Adaptive soft-sensing method for sulfur recovery process based on online learning and twin support vector regression machine

Abstract

The application discloses a sulfur recovery process adaptive soft measurement method based on instant learning and twin support vector regression machine. By using instant learning technology, similar samples to the query point are selected from a historical database based on a similarity criterion fused with Euclidean distance and JS divergence to train a twin support vector regression machine as a local model for soft measurement prediction. A TSVR local model evolution strategy is established according to a similarity measurement threshold to determine whether the TSVR local model needs to be updated to ensure the accuracy of the prediction. Finally, it is determined whether the current query point needs to be stored in the historical database according to a DMI threshold, so that the samples in the historical database can reflect the long-term changes of the sulfur recovery process. The application overcomes the problems of performance deterioration and the like of the traditional sulfur recovery device soft measurement model in the long-term use process, and improves the tracking ability of the sulfur recovery process soft measurement model to the sulfur recovery condition changes.

Description

Technical Field

[0001] This invention relates to a soft measurement method for sulfur recovery processes, and more particularly to an adaptive soft measurement method for sulfur recovery processes based on instant learning and twin support vector regression machines. Background Technology

[0002] Sulfur Recovery Units (SRUs) are crucial components of oil refining processes. They remove environmental pollutants before acidic gases are released into the atmosphere, while simultaneously recovering valuable elemental sulfur as a byproduct. The sulfur recovery process requires real-time monitoring of the concentrations of acidic gases H2S and SO2 in the exhaust gas. However, these two gases are highly corrosive to the hardware instruments, leading to frequent maintenance or replacement, malfunction of the automatic control loop, and deterioration of SRU performance. Therefore, soft sensing technology can be used in engineering to address this issue.

[0003] Soft measurement is a method and algorithm that estimates or predicts certain physical quantities or product quality in industrial processes based on existing measurement data. It can overcome the limitation that some physical quantities cannot be accurately measured within the required time. Soft measurement technology uses mathematical models to predict dominant variables that are difficult to measure, using easily measurable auxiliary variables.

[0004] Data-driven soft sensor modeling methods have been widely used in industrial systems with complex processes, establishing measurement models for difficult-to-measure parameters based on process data and developing soft sensors with high predictive performance. However, current soft sensor models, such as the cyclic echo state network used in sulfur recovery units, mostly employ global modeling methods. This means that the model is built offline based on collected historical samples. Once training is complete, the model's structure and parameters are fixed. During long-term actual online applications, its predictive performance deteriorates with changes in operating conditions; this phenomenon is called concept drift. It is generally caused by changes in the model's input-output relationship due to component aging or changes in the production environment, and is often difficult to predict and quantify. Furthermore, changes in process characteristics (such as catalyst deactivation, process drift, and changes in the state of the chemical plant) can gradually worsen the predictive performance of soft sensor models.

[0005] Due to changes in process characteristics, such as catalyst deactivation, process drift, and changes in the state of the chemical plant, predictive performance gradually deteriorates. Soft sensor modeling methods based on moving window and recursive models update the model using new samples reflecting process changes, enabling them to adapt to new operating conditions. However, these methods are difficult to apply to rapid or transient process changes. Furthermore, recursive soft sensors require time to adapt to new operating regions before achieving good accuracy. To address these issues, an adaptive soft sensor method for sulfur recovery processes based on instantaneous learning and twin support vector regression is proposed as an effective alternative for handling nonlinearities and changes in process characteristics. Summary of the Invention

[0006] Objective: The technical problem this invention aims to solve is to provide an adaptive soft measurement method for sulfur recovery based on real-time learning and twin support vector regression (TSVR). This method addresses the issue that existing soft measurement methods based on global models suffer from gradually deteriorating predictive performance during long-term online applications due to changes in process characteristics such as catalyst deactivation, process drift, and changes in the chemical plant's state, making it difficult to maintain expected high performance. The new method utilizes real-time learning technology to accurately track the process state of the sulfur recovery unit under long-term operation and updates the local model promptly when operating conditions change based on the TSVR local model evolution strategy, effectively improving the predictive performance of H2S and SO2 concentrations in the sulfur recovery unit.

