An online soft measurement method for intrinsic viscosity based on double-gaussian bayesian network

Abstract

The application provides an online soft measurement method for intrinsic viscosity based on a double Gaussian Bayesian network, which comprises a pre-training stage and an online soft measurement stage; before the pre-training stage, historical data are collected from a sensor of a polyester fiber polymerization process, and the historical data are divided into training samples and test samples; in the pre-training stage, a prediction model is trained by using the training samples, and a correction model is trained by using the training samples; in the online soft measurement stage, the trained prediction model and the trained correction model are used to obtain a final prediction value of intrinsic viscosity of a sample to be measured; and the method provided by the application overcomes the shortcomings of poor self-adaptability and performance decline of a traditional offline modeling method, and improves the quality of a polymer and the performance of a finished fiber.

Description

Technical Field

[0001] This invention belongs to the field of chemical fiber production process technology, and relates to an online soft measurement method for intrinsic viscosity based on a dual Gaussian Bayesian network. Background Technology

[0002] Polyester fiber (PET) is a basic raw material for the production of synthetic fibers, films, filaments, and plastic products, possessing advantages such as high strength, wrinkle resistance, and low water absorption. Among various apparel fibers, polyester fiber has gained wider applications than other synthetic fibers due to its excellent elastic recovery, chemical stability, and resource availability. The direct spinning production process of polyester fiber mainly involves four steps: the polymerization reaction of organic diacids and diols to obtain a polymer with a certain viscosity; the polymer is then transported to the spinning workshop for spinning and post-processing to obtain the finished polyester fiber.

[0003] With the development of technologies such as computers, sensors, and data storage, the polyester fiber industry has generated a large amount of industrial data, which is of high value for modeling the polyester fiber production process. Modeling methods for the polyester fiber polymerization process can be divided into data-driven modeling and mechanistic modeling. Mechanistic modeling requires comprehensive and accurate theoretical support and is based on many simplifications and assumptions, thus facing significant obstacles in its application to complex industrial processes. Data-driven modeling only involves the input and output variables of the process and does not require detailed analysis of its internal mechanisms, making it more suitable for modeling polyester fiber polymerization processes with nonlinearity and uncertainty. Data-driven modeling relies heavily on industrial big data. Soft-sensor models can be established by training models using machine learning and deep learning methods and analyzing the nonlinear mapping relationships between input and output variables, and is commonly used in soft-sensor modeling of industrial processes. Since polyester fiber polymerization process data often exhibits severe data collinearity, principal component regression and partial least squares regression methods can effectively eliminate the correlation between variables, describing the characteristics of the object with fewer variables. With the continuous development of technologies such as machine learning, deep learning, and artificial intelligence, algorithms such as Gaussian process regression, support vector machines, and deep neural networks have been widely applied in data-driven modeling.

[0004] Polymerization is the first step in the entire polyester fiber production process, and the quality of the resulting polymer directly affects the quality of the final product. Polymer quality is measured by intrinsic viscosity, but measuring intrinsic viscosity presents several challenges: traditional sensors are difficult to use for direct measurement; viscometer measurements are easily affected by environmental factors, resulting in low accuracy; and laboratory analysis is time-consuming, costly, and subject to time delays. Therefore, timely and accurate soft sensing of the polymer's intrinsic viscosity is crucial. This allows for real-time monitoring of viscosity changes, enabling workers to make appropriate adjustments to operating conditions such as raw materials, catalysts, pressure, and temperature. This is essential for ensuring the smooth operation of the entire production process.

