Online prediction method for density index of outlet oligomer in polyester fiber esterification process

Abstract

The invention relates to an online prediction method for an outlet oligomer density index in a polyester fiber esterification process, and the method specifically comprises the steps: (1) collecting esterification process variables and historical data of the outlet oligomer density index from a sensor in the polyester fiber esterification process, constructing an offline esterification data set, and dividing the offline esterification data set into a training set and a test set; (2) establishing an MC-RSGN model, wherein the MC-RSGN model is composed of a dynamic feature extraction module, a deep noise reduction module and a key index regression module; (3) setting a hyper-parameter initial value of the model, training the MC-RSGN model by adopting the training set, optimizing parameter setting, and verifying the model by using the test set; and (4) inputting real-time production process data collected on site into the model after offline verification for prediction, and outputting a real-time esterification stage outlet oligomer density index prediction value. According to the method, the density index of the outlet oligomer in the polyester fiber esterification process can be quickly, online and accurately predicted.

Description

Technical Field

[0001] The invention belongs to the technical field of intelligent production of chemical fibers, and relates to an online prediction method for an outlet oligomer density index in a polyester fiber esterification process. Background Art

[0002] The polymerization process is the primary link in the research of the entire process of polyester fiber production, and the esterification process is the first step in the polymerization process and also the most important step. The oligomer density of its output product directly affects the subsequent series of chemical reactions and plays an extremely critical role in the quality and performance of the final product fiber. However, due to the complex process mechanism and variable external environmental factors, the system exhibits very complex nonlinear behavior, and it is difficult to adapt to current production needs using simple neural network modeling and prediction. In this regard, the use of deep learning algorithms to model and optimize the polyester fiber esterification process has attracted widespread attention from scholars.

[0003] In today's industrial modeling methods for polyester fiber production processes, due to the high dimensionality, high redundancy, and high noise characteristics of polyester fiber production data, extracting effective features from the collected process data has become the most critical step. In the prior art, in order to reduce the complexity of the model, the extraction of effective features is generally considered from one of the perspectives of capturing data dynamic characteristics or removing noise data.

[0004] In terms of capturing the dynamic characteristics of data, some scholars have constructed dynamic feature extractors from the perspectives of sampling rate, slow feature analysis, and online tracking of fast time-varying processes. However, due to the difficulty of processing high-dimensional data, most existing dynamic feature extraction methods focus on the extraction of global information, while ignoring the hidden relationships that may exist between variables, that is, the network pays more attention to the integrity and coverage of the data, while ignoring the various local correlations of different combinations of variables. Secondly, from the perspective of eliminating redundant features, for filtering and processing noise and useless information, scholars often use complex noise processing networks, such as time-frequency conversion, integrated spatiotemporal noise adaptation, Huber-m estimation, and collaborative noise allocation methods. However, for continuous industrial processes such as the esterification stage of polyester fiber, the dynamic characteristics of data and noise are often closely related. Due to the characteristics of the continuity of the production line, noise will continue to accumulate and transmit over time. When the noise reaches a certain level, it will break the balance between effective information and noise in the data, interfere with the dynamic characteristics of the system, and make the estimation of dynamic characteristics more difficult. At the same time, the dynamic characteristics of the system will also affect the transmission and response of noise in the system. Different system dynamic characteristics will determine the diffusion and attenuation law of noise in the system, thus affecting data collection and processing. Therefore, for the polyester fiber esterification process, it is not comprehensive to consider it from any single perspective.

[0005] In recent years, only a few scholars have considered the above two perspectives at the same time in the research of industrial process systems similar to the polyester fiber esterification process. For example, Reference 1 (Semisupervised dynamic soft sensor based on complementary ensemble empirical mode decomposition and deep learning [J]. Measurement, 2021, 183: 109788.) can effectively remove noise and redundancy while capturing the dynamic characteristics of the data, thereby improving the prediction accuracy of the soft sensor; Reference 2 (Robust low-rank clustering contrastive learning integrating transformer for noisy industrial soft sensors [J]. IEEE Transactions on Instrumentation and Measurement, 72 (2023): 1-10.) proposed a novel robust low-rank clustering contrastive learning (LrCCL) integrated Transformer (LrCCL-T) method to solve the problem of strong process noise and interference in industrial soft sensor modeling. However, due to the complexity of modeling, scholars only considered the extraction of dynamic features from a global perspective, and did not pay attention to the hidden features that may exist locally. At the same time, in terms of noise processing, they only chose simple noise processing methods such as contrastive learning and mode decomposition. However, these methods cannot meet the requirements of accuracy and robustness in modeling the polyester fiber esterification process system in terms of data feature capture and noise processing capabilities.

