A monitoring method and system for batch production of resin reaction process

Abstract

The present invention discloses a method and system for monitoring the progress of an intermittent resin production reaction. The monitoring method includes the following steps: collecting historical data and preprocessing the data; mining key features that affect the molecular weight of reactants based on the preprocessed data; dividing the resin reaction stages and establishing reactant molecular weight prediction models for different time periods based on the key features; establishing a time series prediction model based on the output of the reactant molecular weight prediction model for different time periods; inputting the output of the reactant molecular weight prediction model for different time periods or the reactant molecular weight detected offline into the time series prediction model to generate a curve of reactant molecular weight versus reaction time; and monitoring the reaction progress based on the reactant molecular weight change curve to determine the reaction stop time. This method can reduce the number of sampling and testing times, reduce tedious and repetitive work for employees, and more accurately predict the stop time of the resin reaction.

Description

Technical Field

[0001] The present invention relates to the technical field of production reaction process monitoring, and in particular to a monitoring method and system for intermittent resin production reaction process. Background Art

[0002] Resin production often utilizes batch production, a crucial industrial production method that involves processing raw materials into products that meet quality requirements within a limited timeframe according to pre-designed process flows. Compared to continuous production, batch production is characterized by complexity, non-steady-state behavior, multiple time periods, and significant time lags. Repeated production of the same formula under the same production conditions can result in significant product quality variations. Due to the lack of online molecular weight monitoring instruments for key reactants during the reaction, process engineers must infer the current state of the reaction based on multiple offline molecular weight measurements and, based on experience, estimate the reaction stop time.

[0003] These traditional reaction progress monitoring methods not only suffer from lags and large deviations, but also require multiple manual offline testing operations, resulting in high labor costs. Furthermore, these traditional reaction progress monitoring methods require experienced personnel to perform the calculations, and inexperienced personnel require a certain level of training before they can perform the task, which also incurs certain additional costs.

[0004] Most of the existing chemical reaction process monitoring methods measure reaction rates based on mechanisms, equipment, imaging technology, etc., which are not suitable for monitoring resin reactions. In addition, the cost of relying on mechanism and equipment measurement is relatively high.

[0005] Therefore, new methods need to be explored to monitor the resin reaction process. The present invention aims to monitor the resin reaction process by establishing a mathematical prediction model, reduce the number of sampling and testing times, reduce the tedious and repetitive work of staff, and more accurately predict the resin reaction stop time. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a monitoring method and system for the intermittent production of resin reaction progress, which can reduce the number of sampling and testing times, reduce the tedious and repetitive work of staff, and more accurately predict the resin reaction stop time.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions.

[0008] In a first aspect, the present invention provides a method for monitoring the progress of a batch resin production reaction, comprising the following steps:

[0009] S1. Collect historical data and preprocess the data;

[0010] S2. Mining key features that affect the molecular weight of the reactants based on the preprocessed data, the key features including utility-related features and non-utility-related features after screening and filtering;

[0011] S3. The reaction stage of the resin is divided, and a molecular weight prediction model of the reactant is established at different time periods based on the key characteristics affecting the molecular weight of the reactant;

[0012] S4. Establishing a time series prediction model based on the output of the reactant molecular weight prediction model at different time periods;

[0013] S5. The output of the reactant molecular weight prediction model or the offline detected reactant molecular weight at different time periods is input into the time series prediction model to generate a curve of the reactant molecular weight versus reaction time;

[0014] S6. Monitoring the reaction progress based on a curve showing the change in molecular weight of the reactant versus reaction time, and determining the reaction stop time.

[0015] Furthermore, step S5 specifically includes:

[0016] S51. When the reaction process reaches a certain time period, the real-time data collected from the start of the reaction to the end of the previous time period is used to calculate the similarity between the real-time data and the historical data;

[0017] S52. Select the most similar historical data to supplement the data for the subsequent time period, and call the reactant molecular weight prediction model for calculation;

[0018] S53. Input the output of the reactant molecular weight prediction model into the time series prediction model to generate a curve showing the change of the reactant molecular weight with the reaction time.

[0019] Furthermore, step S5 further includes:

[0020] S54. In the last reaction time period, the actual value of the reactant molecular weight in the previous time period detected manually offline is used to replace the output value of the reactant molecular weight prediction model, and the curve of the reactant molecular weight changing with the reaction time is corrected.

