A method for online updating melt index prediction model coefficients based on data classification sliding window method

Abstract

The application discloses a method for online updating of melt index prediction model coefficients based on a sliding window method of data classification, and the method classifies new generated data according to the high and low melt indexes, establishes sub-databases with the same data length, updates the corresponding sub-databases by using the sliding window method, re-trains a model by using a new database, and obtains new model coefficients, so that the learning ability of the model for new data can be significantly improved. The application can be applied to the field of automatic updating of soft measurement model coefficients, realizes automatic updating of modeling training set data without changing the data structure of the training set, and improves the prediction precision of the soft measurement model for new data.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of automatic updating of melt index prediction models, in particular to a method for online updating of melt index prediction model coefficients based on data classification using a sliding window method. BACKGROUND

[0002] During the long-term production of high-pressure polyethylene, the process variables of the extruder gradually deviate due to various reasons, for example, the carbon deposition and the added opening agent silicon dioxide generated during the production of high-pressure polyethylene can cause the filter screen to be blocked, which can affect the operating parameters of the extruder during normal production, so that the soft-sensing model established based on the operating parameters of the extruder has poor prediction effect.

[0003] This phenomenon of data deviation is essentially a change in the functional relationship between process variables and quality indicators, which is caused by the time-varying nature of the process variables. Because the data samples represent the information of the past working conditions, the model trained based on the past data samples cannot correspond to the current working conditions that have changed, and it is difficult to accurately predict the output variable value. If the model is still established based on these historical data samples to predict the current working conditions, the prediction accuracy of the model will inevitably decrease, and it is difficult to monitor the product quality in real time. Real-time correction and updating of model coefficients are of great significance for real-time and accurate monitoring of the melt index of the product in response to real-time working condition changes.

[0004] Many researchers have chosen various methods to enable the model to track real-time working condition changes and improve the prediction accuracy of the soft-sensing model. Reference 1: Tang H S, Xue S T, Chen R, et al. Online weighted LS-SVM for hysteretic structural system identification [J]. Engineering Structures, 2006, 28(12): 1728-1735. He-Sheng Tang et al. adopted the traditional sliding window method to update the modeling data in real time, and used the LS-SVM modeling method to identify the system parameters. The application of the sliding window method enables the LS-SVM algorithm to track the sudden or slow time-varying changes of the system parameters, so as to detect the damage of the data and evaluate the severity of the data structure damage.

[0005] Reference 2: Wang C, Yu Z, Meng F. Discounted moving window recursive PLS algorithm and its application in polypropylene production process[J]. CIESC Journal, 2013, 64(12): 4592-4598. Wang Chunpeng et al. For the problem of polypropylene melt index prediction model in long-term prediction process, a discounted moving window recursive partial least squares algorithm (DMW-RPLS) is proposed, that is, the coefficients of the soft measurement model established by the discounted moving window recursive partial least squares algorithm are updated, and the online update of the polypropylene soft measurement model is realized.

[0006] Reference 3: Jin H P, Pan B, Chen X G, et al. Ensemble just-in-time learning framework through evolutionary multi-objective optimization for soft sensor development of nonlinear industrial processes[J]. Chemometrics and Intelligent Laboratory Systems, 2019, 184: 153-166. The ensemble model method establishes multiple prediction models, and in the prediction process, the average or weighted combination of multiple model calculation values is obtained, which essentially enhances the learning ability of the model to new data by increasing the number of models. However, due to the increase in the number of models, the modeling efficiency is reduced, and the model calculation cost is increased.

[0007] In practice, the prior art mainly has the following defects:

[0008] The ensemble model method is complex, time-consuming and has high calculation cost in real-time prediction process, which is not considered. The sliding window method is directly used to update the data without distinction, ignoring the problem of unstable data structure in the application process. The problem of unstable data structure refers to the wide range of high-pressure polyethylene melt index in the production process, and the data structure is unstable. Mainly in the window sliding process, the proportion of data of different orders of magnitude in the window changes constantly, which makes the model prediction accuracy unstable. SUMMARY

[0009] In order to overcome the shortcomings of the soft measurement model prediction accuracy changing with the structure of the training set data, and the prediction accuracy of data of different orders of magnitude being unstable, so that the model has higher prediction accuracy for industrial data, the present application proposes a method for online updating the melt index prediction model coefficients based on the sliding window method of data classification.