[0006] Therefore, for the online prediction task of the export oligomer density index of the polyester fiber esterification process with high noise and dynamic characteristics, it is of great significance to study a prediction model that can capture the dynamic characteristics of hidden data, filter noise signals and have lower complexity. Summary of the invention

[0007] The purpose of the present invention is to solve the problems existing in the prior art and to provide an online prediction method for the density index of oligomers at the outlet of the polyester fiber esterification process.

[0008] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0009] An online prediction method for the oligomer density index at the outlet of a polyester fiber esterification process comprises the following steps:

[0010] (1) The historical data of esterification process variables and export oligomer density indicators were collected from sensors in the polyester fiber esterification process, and an offline esterification dataset was constructed and divided into a training set and a test set.

[0011] (2) Establishing the MC-RSGN model; the MC-RSGN model consists of three parts: a dynamic feature extraction module, a deep noise reduction module, and a key indicator regression module;

[0012] The dynamic feature extraction module is used to take the one-dimensional time series data initially collected at a certain moment as input, and expand it into a three-dimensional tensor through data reconstruction and time-delay displacement technology, so as to focus on the local correlation between different variables, break the randomness and order dependence problems brought by single-order input, consider the hidden information and internal connections between multiple variables, and finally use convolution operations to extract features from the three-dimensional tensor data to generate a three-dimensional output matrix;

[0013] The deep denoising module is composed of m stacked residual shrinkage modules. Each residual shrinkage module introduces a soft threshold function in the residual connection to filter out noise-related features. The deep denoising module is used to take the 3D output matrix generated by the dynamic feature extraction module as input X and perform feature mapping to obtain the feature map |X|. A different threshold τ is assigned to each channel in the third dimension of the feature map |X| through the soft threshold function. Then, a residual connection is used to enable the network to learn the difference between the denoising target and the current output. The features related to noise are gradually reduced through the stacked residual shrinkage modules to obtain the final three-dimensional sequence feature Y.

[0014] The key indicator regression module uses a GRU network to predict the final target output. A fully connected layer is connected after the GRU network to output the final predicted value of the MC-RSGN model. The GRU network captures the long-term dependency of the time dimension in the three-dimensional sequence feature Y to achieve selective retention and update of information, and passes its output data through the fully connected layer to obtain the final predicted value of the key indicator.

[0015] (3) Setting the initial values ​​of the model’s hyperparameters, using the training set to train the MC-RSGN model to optimize the parameter settings, and using the test set to perform offline verification on the trained MC-RSGN model;

[0016] (4) The real-time data of the esterification process variables collected by the on-site sensors are input into the MC-RSGN model after offline verification for prediction, and the real-time predicted value of the oligomer density index at the outlet of the esterification stage is output.

[0017] As the preferred technical solution:

[0018] The invention discloses an online prediction method for an outlet oligomer density index of a polyester fiber esterification process. The esterification process variables collected in step (1) include 13 sensor data, namely: esterification kettle pressure, esterification kettle liquid level, esterification kettle agitator current signal, esterification kettle gas phase pipeline temperature, esterification kettle temperature, esterification separation tower top pressure, esterification separation tower middle pressure, esterification separation tower top temperature, esterification separation tower temperature, process tower liquid level adjustment, esterification separation water reflux flow rate, water reflux pump frequency, and separation tower return kettle EG and inlet kettle PTA / EG slurry flow rate; the outlet oligomer density index is: 2P231 outlet oligomer density data.

[0019] In the above-mentioned online prediction method for the oligomer density index at the outlet of the polyester fiber esterification process, the sample ratio of the training set and the test set in step (1) is 8:2.

[0020] In the above-mentioned online prediction method for the oligomer density index at the outlet of the polyester fiber esterification process, the construction steps of the dynamic feature extraction module in step (2) are as follows:

[0021] (2.1) Continuously collect data from n industrial process sensors at the same start time and fixed time interval, and form one-dimensional sample vectors of n process variables. Each vector represents the true value of these n process variables at a certain sampling time t, which is recorded as:

[0022]

[0023] Among them, r1=(1,2,…,n), which indicates the arrangement order of n sensors, and each number represents a sensor number;

[0024] (2.2) Using the time-delay shift technique, the sample vector data within s time steps before each sampling moment is used to dynamically expand the one-dimensional sample of the process variable at the sampling moment into a two-dimensional matrix, that is, each dynamic matrix contains all the data information of the n process variables at the sampling moment t and s time steps before t, and the two-dimensional dynamic matrix is ​​used as the initial matrix R1 at the sampling moment t, expressed as:

[0025]