[0021] Furthermore, the historical data in step S1 includes some or all of batch information, production execution record information and bit number information; the data preprocessing includes abnormal data removal, null value filling, feature extraction or standardization.

[0022] Furthermore, in step S2, screening and filtering non-public engineering related features includes: performing low variance filtering and high correlation elimination on non-public engineering features, and then using an integrated machine learning algorithm to perform integrated combination sorting on the features, and selecting several non-public engineering features with the highest sorting.

[0023] Furthermore, the integrated machine learning algorithm integrates random forest, Pearson correlation coefficient, maximum mutual information and gradient boosting decision tree algorithms. The calculation formula for standardizing the evaluation results of different algorithms is as follows:

[0024]

[0025] In the formula, newS represents the standardized result, s i Represents the correlation coefficient of the algorithm evaluation results, s i ∈S, S is the set of algorithm evaluation results;

[0026] The calculation formula for the integration of the standardized algorithm is as follows:

[0027] r=w1*newSRF+w2*newSP+w3*newSMIC+w4*newSGB

[0028] Where r represents the correlation coefficient set calculated by ensemble selection, w1, w2, w3, and w4 represent weight coefficients, newSRF represents the correlation coefficient set of random forest evaluation, newSP represents the correlation coefficient set of Pearson correlation coefficient evaluation, newSMIC represents the correlation coefficient set of maximum mutual information evaluation, and newSGB represents the correlation coefficient set of gradient boosting decision tree evaluation.

[0029] Furthermore, in step S3, the reaction stage of the resin is divided into at least four different time periods, and a prediction model of the reactant molecular weight in different time periods is established using relevant statistical and machine learning algorithms, wherein the relevant statistical and machine learning algorithms include one or more of partial least squares, support vector machine, random forest, neural network or gradient boosting decision tree.

[0030] Furthermore, in step S4, the outputs of the reactant molecular weight prediction model in different time periods are input as samples into the grayscale prediction model to establish the following time series prediction model:

[0031]

[0032] Among them, a represents the development coefficient, b represents the endogenous control gray number, represents the molecular weight of the reactants in the first time period, To represent the molecular weight of the k+1th reactant in the 1-AGO series, we need to The molecular weight of the reactant in the k+1 time period is obtained by cumulative subtraction calculation. The specific formula is as follows:

[0033]

[0034] in, represents the molecular weight of the reactant in the k+1 time period, represents the molecular weight of the kth reactant in the 1-AGO series.

[0035] In a second aspect, the present invention provides a monitoring system for an intermittent resin production reaction process, comprising:

[0036] Data collection module, used to collect historical data;

[0037] A data preprocessing module, comprising an abnormal data elimination unit, a null value filling unit, a feature extraction unit or a normalization unit, for preprocessing the collected historical data;

[0038] A key feature mining module, comprising a low-variance filtering unit and a high-correlation elimination unit, as well as an integrated combined ranking unit that integrates random forest, Pearson correlation coefficient, maximum mutual information, and gradient boosting decision tree algorithms. This module is used to mine key features that affect reactant molecular weight based on preprocessed data. These key features include utility-related features and non-utility-related features after screening and filtration.

[0039] The reactant molecular weight prediction module includes reactant molecular weight prediction units for each time period, which are used to predict the reactant molecular weight for each time period based on the data corresponding to the mined key features;

[0040] The time series prediction module is used to generate a curve showing the change of reactant molecular weight with reaction time based on the output of the reactant molecular weight prediction module or the reactant molecular weight detected offline, monitor the reaction process, and determine the reaction stop time.

[0041] Furthermore, the reactant molecular weight prediction module also includes a reaction stage division unit and a similarity calculation unit. The reaction stage division unit is used to divide the reaction process into several time periods. The similarity calculation unit is used to calculate the similarity between real-time data and historical data, and select the most similar historical data to supplement the data of subsequent time periods.

