A high-precision sampling control method for industrial thermal process identification and new pollutant emissions

Abstract

The present invention discloses a method for high-precision sampling and control of industrial thermal process identification and new pollutant emissions, including obtaining factory furnace data through a technical interface, establishing a deep autoencoding neural network model, the encoder-decoder is a 2-layer neural network, one layer is a decoder input layer, and one layer is a decoder output layer, a linear activation function must be used, and the above data is used for training, the output of the decoder is obtained, and the output is unsupervised clustered, the cluster center is saved, the autoencoding neural network model decoder is changed to the cluster center, and a system clustering identification model is implemented, the factory real-time data is obtained through a technical interface, and sent to the above identification model, and the identification result is obtained or a control instruction is sent according to the real-time result and the sampling target state. The present invention can be applied to a terminal, a server, or software running in a terminal or a server, and the terminal can be a tablet computer, a laptop computer, a desktop computer, etc.

Description

Technical Field

[0001] The present invention relates to the technical field of environmental monitoring, and in particular to a method for high-precision sampling and control of industrial thermal process identification and new pollutant emissions. Background Art

[0002] The operation trajectory of industrial thermal processes presents high-frequency dynamic characteristics that change over time and are complex. New pollutants (dioxins, PFOS, PFOA, etc.) are gradually gaining attention. In industrial thermal processes, the generation process of new pollutants is more complex than SO 2 , NOx, PM 2.5 Conventional pollutants such as carbon dioxide are complex and cannot be monitored online. They are generated in many types and are trace organic matter, so the sampling time required to obtain their emission data is long.

[0003] At present, the sampling methods for new flue gas pollutants generated by industrial thermal processes mainly include: "Technical Specifications for Monitoring Dioxin Emissions from Hazardous Waste (Including Medical Waste) Incineration Disposal Facilities" (HJ / T 365-2007), "Isotope Dilution High Resolution Gas Chromatography-High Resolution Mass Spectrometry for the Determination of Dioxins in Ambient Air and Waste Gas" (HJ 77.2-2008), which usually require continuous sampling for more than 2 hours. For industrial thermal processes, this sampling time cannot meet the requirements for the fine characterization of new pollutant emissions under high-frequency dynamic characteristics, resulting in the inability to fully confidently determine their emissions in the field of control and emission supervision. Industrial thermal processes are complex nonlinear systems, and state identification faces multiple challenges. These challenges include multivariate interactions, nonlinear dynamic behavior, and the impact of environmental disturbances, which cannot be directly achieved through artificial senses.

[0004] In view of the above problems, the present invention provides an industrial thermal process identification and new pollutant emission high-precision sampling control method for realizing high-precision real-time identification and sampling control of system status. Summary of the invention

[0005] The present invention provides an industrial thermal process identification and new pollutant emission high-precision sampling control method for realizing high-precision real-time identification and sampling control of system status.

[0006] The purpose of the present invention is to provide a method for industrial thermal process identification and new pollutant emission high-precision sampling control, the method comprising the following steps:

[0007] Step 1: Obtain all temperature and air volume data of the factory furnace and conventional pollutant data at the chimney through the technical interface;

[0008] Step 2: After cleaning the collected data, construct the input and output. The input of the model is all the collected data, i.e., all the temperature and air volume data collected in step 1, and the conventional pollutant data at the chimney. The output is the temperature and pollutant data, i.e., the temperature and conventional pollutant data at the chimney collected in step 1.

[0009] Step 3: Establish a deep autoencoder neural network model. The decoder is required to be a 2-layer neural network, one layer is the decoder input layer, and one layer is the decoder output layer. A linear activation function must be used.

[0010] Step 4: Divide the data into training set and test set in a ratio of 8:2, train the model using MAE or MSE as the loss function, and save the normalization parameters of the training set;

[0011] Step 5. After the model is trained, the output MID-X of the encoder output layer is obtained, and the cosine similarity is directly used as the distance measurement method to perform unsupervised clustering on the output. The classification cluster uses the silhouette coefficient and the elbow method to obtain the optimal number of clusters k, and the cluster center of the unsupervised clustering is saved. The last linear layer of the deep neural network is changed into the cluster center to realize the system state identification method model;

[0012] Silhouette Coefficient The calculation formula is:

[0013]

[0014] in, Represents the average distance between sample point i and other points in its cluster. Represents the minimum average distance between sample point i and other clusters;

[0015] The calculation formula of the elbow method is:

[0016]

[0017] in, represents the kth cluster of the classification, yes The sample points in yes The centroid of is the average degree of distortion;

[0018] The specific method of changing the deep neural network decoding layer into a cluster center to construct a system state identification method model is as follows: the sample to be identified is encoded through the encoding layer, and the cosine similarity with each cluster center is calculated, and the category with the minimum value is taken as the output category, which is the real-time system state;

[0019] Step 6: Obtain real-time data from the factory through the technical interface, use step 2 to build the model input and output, and input the system state identification method model in step 5 to obtain the real-time system state;

[0020] Step 7: According to the sampling target and the real-time status of the system, a command is sent to the sampling device through a command sending device to realize the designated state sampling.