[0010] To solve the above technical problems, the technical scheme of the present application is:

[0011] The application provides a method for updating melt index prediction model coefficients on line based on data classification sliding window method, which comprises the following steps:

[0012] (1) arranging original data set of polyolefin extrusion production process according to time sequence of data generation, and dividing the original data set into multiple sub-databases with same dimension according to melt index;

[0013] (2) judging newly generated data, and placing the data in tail part of corresponding sub-database according to melt index, and removing first data in head part of the sub-database to realize database updating;

[0014] (3) judging whether melt index prediction needs to be updated, returning to step (2) if the prediction does not need to be updated, and training melt index prediction model by using the updated database in step (2) if the prediction needs to be updated to obtain a new model coefficient;

[0015] (4) repeating step (2) and step (3) once for each newly generated data, and repeating the steps to realize continuous updating of model coefficients.

[0016] According to the preferred embodiment of the application, each data in the original data set comprises extruder current, die pressure, barrel temperature, extruder rotating speed and actually measured melt index.

[0017] According to the preferred embodiment of the application, the original data set is divided into multiple sub-databases with same dimension according to melt index, and the division is as follows:

[0018] According to melt index MI, the original data set is divided into three categories: A category, MI=30-60 g / 10min; B category, MI=5-30 g / 10min; and C category, MI=0-5 g / 10min;

[0019] The last M data of the three categories are reserved to establish three sub-databases with same length, and the three sub-databases are combined to form a standard database for melt index prediction.

[0020] According to the preferred embodiment of the application, in step (2), for the newly generated data, firstly, the data is determined to be divided into which sub-database according to melt index, and then the data is placed in tail part of the corresponding sub-database, and the first data in head part of the sub-database is removed.

[0021] According to the preferred embodiment of the application, the melt index prediction model in step (3) is as follows:

[0022]

[0023] In the formula, I is the extruder current (A), P is the die pressure of the extruder (kg / cm 2 ), T is the barrel temperature of the extruder (K), N is the extruder speed (rpm); wherein a, b, c, d, e, f, g, h, i are all model coefficients to be fitted.

[0024] According to the preferred embodiment of the present application, the judgment in step (3) whether the melt index needs to be updated is specifically:

[0025] For the newly generated data, the extruder current, die pressure, barrel temperature and extruder speed in the data are substituted into the current melt index prediction model, the preset result is compared with the actual measured value of the melt index in the data, and a prediction relative error is obtained;

[0026] If the prediction relative error is less than the set value, the model is not trained, and the model coefficient is not updated; if the prediction relative error is not less than the set value, the model is trained, and a new model coefficient is obtained to update the model.

[0027] The technical effects of the present application are embodied in:

[0028] 1) The present application is a training set data updating method based on a soft measurement model, which has short running time and can improve the learning ability of the soft measurement model to new data to improve the prediction accuracy of the model to new data.

[0029] 2) The updating method of the present application classifies the data in advance on the basis of the sliding window method, fixes the training set data structure, and ensures that the prediction error of the model to different orders of magnitude of new data remains stable, which can provide strong technical support for accurate monitoring of product quality. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 is a principle diagram of the sliding window method based on data classification;

[0031] Figure 2 is a prediction effect diagram of a model that cannot automatically update the prediction effect of long-term slow time-varying data: (a) test set; (b) promotion set;

[0032] Figure 3 is a prediction effect diagram of a model that automatically updates the prediction effect of long-term slow time-varying data without discrimination: (a) test set; (b) promotion set;

[0033] Figure 4 is a prediction effect diagram of a model that automatically updates the prediction effect of long-term slow time-varying data based on classification: (a) test set; (b) promotion set;

[0034] Figure 5 is the prediction effect diagram of the sudden time-varying data by the non-difference automatic updating model: (a) test set; (b) promotion set;

[0035] Figure 6 is the prediction effect diagram of the sudden time-varying data by the non-difference automatic updating model: (a) test set; (b) promotion set;

[0036] Figure 7 is the prediction effect diagram of the sudden time-varying data by the model automatically updated based on the classification-based sliding window method: (a) test set; (b) promotion set. DETAILED DESCRIPTION

[0037] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0038] Figure 1 is the principle diagram of the present application. When the method for online updating the melt index prediction model coefficients based on the data classification-based sliding window method is applied to industrial data, the following steps are included:

[0039] (1) The original data set is defined as a window queue with a length of N, N∈(1, 2, 3, …, t), and the window queue elements include X1, X2, X3, …, XN. t The content in the element X includes the extruder current, the die pressure, the cylinder temperature, the extruder speed, and the actually measured melt index. According to the high and low of the melt index, the data is divided into three categories: A category (MI=30-60 g / 10 min), a B category (MI=5-30 g / 10 min), and a C category (MI=0-5 g / 10 min). The last 50 values of the data in each category are reserved, and three sub-databases of high, medium, and low are respectively established to be combined into a standard database for melt index prediction.

