Polypropylene melt index prediction model based on dynamic variational autoencoder regressor

Abstract

The present invention discloses a polypropylene melt index prediction model based on a dynamic variational autoencoder regressor. First, the encoder of the model extracts the dynamic spatiotemporal interaction features of polypropylene production process data through multi-scale dynamic graph attention and linear gate attention. Then, the decoder of the model reconstructs the input data into dynamic spatiotemporal interaction features through deconvolution and learns a robust Gaussian latent feature representation space. Finally, the regressor of the model learns the dynamic relationship between the dynamic spatiotemporal interaction features and the output variables, and outputs the predicted value of the polypropylene melt index through a fully connected layer. The polypropylene melt index prediction model based on a dynamic variational autoencoder regressor provided by the present invention improves the prediction accuracy of the polypropylene melt index and is of great significance for the development of high-performance grades and quality monitoring of petrochemical products.

Description

Technical Field

[0001] The present invention relates to the technical field of polypropylene production, and in particular to a polypropylene melt index prediction model based on a dynamic variational autoencoder regressor. Background Art

[0002] As one of the leading processes in the petrochemical industry, the refined and high-quality production of propylene polymerization is the key to ensuring the national economy and people's livelihood. Figure 1 The flow chart of the polypropylene production device provided in Example 1 of the present invention is as follows. Figure 1 As shown in the figure, 1 represents the product purification tank, 2 represents ethylene, 3 represents hydrogen, 4 represents fresh propylene, 5 represents recycled propylene, 6 represents catalyst, and 7 represents purge gas. Industrial polypropylene products are categorized by their melt index (MI) values. Rapid and accurate prediction of the MI during propylene polymerization plays a crucial role in product quality control, grade conversions, new product development, and significantly reducing transition materials in polypropylene production facilities. Currently, the commonly used MI detection method involves manual sampling and offline laboratory analysis, which cannot meet the requirements of real-time industrial production control and poses significant challenges to production quality monitoring and control. Industrial on-site quality control relies primarily on operator experience, leading to production process instability, low product quality, and low economic efficiency. Internationally available advanced control optimization software for olefin polymerization plants is extremely expensive, with limited after-sales service. The accuracy of MI predictions is insufficient to meet the high-quality production requirements of "safe, stable, long-term, full, and excellent" industrial processes. Therefore, the development of high-precision and robust online MI prediction methods for propylene polymerization processes is of great significance.

[0003] Soft sensing can be divided into two categories: process mechanism models and data-driven models. Process mechanism models require a deep understanding of the internal reaction mechanisms of industrial processes. By establishing high-order differential equations such as heat balance, energy balance, and material balance, they are iteratively solved using methods such as finite element analysis. The mechanism model is an ideal soft sensing model. Its advantage is that it can directly reflect the interrelationships between the variables within the process and has strong interpretability. However, when establishing a mechanism model, it is not advisable to make a large number of assumptions, as the actual process is completely different. Due to the multiple sources of raw materials and large fluctuations in raw material composition, the initial test conditions of the model are unstable, and many key process parameters cannot be accurately obtained in actual production, resulting in limitations in mechanism modeling. Data-driven soft sensing is the mainstream of soft sensing technology research. With the continuous development of information science and the application of distributed control systems (DCS) in enterprises, industrial sites have collected and stored massive amounts of data, providing a reliable foundation for establishing data-based soft sensing. Unlike process mechanism models, data-based soft sensing methods rely solely on industrial data and do not require any prior conditions or complex expert knowledge. They only require the selection of appropriate algorithms to capture the nonlinear relationship between easily measurable auxiliary variables and difficult-to-measure target outputs. Summary of the Invention

[0004] To address the limitations and defects of the prior art, the present invention provides a polypropylene melt index prediction model based on a dynamic variational autoencoder regressor, comprising:

[0005] Get a polypropylene dataset with continuous time Where L is the total length of the sequence and M is the number of auxiliary variables;

[0006] The polypropylene dataset is divided into datasets according to the preset time window

[0007] Where N is the number of samples in the dataset and T is the time window;

[0008] The encoder of the dynamic variational autoencoder uses dynamic graph attention to extract the spatial interaction features of auxiliary variables, as shown below:

[0009]

[0010]

[0011] A t =1U t ⊙A glob )W global +A glob (3)

[0012]

