented by a graph structure in graph theory, in which nodes represent each monitoring object, edges represent the relationship between objects, and edges can carry attributes such as weights to represent information such as the strength of the relationship. For example, if there is a material transmission relationship between two workshops, the corresponding edge in the graph can be assigned a weight value based on factors such as the transmission amount.

[0058] In step S320, the example of a chemical park is used to explain that for a certain workshop in the map (map monitoring object), there is a specific involvement relationship between it and the adjacent storage facility (adjacent monitoring object). This implicit representation of the involvement relationship may contain multiple aspects of information. From a spatial perspective, it may include the representation of information such as the distance and relative position between the two after some mathematical transformation; from a production process perspective, it may include a comprehensive representation of factors such as the impact of materials in the storage facility on VOCs emissions during the workshop production process; from an environmental impact perspective, it may involve the environmental conditions around the storage facility (such as temperature, humidity, etc.) on the VOCs emitted by the workshop during the transmission and diffusion process. This implicit representation of the involvement relationship can be in the form of a multidimensional vector, where each dimension represents a quantitative representation of a specific relationship factor. For example, a simple implicit representation vector of the involvement relationship ,in Represents the normalized value of the distance factor. Represents the quantitative value of the factors affecting the production process, Represents the quantitative value of environmental impact factors.

[0059] In step S330, the implicit representation of the target substance index reflects the characteristics related to the specific substance involved in the target monitoring object, such as the comprehensive representation of the concentration characteristics and chemical activity characteristics of specific VOCs substances. Still taking the chemical park as an example, assuming that the workshop of the target monitoring object mainly emits a certain VOCs substance, the implicit representation of its target substance index includes the concentration change trend of this substance, the reaction activity under different environmental conditions and other information. The monitoring system calculates the implicit representation involvement coefficient through a specific algorithm. For example, the method of vector dot product and module length ratio can be used for calculation. Suppose the implicit representation of the involvement relationship is , the target substance index is implicitly expressed as , the implicit representation involvement coefficient k can be expressed as ,here represents the vector dot product, Represents vectors Then, since the implicit representation involvement coefficient is positively correlated with the focusing influence coefficient corresponding to the adjacent monitored object, the focusing influence coefficient can be determined according to a preset mapping relationship. For example, if the implicit representation involvement coefficient k is in the range of 0-0.3, the corresponding focusing influence coefficient f can be set to 0.1k; if k is in the range of 0.3-0.7, f=0.3+0.5(k-0.3); if k>0.7, f=0.8+0.2(k-0.7). This mapping relationship can be adjusted according to the actual monitoring data characteristics and experience.

[0060] In step S340, in the example of the chemical park, for a certain atlas monitoring object in the atlas (such as a workshop), its adjacent monitoring objects (such as surrounding storage facilities) have their own implicit monitoring object representations, which contain various characteristic information of the storage facilities themselves, such as the type of stored material, storage volume, and operating status of the facilities, after feature extraction. The monitoring system weights the implicit monitoring object representations of the adjacent monitoring objects according to the previously determined focusing influence coefficient to obtain the target focused implicit representation corresponding to the atlas monitoring object. Assume that the implicit monitoring object representation of the adjacent monitoring object i is , and its corresponding focusing influence coefficient is , the target focus implicit representation corresponding to the atlas monitoring object Can be The summation here is performed on all adjacent monitoring objects of the atlas monitoring object. This weighted summation method can comprehensively consider the different degrees of influence of each adjacent monitoring object on the atlas monitoring object, thereby obtaining an implicit representation that can better reflect the focus features of the atlas monitoring object in the entire object involvement relationship.

[0061] In step S350, the entire object involvement map includes multiple map monitoring objects, and each map monitoring object has its corresponding target focused implicit representation. The target monitoring object is an object of special concern, such as the main production workshop in a chemical park. The monitoring system accurately extracts the target focused implicit representation corresponding to the target monitoring object from the target focused implicit representations of many map monitoring objects by positioning and identifying in the object involvement map. This process is similar to searching for specific data items in a complex data set. For example, if the object involvement map is stored in the form of a graph database, the monitoring system can find the corresponding target focused implicit representation in the graph database according to the identification information of the target monitoring object (such as the number and name of the workshop, etc.) through a query statement. This extraction operation ensures that subsequent steps can be processed based on the specific focused implicit representation of the target monitoring object, thereby more accurately analyzing the status and impact of the target monitoring object in the entire VOCs monitoring system.

[0062] In summary, steps S310-S350 perform a series of processing on the relationship and feature representation of each monitoring object in the object involvement map through the monitoring system, and gradually determine the target focus implicit representation corresponding to the target monitoring object, which provides a key data basis for further analysis in the subsequent online monitoring and early warning method of volatile organic pollutants. These steps comprehensively consider the complex relationship between monitoring objects, material index characteristics, and the quantitative representation of various influencing factors, which helps to improve the accuracy and reliability of the entire VOCs monitoring and early warning.

[0063] In the actual industrial production environment, there are many variables and complex relationships. The execution of these steps can effectively handle these complexities. For example, in a large petrochemical complex, different production units are interrelated, and the transportation of raw materials, product processing, and waste treatment processes will involve the emission and propagation of VOCs. The target monitoring object may be a key reaction device. The first involved monitoring objects include raw material supply devices, subsequent processing devices, environmental protection treatment devices, and surrounding environmental monitoring points. The object involvement map constructed by the monitoring system can comprehensively describe the relationship between these devices. When obtaining the implicit representation of the involvement relationship, it is necessary to consider multiple factors such as the chemical properties of the materials, reaction conditions, layout of the device, and meteorological conditions. For example, the raw material composition in the raw material supply device may affect the type and concentration of VOCs generated in the reaction device. This influence needs to be quantified through complex chemical engineering calculations and environmental factor analysis in the implicit representation of the involvement relationship.

[0064] For the calculation of implicit representation involvement coefficient and focus influence coefficient, the calculation method and parameter setting may need to be adjusted in different production scenarios. Taking an enterprise with both oil refining and chemical production as an example, the production processes of the oil refining part and the chemical part are different, and the impact mechanism on VOCs emissions is also different. When calculating the implicit representation involvement coefficient of the implicit representation of the involvement relationship and the implicit representation of the target substance index, different weight distribution or calculation functions may be required for the relationship between the oil refining unit and the chemical unit and the target monitoring object. For example, the impact of operating conditions such as temperature and pressure of the oil refining unit on VOCs emissions may account for a large proportion in the calculation, while the impact of the raw material composition and reaction path of the chemical unit on VOCs emissions may be more critical in the calculation. Accordingly, when determining the focus influence coefficient, it is also necessary to set different mapping relationships according to different implicit representation involvement coefficient ranges to accurately reflect the degree of influence of different units on the target monitoring object.

[0065] In the process of determining the target focused implicit representation corresponding to the spectrum monitoring object, due to the large number of devices and adjacent monitoring objects in large enterprises, each adjacent monitoring object has different ways and degrees of influence on the spectrum monitoring object. Take a chemical production area containing multiple reaction towers, separation towers and storage tanks as an example. There are multiple storage tanks (adjacent monitoring objects) around a reaction tower (spectrum monitoring object). The material properties, storage volume and sealing performance of the storage tanks will affect the VOCs emission of the reaction tower. The monitoring system can comprehensively consider these factors and obtain accurate target focused implicit representation through precise focus influence coefficient calculation and weighted summation.

[0066] Finally, the target-focused implicit representation corresponding to the target monitoring object is extracted from the target-focused implicit representations corresponding to the numerous atlas monitoring objects. This process is particularly important in the complex production environment of large enterprises. Due to the complexity of the production process and the massive amount of monitoring data, ensuring the accurate extraction of the target-focused implicit representation of the target monitoring object is the key to subsequent accurate VOCs monitoring and early warning. For example, in an enterprise with multiple production workshops and a complex logistics network, each workshop and logistics node has its own target-focused implicit representation. The monitoring system needs to use efficient database query and data matching technology to accurately find the target-focused implicit representation of the target monitoring object (such as a key supervision workshop) for further analysis and decision-making.

[0067] These steps, through the precise execution of the monitoring system in different industrial scenarios, can effectively handle complex VOCs monitoring relationships and provide strong technical support for environmental protection and the sustainable development of industrial production.

[0068] For the construction of object involvement graph, graph database technology, such as Neo4j, can be used. The monitoring objects can be used as nodes, the relationships between objects as edges, and the relevant attributes (such as the strength and direction of the relationship) can be stored on the edges to build the object involvement graph.

