A model training system, method and online early warning device for volatile organic compound prediction based on machine learning

Abstract

The present application provides a model training system, method and online early warning device for volatile organic compound prediction based on machine learning, wherein the model training system includes: a data acquisition module; a data preprocessing module; a data training module; a model evaluation module. Through the above-mentioned model training system, method and online early warning device for volatile organic compound prediction based on machine learning of the present application, a qualified final prediction model can be obtained, and the qualified final prediction model can be used to predict the predicted value of the standard equipment detection data through the PID sensor detection data and meteorological data obtained in real time at the target monitoring point; detection through PID has the advantages of shorter PID monitoring frequency, small size, low price and easy to deploy in large numbers. In addition, the predicted value of the predicted standard equipment detection data is the mass concentration of each substance, which can be used to determine the specific detection value of each substance for alarm.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a model training system, method, and online early warning device for volatile organic compound prediction based on machine learning. Background Art

[0002] In my country, the mass concentration (μg / m 3 ) is used as the existing alarm limit unit to reflect and alarm the detection concentration of each organic matter.

[0003] Currently, standard equipment such as gas chromatography-mass spectrometry (GC-MS) or gas chromatography-flame ionization detector (GC-FID) is commonly used to detect VOCs concentration. They have the advantages of high sensitivity, wide linear range, and good stability. In particular, they can analyze single VOC species and measure the concentration in terms of mass concentration (μg / m 3 However, such equipment is relatively expensive, takes a long time to analyze, lacks temporal representativeness, has high requirements for the working environment, and has high maintenance costs.

[0004] Photo Ionization Detector (PID) is one of the detection methods for VOCs. It uses ultraviolet lamp (UV) as a light source to break the substance into positive and negative ions (ionization) that can be detected by the detector. The detector measures the charge of the ionized gas and converts it into a current signal. The current is amplified and displays the corresponding concentration value. After being detected, the ions recombine into the original gas and vapor. It is a detector with a wide range of uses. PID has the advantages of shorter monitoring frequency (minutes or even seconds), small size, low price, and easy to deploy in large quantities. However, it cannot monitor a single VOC substance, but responds to a certain type or multiple types of pollutants (such as aromatic hydrocarbons and alkanes). The detection results are the total volume fraction of VOCs, and since the composition ratio of VOCs is unknown, the volume fraction cannot be converted into mass concentration (μg / m 3 ), can only be displayed as a volume percentage and cannot be compared with the existing alarm limit unit of μg / m 3 No corresponding comparison is performed, so that the real alarm function cannot be achieved.

[0005] Therefore, there is an urgent need for a method that can quickly detect and measure the concentration of 3 ) reflects the online VOCs monitoring plan for industrial parks for each substance being tested. Summary of the Invention

[0006] To address the problems in the prior art, this application provides a model training system, model training method, and online early warning device for volatile organic compound prediction based on machine learning. The technical solution of this application is as follows:

[0007] 1. A model training system for volatile organic compound prediction based on machine learning, comprising:

[0008] A data acquisition module, wherein the data acquisition module acquires detection data of target monitoring points as training sets and test sets; wherein the detection data includes standard equipment detection data of volatile organic compound concentration, PID sensor detection data of volatile organic compound concentration, and meteorological data;

[0009] A data preprocessing module, which preprocesses the detection data to obtain a processed training set and a processed test set;

[0010] a data training module, wherein the data training module trains the processed training set to obtain a final prediction model for predicting standard equipment detection data based on PID sensor detection data and meteorological data;

[0011] The model evaluation module compares the predicted value of the standard equipment detection data obtained by using the final prediction model to predict the test set with the actual value of the standard equipment detection data to obtain an evaluation result.

[0012] 2. The model training system according to claim 1, wherein:

[0013] The meteorological data is selected from one or more of wind direction, wind speed, temperature, humidity, and air pressure.

[0014] 3. The model training system according to claim 1, wherein:

[0015] The data preprocessing module includes:

[0016] a data separation component, wherein the data separation component divides the standard equipment detection data of the training set and the standard equipment detection data of the test set into training label data and test label data, and divides the PID sensor detection data and the meteorological data of the training set and the PID sensor detection data and the meteorological data of the test set into training feature data and test feature data;

[0017] A feature standardization component, wherein the feature standardization component standardizes the training feature data and the test feature data to obtain training feature standardized data and test feature standardized data respectively;

[0018] A dimension-raising component is provided for respectively raising the dimension of the training feature normalized data and the test feature normalized data to obtain training feature dimension-raised data and test feature dimension-raised data, so as to respectively obtain a processed training set consisting of training label data and training feature dimension-raised data, and a processed test set consisting of test label data and test feature dimension-raised data.

[0019] 4. The model training system according to item 3, wherein:

[0020] The training feature data and the test feature data are respectively standardized as follows:

[0021] The training feature data and the test feature data are made to have zero mean and unit variance.

[0022] 5. The model training system according to item 4, wherein:

[0023] The training feature data and the test feature data are made to have zero mean and unit variance as follows:

[0024] For each feature j, first calculate the mean µ of all samples j and standard deviation σ j :

[0025]

[0026]

[0027] in:

[0028] m is the number of samples;

[0029] x ij is the original value of the i-th sample on the j-th feature;

[0030] µ j is the mean of the jth feature;

[0031] σ j is the standard deviation of the jth feature.

[0032] Use the mean and standard deviation of the features to transform the original data to standardize it:

[0033]

[0034] Where: z ij is the normalized value of the i-th sample on the j-th feature.

[0035] 6. The model training system of item 3, wherein:

[0036] The training feature standardized data and the test feature standardized data are respectively upgraded to:

[0037] The training feature normalized data and the test feature normalized data are mapped to a high-dimensional space using a polynomial kernel function.

[0038] 7. The model training system according to item 6, wherein:

[0039] The polynomial kernel function is:

[0040]

[0041] in,

[0042] x i is the feature vector representing the i-th sample;

[0043] x j is the feature vector representing the jth sample;

[0044] <x i ,x j > represents a vector x i and x j The inner product of

[0045] d is the degree of the kernel function.

