Cement plant carbon emission monitoring data anomaly detection method and device

Abstract

The invention discloses a cement plant carbon emission monitoring data anomaly detection method and device, and relates to the technical field of carbon emission monitoring, and the method comprises the following steps: S1, collecting CO2 monitoring value, fire coal amount and clinker yield data of a CEMS at regular intervals; s2, constructing a CO2 prediction model, and inputting fire coal amount and clinker yield data into the CO2 prediction model to obtain a CO2 prediction value; s3, an Anomaly-Transform network of a sparse attention mechanism is used to construct an anomaly detection model; s4, the CEMS missing value at each time point in the step S1 is judged, an anomaly detection model is constructed by using an Angle-Transform network of a sparse attention mechanism to detect the difference between CO2 values obtained by the two methods, the situation that abnormal value recognition is insensitive due to tendency errors of single equipment data is avoided, and the detection accuracy is improved. The method considers the association weight in the time dimension of the time sequence, achieves the anomaly detection based on association difference, and greatly improves the calculation speed when the data volume is large through a sparse attention mechanism.

Description

Technical Field

[0001] This invention relates to the field of carbon emission monitoring technology, and in particular to a method and apparatus for detecting anomalies in carbon emission monitoring data of cement plants. Background Technology

[0002] The cement industry, as a high-energy-consuming and high-emission sector, plays a crucial role in global carbon emission control. Accurate monitoring of CO2 emissions from cement plants is essential for developing effective emission reduction strategies and achieving carbon neutrality goals. Currently, there are two main methods for monitoring and calculating CO2 emissions from cement plants. The first method involves using formulas to calculate CO2 emissions from the combustion of fossil fuels and the decomposition of carbonates in cement clinker production. However, this indirect method is susceptible to the influence of fossil fuel and material quality, making it difficult to guarantee accuracy. The second method uses a CEMS (Continuous Emission Monitoring System) to monitor CO2 emissions from cement plants. However, this method is also affected by environmental and equipment conditions, leading to data anomalies, missing data, insufficient sample representativeness, instrument contamination, and data bias. Therefore, it is necessary to process the CO2 data monitored by the CEMS to fill in missing values ​​and correct outliers.

[0003] In terms of outlier identification, common methods include using thresholding, statistical or machine learning methods to detect outliers in CO2 over a period of time and treat these outliers as outliers. However, these methods have significant shortcomings: trend changes in CO2 monitoring values ​​may lead to insensitive outlier identification, making it impossible to detect anomalies in the data in a timely manner; at the same time, these methods do not make full use of other relevant data information, which limits the accuracy and reliability of outlier identification.

[0004] There are two common methods for imputing missing values: one is to analyze CO2 values ​​alone and use statistical or machine learning methods to fill in the missing CO2 values ​​at the time points; the other is to predict the missing CO2 values ​​at the time points using linear or nonlinear models based on other production parameters. However, both methods also have many problems: predicting missing values ​​based on only CO2 fails to fully utilize the value of production data, resulting in low imputation accuracy; while when using other variables to predict missing CO2 values ​​at the time points, the lack of other variables will prevent the imputation work from proceeding normally in the event of regional power outages or system failures, affecting the integrity and continuity of the data.

[0005] In summary, existing methods for monitoring and processing CO2 emissions from cement plants have many problems in terms of accuracy, outlier identification, and missing value imputation, and cannot meet the accuracy and reliability requirements of cement plants for carbon emission monitoring. Therefore, a method and device for detecting anomalies in carbon emission monitoring data from cement plants is proposed. Summary of the Invention

[0006] The purpose of this invention is to solve the problems in the prior art by proposing a method and device for detecting abnormal carbon emission monitoring data in cement plants.

[0007] A method for detecting anomalies in carbon emission monitoring data from a cement plant includes the following steps: S1. Collect CO2 monitoring values, coal consumption and clinker production data from CEMS at regular intervals; S2. Construct a CO2 prediction model by inputting coal consumption and clinker production data into the CO2 prediction model to obtain CO2 prediction values. S3. Construct an anomaly detection model using an Anomaly-Transformer network with a sparse attention mechanism; S4. Determine the missing values ​​of CEMS at each time point in step S1. If there are missing values, fill them with the predicted CO2 values. If there are no missing values, calculate the difference between the predicted CO2 value and the CO2 monitoring value of CEMS at each time point, and input the coal consumption and clinker production as the sequence data for detection into the anomaly detection model. S5. The final result is compared with the preset threshold. If the result exceeds the threshold, the data at that point is considered abnormal. The CO2 predicted value is then used to correct the outlier value of the CO2 monitoring value of the CEMS at that point.