[0005] Currently, scholars have employed various methods to perform soft measurement on the intrinsic viscosity of polyester fibers, establishing corresponding mechanistic and data-driven models with some success. However, these models still have some problems that urgently need to be addressed. Due to the inherent characteristics of the polyester fiber polymerization process, such as multivariability, strong interference, large hysteresis, and strong coupling, these models will experience data drift if used online for too long. Without correction, this will inevitably affect the accuracy of the soft measurement. Furthermore, the data from the polyester fiber polymerization process exhibits significant time-varying and nonlinear characteristics. Once conventional offline models are trained, their structure or parameters are fixed, making it impossible to adjust the models to adapt to changes in production conditions. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing offline models, such as poor adaptability and decreased online performance, by providing an online soft measurement method for intrinsic viscosity that can correct offline predicted values ​​online based on actual production conditions while ensuring sufficient operating efficiency. This invention employs a dual-model approach to correct the output values ​​of the offline model online to compensate for the performance deficiencies of a single model. It monitors model performance through a dynamic variable window, and adjusts model parameters incrementally when performance declines, thereby improving model update efficiency and ultimately optimizing the production process and improving the quality of polyester fiber production.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] A method for online soft measurement of intrinsic viscosity based on a dual Gaussian Bayesian network includes a pre-training phase and an online soft measurement phase.

[0009] Before the pre-training phase, historical data was collected from sensors in the polyester fiber polymerization process and divided into training samples and test samples.

[0010] In the pre-training phase, the prediction model is first trained using training samples, and then the correction model is trained using training samples.

[0011] The prediction model uses Gaussian Bayesian Network I. The input to Gaussian Bayesian Network I is the process parameter X related to intrinsic viscosity, and the output of Gaussian Bayesian Network I is the initial predicted value of intrinsic viscosity. The process of training the prediction model is the process of learning and updating the parameters of the Gaussian Bayesian network I. Except for the training samples, the specific training process is the same as in the literature (Zhu M, Liu S, Jiang JA novel divergence for sensitivity analysis in Gaussian Bayesian networks[J]. International Journal of Approximate Reasoning,2017,90:37-55.). The network structure of the prediction model adopts a multi-input single-output approach, that is, each input variable is the direct parent node of the output variable, and the input variables follow the assumption of conditional independence.

[0012] The calibration model uses a Gaussian Bayesian Network II, and the workflow of the calibration model is as follows:

[0013] (i) will The input is fed into a Gaussian Bayesian network II, which outputs a quadratic prediction value Y' of the intrinsic viscosity; the output of the prediction model is then used. As one of the inputs to the calibration model, it is to make fuller use of the effective information of the prediction model;

[0014] (ii) Input D' = {X', Y'} into Gaussian Bayesian Network II, and have it output the predicted value μ of the bias variable B. BX ', Y represents the true value of the intrinsic viscosity;

[0015] The process of training the calibration model is the process of learning and updating the parameters of the Gaussian Bayesian Network II. Apart from the training samples, the specific training process is the same as in the literature (Zhu M, Liu S, Jiang JA novel divergence for sensitivity analysis in Gaussian Bayesian networks[J]. International Journal of Approximate Reasoning, 2017, 90:37-55.).

[0016] In the prior art, the training of Gaussian Bayes networks is divided into network structure training and network parameter training. Network structure training is to find an optimal network structure through a scoring function. The termination condition is generally that the value of the scoring function no longer changes or the change is less than a threshold. The network structure of the present invention is determined based on the conditional independence assumption and expert experience. Therefore, it is not necessary to train the network structure, but only to train the network parameters.

[0017] This invention can use a correction model to correct the output of a prediction model, overcoming the performance defects of traditional single-model modeling, and making full use of the powerful uncertainty reasoning ability of Gaussian Bayes networks, making it suitable for complex process industrial modeling with strong time-varying characteristics and high coupling.

[0018] In the online soft measurement phase, the final predicted value of the intrinsic viscosity of the sample under test is obtained by using the trained prediction model and the trained correction model.

[0019] As a preferred technical solution:

[0020] As described above, the online soft measurement method for intrinsic viscosity based on a dual Gaussian Bayesian network has a training sample to test sample ratio of 2 to 4:1.

