Methanation furnace gas content prediction method based on human-in-the-loop deep active learning model

Abstract

The invention discloses a methanation furnace gas content prediction method based on a human-in-the-loop deep active learning model. A methanation furnace is one of key devices in a synthetic ammonia production process, and hydrogen as one of production raw materials is generated by a methane decarburization device of the process. In the operation process of the device, carbon elements in gas exist in the form of carbon monoxide and carbon dioxide, and the methanation furnace is mainly used for converting the carbon monoxide and the carbon dioxide into methane. Therefore, effective measurement of the contents of residual carbon monoxide and carbon dioxide at the outlet of the device is crucial for quality control of the methanation furnace, however, the conventional measurement method is expensive, time-consuming and labor-consuming. Although a data-driven prediction method can provide a real-time measurement result, available modeling samples are very limited in an actual process, and the performance of a prediction model is seriously limited. According to the method, the deep active learning model of the human-in-the-loop is designed, the samples with the large information content are selectively marked manually, the cost of data-driven modeling is greatly reduced, meanwhile, the model prediction performance is improved, and the method has important value for improving the methanation furnace gas content prediction effect.

Description

Technical Field

[0001] The present invention belongs to the field of data-driven modeling and application of complex industrial production processes, and in particular relates to a method for predicting gas content in a methanation furnace based on a human-in-the-loop deep active learning model. Background Art

[0002] In the process of synthetic ammonia production, the methanogen is a very critical device, in which hydrogen, as one of the raw materials, is mainly produced by the methane decarbonization device of the process. In the actual operation of the device, the carbon element contained in the gas mainly exists in the form of carbon monoxide and carbon dioxide, and the main function of the methanogen is to convert the carbon monoxide and carbon dioxide therein into methane and transport it to other devices for subsequent treatment. For the methanogen, effectively measuring the residual carbon monoxide and carbon dioxide content at the outlet of the device is crucial for the quality control of the methanogen, and is also directly related to the operation efficiency of subsequent related devices. However, conventional measurement methods are expensive and time-consuming. The more popular method is based on data-driven modeling, which predicts the residual carbon monoxide and carbon dioxide content at the outlet of the device in real time through the variables that are easier to measure in the device. Although the current data-driven prediction method can provide real-time measurement results, in actual processes, the modeling samples we can use are usually very limited, which seriously limits the performance and application effect of regression or prediction models. Therefore, there is an urgent need for a new technology to improve the modeling performance and actual use effect of data-driven prediction models. Summary of the invention

[0003] The present invention designs a deep active learning model with humans in the loop to selectively manually mark samples with large information content, thereby greatly reducing the cost of data-driven modeling, while improving the performance of model prediction, and achieving improvement in the actual application effect of gas content prediction in methanation furnaces.

[0004] The objective of the present invention is achieved through the following technical solutions:

[0005] A method for predicting gas content in a methanogen based on a human-in-the-loop deep active learning model, characterized in that it comprises the following steps:

[0006] (1) Through various automatic or manual measuring instruments installed in the methanogenization furnace, data samples of the device under normal operation are collected. Among all these measurement variables, based on engineering experience, the measurement variables closely related to the content in the furnace are selected as the input variables of the subsequent prediction model, and the data sample set corresponding to the input variables is obtained as X∈R n×m . Where n is the number of input variable sample set data, and m is the number of selected conventional measurement variables.

[0007] (2) Through manual gas sampling and analysis in the analysis room, a small amount of carbon monoxide and carbon dioxide gas content measurement data is obtained and used as the output variable of the prediction model. The corresponding data sample set is recorded as Among them, n lb Represents the number of labeled sample data.

[0008] (3) Based on labeled data sample set Extract the corresponding data set from the input variable sample set and record it as The initial modeling data sample set is composed of

[0009] (4) Based on the initial labeled sample set A deep partial least squares regression model is established. By extracting key feature information from the data set, a regression relationship between input variables and output variables (i.e., carbon monoxide and carbon dioxide gas content) is established, and the obtained model parameters are stored in the model database for future use.

[0010] (5) Starting from the key features extracted by the deep partial least squares regression model, a feature data description model is constructed. n×m Select unlabeled data samples with large information content, manually mark them, and update the labeled data sample set to prepare for the next model training.

[0011] (6) Based on the updated labeled training data sample set, the deep partial least squares regression model is re-established, and the obtained model parameters are stored in the model database for future use.