[0008] Preferably, in step S2, the CO2 prediction model includes using coal consumption and clinker production as independent variables to establish an SVR to predict CO2 values. When introducing a kernel function, the formula is as follows:

[0009] in This represents the kernel function, which is selected based on specific needs. These are dual variables, satisfying... .

[0010] Preferably, in step S3, the anomaly detection model includes: (1) The difference between the predicted CO2 value and the monitored CO2 value is denoted as . ; (2) The collected coal consumption, clinker production and splicing as the independent variable X ; (3) To X The encoding process consists of two parts: the first part uses convolution for value encoding, and the second part uses position encoding. Dimensional expansion is performed, and the two parts are then added together to obtain the encoded value. ; (4) Obtained by linear transformation , , The model uses Sigma and refactors it into n-head attention to enable the model to observe and capture different features of the data from different perspectives. (5) Calculation , Sparse dot product attention score; (6) Calculate the sparse attention A; (7) Establish the Gaussian prior probability density function based on sigma; (8) Reconstruct the attention-weighted output A into a single attention, perform a linear transformation, add the residual, and output the result, denoted as output; (9) Loss function; (10) Train the model using normal data.

[0011] Preferably, in step (3), the encoding process is as follows: a. Use d one-dimensional convolutions with kernel size 3 and stride 1 to extract feature information from the model input data. , Feature information Represented as: ; b. Perform position encoding on the model input data to obtain the data's position information. Location information Represented as:

[0012] in, The calculation method is as follows , i The value range is [0, m). j The value range is [0, d), where m is the number of sequences in the model input data, and d represents the dimension of the positional information and is the same as the number of convolutional kernels used. c. Transfer location information and feature information Adding them together yields the complete data features. The dimensions of the data features are (m,n,d), and are represented as follows: + .

[0013] Preferably, in step (5), the calculation 、 The sparse dot product attention score is expressed by the formula:

[0014] in, This indicates the dimension of each head of data.

[0015] Preferably, in step (6), the formula for calculating sparse attention A is:

[0016] in, Includes only sparse metrics The largest u queries in the middle.

[0017] Preferably, in step (7), the Gaussian prior probability density function is established based on sigma, and the formula is expressed as:

[0018] in, x The distance from one feature to other features is represented as: .

[0019] Preferably, in step (9), the loss function includes: a. Calculate the sparse attention A and Gaussian probability density The two-way KL divergence losses loss1 and loss2, the KL divergence formula can be expressed as:

[0020] b. Calculate the mean squared error of the output and input as the reconstruction loss, denoted as recloss, which can be expressed by the following formula:

[0021] c. Calculate the difference between recloss and loss1 and loss2 respectively, and perform backpropagation to update the model parameters.

[0022] Preferably, in step (10), normal data is used to train the model, and when performing detection, the data to be detected is used as input and the model output is used as output. a. Calculate the MSE of the input and output, denoted as loss; b. Calculate the sparse attention A and the Gaussian probability density. The bidirectional KL loss, loss1, loss2; c. Calculate the softmax of -loss1-loss2 and multiply it by the loss. The formula is as follows: .

[0023] A device for detecting abnormal carbon emission monitoring data in a cement plant includes a monitoring module, a data acquisition module, and a processing module. The monitoring module is used to monitor and acquire CO2 monitoring values, and includes a stainless steel flue gas sampling probe, a heat tracing pipeline, a vacuum pump, a Pitot tube flow meter, a thermobaric sensor, a dust meter, and a gas analyzer. The acquisition module obtains data on coal consumption and clinker production, including data from Schenker scales; The processing module is used to receive data acquired by the monitoring module and the acquisition module, and to construct CO2 prediction models and anomaly detection models for calculation. It includes a processor and a memory for storing processor-executable instructions.