[0021] As described above, the online soft measurement method for intrinsic viscosity based on a dual Gaussian Bayesian network requires process parameters with a Pearson correlation coefficient higher than 0.6 to be associated with intrinsic viscosity. When using the Pearson correlation coefficient for feature selection, this invention found that most features had a very low correlation with intrinsic viscosity (correlation coefficient less than 0.3), with only a few features maintaining a correlation coefficient of around 0.6 to 0.7. Considering that the focus of this invention is not on feature selection, the threshold is simply set to 0.6.

[0022] The above describes an online soft measurement method for intrinsic viscosity based on a dual Gaussian Bayesian network. Intrinsic viscosity is a key indicator of polymer quality. The polymer is the final product of the polymerization process, which includes esterification, pre-condensation, and final condensation units, specifically comprising five parts, such as... Figure 1 As shown:

[0023] (1) Terephthalic acid feed and slurry ratio: The crude terephthalic acid produced by the refined terephthalic acid oxidation unit enters the slurry tank through the relevant conveying device, is dissolved in water, and is then transported to the preheater by a booster pump;

[0024] (2) Esterification: The esterification reactor consists of a steam separator and a heat exchanger. The mixture of terephthalic acid and ethylene glycol enters the heat exchanger and mixes with the oligomers in the tubes to undergo an esterification reaction. The by-products generated by the reaction enter the steam separator for recycling, and the main product enters the injection system from the discharge port at the bottom of the circulation pipe.

[0025] (3) Oligomer transport and additive injection: The oligomers generated by the esterification reaction are transported to the prepolymerization reactor through pipelines, and additives are added to the pipelines, mainly including matting agents, catalysts and diethylene glycol;

[0026] (4) Prepolymerization: Under vacuum conditions, the product of the esterification part enters the bottom of the prepolymerization reactor and rises layer by layer along the riser pipe. During the rise, the first polymerization is carried out. The ethylene glycol vapor generated during the reaction can play a stirring role. After the reaction is completed, the prepolymer enters the final polymerization reactor at the material outlet.

[0027] (5) Final polycondensation: The prepolymer is transported to the inlet of the final polycondensation reactor by a metering pump. During the transport process in the reactor, the stirrer continuously stirs it to complete the final polycondensation reaction. The intrinsic viscosity also gradually increases, and finally a high-quality finished polymer is obtained.

[0028] The process parameters related to intrinsic viscosity are TPA rotary valve speed, slurry preparation tank current, online slurry injection pressure, and CAT spray tank temperature.

[0029] The above-described method for online soft measurement of intrinsic viscosity based on dual Gaussian Bayesian networks, wherein the parameters of Gaussian Bayesian network I are the mean and variance of Gaussian Bayesian network I; and the parameters of Gaussian Bayesian network II are the mean and variance of Gaussian Bayesian network II.

[0030] The above-described method for online soft measurement of intrinsic viscosity based on a dual Gaussian Bayesian network employs incremental learning to fine-tune the parameters of the Gaussian Bayesian network during the online soft measurement stage. It monitors model performance through a dynamic variable window and updates the model when its performance deteriorates, thereby avoiding the inefficiency caused by frequent updates.

[0031] The online soft measurement method for intrinsic viscosity based on a dual Gaussian Bayesian network, as described above, has the following process for the online soft measurement stage:

[0032] (a) The test sample at time t is input into the trained prediction model and the trained calibration model, respectively. The prediction model outputs the initial predicted value of the intrinsic viscosity. The predicted value of the output bias variable B of the calibration model

[0033] (b) Use a window to monitor the performance of the calibration model. The start and end positions of the window are two time points. Assume the start time is t1 and the end time is t2. Initially, both t1 and t2 point to 0, and the window length is 0. Time t is traversed from 1. If the calibration model does not need to be updated, t1 remains unchanged and t2 is incremented by 1. If the calibration model needs to be updated, the calibration model is updated with the test samples from the time period t1 to t2, and the test samples at time t are re-predicted. Then, both t1 and t2 are reset to t, and the window length is 0. The window length will change dynamically according to the performance of the calibration model. It is called a dynamic variable window.