[0026] (2.3) The data arrangement of the initial matrix is ​​shuffled multiple times according to certain rules, and each newly formed matrix is ​​retained as an additional matrix of the initial matrix; the shuffling of the data arrangement of the initial matrix is ​​performed according to the dimension of the process variables, that is, changing the topological order of the process variables on the variable axis in the two-dimensional matrix; for the (k-1)th order shuffling, the kth column data in the initial matrix is ​​exchanged with the 1st column data, and the corresponding arrangement order of the n sensors is r k =(k,2,…,1,…,n), and the new two-dimensional matrix is ​​denoted as Rk , expressed as:

[0027]

[0028] When k is the same as n, it means that each process variable has undergone a scrambling operation. In this case, the order of the n sensors is r. k =(k,2,3,…,k-1,1);

[0029] (2.4) The initial matrix and its corresponding additional matrix are superimposed channel by channel to form a three-dimensional tensor as the input at time t, and a convolution operation is performed to construct a dynamic feature extraction module. The output feature of the dynamic feature extraction module is X = [:,:,k].

[0030] In the above-mentioned online prediction method for the density index of oligomers at the outlet of the polyester fiber esterification process, the soft threshold function introduced in the residual connection in step (2) is as follows:

[0031]

[0032] Among them, τ is the channel threshold, the features within the threshold interval [-τ, τ] are set to 0, and the features outside the threshold interval [-τ, τ] gradually shrink to 0; by using a soft threshold, the network can represent weakly correlated features as attributes close to zero and retain useful negative features.

[0033] In the above-mentioned online prediction method for the oligomer density index at the outlet of the polyester fiber esterification process, the threshold τ in step (2) is a parameter that can be automatically obtained by the network through learning and training, and the soft threshold allows each channel to have an independent threshold. The threshold calculation method for each channel is as follows:

[0034] First, feature mapping is performed on the input feature X of the deep denoising module to obtain the feature map |X|, and global average pooling is used to generate a one-dimensional vector of the feature map |X|;

[0035] Secondly, the one-dimensional vector of the obtained feature map |X| is input into the two-layer network to determine the contraction parameter of each channel; the internal structure of the two-layer network is arranged in order: the first fully connected layer, batch normalization, RELU activation function, the second fully connected layer and the sigmoid function. The number of neurons in the second fully connected layer of the two-layer network corresponds to the number of channels of the input feature map |X|; after the second fully connected layer in the two-layer network, the sigmoid function is applied to scale the parameters in the range of (0,1), and the formula is as follows:

[0036]

[0037] In the formula, α iThe i-th element of the scaling parameter α corresponds to the scaling parameter of the i-th channel of the feature map |X|, z i is the i-th element output by the second fully connected layer in the two-layer network;

[0038] Next, calculate the average value of the one-dimensional vector of the feature map |X|, as follows:

[0039]

[0040] Among them, |X i | is the feature of the i-th channel of the feature map |X|, w, h, and c are the width, height, and number of channels of the feature map |X|, respectively;

[0041] Finally, the threshold calculation formula for each channel is as follows:

[0042] τ i =α i ×average|X w,h,c | w,h ;

[0043] In the formula, τ i is the threshold of the i-th channel of the feature map |X|.

[0044] In the above-mentioned online prediction method for the oligomer density index at the outlet of the polyester fiber esterification process, in step (2), each residual shrinkage module calculates the threshold value τ of each channel. i The features are reduced and deleted and input into the residual connection to reduce the noise-related features. The output of the previous residual shrinkage module is used as the input of the next residual shrinkage module. By stacking the residual shrinkage modules, the noise signal is deeply filtered, and finally the output three-dimensional sequence feature Y of the deep denoising module is obtained.

[0045] In the above-mentioned online prediction method for the density index of oligomers at the outlet of the polyester fiber esterification process, the calculation formula of the key index regression module in step (2) is as follows:

[0046] R t =δ(Y t W xr +H t-1 W hr +b r );

[0047] Z t =δ(Y t W xz +H t-1 W hz +b z );

[0048]

[0049] Among them, Y t represents the input features at the current time t, R t represents the reset gate at time t, Z t represents the update gate at time t, represents the candidate state at time t, H t represents the hidden state at time t, H t-1 Represents the hidden state of the previous moment, that is, the output of the hidden layer at the previous moment; W xr , W hr , W xz , W hz , W xh and W hh is the weight matrix, b r , b z and b h is the bias vector; δ represents the sigmoid function.

[0050] In the above-mentioned online prediction method for the outlet oligomer density index of the polyester fiber esterification process, in step (3), the parameters that can be optimized include: the time step s used for the historical information data, the number of times the initial two-dimensional matrix is ​​shuffled k, the number of stacked residual shrinkage modules m, the learning rate, batch size and number of iterations of the model training, and the initial values ​​of these parameters are set according to the empirical parameter adjustment method.