[0024] Compared with existing technologies, the advantages of this invention are: 1. This invention uses an anomaly-transformer network with sparse attention mechanism to construct an anomaly detection model to detect the difference in CO2 values ​​obtained by two methods. This avoids the insensitivity of anomaly identification caused by trend errors in data from a single device. Furthermore, this method considers the correlation weights in the time dimension of the time series to achieve anomaly detection based on correlation differences. At the same time, the sparse attention mechanism significantly improves the computation speed when dealing with large amounts of data.

[0025] 2. This invention combines the CO2 prediction value calculated from raw material consumption with the CO2 monitoring value from CEMS, thereby reducing the inaccuracies caused by calculating CO2 from raw material consumption and the data anomalies and missing data caused by equipment factors. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the method for detecting anomalies in carbon emission monitoring data in this invention.

[0027] Figure 2 This is a flowchart illustrating steps S4 and S5 in this invention.

[0028] Figure 3 This is a diagram of the sampling probe in the monitoring module of the present invention.

[0029] Figure 4 This is a diagram of the gas analyzer in the monitoring module of the present invention. Detailed Implementation

[0030] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0031] Reference Figure 1-2 As shown, a method for detecting anomalies in carbon emission monitoring data from a cement plant includes the following steps: S1. Collect CO2 monitoring values, coal consumption and clinker production data from CEMS at regular intervals; S2. Construct a CO2 prediction model by inputting coal consumption and clinker production data into the CO2 prediction model to obtain CO2 prediction values. In cement clinker production, CO2 emission sources include fossil fuel combustion and process emissions (CO2 produced by carbonate decomposition). CO2 prediction models include using coal consumption and clinker production as independent variables to establish a kernel function (SVR) to predict CO2 values. When introducing the kernel function, the formula is as follows:

[0032] in This represents the kernel function, which is selected based on specific needs. These are dual variables, satisfying... .

[0033] S3. An anomaly detection model is constructed using an Anomaly-Transformer network with a sparse attention mechanism. This method considers the correlation weights in the time dimension of the time series to achieve anomaly detection based on correlation differences. At the same time, the sparse attention mechanism greatly improves the computation speed when dealing with large amounts of data. Anomaly detection models include: (1) The difference between the predicted CO2 value and the monitored CO2 value is denoted as . ; (2) The collected coal consumption, clinker production and splicing as the independent variable X ; (3) To X The encoding process consists of two parts: the first part uses convolution for value encoding, and the second part uses position encoding. Dimensional expansion is performed, and the two parts are then added together to obtain the encoded value. The encoding process is as follows: a. Use d one-dimensional convolutions with kernel size 3 and stride 1 to extract feature information from the model input data. Feature information Represented as: ; b. Perform position encoding on the model input data to obtain the data's position information. Location information Represented as:

[0034] in, The calculation method is as follows , i The value range is [0, m). j The value range is [0, d), where m is the number of sequences in the model input data, and d represents the dimension of the positional information and is the same as the number of convolutional kernels used. c. Transfer location information and feature information Adding them together yields the complete data features. The dimensions of the data features are (m,n,d), and are represented as follows: + .

[0035] (4) Obtained by linear transformation , , The model uses Sigma and refactors it into n-head attention to enable the model to observe and capture different features of the data from different perspectives. (5) Calculation , The sparse dot product attention fraction formula is expressed as:

[0036] in, This indicates the dimension of each head of data.

[0037] (6) Calculate the sparse attention A, the formula is:

[0038] in, Includes only sparse metrics The largest u queries in the middle.

[0039] (7) Based on sigma, establish the Gaussian prior probability density function, which is expressed as:

[0040] in, x The distance from one feature to other features is represented as: .

[0041] (8) Reconstruct the attention-weighted output A into a single attention, perform a linear transformation, add the residual, and output the result, denoted as output; (9) Loss function, including: a. Calculate the sparse attention A and Gaussian probability density The two-way KL divergence losses loss1 and loss2, the KL divergence formula can be expressed as:

[0042] b. Calculate the mean squared error of the output and input as the reconstruction loss, denoted as recloss, which can be expressed by the following formula:

[0043] c. Calculate the difference between recloss and loss1 and loss2 respectively, and perform backpropagation to update the model parameters.

[0044] (10) Use normal data to train the model. When performing detection, use the data to be detected as input and the model output as output. a. Calculate the MSE of the input and output, denoted as loss; b. Calculate the sparse attention A and the Gaussian probability density. The bidirectional KL loss, loss1, loss2; c. Calculate the softmax of -loss1-loss2 and multiply it by the loss. The formula is as follows: .