[0034] Determine the rate of change Δμ of the conditional mean at time t t Is it less than or equal to the statistical limit SL? To correct the quadratic prediction of intrinsic viscosity output by the model in step (i), The intrinsic viscosity is represented by y. i Let represent the true value of the characteristic viscosity of the training sample with index i, H represent the number of training samples, and SL is determined according to the distribution of each data point in the training set, so that the proportion of the number of training samples within the statistical limit is θ, where θ is 0.95 (i.e., the number of training samples within the statistical limit accounts for 0.95 of the total number of training samples, and the number of training samples outside the statistical limit accounts for 0.05).

[0035] If so, add the sample to be tested to the dynamically variable window, increment the window length by 1, and proceed to the next step;

[0036] If not, then the mean of the Gaussian Bayesian Network II (a Gaussian Bayesian Network can be defined as a pair (G,P), where G represents a directed acyclic graph consisting of nodes and directed edges, nodes represent random variables, directed edges represent dependencies between variables, and P defines the joint probability density between variables, which is a multivariate normal distribution N(μ,Σ), where μ is the mean of the Gaussian Bayesian Network) is updated according to the following formula, while the window length of the dynamic variable window is reset to 0;

[0037]

[0038]

[0039]

[0040]

[0041] In the formula, h represents the number of times the mean of the Gaussian Bayesian Network II is updated, with an initial value of 0, and μ h To update the mean of the previous Gaussian Bayesian Network II, μ h+1This is the mean of the updated Gaussian Bayesian Network II. β represents the mean of all test samples stacked within a dynamically variable window (test samples within the time range corresponding to the start and end points of the window are the stacked samples within the window; the mean of the stacked samples is a vector, where each element represents the mean of the corresponding input variable among the test samples within the window). h and α h M represents the corresponding weights. h N represents the number of test samples used for the h-th update (the number of test samples for the h-th update is numerically equal to the length of the dynamically variable window). h M represents the sum of the number of training and test samples consumed before the h-th update, where H represents the number of training samples and M represents the number of test samples. i This represents the number of test samples used for the i-th update;

[0042] After the update is complete, the calibration model is used again to predict the sample to be tested at time t (since the prediction model has not been updated, it is not necessary to re-input into the prediction model, that is, it is not necessary to re-use the prediction model to predict the sample to be tested at time t) and then proceed to the next step.

[0043] Incremental learning aims to learn new knowledge while retaining most of the previously learned old knowledge. It continuously adjusts the structure or parameters of the model through new data, making full use of the rich information brought by online data. The calibration model can compensate for the output of the prediction model to a certain extent, making the prediction results closer to the actual output. Since the structure or parameters of the dual Gaussian Bayesian network model do not change after training, and it cannot effectively utilize the information contained in the new samples, it is difficult to make corresponding adjustments to the changes in production conditions. The prediction performance of the entire model will deviate from the expected performance with continuous online application. In order to adjust the model parameters according to the changes in working conditions during the prediction process, incremental learning (i.e., updating) is used to update the mean of the calibration model based on the offline calibration model of the dual Gaussian Bayesian network.

[0044] Updating the model using incremental learning can combine the information contained in the new samples with the original training results, and greatly reduce the model update time by fine-tuning the mean of the calibration model, thus meeting the real-time requirements of online modeling.

[0045] A dynamically variable window controls the number of test samples used for each update. Evaluating model performance involves considering both whether its accuracy meets expectations and whether the algorithm's time and space complexity are excessive. Time efficiency is particularly crucial for online models; frequent adjustments inevitably impact online prediction performance. For calibration models, the mean should be updated selectively. Specifically, when operating conditions are stable and performance indicators are within controllable ranges, the current model meets production needs without adjustment. When operating conditions change significantly or performance indicators exceed limits, the current model cannot meet production requirements, necessitating adjustments to adapt and ensure sufficient prediction accuracy. A dynamically variable window monitors model performance; if expected performance is met, the model is not updated; otherwise, the test samples within the window are used to update the model's mean. The conditional mean of the samples reflects the accuracy of the model's inference; therefore, in this invention, the rate of change of the mean, Δμ, is selected. t As a performance indicator, the statistical limit SL in this invention determines the length of the variable window. When SL is too large, too many samples accumulate in the window, making it impossible to make timely and effective adjustments to changes in operating conditions. When SL is too small, it will lead to some unnecessary updates, thus affecting the model efficiency. Therefore, the size of the statistical limit affects the update frequency of the model.