[0069] When calculating the implicit representation involvement coefficient and the focus influence coefficient, in addition to the vector calculation method mentioned above, an algorithm based on machine learning can also be used for optimization. For example, a decision tree algorithm can be used to analyze historical monitoring data, and automatically generate the calculation rules of the implicit representation involvement coefficient and the mapping relationship of the focus influence coefficient according to different situations. The decision tree algorithm builds a decision tree by learning a large amount of known monitoring data (including implicit representation of the involvement relationship, implicit representation of the target substance index, and the corresponding actual monitoring results). When calculating the new implicit representation involvement coefficient and the focus influence coefficient, the decision tree is traversed according to the input features (implicit representation of the involvement relationship and implicit representation of the target substance index) to obtain the corresponding results.

[0070] Data mining technology can be used in the process of determining the target-focused implicit representation corresponding to the graph monitoring object and extracting the target-focused implicit representation corresponding to the target monitoring object. For example, cluster analysis can be used to classify graph monitoring objects in order to better understand the roles and impacts of different types of graph monitoring objects in the entire VOCs monitoring system. For extracting the target-focused implicit representation corresponding to the target monitoring object, indexing technology can be used to improve query efficiency. In the graph database, specific indexes can be established for the target monitoring object so that the monitoring system can quickly locate and extract its corresponding target-focused implicit representation.

[0071] As an implementation mode, the target focused implicit representation corresponding to the atlas monitoring object is obtained by processing the focused implicit representation extraction network, and the focused implicit representation extraction network includes one or more target intermediate network components. Step S340 determines the target focused implicit representation corresponding to the atlas monitoring object through the focused influence coefficient corresponding to the adjacent monitoring object and the monitoring object implicit representation of the adjacent monitoring object, including:

[0072] Step S341: loading the monitoring object implicit representation and the involved relationship implicit representation of the adjacent monitoring object into the target intermediate network component, and executing to obtain the first focused implicit representation corresponding to the atlas monitoring object;

[0073] Step S342: Determine the target focus implicit representation corresponding to the atlas monitoring object through the first focus implicit representation corresponding to the atlas monitoring object and the focus influence coefficient corresponding to the corresponding adjacent monitoring object.

[0074] In step S341, the implicit representation of the monitoring object of the adjacent monitoring object is an abstract representation of the characteristics of the adjacent monitoring object obtained after processing in the previous steps. For example, in a VOCs monitoring scenario in a city, assuming that the target monitoring object is a large chemical enterprise located in the city center, its adjacent monitoring objects may include surrounding small factories, warehouses, and transportation hubs. For a small factory (adjacent monitoring object), its implicit representation of the monitoring object may include the production type of the factory (such as whether it involves chemical raw material processing, etc.), production scale (expressed in some quantitative indicators), equipment operation status (such as the newness of the equipment, maintenance frequency and other related factors after the quantified value of the comprehensive) and other information after feature extraction and conversion. This implicit representation exists in a specific mathematical structure, which may be in the form of a vector or matrix, which is convenient for the monitoring system to perform subsequent processing.

[0075] The implicit representation of the involved relationship is a form of representation that characterizes the relationship between the atlas monitoring object and the adjacent monitoring object. Continuing with the above example, the implicit representation of the involved relationship between the chemical enterprise (atlas monitoring object) and the surrounding small factories (adjacent monitoring objects) may include a comprehensive representation of factors such as the material exchange relationship between them (if any, the frequency and quantity of the exchange are represented by some quantitative indicators), geographical distance (normalized value), wind direction influence (for example, the quantitative value of the wind direction's influence on VOCs diffusion is determined based on local meteorological data). Similarly, this implicit representation of the involved relationship also exists in a mathematical structure suitable for computer processing, such as a vector form.

[0076] The target intermediate network component is a part of the entire network structure. It is designed to perform specific processing on the input implicit representation of the monitored object and the implicit representation of the involved relationship to obtain the first focused implicit representation corresponding to the atlas monitored object. This target intermediate network component can be a module based on a neural network architecture, such as a simple multi-layer perceptron (MLP). The monitoring system loads the implicit representation of the monitored object and the implicit representation of the involved relationship of the adjacent monitored object as input into this MLP. Assume that the implicit representation of the monitored object is a vector , the implied relation is implicitly represented as a vector , the number of neurons in the input layer of MLP is based on and is determined by the sum of the dimensions.

[0077] Inside the MLP, the input data is processed through a series of linear transformations and nonlinear activation function operations. For example, for the connection from the input layer to the first hidden layer, there is a weight matrix and the bias vector , the linear transformation formula is ,in Indicates that The vector is concatenated by column. Then, a nonlinear activation function (such as ReLU function y=max(0,x)) is used to Process and obtain This process is carried out in sequence between hidden layers. After processing through multiple hidden layers, the output vector is finally obtained, which is the first focused implicit representation corresponding to the atlas monitoring object. .

[0078] In step S342, after obtaining the first focused implicit representation corresponding to the atlas monitoring object Afterwards, the monitoring system needs to combine the focus influence coefficients corresponding to the adjacent monitoring objects to determine the final target focus implicit representation. The focus influence coefficient reflects the importance of the impact of the adjacent monitoring object on the atlas monitoring object. Taking the previous example of the chemical company and the surrounding small factories, if a certain VOCs substance produced in the production process of the small factory has a greater impact on the VOCs emissions of the chemical company (for example, because of wind direction and distance factors, the substances emitted by the small factory are easy to diffuse to the chemical company area and react with the substances emitted by it), then the focus influence coefficient corresponding to this small factory will be relatively large.

[0079] Assuming there are n adjacent monitoring objects, the focusing influence coefficient corresponding to the i-th adjacent monitoring object is , and its corresponding first focus implicitly expressed as The monitoring system determines the target focus implicit representation corresponding to the atlas monitoring object by weighted summation. The formula can be expressed as . This weighted summation method can comprehensively consider the different degrees of influence of each adjacent monitoring object on the atlas monitoring object, so that the final target focus implicit representation can more accurately reflect the focus characteristics of the atlas monitoring object in the entire monitoring system. For example, in the VOCs monitoring of a large industrial area, there are many complex monitoring object relationships. For example, there are multiple different types of factories, warehouses, transportation sites, etc., and there are many factors that influence each other, including the interaction of production processes, the diffusion and transmission of pollutants, the flow of materials, etc. Steps S341 and S342 can extract useful information from these complex relationships through the precise processing of the monitoring system, and integrate it into the target focus implicit representation corresponding to the atlas monitoring object. This helps to accurately analyze the target monitoring object in the subsequent analysis. For example, when judging whether the VOCs emissions of the target monitoring object are abnormal, the influence of the surrounding related monitoring objects can be comprehensively considered to improve the accuracy of the judgment.

[0080] In the actual implementation process, in order to optimize the performance of the target intermediate network component, a variety of technical means can be used. For example, when training the target intermediate network component, the back propagation algorithm can be used to adjust the weight of the network. The network is trained with a large number of known monitoring data samples (including implicit representations of monitoring objects, implicit representations of involved relationships, and corresponding actual monitoring results). In each iteration, the gradient is calculated using the back propagation algorithm based on the error between the predicted result and the actual result, and then the weight of the network is updated, so that the network can better fit the data and improve the prediction accuracy of the first focused implicit representation.

[0081] In addition, the determination of the focus influence coefficient can be adjusted according to the actual distribution of monitoring data and expert experience. For example, if it is found that the wind direction in a certain area has a more significant impact on the diffusion of VOCs than other areas, then when calculating the focus influence coefficient corresponding to the implicit representation of the involved relationship involving the wind direction factor, its weight can be appropriately increased. At the same time, as the monitoring data continues to accumulate, the focus influence coefficient can be regularly re-evaluated and adjusted to adapt to changes in the environment and production process.

[0082] As an implementation mode, step S341, loading the monitoring object implicit representation and the involved relationship implicit representation of the adjacent monitoring object into the target intermediate network component, and executing to obtain the first focused implicit representation corresponding to the atlas monitoring object includes:

[0083] Step S3411: determining a target operation mode through the monitoring object involvement relationship between the graph monitoring object and the adjacent monitoring objects, the target operation mode being summation or difference;

[0084] Step S3412: performing operations on the monitoring object implicit representation and the involved relationship implicit representation of the adjacent monitoring object through the target operation mode to obtain the calculated implicit representation corresponding to the atlas monitoring object;

[0085] Step S3413: Take the intermediate network component parameters in the target intermediate network component to process the calculated implicit representation to obtain the first focused implicit representation corresponding to the atlas monitoring object.