[0040] (2) Ensure that the training sample amount is sufficient. In actual melt index prediction model modeling, the training sample amount is set to 150, that is, the sample amount of the three sub-databases is set to 50. Before model training, the sample amount in the window is judged. If the sample amount of a sub-database is less than 50, the model cannot be trained. If the sample amount of all sub-databases is 50, the model can be trained to obtain the model coefficients.

[0041] The auxiliary variables of the polyethylene melt index soft measurement model studied in the present application include the extruder current I (A), the die pressure P (kg / cm2) of the extruder, the cylinder temperature T (℃) of the extruder, the extruder speed N (rpm), and the melt index MI (g / 10 min). 2), barrel temperature T(K) of the extruder, rotation speed N(rpm) of the extruder, and the main variable is the melt index MI(g / 10min) of the polyethylene, and the model form is shown in formula 1.

[0042]

[0043] wherein a, b, c, d, e, f, g, h, i are model coefficients to be fitted.

[0044] (3) For the newly generated data in the process, firstly, according to the high and low of the melt index, determine which sub-database should be divided into; then place it at the tail of the corresponding sub-database, and delete the first data at the head of the sub-database, that is, the sample size of each sub-database is always 50.

[0045] (4) Substitute the model coefficients obtained in step (2) into the model, and use the corresponding auxiliary variables of the new data to predict the melt index value, save the predicted value; according to the actual measured value of the melt index in the new data, calculate the prediction relative error of the melt index, and save it;

[0046] If the prediction relative error is greater than 8%, it is considered that the prediction accuracy of the current model is low, and the latest data updated in step (3) is used to train the model to obtain new model coefficients and update the model. If the prediction relative error is less than 8%, the accuracy of the current model meets the requirements, and the model coefficients are not updated, and the last model coefficients are continued to be used. It should be noted that at this time, although the model coefficients are not updated, the sub-database is still updated. That is, each new data needs to update the sub-database, and needs to calculate the prediction value according to the new data, and calculate the prediction relative error of the prediction value and the measured value, and save it.

[0047] (5) In the subsequent process, repeat steps (3) and (4) to realize the update of the melt index soft measurement model coefficients of the polyethylene.

[0048] The following is the technical scheme for improving the monitoring accuracy of the melt index prediction model based on the operation parameters of the extruder based on the automatic updating program, which combines the drawings and six specific application examples to illustrate the effectiveness of the technical scheme provided by the above embodiments of the present application. Specific embodiment one

[0050] The production process data of the extruder which changed slowly in 3 months were selected for simulation, and the data were divided into training set, test set and promotion set according to time sequence, wherein the training set contained 1000 data points, the test set contained 650 data points, and the promotion set contained 1000 data points. The data in the promotion set and the data in the training set and the test set were derived from two different batches. The model coefficients were trained based on the data in the training set, and the data in the test set and the promotion set were predicted by the static model which could not be automatically updated. The prediction effect of the test set is shown in Figure 2 (a), and the prediction effect of the promotion set is shown in Figure 2 (b).

[0051] From Figure 2 (a), it can be seen that the average prediction error of the test set by the model which could not be automatically updated was 0.94, the average relative error was 8.26%, and the root mean square error was 0.004. From Figure 2 (b), it can be seen that the average prediction error of the promotion set by the model was 1.45, the average relative error was 10.52%, and the root mean square error was 0.004. With the passage of time, the deviation between the predicted value and the actual measured value of the model gradually increased. It can be seen that the prediction effect of the data by the static model which could not be automatically updated was poor. Specific embodiment two

[0053] The production process data of the extruder which changed slowly in 3 months were selected for simulation, and the data were divided into training set, test set and promotion set according to time sequence, wherein the training set contained 1000 data points, the test set contained 650 data points, and the promotion set contained 1000 data points. The data in the promotion set and the data in the training set and the test set were derived from two different batches. The data in the test set and the promotion set were predicted by the model which was automatically updated without distinction. The model which was automatically updated without distinction means that the model coefficients were updated without classification of the data by the traditional sliding window method. The prediction effect of the test set is shown in Figure 3 (a), and the prediction effect of the promotion set is shown in Figure 3 (b).

