Parameter identification method for the process of homogenization and viscosity-increasing reactor of recycled polyester

Abstract

The present invention relates to the field of data processing and identification, and specifically to a parameter identification method for a process of a homogenizing and viscosity-increasing reactor for recycled polyester. The method comprises obtaining a set of historical process parameters of the reactor after each reaction is completed and the corresponding product effects; performing permutations and combinations on the parameters to obtain multiple subspaces, and obtaining the optimal subspace corresponding to each standard model; calculating the noise compliance coefficient corresponding to the optimal subspace; calculating the attention weight of each optimal subspace to obtain the comprehensive attention weight of each parameter; constructing a set of historical process parameters for training, using the set of historical process parameters for training as data for a neural network, using the product effects as the output of the neural network to train the neural network, and using the trained neural network to identify the process parameters of the reactor. The present invention divides the original set of historical process parameters by subspaces and obtains the attention weight of each parameter, thereby training the neural network and achieving accurate and adaptable parameter identification.

Description

Technical Field

[0001] The invention relates to the field of data processing and identification, and in particular to a parameter identification method for a process of a homogenization and viscosity-increasing reactor for recycled polyester. Background Art

[0002] Polyester regeneration is the main method for processing and recycling waste PET and other materials. The process itself is a homogenization and viscosity-increasing process. The process reaction process controls the reactor temperature, pressure, viscosity-increasing agent addition, reaction time and other parameters during the process to ensure the product's characteristics and effects. The existing process generally relies on manual experience control and has the problem of relying on the operator's experience. Once the operator changes, the product's characteristics and effects will become unstable, affecting the product quality.

[0003] To address these issues, a method for identifying reactor process parameters is needed. This method can identify errors in reactor process parameter settings, providing early warnings of these errors and providing data support for subsequent automated process parameter control. Existing methods for parameter identification typically use models to fit historical reactor operating data. This method identifies parameters and predicts the performance of reactor product characteristics. While this approach offers good accuracy, it relies on the accuracy of historical data collection. Various factors can influence the actual collection of reactor process parameter data and product characteristics, leading to data deviations. If these deviations are used to fit the model, the model's predictions will be compromised, defeating the purpose of parameter identification.

[0004] Therefore, the present invention provides a parameter identification method for the process of recycled polyester homogenization and viscosity increasing reactor. By broadening the number of parameters and increasing the data dimension to enhance the fitting accuracy, the subspace data characteristics are used to identify and screen the data dimensions of different process parameters, avoiding the dimensionality disaster caused by excessively high data dimensions, making the parameter identification process more scalable and accurate, and adapting to the changes in the number of process parameters brought about by process updates. Summary of the Invention

[0005] The present invention provides a parameter identification method for a process of a homogenizing and tackifying reactor for recycled polyester to solve existing problems. The method comprises the following steps: obtaining a historical process parameter set of the reactor after each reaction is completed and a corresponding product effect; arranging and combining the parameters to obtain a plurality of subspaces, and obtaining an optimal subspace corresponding to each standard model; calculating a noise compliance coefficient corresponding to the optimal subspace; calculating an attention weight of each optimal subspace to obtain a comprehensive attention weight of each parameter; constructing a training historical process parameter set, using the training historical process parameter set as data of a neural network, using the product effect as output of the neural network to train the neural network, and using the trained neural network to identify the process parameters of the reactor.

[0006] According to the technical means proposed in the present invention, while broadening the number of parameters and increasing the data dimension to enhance the fitting accuracy, the subspace data characteristics are used to identify and screen the data dimensions of different process parameters, and the attention weights of each data dimension of the original historical process parameter set are set through the subspace division effect of the original historical process parameter set. While improving the accuracy of the model, the impact of the dimensionality disaster of high-dimensional data is reduced, so that the abnormal and normal states of the process parameters of the reactor can be identified more accurately, and the characteristic effects of the product can be predicted, thereby realizing accurate and adaptable parameter identification.

[0007] The present invention adopts the following technical solution, a parameter identification method for the process of homogenizing and tackifying a recycled polyester reactor, comprising:

[0008] Obtain the historical process parameter set of the reactor and the corresponding product effect after each reaction is completed.