[0086] In step S3411, in the complex system of VOCs monitoring, there are various types of monitoring object involvement relationships between the spectrum monitoring object and the adjacent monitoring objects. These relationships reflect their connections in the generation, propagation, and mutual influence of VOCs. For example, in a monitoring scenario of a chemical park, the spectrum monitoring object is a large chemical production workshop, and the adjacent monitoring object may be a warehouse for storing specific raw materials, environmental protection equipment for treating exhaust gas, or other workshops located downwind.

[0087] If the material flow relationship of VOCs is taken into account, when a certain raw material in the warehouse is directly input into the chemical production workshop during the production process, and the input amount of this raw material is positively correlated with the VOCs generated in the workshop, this monitoring object involvement relationship may be more suitable for the summation target operation mode. For example, the more raw materials input into the warehouse, the more VOCs substances are generated in the workshop. At this time, the monitoring system can be determined as the summation operation mode.

[0088] On the contrary, if the relationship between the workshop and the environmental protection equipment is taken into account, there is a counteracting relationship between the processing capacity of the environmental protection equipment and the amount of VOCs generated by the workshop. The amount of VOCs generated by the workshop minus the amount processed by the environmental protection equipment is the actual amount of VOCs discharged into the environment. In this case, the monitoring system may determine the difference as the target operation mode.

[0089] This method of determining the target operation mode according to the relationship involved in different monitoring objects will help to perform reasonable operations on the relevant data according to the specific relationship characteristics, so as to better reflect the characteristics of the spectrum monitoring object in the entire VOCs monitoring system.

[0090] In step S3412, it is assumed that the monitoring object of the adjacent monitoring object is implicitly represented as a vector , whose dimension is n, representing n features of adjacent monitoring objects; the involved relationship is implicitly represented as a vector , with dimension m (m may be different from n because they represent different aspects of information).

[0091] When the target operation mode is summation, if n=m, the implicit representation after the operation is The formula can be For example, in the above chemical workshop and warehouse example, if A dimension in represents the production capacity-related characteristics of the workshop itself, The corresponding dimension in represents the characteristics of the impact of warehouse raw material input on workshop production capacity. The characteristics corresponding to the new vector dimension obtained after summation combine the impact of both.

[0092] When the target operation mode is difference, if n=m, then Taking workshops and environmental protection equipment as an example, if one dimension in \vec{I} represents the VOCs concentration-related characteristics generated by the workshop, The corresponding dimension in represents the characteristics of the impact of the environmental protection equipment's treatment capacity on the concentration. The characteristics corresponding to the vector dimension obtained after taking the difference reflect the characteristics related to the actual VOCs concentration emitted by the workshop after treatment by the environmental protection equipment.

[0093] The implicit representation of the graph monitoring object obtained by this operation integrates the information in the implicit representation of the monitoring object of the adjacent monitoring object and the implicit representation of the involved relationship, providing a more comprehensive basis for subsequent further processing.

[0094] In step S3413, the target intermediate network component is a network structure specially designed to process such data, which includes intermediate network component parameters. The intermediate network component parameters include parameters such as weight matrix and bias vector, which determine how the network converts and processes the input data.

[0095] For example, the target intermediate network component can be a simple linear neural network layer. Assume that after the operation, it is implicitly represented as a vector , with a dimension of p, the weight matrix of the intermediate network component is W, with a dimension of q×p (q is the output dimension, set as needed), and the bias vector is \vec{b}, with a dimension of q.

[0096] The monitoring system uses the linear transformation formula After processing the implicit representation after the operation, we get It is the first focused implicit representation corresponding to the atlas monitoring object.

[0097] In practical applications, these intermediate network component parameters need to be trained to determine appropriate values. For example, a large amount of historical monitoring data can be used, including implicit representations of known adjacent monitoring objects, implicit representations of involved relationships, and actual state data of corresponding graph monitoring objects (such as actual VOCs emissions, etc.). By minimizing the error between the predicted results and the actual results, for example, using the mean square error (MSE) as the loss function, the formula is , where N is the number of samples, is the predicted value, is the actual value.

[0098] Using optimization algorithms, such as the stochastic gradient descent (SGD) algorithm, the parameters of the intermediate network components are adjusted according to the loss function. In each iteration, the gradient of the loss function to the parameter is calculated, and then the parameter is updated according to a certain step size, so that the loss function gradually decreases, thereby obtaining the appropriate parameters of the intermediate network components, and then obtaining the first focused implicit representation corresponding to the accurate atlas monitoring object.

[0099] In VOCs monitoring of a large industrial area, there are many different types of monitoring objects and complex relationships. For example, there are multiple factories with different production processes, various types of storage facilities, and complex meteorological conditions. Steps S3411-S3413 can perform reasonable calculations based on the specific monitoring object involvement relationships through precise processing of the monitoring system, and effectively convert and integrate data through the target intermediate network components.

[0100] When actually building and optimizing this process, some advanced technical means can also be used. For example, regularization technology can be used to prevent overfitting of intermediate network components. Regularization terms, such as L1 regularization or L2 regularization, are added when calculating the loss function. This can limit the range of values ​​of the parameters of the intermediate network components during training, avoiding the model being too complex and resulting in a decrease in generalization ability on new data.

[0101] In addition, batch normalization technology can be used to normalize the implicit representation after operation before inputting it into the target intermediate network component. Batch normalization can accelerate the network training process and improve the stability and convergence speed of the network. Its basic idea is to normalize the data of each batch so that the mean of the data is 0 and the variance is 1.

[0102] As an implementation mode, step S3413, taking the intermediate network component parameter in the target intermediate network component to process the calculated implicit representation to obtain the first focused implicit representation corresponding to the atlas monitoring object, includes:

[0103] Step S34131: Obtain an implicit representation of an execution result corresponding to an adjacent monitoring object obtained by executing a previous intermediate network component corresponding to a target intermediate network component in the focused implicit representation extraction network;

[0104] Step S34132: The target intermediate network component uses the first intermediate network component parameter to process the implicit representation after the operation and the implicit representation of the execution result to obtain the first focused implicit representation corresponding to the atlas monitoring object.

[0105] In step S34131, in the complex network structure of the entire volatile organic pollutant monitoring and early warning, the focused implicit representation extraction network is a system composed of multiple components, in which the target intermediate network component does not exist in isolation, but is interconnected and coordinated with other intermediate network components. The implicit representation of the execution result corresponding to the adjacent monitoring object obtained by executing the previous intermediate network component is the intermediate result of the network in the process of processing data.

[0106] For example, in a VOCs monitoring scenario in a chemical park, it is assumed that the focused implicit representation extraction network aims to analyze the impact of the complex relationship between the various workshops (monitoring objects) in the chemical park on a specific workshop (spectrum monitoring object). The target intermediate network component may be a layer in the network, which is used to further integrate and transform the previously processed information. When the previous intermediate network component processes the data of an adjacent workshop (adjacent monitoring object), it will obtain an implicit representation of the execution result related to the adjacent workshop.

[0107] This implicit representation of the execution result contains the comprehensive information after the previous intermediate network component processes the data of the adjacent monitoring object. For example, the previous intermediate network component may have performed preliminary feature fusion and analysis on the production scale, production process type, equipment operation status and other information of the adjacent workshop (this information has been converted into implicit representation of the monitoring object through feature extraction and a series of operations). Assuming that the adjacent workshop is a workshop that produces plastic products, the implicit representation of its execution result may contain feature information related to the potential risk of VOCs emissions in the production process of the workshop after processing by the previous intermediate network component, such as the comprehensive characteristics of the impact of chemical reactions involved in the production process on VOCs generation, the comprehensive characteristics of the impact of equipment operation stability on VOCs leakage risk, etc. These features are represented in a specific mathematical structure (such as vector form).

[0108] In step S34132, the first intermediate network component parameters in the target intermediate network component include parameters such as weight matrix and bias vector, which play a key role in the data processing process.

[0109] Continuing with the example of the chemical park mentioned above, the implicit representation after operation is the result of operating the implicit representation of the monitoring object and the implicit representation of the involvement relationship of the adjacent monitoring object after determining the target operation mode (sum or difference) according to the monitoring object involvement relationship between the spectrum monitoring object and the adjacent monitoring object in the previous step. It is assumed that this implicit representation after operation contains comprehensive information about the relationship between the adjacent workshop and the spectrum monitoring object (for example, a major chemical production workshop) under the current processing step, such as the amount of material transfer between the two, the influence of distance factors on the diffusion of VOCs, and other characteristic information after operation.

[0110] The target intermediate network component takes the implicit representation after operation and the implicit representation of the execution result obtained by the previous intermediate network component as input and processes it using the first intermediate network component parameter. Assume that the implicit representation after operation is a vector , the execution result is implicitly represented as a vector , the weight matrix of the target intermediate network component is W (the dimension is determined by the dimension of the input vector, assuming , where n is The dimension of m is dimension), the bias vector is (dimension is q).