[0046] 8. The model training system of item 1, wherein:

[0047] The data training module includes:

[0048] A base model component, the base model component including two or more base training models, each base training model is trained on the processed training set to obtain two or more base prediction models;

[0049] An integrated model component integrates two or more base prediction models into a final prediction model, and the final prediction model can obtain a final prediction result based on the prediction results of the two or more base prediction models.

[0050] 9. The model training system of item 8, wherein:

[0051] The base training model is selected from two or more models of a linear regression model, a gradient boosting regression model, a support vector regression model, and a decision tree regression model.

[0052] 10. The model training system of item 8, wherein:

[0053] The final prediction model can obtain the final prediction result according to the prediction results of two or more base prediction models:

[0054] Two or more base prediction models independently predict to obtain prediction results respectively, and the majority voting result or the average prediction result of the prediction results of each base prediction model is used as the final prediction result.

[0055] 11. The model training system of item 1, wherein:

[0056] The model evaluation module includes:

[0057] A prediction component capable of using a final prediction model to obtain a predicted value of standard equipment test data from the processed test set;

[0058] An evaluation component compares the predicted value of the standard equipment detection data with the actual value of the standard equipment detection data to obtain an evaluation result.

[0059] 12. The model training system of item 11, wherein:

[0060] The evaluation component compares the predicted value of the standard equipment detection data with the actual value of the standard equipment detection data to obtain an evaluation result:

[0061] The evaluation component calculates the mean square error, determination coefficient, root mean square error and / or mean absolute error between the predicted value of the standard equipment detection data and the true value of the standard equipment detection data as an evaluation result.

[0062] 13. A model training method for volatile organic compound prediction based on machine learning, comprising:

[0063] a data acquisition step, obtaining detection data of target monitoring points and dividing the data into a training set and a test set; wherein the detection data includes standard equipment detection data of volatile organic compound concentration, PID sensor detection data of volatile organic compound concentration, and meteorological data;

[0064] A data preprocessing step of preprocessing the detection data to obtain a processed training set and a processed test set;

[0065] a data training step of training the processed training set to obtain a final prediction model for predicting standard equipment detection data based on PID sensor detection data and meteorological data;

[0066] The model evaluation step compares the predicted value of the standard equipment test data obtained by using the final prediction model to predict the test set with the actual value of the standard equipment test data to obtain the evaluation result.

[0067] 14. The model training method of item 13, wherein:

[0068] The meteorological data is selected from one or more of wind direction, wind speed, temperature, humidity, and air pressure.

[0069] 15. The machine learning-based model training method of item 13, wherein:

[0070] The data preprocessing step includes:

[0071] a data separation sub-step of dividing the standard equipment detection data of the training set and the standard equipment detection data of the test set into training label data and test label data, respectively, and dividing the PID sensor detection data and the meteorological data of the training set and the PID sensor detection data and the meteorological data of the test set into training feature data and test feature data, respectively;

[0072] A feature standardization sub-step of standardizing the training feature data and the test feature data to obtain training feature standardized data and test feature standardized data respectively;

[0073] In the dimensionality upgrading sub-step, the dimensions of the training feature standardized data and the test feature standardized data are upgraded to obtain training feature upgraded dimension data and test feature upgraded dimension data, respectively, so as to obtain a processed training set consisting of training label data and training feature upgraded dimension data, and a processed test set consisting of test label data and test feature upgraded dimension data, respectively.

[0074] 16. The model training method of item 15, wherein:

[0075] The training feature data and the test feature data are respectively standardized as follows:

[0076] The training feature data and the test feature data are made to have zero mean and unit variance.

[0077] 17. The model training method of item 16, wherein:

[0078] The training feature data and the test feature data are made to have zero mean and unit variance as follows:

[0079] For each feature j, first calculate the mean µ of all samples j and standard deviation σ j :

[0080]

[0081]

[0082] in:

[0083] m is the number of samples;

[0084] x ij is the original value of the i-th sample on the j-th feature;

[0085] µ j is the mean of the jth feature;

[0086] σ j is the standard deviation of the jth feature;

[0087] Use the mean and standard deviation of the features to transform the original data to standardize it:

[0088]

[0089] Where: z ij is the normalized value of the i-th sample on the j-th feature.

[0090] 18. The model training method of item 15, wherein:

[0091] The training feature standardized data and the test feature standardized data are respectively upgraded to:

[0092] The training feature normalized data and the test feature normalized data are mapped to a high-dimensional space using a polynomial kernel function.

[0093] 19. The model training method of item 18, wherein:

[0094] The polynomial kernel function is:

[0095]

[0096] in,

[0097] x i is the feature vector representing the i-th sample;

[0098] x j is the feature vector representing the jth sample;

[0099] <x i ,x j > represents a vector x i and x j The inner product of

[0100] d is the degree of the kernel function.

[0101] 20. The model training method of item 13, wherein:

[0102] The data training step includes:

[0103] In a base model sub-step, the processed training set is trained respectively by two or more base training models to obtain two or more base prediction models;

[0104] In the integrated model sub-step, the final prediction model can obtain the final prediction result based on the prediction results of more than two base prediction models.

[0105] 21. The model training method of item 20, wherein:

[0106] The base training model is selected from two or more models of a linear regression model, a gradient boosting regression model, a support vector regression model, and a decision tree regression model.

[0107] 22. The model training method of item 20, wherein:

[0108] The final prediction model can obtain the final prediction result according to the prediction results of two or more base prediction models:

[0109] Two or more base prediction models independently predict to obtain prediction results respectively, and the majority voting result or the average prediction result of the prediction results of each base prediction model is used as the final prediction result.

[0110] 23. The model training method of item 13, wherein:

[0111] The model evaluation step includes:

[0112] In the prediction sub-step, the final prediction model is used to obtain the predicted value of the standard equipment test data from the processed test set;

[0113] The evaluation sub-step compares the predicted value of the standard equipment detection data with the actual value of the standard equipment detection data to obtain an evaluation result.