[0012] (7) Steps (6) and (7) are continuously repeated to obtain a deep partial least squares regression model with good algorithm convergence and performance, and the final model parameters are stored in the model database for future use.

[0013] (8) Using the final deep partial least squares regression model obtained by the regression method, calculate the new data sample x obtained under the current operating conditions of the methanogen device new The corresponding carbon monoxide and carbon dioxide gas contents are used for performance evaluation and optimization of the device.

[0014] Wherein, the step (4) is specifically as follows: based on the initial labeled sample set A deep partial least squares regression model is established to extract key feature information from the data set and establish the regression relationship between the input variables and the output variables (i.e., carbon monoxide and carbon dioxide gas content). The deep partial least squares regression model is expressed as follows:

[0015]

[0016] Among them, T (l) ,l=1,2,...,L is the key characteristic variable matrix P extracted by the model (l) ,l=1,2,...,L and Q (l) ,l=1,2,...,L are the load matrices corresponding to the input variables and output variables respectively, W (l) ,l=1,2,...,L is the variable weight matrix of the model, E (l) ,l=1,2,...,L and F (l) ,l=1,2,...,L are the residual matrices corresponding to the input variables and output variables respectively, and L is the number of layers of the deep model.

[0017] The step (5) is specifically as follows: starting from the key features extracted by the deep partial least squares regression model, a feature data description model is constructed, and the input variable data set X∈R n×m Select unlabeled data samples with large information content, manually label them, and update the labeled data sample set to prepare for the next model training. Among them, the optimization function of the feature data description model is as follows:

[0018]

[0019] in To enhance the key feature vector, R and a are the radius and center of the hypersphere corresponding to the feature data description model in high-dimensional space, C and ξ i ,i=1,2,...,n lb are adjustment parameters and relaxation coefficients respectively, and the optimized radius and center of the sphere are:

[0020]

[0021] in, and are the nonlinear mapping function and kernel function respectively, α i ,i=1,2,...,n lb For each sample corresponding to the Langrange coefficient, next, evaluate the input variable data set X∈R n×m The distance between each unlabeled sample in the high-dimensional feature space and the center of the hypersphere is expressed as follows:

[0022]

[0023] in, The key features corresponding to each unlabeled sample are mapped in the high-dimensional feature space. The distance indicators corresponding to different unlabeled samples are arranged from large to small. The S samples with the largest distance are selected and sent to manual labeling. The labeled training sample set is updated and recorded as The step (6) is specifically as follows: based on the updated labeled training data sample set, a deep partial least squares regression model is re-established, and the architecture is as follows:

[0024] The updated deep partial least squares regression model is established as follows:

[0025]

[0026] in, Matrix of key feature variables extracted for updating the model and are the load matrices corresponding to the input variables and output variables, To update the variable weight matrix of the model, and are the residual matrices corresponding to the input variables and output variables respectively, and L is the number of layers of the deep model.

[0027] The step (8) is specifically as follows: using the final deep partial least squares regression model obtained by the human in the regression method, the new data sample x obtained under the current operating condition of the methanogen device is calculated new The corresponding carbon monoxide and carbon dioxide gas contents are calculated as follows:

[0028]

[0029] in, For the new data x new The corresponding key feature vectors of each hidden layer of the deep model, is the enhanced feature vector, It is the predicted value of final carbon monoxide and carbon dioxide gas content.

[0030] An electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method for predicting the gas content in a methanogenerator based on a deep active learning model of a human in the loop is implemented.

[0031] A computer-readable storage medium stores computer instructions, which, when executed by a processor, implement the method for predicting gas content in a methanogen based on a deep active learning model of a human in the loop.

[0032] Beneficial effects of the present invention:

[0033] The present invention designs a deep active learning technology of human in the loop. By constructing a data description analysis model based on enhanced features, it selectively evaluates the unlabeled samples corresponding to the process data variables, selects samples with large information content in the high-dimensional feature space, and sends them to technicians for manual labeling, thereby greatly reducing the number and cost of data labeling and effectively improving the efficiency of data-driven modeling. At the same time, due to the selection of data modeling samples with large information content, the prediction performance of the constructed data-driven regression model has also been effectively improved, greatly improving the online prediction effect of the carbon monoxide and carbon dioxide gas content of the methanogen. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 :Methanation furnace production unit flow chart;

[0035] Figure 2 : Results of online prediction of gas content by human-in-the-loop deep active learning model;

[0036] Figure 3 : Results of the basic depth partial least squares regression model for online prediction of gas content. DETAILED DESCRIPTION

[0037] Aiming at the problem of online detection of carbon monoxide and carbon dioxide gas contents in a methanation furnace production device, the present invention uses conventional process variables that are easy to measure in the device, utilizes a deep active learning technology with humans in the loop, constructs a data description and analysis model based on enhanced features, selects samples with large information content for manual labeling, and establishes a high-performance deep partial least squares regression model to achieve online prediction of carbon monoxide and carbon dioxide gas contents.