[0009] The parameters in each historical process parameter set are arranged and combined to obtain multiple subspaces, and all subspaces of each historical process parameter set are input into the supervised neural network to obtain the optimal subspace in each historical process parameter set.

[0010] The information entropy of the parameters in each historical process parameter set is obtained, and the noise conformity coefficient of the corresponding optimal subspace is calculated according to the information entropy of the parameters in the optimal subspace of each historical process parameter set and the information entropy of the parameters in all non-optimal subspaces of the parameter set.

[0011] The attention weight of each optimal subspace is calculated according to the noise conformity coefficient of each optimal subspace, and the attention weight of the optimal subspace is used as the attention weight of each parameter in the optimal subspace. The comprehensive attention weight of each parameter is obtained according to the sum of the attention weights of each parameter in all optimal subspaces in each historical process parameter set.

[0012] The comprehensive attention weight of each parameter is used to correct the historical process parameters of the reactor after each reaction is completed, and the corrected process parameters are input into the neural network as the training historical process parameter set. The product effect corresponding to the historical process parameters of the reactor after each reaction is completed is used as the output of the neural network to train the neural network, and the trained neural network is used to identify the reactor process parameters.

[0013] Furthermore, a parameter identification method for the process of homogenizing and viscosity-increasing reactor of recycled polyester is provided, wherein a method for constructing multiple subspaces using different numbers of parameters in the historical process parameter set is as follows:

[0014] The number of parameters in the historical process parameter set is K, the maximum number of parameters in the subspace is set to B, the parameters in the historical process parameter set are permuted and combined, and the parameters after permutation and combination are projected into the basis vectors of the subspace. The number of subspaces obtained is

[0015] Furthermore, a parameter identification method for the process of homogenizing and tackifying a recycled polyester reactor is provided. The supervised neural network uses a constructed loss function. The loss function is constructed by fitting the subspace of each standard model, and the expression is:

[0016]

[0017] in, represents the model coefficient of the αth standard model, D τ represents the subspace with τ number of basis vectors, (Y i ) τ It represents the parameters in the original historical process parameter set after being projected into the subspace with the number of basis vectors τ. represents the fitting result of the αth standard model to the subspace with τ basis vectors, D τ T represents the transposed matrix of the subspace with τ as the number of basis vectors, I represents the identity matrix, represents the product effect corresponding to the i-th historical process parameter set, N represents the total number of N historical process parameter sets, K is the number of parameters in the historical process parameter set, B is the maximum number of parameters in the subspace, represents the number of subspaces in the i-th historical process parameter set, loss α represents the neural network fitted using the αth standard model.

[0018] Furthermore, a parameter identification method for a recycled polyester homogenization and viscosity-increasing reactor process, wherein the supervised neural network also includes:

[0019] When the loss function value of the supervised neural network is minimized, the result of the first loss term in the loss function is used as the fitting degree of each optimal subspace. The expression of the first loss term is:

[0020]

[0021] in, represents the model coefficient of the αth standard model, (Y i ) τ It represents the parameters in the original historical process parameter set after being projected into the subspace with the number of basis vectors τ. represents the fitting result of the αth standard model to the subspace with τ basis vectors, represents the product effect corresponding to the i-th historical process parameter set, and N represents a total of N historical process parameter sets.

[0022] Furthermore, a parameter identification method for the process of homogenization and viscosity-increasing reactor of recycled polyester is proposed. The expression for calculating the noise compliance coefficient of the optimal subspace is:

[0023]

[0024] Among them, Z α represents the noise compliance coefficient of the optimal subspace corresponding to the αth standard model, Q αBest represents the information entropy of the parameters in the optimal subspace corresponding to the αth standard model, Q ξ Represents the information entropy of parameters in all non-optimal subspaces corresponding to the αth standard model.

[0025] Furthermore, a parameter identification method for the process of homogenization and viscosity-increasing reactor of recycled polyester is proposed, and the method for calculating the attention weight of each optimal subspace is as follows:

[0026] The sum of the partition eigenvalues ​​of all subspaces is obtained according to the noise conformity coefficient and fitting degree of all optimal subspaces. The attention weight of each optimal subspace is calculated according to the sum of the partition eigenvalues. The expression for calculating the sum of the partition eigenvalues ​​is:

[0027]

[0028] The expression for calculating the attention weight of the optimal subspace is:

[0029]

[0030] Among them, W α represents the attention weight of the optimal subspace corresponding to the αth standard model, Z α represents the noise compliance coefficient of the optimal subspace corresponding to the αth standard model, H αIt represents the degree of fit of the optimal subspace corresponding to the αth standard model, A is the number of standard models, and T is the sum of the partition eigenvalues ​​of all optimal subspaces.