[0051] In the above-mentioned online prediction method for the outlet oligomer density index of the polyester fiber esterification process, in step (3), MAE is used as the evaluation index of model training, and the training set is input into the MC-RSGN model for training. When MAE is less than 0.02, the training is stopped, and the test set is input into the trained model for verification; otherwise, the model parameter settings are optimized and adjusted, and the training is restarted.

[0052] The principle of the present invention is as follows:

[0053] Considering the complex mechanism of polyester fiber esterification process and the high redundancy and high noise characteristics of the data, the present invention processes the data features from two aspects: dynamic feature extraction and noise signal removal, and builds a targeted prediction model. The model mines the hidden information and dynamic features between data from the perspective of local correlation, and uses soft threshold residual connection to filter noise signals, thereby enhancing the feature representation ability of the model.

[0054] The present invention is based on the GRU network, and through its flexible gating mechanism, it effectively captures the long-term dependencies in the time series data, realizes the selective retention and update of information, and enables the network to adapt to the time series pattern of the polyester fiber esterification process data. On this basis, two functional modules of dynamic feature extraction and deep noise reduction are added. The dynamic feature extraction module focuses on the local correlation between different variables through time lag displacement technology and data reconstruction, breaks the problems of randomness and sequential dependence brought by single sequential input, considers the hidden information and intrinsic connections between multiple variables, and captures the dynamic characteristics between data; the deep noise reduction module performs deep filtering on high-noise data, introduces soft thresholds into the residual connection, enables the network to learn the difference between the denoising target and the current output, and gradually reduces the noise-related features by stacking the residual shrinkage module, achieving a balance between noise and data features. Finally, the GRU network and the fully connected layer are connected to achieve accurate prediction of the export oligomer density index of the polyester fiber esterification process.

[0055] Beneficial effects:

[0056] (1) The online prediction method for the oligomer density index at the outlet of the polyester fiber esterification process of the present invention is based on the two perspectives of dynamic feature extraction and noise signal removal, and aims to construct a prediction model for the oligomer density index at the outlet of the polyester fiber esterification process with high redundancy and high noise characteristics. By fully mining the hidden correlations in complex nonlinear data and filtering invalid features, the problem of inaccurate modeling of complex processes caused by considering feature extraction from a single perspective is solved.

[0057] (2) The online prediction method for the oligomer density index at the outlet of the polyester fiber esterification process of the present invention is improved based on the GRU network model, making the model more suitable for the prediction of the polyester fiber esterification process, further improving the prediction accuracy of the algorithm on the oligomer density index at the outlet of the polyester fiber esterification process, and realizing real-time online prediction, which has a certain guiding role in actual production. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a flow chart of an online prediction method for the density index of the outlet oligomers in the polyester fiber esterification process;

[0059] Figure 2 It is a schematic diagram of the construction process of the three-dimensional tensor in the dynamic feature extraction module;

[0060] Figure 3 This is the internal structure diagram of the deep noise reduction module;

[0061] Figure 4 This is the structural diagram of the MC-RSGN model;

[0062] Figure 5It is a schematic diagram of the system structure and sensor distribution of the esterification process;

[0063] Figure 6 This is a simulation result diagram of the MC-RSGN model of the present invention on the polyester fiber esterification process data set. DETAILED DESCRIPTION

[0064] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the appended claims of the application equally.

[0065] An online prediction method for the oligomer density index at the outlet of polyester fiber esterification process, such as Figure 1 As shown, the following steps are included:

[0066] (1) The historical data of esterification process variables and export oligomer density indicators were collected from sensors in the polyester fiber esterification process, and an offline esterification dataset was constructed and divided into a training set and a test set.

[0067] The collected esterification process variables include 13 sensor data, namely: esterification kettle pressure, esterification kettle liquid level, esterification kettle agitator current signal, esterification kettle gas phase pipeline temperature, esterification kettle temperature, esterification separation tower top pressure, esterification separation tower middle pressure, esterification separation tower top temperature, esterification separation tower temperature, process tower liquid level adjustment, esterification separation water reflux flow, water reflux pump frequency and separation tower return kettle EG and inlet kettle PTA / EG slurry flow; outlet oligomer density index is: 2P231 outlet oligomer density data, polyester fiber esterification process variables and outlet oligomer density index such as Figure 5 As shown in Table 1:

[0068] Table 1 Description of polyester fiber esterification process variables and export oligomer density indicators

[0069]

[0070]

[0071] The data for the implementation of the present invention are collected from an actual polyester esterification production line of a domestic polyester factory. The factory process adopts the DuPont three-reactor polymerization process, and the density of the output oligomers in the esterification stage of the polymerization process is used as the key indicator for prediction. The data includes 13 process variables related to it, which are collected by DCS (such as Figure 5As shown in Figure 2, the total number of sensors n used is 14, and the sensor data of two consecutive days on the DCS is used. The sampling interval is 1 minute, with a total of 2880 samples. The 2880 samples are divided into training sets and test sets according to the ratio of 8:2, of which 2304 data points are used as the training set and 576 data points are used as the test set.