[0114] 24. The model training method of item 23, wherein:

[0115] Comparing the predicted value of the standard equipment test data with the actual value of the standard equipment test data to obtain the evaluation result is:

[0116] The mean square error, determination coefficient, mean absolute error and / or mean absolute percentage error between the predicted value of the standard equipment test data and the true value of the standard equipment test data are calculated as the evaluation result.

[0117] 25. A volatile organic compound online early warning device based on machine learning, comprising:

[0118] More than one detection module, each of which includes a PID sensor, and each detection module detects the concentration of volatile organic compounds at a corresponding target monitoring point to obtain PID sensor detection data;

[0119] A data acquisition module, wherein the data acquisition module acquires the detection data of the PID sensor and the meteorological data of the target monitoring point as prediction data;

[0120] A prediction module, wherein the prediction module is deployed with a final prediction model, and the prediction module obtains a prediction result of a corresponding target monitoring point from the prediction data based on the final prediction model;

[0121] Wherein, the final prediction model is obtained by the model training system described in any one of items 1 to 12 or the model training method described in any one of items 13 to 24.

[0122] 26. The online early warning device for volatile organic compounds based on machine learning as described in item 25, wherein:

[0123] The detection module further includes a meteorological detection device, which detects meteorological data of the target monitoring point.

[0124] 27. The machine learning-based online early warning device for volatile organic compounds according to item 25 or 26, wherein:

[0125] The meteorological data is selected from one or more of wind direction, wind speed, temperature, humidity, and air pressure.

[0126] Through the above-mentioned machine learning-based model training system, method and online early warning device for volatile organic compound prediction of the present application, a qualified final prediction model can be obtained, and the qualified final prediction model can be used to predict the predicted value (μg / m2) of the standard equipment detection data through the PID sensor detection data (ppb) and meteorological data obtained in real time at the target monitoring point. 3 PID detection has the advantages of shorter monitoring frequency (minutes or even seconds), small size, low price, and easy deployment of large numbers of points. In addition, the predicted value of the standard equipment detection data is the mass concentration of each substance (μg / m 3 ), which can be used to determine the specific detection value of each substance for alarm.

[0127] The above description is only an overview of the technical solution of the present application. In order to make the technical means of the present application clearer and easier to understand, so that those skilled in the art can implement it according to the contents of the specification, and to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are illustrated below. BRIEF DESCRIPTION OF THE DRAWINGS

[0128] Figure 1 : A schematic diagram of a model training system for volatile organic compound prediction based on machine learning in an embodiment of the present application;

[0129] Figure 2: A schematic diagram of a model training method for volatile organic compound prediction based on machine learning in an embodiment of the present application;

[0130] Figure 3 : Schematic diagram of an online early warning device for volatile organic compounds based on machine learning in an embodiment of the present application. DETAILED DESCRIPTION

[0131] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described herein.

[0132] Application Overview

[0133] As mentioned above, for online VOCs monitoring in industrial parks, standard equipment such as gas chromatography-mass spectrometry (GC-MS) or gas chromatography-flame ionization detector (GC-FID) is generally used for detection. Although it can detect and analyze individual VOC species, it is based on the existing alarm limit unit of mass concentration (μg / m 3 ) to reflect, but such equipment is relatively expensive, has high maintenance costs, and the analysis time is long, resulting in insufficient time representativeness. In addition, PID is also a commonly used VOCs detection solution, but it can only detect the total volume fraction of VOCs, not a single VOC, and cannot detect VOC by mass concentration (μg / m 3 ) reflects the detection results of a single VOC and is used for VOCs over-limit alarm. Since the composition ratio of VOCs is unknown in the PID monitoring solution, technicians generally cannot accurately convert the PID monitoring results (volume fraction) into mass concentration (μg / m 3 ), and is used to detect the bias of single VOC mass concentration in VOCs and its over-limit alarm.

[0134] For online VOCs monitoring in industrial parks, the following technical solution not only has the advantages of PID's shorter monitoring frequency (minutes or even seconds), small size, low price, and ease of large-scale deployment, but also has the ability to actually detect the mass concentration (μg / m3) of individual substances in VOCs compared to standard equipment. 3 ) and the level of over-limit alarm.

[0135] Exemplary Systems

[0136] Figure 1 Illustration of a machine learning-based model training system for volatile organic compound prediction.

[0137] like Figure 1As shown, this embodiment provides a model training system for volatile organic compound prediction based on machine learning, which includes:

[0138] A data acquisition module, wherein the data acquisition module acquires detection data of target monitoring points as training sets and test sets; wherein the detection data includes standard equipment detection data of volatile organic compound concentration, PID sensor detection data of volatile organic compound concentration, and meteorological data;

[0139] A data preprocessing module, which preprocesses the detection data to obtain a processed training set and a processed test set;

[0140] a data training module, wherein the data training module trains the processed training set to obtain a final prediction model for predicting standard equipment detection data based on PID sensor detection data and meteorological data;

[0141] The model evaluation module compares the predicted value of the standard equipment detection data obtained by using the final prediction model to predict the test set with the actual value of the standard equipment detection data to obtain an evaluation result.

[0142] Regarding the collection time and quantity of detection data, it can be recent historical data of the target monitoring point, such as historical data within a month, a quarter, or a year. In addition, those skilled in the art will recognize that it is necessary to collect an appropriate amount of sample data and an appropriate distribution of sample collection times to facilitate subsequent machine learning training and testing.

[0143] In this application, “standard equipment” refers to equipment that can detect the mass concentration (μg / m 3 ) equipment, specifically, gas chromatography-mass spectrometry (GC-MS) and gas chromatography-flame ionization detector (GC-FID) which are commonly used at present.