[0038] Embodiment: The main steps of the technical solution adopted by the present invention are as follows: A method for predicting gas content in a methanogen based on a deep active learning model of a human in the loop, the method is specifically as follows:

[0039] The first step is to collect data samples of the device under normal operation through various automatic or manual measuring instruments installed in the methanogenization furnace. Among all these measurement variables, based on engineering experience, the measurement variables closely related to the content in the furnace are selected as the input variables of the subsequent prediction model, and the data sample set corresponding to the input variables is obtained as X∈R n×m . Where n is the number of input variable sample set data, and m is the number of selected conventional measurement variables.

[0040] In the second step, a small amount of carbon monoxide and carbon dioxide gas content measurement data is obtained through manual gas sampling and analysis in the analysis room, which is used as the output variable of the prediction model. The corresponding data sample set is recorded as Among them, nlb Represents the number of labeled sample data.

[0041] The third step is based on the labeled data sample set Extract the corresponding data set from the input variable sample set and record it as The initial modeling data sample set is composed of

[0042] The fourth step is based on the initial labeled sample set A deep partial least squares regression model is established to extract key feature information from the data set and establish the regression relationship between the input variables and the output variables (i.e., carbon monoxide and carbon dioxide gas content). The deep partial least squares regression model is expressed as follows:

[0043]

[0044] Among them, T (l) ,l=1,2,...,L is the key characteristic variable matrix P extracted by the model (l) ,l=1,2,...,L and Q (l) ,l=1,2,...,L are the load matrices corresponding to the input variables and output variables respectively, W (l) ,l=1,2,...,L is the variable weight matrix of the model, E (l) ,l=1,2,...,L and F (l) ,l=1,2,...,L are the residual matrices corresponding to the input variables and output variables respectively, and L is the number of layers of the deep model.

[0045] The fifth step starts from the key features extracted by the deep partial least squares regression model and constructs a feature data description model. n×m Select unlabeled data samples with large information content, manually mark them, and update the labeled data sample set to prepare for the next model training.

[0046] Among them, the optimization function of the feature data description model is as follows:

[0047]

[0048] in To enhance the key feature vector, R and a are the radius and center of the hypersphere corresponding to the feature data description model in high-dimensional space, C and ξ i ,i=1,2,...,n lb are adjustment parameters and relaxation coefficients respectively. The optimized radius and sphere center are:

[0049]

[0050] in, and are the nonlinear mapping function and kernel function respectively, α i ,i=1,2,...,n lb is the Langrange coefficient corresponding to each sample. Next, evaluate the input variable data set X∈R n×m The distance between each unlabeled sample in the high-dimensional feature space and the center of the hypersphere is expressed as follows:

[0051]

[0052] in, is the mapping of the key features corresponding to each unlabeled sample in the high-dimensional feature space. Arrange the distance indicators corresponding to different unlabeled samples from large to small, select the S samples with the largest distance, send them to manual labeling, update the labeled training sample set, and record it as

[0053] The sixth step is to rebuild the deep partial least squares regression model based on the updated labeled training data sample set. The architecture is as follows:

[0054] The updated deep partial least squares regression model is established as follows:

[0055]

[0056] in, Matrix of key feature variables extracted for updating the model and are the load matrices corresponding to the input variables and output variables, To update the variable weight matrix of the model, and are the residual matrices corresponding to the input variables and output variables respectively, and L is the number of layers of the deep model.

[0057] The seventh step is to repeat steps (6) and (7) to obtain a deep partial least squares regression model with good algorithm convergence and performance, and store the final model parameters in the model database for future use.