[0031] Furthermore, a parameter identification method for the process of homogenizing and tackifying the recycled polyester reactor is provided, and the method for obtaining the attention weight of each parameter is as follows:

[0032] The attention weight of each optimal subspace is used as the attention weight of each parameter in the optimal subspace, and the sum of the attention weights of each parameter in each historical process parameter set in the optimal subspace is calculated to obtain the comprehensive attention weight of each parameter in the historical process parameter set.

[0033] Furthermore, a parameter identification method for a recycled polyester homogenization and viscosity-increasing reactor process and a method for training a neural network are as follows:

[0034] Construct a training historical process parameter set X′ based on each parameter in the original historical process parameter set and its corresponding comprehensive attention weight i =[w1c1,w2c2,…,w K c K ], the training historical process parameter set is used as the input of the neural network, the product effect of the original historical process parameter set is used as the output of the neural network to train the neural network, and the reactor process parameters are identified based on the trained FC neural network;

[0035] The loss function of the trained neural network is:

[0036]

[0037] in, is the product effect of the i-th historical process parameter set output by the trained neural network, represents the original product effect corresponding to the original i-th historical process parameter set, and N represents the number of historical process parameter sets.

[0038] The beneficial effects of the present invention are: according to the technical means proposed in the present invention, while broadening the number of parameters and increasing the data dimension to enhance the fitting accuracy, the subspace data characteristics are used to identify and screen the data dimensions of different process parameters, and the attention weights of each data dimension of the original historical process parameter set are set through the subspace division effect of the original historical process parameter set, which improves the accuracy of the model while reducing the impact of the dimensionality disaster of high-dimensional data, so that the abnormal state and normal state of the process parameters of the reactor can be identified more accurately, and the characteristic effects of the product can be predicted, thereby realizing accurate and adaptable parameter identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0040] Figure 1 This is a structural schematic diagram of a parameter identification method for a recycled polyester homogenization and viscosity-increasing reactor process according to an embodiment of the present invention. DETAILED DESCRIPTION

[0041] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0042] like Figure 1 As shown, a schematic diagram of a parameter identification method for a homogenization and viscosity-increasing reactor process of recycled polyester according to an embodiment of the present invention is provided, including:

[0043] 101. Obtain the historical process parameter set of the reactor and the corresponding product effect after each reaction is completed.

[0044] The present invention utilizes sensors set on the reactor to obtain historical process parameter data of the reactor, including data such as temperature, pressure, viscosity enhancer dosage, reaction duration, and product characteristic effects.

[0045] The product characteristic effect data is used as the label dimension, and each other process parameter data is used as the data dimension to construct a multidimensional historical process parameter set.

[0046] The reactor process parameter data obtained by the sensor is expressed as X i , which is a K-dimensional vector, namely:

[0047] X i =[c1,c2,…,c K ]

[0048] Among them, each dimension c represents the value of a parameter.

[0049] The product feature effect is a separate dimension as a label dimension, that is, each X i Each corresponds to a label value Constructing a high-dimensional historical process parameter set Where i = 1, 2, ..., N, N is the number of high-dimensional data points.

[0050] 102. The parameters in each historical process parameter set are arranged and combined to obtain multiple subspaces, and all subspaces of each historical process parameter set are input into a supervised neural network to obtain the optimal subspace in each historical process parameter set.

[0051] For high-dimensional historical process parameter sets, there are data of multiple dimensions. These data of different dimensions have different degrees of influence on the establishment of parameter identification models, that is, data of some dimensions are important, while some have little influence on the parameter identification model. High-dimensional data is not necessarily evenly distributed in all dimensions in high-dimensional space, that is, there are situations where data distribution forms are different in different subspaces. When identifying the importance of data of various dimensions, whether the distribution of all data in multiple subspaces conforms to the standard parameter model is an important basis. This is because if the distribution of data can be well fitted by a standard parameter model, it means that the data is not noise, but has a certain distribution law.