[0111] The monitoring system uses the linear transformation formula (in Indicates that Processing these two input vectors by column concatenation, we get It is the first focused implicit representation corresponding to the atlas monitoring object.

[0112] In actual operation, determining the values ​​of these first intermediate network component parameters is a key link. A feasible technical means is to optimize through training data. For example, a large amount of monitoring data of different workshops in a chemical park under different production conditions is collected. These data include various relationship data between known adjacent workshops and graph monitoring workshops (used to generate implicit representations after operations and implicit representations of execution results) and the actual VOCs emissions of the corresponding graph monitoring workshops (as target output).

[0113] An optimization algorithm, such as stochastic gradient descent (SGD), is used to minimize the error between the predicted result (the VOCs emission prediction value obtained by the above formula and subsequent processing) and the actual result (the actual VOCs emission situation). During each iteration, the parameters of the first intermediate network component are adjusted according to the loss function (such as mean square error).

[0114] In addition, model selection techniques can be used to determine the optimal structure and parameter settings of the target intermediate network components. For example, a cross-validation method can be used to divide the data set into a training set, a validation set, and a test set. During the training process, the model performance under different structures and parameter settings is evaluated on the validation set (such as based on the value of the loss function), and the model structure and parameter settings with the best performance are selected, and then the final performance evaluation is performed on the test set.

[0115] As an implementation manner, the target focused implicit representation corresponding to the atlas monitoring object is obtained by executing a focused implicit representation extraction network, and the focused implicit representation extraction network includes one or more target intermediate network components. Step S330, obtaining an implicit representation involvement coefficient through an implicit representation of an involvement relationship and an implicit representation of a target material index, and determining a focused influence coefficient corresponding to an adjacent monitoring object through the implicit representation involvement coefficient, includes:

[0116] Step S331: taking the second intermediate network component parameter in the target intermediate network component to process the implicit representation of the involved relationship to obtain the core implicit representation;

[0117] Step S332: taking the third intermediate network component parameter in the target intermediate network component to process the target material index implicit representation to obtain the search implicit representation;

[0118] Step S333: Obtaining an implicit representation involvement coefficient by determining the core implicit representation and the search implicit representation;

[0119] Step S334: determining the focus influence coefficient corresponding to the adjacent monitoring object by implicitly expressing the involvement coefficient, wherein the implicitly expressing involvement coefficient is positively correlated with the focus influence coefficient corresponding to the adjacent monitoring object.

[0120] In step S331, in the entire VOCs monitoring and early warning network architecture, the target intermediate network component is a part dedicated to processing relevant data to obtain useful information, in which the second intermediate network component parameter plays a key role. The implicit representation of the involvement relationship is a data form that characterizes the involvement relationship of the monitored objects, and contains comprehensive information on multiple factors such as the physical distance between the monitored objects, the influence of wind direction, and the association of production processes.

[0121] For example, in a VOCs monitoring scenario in a chemical park, suppose there is a production workshop (target monitoring object) and a raw material storage warehouse (adjacent monitoring object). The implicit representation of the relationship between them may include the distance between the two (for example, expressed in a normalized numerical value), the impact of wind direction on the possible diffusion of VOCs (quantitative value calculated based on meteorological data and park layout), and the degree of correlation between raw material supply and workshop production process (such as the potential impact of the stability and timeliness of raw material supply on VOCs emissions in the workshop production process, expressed in a specific quantitative way).

[0122] The monitoring system processes the implicit representation of the involvement relationship using the second intermediate network component parameter in the target intermediate network component. The second intermediate network component parameter may include a weight matrix, a bias vector, etc. Assume that the implicit representation of the involvement relationship is a vector , with dimension n, the weight matrix of the second intermediate network component is , the dimension is m×n (m is the output dimension set according to the processing requirements), and the bias vector is , dimension is m. The monitoring system is transformed by linear formula Get the core implicit representation This core implicitly represents It is the result of specific network processing of the implicit representation of the involvement relationship. It reflects the involvement relationship between the monitored objects in a new form, which is more conducive to the subsequent association calculation with the implicit representation of the target material index.

[0123] In step S332, the implicit representation of the target substance index is an implicit representation of the characteristics related to the specific substance involved in the target monitoring object, such as the concentration change trend of a specific VOCs substance, the chemical activity characteristics under different environmental conditions and other information after processing.

[0124] Continuing with the example of the chemical park, the target substance indicators involved in the target monitoring workshop may implicitly include the fluctuation characteristics of the concentration of a certain key VOCs substance emitted by the workshop in different time periods (expressed as a numerical value after mathematical transformation), the generation and transformation of the substance in different production links in the workshop (expressed as a comprehensive quantitative indicator), and the impact of the chemical activity changes of the substance under specific meteorological conditions (such as temperature and humidity range) on the total VOCs emissions (expressed as a specific numerical value), etc.

[0125] The monitoring system uses the third intermediate network component parameter in the target intermediate network component to process the implicit representation of the target material index. Assume that the implicit representation of the target material index is a vector , with dimension p, the weight matrix of the third intermediate network component is , the dimension is q×p (q is the output dimension set according to processing requirements), and the bias vector is , with dimension q. Through the linear transformation formula Get the search implicit representation This search implicitly means It is the result of a specific network processing of the implicit representation of the target material index, which is in a way that is different from the core implicit representation. The corresponding forms exist to facilitate the subsequent association operations between the two.

[0126] In step S333, the core implicitly represents and search for implicit representations They are all specific forms of representation obtained after the previous steps. In order to determine the degree of association between the two, it is necessary to calculate the implicit representation involvement coefficient.

[0127] One possible calculation method can be based on the vector similarity calculation method. For example, the cosine similarity formula can be used to calculate the implicit representation involvement coefficient C. The value of the implicit representation involvement coefficient C is between -1 and 1, which reflects the degree of association between the core implicit representation and the search implicit representation. When C is close to 1, it means that there is a strong positive correlation between the two; when C is close to -1, it means that there is a strong negative correlation between the two; when C is close to 0, it means that the correlation between the two is weak.

[0128] In step S334, the focus influence coefficient is used to measure the influence of the adjacent monitoring object on the target monitoring object, and it has a positive correlation with the implicit representation involvement coefficient. This positive correlation means that when the implicit representation involvement coefficient is larger, the focus influence coefficient corresponding to the adjacent monitoring object is also larger, indicating that the influence of the adjacent monitoring object on the target monitoring object is more significant.

[0129] For example, in a chemical park, if the implicit representation involvement coefficient C between the raw material storage warehouse and the production workshop is high, indicating that the relationship between them is close, then the raw material storage warehouse as an adjacent monitoring object should have a larger focusing influence coefficient on the production workshop (target monitoring object). The monitoring system can determine the focusing influence coefficient based on the pre-set mapping relationship.

[0130] One mapping relationship can be a linear mapping. Assuming that the implicit representation involvement coefficient C has a value range of [-1,1] and the focusing influence coefficient F has a value range of [0,1], the mapping relationship can be set to F=(C+1) / 2. In this way, when C=-1, F=0, indicating that the adjacent monitoring object has almost no influence on the target monitoring object; when C=1, F=1, indicating that the adjacent monitoring object has the greatest influence on the target monitoring object. Of course, in practical applications, this mapping relationship can be designed in a more complex way according to the specific characteristics and requirements of the monitoring data, such as using piecewise functions or nonlinear mapping relationships.

[0131] In the actual VOCs monitoring and early warning system, in order to optimize the parameter determination process in these steps, a variety of technical means can be used. For example, when determining the second intermediate network component parameter and the third intermediate network component parameter, a training method based on a large amount of historical monitoring data can be used. Collect monitoring data from chemical parks at different times and under different production conditions, including the data on the relationship between monitoring objects, target substance index data, and the corresponding actual VOCs emission result data.

[0132] Based on these data, the back propagation algorithm is used to adjust the parameters of the second intermediate network component and the third intermediate network component. In each iteration, according to the error between the predicted result (the core implicit representation calculated by the current parameter and the implicit representation search, and then the implicit representation involvement coefficient and the focus influence coefficient, and then the predicted VOCs emission obtained in the subsequent steps) and the actual result (the actual VOCs emission), the gradient is calculated and the parameters are updated, so that the model can better fit the data and improve the ability to accurately calculate the focus influence coefficient.

[0133] In addition, the mapping relationship between the implicit representation involvement coefficient and the focused influence coefficient can also be dynamically adjusted according to the actual monitoring data. For example, with the improvement of production processes and the update of environmental protection equipment in the chemical park, the relationship between monitoring objects and the impact pattern on VOCs emissions may change. By regularly analyzing new monitoring data, re-evaluating and adjusting the mapping relationship, it is ensured that the focused influence coefficient can accurately reflect the actual impact of adjacent monitoring objects on the target monitoring object.