[0058] The eighth step is to use the final deep partial least squares regression model obtained by the regression method to calculate the new data sample x obtained under the current operating conditions of the methanogen device. new The corresponding carbon monoxide and carbon dioxide gas contents are calculated as follows:

[0059]

[0060] in, For the new data x newThe corresponding key feature vectors of each hidden layer of the deep model, is the enhanced feature vector, It is the predicted value of final carbon monoxide and carbon dioxide gas content.

[0061] The following is a specific practical application case to illustrate the effectiveness of the present invention. The main process of the methanogenization furnace production device is as follows: Figure 1 As shown in Table 1, in order to establish an effective prediction regression model for the content of carbon monoxide and carbon dioxide gas at the outlet of the device, we selected 10 conventional measurement variables closely related to the content of carbon monoxide and carbon dioxide gas as the input of the prediction model through the flow chart of the device and combined with the operating experience of process engineers, including temperature, pressure and flow rate, etc. The information of the variables is shown in Table 1.

[0062] Table 1: Methanation furnace related variable information

[0063] Label name U1 MF outlet process gas volume U2 MF inlet gas pressure U3 F3 inlet pressure U4 MF bed upper layer temperature U5 MF bed middle temperature U6 MF bed lower layer temperature U7 MF outlet gas temperature U8 F2 outlet synthesis gas temperature U9 F2 level U10 MF inlet gas temperature Y <![CDATA[MF outlet CO + CO2 content]]>

[0064] For this production device, we collected a total of 1,000 data samples corresponding to the input variables, but only 50 samples contained both the measured values ​​of the input variables and the carbon monoxide and carbon dioxide gas content values, that is, the number of labeled data samples was only 50. Next, the implementation steps of the present invention are described in detail in conjunction with this specific process:

[0065] 1. Data preparation and initial modeling

[0066] Based on a modeling sample set consisting of 50 labeled data samples, an initial deep partial least squares regression analysis model is established, the appropriate number of latent variables and the number of hidden layers of the deep learning model are selected, and the relevant parameters of the model are stored in the model library for future use.

[0067] 2. Human-in-the-loop sample information evaluation and model cyclic update iteration

[0068] Based on the sample set corresponding to 1,000 unlabeled input variables, the information content of each data sample is evaluated by enhancing the feature data description model. At each model update step, the 10 samples with the largest information content are selected to update the deep regression model. Through the cyclic iterative update mode, a deep partial minimum regression analysis model with superior performance is finally obtained, and the relevant model parameters are stored in the model database.

[0069] 3. Online estimation of carbon monoxide and carbon dioxide gas content

[0070] Using the high-performance deep partial least squares regression model established in a human-in-the-loop manner, we first extract the latent variable information of the current new data in each hidden layer of the deep model, further obtain the enhanced key feature vectors, and input them into the final partial least squares regression model to obtain the online prediction values ​​of carbon monoxide and carbon dioxide gas content. Figure 2 The online prediction results of the method of the present invention for 1000 test samples are given. Compared with the prediction results obtained by the basic deep partial least squares model (such as Figure 3 As shown in the figure, the deep active learning model of human in the loop achieved significant improvement and obtained more accurate online prediction results of gas content.

[0071] The above embodiments are used to illustrate the present invention rather than to limit the present invention. Any modification and change made to the present invention within the spirit of the present invention and the protection scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A method for predicting gas content in a methanogen based on a human-in-the-loop deep active learning model, characterized in that: The following steps are involved: (1) Collect data samples under normal operation through the methanogen device, (2) Through manual gas sampling and analysis in the analysis room, a small amount of carbon monoxide and carbon dioxide gas content measurement data is obtained and used as the output variable of the prediction model. The corresponding data sample set is recorded as Among them, nlb represents the number of labeled sample data, (3) Based on labeled data sample set Extract the corresponding data set from the input variable sample set and record it as The initial modeling data sample set is composed of (4) Based on the initial labeled sample set A deep partial least squares regression model is established. By extracting key feature information from the data set, the regression relationship between the input variables and the output variables (i.e., carbon monoxide and carbon dioxide gas content) is established, and the obtained model parameters are stored in the model database for future use. (5) Starting from the key features extracted by the deep partial least squares regression model, a feature data description model is constructed. n×m Select unlabeled data samples with large information content, manually label them, and update the labeled data sample set to prepare for the next model training. (6) Based on the updated labeled training data sample set, the deep partial least squares regression model is re-established, and the obtained model parameters are stored in the model database for future use. (7) Repeat steps (6) and (7) to obtain a deep partial least squares regression model with good algorithm convergence and performance, and store the final model parameters in the model database for future use. (8) Using the final deep partial least squares regression model obtained by the regression method, calculate the new data sample x obtained under the current operating conditions of the methanogen device new The corresponding carbon monoxide and carbon dioxide gas contents are used for performance evaluation and optimization of the device.