[0007] Technical solution: The adaptive soft measurement method for sulfur recovery process based on real-time learning and twin support vector regression machine described in this invention includes the following steps:

[0008] Offline testing phase:

[0009] (1) Obtain historical datasets of sulfur recovery process and preprocess the above data: remove abnormal data and perform data normalization.

[0010] (2) Based on the pre-processed historical dataset of the sulfur recovery process obtained in step (1), perform data segmentation: select the first M samples of the dataset as the historical database of the sulfur recovery process {X i Y i}={x i1 x i2 , ..., x i5 y i1 y i2}, i = 1, 2, ..., M, and N samples are uniformly selected from the remaining samples in the dataset as the query sample set {X} for the sulfur recovery process. q Y q}={x q1 x q2 , ..., x q5 y q1 yq2}, q = 1, 2, ..., N, where x i1 and x q1 Represents gas flow data for the MEA area; x i2 and x q2 Represents the first airflow data for the MEA area; x i3 and x q3 Represents the second airflow data for the MEA area; x i4 and x q4 Represents gas flow data for the SWS area; x i5 and x q5 Represents airflow data for the SWS area; y i1 and y q1 Represents H2S concentration data; y i2 and y q2 This represents SO2 concentration data.

[0011] (3) Obtain query point X from the sulfur recovery process query sample set obtained in step (2). q Then, the number of times the query point X is retrieved is used to determine whether a trained TSVR local model already exists. If it is the first time retrieving query point X... q If a prediction is made, it means there is no pre-trained TSVR local model, proceed to step (4); otherwise, skip to step (9).

[0012] (4) Calculate the similarity metric for each input data x in the historical database of the sulfur recovery process using a fusion of Euclidean distance and JS divergence. i With query point X q Similarity measurement between i = 1, 2, ..., M.

[0013] (5) Input data x from the historical database of sulfur recovery process i Similarity measure calculated in step (4) Merging i = 1, 2, ..., M.

[0014] (6) Using the results obtained in step (5) i = 1, 2, ..., M, according to Sort the samples in descending order and select the top w samples as the training set for the TSVR local model.

[0015]

[0016] Among them, Ω w For the local model training set, This represents samples X1 and X in the historical dataset of the sulfur recovery process. q Similarity metric.

[0017] (7) According to the TSVR local model evolution strategy described in step (9), The threshold is updated to the value calculated in step (6).

[0018] (8) Establish a local model of Twin Support Vector Regression (TSVR), set two penalty coefficients and ε value for TSVR, select a Gaussian kernel as the kernel function and set the Gaussian kernel width, train the TSVR local model using the training set obtained in step (6), and input X. q Make a prediction and obtain the predicted value. Skip to step (10).

[0019] (9) Establish a threshold based on similarity metric The TSVR local model evolution strategy: Calculate the similarity metric between the current query point and the query point used in the previous TSVR local model evolution. The calculation method is the same as step (4). Determine the inequality conditions. If the conditions are met, it means that the current working conditions are stable. The predicted value is obtained directly using the trained TSVR local model, and the process proceeds to step (10). If the conditions are not met, it means that the working conditions have changed. The TSVR local model evolution operation is triggered, and the process jumps to step (4). The TSVR local model training sample set is re-selected using the current query point to update the TSVR local model.

[0020] (10) Use a Database Monitoring Index (DMI) strategy to determine whether to include X. q Add to the historical database of sulfur recovery processes: Calculate X q The DMI value is compared with all samples in the historical database of sulfur recovery processes. When the smallest DMI value is greater than the PDMI threshold, X is considered... q Add the sample to the historical database of sulfur recovery processes; otherwise, discard the sample. The definition of DMI is as follows:

[0021]

[0022] (11) Obtain the next query point X q Return to step (3).