[0046] (c) Calculate the final predicted value of the intrinsic viscosity at time t. When time t+1 arrives, let t = t+1, and then return to step (a).

[0047] Beneficial effects

[0048] (1) The present invention provides an online soft measurement method for intrinsic viscosity based on a dual Gaussian Bayesian network. By utilizing the powerful uncertainty reasoning ability of the Gaussian Bayesian network to learn the posterior distribution of the bias, it can correct the offline predicted value of intrinsic viscosity (i.e. the output value obtained by the prediction model), overcome the problem of performance degradation of traditional offline models in online applications, improve the prediction accuracy of intrinsic viscosity, and produce low-cost, high-quality polyester fiber products.

[0049] (2) The present invention provides an online soft measurement method for characteristic viscosity based on a dual Gaussian Bayesian network. It uses incremental learning to update the calibration model parameters, and can learn rich information from new samples based on existing knowledge. By monitoring the model performance through a dynamic variable window, the model update frequency is reduced, which can adapt to real-time changes in working conditions and help improve the economic benefits and market competitiveness of enterprises. Attached Figure Description

[0050] Figure 1 This is a flow chart of DuPont's three-reactor process for polyester fiber polymerization.

[0051] Figure 2 To correct the model network structure diagram;

[0052] Figure 3 This is a schematic diagram of a dynamically variable window;

[0053] Figure 4 This is a graph showing the trend of TPA rotary valve rotation speed during the production process;

[0054] Figure 5 A graph showing the trend of current variation in the slurry preparation tank during the production process;

[0055] Figure 6 This is a graph showing the trend of online slurry injection pressure during the production process.

[0056] Figure 7 A graph showing the temperature variation trend of the CAT spray tank during the production process;

[0057] Figure 8 This is a schematic diagram of the process for an online soft measurement method of intrinsic viscosity based on a dual Gaussian Bayesian network. Detailed Implementation

[0058] The present invention will be further described below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0059] An online soft measurement method for intrinsic viscosity based on a dual Gaussian Bayesian network, such as... Figure 8 As shown, this includes the pre-training phase and the online soft measurement phase, which are carried out sequentially.

[0060] Before the pre-training phase, historical data was collected from sensors in the polyester fiber polymerization process and divided into training samples and test samples; the ratio of training samples to test samples was 2 to 4:1.

[0061] In the pre-training phase, the prediction model is first trained using training samples, and then the correction model is trained using training samples.

[0062] The prediction model uses Gaussian Bayesian Network I. The input to Gaussian Bayesian Network I is the process parameter X related to intrinsic viscosity, and the output of Gaussian Bayesian Network I is the initial predicted value of intrinsic viscosity. The process of training the prediction model is the process of learning and updating the parameters of Gaussian Bayesian Network I. Except for the training samples, the specific training process is the same as in the literature (Zhu M, Liu S, Jiang JA novel divergence for sensitivity analysis in Gaussian Bayesian networks[J]. International Journal of Approximate Reasoning, 2017, 90:37-55.). Among them, the process parameters related to intrinsic viscosity are the TPA rotary valve speed, slurry preparation tank current, online slurry injection pressure, and CAT spray tank temperature; the parameters of Gaussian Bayesian Network I are the mean and variance of Gaussian Bayesian Network I.

[0063] The calibration model uses Gaussian Bayesian Network II, such as Figure 2 As shown, the workflow for calibrating the model is as follows:

[0064] (i) will The input is fed into a Gaussian Bayesian network II, which outputs a quadratic prediction value Y' of the intrinsic viscosity.