[0052] The standard model includes (polynomial models, multidimensional mixed Gaussian models, and other parameter models that can generally be used to fit high-dimensional data.) The better the data in the subspace of the original historical process parameter set can be fitted by the standard model, the more important the dimension of the original historical process parameter set divided by the subspace is. Therefore, a self-supervised neural network is used to divide the optimal subspace based on the effect of the standard model fitting the data in the subspace. The process is as follows:

[0053] Set the standard model number hyperparameter A, with M α represents a standard model (α=1,2,…,A).

[0054] Set the subspace dimension limit hyperparameter B (B≤K), then the data in each subspace are the projections of the original historical process parameter set on each subspace, and the number of subspaces is at most species, then the basis vectors of a subspace can be D τ =[v1,v2,…,v B ] T Indicates that the data in this subspace can be obtained from the original data X i Projected onto the basis vector, (Y i ) τ express, The product effect remains unchanged. That is, every time there is a data (Y i ) τ Each of them will correspond to a label In this way, a subspace historical process parameter set is constructed.

[0055] This subspace historical process parameter set is used as training data to input into the self-supervised neural network, and the self-supervised neural network outputs a D Best (A set of basis vectors representing the subspace) Each standard model corresponds to a self-supervised network.

[0056] The supervised neural network uses a constructed loss function, which is constructed by fitting the subspace of each standard model, and the expression is:

[0057]

[0058] in, represents the model coefficient of the αth standard model, D τ represents the subspace with τ number of basis vectors, (Y i ) τ It represents the parameters in the original historical process parameter set after being projected into the subspace with the number of basis vectors τ. represents the fitting result of the αth standard model to the subspace with τ basis vectors, D τ T represents the transposed matrix of the subspace with τ as the number of basis vectors, I represents the identity matrix, represents the product effect corresponding to the i-th historical process parameter set, N represents the total number of N historical process parameter sets, K is the number of parameters in the historical process parameter set, B is the maximum number of parameters in the subspace, represents the number of subspaces in the i-th historical process parameter set, loss α represents the neural network fitted using the αth standard model.

[0059] The method of constructing multiple subspaces using different numbers of parameters in the historical process parameter set is as follows:

[0060] The number of parameters in the historical process parameter set is K, the maximum number of parameters in the subspace is set to B, the parameters in the historical process parameter set are permuted and combined, and the parameters after permutation and combination are projected into the basis vectors of the subspace. The number of subspaces obtained is

[0061] The method of using supervised neural networks to obtain the optimal subspace corresponding to each standard model is:

[0062] Each standard model corresponds to a supervised neural network, and the subspace corresponding to the minimum loss function of each supervised neural network is taken as the optimal subspace of the standard model.

[0063] For the standard model M α , and its corresponding self-supervised neural network will output a loss αThe smallest corresponding subspace D αBest . This subspace is M α The optimal subspace of .

[0064] 103. Obtain the information entropy of the parameters in each historical process parameter set, and calculate the noise conformity coefficient of the corresponding optimal subspace based on the information entropy of the parameters in the optimal subspace of each historical process parameter set and the information entropy of the parameters in all non-optimal subspaces of the parameter set.

[0065] For a standard model corresponding to the optimal subspace D αBest The more the information entropy of the data of the remaining data dimensions projected onto the original historical process parameter set outside the optimal subspace conforms to the distribution of noise, the better the dimensional division effect of the optimal subspace on the original historical process parameter set is, and the more the standard model corresponding to the optimal subspace can fit the characteristics of the original historical process parameter set. Noise is relative, that is, the less the projected data within the optimal subspace conforms to the characteristics of noise, the more it can explain the aforementioned characteristics.

[0066] Based on the above logic, the relative noise compliance coefficient Z of the optimal subspace corresponding to a standard model is as follows α The calculation of the noise compliance coefficient of the optimal subspace is:

[0067]

[0068] Among them, Z α represents the noise compliance coefficient of the optimal subspace corresponding to the αth standard model, Q αBest represents the information entropy of the parameters in the optimal subspace corresponding to the αth standard model, Q ξ Represents the information entropy of parameters in all non-optimal subspaces corresponding to the αth standard model.