[0072] (2) Figure 4 As shown, a MC-RSGN model is established; the MC-RSGN model consists of three parts: a dynamic feature extraction module, a deep noise reduction module, and a key indicator regression module;

[0073] The dynamic feature extraction module is used to take the one-dimensional time series data initially collected at a certain moment as input, and expand it into a three-dimensional tensor through data reconstruction and time-delay displacement technology (the construction process of the three-dimensional tensor is as follows Figure 2 As shown in Figure 2), the convolution operation is used to extract features from the 3D tensor data and generate a 3D output matrix. The construction steps of the dynamic feature extraction module are as follows:

[0074] (2.1) Continuously collect data from n industrial process sensors at the same start time and fixed time interval, and form one-dimensional sample vectors of n process variables. Each vector represents the true value of these n process variables at a certain sampling time t, which is recorded as:

[0075]

[0076] Among them, r1=(1,2,…,n), which indicates the arrangement order of n sensors, and each number represents a sensor number;

[0077] (2.2) Using the time-delay shift technique, the sample vector data within s time steps before each sampling moment is used to dynamically expand the one-dimensional sample of the process variable at the sampling moment into a two-dimensional matrix, that is, each dynamic matrix contains all the data information of the n process variables at the sampling moment t and s time steps before t, and the two-dimensional dynamic matrix is ​​used as the initial matrix R1 at the sampling moment t, expressed as:

[0078]

[0079] (2.3) The data arrangement of the initial matrix is ​​shuffled multiple times according to certain rules, and each newly formed matrix is ​​retained as an additional matrix of the initial matrix; the shuffling of the data arrangement of the initial matrix is ​​performed according to the dimension of the process variables, that is, changing the topological order of the process variables on the variable axis in the two-dimensional matrix; for the (k-1)th order shuffling, the kth column data in the initial matrix is ​​exchanged with the 1st column data, and the corresponding arrangement order of the n sensors is r k =(k,2,…,1,…,n), and the new two-dimensional matrix is ​​denoted as Rk , expressed as:

[0080]

[0081] When k is the same as n, it means that each process variable has undergone a scrambling operation. In this case, the order of the n sensors is r. k =(k,2,3,…,k-1,1);

[0082] (2.4) The initial matrix and its corresponding additional matrix are superimposed channel by channel to form a three-dimensional tensor as the input at time t, and a convolution operation is performed to construct a dynamic feature extraction module. The output feature of the dynamic feature extraction module is X = [:,:,k].

[0083] The internal structure of the deep noise reduction module is as follows Figure 3 As shown in FIG. 1 , the deep denoising module is composed of m stacked residual shrinkage modules. Each residual shrinkage module introduces a soft threshold function in the residual connection to filter out noise-related features. The deep denoising module is used to take the 3D output matrix generated by the dynamic feature extraction module as input X and perform feature mapping to obtain a feature map |X|. A different threshold τ is assigned to each channel in the third dimension of the feature map |X| through a soft threshold function, and the threshold τ of each channel is calculated. i Each residual shrinkage module is based on the threshold τ of each channel calculated i The features are reduced and deleted, and input into the residual connection to reduce the noise-related features, so that the network learns the difference between the denoising target and the current output, and uses the output of the previous residual shrinkage module as the input of the next residual shrinkage module, and through the stacked residual shrinkage modules, the noise-related features are gradually reduced to obtain the three-dimensional sequence feature Y;

[0084] The soft threshold function introduced in the residual connection is as follows:

[0085]

[0086] Among them, τ is the channel threshold, the features within the threshold interval [-τ,τ] are set to 0, and the features outside the threshold interval [-τ,τ] gradually shrink to 0;

[0087] The threshold τ is a parameter that can be automatically obtained by the network through learning and training, and the soft threshold allows each channel to have an independent threshold. The threshold calculation method for each channel is as follows:

[0088] First, feature mapping is performed on the input feature X of the deep denoising module to obtain the feature map |X|, and global average pooling is used to generate a one-dimensional vector of the feature map |X|;