[0021] Furthermore, the temperature data in step one includes at least the furnace temperature, the flue gas temperature near the flue gas treatment facility, and the flue gas temperature at the chimney outlet; the air volume data includes at least the primary air volume and secondary air volume in the furnace, and the chimney air volume; and the conventional pollutants at the chimney include at least particulate matter, nitrogen oxides, sulfur dioxide, and carbon monoxide.

[0022] Furthermore, in the deep autoencoder neural network model in step 3, x i ∈X, i=1,2,3…n is the input data set X, y in S2 i ∈Y, i=1,2,3…m is the output data set Y in S2, hidden layer ×l, l=1,2,3…, adjusted according to the actual training situation.

[0023] Furthermore, the formula of the MSE loss function in step 4 is:

[0024]

[0025] Where n represents the number of samples, represents the true value, is the predicted value.

[0026] Furthermore, the formula for obtaining the standardized parameters in the training set is:

[0027]

[0028] Among them, Z is the standardized result, is the mean of the input samples in the training set, The standard deviation of the input samples for the training set, and The standardized parameters are:

[0029]

[0030]

[0031] in, represents the input sample, i=1,2,3…n is the number of training set samples.

[0032] Furthermore, the cosine similarity calculation formula in step 5 is:

[0033]

[0034] in, represents the output vector A∈MID-X of the encoder output layer, Represents the output vector B∈MID-X of the encoder output layer, where i=1,2,3…n is the number of samples in the training set.

[0035] The present invention has the following advantages: the present invention can be applied to a terminal, a server, or software running in a terminal or a server. The terminal can be a tablet computer, a laptop computer, a desktop computer, etc., but is not limited thereto. The server can be an independent physical server, or a server cluster or a distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a flowchart of the present invention;

[0037] Figure 2 is a flowchart of the controller of the present invention;

[0038] Figure 3 It is the loss curve diagram of the training set and the test set of the present invention;

[0039] Figure 4 Determine the optimal number of clusters for the present invention;

[0040] Figure 5 This is a model structure diagram of the deep autoencoding neural network of the present invention. DETAILED DESCRIPTION

[0041] The present invention provides a method for industrial thermal process identification and new pollutant emission high-precision sampling control, the method comprising the following steps:

[0042] Step 1: Obtain all temperature and air volume data of the factory furnace and conventional pollutant data at the chimney through the technical interface;

[0043] Step 2: After cleaning the collected data, construct the input and output. The input of the model is all the collected data, i.e., all the temperature and air volume data collected in step 1, and the conventional pollutant data at the chimney. The output is the temperature and pollutant data, i.e., the temperature and conventional pollutant data at the chimney collected in step 1.

[0044] Step 3: Establish a deep autoencoder neural network model. The decoder is required to be a 2-layer neural network, one layer is the decoder input layer, and one layer is the decoder output layer. A linear activation function must be used.

[0045] Step 4: Divide the data into training set and test set in a ratio of 8:2, train the model using MAE or MSE as the loss function, and save the normalization parameters of the training set;

[0046] Step 5. After the model is trained, the output MID-X of the encoder output layer is obtained, and the cosine similarity is directly used as the distance measurement method to perform unsupervised clustering on the output. The classification cluster uses the silhouette coefficient and the elbow method to obtain the optimal number of clusters k, and the cluster center of the unsupervised clustering is saved. The last linear layer of the deep neural network is changed into the cluster center to realize the system state identification method model;

[0047] Silhouette coefficient The calculation formula is:

[0048]

[0049] in, Represents the average distance between sample point i and other points in its cluster. Represents the minimum average distance between sample point i and other clusters;

[0050] The calculation formula of the elbow method is:

[0051]

[0052] in, represents the kth cluster of the classification, yes The sample points in yes The centroid of is the average degree of distortion;

[0053] The specific method of changing the deep neural network decoding layer into a cluster center to construct a system state identification method model is as follows: the sample to be identified is encoded through the encoding layer, and the cosine similarity with each cluster center is calculated, and the category with the minimum value is taken as the output category, which is the real-time system state;

[0054] Step 6: Obtain real-time data from the factory through the technical interface, use step 2 to build the model input and output, and input the system state identification method model in step 5 to obtain the real-time system state;

[0055] Step 7: According to the sampling target and the real-time status of the system, a command is sent to the sampling device through a command sending device to realize the designated state sampling.