[0144] Meteorological data can be selected from one or more of wind direction, wind speed, temperature, humidity, and air pressure. In this embodiment, the meteorological data selected by the inventors includes wind direction, wind speed, temperature, humidity, and air pressure at the target monitoring point, which can achieve excellent prediction results for individual VOC mass concentrations (especially for individual VOC mass concentrations of benzene compounds). Benzene compounds are organic compounds containing a benzene ring structure, also known as aromatic compounds, such as toluene and xylene.

[0145] Through the above-mentioned machine learning-based model training system for volatile organic compound prediction of the present application (hereinafter referred to as "the system of the present application"), the detection data (including training set and test set) of the target monitoring point can be obtained through the data acquisition module for subsequent processing; the detection data (including training set and test set) can be preprocessed (such as standardization, etc.) through the data preprocessing module for subsequent machine learning; through the data training module, the PID sensor detection data and meteorological data in the processed training set are regressed with the standard equipment detection data to construct a final prediction model that predicts the standard equipment detection data from the PID sensor detection data and meteorological data; the model evaluation module can evaluate the final prediction model. Therefore, the qualified final prediction model can be used to predict the predicted value (μg / m 3 ), which is detected by PID, has the advantages of shorter monitoring frequency (minutes or even seconds), small size, low price, and easy to deploy in large quantities. In addition, the predicted value of the standard equipment test data is the mass concentration of each substance (μg / m 3 ), which can be used to determine the specific detection value of each substance for alarm, thereby solving the problems of the above-mentioned prior art.

[0146] In addition, by setting the above-mentioned PID sensors at multiple points in the industrial park, the system of this application can predict the VOCs prediction results at multiple points separately, so as to facilitate timely and comprehensive understanding of the pollution status and trends in the entire park.

[0147] In addition, it should be noted that the correction coefficient is an important parameter of the PID sensor. The correction coefficient is a measure of the sensitivity of the PID to a specific gas. The system of the present application has a better prediction effect for compounds with lower correction coefficients. Since the correction coefficients of the main VOCs (such as benzene series, etc.) for PID sensors are relatively low, the system of the present application can be used in most industrial parks. Of course, for special scenarios, those skilled in the art can decide whether to adopt the system of the present application based on the industrial distribution of the industrial park. For example, benzene series compounds are the main VOCs in the industrial park where the applicant is conducting experiments, and the system of the present application has a good prediction effect for benzene series compounds.

[0148] Regarding the data preprocessing module, Figure 1 As shown, it includes:

[0149] a data separation component, wherein the data separation component divides the standard equipment detection data of the training set and the standard equipment detection data of the test set into training label data and test label data, and divides the PID sensor detection data and the meteorological data of the training set and the PID sensor detection data and the meteorological data of the test set into training feature data and test feature data;

[0150] A feature standardization component, wherein the feature standardization component standardizes the training feature data and the test feature data to obtain training feature standardized data and test feature standardized data respectively;

[0151] A dimension-raising component is provided for respectively raising the dimension of the training feature normalized data and the test feature normalized data to obtain training feature dimension-raised data and test feature dimension-raised data, so as to respectively obtain a processed training set consisting of training label data and training feature dimension-raised data, and a processed test set consisting of test label data and test feature dimension-raised data.

[0152] Before subsequent training or testing of the test data used as training and test sets, they must first be processed to achieve better training and testing results. Specifically, in this solution, the training and test sets are first divided. Using a data separation component, the training set is divided into training label data (corresponding to standard equipment test data) and training feature data (corresponding to PID sensor test data and meteorological data), and the test set is divided into test label data (corresponding to standard equipment test data) and test feature data (corresponding to PID sensor test data and meteorological data). This facilitates the construction of a model that predicts standard equipment test data from PID sensor test data and meteorological data. Furthermore, to achieve better training results, the training and test feature data are further standardized using a feature normalization component to accelerate the convergence of the machine learning algorithm and improve model performance. Furthermore, the standardized data (training feature normalized data and test feature normalized data) needs to be further dimensionally upgraded, projecting sample points to a higher dimension to separate mixed sample points, thereby preventing underfitting and improving model accuracy.

[0153] In this embodiment, standardizing the training feature data and the test feature data specifically refers to: making the training feature data and the test feature data have zero mean and unit variance, specifically:

[0154] For each feature j, first calculate the mean µ of all samples j and standard deviation σ j :

[0155]

[0156]

[0157] in:

[0158] m is the number of samples;

[0159] x ij is the original value of the i-th sample on the j-th feature;

[0160] µ j is the mean of the jth feature;

[0161] σ j is the standard deviation of the jth feature.

[0162] Use the mean and standard deviation of the features to transform the original data to standardize it:

[0163]

[0164] Where: z ij is the normalized value of the i-th sample on the j-th feature.

[0165] In addition, the dimension of the training feature standardized data and the test feature standardized data are respectively increased as follows: the training feature standardized data and the test feature standardized data are mapped to a high-dimensional space using a polynomial kernel function.

[0166] Wherein, the polynomial kernel function is:

[0167]

[0168] in,

[0169] x i is the feature vector representing the i-th sample;

[0170] x j is the feature vector representing the jth sample;

[0171] <x i ,x j > represents a vector x i and x j The inner product of

[0172] d is the degree of the kernel function.

[0173] A polynomial kernel function with a lower degree d may not be able to adequately capture complex relationships in the data, resulting in insufficient model fit. A polynomial kernel function with a higher degree d provides stronger fit and can fit more complex data patterns, but generally results in a more complex model and requires careful adjustment to avoid overfitting. In this example, d is set to 3.

[0174] Regarding the data training module, such as Figure 1 As shown, it includes:

[0175] A base model component, the base model component including two or more base training models, each base training model is trained on the processed training set to obtain two or more base prediction models;

[0176] An integrated model component integrates two or more base prediction models into a final prediction model, and the final prediction model can obtain a final prediction result based on the prediction results of the two or more base prediction models.