[0089] Secondly, the one-dimensional vector of the obtained feature map |X| is input into the two-layer network to determine the contraction parameter of each channel; the internal structure of the two-layer network is arranged in order: the first fully connected layer, batch normalization, RELU activation function, the second fully connected layer and the sigmoid function. The number of neurons in the second fully connected layer of the two-layer network corresponds to the number of channels of the input feature map |X|; after the second fully connected layer in the two-layer network, the sigmoid function is applied to scale the parameters in the range of (0,1), and the formula is as follows:

[0090]

[0091] In the formula, α i The i-th element of the scaling parameter α corresponds to the scaling parameter of the i-th channel of the feature map |X|, z i is the i-th element output by the second fully connected layer in the two-layer network;

[0092] Next, calculate the average value of the one-dimensional vector of the feature map |X|, as follows:

[0093]

[0094] Among them, |X i | is the feature of the i-th channel of the feature map |X|, w, h, and c are the width, height, and number of channels of the feature map |X|, respectively;

[0095] Finally, the threshold calculation formula for each channel is as follows:

[0096] τ i =α i ×average|X w,h,c | w,h ;

[0097] In the formula, τ i is the threshold of the i-th channel of the feature map |X|.

[0098] The key indicator regression module uses a GRU network to predict the final target output. A fully connected layer is connected after the GRU network to output the final predicted value of the MC-RSGN model. The GRU network captures the long-term dependency of the time dimension in the three-dimensional sequence feature Y to achieve selective retention and update of information, and passes its output data through the fully connected layer to obtain the final predicted value of the key indicator.

[0099] The calculation formula of the key indicator regression module is as follows:

[0100] R t =δ(Y t W xr +H t-1 Whr +b r );

[0101] Z t =δ(Y t W xz +H t-1 W hz +b z );

[0102]

[0103] Among them, Y t represents the input features at the current time t, R t represents the reset gate at time t, Z t represents the update gate at time t, represents the candidate state at time t, H t represents the hidden state at time t, H t-1 Indicates the hidden state at the previous moment; W xr , W hr , W xz , W hz , W xh and W hh is the weight matrix, b r , b z and b h is the bias vector; δ represents the sigmoid function.

[0104] (3) Set the initial values ​​of the model's hyperparameters, use the training set to train the MC-RSGN model to optimize the parameter settings, and use the test set to verify the trained MC-RSGN model offline; use MAE as the evaluation indicator for model training, input the training set into the MC-RSGN model for training, stop training when MAE is less than 0.02, and input the test set into the trained model for verification; otherwise, optimize the model parameter settings and restart training;

[0105] The parameters that can be optimized include: the time step s used for historical information data, the number of times the initial two-dimensional matrix is ​​shuffled k, the number of stacked residual shrinkage modules m, the learning rate, batch size and number of iterations of model training, and the initial values ​​of these parameters are set according to the empirical parameter adjustment method as follows: s = 2, k = 5, m = 4, the learning rate is 0.01, the batch size is 64, and the number of iterations is 50;

[0106] The values ​​of the optimizable parameters were changed in turn, and the experimental results were compared. Multiple experiments were conducted through the control variable method, and the final hyperparameter settings of the model were determined as follows: the time step s used for the historical information data was 4, the number of times the initial two-dimensional matrix was shuffled k was 8, and the number of stacked residual shrinkage modules m was 4. In addition to the hyperparameters of the network structure, during the network training process, the initial learning rate was set to 0.01, the batch size was 128, and all samples were traversed in each epoch. The epoch size was set to 70 times by adjustment, and the Adam optimizer was used. The experimental results are shown in Figure 2. Figure 6 As shown in the figure, the image shows the degree of fit between the predicted value of the MC-RSGN model and the true value, which intuitively reflects the prediction effect of the model. The higher the overlap between the two curves, the better the model accuracy, which also means the better performance in the prediction task. Figure 6 It can be seen that, whether looking at the overall situation or focusing on the details at the inflection point, the MC-RSGN model can well capture the trend of dynamic changes in the data, has a certain anti-interference ability for noise and outliers in the data, and has a good advantage in processing noisy data.

[0107] (4) The real-time data of the esterification process variables collected by the on-site sensors are input into the MC-RSGN model after offline verification for prediction, and the real-time predicted value of the oligomer density index at the outlet of the esterification stage is output.

[0108] In order to verify the effectiveness of the online prediction method for the outlet oligomer density index of the polyester fiber esterification process proposed in the present invention, a comparative experiment was also carried out in this section. In the comparative experiment, all the parameter settings of the above experiment remained unchanged, and ResNet was selected as the basic comparison algorithm. The algorithm also has a residual network structure and has obvious advantages in processing industrial data containing high noise. Two algorithms were used for 10 independent experiments, and RMSE and MAE were used as evaluation indicators of the prediction results. The maximum, minimum and average values ​​of the indicators were compared respectively. The experimental results and comparisons are shown in Table 2.