[0042] Compared with the prior art, the method and system for monitoring the progress of intermittent resin production reaction of the present invention have the following beneficial effects:

[0043] A data-driven model is built using historical production data, dividing the resin reaction into multiple stages. Predictive models for reactant molecular weights in different time periods are then established. Based on the output of the predictive models, a time series prediction model is then built. This generates a curve showing the molecular weight versus reaction time. The curve is then used to monitor the reaction progress and determine the reaction stop time. This reduces the number of sampling and testing times, alleviates tedious and repetitive employee work, and allows for more accurate predictions of the resin reaction stop time. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a main flow chart of the method for monitoring the progress of the intermittent resin production reaction in this embodiment;

[0045] Figure 2 This is a flow chart of the relevant feature selection in the method for monitoring the progress of the intermittent resin production reaction in this embodiment;

[0046] Figure 3 Schematic diagram of a curve showing the change in molecular weight of reactants versus reaction time in the method for monitoring the progress of a batch resin production reaction in this embodiment;

[0047] Figure 4 This is a comparison chart of the endpoint detection value of the molecular weight of the reactants in the resin reaction process of this embodiment and the stop reaction time of the traditional reaction process monitoring;

[0048] Figure 5 Another comparison chart of the endpoint detection value of the molecular weight of the reactants in the resin reaction process of this embodiment and the stop reaction time of the traditional reaction process monitoring;

[0049] Figure 6 This is a diagram of the structure of the monitoring system for the intermittent production of resin reaction progress provided by the present invention. DETAILED DESCRIPTION

[0050] To clearly illustrate the objectives, technical solutions, and advantages of the present invention, the following will refer to the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions of the present invention through implementation methods. Obviously, the embodiments described are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0051] like Figure 1 As shown; the monitoring method for the intermittent production resin reaction process provided by the present invention mainly comprises the following steps:

[0052] S1. Collect historical data and preprocess the data.

[0053] The collected data mainly include batch information, production execution record information, and position number information, among which the position number information mainly includes kettle temperature, hot oil jacket inlet oil temperature, cold oil jacket inlet oil temperature, kettle top temperature, etc.

[0054] Specifically, batch information and production execution record information come from the MES (Manufacturing Execution System), a management system for workshop production. In the process from the issuance of work orders to the completion of finished products, the manufacturing execution system plays the role of transmitting information to optimize production activities.

[0055] Data processing includes the removal of abnormal data, filling in null values, feature extraction, and standardization.

[0056] Abnormal data is eliminated using the 3δ method, also known as the Laida criterion. The 3δ method assumes that a set of test data contains only random errors, calculates the standard deviation, and then determines an interval based on a certain probability. It assumes that any error exceeding this interval is not a random error but a gross error, and data containing this error is eliminated.

[0057] Null value filling is done with the mean. Missing data is a common problem in data analysis. When the missing percentage is small, the missing records can be discarded or manually processed. In real-world data, missing data often accounts for a considerable proportion. However, discarding missing records in this situation would result in a significant loss of information, leading to systematic differences between incomplete and complete observations. The mean is a quantity that indicates the trend in a set of data. Using the mean to fill in missing data reduces the error rate in data analysis.

[0058] Feature extraction is to extract the mean value of the position number data within the production activity time range recorded in the production execution record, the standard deviation of the position number data, the activity duration, the specific value of each position number data at the start of the offline detection activity, the feeding time, etc.

[0059] Standardization primarily involves normalizing data. Different evaluation indicators often have different dimensions and units, which can affect the results of data analysis. To eliminate the dimensional effects between indicators, data standardization is necessary to achieve comparability. After data standardization, the raw data is brought to the same order of magnitude across all indicators, making it suitable for comprehensive comparative evaluation. In this example, the maximum-minimum standardization method is used, utilizing the maximum and minimum values ​​in a data column for normalization, mapping the data values ​​to a value between 0 and 1.

[0060] S2. Use correlation analysis algorithm based on preprocessed data to mine key features that affect the molecular weight of reactants.

[0061] The process of relevant feature selection is as follows Figure 2 As shown in the figure, the relevant features are divided into utility-related features and non-utility-related features. Specifically, utility-related features include the inlet and outlet temperatures and flow rates of cold oil and hot oil, and non-utility-related features include the duration of activities in the production execution record, the average temperature of the activity kettle, the amount of material fed, the properties of the raw materials, etc.

[0062] Since the resin reaction is an exothermic reaction, long-term experience shows that there is a high correlation between utility-related features and the molecular weight of the reactants. When performing feature selection, all utility-related features are selected.