[0013] in, is the trend relationship between the i-th process variable and the j-th process variable,

[0014] is the local interaction matrix at time t, U t is the global trend information control gate, W loc 、W glob 、

[0015] is a learnable parameter, A t is the fused multi-scale interaction matrix, ⊙ is the Hadamard product, and the Pearson correlation coefficient is used to calculate the process variable A glob The trend matrix, is the spatial interaction feature of the auxiliary variables at time t, S represents the number of heads, Represents the interaction matrix A t The element in row i and column j of represents the eigenvector of the jth process variable;

[0016] The encoder of the dynamic variational autoencoder uses a convolutional neural network to extract the temporal interaction features of the auxiliary variables, which are expressed as follows:

[0017]

[0018] in, is the filter, is the time interaction feature of the auxiliary variables, is the characteristic at time t-k+1;

[0019] The temporal interaction feature and the dynamic spatial interaction feature are concatenated using an addition method, and the decoder of the dynamic variational autoencoder uses deconvolution to reconstruct input data to learn a robust Gaussian latent feature representation space for the dynamic spatiotemporal interaction feature;

[0020] The regressor uses dynamic linear gate attention to learn the dynamic relationship between the dynamic spatiotemporal interaction features and the output variable, and outputs the predicted value of the polypropylene melt index through a fully connected layer. The expression is as follows:

[0021] Attention(Q,K,V)=relu(QK T )V (6)

[0022] MultiHead(Q,K,V)=Concat(head1,…,head h )W o (7)

[0023] Among them, Q is the query, K is the key, V is the value, relu is the activation function, Concat is the vector splicing function, W o is a learnable parameter, and head. represents the features extracted in different subspaces.

[0024] Optionally, the loss function of the polypropylene melt index prediction model has three loss terms, and the expressions of the loss terms are as follows:

[0025]

[0026]

[0027]

[0028] in, is the KL divergence between the Gaussian latent feature representation space and the standard normal distribution, is the reconstruction loss of the decoder, is the Gaussian loss function of the regressor, y i is the label, μ en (·) represents the mean of the encoder mapping the auxiliary variables to the Gaussian latent space, σ en (·) represents the variance of the encoder mapping the auxiliary variables to the Gaussian latent space, μ de (·) represents the original data recovered by the decoder, μ reg (·) represents the key indicator predicted by the regressor, σ reg (·) represents the confidence of the regressor prediction.

[0029] The present invention has the following beneficial effects:

[0030] The present invention provides a polypropylene melt index prediction model based on a dynamic variational autoencoder regressor. First, the encoder of the model extracts dynamic spatiotemporal interaction features of polypropylene production process data through multi-scale dynamic graph attention and linear gate attention. Then, the decoder of the model reconstructs the input data into dynamic spatiotemporal interaction features through deconvolution and learns a robust Gaussian latent feature representation space. Finally, the regressor of the model learns the dynamic relationship between the dynamic spatiotemporal interaction features and the output variables, and outputs the predicted value of the polypropylene melt index through a fully connected layer. The polypropylene melt index prediction model based on a dynamic variational autoencoder regressor provided by the present invention improves the prediction accuracy of the polypropylene melt index and is of great significance for the development of high-performance grades and quality monitoring of petrochemical products. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a flow chart of a polypropylene production device provided in Example 1 of the present invention.

[0032] Figure 2This is a flow chart of a polypropylene melt index prediction model based on a dynamic variational autoencoder regressor provided in Example 1 of the present invention.

[0033] Figure 3 This is a schematic diagram comparing the actual value and predicted value of the polypropylene melt index provided in Example 1 of the present invention.

[0034] Figure 4a-4b This is a schematic diagram comparing input data and reconstructed data provided in the first embodiment of the present invention.

[0035] The reference numerals in the figures are: 1 represents a product purification chamber, 2 represents ethylene, 3 represents hydrogen, 4 represents fresh propylene, 5 represents regenerated propylene, 6 represents a catalyst, and 7 represents purified gas. DETAILED DESCRIPTION

[0036] In order to enable those skilled in the art to better understand the technical solution of the present invention, the polypropylene melt index prediction model based on the dynamic variational autoencoder regressor provided by the present invention is described in detail below with reference to the accompanying drawings.