[0056] In this embodiment, the temperature data in step one includes at least the furnace temperature, the flue gas temperature near the flue gas treatment facility, and the flue gas temperature at the chimney outlet; the air volume data includes at least the primary air volume and secondary air volume in the furnace, and the chimney air volume; and the conventional pollutants at the chimney include at least particulate matter, nitrogen oxides, sulfur dioxide, and carbon monoxide.

[0057] In this embodiment, in the deep autoencoder neural network model in step 3, x i ∈X, i=1,2,3…n is the input data set X, y in S2 i ∈Y, i=1,2,3…m is the output data set Y in S2, hidden layer ×l, l=1,2,3…, adjusted according to the actual training situation.

[0058] In this embodiment, the formula of the MSE loss function in step 4 is:

[0059]

[0060] Where n represents the number of samples, represents the true value, is the predicted value.

[0061] In this embodiment, the formula for obtaining the standardized parameters in the training set is:

[0062]

[0063] Among them, Z is the standardized result, is the mean of the input samples in the training set, The standard deviation of the input samples for the training set, and The standardized parameters are:

[0064]

[0065]

[0066] in, represents the input sample, i=1,2,3…n is the number of training set samples.

[0067] In this embodiment, the cosine similarity calculation formula in step 5 is:

[0068]

[0069] in, represents the output vector A∈MID-X of the encoder output layer, Represents the output vector B∈MID-X of the encoder output layer, where i=1,2,3…n is the number of samples in the training set.

[0070] In this embodiment, taking the municipal solid waste incineration power plant as an example, the upper median temperature of the combustion chamber, the oxygen content of the boiler, the flue gas temperature at the economizer outlet, the flue gas temperature at the third superheater of the furnace, the bag outlet temperature (taken from the SCR inlet temperature), the inlet temperature of the furnace bag system, the average temperature of the upper middle section, T10, T11, T12, T20, T21, T22, average temperature, DCS temperature, oxygen content, flue gas temperature, flue gas humidity, particulate matter, sulfur dioxide, nitrogen oxides, carbon monoxide, hydrogen chloride, #1 main steam flow, #1 furnace primary drying air volume, #1 furnace primary gasification air volume, secondary air flow measurement, flue gas flow rate, flue gas pressure, cumulative flow and other furnace performance parameters are obtained.

[0071] In this embodiment, after removing the abnormal values, the upper median temperature of the combustion chamber, the oxygen content of the boiler, the flue gas temperature at the economizer outlet, the flue gas temperature at the third superheater of the furnace, the bag outlet temperature (taken from the SCR inlet temperature), the inlet temperature of the bag system of the furnace, the average temperature of the upper middle section, T10, T11, T12, T20, T21, T22, average temperature, DCS temperature, oxygen content, flue gas temperature, flue gas humidity, particulate matter, sulfur dioxide, nitrogen oxides, carbon monoxide, hydrogen chloride, #1 main steam flow, #1 furnace primary drying air volume, #1 furnace primary gasification air volume, secondary air flow measurement, flue gas flow rate, flue gas pressure, and cumulative flow are used as input, and the average temperature of the upper middle section, oxygen content, particulate matter, sulfur dioxide, nitrogen oxides, carbon monoxide, hydrogen chloride, and flue gas temperature are used as output to obtain a constructed data set.

[0072] In this embodiment, the training set and the test set are divided into 8:2, and the training set is standardized using the z score to obtain the standardized parameters shown in Table 1. The test set is standardized according to the standardized parameters in Table 1, and finally the standardized training set and test set are obtained.

[0073] Table 1

[0074]

[0075] In this embodiment, a residual neural network is constructed, and the last layer is a three-layer linear layer. The output of the third-to-last linear layer is taken as the intermediate output. The neural network is trained using the above training set with MSE as the loss function. After 300 epochs, the trained neural network model and its parameters are obtained.

[0076] In this embodiment, the trained residual neural network model is used to infer the new data set, the output of the third linear layer from the bottom is obtained, and the k-means method is used for unsupervised clustering. The silhouette coefficient or elbow method is used to obtain the optimal number of clusters, which is 14 categories. Figure 2 As shown, save the cluster centers.