[0063] Non-public engineering features are first filtered through low variance and eliminated through high correlation. Low variance filtering calculates the variance corresponding to each feature value in the sample. If it is below a preset threshold, filtering is performed. In this embodiment, the high correlation calculation method in the high correlation elimination operation uses the Pearson correlation coefficient to evaluate whether there is a relatively high linear correlation between features. If so, they need to be eliminated. The eliminated features are then sorted by feature integration and combination correlation using an integrated machine learning algorithm, and finally the non-public engineering projects with the highest ranking are selected.

[0064] It should be noted that the integrated machine learning algorithm in this embodiment is an algorithm that integrates random forest, Pearson correlation coefficient, maximum mutual information, and gradient boosting decision tree. The calculation formula for standardizing the evaluation results of different integrated algorithms is as follows:

[0065]

[0066] In the formula, newS represents the standardized result, s i Represents the correlation coefficient of the algorithm evaluation results, s i ∈S, S is the set of algorithm evaluation results.

[0067] The calculation formula for the integration of the standardized algorithm is as follows:

[0068] r=w1*newSRF+w2*newSP+w3*newSMIC+w4*newSGB

[0069] Where r represents the correlation coefficient set calculated by ensemble selection, w1, w2, w3, and w4 represent weight coefficients, newSRF represents the correlation coefficient set of random forest evaluation, newSP represents the correlation coefficient set of Pearson correlation coefficient evaluation, newSMIC represents the correlation coefficient set of maximum mutual information evaluation, and newSGB represents the correlation coefficient set of gradient boosting decision tree evaluation.

[0070] S3. Divide the resin reaction stages and establish a reactant molecular weight prediction model for different time periods based on the key characteristics that affect the molecular weight of the reactants.

[0071] In conventional reaction progress monitoring methods, the number of offline detections is divided into different time periods, wherein the number of divided time periods is at least four, so as to more accurately estimate the time when the resin reaction ends. In this embodiment, based on conventional reaction progress monitoring methods, the resin reaction time period is divided, and a reactant molecular weight prediction model for different time periods is established based on relevant statistical and machine learning algorithms to more accurately predict the future change in reactant molecular weight.

[0072] Different reactant molecular weight prediction models are established in different time periods using statistical and machine learning algorithms, including one or more of partial least squares, support vector machine, random forest, neural network, or gradient boosting decision tree.

[0073] In this embodiment, the partial least squares method is preferably used to establish a prediction model. Partial least squares regression is a new multivariate statistical data analysis method that combines the advantages of principal component analysis, canonical correlation analysis, and multiple linear regression analysis. The variables X and Y of the partial least squares method are decomposed as follows:

[0074]

[0075]

[0076] Where t1 and u1 are the first pair of score vectors, p1 and q1 are the corresponding loading vectors, and E1 and F1 are the residual matrices. The eigenvectors t1 and u1 selected by partial least squares are those with the largest covariance in the row t = Xw, u = Yc, ||w|| = ||c|| = 1. The eigenvectors are linked using a univariate linear regression model:

[0077] u1=b1t1+r1

[0078] Where b1 is the regression factor;

[0079]

[0080] Then calculate the residual matrix of X and Y:

[0081]

[0082]

[0083] Next, the same method is used to extract the second pair of eigenvectors from E1 and F1, and decompose them. The above process is repeated until the required a principal component eigenvectors are extracted.

[0084] After the reactant molecular weight prediction model is established, it is preferably evaluated using R-squared and maximum absolute error. R-squared is used to determine the degree of fit of the model, and maximum absolute error is used to evaluate the accuracy of the predicted value. The model with the highest score is selected as the final prediction model, where the scoring is based on the fact that the maximum absolute error is smaller when the R-squared values ​​are similar.

[0085] S4. Establish a time series prediction model based on the output of the reactant molecular weight prediction model in different time periods.

[0086] In this embodiment, due to the relationship between the molecular weight of the reactants and the reaction time, the outputs of the molecular weight prediction models for the reactants at different time periods are used as samples to input into a pre-set prediction model to establish a time series prediction model. In this embodiment, a grayscale prediction model is preferably selected as the pre-set model for establishing the time series prediction model. Establishing the grayscale prediction model does not require a large number of samples, nor does it require a regular distribution of samples. It has the advantages of low computational workload, high accuracy, and can be used for recent, short-term, and medium- to long-term predictions.