[0177] In this embodiment, the training set is trained separately through multiple base training models to obtain different base prediction models, so that the relationship between features and labels can be established from different dimensions. Afterwards, the base prediction models are integrated through the integrated model component. Therefore, when making predictions, predictions are actually made jointly through different dimensions to increase the accuracy of the prediction results.

[0178] The base training models are selected from two or more models selected from a linear regression model, a gradient boosting regression model, a support vector regression model, and a decision tree regression model. The final prediction model can obtain a final prediction result based on the prediction results of the two or more base prediction models: the two or more base prediction models independently predict to obtain prediction results respectively, and the majority voting result or the average prediction result of the prediction results of the base prediction models is used as the final prediction result.

[0179] In this embodiment, the above four base training models are used simultaneously for training. Among them, the linear regression model attempts to learn the linear relationship between the data, the gradient boosting regression model captures the nonlinear relationship in the data, the support vector regression model emphasizes the model's prediction accuracy for certain specific areas, and the decision tree regression model adjusts the model's attention to extreme values ​​or important areas. The prediction models trained by these base training models participate in the prediction process together. When making predictions, each model will independently give its own prediction results, and the integrated model component calculates the final prediction output based on these results. This makes it possible to achieve very good prediction results for a single VOC mass concentration (it can achieve very good prediction results for a single VOC mass concentration (especially benzene compounds)).

[0180] Regarding the model evaluation module, such as Figure 1 As shown, it includes:

[0181] A prediction component capable of using a final prediction model to obtain a predicted value of standard equipment detection data from a processed test set (specifically, test feature normalized data);

[0182] An evaluation component compares the predicted value of the standard equipment detection data with the true value of the standard equipment detection data (test label data) to obtain an evaluation result.

[0183] Specifically, the evaluation component compares the predicted value of the standard equipment detection data with the true value of the standard equipment detection data to obtain an evaluation result, which means that the evaluation component calculates at least one of the mean squared error (MSE), determination coefficient (R² score), root mean squared error (RMSE) and / or mean absolute error (MAE) between the model predicted value and the true value of the test set as the evaluation result.

[0184] This example uses all four evaluation metrics simultaneously to generate results, providing different perspectives on model performance evaluation, helping to understand the model's predictive capabilities. MSE and RMSE indicate the extent to which the predicted value deviates from the true value; the R² score indicates how much data variability is explained by the model; MAE provides a robust measure of the magnitude of the prediction error; and MAPE provides the ratio of the prediction error to the true value. This helps to achieve a better-performing final prediction model, thereby increasing the accuracy of prediction results.

[0185] Exemplary Methods

[0186] Figure 2 Illustration of a model training method for volatile organic compound prediction based on machine learning.

[0187] like Figure 2 As shown, this embodiment provides an online early warning method for volatile organic compounds based on machine learning, which includes:

[0188] a data acquisition step of acquiring detection data of target detection points as training sets and test sets; wherein the detection data includes standard equipment detection data of volatile organic compound concentration, PID sensor detection data of volatile organic compound concentration, and meteorological data;

[0189] A data preprocessing step of preprocessing the detection data to obtain a processed training set and a processed test set;

[0190] a data training step of training the processed training set to obtain a final prediction model for predicting standard equipment detection data based on PID sensor detection data and meteorological data;

[0191] The model evaluation step compares the predicted value of the standard equipment test data obtained by using the final prediction model to predict the test set with the actual value of the standard equipment test data to obtain the evaluation result.

[0192] The collection time and quantity of test data and the meaning of standard equipment have been introduced above and will not be repeated here.

[0193] Meteorological data can be selected from one or more of wind direction, wind speed, temperature, humidity, and air pressure. In this embodiment, the meteorological data selected by the inventors includes wind direction, wind speed, temperature, humidity, and air pressure at the target monitoring point, which can achieve excellent prediction results for individual VOC mass concentrations (especially for benzene-based compounds). Benzene-based compounds are organic compounds containing a benzene ring structure, also known as aromatic compounds. For VOCs, examples include toluene and xylene.

[0194] Through the above-mentioned model training method for volatile organic compound prediction based on machine learning of the present application (hereinafter referred to as "the method of the present application"), the detection data (including training set and test set) of the target monitoring point can be obtained through the data acquisition step for subsequent processing; the detection data (including training set and test set) can be preprocessed (such as standardization, etc.) through the data preprocessing step for subsequent machine learning; through data training, etc., the PID sensor detection data and meteorological data in the processed training set are regressed with the standard equipment detection data to construct a final prediction model that predicts the standard equipment detection data from the PID sensor detection data and meteorological data; model evaluation, etc. can evaluate the final prediction model. Therefore, the qualified final prediction model can be used to predict the predicted value (μg / m 3 ), which is detected by PID, has the advantages of shorter monitoring frequency (minutes or even seconds), small size, low price, and easy to deploy in large quantities. In addition, the predicted value of the standard equipment test data is the mass concentration of each substance (μg / m 3 ), which can be used to determine the specific detection value of each substance for alarm, thereby solving the problems of the above-mentioned prior art.

[0195] In addition, the above-mentioned PID sensors are set at multiple points in the industrial park. The method of the present application can predict the VOCs prediction results at multiple points respectively, so as to facilitate timely and comprehensive understanding of the pollution status and trends in the entire park.

[0196] Regarding data preprocessing steps, such as Figure 2 As shown, it includes:

[0197] a data separation sub-step of dividing the standard equipment detection data of the training set and the standard equipment detection data of the test set into training label data and test label data, respectively, and dividing the PID sensor detection data and the meteorological data of the training set and the PID sensor detection data and the meteorological data of the test set into training feature data and test feature data, respectively;

[0198] A feature standardization sub-step of standardizing the training feature data and the test feature data to obtain training feature standardized data and test feature standardized data respectively;

[0199] In the dimensionality upgrading sub-step, the dimensions of the training feature standardized data and the test feature standardized data are upgraded to obtain training feature upgraded dimension data and test feature upgraded dimension data, respectively, so as to obtain a processed training set consisting of training label data and training feature upgraded dimension data, and a processed test set consisting of test label data and test feature upgraded dimension data, respectively.