[0077] In this embodiment, the standardized parameters, neural network structure and parameters, clustering centers and corresponding programs are burned into the Raspberry Pi 4B model, and the real-time factory data is obtained through the onboard network interface. After identification, the control instructions are implemented through the GPIO (general input and output) interface. The output control instructions are switch quantities, 1 represents the start of sampling, and 0 represents the pause of sampling, as shown in Table 2.

[0078] Table 2

[0079]

[0080] Although the specific implementation of the present invention is described in detail in conjunction with the drawings, it should not be understood as limiting the scope of protection of this patent. Within the scope described in the claims, various modifications and variations that can be made by those skilled in the art without creative work still fall within the scope of protection of this patent.

Claims

1. A method for high-precision sampling and control of industrial thermal process identification and new pollutant emissions, characterized by: The method comprises the following steps: Step 1: Obtain all temperature and air volume data of the factory furnace and conventional pollutant data at the chimney through the technical interface; Step 2: After cleaning the collected data, construct the input and output. The input of the model is all the collected data, i.e., all the temperature and air volume data collected in step 1, and the conventional pollutant data at the chimney. The output is the temperature and pollutant data, i.e., the temperature and conventional pollutant data at the chimney collected in step 1. Step 3: Establish a deep autoencoder neural network model. The decoder is required to be a 2-layer neural network, one layer is the decoder input layer, and one layer is the decoder output layer. A linear activation function must be used. Step 4: Divide the data into training set and test set in a ratio of 8:2, train the model using MAE or MSE as the loss function, and save the normalization parameters of the training set; Step 5. After the model is trained, the output MID-X of the encoder output layer is obtained, and the cosine similarity is directly used as the distance measurement method to perform unsupervised clustering on the output. The classification cluster uses the silhouette coefficient and the elbow method to obtain the optimal number of clusters k, and the cluster center of the unsupervised clustering is saved. The last linear layer of the deep neural network is changed into the cluster center to realize the system state identification method model; The silhouette coefficient The calculation formula is: ； in, Represents the average distance between sample point i and other points in its cluster. Represents the minimum average distance between sample point i and other clusters; The elbow method calculation formula is: ； in, represents the kth cluster of the classification, yes The sample points in yes The centroid of is the average degree of distortion; The specific method of changing the deep neural network decoding layer into a cluster center to construct a system state identification method model is as follows: after the sample to be identified is encoded through the encoding layer, the cosine similarity with each cluster center is calculated, and the category with the minimum value is taken as the output category, and the output category is the real-time system state; Step 6: Obtain real-time data from the factory through the technical interface, use step 2 to build the model input and output, and input the system state identification method model in step 5 to obtain the real-time system state; Step 7: According to the sampling target and the real-time status of the system, a command is sent to the sampling device through a command sending device to realize the designated state sampling.

2. The method for industrial thermal process identification and new pollutant emission high-precision sampling control as claimed in claim 1, characterized in that: The temperature data in step one includes at least the furnace temperature, the flue gas temperature near the flue gas treatment facility, and the chimney flue gas temperature; the air volume data includes at least the primary air volume and secondary air volume in the furnace, and the chimney air volume; the conventional pollutants at the chimney include at least particulate matter, nitrogen oxides, sulfur dioxide, and carbon monoxide.

3. The method for industrial thermal process identification and new pollutant emission high-precision sampling control as claimed in claim 1, characterized in that: In the deep autoencoder neural network model in step 3, x i ∈X, i=1,2,3…n is the input data set X, y in S2 i ∈Y, i=1,2,3…m is the output data set Y in S2, hidden layer ×l, l=1,2,3…, adjusted according to the actual training situation.

4. The method for industrial thermal process identification and new pollutant emission high-precision sampling control as claimed in claim 1, characterized in that: The formula of the MSE loss function in step 4 is: ； Where n represents the number of samples, represents the true value, is the predicted value.

5. The method for industrial thermal process identification and new pollutant emission high-precision sampling control as claimed in claim 1, characterized in that: The formula for obtaining the standardized parameters in the training set is: ； Among them, Z is the standardized result, is the mean of the input samples in the training set, The standard deviation of the input samples for the training set, and The standardized parameters are: ； ； in, represents the input sample, i=1,2,3…n is the number of training set samples.

6. The method for industrial thermal process identification and new pollutant emission high-precision sampling control as claimed in claim 1, characterized in that: The cosine similarity calculation formula in step 5 is: ； in, represents the output vector A∈MID-X of the encoder output layer, Represents the output vector B∈MID-X of the encoder output layer, where i=1,2,3…n is the number of samples in the training set.

Images