[0037] Example 1

[0038] The problem of super-multidimensional, nonlinear and temporal dynamics of polypropylene production process data. First, the encoder of the polypropylene melt index prediction model based on the dynamic variational auto-encoder (DVAE) regressor adopts a dynamic gate mechanism to adaptively fuse the global trend interaction features and local spatial interaction features of the auxiliary variables to obtain multi-scale spatial interaction features. Then, the temporal interaction features of the polypropylene chemical data are extracted by replacing the original softmax with radial transformation and linear attention, which can effectively reduce the computational complexity. At the same time, the temporal multi-scale spatial interaction features and the temporal interaction features are fused to obtain the spatiotemporal interaction features. Secondly, the decoder of the polypropylene melt index prediction model based on the dynamic variational auto-encoder regressor reconstructs the input data into dynamic spatiotemporal interaction features through deconvolution and learns a robust Gaussian latent feature representation space. Finally, the dynamic regressor learns the dynamic relationship between the dynamic spatiotemporal interaction features and the output variables, and outputs the predicted value through the fully connected layer. The polypropylene melt index prediction model based on the dynamic variational auto-encoder regressor provided in this embodiment improves the accuracy of soft sensor prediction, which is of great significance for the development of high-performance grades and quality monitoring of petrochemical products.

[0039] The polypropylene melt index prediction model based on the dynamic variational autoencoder regressor provided in this embodiment adopts a dynamic gate mechanism to adaptively fuse the global trend interaction characteristics and local spatial interaction characteristics of auxiliary variables to obtain multi-scale spatial interaction characteristics.

[0040] The polypropylene melt index prediction model based on a dynamic variational autoencoder regressor provided in this embodiment uses radial transformation and linear attention instead of the original softmax to extract temporal interaction features of complex chemical data, effectively reducing computational complexity. Simultaneously, it fuses temporal multiscale spatial interaction features with temporal interaction features to obtain spatiotemporal interaction features.

[0041] The polypropylene melt index prediction model based on dynamic variational autoencoder regressor provided in this embodiment first uses dynamic variational autoencoder to extract the dynamic spatiotemporal interaction features of auxiliary variables in Gaussian latent space, and then the dynamic regressor predicts key indicators based on the dynamic spatiotemporal interaction features. This embodiment uses a continuous time polypropylene dataset Where L is the total length of the sequence and M is the number of auxiliary variables. The dataset is divided into fixed time windows to obtain the dataset Where N is the number of samples in the dataset, and T is the time window size. The encoder of the dynamic variational autoencoder uses dynamic graph attention and convolutional neural networks (CNNs) to extract the temporal interaction features and spatial dynamic interaction features of auxiliary variables, respectively. The calculation formula for the spatial interaction features of auxiliary variables extracted by dynamic graph attention at time t is as follows:

[0042]

[0043]

[0044] A t =(U t ⊙A glob )W glo,al +A glob (3)

[0045]

[0046] Among them, A glob is the trend matrix of the process variable, is the local interaction matrix at time t, A t is the fused multi-scale interaction matrix, and ⊙ represents the Hadamard product. In this embodiment, the Pearson correlation coefficient is used to calculate the process variable A glob The trend matrix, is the spatial interaction characteristic of the auxiliary variables at time t.

[0047] The calculation formula for extracting temporal interaction features using the convolutional neural network provided in this embodiment is as follows:

[0048]

[0049] in, is the filter, is the temporal interaction feature of the auxiliary variable. Finally, the temporal interaction feature and the dynamic spatial interaction feature are concatenated using an additive method. The decoder of the dynamic variational autoencoder uses deconvolution to reconstruct the input data and learn a robust Gaussian latent feature representation space for the dynamic spatiotemporal interaction feature.

[0050] The regressor provided in this embodiment uses dynamic linear gate attention to learn the dynamic relationship between dynamic spatiotemporal interaction features and output variables, and outputs the predicted value through the fully connected layer. The expression is as follows:

[0051] Attention(Q,K,V)=relu(QK T )V (6)

[0052] MultiHead(Q,K,V)=Concat(head1,…,head h )W o (7)

[0053] Among them, Q, K and V are query, key and value respectively, d k is the dimension of the input features.