[0069] Z α It is a normalized data. The closer it is to 0, the worse the optimal subspace is in dividing the dimensions of the original historical process parameter set. Conversely, the closer it is to 1, the better.

[0070] The method to obtain the degree of fit of each optimal subspace is:

[0071] For a standard model M α , which corresponds to an optimal subspace D αBest , then when the original data is projected into the optimal subspace, the standard model M αThe standard fitting degree of the data in the optimal subspace is calculated to measure the gap between the standard model fitting result and the label. When the loss function value of the supervised neural network is minimized, the result of the first loss term in the loss function is used as the fitting degree of each optimal subspace. The expression of the first loss term is:

[0072]

[0073] in, represents the model coefficient of the αth standard model, (Y i ) τ It represents the parameters in the original historical process parameter set after being projected into the subspace with the number of basis vectors τ. represents the fitting result of the αth standard model to the subspace with τ basis vectors, represents the product effect corresponding to the i-th historical process parameter set, and N represents a total of N historical process parameter sets.

[0074] 104. The attention weight of each optimal subspace is calculated according to the noise conformity coefficient of each optimal subspace, and the attention weight of the optimal subspace is used as the attention weight of each parameter in the optimal subspace. The comprehensive attention weight of each parameter is obtained according to the sum of the attention weights of each parameter in all optimal subspaces in each historical process parameter set.

[0075] The method for calculating the attention weight of each optimal subspace is:

[0076] The sum of the partition eigenvalues ​​of all subspaces is obtained according to the noise conformity coefficient and the degree of fit of all optimal subspaces. The attention weight of each optimal subspace is calculated according to the sum of the partition eigenvalues. The expression for calculating the sum of the partition eigenvalues ​​is:

[0077]

[0078] Where T is the sum of the eigenvalues ​​of all optimal subspace partitions, which is an intermediate representation and only serves for weight calculation.

[0079] The expression for calculating the attention weight of the optimal subspace is:

[0080]

[0081] Among them, W α represents the attention weight of the optimal subspace corresponding to the αth standard model, Z α represents the noise compliance coefficient of the optimal subspace corresponding to the αth standard model, H αIt represents the degree of fit of the optimal subspace corresponding to the αth standard model, A is the number of standard models, and T is the sum of the partition eigenvalues ​​of all optimal subspaces.

[0082] Each optimal subspace is essentially a set of dimensions of the original historical process parameter set, so W α As the weight of each dimension in each optimal subspace, that is, each dimension of the original historical process parameter set can be determined by the attention weight of the optimal subspace in which it is located.

[0083] It is worth noting that an optimal subspace can correspond to B dimensions of the original historical process parameter sets, and each dimension of the original historical process parameter set can belong to at most A optimal subspaces at the same time. The attention weight of each dimension of the original historical process parameter set can be expressed as the sum of the attention weights of the optimal subspaces to which it belongs at the same time, denoted as w k (K=1,2,…,K), if a dimension of the original historical process parameter set does not belong to any optimal subspace, its weight is 0, so w k It is a number between 0 and 1. The larger it is, the more attention should be paid to this data dimension.

[0084] The method to obtain the comprehensive attention weight of each parameter is:

[0085] The attention weight of each optimal subspace is used as the attention weight of each parameter in the optimal subspace, and the sum of the attention weights of each parameter in the historical process parameter set in the optimal subspace is calculated to obtain the comprehensive attention weight of each parameter in the historical process parameter set.

[0086] So far, for each data dimension on the original historical process parameter set, there is a corresponding comprehensive attention weight w k (K = 1, 2, ..., K) Its value range is 0-1. The larger the value, the more important the dimension. The data dimensions of the original historical process parameter set represent various parameters of the reactor process. This completes the identification of the importance of the reactor process parameters.

[0087] 105. The historical process parameters of the reactor after each reaction are corrected using the comprehensive attention weight of each parameter. The corrected process parameters are input into the neural network as the training historical process parameter set. The product effects corresponding to the historical process parameters of the reactor after each reaction are used as the output of the neural network to train the neural network. The trained neural network is used to identify the reactor process parameters.