[0109] Table 2 Statistical results of MC-RSGN model and comparison algorithms

[0110]

[0111] The data in Table 2 reflect the prediction accuracy of the two models. The three indicators of maximum, minimum and average can comprehensively show the randomness and statistical results of the model in each independent experiment. The experiment shows that the MC-RSGN proposed in the present invention has the best performance, whether from the results of a single experiment or the statistical results of multiple experiments. In the prediction task of the density of oligomers at the outlet of the polyester fiber esterification process, the accuracy is better than the traditional ResNet model, and it has better prediction results.

Claims

1. An online prediction method for the oligomer density index at the outlet of polyester fiber esterification process, characterized in that The steps include: (1) The historical data of esterification process variables and export oligomer density indicators were collected from sensors in the polyester fiber esterification process, and an offline esterification dataset was constructed and divided into a training set and a test set. (2) Establishing the MC-RSGN model; the MC-RSGN model consists of three parts: a dynamic feature extraction module, a deep noise reduction module, and a key indicator regression module; The dynamic feature extraction module is used to take the one-dimensional time series data initially collected at a certain moment as input, expand it into a three-dimensional tensor through data reconstruction and time-delay displacement technology, and use convolution operation to extract features from the three-dimensional tensor data to generate a three-dimensional output matrix; The deep denoising module consists of m stacked residual shrinkage modules, each of which introduces a soft threshold function in the residual connection to filter out noise-related features; The deep denoising module is used to take the 3D output matrix generated by the dynamic feature extraction module as input X and perform feature mapping to obtain the feature map |X|. A different threshold τ is assigned to each channel in the third dimension of the feature map |X| through a soft threshold function. Then, a residual connection is used to enable the network to learn the difference between the denoising target and the current output. The features related to noise are gradually reduced through stacked residual shrinkage modules, and finally the three-dimensional sequence feature Y is obtained. The key indicator regression module uses a GRU network to predict the final target output. A fully connected layer is connected after the GRU network to output the final predicted value of the MC-RSGN model. The GRU network captures the long-term dependency of the time dimension in the three-dimensional sequence feature Y to achieve selective retention and update of information, and passes its output data through the fully connected layer to obtain the final predicted value of the key indicator. (3) Setting the initial values ​​of the model’s hyperparameters, training the MC-RSGN model using the training set to optimize the parameter settings, and using the test set to perform offline verification on the trained MC-RSGN model; (4) The real-time data of the esterification process variables collected by the on-site sensors are input into the MC-RSGN model after offline verification for prediction, and the real-time predicted value of the outlet oligomer density index of the esterification stage is output.

2. The method for online prediction of oligomer density index at the outlet of polyester fiber esterification process according to claim 1, characterized in that: The esterification process variables collected in step (1) include 13 sensor data, namely: esterification kettle pressure, esterification kettle liquid level, esterification kettle agitator current signal, esterification kettle gas phase pipeline temperature, esterification kettle temperature, esterification separation tower top pressure, esterification separation tower middle pressure, esterification separation tower top temperature, esterification separation tower temperature, process tower liquid level adjustment, esterification separation water reflux flow, water reflux pump frequency and separation tower return kettle EG and inlet kettle PTA / EG slurry flow; outlet oligomer density index is: 2P231 outlet oligomer density data.

3. The method for online prediction of oligomer density index at the outlet of polyester fiber esterification process according to claim 1, characterized in that: The sample ratio of the training set and the test set in step (1) is 8:

2.

4. The method for online prediction of oligomer density index at the outlet of polyester fiber esterification process according to claim 1, characterized in that: The construction steps of the dynamic feature extraction module in step (2) are as follows: (2.1) Continuously collect data from n industrial process sensors at the same start time and fixed time interval, and form one-dimensional sample vectors of n process variables. Each vector represents the true value of these n process variables at a certain sampling time t, which is recorded as: Among them, r1=(1,2,…,n), which indicates the arrangement order of n sensors, and each number represents a sensor number; (2.2) Using the time-delay shift technique, the sample vector data within s time steps before each sampling moment is used to dynamically expand the one-dimensional sample of the process variable at the sampling moment into a two-dimensional matrix, that is, each dynamic matrix contains all the data information of the n process variables at the sampling moment t and s time steps before t, and the two-dimensional dynamic matrix is ​​used as the initial matrix R1 at the sampling moment t, expressed as: (2.3) The data arrangement of the initial matrix is ​​shuffled multiple times according to certain rules, and each newly formed matrix is ​​retained as an additional matrix of the initial matrix; the shuffling of the data arrangement of the initial matrix is ​​performed according to the dimension of the process variables, that is, changing the topological order of the process variables on the variable axis in the two-dimensional matrix; for the (k-1)th order shuffling, the kth column data in the initial matrix is ​​exchanged with the 1st column data, and the corresponding arrangement order of the n sensors is r k =(k,2,…,1,…,n), and the new two-dimensional matrix is ​​denoted as R k , expressed as: When k is the same as n, it means that each process variable has undergone a scrambling operation. In this case, the order of the n sensors is r. k =(k,2,3,…,k-1,1); (2.4) The initial matrix and its corresponding additional matrix are superimposed channel by channel to form a three-dimensional tensor as the input at time t, and a convolution operation is performed to construct a dynamic feature extraction module. The output feature of the dynamic feature extraction module is X = [:,:,k].