[0065] (ii) Input D' = {X', Y'} into Gaussian Bayesian Network II, and have it output the predicted value μ of the bias variable B. BX ', Y represents the true value of the intrinsic viscosity;

[0066] The process of training the calibration model is the process of learning and updating the parameters of Gaussian Bayesian Network II. Except for the training samples, the specific training process is the same as in the literature (Zhu M, Liu S, Jiang JA novel divergence for sensitivity analysis in Gaussian Bayesian networks[J]. International Journal of Approximate Reasoning, 2017, 90:37-55.); where the parameters of Gaussian Bayesian Network II are the mean and variance of Gaussian Bayesian Network II.

[0067] The process for the online soft measurement phase is as follows:

[0068] (a) The test sample at time t is input into the trained prediction model and the trained calibration model, respectively. The prediction model outputs the initial predicted value of the intrinsic viscosity. The predicted value of the output bias variable B of the calibration model

[0069] (b) such as Figure 3As shown, a window is used to monitor the performance of the calibration model. The start and end positions of the window are two time points. Assume the start time is t1 and the end time is t2. Initially, both t1 and t2 point to 0, and the window length is 0. Time t is traversed starting from 1. If the calibration model does not need to be updated, t1 remains unchanged, and t2 is incremented by 1. If the calibration model needs to be updated, the calibration model is updated with the test samples from the time interval t1 to t2, and the test samples at time t are re-predicted. Then, both t1 and t2 are reset to t, and the window length is 0. The window length will change dynamically according to the performance of the calibration model, which is called a dynamic variable window.

[0070] Determine the rate of change Δμ of the conditional mean at time t t Is it less than or equal to the statistical limit SL? To correct the quadratic prediction of intrinsic viscosity output by the model in step (i), The intrinsic viscosity is represented by y. i Let represent the true value of the characteristic viscosity of the training sample with index i, H represent the number of training samples, and SL be determined according to the distribution of each data point in the training set, so that the proportion of training samples within the statistical limit is θ, where θ is 0.95.

[0071] If so, add the sample to be tested to the dynamically variable window, increment the window length by 1, and proceed to the next step;

[0072] If not, then update the mean of the Gaussian Bayesian Network II using the test samples within the dynamically variable window according to the following formula, while resetting the window length of the dynamically variable window to 0.

[0073]

[0074]

[0075]

[0076]

[0077] In the formula, h represents the number of times the mean of the Gaussian Bayesian Network II is updated, with an initial value of 0, and μ h To update the mean of the previous Gaussian Bayesian Network II, μ h+1 This is the mean of the updated Gaussian Bayesian Network II. β represents the mean of all test samples stacked within a dynamically variable window. h and α h M represents the corresponding weights. h N represents the number of test samples used for the h-th update. hM represents the sum of the number of training and test samples consumed before the h-th update, where H represents the number of training samples and M represents the number of test samples. i This represents the number of test samples used for the i-th update;

[0078] After the update is complete, the calibration model is used again to predict the sample to be tested at time t and then proceed to the next step.

[0079] (c) Calculate the final predicted value of the intrinsic viscosity at time t. When time t+1 arrives, let t = t+1, and then return to step (a).

[0080] The present invention will now be illustrated with a specific example to illustrate an online soft measurement method for intrinsic viscosity based on a dual Gaussian Bayesian network:

[0081] Example 1

[0082] In the modeling of the polyester fiber polymerization process, the sensor sampling interval was 1 minute, and the dataset included features such as temperature, pressure, flow rate, and liquid level. Based on the Pearson correlation coefficient, features with a high correlation coefficient with intrinsic viscosity were selected as input variables. Ultimately, four feature variables were chosen as input variables: TPA (terephthalic acid) rotary valve speed, slurry preparation tank current, online slurry injection pressure, and CAT (catalyst) injection tank temperature. The intrinsic viscosity of the polymer was used as the output variable. After preprocessing the data, 2400 samples were selected, with the first 1800 samples used as the training set and the last 600 samples used as the test set. The trends of the input variables are shown below. Figures 4-7 As shown.