[0134] As an implementation mode, the target VOCs monitoring data includes a plurality of monitoring data items, the target VOCs monitoring data implicit representation includes a monitoring data item implicit representation sequence, and the monitoring data item implicit representation sequence includes monitoring data item implicit representations corresponding to each monitoring data item; step S400, implicitly integrating the target VOCs monitoring data implicit representation with the target focused implicit representation corresponding to the target monitoring object to obtain a target integrated implicit representation, including:

[0135] Step S410: performing an implicit representation combination operation on the implicit representation of the monitoring data item corresponding to the target monitoring data item through the target focused implicit representation corresponding to the target monitoring object, so as to obtain a combined implicit representation corresponding to the target monitoring data item;

[0136] Step S420: taking the combined implicit representation corresponding to the target monitoring data item, iterating the monitoring data item implicit representation corresponding to the target monitoring data item in the monitoring data item implicit representation sequence, and obtaining an iterated monitoring data item implicit representation sequence;

[0137] Step S430: using an integrated neural network to perform implicit representation integration on the iterated implicit representation sequence of monitoring data items and the implicit representation of target substance indicators, to obtain an implicit representation sequence of monitoring data items after implicit representation integration and an implicit representation of target substance indicators after implicit representation integration.

[0138] In step S410, the implicit representation of the target VOCs monitoring data includes implicit representations of many monitoring data items, which are feature representations of various monitoring data items (such as the concentration of different types of VOCs substances, data items related to meteorological conditions, etc.) obtained after performing feature extraction and other operations on the original VOCs monitoring data. The target focused implicit representation corresponding to the target monitoring object is a focused representation after comprehensively considering many factors related to the target monitoring object (such as the impact of the monitoring object, etc.).

[0139] For example, in a VOCs monitoring scenario in a city, the target VOCs monitoring data may contain data from multiple monitoring points, where the monitoring data items of one monitoring point may include the concentration of benzene (a VOCs substance), the concentration of toluene, the temperature, humidity, wind speed, etc. After feature extraction, the implicit representation of the monitoring data item corresponding to the concentration of benzene may be a multidimensional vector, in which the dimensions may represent different features, such as the trend characteristics of the change of benzene concentration in different time periods, and the characteristics associated with the concentration of other substances. The target monitoring object is a specific chemical enterprise, and its target focus implicit representation may include the comprehensive characteristics of the impact of the enterprise's own production status on the surrounding VOCs concentration, the impact characteristics of surrounding related enterprises (as involved monitoring objects) on its VOCs emissions, etc.

[0140] When the monitoring system performs implicit representation combination operations, it can adopt multiple methods. One possible method is splicing operation. Assume that the monitoring data item corresponding to the target monitoring data item is implicitly represented as a vector , the target focus corresponding to the target monitoring object is implicitly represented as a vector ,and The dimension is n, The dimension of is m. The combined implicit representation obtained by the concatenation operation is is a vector of dimension n+m, that is =[ ; ]. This concatenation operation combines the implicit representation corresponding to the target monitoring data item with the target focus implicit representation in a new vector space. The new vector contains two aspects of information, including the characteristics of the monitoring data item itself and the characteristics of the relevant influencing factors of the target monitoring object.

[0141] Another possible combination operation is weighted summation. and Set weights separately ( ), then the combination implicitly represents This weighted summation method can adjust the importance of the two parts in the combination according to the actual situation. For example, if it is believed that the impact of the target focused implicit representation on the target monitoring data item is relatively small, w1 can be set larger and w2 smaller.

[0142] In step S420, the monitoring data item implicit representation sequence is a sequence composed of the implicit representations of the individual monitoring data items. In the previous steps, the combined implicit representation corresponding to the target monitoring data item has been obtained, and now the combined implicit representation is used to iteratively update the monitoring data item implicit representation corresponding to the target monitoring data item in the entire monitoring data item implicit representation sequence.

[0143] Continuing with the example of urban VOCs monitoring mentioned above, assume that the initial implicit representation sequence of monitoring data items includes implicit representations of multiple monitoring points and multiple monitoring data items (such as multiple VOCs substance concentrations and meteorological condition data items). For a specific target monitoring data item (such as the implicit representation of the monitoring data item corresponding to the benzene concentration at a certain monitoring point), its combined implicit representation has been obtained. The monitoring system uses this combined implicit representation to update the implicit representation of the monitoring data item corresponding to the original target monitoring data item through an iterative algorithm.

[0144] An iterative algorithm can be based on a weighted average approach. Suppose the monitoring data item corresponding to the original target monitoring data item is implicitly expressed as , the combination is implicitly expressed as , the monitoring data item after iteration is implicitly represented The formula can be Get, where is a weight coefficient between 0 and 1, which determines the relative importance of the original implicit representation and the combined implicit representation in the iterative process. =0.8, it means that in the iteration process, it is more inclined to retain the characteristics of the implicit representation of the original monitoring data items, and also integrates some new information in the combined implicit representation.

[0145] By performing such an iterative operation on the implicit representation of the monitoring data items corresponding to the target monitoring data items in the implicit representation sequence of monitoring data items, the obtained iterative implicit representation sequence of monitoring data items integrates the relevant information in the target focused implicit representation to a certain extent, so that the entire sequence can better reflect the comprehensive situation of the target monitoring object in the entire VOCs monitoring system.

[0146] In step S430, the integrated neural network is a neural network structure specially designed for implicit representation integration. The implicit representation sequence of the monitoring data items after iteration contains the implicit representations of each monitoring data item after processing in the previous steps, and these implicit representations have integrated the relevant information of the target focus implicit representation. The implicit representation of the target substance index is an implicit representation of the characteristics related to the specific substance involved in the target monitoring object.

[0147] For example, in VOCs monitoring of chemical enterprises, the implicit representation of target substance indicators may include the concentration characteristics and chemical activity characteristics of specific VOCs substances (such as styrene) emitted by the enterprise after feature extraction. The integrated neural network processes the implicit representation sequence of the iterative monitoring data items and the implicit representation of the target substance indicators as input.

[0148] The integrated neural network contains multiple neuron layers, and the neurons are connected by weights. Assume that the monitoring data item after iteration implicitly represents a monitoring data item in the sequence implicitly represented as a vector , the target material index is implicitly represented as a vector At the input layer, these vectors are fed into the network and then transformed and fused through the hidden layers.

[0149] In the hidden layer, each neuron performs a linear combination of the input data and adds a bias, and then performs a nonlinear transformation through the activation function. For example, for the jth neuron in the hidden layer, its input is and The combination vector , let the weight matrix be , the bias is , then the result after linear combination is , and then through the activation function (such as ReLU function f(x)=max(0,x)) get .

[0150] After being processed by multiple hidden layers, the output layer obtains the implicit representation sequence of the monitoring data items after implicit representation integration and the implicit representation of the target substance index after implicit representation integration. For example, if the output layer has two parts, one part outputs the implicit representation of the monitoring data items after implicit representation integration and the implicit representation of a monitoring data item in the implicit representation sequence , the other part of the output implicitly represents the integrated target material index implicitly represents .

[0151] When training the integrated neural network, a large amount of known VOCs monitoring data (including implicit representation sequences of monitoring data items, implicit representations of target substance indicators, and corresponding actual VOCs test results) can be used as training sets. By defining a loss function (such as mean square error) and using an optimization algorithm (such as a stochastic gradient descent algorithm) to adjust the weights in the network, the network can accurately integrate the input iterative implicit representation sequences of monitoring data items and implicit representations of target substance indicators, and obtain the results of implicit representation integration that conform to the actual situation.

[0152] In order to better utilize the time series information in the monitoring data, a recurrent neural network (RNN) or its variants (such as a long short-term memory network LSTM or a gated recurrent unit GRU) can be incorporated into the integrated neural network. For example, when processing monitoring data items that implicitly represent a sequence, the memory unit of the LSTM is used to capture the long-term dependencies in the sequence. LSTM controls the flow of information through input gates, forget gates, and output gates, and can better process long-term and short-term information in time series data.

[0153] In practical applications, with the continuous accumulation of monitoring data and changes in the monitoring environment, it may be necessary to dynamically adjust the integrated neural network. For example, when a new enterprise settles in the chemical park or the production process of the enterprise undergoes major changes, the relationship between the target monitoring object and other monitoring objects and the emission characteristics of VOCs may change. At this time, the integrated neural network can be retrained with new monitoring data to update the weights in the network to adapt to the new monitoring situation.