[0054] In order to simultaneously constrain the encoder, decoder, and regressor of the polypropylene melt index prediction model based on the dynamic variational autoencoder regressor, this embodiment provides three loss terms for the loss function of the polypropylene melt index prediction model, as shown in the following expression:

[0055]

[0056]

[0057]

[0058] in, It is the KL divergence between the learned Gaussian latent space and the standard normal distribution, which ensures that the dynamic variational autoencoder regression model has the ability to generate data. is the reconstruction loss of the decoder, which ensures that the dynamic variational autoencoder regression model can learn a robust Gaussian latent representation space. It is the Gaussian loss function of the regressor, which can reduce the learning cost of converting from the Gaussian latent representation space to the target space and achieve accurate soft measurement modeling.

[0059] Figure 2 This is a flow chart of the polypropylene melt index prediction model based on the dynamic variational autoencoder regressor provided in the first embodiment of the present invention. Figure 2As shown, the encoder of the dynamic variational autoencoder consists of a dynamic graph attention and a convolutional neural network. The dynamic graph attention uses time-gated linear units to extract multi-scale spatial interaction features of auxiliary variables. The convolutional neural network extracts temporal interaction features of the auxiliary variable series. The merging layer adds the temporal and spatial interaction features to obtain spatiotemporal interaction features. The dynamic graph attention, convolutional neural network, and merging layer together constitute the spatiotemporal module. At the same time, the dynamic graph attention of the last layer of the encoder maps the spatiotemporal features extracted by the last spatiotemporal module into a Gaussian latent representation space. The decoder of the dynamic variational autoencoder is established through transposed convolution, which reconstructs the input data and learns a robust Gaussian latent representation space for spatiotemporal features. The regressor learns the dynamic relationship between the extracted spatiotemporal features and the output variables through dynamic linear gate attention and uses a fully connected layer to predict key indicators. The polypropylene melt index prediction model based on the dynamic variational autoencoder regressor proposed in this embodiment consists of four spatiotemporal modules, four deconvolution modules, and four dynamic linear gate attention modules.

[0060] Figure 3 This is a schematic diagram comparing the actual value and predicted value of the polypropylene melt index provided in Example 1 of the present invention. Figure 3 The polypropylene melt index prediction model based on the dynamic variational autoencoder regressor proposed in this embodiment can not only accurately predict key indicators, but also provide a confidence level for each predicted value, providing a reference for related operations. Figure 4a-4b A schematic diagram comparing the input data and reconstructed data provided in Example 1 of the present invention. The distributions of the reconstructed data and the original data are almost indistinguishable, demonstrating that the proposed model can learn robust spatiotemporal features in the Gaussian latent representation space using GL as an optimization term.

[0061] The polypropylene melt index prediction model based on a dynamic variational autoencoder regressor, provided in this embodiment, learns the dynamic relationship between spatiotemporal interaction features and output variables, maps the prediction results to a Gaussian subspace, and then uses the Gaussian loss function as the optimization objective to provide a prediction confidence level for each prediction result. By combining the prediction results and confidence levels of key indicators, online product quality monitoring and control can be achieved.

[0062] This example uses polypropylene melt index soft measurement to verify the effectiveness of the polypropylene melt index prediction model based on dynamic variational autoencoder regressor. The input auxiliary variables of the polypropylene melt index prediction model based on dynamic variational autoencoder regressor proposed in this example are shown in Table 1.

[0063] Table 1 List of auxiliary variables

[0064]

[0065] In order to verify the effectiveness of the polypropylene melt index prediction model based on the dynamic variational autoencoder regressor provided in this embodiment, this embodiment uses five evaluation indicators, namely RMSE, MAE, RSE, MAPE, and CORR, to evaluate the prediction effect of the polypropylene melt index prediction model based on the dynamic variational autoencoder regressor. The comparison results are shown in Table 2.

[0066] Table 2 Comparison results of polypropylene melt index prediction models based on dynamic variational autoencoder regressor

[0067]

[0068] The ELM and MLP provided in this embodiment are traditional shallow neural networks. Compared to the best shallow neural network model, the MLP, the worst deep learning model, the LSTM, achieved a 16% reduction in RMSE. Therefore, deep learning models outperform shallow neural network models in soft sensors. Compared to CNN+LSTM, the RMSE of CIVAE decreased by 6%, demonstrating the superiority of VAE over general deep learning models. Compared to CIVAE, the polypropylene melt index prediction model based on a dynamic variational autoencoder regressor provided in this embodiment achieved state-of-the-art results and demonstrated a 15% reduction in RMSE.