[0083] The predicted intrinsic viscosity values ​​were compared with the actual values. The baseline model was a single Gaussian Bayesian network model (for offline prediction using a single Gaussian Bayesian network, i.e., the prediction model in this invention). The comparison results are shown in Table 1. Overall, the online calibration model using a dual Gaussian Bayesian network (i.e., the method of this invention) outperformed the other two models in all indicators. The goodness of fit of the offline calibration model using a dual Gaussian Bayesian network (which differs from this invention in that the calibration model was not updated) was approximately 48% higher than that of the single Gaussian Bayesian network model, while the goodness of fit of the online calibration model using a dual Gaussian Bayesian network was approximately 4% higher than that of the offline calibration model using a dual Gaussian Bayesian network. This indicates that the online calibration model using a dual Gaussian Bayesian network, while addressing the problem of low offline prediction accuracy, can further improve the predictive performance of the model by adjusting the model parameters online.

[0084] To further verify the superiority of the dual Gaussian Bayesian network online calibration model, it was compared with other online calibration models. Instant learning (reference: Yuan XF, Ge ZQ, Huang B, et al. Semisupervised JITL framework for nonlinear industrial soft sensing based on locally semisupervised weighted PCR[J]. IEEE Transactions on Industrial Informatics, 2016, 13(2): 532-541.) performs online soft sensing through local modeling. Bias correction (reference: Efron B. Bootstrap Methods: Another Look at the Jackknife[J]. The Annals of Statistics, 1979, 7(1): 1-26.) corrects the current output value by adding the predicted value to the deviation from the previous time step. The comparison results are shown in Table 1. Real-time learning constructs local models using similar sample sets to achieve online modeling, overcoming the drawback of unchanging global models and effectively improving the prediction accuracy of offline models. However, since it requires rebuilding the model for each prediction, the algorithm has high time complexity. Bias correction correlates the current predicted value with the prediction bias from the previous time step, but does not consider the impact of the overall error distribution on the experimental results. The dual Gaussian Bayesian network online correction model corrects offline predicted values ​​by calculating the conditional mean of the bias and updates model parameters using incremental learning and a dynamically variable window, significantly improving the model's prediction accuracy and operating efficiency. Experimental results show that the online soft measurement method for characteristic viscosity based on dual Gaussian Bayesian networks in this invention outperforms traditional modeling methods and effectively solves the problems of low accuracy and poor dynamism in offline modeling of industrial processes.

[0085] Table 1 Comparison of prediction errors of different models (MAE is mean absolute error, MSE is mean squared error, R...) 2 (For the coefficient of determination)

[0086]

Claims

1. A method for online soft measurement of intrinsic viscosity based on a double Gaussian Bayesian network, characterized in that, The method comprises a pre-training stage and an online soft measurement stage in sequence. Before the pre-training stage, historical data is collected from sensors of a polyester fiber polymerization process and divided into training samples and test samples; In the pre-training stage, the prediction model is trained using the training samples, and then the correction model is trained using the training samples; The prediction model adopts a Gaussian Bayesian network I, an input of the Gaussian Bayesian network I is a process parameter X related to the intrinsic viscosity, and an output of the Gaussian Bayesian network I is a first prediction value of the intrinsic viscosity A process of training the prediction model, that is, a process of learning and updating parameters of the Gaussian Bayesian network I; The correction model adopts a Gaussian Bayesian network II, and the working process of the correction model is as follows: (i) inputting to the Gaussian Bayesian network II, from which a second predicted value Y' of the intrinsic viscosity is output; (ii) inputting D' = {X', Y'} into the Gaussian Bayesian network II, from which a predicted value μ of the bias variable B is output B丨X' , Y is the true value of the intrinsic viscosity; The process of training the correction model is a process of learning and updating the parameters of the Gaussian Bayesian network II. In the online soft measurement stage, the trained prediction model and the trained correction model are used to obtain the final predicted value of the intrinsic viscosity of the to-be-measured sample. 2.The online soft-sensing method of intrinsic viscosity based on double Gaussian Bayesian network according to claim 1, wherein, The number ratio of the training samples to the test samples is 2-4:

1. 3.The online soft-sensing method of intrinsic viscosity based on double Gaussian Bayesian network according to claim 1, characterized in that, The process parameters related to the intrinsic viscosity are process parameters with a Pearson correlation coefficient with the intrinsic viscosity higher than 0.

6. 4.The online soft-sensing method of intrinsic viscosity based on double Gaussian Bayesian network according to claim 3, characterized in that, The process parameters related to the intrinsic viscosity are TPA rotary valve speed, slurry preparation tank current, online slurry injection pressure, and CAT spray tank temperature.

5. The method of claim 1, wherein the method is based on a dual Gaussian Bayesian network. The parameters of the Gaussian Bayesian network I are the mean and variance of the Gaussian Bayesian network I; and the parameters of the Gaussian Bayesian network II are the mean and variance of the Gaussian Bayesian network II.

6. The method according to claim 1, wherein, The process of the online soft measurement stage is as follows: (a) the sample to be tested at time t is input into the trained prediction model and the trained correction model respectively, the prediction model outputs the initial predicted value of the intrinsic viscosity the correction model outputs the predicted value of the bias variable B (b) A window is used to monitor the performance of the correction model, the start and end positions of the window are two time points, assuming that the start time is t1 and the end time is t2, initially, t1 and t2 point to 0, the window length is 0, and the time t starts from 1, if the correction model does not need to be updated, t1 remains unchanged and t2 is increased by 1; if the correction model needs to be updated, the test samples in the t1-t2 time period are used to update the correction model, the to-be-measured sample at time t is predicted again, then t1 and t2 are reset to t, and the window length is 0, the window length dynamically changes according to the performance of the correction model, which is called a dynamically variable window; a rate of change of the condition mean value at time t, Δμ t whether it is less than or equal to a statistical limit SL, to correct the secondary predicted value of the intrinsic viscosity output by the model at the stage of step (i), denotes the mean value of the intrinsic viscosity, y i denotes the true value of the intrinsic viscosity of the training sample with index i, H denotes the number of training samples, SL is determined according to the distribution of each data point in the training set, so that the proportion of the number of training samples within the statistical limit is θ, and θ is 0.95; If yes, the to-be-measured sample is added to the dynamically variable window to increase the window length by 1, and the next step is entered; If no, the mean of the Gaussian Bayesian network II is updated using the test samples in the dynamically variable window according to the following formula, and the window length of the dynamically variable window is reset to 0; where h represents the number of updates of the mean of Gaussian Bayesian Network II, the initial value is 0, μ h is the mean of Gaussian Bayesian Network II before updating, h+1 is the mean of Gaussian Bayesian Network II after updating, represents the mean of all test samples accumulated in the dynamic variable window, β h and α h respectively represent the corresponding weights, M h represents the number of test samples for the hth update, N h represents the sum of the number of training samples and test samples consumed before the hth update, H represents the number of training samples, M i represents the number of test samples for the ith update; After the update is completed, the to-be-measured sample at time t is predicted again using the correction model, and the next step is jumped to; If yes, the to-be-measured sample is added to the dynamically variable window to increase the window length by 1, and the next step is entered; If no, the mean of the Gaussian Bayesian network II is updated using the test samples in the dynamically variable window according to the following formula, and the window length of the dynamically variable window is reset to 0; After the update is completed, the to-be-measured sample at time t is predicted again using the correction model, and the next step is jumped to; (c) calculating a final predicted value of the intrinsic viscosity at time t When time t+1 arrives, let t = t+1 and return to step (a).

Images