[0023] After iterating the above method N times, all samples in the query sample set are queried, the MSE is calculated, and the weighting factor λ of the TSVR local model hyperparameters and similarity metric criterion, as well as the DMI threshold, are evaluated. Multiple tests are conducted to find the optimal hyperparameters.

[0024] Online application stage:

[0025] Furthermore, after offline testing is completed and each parameter is determined, the sulfur recovery adaptive soft measurement method is applied to the sulfur recovery soft measurement system for online prediction, including the following steps:

[0026] (1) Remove outliers from the historical dataset of the sulfur recovery process and save it as a historical database.

[0027] (2) Connect the signals from the various auxiliary variable sensors of the sulfur recovery unit to the soft measurement system, sample the data from the hardware sensors in real time, and combine the sampled data from each sensor into X. q ={x q1 x q2 , ..., x q5 The array format} facilitates subsequent operations, where x q1 ~x q5 These are, respectively, the gas flow sensor data for the MEA area, the first air flow sensor data for the MEA area, the second air flow sensor data for the MEA area, the gas flow sensor data for the SWS area, and the air flow sensor data for the SWS area.

[0028] (3) Obtain the query point X from step (2). q Then, it is compared with the auxiliary variable X in the historical database. i The data is normalized together. The method described in step (3) of the offline testing phase is used to determine whether there is a trained TSVR local model. If not, proceed to step (4). If there is, jump to step (9).

[0029] Steps (4) to (10) are the same as steps (4) to (10) in the offline testing phase.

[0030] (11) Waiting for the next query point X q Return to step (2).

[0031] Furthermore, the similarity metric criterion based on the fusion of Euclidean distance and JS divergence described in step (4) of the offline testing phase comprehensively considers Euclidean distance and JS divergence as similarity metrics that reflect the local domain information of sample points. This can effectively avoid the problem of inaccuracy in traditional Euclidean distance metrics and more accurately identify the differences between samples.

[0032] The formula for calculating Euclidean distance is as follows:

[0033]

[0034] JS divergence is a metric improved from KL divergence, and its calculation formula is as follows:

[0035]

[0036] Where KL(·) is the formula for calculating KL divergence, and KL divergence measures two individual probabilities X. i (x) and X q (x) The way in which the distributions differ, assuming x∈{x1, x2, ..., x... n Given n possible states, and p(x) and g(x) representing two probability distributions for different values ​​of the random discrete variable x, the KL divergence formula is as follows:

[0037]

[0038] In summary, the similarity metric based on the fusion of Euclidean distance and JS divergence is calculated using the following formula:

[0039]

[0040] Where λ is a weighting factor, and its value ranges from λ∈[0,1]. The larger λ is, the greater the role played by the JS metric.

[0041] Furthermore, in step (5), the soft measurement model adaptive evolution algorithm is described, where w = 5.

[0042] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0043] Existing global model-based soft measurement models for sulfur recovery exhibit a gradual decrease in prediction accuracy over long-term practical applications due to changes in catalyst activity and alterations in the sulfur recovery process conditions. The soft measurement method proposed in this invention, based on real-time learning and twin support vector regression, can establish a local model for prediction based on changes in current operating conditions, thereby improving prediction accuracy over long-term applications. Attached Figure Description

[0044] Figure 1 This is a flowchart of the sulfur recovery soft measurement adaptive evolution algorithm based on real-time learning and twin support vector regression machine of the present invention;

[0045] Figure 2 This is the overall architecture diagram of an industrial application that encapsulates a soft-measurement adaptive evolution algorithm for sulfur recovery;

[0046] Figure 3 This is a comparison between the actual value of hydrogen sulfide concentration during the sulfur recovery process and the value predicted by this method.

[0047] Figure 4 This is a comparison between the actual value of sulfur dioxide concentration during the sulfur recovery process and the value predicted by this method. Detailed Implementation

[0048] This embodiment uses the SRU sulfur recovery unit dataset, which contains 5 auxiliary variables and 2 dominant variables. The meaning of each variable is explained in Table 1.