[0045] S4. Determine the missing values ​​of CEMS at each time point in step S1. If there are missing values, fill them with the predicted CO2 values. If there are no missing values, calculate the difference between the predicted CO2 value and the CO2 monitoring value of CEMS at each time point, and input the coal consumption and clinker production as the sequence data for detection into the anomaly detection model. S5. The final result is compared with the preset threshold. If the result exceeds the threshold, the data at that point is considered abnormal. The CO2 predicted value is then used to correct the outlier value of the CO2 monitoring value of the CEMS at that point.

[0046] A device for detecting abnormal carbon emission monitoring data in a cement plant includes a monitoring module, a data acquisition module, and a processing module. The monitoring module is used to monitor and acquire CO2 monitoring values, and includes a stainless steel flue gas sampling probe, a heat tracing pipeline, a vacuum pump, a Pitot tube flow meter, a thermobaric sensor, a dust meter, and a gas analyzer. A sampling probe is installed at the vertical section of the exhaust pipe, with the probe tail connected to a sampling pipeline. The pipeline is inclined downwards at ≥5° to prevent water accumulation. The heating pipeline is heated to a temperature ≥10° above the flue gas dew point to prevent condensation. An air pump provides negative pressure for the sampling circuit. A Pitot tube flow meter measures and controls the flue gas sampling flow rate and monitors the gas emission from the exhaust pipe. A temperature and pressure sensor monitors the gas temperature and pressure to correct for flue gas concentration. The flue gas obtained through the sampling probe is sent to a gas analyzer to measure the concentration of gaseous pollutants in the flue gas. The acquisition module obtains data on coal consumption and clinker production, including data from Schenker scales; The processing module is used to receive data acquired by the monitoring module and the acquisition module, and to construct CO2 prediction models and anomaly detection models for calculation. It includes a processor and a memory for storing processor-executable instructions.

[0047] Example During the data acquisition phase, CO2 monitoring values, coal consumption, and clinker production data of CEMS are obtained every 15 minutes through corresponding monitoring and metering equipment. Input the coal consumption and clinker production data at each time point into the CO2 prediction model to obtain the CO2 prediction value at the corresponding time point; The collected data was input into the anomaly detection model to obtain the results at each time point. Taking a threshold of 0.1 as an example, the results considered as outliers are shown in Table 1 below: Table 1: Examples of Anomalies in Carbon Emission Monitoring Data

[0048] The data table shows that at a certain point in time, the coal consumption was 0.010135042 t / s, the clinker production was 0.075560592 t / s, and the CO2 monitoring value was 135.4414757 m³. 3 / s, CO2 predicted value is 134.4920075m 3 / s, the difference is 0.949468156m 3 / s, the model result is 0.6919389. Since this result is greater than the threshold of 0.1, the data at this point is determined to be abnormal. Therefore, the CO2 prediction value at this time point needs to be used to correct the outlier value of the CO2 monitoring value of the CEMS at this point.

[0049] As is known from common technical knowledge, this invention can be implemented through other embodiments that do not depart from its spirit or essential characteristics. Therefore, the disclosed embodiments described above are merely illustrative in all respects and are not the only ones. All modifications within the scope of this invention or its equivalents are included in this invention.

Claims

1. A method for detecting anomalies in carbon emission monitoring data of a cement plant, characterized in that: The method comprises the following steps: S1, collecting CO2 monitoring values of CEMS, coal consumption and clinker output data every certain time; S2, constructing a CO2 prediction model, inputting the coal consumption and clinker output data into the CO2 prediction model to obtain CO2 prediction values; S3, constructing an anomaly detection model using an Anomaly-Transformer network with a sparse attention mechanism; S4, judging the missing values of CEMS at each time point in step S1, if there are missing values, filling the missing values using the CO2 prediction values, if there are no missing values, calculating the difference between the CO2 prediction values and the CO2 monitoring values of CEMS at each time point, and taking the coal consumption and clinker output as sequence data for inputting into the anomaly detection model; S5, comparing the final result with a pre-set threshold value, if the result exceeds the threshold value, considering that the data at this point is abnormal, and using the CO2 prediction value to correct the abnormal value of the CO2 monitoring value of CEMS at this point. 2.The method for monitoring data anomaly of carbon emission of a cement plant according to claim 1, characterized in that: In the step S2, the CO2 prediction model comprises using the coal consumption and clinker output as independent variables to establish an SVR to predict the CO2 value, and when a kernel function is introduced, the formula is as follows: wherein represents a kernel function, selected according to specific needs, is a dual variable, satisfying .