[0054] From Figure 3 (a), it can be seen that the average prediction error of the test set by the model which was automatically updated without distinction was 1.22, the average relative error was 7.53%, and the root mean square error was 0.004. From Figure 3(b) can be seen, after the model is updated without discrimination, the prediction accuracy of the model of embodiment one is significantly improved, the prediction average error is reduced by 0.45, the average relative error is reduced by 5.54%, and the root mean square error is reduced by 0.002, so the model coefficient updated without discrimination can effectively improve the prediction effect of the model on data. However, the model prediction value and the actual measured value have obvious deviation in the process of switching from low melt index to medium melt index and from medium melt index to high melt index, and the prediction effect is poor. Embodiment three

[0056] In this embodiment, a group of extruder production process data which changes slowly in 3 months of actual operation is selected for simulation, and the data is divided into a training set, a test set and a promotion set in time sequence, wherein the training set contains 1000 data points, the test set contains 650 data points, and the promotion set contains 1000 data points. The data in the promotion set and the data in the training set and the test set are derived from two different batches. The model automatically updated based on the classification sliding window method proposed in the application is used to predict the data in the test set and the promotion set. The prediction effect of the test set in this embodiment is shown in Figure 4 (a), and the prediction effect of the promotion set is shown in Figure 4 (b).

[0057] From Figure 4 (a) can be seen, the prediction accuracy of the classification automatic update model is higher than that of embodiment one and embodiment two on the test set, and the prediction relative error is smaller. The prediction average error is 0.88, the average relative error is 6.11%, and the root mean square error is 0.003. It can be seen that the classification automatic update model has better prediction effect on the overall data. From Figure 3 (b) can be seen, the average error of the classification automatic update on the promotion set is 0.86, the average relative error is 6.05%, and the root mean square error is 0.002. The prediction value and the actual measured value can basically coincide, so the prediction effect of the classification automatic update model on the promotion set data is the best. Embodiment four

[0059] In this embodiment, a group of extruder production process data which changes greatly in a short time before and after the extruder is repaired is selected for simulation, and the data is divided into a training set, a test set and a promotion set in time sequence, wherein the training set contains 1000 data points, the test set contains 650 data points, and the promotion set contains 1000 data points. The data in the promotion set and the data in the training set and the test set are derived from two different batches. A static model which cannot be automatically updated is used to predict the data in the test set and the promotion set. The prediction effect of the test set in this embodiment is shown in Figure 5(a) and (b) show that the prediction effect of the promotion set is as shown in Figure 5 (b) shows.

[0060] From Figure 5 As can be seen from (a), the average relative error of the model for predicting the test set cannot be automatically updated below 10%, and the predicted value and the measured value can be well coincided, which shows that the prediction effect of the model for the test set data is good. As can be seen from Figure 5 (b), when the soft measurement model with fixed coefficients is used to predict the promotion set, the prediction result of the soft measurement model after the mutation of the operating point deviates far from the normal production range, and cannot accurately predict the product melt index. It can be seen that the model of the model that cannot be automatically updated is difficult to accurately predict the mutation data. Specific embodiment five

[0062] In this embodiment, a section of the extruder maintenance before and after, in a short time, the production process data of the extruder that changes greatly is simulated, and the data is divided into training set, test set and promotion set according to time sequence, wherein the training set contains 1000 data points, the test set contains 650 data points, and the promotion set contains 1000 data points. The data in the promotion set and the data in the training set and the test set are derived from two different batches. The model of no difference automatic update is used to predict the data of the test set and the promotion set. The model of no difference automatic update refers to the sliding window method in the traditional sense without classifying the data to update the model coefficients. The prediction effect of the test set in this embodiment is as shown in Figure 6 (a), and the prediction effect of the promotion set is as shown in Figure 6 (b).