[0049] Table 1. Auxiliary and dominant variables of the SRU soft sensor model.

[0050]

[0051] The specific implementation steps of this method are as follows:

[0052] Step 1: First, design an adaptive soft measurement algorithm for the sulfur recovery process based on instant learning and twin support vector regression machine, and conduct offline testing.

[0053] Step 1-1: Obtain the SRU sulfur recovery unit dataset and preprocess the data: remove abnormal data and perform data normalization.

[0054] Step 1-2: Based on the preprocessed SRU sulfur recovery unit dataset obtained in Step 1-1, perform data segmentation: randomly select the first 9000 data points from the dataset as the historical database of the sulfur recovery process {X}. i Y i}={x i1 x i2 , ..., x i5 y q1 y q2}, i = 1, 2, ..., 9000, randomly select 100 data points from the remaining samples in the dataset and query the sample set {X} for the sulfur recovery process. q Y q}={x q1 x q2 , ..., x q5 y q1 y q2}, q = 1, 2, ..., 100.

[0055] Steps 1-3: Obtain query point X from the sulfur recovery process query sample set obtained in step (2). q Next, determine whether a pre-trained TSVR local model already exists. If not, proceed to steps 1-4; if so, skip to steps 1-9.

[0056] Steps 1-4: Calculate the similarity criterion based on the fusion of Euclidean distance and JS divergence for each input data X in the historical database of the sulfur recovery process. i With query point X q Similarity measurement between i = 1, 2, ..., 9000.

[0057] The similarity criterion based on the fusion of Euclidean distance and JS divergence is calculated using the following formula:

[0058]

[0059] Where λ is a weighting factor, and its value ranges from λ∈[0,1]. The larger λ is, the greater the role of the JS metric. λ is determined during offline testing.

[0060] The formula for calculating Euclidean distance is:

[0061]

[0062] The formula for calculating JS divergence is:

[0063]

[0064] The formula for calculating KL divergence is:

[0065]

[0066] Steps 1-5: Input data x from the historical database of sulfur recovery process i Similarity measure calculated in step (4) Merging i = 1, 2, ..., 9000.

[0067] Steps 1-6: Using the results obtained in Steps 1-5 i = 1, 2, ..., 9000, according to Sort the samples in descending order and select the top w (determined empirically) samples as the training set for the TSVR local model.

[0068]

[0069] Among them, Ω w For the local model training set, This represents samples X1 and X in the historical dataset of the sulfur recovery process. q Similarity metric.

[0070] Steps 1-7: According to the TSVR local model evolution strategy described in steps 1-9, ... The threshold is updated to the value calculated in steps 1-6.

[0071] Steps 1-8: Train the Twin Support Vector Regression (TSVR) machine using the TSVR local model training set obtained in Steps 1-6. Set two penalty coefficients and ε values ​​for the TSVR machine, select a Gaussian kernel as the kernel function and set the Gaussian kernel width. The Twin Support Vector Regression machine will generate a pair of non-parallel functions on both sides of the training data points, and determine the ε-insensitive upper and lower bounds of the regression function respectively.

[0072] For non-linear training data, TSVR considers two kerneled, non-parallel functions as shown below to determine the final regression function.

[0073] f1(x)=K(x T A T w1+b1

[0074] f2(x)=K(x T A T w2+b2

[0075] Where f1(x) represents the insensitive lower bound of ε1 of the training data, and f2(x) represents the insensitive upper bound of ε2 of the training data.

[0076] These two non-parallel functions can be obtained by solving the following pair of optimization problems.