[0087] Using the GM(1,1) model in the grayscale prediction model, a single variable first-order differential prediction model is established based on a small amount of incomplete information. The principle of the grayscale prediction model is as follows:

[0088] The output of the reactant molecular weight prediction model at different time periods is used as the modeling input data, and the corresponding sequence data is:

[0089]

[0090] After one accumulation, we get our 1-AGO series:

[0091]

[0092] in, is the initial model of the grayscale prediction model.

[0093] The formula for generating the adjacent mean series is as follows:

[0094]

[0095] The generated sequence of adjacent means:

[0096]

[0097] Therefore is the initial form of the grey prediction model, let And there are:

[0098]

[0099] Solving using the least squares method yields:

[0100]

[0101] Solving the differential equation yields

[0102] Among them, a represents the development coefficient, b represents the endogenous control gray number, represents the molecular weight of the reactants in the first time period, To represent the molecular weight of the k+1th reactant in the 1-AGO series, we need to The molecular weight of the reactant in the k+1 time period is obtained by cumulative subtraction calculation. The specific formula is as follows:

[0103]

[0104] in, represents the molecular weight of the reactant in the k+1 time period, represents the molecular weight of the kth reactant in the 1-AGO series.

[0105] S5. The output of the reactant molecular weight prediction model or the offline detected reactant molecular weight at different time periods is input into the time series prediction model to generate a curve of the reactant molecular weight versus reaction time; step S5 specifically comprises:

[0106] S51. When the reaction process reaches a certain time period, based on the real-time data collected from the start of the reaction to the end of the previous time period, the real-time data includes temperature, pressure and other real-time data collected, and the similarity between the real-time data and the historical data is calculated;

[0107] The similarity between data is calculated using the distance calculation formula between vectors as follows:

[0108] s=d(x q ,x k )

[0109] In the formula, s represents the distance between vectors, d(x q ,x k ) represents the Euclidean distance, x q It represents real-time new data, x k It represents historical data. For example, if the response time is 400 minutes, x q is the real-time position data of the first 90 minutes of the reaction phase, then x k It must also be the historical data for the first 90 minutes.

[0110] S52. Select the most similar historical data to supplement the data for the subsequent time period, and call the reactant molecular weight prediction model for calculation;

[0111] For example, a reaction time period starts at the 90th minute, that is, the previous time period ends at the 90th minute. The real-time data corresponding to the first 90 minutes and the historical data of the corresponding time period are used to calculate the similarity. The last time period ends at the 240th minute, so the data of 90 minutes and 240 minutes are supplemented by historical data.

[0112] That is, during the monitoring process of the entire reaction process, after the reaction reaches the first time period, the real-time data of the first time period and the similar historical data corresponding to the subsequent time periods are input into the reactant molecular weight prediction model, and the reactant molecular weight is predicted only based on the real-time data of the first time period; after the reaction reaches the second time period, the real-time data of the first time period, the second time period and the similar historical data corresponding to the subsequent time period are input into the reactant molecular weight prediction model, and the reactant molecular weight is predicted based on the real-time data of the second time period and the first time period, and so on for subsequent time periods.

[0113] S53. The output of the reactant molecular weight prediction model is input into the time series prediction model to generate a curve of the reactant molecular weight versus reaction time;

[0114] Because the reaction process is divided into multiple time periods, the corresponding historical data is replaced by newly collected real-time data at the end of each time period, and the reactant molecular weight prediction model is called for recalculation. The reactant molecular weight prediction model output changes, and the reactant molecular weight vs. reaction time curve is updated. The curve is updated once each time period is reached, reducing the accumulation of errors.

[0115] S54. In the final reaction time period, the actual value of the reactant molecular weight detected in the previous stage is used to replace the output value of the reactant molecular weight prediction model and input into the time series prediction model to correct the curve of the reactant molecular weight changing with the reaction time, thereby completely generating the curve of the reactant molecular weight changing with the reaction time, such as Figure 3 shown.

[0116] S6. Monitoring the reaction progress based on a curve showing the change in molecular weight of the reactant versus reaction time, and determining the reaction stop time.

[0117] In actual application, 91 batches of historical data from a resin company were collected, 76 batches of which were used as training data for the model, and the remaining 15 batches of data were used as test data for the model. The endpoint detection value of the product reactant molecular weight (ep) and the traditional reaction process monitoring stop reaction time were compared and verified. The specific comparison results are as follows: Figure 4 、 Figure 5 As shown in the table below: ep = 0.8875 is the optimal molecular weight value of the reactant when the reaction stops.