3. The method for monitoring data anomaly of carbon emission of a cement plant according to claim 1, characterized in that: In the step S3, the anomaly detection model comprises: (1) using the difference between the predicted value of CO2 and the monitored value of CO2, denoted as ; (2) the collected amount of coal, the amount of clinker production and splicing as an independent variable X ; (3) To encode the X where the encoding is divided into two parts, the first part encodes the value using convolution, and the second part encodes the position, expands the dimension, and adds the two to obtain the encoded value ; (4) to perform linear changes to obtain , , , sigma, and reconstruct into n-head attention, so that the model can observe and capture different features of the data from different angles; (5) compute , sparse dot-product attention scores; (6) calculating sparse attention A; (7) establishing a Gaussian prior probability density function according to sigma; (8) reconstructing the attention weighted output A into a head attention, and performing linear transformation, outputting after increasing the residual, denoted as output; (9) loss function; (10) training the model using normal data.

4. The method for monitoring data anomaly of carbon emission of a cement plant according to claim 3, characterized in that: In the step (3), the encoding process is as follows: a. using d one-dimensional convolutions with a kernel size of 3 and a stride of 1 to extract feature information from the model input data , feature information is represented as: ; b. positionally encoding the model input data to obtain positional information of the data positional information is expressed as: wherein, The calculation method is , i The value range is [0, m), j The value range is [0, d), m is the number of sequences in the model input data, and d represents the dimension of the position information and is the same as the number of convolution kernels used. c. adding the location information and the feature information to obtain the complete data feature , the dimension of the data feature is (m, n, d), and is expressed as: + .

5. The method for monitoring data anomaly of carbon emission of a cement plant according to claim 3, characterized in that: In the step (5), the sparse dot-product attention score of the formula is calculated as follows: , ​ wherein, represents the dimension of each head of data.

6. The method for monitoring data anomaly of carbon emission of a cement plant according to claim 3, characterized in that: In the step (6), the formula for calculating sparse attention A is as follows: wherein, only containing sparsity measures the u largest queries in the set.

7. The method for monitoring data anomaly of carbon emission of a cement plant according to claim 3, characterized in that: In the step (7), the Gaussian prior probability density function is established according to sigma, and the formula is as follows: wherein, x represents the distance of one feature to other features, represented as: 。 8.The method of claim 1, wherein: In the step (9), the loss function comprises: a、compute the bidirectional KL divergence loss loss1, loss2 of sparse attention A and Gaussian probability density The formula of KL divergence can be expressed as: b, calculating the mean square error of output and input as the reconstruction loss, denoted as recloss, and the formula is as follows: c, calculating the difference between recloss and loss1 and loss2 respectively, and performing back propagation to update the model parameters.

9. The method for monitoring data anomaly of carbon emission of a cement plant according to claim 1, characterized in that: In the step (10), the model is trained using normal data, and when detecting, the data to be detected is taken as input, and the model output is taken as output; a, calculating the MSE of input and output, denoted as loss; b. computing a bidirectional KL loss loss1, loss2 of the sparse attention A and the Gaussian probability density b. computing a bidirectional KL loss loss1, loss2 of the sparse attention A and the Gaussian probability density c, calculating the softmax of -loss1-loss2, and multiplying it by loss, and the formula is as follows: 。 10.A device for detecting abnormality of carbon emission monitoring data of a cement plant, characterized in that: The method comprises a monitoring module, a collecting module and a processing module; The monitoring module is used for monitoring and acquiring CO2 monitoring values, and comprises a stainless steel flue gas sampling probe, a heat tracing pipeline, a gas suction pump, a Pitot tube flowmeter, a thermal pressure sensor, a dust measuring instrument and a gas analyzer; The collecting module acquires coal consumption and clinker output data, and comprises a Schenke scale; The processing module is used for receiving the data acquired by the monitoring module and the collecting module, and constructing a CO2 prediction model and an anomaly detection model for calculation, and comprises a processor and a memory for storing processor executable instructions.

Images