[0200] Before subsequent training or testing of the test data used as training and test sets, they must first be processed to achieve better training and testing results. Specifically, in this solution, the training and test sets are first divided. Through a data separation step, the training set is divided into training label data (corresponding to standard equipment test data) and training feature data (corresponding to PID sensor test data and meteorological data), and the test set is divided into test label data (corresponding to standard equipment test data) and test feature data (corresponding to PID sensor test data and meteorological data). This facilitates the construction of a model that predicts standard equipment test data from PID sensor test data and meteorological data. In addition, to achieve better training results, the training feature data and test feature data are further standardized through a feature standardization step to accelerate the convergence of many machine learning algorithms and improve model performance. Furthermore, the standardized data (training feature standardized data and test feature standardized data) needs to be further dimensionally upgraded, projecting sample points to a higher dimension to separate mixed sample points, thereby preventing underfitting and improving model accuracy.

[0201] In this embodiment, standardizing the training feature data and the test feature data specifically refers to: making the training feature data and the test feature data have zero mean and unit variance, specifically:

[0202] For each feature j, first calculate the mean µ of all samples j and standard deviation σ j :

[0203]

[0204]

[0205] in:

[0206] m is the number of samples;

[0207] x ij is the original value of the i-th sample on the j-th feature;

[0208] µ j is the mean of the jth feature;

[0209] σ j is the standard deviation of the jth feature.

[0210] Use the mean and standard deviation of the features to transform the original data to standardize it:

[0211]

[0212] Where: z ij is the normalized value of the i-th sample on the j-th feature.

[0213] In addition, respectively increasing the dimension of the training feature standardized data and the test feature standardized data specifically refers to: using a polynomial kernel function to map the training feature standardized data and the test feature standardized data to a high-dimensional space.

[0214] Wherein, the polynomial kernel function is:

[0215]

[0216] in,

[0217] x i is the feature vector representing the i-th sample;

[0218] x j is the feature vector representing the jth sample;

[0219] <x i ,x j > represents a vector x i and x j The inner product of

[0220] d is the degree of the kernel function.

[0221] A polynomial kernel function with a lower degree d may not be able to adequately capture complex relationships in the data, resulting in insufficient model fit. A polynomial kernel function with a higher degree d provides stronger fit and can fit more complex data patterns, but generally results in a more complex model and requires careful adjustment to avoid overfitting. In this example, d is 3.

[0222] Regarding the data training steps, such as Figure 2 As shown, it includes:

[0223] In a base model sub-step, the processed training set is trained respectively by two or more base training models to obtain two or more base prediction models;

[0224] In the integrated model sub-step, the final prediction model can obtain the final prediction result based on the prediction results of more than two base prediction models.

[0225] In this embodiment, the training set is trained separately through multiple base training models to obtain different base prediction models, so that the relationship between features and labels can be established from different dimensions. Afterwards, the base prediction models are integrated through the integrated model component. Therefore, when making predictions, predictions are actually made jointly through different dimensions to increase the accuracy of the prediction results.

[0226] The base training models are selected from two or more models selected from a linear regression model, a gradient boosting regression model, a support vector regression model, and a decision tree regression model. The final prediction model can obtain a final prediction result based on the prediction results of the two or more base prediction models: the two or more base prediction models independently predict to obtain prediction results respectively, and the majority voting result or the average prediction result of the prediction results of the base prediction models is used as the final prediction result.

[0227] In this embodiment, the above four base training models are used simultaneously for training. Among them, the linear regression model attempts to learn the linear relationship between the data, the gradient boosting regression model captures the nonlinear relationship in the data, the support vector regression model emphasizes the model's prediction accuracy for certain specific areas, and the decision tree regression model adjusts the model's attention to extreme values ​​or important areas. The prediction models trained by these base training models participate in the prediction process together. When making predictions, each model will independently give its own prediction results, and the final prediction output is calculated based on these results. This makes it possible to achieve very good prediction results for a single VOC mass concentration (it can achieve very good prediction results for a single VOC mass concentration (especially benzene compounds)).

[0228] Regarding the model evaluation step, such as Figure 2 As shown, including:

[0229] In the prediction sub-step, the final prediction model is used to obtain the predicted value of the standard equipment test data from the processed test set (specifically, the test feature standardized data);

[0230] The evaluation sub-step compares the predicted value of the standard equipment detection data with the true value of the standard equipment detection data (test label data) to obtain an evaluation result.

[0231] Specifically, the evaluation step of comparing the predicted value of the standard equipment detection data with the true value of the standard equipment detection data to obtain an evaluation result means that the evaluation step calculates at least one of the mean squared error (MSE), determination coefficient (R² score), root mean squared error (RMSE) and / or mean absolute error (MAE) between the model predicted value and the true value of the test set as the evaluation result.

[0232] This example uses all four evaluation metrics simultaneously to generate results, providing different perspectives on model performance evaluation, helping to understand the model's predictive capabilities. MSE and RMSE indicate the extent to which the predicted value deviates from the true value; the R² score indicates how much data variability is explained by the model; MAE provides a robust measure of the magnitude of the prediction error; and MAPE provides the ratio of the prediction error to the true value. This helps to achieve a better-performing final prediction model, thereby increasing the accuracy of prediction results.

[0233] Exemplary Device

[0234] Figure 3 The diagram shows an online early warning device for volatile organic compounds based on machine learning.

[0235] like Figure 3 As shown, this embodiment provides a volatile organic compound online early warning device based on machine learning, which includes:

[0236] One or more (e.g., 2, 3, 4, 5, 10, 20, 30, 40, 50 or more) detection modules, each of which includes a PID sensor, and each of which detects the concentration of volatile organic compounds at a corresponding target monitoring point to obtain PID sensor detection data;

[0237] A data acquisition module, wherein the data acquisition module acquires the detection data of the PID sensor and the meteorological data of the target monitoring point as prediction data;

[0238] A prediction module, wherein the prediction module is deployed with a final prediction model, and the prediction module obtains a prediction result of a corresponding target monitoring point from the processed prediction data based on the final prediction model;

[0239] Wherein, the final prediction model is obtained by the above-mentioned model training system or the above-mentioned model training method.