[0069] After deleting different modules of the polypropylene melt index prediction model based on the dynamic variational autoencoder regressor provided in this embodiment, the influence of each module of the proposed model on the prediction results of key indicators was further explored. The results are shown in Table 3.

[0070] Table 3 Ablation experiment results

[0071]

[0072] Table 3 shows that after removing the spatiotemporal module, Gaussian loss, and variational autoencoder, the RMSE of the polypropylene melt index prediction model based on the dynamic variational autoencoder regressor provided in this embodiment increased by 50%, 13%, and 25%, respectively. Therefore, the spatiotemporal module has the greatest impact on the results of the polypropylene melt index prediction model based on the dynamic variational autoencoder regressor. Spatiotemporal features reflect the true characteristics of polypropylene production data and are of great significance for improving the prediction of key indicators.

[0073] This example demonstrates the effectiveness and accuracy of a polypropylene melt index prediction model based on a dynamic variational autoencoder regressor for predicting the melt index during polypropylene production. Rapid and accurate prediction of the melt index during propylene polymerization plays a crucial role in product quality control, brand conversion, new product development, and significantly reducing transition material in polypropylene production facilities.

[0074] It will be understood that the above embodiments are merely exemplary embodiments for illustrating the principles of the present invention, and the present invention is not limited thereto. Those skilled in the art will appreciate that various modifications and improvements can be made without departing from the spirit and substance of the present invention, and such modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A polypropylene melt index prediction model based on dynamic variational autoencoder regressor, characterized in that: include: Get a polypropylene dataset with continuous time Where L is the total length of the sequence and M is the number of auxiliary variables; The polypropylene dataset is divided into datasets according to the preset time window Where N is the number of samples in the dataset and T is the time window; The encoder of the dynamic variational autoencoder uses dynamic graph attention to extract the spatial interaction features of auxiliary variables, as shown below: A t =(U t ⊙A glob )W global +A glob (3) in, is the trend relationship between the i-th process variable and the j-th process variable, is the local interaction matrix at time t, U t is the global trend information control gate, W loc 、W glo, 、 is a learnable parameter, A t is the fused multi-scale interaction matrix, ⊙ is the Hadamard product, and the Pearson correlation coefficient is used to calculate the process variable A glob The trend matrix, is the spatial interaction feature of the auxiliary variables at time t, S represents the number of heads, Represents the interaction matrix A t The element in row i and column j of represents the eigenvector of the jth process variable; The encoder of the dynamic variational autoencoder uses a convolutional neural network to extract the temporal interaction features of the auxiliary variables, which are expressed as follows: in, is the filter, is the time interaction feature of the auxiliary variables, is the characteristic at time t-k+1; The temporal interaction feature and the dynamic spatial interaction feature are concatenated using an addition method, and the decoder of the dynamic variational autoencoder uses deconvolution to reconstruct the input data to learn a robust Gaussian latent feature representation space for the dynamic spatiotemporal interaction feature; The regressor uses dynamic linear gate attention to learn the dynamic relationship between the dynamic spatiotemporal interaction features and the output variable, and outputs the predicted value of the polypropylene melt index through a fully connected layer. The expression is as follows: Attention(Q,K,V)=relu(QK T )V (6) MultiHead(Q,K,V)=Concat(head1,...,head h )W o (7) Among them, Q is the query, K is the key, V is the value, relu is the activation function, Concat is the vector splicing function, W o is a learnable parameter, and head. represents the features extracted in different subspaces.

2. The polypropylene melt index prediction model based on dynamic variational autoencoder regressor according to claim 1, characterized in that: The loss function of the polypropylene melt index prediction model has three loss terms, and the expressions of the loss terms are as follows: in, is the KL divergence between the Gaussian latent feature representation space and the standard normal distribution, is the reconstruction loss of the decoder, is the Gaussian loss function of the regressor, y i is the label, μ en (·) represents the mean of the encoder mapping the auxiliary variables to the Gaussian latent space, σ en (·) represents the variance of the encoder mapping the auxiliary variables to the Gaussian latent space, μ de (·) represents the original data recovered by the decoder, μ reg (·) represents the key indicator predicted by the regressor, σ reg (·) represents the confidence of the regressor prediction.

Images