[0063] From Figure 6 As can be seen from (a), the average relative error of the model for predicting the test set cannot be automatically updated below 10%, and the predicted value and the measured value can be well coincided, which shows that the prediction effect of the model for the test set data is good. As can be seen from Figure 6 (b), when the soft measurement model with fixed coefficients is used to predict the promotion set, the prediction result of the soft measurement model after the mutation of the operating point deviates far from the normal production range, and cannot accurately predict the product melt index. It can be seen that the model of the model that cannot be automatically updated is difficult to accurately predict the mutation data. Specific embodiment six

[0065] The embodiment selects a section of the extruder before and after maintenance, and simulates the production process data of the extruder which changes greatly in a short time. The data is divided into a training set, a test set and a promotion set according to time sequence, wherein the training set contains 1000 data points, the test set contains 650 data points, and the promotion set contains 1000 data points. The data in the promotion set and the data in the training set and the test set are derived from two different batches. The model automatically updated based on the classification sliding window method is used to predict the data of the test set and the promotion set. The prediction effect of the test set is shown in Figure 7 (a), and the prediction effect of the promotion set is shown in Figure 7 (b).

[0066] From Figure 7 (a), it can be seen that the model automatically updated based on the classification sliding window method has an average relative error of less than 10% in predicting the test set, and the predicted value and the measured value can well coincide, which indicates that the prediction effect of the model on the test set data is good. From Figure 7 (b), it can be seen that the classification sliding window method has good prediction effect in the process of updating and correcting the model coefficients.

[0067] The method of updating the model coefficients online based on the data classification sliding window method provided by the above embodiment of the present application can make the model have good prediction ability for both slow change process and mutation process, can consider the prediction of normal long-term production and extruder maintenance, is safe, simple, timely and reliable, can improve the prediction precision of the soft measurement model, provides effective monitoring basis for safe and stable production of the production process, and improves product quality and industrial production economic benefits.

[0068] The specific embodiments of the present application are described above. It should be understood that the present application is not limited to the above specific embodiments, and various modifications or changes can be made by those skilled in the art within the scope of the claims, which does not affect the essential content of the present application.

Claims

1. A method for updating the coefficients of a melt index prediction model on line based on a data classification sliding window method, comprising the following steps: (1) arranging the original data set of a polyolefin extrusion production process in the order of data generation time, and dividing the original data set into a plurality of sub-databases with the same dimensions according to the level of melt index; said dividing the original data set into a plurality of sub-databases with the same dimensions according to the level of melt index is specifically: dividing the original data set into three categories according to the level of melt index MI: category A, MI = 30-60 g / 10 min; category B, MI = 5-30 g / 10 min; and category C, MI = 0-5 g / 10 min; retaining the last M data of the three categories to establish three sub-databases with the same length, and combining the three sub-databases into a standard database for melt index prediction; (2) judging the newly generated data, and placing it at the tail of the corresponding sub-database according to the level of melt index, and removing the first data at the head of the sub-database to update the database; (3) judging whether the melt index prediction needs to be updated, if not, returning to step (2); if yes, training the melt index prediction model using the updated database of step (2) to obtain a new model coefficient; said judging whether the melt index prediction needs to be updated in step (3) is specifically: for the newly generated data, substituting the extruder current, die pressure, barrel temperature and extruder speed in the data into the current melt index prediction model, comparing the preset result with the actual measured value of the melt index in the data to obtain a prediction relative error; if the prediction relative error is less than a set value, the model is not trained and the model coefficient is not updated; if the prediction relative error is not less than the set value, the model is trained and a new model coefficient is obtained to update the model; (4) repeating steps (2) and (3) once for each new data, and repeating the cycle to realize continuous updating of the model coefficient.

2. The method of claim 1, wherein, each data in the original data set comprises an extruder current, a die pressure, a barrel temperature, an extruder speed and an actually measured melt index.

3. The method of claim 1, wherein, in step (2), for the newly generated data, first determine which sub-database it should be divided into according to the level of melt index, and then place it at the tail of the corresponding sub-database and remove the first data at the head of the sub-database.

4. The method of claim 1, wherein, the melt index prediction model in step (3) is as follows: ； In the formula, I is the current of the extruder (A), P is the die pressure of the extruder (kg / cm 2 ), T is the barrel temperature of the extruder (K), and N is the rotational speed of the extruder (rpm); wherein a, b, c, d, e, f, g, h, and i are model coefficients to be fitted.

5. The method of claim 1, wherein, the set value is preferably 8%.

Images