[0240] If the park is small and the weather conditions are simple, a single set of weather monitoring equipment may be used. If the park is large, a weather monitoring device may be installed in each detection module to detect weather data at the target monitoring point. Specifically, the weather data may be one or more selected from wind direction, wind speed, temperature, humidity, and air pressure. Preferably, these three data are simultaneously included.

[0241] Test example

[0242] The inventors of this application adopt the system of the above embodiment (see Figure 1 ) and methods (see Figure 2 ) to obtain the final prediction model, and deploy the final prediction model on the above-mentioned online early warning device (see Figure 3 ), meteorological data includes wind direction, wind speed, temperature, humidity, and air pressure. The applicant analyzed the concentrations monitored by standard equipment (FID detection concentrations) at the time of warnings issued by standard equipment in the industrial park under study. PID sensors deployed in the online warning device acquired detection data and meteorological data, and further predicted the concentrations based on the standard detection values. Detailed statistics are shown in Table 1.

[0243] Among them, the standard equipment is the SynspecGC955 series 615 / 815 toxic and hazardous volatile organic compound analyzer (referred to as "GC-FID equipment") from SYNSPEC of the Netherlands; the PID sensor is the MiniPID 2 from ION Science of the UK.

[0244] Table 1: Statistics of standard equipment warning status and online warning device prediction status.

[0245]

[0246] As can be seen in Table 1, the standard equipment issued five alarms for a single VOC in one month. The online warning device of this application was able to generate an alarm when the predicted concentration reached or exceeded the species alarm concentration on the first four occasions. Even when the species alarm concentration was not reached on the fifth occasion, the predicted concentration was extremely close to that of the standard equipment. Based on the predicted concentration, the staff could determine that the concentration of the corresponding single VOC at the target monitoring point of the PID sensor was high, providing a reliable basis for subsequent processing.

[0247] Although the embodiments of the present application are described above, the present application is not limited to the above-mentioned specific embodiments and application fields. The above-mentioned specific embodiments are merely illustrative and instructive, and not restrictive. A person of ordinary skill in the art, guided by this specification and without departing from the scope of protection of the claims of this application, may also devise various forms, all of which fall within the scope of protection claimed in this application.

Claims

1. A model training system based on machine learning for predicting volatile organic compounds in industrial parks, wherein: include: A data acquisition module, wherein the data acquisition module acquires detection data of target monitoring points as training sets and test sets; wherein the detection data includes standard equipment detection data of volatile organic compound concentration, PID sensor detection data of volatile organic compound concentration, and meteorological data; the meteorological data includes wind direction, wind speed, temperature, humidity, and air pressure; the standard equipment is a device capable of detecting the mass concentration of each individual substance in the volatile organic compound; and the standard equipment detection data is the mass concentration of each substance in the volatile organic compound; A data preprocessing module, which preprocesses the detection data to obtain a processed training set and a processed test set; a data training module, wherein the data training module trains the processed training set to obtain a final prediction model for predicting standard equipment detection data based on PID sensor detection data and meteorological data; A model evaluation module, wherein the model evaluation module compares the predicted value of the standard equipment detection data obtained by using the final prediction model to predict the test set with the actual value of the standard equipment detection data to obtain an evaluation result; Among them, the data preprocessing module includes: a data separation component, which divides the standard equipment detection data of the training set and the standard equipment detection data of the test set into training label data and test label data, and divides the PID sensor detection data and the meteorological data of the training set and the PID sensor detection data and the meteorological data of the test set into training feature data and test feature data; a feature standardization component, which standardizes the training feature data and the test feature data to obtain training feature standardized data and test feature standardized data; a dimension upgrading component, which upgrades the training feature standardized data and the test feature standardized data to obtain training feature upgraded dimension data and test feature upgraded dimension data, so as to obtain a processed training set consisting of training label data and training feature upgraded dimension data, and a processed test set consisting of test label data and test feature upgraded dimension data.

2. The model training system according to claim 1, wherein: The training feature data and the test feature data are respectively standardized as follows: The training feature data and the test feature data are made to have zero mean and unit variance.

3. The model training system according to claim 2, wherein: The training feature data and the test feature data are made to have zero mean and unit variance as follows: For each feature j ， First calculate the mean µ of all samples j and standard deviation σ j : ； ； in: m is the number of samples; x ij is the original value of the i-th sample on the j-th feature; µ j is the mean of the jth feature; σ j is the standard deviation of the jth feature; Use the mean and standard deviation of the features to transform the original data to standardize it: ； Where: z ij is the normalized value of the i-th sample on the j-th feature.

4. The model training system according to claim 1, wherein: The training feature standardized data and the test feature standardized data are respectively upgraded to: The training feature normalized data and the test feature normalized data are mapped to a high-dimensional space using a polynomial kernel function.

5. The model training system according to claim 4, wherein: The polynomial kernel function is: ； in, x i is the feature vector representing the i-th sample; x j is the feature vector representing the jth sample; <x i ,x j > represents a vector x i and x j The inner product of d is the degree of the kernel function.

6. The model training system according to claim 1, wherein: The data training module includes: A base model component, the base model component including two or more base training models, each base training model is trained on the processed training set to obtain two or more base prediction models; An integrated model component integrates two or more base prediction models into a final prediction model, and the final prediction model can obtain a final prediction result based on the prediction results of the two or more base prediction models.

7. The model training system according to claim 6, wherein: The base training model is selected from two or more models of a linear regression model, a gradient boosting regression model, a support vector regression model, and a decision tree regression model.

8. The model training system according to claim 6, wherein: The final prediction model can obtain the final prediction result according to the prediction results of two or more base prediction models: Two or more base prediction models independently predict to obtain prediction results respectively, and the majority voting result or the average prediction result of the prediction results of each base prediction model is used as the final prediction result.