[0154] As an implementation method, step S500, determining the VOCs detection result corresponding to the target VOCs monitoring data through the target integrated implicit representation, includes:

[0155] Step S510: Load the implicit representation of the target substance index after the implicit representation is integrated into the VOCs anomaly detection network that has been debugged in advance to obtain the VOCs anomaly detection result corresponding to the target VOCs monitoring data.

[0156] The implicit representation of the target substance index after implicit representation integration is the result obtained through the previous series of steps. It integrates multiple information such as the implicit representation of the target VOCs monitoring data and the implicit representation of the target focus corresponding to the target monitoring object. For example, in a VOCs monitoring scenario in a chemical park, the implicit representation of the target substance index may contain the integrated information such as the concentration characteristics, chemical activity characteristics, and interaction characteristics of specific VOCs substances (such as benzene, toluene, etc.). This information exists in a specific mathematical structure (such as a vector form), in which each dimension represents the quantitative representation of the relevant characteristics after complex processing.

[0157] The pre-debugged VOCs anomaly detection network is a neural network model specifically used to detect whether VOCs are abnormal. This network has been trained with a large amount of known VOCs data (including normal and abnormal VOCs monitoring data). During the training process, the network learned the relationship between the implicit representation of the target substance index and the abnormal state of VOCs in different situations.

[0158] The monitoring system loads the implicit representation of the target substance index after the implicit representation is integrated into the VOCs anomaly detection network. For example, assume that the VOCs anomaly detection network is a multi-layer perceptron (MLP) structure. The number of neurons in the input layer of the network is determined according to the dimension of the implicit representation of the target substance index after the implicit representation is integrated. When the input data enters the network, the information is processed and converted through the hidden layer.

[0159] In the hidden layer, each neuron performs a linear combination of the input data and adds a bias, and then performs a nonlinear transformation through the activation function. Assume that the input of the i-th neuron in the hidden layer is x i (a dimension or combination of the implicit representation of the target material indicator after the implicit representation integration), with a weight of w ij (the weight connecting the input layer to the i-th hidden layer neuron), with a bias of b i , then the output y of this neuron is i The formula can be Calculated, where f is the activation function, such as the ReLU function f(x)=max(0,x).

[0160] After being processed by multiple hidden layers, the VOCs anomaly detection results are obtained in the output layer. If the output layer is a binary classification problem (i.e., judging whether VOCs are abnormal, normal is 0, abnormal is 1), a single value is output. For example, if the output value is close to 0, it means that according to the implicit representation of the target substance index after the integration of the input implicit representation, the monitoring system determines that the VOCs situation corresponding to the target VOCs monitoring data is normal; if the output value is close to 1, it is judged to be abnormal.

[0161] For multi-classification, for example, to distinguish different degrees of VOCs abnormalities (mild abnormality, moderate abnormality, severe abnormality), the output layer can have multiple neurons, each corresponding to an abnormality category. At this time, the activation value of the output neuron can be used to determine which category the abnormality belongs to.

[0162] In an actual industrial environment, such as a large chemical company with multiple production workshops, the VOCs emitted by different workshops are complex and diverse. Taking one of the workshops that produces organic chemicals as an example, the production process, equipment operation status, and the use of raw materials in the workshop will affect the emission of VOCs. The implicit representation of the target substance index after the implicit representation obtained in the previous steps contains comprehensive information about these complex factors. When the monitoring system loads it into the VOCs anomaly detection network, the network can accurately determine whether there is an abnormality in the VOCs emission of the workshop based on the previously learned pattern.

[0163] In order to build and train this VOCs anomaly detection network, a variety of technical means can be used. In terms of data collection, it is necessary to collect a large amount of VOCs monitoring data under different working conditions, including data during normal production, data during equipment failure, data during different raw material ratios, etc. These data need to be preprocessed, such as data cleaning (removing outliers, erroneous data, etc.), data normalization (mapping data to a specific interval to facilitate network processing), and other operations.

[0164] When training a network, you can use the cross entropy loss function (for multi-classification problems) or the binary cross entropy loss function (for binary classification problems) to measure the error between the predicted results and the actual results. For example, for a binary classification problem, the binary cross entropy loss function , where y is the actual label (0 or 1), is the predicted value of the network. By minimizing this loss function, the weight of the network is adjusted using an optimization algorithm (such as the stochastic gradient descent algorithm) so that the network can accurately classify the implicit representation of the target substance index after integrating the input implicit representation, thereby obtaining accurate VOCs anomaly detection results.

[0165] As an implementation manner, step S410, performing an implicit representation combination operation on the implicit representation of the monitoring data item corresponding to the target monitoring data item through the target focused implicit representation corresponding to the target monitoring object, to obtain the combined implicit representation corresponding to the target monitoring data item, includes:

[0166] Step S411: The target focus implicit representation corresponding to the target monitoring object and the monitoring data item implicit representation corresponding to the target monitoring data item are spliced ​​together to obtain a combined implicit representation corresponding to the target monitoring data item.

[0167] The target focused implicit representation corresponding to the target monitoring object is a feature representation form obtained after comprehensive consideration of multiple factors in the previous steps. It reflects a focus feature of the target monitoring object in the entire VOCs monitoring system. For example, in a VOCs monitoring scenario in a chemical park, the target monitoring object is a large chemical production workshop. Its target focused implicit representation may include the comprehensive characteristics of the impact of factors such as the workshop’s own production scale, production process type, and equipment operation status on the workshop’s VOCs emissions. It also includes the impact characteristics of surrounding related workshops (as involved monitoring objects) on the workshop’s VOCs emissions. After a series of complex calculations and processing, these features exist in a specific mathematical structure (such as a vector form), assuming that the vector quantity , where fi represents the quantized values ​​of different features.

[0168] The implicit representation of the monitoring data item corresponding to the target monitoring data item is part of the implicit representation of the target VOCs monitoring data. It is a feature representation obtained after feature extraction for a specific monitoring data item (such as the concentration of a certain VOCs substance, a parameter in meteorological conditions, etc.). Taking the benzene concentration monitoring data item in the chemical park as an example, its implicit representation of the monitoring data item may include the changing trend characteristics of the benzene concentration in different time periods, the characteristics associated with the concentration of other VOCs substances, and the characteristics of the impact of meteorological conditions on the benzene concentration, etc., which are also expressed in vector form, assuming that the vector .

[0169] When the monitoring system performs the forward and backward splicing operation, it combines the implicit representation of the target focus and the implicit representation of the monitoring data item in order into a new vector. Assume that the combined implicit representation after splicing is the vector ,but , whose dimension is m+n. This front-to-back splicing operation is a simple and effective combination method, which makes the combined implicit representation contain both the focused feature information of the target monitoring object and the feature information of the target monitoring data item itself.

[0170] As an implementation mode, the method is implemented by a neural network, and the method also includes a training process of the neural network, including:

[0171] Step S10: Obtain VOCs monitoring data samples and VOCs detection result labels corresponding to the VOCs monitoring data samples;

[0172] Step S20: loading the VOCs monitoring data sample into the feature extraction network to obtain an implicit representation of the VOCs monitoring data sample, wherein the implicit representation of the VOCs monitoring data sample includes an implicit representation of a sample material index corresponding to the VOCs monitoring data sample;

[0173] Step S30: obtaining a sample monitoring object corresponding to the VOCs monitoring data sample, and determining a second involved monitoring object corresponding to the sample monitoring object;

[0174] Step S40: loading the monitoring object implicit representation corresponding to the second involved monitoring object into the focused implicit representation extraction network, and determining the sample focused implicit representation corresponding to the sample monitoring object through the monitoring object implicit representation and the corresponding focused influence coefficient, wherein the focused influence coefficient is obtained through the involvement coefficient between the sample material indicator implicit representation and the involved relationship implicit representation, and the involved relationship implicit representation is an implicit representation characterizing the involved relationship of the monitoring object;

[0175] Step S50: loading the VOCs monitoring data sample implicit representation and the sample focused implicit representation corresponding to the sample monitoring object into the integrated neural network for implicit representation integration to obtain the sample integrated implicit representation;

[0176] Step S60: Processing the sample integrated implicit representation based on the application layer neural network to obtain the sample VOCs detection result;

[0177] Step S70: Optimize the network configuration variables of the application layer neural network through the sample VOCs detection results and the VOCs detection result labels, or optimize the network configuration variables of the application layer neural network and the pre-trained network, the pre-trained network includes a feature extraction network, a focused implicit representation extraction network and an integrated neural network.