[0088] The purpose of the parameter identification is to use the model and historical data to predict the system output results. For the reactor process parameters, due to the large number of parameters, it is more complicated to establish a mathematical model, and it requires a mixed construction of multiple models. The model has too many parameters and is difficult to fit. Therefore, it is necessary to use an artificial neural network as a model for fitting the reactor process parameters to improve the fitting ability of the reactor process parameter data. However, the artificial neural network is easily disturbed by noise data and erroneous data. Therefore, in the above steps, the dimensional importance of the reactor process parameter data is identified and the attention weight is set at the same time. Based on this logic, the reactor process status is identified.

[0089] According to the comprehensive attention weight w of each data dimension of the original historical process parameter set k The training process of the artificial neural network model is designed as follows:

[0090] For the original process parameter historical process parameter set and the label historical process parameter set, according to the comprehensive attention weight w of each dimension of the original historical process parameter set obtained k , using this comprehensive attention weight to construct a new set of historical process parameters, namely X′ i =[w1c1,w2c2,…,w K c K ] Each X′ i The corresponding label is still X′ i (i=1,2,…,N), As training data, build the FC network.

[0091] The method for training the FC neural network is:

[0092] Construct a training historical process parameter set X′ based on each parameter in the original historical process parameter set and its corresponding attention weight i =[w1c1,w2c2,…,w K c K ], the training history process parameter set is used as the input of the FC neural network, the product effect is used as the output of the neural network to train the neural network, and the reactor process parameters are identified based on the trained FC neural network;

[0093] The loss function of the trained FC neural network is:

[0094]

[0095] in, is the product effect of the i-th historical process parameter set output by the trained neural network, represents the original product effect corresponding to the original i-th historical process parameter set, and N represents the number of historical process parameter sets.

[0096] After the network is trained, it is programmed into a processor to construct an electronic device for processing. The electronic device includes a sensor, a processor, and an output display device.

[0097] This electronic device is used to identify the parameter status (label value) of the reactor. The process parameters of the reactor can be collected by using sensors or manually input into the processor of the electronic device. The trained artificial neural network programmed in the processor can output a label of the process parameter status of the reactor, thereby realizing the monitoring and parameter identification of the reactor process, and further providing a basis for the automatic control of the process parameters of the reactor.

[0098] According to the technical means proposed in the present invention, while broadening the number of parameters and increasing the data dimension to enhance the fitting accuracy, the subspace data characteristics are used to identify and screen the different process parameter data dimensions, and the attention weights of each data dimension of the original historical process parameter set are set by the subspace division effect of the original historical process parameter set. While improving the accuracy of the model, the impact of the dimensionality disaster of high-dimensional data is reduced, so that the abnormal and normal states of the process parameters of the reactor can be more accurately identified, and the characteristic effects of the product can be predicted, thereby achieving accurate and adaptable parameter identification.

[0099] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A parameter identification method for a recycled polyester homogenization and viscosity-increasing reactor process, characterized in that: include: Obtain the historical process parameter set of the reactor and the corresponding product effect after each reaction is completed; The parameters in each historical process parameter set are arranged and combined to obtain multiple subspaces, and all subspaces of each historical process parameter set are input into a supervised neural network to obtain the optimal subspace in each historical process parameter set; Obtaining the information entropy of the parameters in each historical process parameter set, and calculating the noise coincidence coefficient of the corresponding optimal subspace based on the information entropy of the parameters in the optimal subspace of each historical process parameter set and the information entropy of the parameters in all non-optimal subspaces of the parameter set; Calculate the attention weight of each optimal subspace according to the noise coincidence coefficient of each optimal subspace, use the attention weight of the optimal subspace as the attention weight of each parameter in the optimal subspace, and obtain the comprehensive attention weight of each parameter according to the sum of the attention weights of each parameter in all optimal subspaces in each historical process parameter set; The comprehensive attention weight of each parameter is used to correct the historical process parameters of the reactor after each reaction is completed. The corrected process parameters are input into the neural network as the training historical process parameter set. The product effect corresponding to the historical process parameters of the reactor after each reaction is completed is used as the output of the neural network to train the neural network, and the trained neural network is used to identify the reactor process parameters. The expression for calculating the noise compliance coefficient of the optimal subspace is: ； in, Indicates the The noise compliance coefficient of the optimal subspace corresponding to the standard model is, Indicates the The information entropy of the parameters in the optimal subspace corresponding to the standard model, Indicates the The information entropy of parameters in all non-optimal subspaces corresponding to the standard model; The method for calculating the attention weight of each optimal subspace is: The sum of the partition eigenvalues ​​of all subspaces is obtained according to the noise conformity coefficient and fitting degree of all optimal subspaces. The attention weight of each optimal subspace is calculated according to the sum of the partition eigenvalues. The expression for calculating the sum of the partition eigenvalues ​​is: ； The expression for calculating the attention weight of the optimal subspace is: ； in, Indicates the The attention weight of the optimal subspace corresponding to the standard model, Indicates the The noise compliance coefficient of the optimal subspace corresponding to the standard model is, Indicates the The degree of fitting of the optimal subspace corresponding to the standard models, A is the number of standard models, is the sum of the partition eigenvalues ​​of all optimal subspaces.