5. The method for online prediction of the oligomer density index at the outlet of the polyester fiber esterification process according to claim 4, characterized in that: The soft threshold function introduced in the residual connection in step (2) is as follows: Among them, τ is the channel threshold, the features within the threshold interval [-τ,τ] are set to 0, and the features outside the threshold interval [-τ,τ] gradually shrink to 0.

6. The method for online prediction of oligomer density index at the outlet of polyester fiber esterification process according to claim 5, characterized in that: The threshold τ in step (2) is a parameter that can be automatically obtained by the network through learning and training, and the soft threshold allows each channel to have an independent threshold. The threshold calculation method for each channel is as follows: First, feature mapping is performed on the input feature X of the deep denoising module to obtain the feature map |X|, and global average pooling is used to generate a one-dimensional vector of the feature map |X|; Secondly, the one-dimensional vector of the obtained feature map |X| is input into the two-layer network to determine the shrinkage parameter of each channel; the number of neurons in the second fully connected layer of the two-layer network corresponds to the number of channels of the input feature map |X|; after the second fully connected layer in the two-layer network, the sigmoid function is applied to scale the parameters in the range of (0,1), and the formula is as follows: In the formula, α i The i-th element of the scaling parameter α corresponds to the scaling parameter of the i-th channel of the feature map |X|, z i is the i-th element output by the second fully connected layer in the two-layer network; Next, calculate the average value of the one-dimensional vector of the feature map |X|, as follows: Among them, |X i | is the feature of the i-th channel of the feature map |X|, w, h, and c are the width, height, and number of channels of the feature map |X|, respectively; Finally, the threshold calculation formula for each channel is as follows: t i =a i ×average|X w,h,c | w,h ; In the formula, τ i is the threshold of the i-th channel of the feature map |X|.

7. The method for online prediction of oligomer density index at the outlet of polyester fiber esterification process according to claim 6, characterized in that: In step (2), each residual shrinkage module calculates the threshold τ of each channel i The features are reduced and deleted and input into the residual connection to reduce the noise-related features. The output of the previous residual shrinkage module is used as the input of the next residual shrinkage module. By stacking the residual shrinkage modules, the noise signal is deeply filtered, and finally the output three-dimensional sequence feature Y of the deep denoising module is obtained.

8. The method for online prediction of oligomer density index at the outlet of polyester fiber esterification process according to claim 1, characterized in that: The calculation formula of the key indicator regression module in step (2) is as follows: R t =δ(Y t W xr +H t-1 W hr +b r ); Z t =δ(Y t W xz +H t-1 W hz +b z ); Among them, Y t represents the input features at the current time t, R t represents the reset gate at time t, Z t represents the update gate at time t, represents the candidate state at time t, H t represents the hidden state at time t, H t-1 Indicates the hidden state at the previous moment; W xr , W hr , W xz , W hz , W xh and W hh is the weight matrix, b r , b z and b h is the bias vector; δ represents the sigmoid function.

9. The method for online prediction of oligomer density index at the outlet of polyester fiber esterification process according to claim 1, characterized in that: In step (3), the parameters that can be optimized include: the time step s used for the historical information data, the number of times the initial two-dimensional matrix is ​​shuffled k, the number of stacked residual shrinkage modules m, the learning rate, batch size and number of iterations for model training, and the initial values ​​of these parameters are set according to the empirical parameter adjustment method.

10. The method for online prediction of oligomer density index at the outlet of polyester fiber esterification process according to claim 1, characterized in that: In step (3), MAE is used as the evaluation index of model training, and the training set is input into the MC-RSGN model for training. When MAE is less than 0.02, the training is stopped, and the test set is input into the trained model for verification; Otherwise, optimize the model parameter settings and restart training.

Images