[0118]

[0119]

[0120] The average error of molecular weight is 0.01124, the maximum error of molecular weight is 0.0254, the average error of time is 4.3964, and the maximum error of time is 8.1166. The batch with the largest predicted time error is H802110-040014, but the actual value of this batch, 0.912, is higher than the optimal molecular weight value of 0.8875, and it actually needs to react for a while longer.

[0121] In summary, the method for monitoring the progress of intermittent resin production reaction in this embodiment has high prediction accuracy, reduces the tedious and repetitive work of employees, and reduces labor costs.

[0122] like Figure 6 As shown, the present invention also provides a monitoring system for the reaction process of intermittent resin production, comprising:

[0123] Data collection module, used to collect historical data;

[0124] A data preprocessing module, comprising an abnormal data elimination unit, a null value filling unit, a feature extraction unit or a normalization unit, for preprocessing the collected historical data;

[0125] A key feature mining module, comprising a low-variance filtering unit and a high-correlation elimination unit, as well as an integrated combined ranking unit that integrates random forest, Pearson correlation coefficient, maximum mutual information, and gradient boosting decision tree algorithms. This module is used to mine key features that affect reactant molecular weight based on preprocessed data. These key features include utility-related features and non-utility-related features after screening and filtration.

[0126] The reactant molecular weight prediction module includes reactant molecular weight prediction units for each time period, which are used to predict the reactant molecular weight for each time period based on the data corresponding to the mined key features;

[0127] The time series prediction module is used to generate a curve showing the change of reactant molecular weight with reaction time based on the output of the reactant molecular weight prediction module or the reactant molecular weight detected offline, monitor the reaction process, and determine the reaction stop time.

[0128] In this embodiment, the reactant molecular weight prediction module also includes a reaction stage division unit and a similarity calculation unit. The reaction stage division unit is used to divide the reaction process into several time periods. The similarity calculation unit is used to calculate the similarity between real-time data and historical data, and select the most similar historical data to supplement the data of subsequent time periods.

[0129] The above are merely preferred embodiments of the present invention and the technical principles employed. Those skilled in the art will appreciate that the present invention is not limited to the specific embodiments herein, and that various obvious changes, readjustments, and substitutions are possible for those skilled in the art without departing from the scope of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments and may include many other equivalent embodiments without departing from the scope of the present invention. The scope of the present invention is determined by the scope of the appended claims.

Claims

1. A method for monitoring the progress of a batch production resin reaction, characterized in that: The steps include: S1. Collect historical data and preprocess the data; S2. Mining key features that affect the molecular weight of reactants based on preprocessed data. The key features include utility-related features and filtered non-utility-related features. Filtering non-utility-related features includes: filtering non-utility-related features for low variance and eliminating high correlation, then using an integrated machine learning algorithm to perform integrated combination sorting on the features, selecting several non-utility-related features with the highest sorting. The integrated machine learning algorithm integrates random forest, Pearson correlation coefficient, maximum mutual information, and gradient boosting decision tree algorithms. The calculation formula for standardizing the evaluation results of different algorithms is as follows: Where, It represents the normalized result. represents the correlation coefficient of the algorithm evaluation results, , is the set of algorithm evaluation results; The calculation formula for the integration of the standardized algorithm is as follows: Where, represents the set of correlation coefficients calculated by ensemble selection, 、 、 、 represents the weight coefficient, represents the set of correlation coefficients evaluated by random forest, represents the set of correlation coefficients evaluated by the Pearson correlation coefficient, represents the set of correlation coefficients for maximum mutual information evaluation, Represents the set of correlation coefficients for gradient boosting decision tree evaluation; S3. The reaction stage of the resin is divided, and a molecular weight prediction model of the reactant is established at different time periods based on the key characteristics affecting the molecular weight of the reactant; S4. Establishing a time series prediction model based on the output of the reactant molecular weight prediction model at different time periods; S5. Inputting the output of the reactant molecular weight prediction model at different time periods or the reactant molecular weight detected offline into the time series prediction model to generate a curve of reactant molecular weight versus reaction time; including: S51. When the reaction process reaches a certain time period, the real-time data collected from the start of the reaction to the end of the previous time period is used to calculate the similarity between the real-time data and the historical data; S52. Select the most similar historical data to supplement the data for the subsequent time period, and call the reactant molecular weight prediction model for calculation; S53. The output of the reactant molecular weight prediction model is input into the time series prediction model to generate a curve of the reactant molecular weight versus reaction time; S6. Monitoring the reaction progress based on a curve showing the change in molecular weight of the reactant versus reaction time, and determining the reaction stop time.