9. The model training system according to claim 1, wherein: The model evaluation module includes: A prediction component capable of using a final prediction model to obtain a predicted value of standard equipment test data from the processed test set; An evaluation component compares the predicted value of the standard equipment detection data with the actual value of the standard equipment detection data to obtain an evaluation result.

10. The model training system according to claim 9, wherein: The evaluation component compares the predicted value of the standard equipment detection data with the actual value of the standard equipment detection data to obtain an evaluation result: The evaluation component calculates the mean square error, determination coefficient, root mean square error and / or mean absolute error between the predicted value of the standard equipment detection data and the true value of the standard equipment detection data as an evaluation result.

11. A model training method based on machine learning for predicting volatile organic compounds in industrial parks, wherein: include: a data acquisition step, obtaining detection data of target monitoring points and dividing the data into a training set and a test set; wherein the detection data includes standard equipment detection data of volatile organic compound concentration, PID sensor detection data of volatile organic compound concentration, and meteorological data; the meteorological data includes wind direction, wind speed, temperature, humidity, and air pressure; the standard equipment is a device capable of detecting the mass concentration of each individual substance in the volatile organic compound; and the standard equipment detection data is the mass concentration of each substance in the volatile organic compound; A data preprocessing step of preprocessing the detection data to obtain a processed training set and a processed test set; a data training step of training the processed training set to obtain a final prediction model for predicting standard equipment detection data based on PID sensor detection data and meteorological data; The model evaluation step compares the predicted value of the standard equipment test data obtained by using the final prediction model to predict the test set with the actual value of the standard equipment test data to obtain an evaluation result; Among them, the data preprocessing step includes: a data separation sub-step, dividing the standard equipment detection data of the training set and the standard equipment detection data of the test set into training label data and test label data, respectively, and dividing the PID sensor detection data and the meteorological data of the training set and the PID sensor detection data and the meteorological data of the test set into training feature data and test feature data, respectively; a feature standardization sub-step, standardizing the training feature data and the test feature data to obtain training feature standardized data and test feature standardized data, respectively; a dimension upgrading sub-step, upgrading the dimensions of the training feature standardized data and the test feature standardized data to obtain training feature upgraded dimension data and test feature upgraded dimension data, respectively, so as to obtain a processed training set consisting of training label data and training feature upgraded dimension data, and a processed test set consisting of test label data and test feature upgraded dimension data, respectively.

12. The model training method according to claim 11, wherein: The training feature data and the test feature data are respectively standardized as follows: The training feature data and the test feature data are made to have zero mean and unit variance.

13. The model training method according to claim 12, wherein: The training feature data and the test feature data are made to have zero mean and unit variance as follows: For each feature j ， First calculate the mean µ of all samples j and standard deviation σ j : ； ； in: m is the number of samples; x ij is the original value of the i-th sample on the j-th feature; µ j is the mean of the jth feature; σ j is the standard deviation of the jth feature; Use the mean and standard deviation of the features to transform the original data to standardize it: ； Where: z ij is the normalized value of the i-th sample on the j-th feature.

14. The model training method according to claim 11, wherein: The training feature standardized data and the test feature standardized data are respectively upgraded to: The training feature normalized data and the test feature normalized data are mapped to a high-dimensional space using a polynomial kernel function.

15. The model training method according to claim 14, wherein: The polynomial kernel function is: ； in, x i is the feature vector representing the i-th sample; x j is the feature vector representing the jth sample; <x i ,x j > represents a vector x i and x j The inner product of d is the degree of the kernel function.

16. The model training method according to claim 11, wherein: The data training step includes: In a base model sub-step, the processed training set is trained respectively by two or more base training models to obtain two or more base prediction models; In the integrated model sub-step, the final prediction model can obtain the final prediction result based on the prediction results of more than two base prediction models.

17. The model training method according to claim 16, wherein: The base training model is selected from two or more models of a linear regression model, a gradient boosting regression model, a support vector regression model, and a decision tree regression model.

18. The model training method according to claim 16, wherein: The final prediction model can obtain the final prediction result according to the prediction results of two or more base prediction models: Two or more base prediction models independently predict to obtain prediction results respectively, and the majority voting result or the average prediction result of the prediction results of each base prediction model is used as the final prediction result.

19. The model training method according to claim 11, wherein: The model evaluation step includes: In the prediction sub-step, the final prediction model is used to obtain the predicted value of the standard equipment test data from the processed test set; The evaluation sub-step compares the predicted value of the standard equipment detection data with the actual value of the standard equipment detection data to obtain an evaluation result.

20. The model training method according to claim 19, wherein: Comparing the predicted value of the standard equipment test data with the actual value of the standard equipment test data to obtain the evaluation result is: The mean square error, determination coefficient, mean absolute error and / or mean absolute percentage error between the predicted value of the standard equipment test data and the true value of the standard equipment test data are calculated as the evaluation result.

21. A machine learning-based online early warning device for volatile organic compounds in industrial parks, wherein: include: More than one detection module, each of which includes a PID sensor, and each detection module detects the concentration of volatile organic compounds at a corresponding target monitoring point to obtain PID sensor detection data; A data acquisition module, wherein the data acquisition module acquires the detection data of the PID sensor and the meteorological data of the target monitoring point as prediction data; A prediction module, wherein the prediction module is deployed with a final prediction model, and the prediction module obtains a prediction result of a corresponding target monitoring point from the prediction data based on the final prediction model; Wherein, the final prediction model is obtained by the model training system according to any one of claims 1 to 10 or the model training method according to any one of claims 11 to 20.

22. The online early warning device for volatile organic compounds in an industrial park based on machine learning as claimed in claim 21, wherein: The detection module also includes a meteorological detection device, which detects meteorological data of the target monitoring point; The meteorological data includes wind direction, wind speed, temperature, humidity, and air pressure.

Images