[0077] min

[0078] stY-(K(A,A T w1+eb1)≥eε1-ξ,ξ≥0

[0079] min

[0080] st(K(A,A T )w2+eb2)-Y≥eε2-η,η≥0

[0081] Based on the KKT conditions and by introducing Lagrange multipliers, the dual optimization problem of the above equation is:

[0082] max

[0083] st0≤α≤C1e

[0084] max

[0085] st0≤γ≤C2e

[0086] Where, H = [K(A, A T From [e], we can obtain the vector values ​​of f1(x) and f2(x):

[0087] [w1b1] T =(H T H) -1 H T (fa)

[0088] [w2b2] T =(H T H) -1 H T (h+γ)

[0089] Therefore, the regression function of STVR can be expressed as:

[0090]

[0091] After obtaining the regression function, the input X is... q Make a prediction and obtain the predicted value. and the true value Y i To compare the predicted results with the actual values, multiple offline tests were conducted using TSVR local model training sets of different sizes to determine the two penalty coefficients, ε value, and Gaussian kernel width of TSVR. The number of samples w in the local model training set was set to 5. The formula for calculating MSE is as follows:

[0092]

[0093] Where N is the number of samples in the sulfur recovery query sample set. After obtaining the predicted value, proceed to steps 1-10.

[0094] Steps 1-9: Establish thresholds based on similarity metrics The TSVR local model evolution strategy: Calculate the similarity metric between the current query point and the query point used in the previous TSVR local model evolution. The calculation method is the same as step (4). Judgment If the conditions are met, it means that the current working conditions are stable. The predicted value is obtained directly using the trained TSVR local model, and the process proceeds to step (10). If the conditions are not met, it means that the working conditions have changed. The TSVR local model evolution operation is triggered, and the process jumps to step (4). The TSVR local model training sample set is re-selected using the current query point to update the TSVR local model.

[0095] Steps 1-10: Use the Database Monitoring Index (DMI) strategy to determine whether to add X. q Add to the historical database of sulfur recovery processes: Calculate X q The DMI value compared with all samples in the historical database of sulfur recovery processes, where the smallest DMI value is greater than P.DMI When the threshold is reached, X q Add the sample to the historical database of sulfur recovery processes; otherwise, discard the sample. The definition of DMI is as follows:

[0096]

[0097] Steps 1-11: Obtain the next query point X q Go back to steps 1-3.

[0098] After repeating the above steps N times, all samples in the query sample set have been queried, the MSE has been calculated, and the weighting factor λ of the TSVR local model hyperparameters and similarity metric criteria, as well as the DMI threshold, have been evaluated. Multiple tests are conducted to find the optimal hyperparameters.

[0099] Step 2: Modify the sulfur recovery adaptive soft measurement algorithm based on instant learning and twin support vector regression machine from Step 1, write a Python script to complete the industrial APP backend development, and design the frontend interface for display. This enables online prediction. The specific operation steps are as follows:

[0100] Step 2-1: Upload the SRU sulfur recovery unit dataset to the big data module of the plant operating system platform to establish a sulfur recovery data sample set.

[0101] Step 2-2: Create object instances and object attributes for the parameters involved in the sulfur recovery soft measurement in the factory operating system platform. Use double type for attribute values ​​and check the historical value option so that the object attributes can save historical attribute values, which will facilitate the development of historical trend charts and monitoring data reports later.

[0102] In the factory operating system platform, the industrial app backend only manipulates data, while the frontend obtains data via API. A unified format is defined for both frontend and backend data, ensuring consistency in data names and types. This allows the backend to match data from the code to the attributes of custom-created objects, while the frontend establishes data connections through these attributes, enabling information exchange between the two. Therefore, backend algorithms can dynamically modify parameter attribute values ​​through parameter configuration, and the frontend can also connect to backend parameter attribute values ​​via API.

[0103] Steps 2-3: After completing the APP backend configuration in the above steps, complete the algorithm flow setup and Python script writing in the big data module of the factory operating system platform.

[0104] The big data module of the factory operating system platform requires a certain amount of historical data samples to perform predictive calculations. Therefore, using the sulfur recovery data sample set established in step 2-1 and the attribute names corresponding to the instance attributes, the factory operating system platform can read the data from the sample set into the big data module, facilitating subsequent predictive calculation operations. Furthermore, in actual operation, the factory operating system platform can also upload data collected by sensors to the big data module through the data acquisition platform, enabling the online operation of soft measurement algorithms.