[0178] In step S10, the VOCs monitoring data sample is a data set obtained by monitoring volatile organic pollutants in the actual environment. These data can come from multiple monitoring points and contain various VOCs-related information. For example, multiple monitoring stations are set up in a chemical park, and the data collected by each monitoring station include the concentration of different VOCs substances (such as benzene, toluene, xylene, etc.), meteorological conditions (temperature, humidity, wind speed, wind direction, etc.) and time information. These raw data constitute the VOCs monitoring data sample.

[0179] The VOCs test result label is an identifier corresponding to the monitoring data sample and represents the true state of VOCs. For example, if the VOCs emission in the environment corresponding to a monitoring data sample is within the normal range, its label may be marked as "normal"; if there is abnormal emission (such as the concentration of a certain VOCs substance exceeds the specified threshold), the label is "abnormal". This label provides the target output for the subsequent neural network training, allowing the monitoring system to learn the relationship between the monitoring data and the true results.

[0180] In step S20, the feature extraction network is a neural network structure or algorithm specially designed to mine key features from raw data. Taking the principal component analysis (PCA) algorithm as an example, it can be used as a way to implement the feature extraction network. The monitoring system inputs the acquired VOCs monitoring data samples (assuming it is in a matrix form, each row represents a sample, and each column represents a feature) into the PCA algorithm. PCA calculates the covariance matrix of the data, then performs eigenvalue decomposition on the covariance matrix, selects the eigenvectors corresponding to the main eigenvalues, and projects the original data into a low-dimensional space. The obtained low-dimensional representation is the implicit representation of the VOCs monitoring data sample.

[0181] In step S30, the sample monitoring object is an object that needs to be focused on or analyzed from the VOCs monitoring data sample. For example, in the monitoring scenario of a chemical park, each chemical workshop can be regarded as a sample monitoring object. Determining the second involved monitoring object corresponding to the sample monitoring object is to find other objects that have an associated relationship with the sample monitoring object. This association relationship can be based on a variety of factors.

[0182] Taking the physical space relationship as an example, if a workshop (sample monitoring object) is located downwind of another workshop (second involved monitoring object), then the wind direction factor makes there a correlation between them, and the downwind workshop may be affected by the VOCs emitted by the upwind workshop. From the perspective of production process relationship, if the output of one workshop is the input raw material of another workshop, then there is also a correlation between the two workshops. The monitoring system determines this correlation by analyzing the relevant information in the monitoring data sample (such as workshop location information, production process information, etc.), thereby finding the second involved monitoring object.

[0183] In step S40, the focused implicit representation extraction network is a network structure used to process the implicit representation of the monitored object to obtain a specific focused representation. Assume that the second involved monitored object is a plurality of workshops, each workshop has its corresponding monitored object implicit representation (which is obtained through the previous processing similar to step S20, and contains the abstract representation of various feature information of the workshop itself).

[0184] For example, the implicit representation of a monitoring object in a workshop may include the comprehensive characteristics of the impact of factors such as the state of the workshop's production equipment, the type of production process, and the number of workers on VOCs emissions. The monitoring system loads these implicit representations of monitoring objects into the focused implicit representation extraction network.

[0185] First, the focus influence coefficient needs to be calculated. This involves the involvement coefficient between the implicit representation of the sample material index and the implicit representation of the involvement relationship. Assuming that the implicit representation of the sample material index focuses on the concentration change characteristics of a certain VOCs substance, the implicit representation of the involvement relationship includes factors such as the distance between workshops and wind direction that affect the transmission of VOCs. The involvement coefficient k is calculated by a specific algorithm (such as a vector-based similarity calculation method), and then the focus influence coefficient is determined based on this involvement coefficient (for example, using a certain mapping relationship, such as the focus influence coefficient f=0.5k+0.1).

[0186] Finally, the monitoring system determines the sample focused implicit representation corresponding to the sample monitoring object according to the implicit representation of the monitoring object and the corresponding focusing influence coefficient. For example, assuming that there are n second involved monitoring objects, the monitoring object implicit representation of the i-th second involved monitoring object is , and its corresponding focusing influence coefficient is , then the sample focus implicit representation corresponding to the sample monitoring object It can be obtained by weighted summation, that is, .

[0187] In step S50, the integrated neural network is a network structure specifically used to integrate different implicit representations. The monitoring system loads the VOCs monitoring data sample implicit representation (including various feature information of the monitoring data sample) and the sample focused implicit representation corresponding to the sample monitoring object (taking into account the influence of other objects related to the sample monitoring object) as input into the integrated neural network.

[0188] For example, VOCs monitoring data samples are implicitly represented as vectors \vec{A}, and sample focus is implicitly represented as vectors The integrated neural network processes through multiple neuron layers. After the input layer receives these two vectors, the neurons in the hidden layer convert and fuse the information. Assume that the input of the jth neuron in the hidden layer is [ ; ] (the result of concatenating the two vectors), the weight matrix is , the bias is , then the output of the neuron The formula can be Calculated, where f is the activation function, such as the ReLU function f(x)=max(0,x). After being processed by multiple hidden layers, the sample integrated implicit representation is obtained in the output layer.

[0189] In step S60, the application layer neural network is a network structure constructed for a specific VOCs detection task. It receives the sample integrated implicit representation as input and outputs the sample VOCs detection result. For example, if the application layer neural network is a simple multi-layer perceptron (MLP), the sample integrated implicit representation is a vector The input layer of the MLP receives Finally, the information is processed through the hidden layer.

[0190] Assume that the number of neurons in the hidden layer is m. For each neuron k in the hidden layer, its input is The weights of each element are , with a bias of b k , then the output of neuron k is The formula can be Calculated, where yes The elements of , f is the activation function. After being processed by the hidden layer, the sample VOCs detection results are obtained in the output layer. If it is a binary classification problem (such as normal and abnormal), the output layer may have only one neuron, and the output value close to 1 indicates abnormality, and close to 0 indicates normality; if it is a multi-classification problem (such as different degrees of abnormality), the output layer has multiple neurons, and the category is determined according to the size of the output value.

[0191] In step S70, the monitoring system adjusts the parameters of the network according to the difference between the sample VOCs detection result and the real VOCs detection result label to improve the accuracy of the network. If only the application layer neural network is optimized, for example, for the above-mentioned MLP, a back propagation algorithm can be used.

[0192] The monitoring system calculates the gradient of the loss function with respect to network configuration variables (such as weights and biases), and then updates these variables according to the gradient descent algorithm. If you want to optimize the network configuration variables of the application layer neural network and the pre-trained network, you need to consider the joint optimization of the entire network structure. This may involve more complex algorithms, such as gradient-based optimization algorithms, while considering the relationships between the various networks. For example, when updating the parameters of the feature extraction network, it is necessary to consider its impact on subsequent networks (such as the focused implicit representation extraction network, the integrated neural network, and the application layer neural network). Through this joint optimization, the entire neural network system can be made more accurate and efficient in VOCs monitoring and early warning tasks.

[0193] In a VOCs monitoring scenario in a large industrial area, such as one that includes multiple chemical companies, different types of production workshops, and complex meteorological and geographical environments, this neural network training process is particularly important. Factors such as the production process, equipment status, and raw material usage of different workshops will affect the emission of VOCs, while meteorological conditions (such as wind direction, temperature, and humidity) will affect the diffusion of VOCs. Through the above-mentioned neural network training process, the monitoring system can learn the relationship between these complex factors, thereby accurately monitoring and warning VOCs.

[0194] In actual operation, in order to improve the training effect, some technical means can also be used. For example, in the data preprocessing stage, the VOCs monitoring data samples can be normalized and the data can be mapped to a specific interval (such as [0,1] or [-1,1]), which helps to improve the training speed and stability of the neural network. At the same time, regularization techniques (such as L1 regularization or L2 regularization) can be used to prevent network overfitting and improve the generalization ability of the network. In addition, during the training process, batch training can be used to divide a large number of monitoring data samples into multiple batches for training, which helps to reduce memory usage and improve training efficiency.

[0195] The present application embodiment provides a monitoring system, such as Figure 2 As shown, the monitoring system 100 includes: a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, such as through a bus 102. Optionally, the monitoring system 100 may also include a transceiver 104. It should be noted that in actual applications, the transceiver 104 is not limited to one, and the structure of the monitoring system 100 does not constitute a limitation on the embodiments of the present application.

[0196] An embodiment of the present application provides a monitoring system, and the monitoring system in the embodiment of the present application includes: one or more processors; a memory; one or more computer programs, wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more programs are executed by the processor, the above-mentioned online monitoring and early warning method for volatile organic pollutants is implemented.