2. According to claim 1, a method for predicting gas content in a methanogenerator based on a human-in-the-loop deep active learning model, characterized in that: The step (4) is specifically as follows: based on the initial labeled sample set A deep partial least squares regression model is established. By extracting key feature information from the data set, the regression relationship between the input variables and the output variables (i.e., carbon monoxide and carbon dioxide gas content) is established. The deep partial least squares regression model is expressed as follows: Among them, T (l) ,l=1,2,...,L is the key characteristic variable matrix P extracted by the model (l) ,l=1,2,...,L and Q (l) ,l=1,2,...,L are the load matrices corresponding to the input variables and output variables respectively, W (l) ,l=1,2,...,L is the variable weight matrix of the model, E (l) ,l=1,2,...,L and F (l) ,l=1,2,...,L are the residual matrices corresponding to the input variables and output variables respectively, and L is the number of layers of the deep model.

3. According to claim 1, a method for predicting gas content in a methanogenerator based on a human-in-the-loop deep active learning model, characterized in that: The step (5) is specifically as follows: starting from the key features extracted by the deep partial least squares regression model, a feature data description model is constructed, and the input variable data set X∈R n×m Unlabeled data samples with large information content are selected, manually labeled, and the labeled data sample set is updated to prepare for the next model training. The optimization function of the feature data description model is as follows: in To enhance the key feature vector, R and a are the radius and center of the hypersphere corresponding to the feature data description model in high-dimensional space, C and ξ i ,i=1,2,...,n lb are adjustment parameters and relaxation coefficients respectively, and the optimized radius and center of the sphere are: in, and are the nonlinear mapping function and kernel function respectively, α i ,i=1,2,...,n lb For each sample corresponding to the Langrange coefficient, next, evaluate the input variable data set X∈R n×m The distance between each unlabeled sample in the high-dimensional feature space and the center of the hypersphere is expressed as follows: in, The key features corresponding to each unlabeled sample are mapped in the high-dimensional feature space. The distance indicators corresponding to different unlabeled samples are arranged from large to small. The S samples with the largest distance are selected and sent to manual labeling. The labeled training sample set is updated and recorded as 4. According to claim 1, a method for predicting gas content in a methanogen based on a human-in-the-loop deep active learning model, characterized in that: The step (6) is specifically as follows: based on the updated labeled training data sample set, a deep partial least squares regression model is re-established, and the architecture is as follows: The updated deep partial least squares regression model is established as follows: in, Matrix of key feature variables extracted for updating the model and are the load matrices corresponding to the input variables and output variables, is the variable weight matrix of the updated model, and are the residual matrices corresponding to the input variables and output variables respectively, and L is the number of layers of the deep model.

5. According to claim 1, a method for predicting gas content in a methanogen based on a human-in-the-loop deep active learning model, characterized in that: The step (8) is specifically as follows: using the final deep partial least squares regression model obtained by the human in the regression method, the new data sample x obtained under the current operating condition of the methanogen device is calculated new The corresponding carbon monoxide and carbon dioxide gas contents are calculated as follows: in, For the new data x new The corresponding key feature vectors of each hidden layer of the deep model, is the enhanced feature vector, It is the predicted value of final carbon monoxide and carbon dioxide gas content.

6. According to the method for predicting gas content in a methanogenerator based on a deep active learning model of a human in the loop, in step (1), various automatic or manual measuring instruments installed in the methanogenerator are used to collect data samples of the device under normal operation. Among all these measured variables, based on engineering experience, the measured variables closely related to the content in the furnace are selected as input variables of the subsequent prediction model, and the data sample set corresponding to the input variables is obtained as X∈R n×m ,in, n is the number of input variable sample set data, and m is the number of selected conventional measurement variables.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, it implements the method for predicting gas content in a methanator based on a human-in-the-loop deep active learning model as described in any one of claims 1 to 6 above.

8. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the computer instructions are executed by the processor, the method for predicting the gas content in a methanator based on a human-in-the-loop deep active learning model as described in any one of claims 1 to 6 is implemented.

Images