2. The method for monitoring the progress of a batch resin production reaction according to claim 1, wherein: Step S5 further includes: S54. In the last reaction time period, the actual value of the reactant molecular weight at the end of the previous time period detected manually offline is used to replace the output value of the reactant molecular weight prediction model, and the curve of the reactant molecular weight changing with the reaction time is corrected.

3. The method for monitoring the progress of batch resin production reaction according to claim 1, wherein: The historical data in step S1 includes some or all of batch information, production execution record information and bit number information; the data preprocessing includes abnormal data removal, null value filling, feature extraction or standardization.

4. The method for monitoring the progress of intermittent resin production reaction according to claim 1, wherein: In step S3, the reaction stage of the resin is divided into at least four different time periods, and a prediction model of the reactant molecular weight in different time periods is established using relevant algorithms of statistics and machine learning, wherein the relevant algorithms of statistics and machine learning include one or more of partial least squares, support vector machine, random forest, neural network or gradient boosting decision tree.

5. The method for monitoring the progress of batch resin production reaction according to claim 1, wherein: In step S4, the outputs of the reactant molecular weight prediction model in different time periods are input as samples into the grayscale prediction model to establish the following time series prediction model: Among them, a represents the development coefficient, b represents the endogenous control gray number, represents the molecular weight of the reactants in the first time period, To represent the molecular weight of the k+1th reactant in the 1-AGO series, we need to The molecular weight of the reactant in the k+1 time period is obtained by cumulative subtraction calculation. The specific formula is as follows: in, represents the molecular weight of the reactant in the k+1 time period, represents the molecular weight of the kth reactant in the 1-AGO series.

6. A monitoring system for the reaction process of intermittent resin production, comprising: Data collection module, used to collect historical data; Data preprocessing module, used to preprocess the collected historical data; The key feature mining module is used to mine key features that affect the molecular weight of reactants based on preprocessed data. The key features include utility-related features and non-utility-related features that have been screened and filtered. Screening and filtering non-utility-related features includes: low-variance filtering and high-correlation elimination of non-utility-related features, and then using an integrated machine learning algorithm to integrate and combine the features and select several non-utility-related features with the highest ranking. The integrated machine learning algorithm integrates random forest, Pearson correlation coefficient, maximum mutual information, and gradient boosting decision tree algorithms. The calculation formula for standardizing the evaluation results of different algorithms is as follows: Where, It represents the normalized result. represents the correlation coefficient of the algorithm evaluation results, , is the set of algorithm evaluation results; The calculation formula for the integration of the standardized algorithm is as follows: Where, represents the set of correlation coefficients calculated by ensemble selection, 、 、 、 represents the weight coefficient, represents the set of correlation coefficients evaluated by random forest, represents the set of correlation coefficients evaluated by the Pearson correlation coefficient, represents the set of correlation coefficients for maximum mutual information evaluation, Represents the set of correlation coefficients for gradient boosting decision tree evaluation; The reactant molecular weight prediction module is used to predict the reactant molecular weight based on the data corresponding to the mined key features. The module includes: when the reaction progress reaches a certain time period, based on the real-time data collected from the start of the reaction to the end of the previous time period, calculating the similarity between the real-time data and the historical data; selecting the most similar historical data to supplement the data of the subsequent time period, and calling the reactant molecular weight prediction model for calculation; inputting the output of the reactant molecular weight prediction model into the time series prediction model to generate a curve of the reactant molecular weight changing with the reaction time; The time series prediction module is used to generate a curve showing the change of reactant molecular weight with reaction time based on the output of the reactant molecular weight prediction module or the reactant molecular weight detected offline, monitor the reaction process, and determine the reaction stop time.

7. The monitoring system for batch resin production reaction process according to claim 6, characterized in that: The reactant molecular weight prediction module also includes a reaction stage division unit and a similarity calculation unit. The reaction stage division unit is used to divide the reaction process into several time periods. The similarity calculation unit is used to calculate the similarity between real-time data and historical data, and select the most similar historical data to supplement the data of subsequent time periods.

Images