2. The parameter identification method for a homogenization and viscosity-increasing reactor process of recycled polyester according to claim 1, characterized in that: The method of constructing multiple subspaces using different numbers of parameters in the historical process parameter set is as follows: The number of parameters in the historical process parameter set is K, the maximum number of parameters in the subspace is set to B, the parameters in the historical process parameter set are permuted and combined, and the parameters after permutation and combination are projected into the basis vectors of the subspace. The number of subspaces obtained is .

3. The parameter identification method for a homogenization and viscosity-increasing reactor process of recycled polyester according to claim 1, characterized in that: The supervised neural network uses a constructed loss function, which is constructed by fitting the subspace of each standard model, and the expression is: ； in, Indicates the The model coefficients of the standard model, The number of basis vectors is The subspace of Indicates the number of parameters projected onto basis vectors in the original historical process parameter set is The parameters after the subspace, Indicates the The number of basis vectors for a standard model is The fitting results of the subspace of The number of basis vectors is The transposed matrix of the subspace of represents the identity matrix, represents the product effect corresponding to the i-th historical process parameter set, N represents the total number of N historical process parameter sets, K is the number of parameters in the historical process parameter set, B is the maximum number of parameters in the subspace, represents the number of subspaces in the i-th historical process parameter set, Indicates the use of A neural network that fits a standard model.

4. The parameter identification method for a homogenization and viscosity-increasing reactor process of recycled polyester according to claim 3, characterized in that: The supervised neural network also includes: When the loss function value of the supervised neural network is minimized, the result of the first loss term in the loss function is used as the fitting degree of each optimal subspace. The expression of the first loss term is: ； in, Indicates the The model coefficients of the standard model, Indicates the number of parameters projected onto basis vectors in the original historical process parameter set is The parameters after the subspace, Indicates the The number of basis vectors for a standard model is The fitting results of the subspace of represents the product effect corresponding to the i-th historical process parameter set, and N represents a total of N historical process parameter sets.

5. The parameter identification method for a homogenization and viscosity-increasing reactor process of recycled polyester according to claim 1, characterized in that: The method to obtain the attention weight of each parameter is: The attention weight of each optimal subspace is used as the attention weight of each parameter in the optimal subspace, and the sum of the attention weights of each parameter in each historical process parameter set in the optimal subspace is calculated to obtain the comprehensive attention weight of each parameter in the historical process parameter set.

6. The parameter identification method for a homogenization and viscosity-increasing reactor process of recycled polyester according to claim 1, characterized in that: The method for training a neural network is: Construct a training historical process parameter set based on each parameter in the original historical process parameter set and its corresponding comprehensive attention weight ,in, is the comprehensive attention weight of the first parameter in the i-th historical process parameter set; is the first parameter in the i-th historical process parameter set; is the comprehensive attention weight of the second parameter in the i-th historical process parameter set; is the second parameter in the i-th historical process parameter set; is the comprehensive attention weight of the Kth parameter in the i-th historical process parameter set; is the Kth parameter in the i-th historical process parameter set; The training history process parameter set is used as the input of the neural network, and the product effect of the original history process parameter set is used as the output of the neural network to train the neural network, and the reactor process parameters are identified based on the trained FC neural network; The loss function of the trained FC neural network is: ； in, is the product effect of the i-th historical process parameter set output by the trained neural network, represents the original product effect corresponding to the original i-th historical process parameter set, and N represents the number of historical process parameter sets.

Images