Claims

1. A method for online monitoring and early warning of volatile organic pollutants, characterized in that: The method comprises: Acquire target VOCs monitoring data to be detected, perform feature extraction on the target VOCs monitoring data, and obtain an implicit representation of the target VOCs monitoring data, wherein the implicit representation of the target VOCs monitoring data includes an implicit representation of a target substance index corresponding to the target VOCs monitoring data; Acquire a target monitoring object in the target VOCs monitoring data, and determine a first involved monitoring object corresponding to the target monitoring object; Determine the target focus implicit representation corresponding to the target monitoring object through the first monitoring object implicit representation and the corresponding focus influence coefficient, wherein the focus influence coefficient is obtained through the involvement coefficient between the target material index implicit representation and the involvement relationship implicit representation, and the involvement relationship implicit representation is an implicit representation that characterizes the monitoring object involvement relationship; Integrating the implicit representation of the target VOCs monitoring data with the target focused implicit representation corresponding to the target monitoring object to obtain a target integrated implicit representation; Determine the VOCs detection result corresponding to the target VOCs monitoring data through the target integrated implicit representation; The determining of the target focus implicit representation corresponding to the target monitoring object by using the first monitoring object implicit representation involving the monitoring object and the corresponding focus influence coefficient includes: Acquire an object involvement map formed by the first involved monitoring object and the target monitoring object; For a graph monitoring object in the object involvement graph, obtaining an implicit representation of an involvement relationship representing an involvement relationship between the graph monitoring object and an adjacent monitoring object; Obtaining an implicit representation involvement coefficient through the implicit representation of the involvement relationship and the implicit representation of the target material index, and determining a focusing influence coefficient corresponding to the adjacent monitoring object through the implicit representation involvement coefficient; Determine the target focus implicit representation corresponding to the atlas monitoring object through the focus influence coefficient corresponding to the adjacent monitoring object and the monitoring object implicit representation of the adjacent monitoring object; Extracting the target focus implicit representation corresponding to the target monitoring object from the target focus implicit representations corresponding to each graph monitoring object of the object involvement graph; The target focused implicit representation corresponding to the atlas monitoring object is obtained by executing a focused implicit representation extraction network, the focused implicit representation extraction network includes one or more target intermediate network components, the implicit representation involvement coefficient is obtained by the implicit representation of the involvement relationship and the implicit representation of the target material index, and the focusing influence coefficient corresponding to the adjacent monitoring object is determined by the implicit representation involvement coefficient, including: Taking a second intermediate network component parameter in the target intermediate network component to process the implicit representation of the involvement relationship to obtain a core implicit representation; Taking the third intermediate network component parameter in the target intermediate network component to process the target material index implicit representation to obtain the search implicit representation; Determining an implicit representation involvement coefficient by using the core implicit representation and the search implicit representation; The focusing influence coefficient corresponding to the adjacent monitoring object is determined by the implicit representation involvement coefficient, and the implicit representation involvement coefficient is positively correlated with the focusing influence coefficient corresponding to the adjacent monitoring object.

2. The method according to claim 1, characterized in that The target focused implicit representation corresponding to the atlas monitoring object is obtained by processing a focused implicit representation extraction network, wherein the focused implicit representation extraction network includes one or more target intermediate network components, and the target focused implicit representation corresponding to the atlas monitoring object is determined by using the focused influence coefficient corresponding to the adjacent monitoring object and the monitoring object implicit representation of the adjacent monitoring object, including: Loading the monitoring object implicit representation of the adjacent monitoring object and the implicit representation of the involved relationship into the target intermediate network component, and executing to obtain the first focused implicit representation corresponding to the atlas monitoring object; The target focus implicit representation corresponding to the atlas monitoring object is determined by the first focus implicit representation corresponding to the atlas monitoring object and the focus influence coefficient corresponding to the corresponding adjacent monitoring object.

3. The method according to claim 2, characterized in that The step of loading the monitoring object implicit representation of the adjacent monitoring object and the implicit representation of the involved relationship into the target intermediate network component and executing to obtain the first focused implicit representation corresponding to the graph monitoring object includes: Determine a target operation mode through a monitoring object involvement relationship between the atlas monitoring object and the adjacent monitoring objects, wherein the target operation mode is summing or difference; The implicit representation of the monitoring object of the adjacent monitoring object and the implicit representation of the involved relationship are operated by the target operation mode to obtain the implicit representation after operation corresponding to the atlas monitoring object; The intermediate network component parameter in the target intermediate network component is used to process the calculated implicit representation to obtain a first focused implicit representation corresponding to the atlas monitoring object.

4. The method according to claim 3, characterized in that The step of taking the intermediate network component parameter in the target intermediate network component to process the calculated implicit representation to obtain the first focused implicit representation corresponding to the graph monitoring object includes: Obtaining an implicit representation of an execution result corresponding to the adjacent monitoring object obtained by executing a previous intermediate network component corresponding to the target intermediate network component in the focused implicit representation extraction network; The target intermediate network component uses the first intermediate network component parameter to process the post-operation implicit representation and the execution result implicit representation to obtain a first focused implicit representation corresponding to the graph monitoring object.

5. The method according to claim 4, characterized in that The target VOCs monitoring data includes a plurality of monitoring data items, the target VOCs monitoring data implicit representation includes a monitoring data item implicit representation sequence, and the monitoring data item implicit representation sequence includes monitoring data item implicit representations corresponding to each monitoring data item; The implicitly integrating the target VOCs monitoring data implicit representation with the target focused implicit representation corresponding to the target monitoring object to obtain the target integrated implicit representation includes: Performing an implicit representation combination operation on the implicit representation of the monitoring data item corresponding to the target monitoring data item through the target focused implicit representation corresponding to the target monitoring object, so as to obtain a combined implicit representation corresponding to the target monitoring data item; Taking the combined implicit representation corresponding to the target monitoring data item, iterating the monitoring data item implicit representation corresponding to the target monitoring data item in the monitoring data item implicit representation sequence, and obtaining an iterated monitoring data item implicit representation sequence; An integrated neural network is used to implicitly integrate the implicit representation sequence of the iterated monitoring data items and the implicit representation of the target substance index to obtain an implicit representation sequence of the monitoring data items after implicit representation integration and an implicit representation of the target substance index after implicit representation integration.

6. The method according to claim 5, characterized in that Determining the VOCs detection result corresponding to the target VOCs monitoring data through the target integrated implicit representation includes: The implicit representation of the target substance index after the implicit representation is integrated is loaded into the VOCs anomaly detection network that has been debugged in advance to obtain the VOCs anomaly detection result corresponding to the target VOCs monitoring data; The step of performing an implicit representation combination operation on the implicit representation of the monitoring data item corresponding to the target monitoring data item through the target focused implicit representation corresponding to the target monitoring object to obtain the combined implicit representation corresponding to the target monitoring data item includes: The combined implicit representation corresponding to the target monitoring data item is obtained by splicing the target focused implicit representation corresponding to the target monitoring object and the monitoring data item implicit representation corresponding to the target monitoring data item forward and backward.

7. The method according to claim 6, characterized in that The method is implemented by a neural network, and the method also includes: Obtaining a VOCs monitoring data sample and a VOCs detection result label corresponding to the VOCs monitoring data sample; Loading the VOCs monitoring data sample into a feature extraction network to obtain an implicit representation of the VOCs monitoring data sample, wherein the implicit representation of the VOCs monitoring data sample includes an implicit representation of a sample material index corresponding to the VOCs monitoring data sample; Acquire a sample monitoring object corresponding to the VOCs monitoring data sample, and determine a second involved monitoring object corresponding to the sample monitoring object; The monitoring object implicit representation corresponding to the second involved monitoring object is loaded into the focused implicit representation extraction network, and the sample focused implicit representation corresponding to the sample monitoring object is determined through the monitoring object implicit representation and the corresponding focused influence coefficient, wherein the focused influence coefficient is obtained through the involvement coefficient between the sample material indicator implicit representation and the involved relationship implicit representation, and the involved relationship implicit representation is an implicit representation characterizing the involved relationship of the monitoring object; Loading the VOCs monitoring data sample implicit representation and the sample focused implicit representation corresponding to the sample monitoring object into an integrated neural network for implicit representation integration to obtain a sample integrated implicit representation; The sample integrated implicit representation is processed based on an application layer neural network to obtain a sample VOCs detection result; The network configuration variables of the application layer neural network are optimized by using the sample VOCs detection results and the VOCs detection result labels, or the network configuration variables of the application layer neural network and the pre-trained network are optimized, and the pre-trained network includes the feature extraction network, the focused implicit representation extraction network and the integrated neural network.

8. A monitoring system, characterized in that: include: one or more processors; Memory; one or more computer programs; The one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method according to any one of claims 1 to 7 is implemented.

Images