[0105] In the experimental modeling view of the big data module's experimental management, an algorithm flowchart is built. First, sample data is connected to the sample set input component, then processed by the missing data preprocessing component, and finally fed into the Python script component. Python code is written within this component. The soft sensor model adaptive evolution algorithm described in step 1 is rewritten according to the script requirements: the data input in the Python script code is defined using `in_data[m][n]`, where `m` and `n` are natural numbers. The order of `m` corresponds to the object attributes in the object instance, and the order of `n` corresponds to historical data, with a value of 0 corresponding to the latest data. Finally, the prediction results are output using the built-in variables `out_data.append(H2S)` and `out_data.append(SO2)` in the factory operating system platform.

[0106] After completing steps 2-4, the algorithm flowchart, and the Python script design, debug the model in the experimental setup. The debugging results can be observed in the insight view. Once satisfactory results are obtained, publish the experimental model, create a new computation task, and set the computation period.

[0107] Steps 2-5: After completing the backend design of the industrial APP, design and develop the frontend interface of the industrial APP.

[0108] By utilizing the business designer module within the factory operating system platform and its built-in components such as data connectivity, alarm views, data reports, and trend charts, interface design can be completed quickly. The front-end interface includes real-time displays of predicted H2S and SO2 values, a simplified process flow diagram of the sulfur recovery unit, historical trend charts of various variable parameters, and data reports. Assuming consistency in name and type between the data in object properties and the data in the code, the data connectivity component can connect the data defined in the APP backend to the frontend, ensuring successful data connection and allowing for modification.

[0109] Steps 2-6: After the industrial APP design is completed, perform functional debugging to ensure that each function of the industrial APP meets the design requirements.

[0110] The prediction results for 100 samples in the query point sample set in this embodiment are as follows: Figure 2 and Figure 3As shown, the evaluation index used is the Mean Square Error (MSE). The MSE between the predicted and actual values ​​of H2S concentration is 0.000333, and the MSE between the predicted and actual values ​​of SO2 concentration is 0.000562. The results of this embodiment verify the effectiveness of the method.

Claims

1. An adaptive soft measurement method for sulfur recovery processes based on instant learning and twin support vector regression machines, characterized in that, Includes the following steps: (1) Obtain historical dataset of sulfur recovery process and preprocess the above data: remove abnormal data and normalize the auxiliary variables; (2) Based on the pre-processed historical dataset of sulfur recovery process obtained in step (1), perform data segmentation: select the first M samples of the dataset as the historical database of sulfur recovery process. = , 1, 2, M, and N samples are uniformly selected from the remaining samples in the dataset as the query sample set for the sulfur recovery process. , =1, 2, N, where and Represents gas flow data for the MEA area; and This represents the first airflow data for the MEA area; and This represents the second airflow data for the MEA area; and Represents gas flow data for the SWS area; and Represents airflow data for the SWS area; and Represents H2S concentration data; and Represents SO2 concentration data; (3) Obtain query points from the sulfur recovery process query sample set obtained in step (2). Then, the number of times the query point is obtained determines whether a trained TSVR local model already exists. If it is the first time the query point is obtained... If a prediction is made, it means there is no pre-trained TSVR local model, proceed to step (4); otherwise, skip to step (9). (4) Calculate the similarity criterion based on the fusion of Euclidean distance and JS divergence for each input data in the historical database of sulfur recovery process. With query point Similarity measurement between , =1, 2, M; (5) Input data from the historical database of sulfur recovery process Similarity measure calculated in step (4) Merging , =1, 2, M; (6) Using the results obtained in step (5) , =1, 2, M, according to The numerical values ​​are sorted in descending order, and the top w samples are selected as the training set for the TSVR local model. ； in, For the local model training set, This represents a sample from the historical dataset of the sulfur recovery process. and Similarity measure; (7) According to the TSVR local model evolution strategy described in step (9), the... The threshold is updated to the value calculated in step (6). ; (8) Establish a local model of twin support vector regression (TSVR), and set two penalty coefficients for TSVR. The value is selected, a Gaussian kernel is chosen as the kernel function and the Gaussian kernel width is set. The model is trained using the TSVR local model training set obtained in step (6) and input. Make predictions and obtain predicted values. Skip to step (10); (9) Establish a threshold based on similarity measurement The TSVR local model evolution strategy: Calculate the similarity metric between the current query point and the query point used in the previous TSVR local model evolution. The calculation method is the same as step (4); determine the inequality conditions. If the conditions are met, it means that the current working condition is stable. The predicted value is obtained directly using the trained TSVR local model and proceeds to step (10). If the conditions are not met, it means that the working condition has changed and the TSVR local model evolution operation is triggered, that is: jump to step (4) and use the current query point to re-select the TSVR local model training sample set to update the TSVR local model. (10) Use the database monitoring index-based DMI strategy to determine whether to... Add to the historical database of sulfur recovery processes: Calculation The DMI value between all samples in the historical database of sulfur recovery processes, where the minimum DMI value is greater than At the threshold, Add the sample to the historical database of sulfur recovery processes; otherwise, discard the sample. The definition of DMI is as follows: ； The large number of samples in the historical database of sulfur recovery processes increases the computational complexity of DMI values. To reduce this complexity, the DMI value for each sample is immediately compared with... The threshold is compared, if it is greater than The threshold immediately terminates the database monitoring index strategy, directly setting it to [the appropriate threshold]. Add to the historical database of sulfur recovery processes; (11) Obtain the next query point Return to step (3); After looping the above method N times, all samples in the query sample set are queried once, the MSE is calculated, and the weighting factor λ of the TSVR local model hyperparameters and similarity metric criteria, as well as the DMI threshold, are evaluated. Multiple tests were conducted to find the optimal hyperparameters.

2. The adaptive soft measurement method for sulfur recovery process based on real-time learning and twin support vector regression machine according to claim 1, characterized in that, After offline testing is completed and all parameters are determined, the query points in the adaptive soft measurement method for the sulfur recovery process will be used. The source is modified to the actual measurement data of the hardware instruments in the sulfur recovery unit, and applied to the sulfur recovery software measurement system for online prediction. The online prediction method includes the following steps: (1) Remove outliers from the historical dataset of the sulfur recovery process and store it as a historical database; (2) Connect the signals of each auxiliary variable sensor of the sulfur recovery unit to the soft measurement system, sample the sensor data in real time, and combine the sampled data of each sensor into a... The array format, where These are data from the gas flow sensor in the MEA area, the first air flow sensor in the MEA area, the second air flow sensor in the MEA area, the gas flow sensor in the SWS area, and the air flow sensor in the SWS area. (3) Obtain the query point from step (2) Then, auxiliary variables in the historical database The data is normalized together, and the method described in step (3) of claim 1 is used to determine whether there is a trained TSVR local model. If not, proceed to step (4); if so, jump to step (9). Steps (4) to (10) are the same as (4) to (10) in claim 1; (11) Waiting for the next query point Return to step (2).

3. The adaptive soft algorithm for sulfur recovery based on instant learning and twin support vector regression machine as described in claim 1 The measurement method is characterized by, The formula for calculating the Euclidean distance is as follows: ； JS divergence is a metric improved from KL divergence, and its calculation formula is as follows: ； in Here is the formula for calculating KL divergence, which measures two individual probabilities. and The way the distribution of differences is assumed have A number of possible states, and and Represents random discrete variables Given two probability distributions with different values, the KL divergence formula is as follows: ； In summary, the similarity criterion based on the fusion of Euclidean distance and JS divergence is calculated as follows: ； in, The weighting factor has a range of values. , The larger the value, the greater the impact of the JavaScript metric. Determined during the offline testing phase, and thus obtained The higher the similarity value, the higher the similarity between the samples.

Images