A method for predicting intrinsic viscosity of polyester fiber during polymerization

Abstract

The present invention relates to a method for predicting the characteristic viscosity of a polyester fiber polymerization process, comprising: obtaining production process data of a polyester fiber polymerization workshop, and preprocessing the production process data. According to a feature selection method and manual experience, variables that have a greater impact on the characteristic viscosity of the polyester fiber polymerization process are selected to form a feature subset. A deep learning model for predicting the characteristic viscosity of the polyester fiber polymerization process is established. With the feature subset as input, the model for predicting the characteristic viscosity of the polyester fiber polymerization process is trained, and the parameters of the neural network model are optimized using the gradient descent method, and the best model for predicting the characteristic viscosity of the polyester fiber polymerization process is obtained through training. The obtained model for predicting the characteristic viscosity of the polyester fiber polymerization process is deployed in a polyester workshop, and predictions are made based on the real-time production data of the workshop. The present application realizes the prediction of the characteristic viscosity of the polyester fiber polymerization process through a deep learning algorithm, so that workshop staff can adjust process parameters in real time.

Description

Technical Field

[0001] The present invention relates to a method for predicting the intrinsic viscosity of a polyester fiber polymerization process, and in particular to a method for predicting the intrinsic viscosity of a polyester fiber polymerization process based on a TCN-GRU of an Attention mechanism, belonging to the technical field of intelligent production prediction of polyester fibers. Background Art

[0002] The textile industry is one of the pillar industries in my country, with early development and large scale. Polyester fiber is one of the important raw materials in the textile industry. Its main production process includes four parts: polyester fiber melt production, transportation, spinning and weaving processing. Among them, the melt polymerization process is the primary link in the production of polyester fiber. The intrinsic viscosity of the polyester fiber polymerization process is an important parameter to measure the quality of polyester fiber, and its stability is an important indicator to measure the production quality.

[0003] In the polyester fiber production process, field engineers generally monitor and adjust individual key indicators based on work experience. This method often relies on production experience and lacks relevant theoretical guidance, which cannot meet the needs of intelligent production. Therefore, there is an urgent need for a prediction method that can predict the intrinsic viscosity of the polyester fiber polymerization process according to the set production parameter requirements without online debugging, which can guide its production. Summary of the invention

[0004] The technical problem to be solved by the present invention is: currently, during the polymerization process of polyester fibers, engineers generally monitor and adjust the characteristic viscosity index based on work experience, lack relevant theoretical guidance, and cannot meet the needs of intelligent production; secondly, the online debugging process may cause a large change in the characteristic viscosity of the polyester fiber melt, which cannot meet the requirements of subsequent production processes, resulting in large economic losses.

[0005] In order to solve the above technical problems, the technical solution of the present invention is to provide a method for predicting the intrinsic viscosity of polyester fiber polymerization process, which is characterized by being a method for predicting the intrinsic viscosity of polyester fiber polymerization process based on a temporal convolutional network TCN-gated neural network GRU of an attention mechanism, comprising the following steps:

[0006] S1 obtains process data of a polyester fiber polymerization production workshop, where the production process data is data of various production process parameters, and each production process parameter is a variable; preprocesses the obtained production process data, and the preprocessing method includes: processing abnormal values ​​of the production process data; processing noise of the production process data; and processing inconsistent dimensions of the production process data;

[0007] S2, based on the production process data obtained in step S1, selects variables having a greater influence on the intrinsic viscosity of the polyester fiber polymerization process from all variables according to a feature selection method and manual experience, all the selected variables constitute a feature subset, and each variable in the feature subset is defined as a feature;

[0008] S3 establishes a deep learning model for predicting the intrinsic viscosity of polyester fiber polymerization process, which includes an input layer, a TCN residual network, a GRU unit layer, an Attention layer, and an output layer;

[0009] The data input through the input layer is first subjected to a one-dimensional dilated convolution operation in the TCN residual network, as shown in the following formula:

[0010]

[0011] In the formula, x s-di is the input sequence, x represents the sth convolution object, di represents the expansion coefficient of the i-th convolution kernel, x-di ensures that only past inputs can be convolved; f(i) represents the filter of the i-th convolution kernel; k represents the size of the convolution kernel;

[0012] After the data is subjected to a one-dimensional dilated convolution operation, the weights are normalized, the ReLU function is used as the activation function, and finally the Dropout operation is used to randomly discard neurons;

[0013] The TCN residual network extracts the temporal features of the data and then inputs them into the GRU unit layer. The GRU unit layer consists of an update gate and a reset gate, as shown in the following formula:

[0014] r t =σ(W r [h t-1 ,u t ])

[0015] z t =σ(W z [h t-1 ,u t ])

[0016]

[0017]

[0018] Where: u t Indicates the input information at the current moment; h t-1 represents the hidden layer state at time t-1; h t represents the hidden layer state at the tth moment; r t Represents the reset gate value, which is used to control the degree of information forgetting at the tth moment; z trepresents the update gate value, which indicates the updated value of the hidden layer information at time t-1 and the input information at time t; represents the alternative state; σ represents the sigmoid activation function; tanh represents the hyperbolic tangent function; W r is to reset the gate weight matrix; W z is the update gate weight matrix; W h is the candidate state weight matrix.

[0019] The input of the Attention layer is the feature vector processed by the GRU unit layer. The Attention layer assigns different weights to the features in the feature vector to highlight more critical influencing factors and help the model improve prediction accuracy.

[0020] The weight coefficient calculation formula of the Attention layer is:

[0021] v t =∑a t h t

[0022] In the formula, v t Represents the output of the Attention layer at time t; e t =utanh(wh t +b), e t It is represented as the feature vector h output by the GRU unit layer at time t t The determined attention probability distribution value; u and w are weight coefficients, and b is the bias coefficient;

[0023] Finally, the output layer calculates the output of the Attention layer through the fully connected layer to obtain the prediction result;

[0024] S4 uses the feature subset as input to train the deep learning model for predicting the characteristic viscosity of the polyester fiber polymerization process established in step S3, optimizes the parameters of the deep learning model using the gradient descent method, and obtains the best prediction model for predicting the characteristic viscosity of the polyester fiber polymerization process through training, wherein the feature data in the feature subset is imported into the input layer of the deep learning model, the input feature is extracted through the TCN residual network for time series features, and a feature vector is output, and then the feature vector output by the TCN residual network is input into the GRU unit layer and the Attention layer, the change law of the polymerization process is learned from the feature vector, the model prediction is realized, and finally the prediction result is obtained through the output layer;

[0025] S5 deployed the optimal prediction model in the polyester workshop and made predictions based on the workshop's real-time production data.

[0026] Preferably, when processing outliers in the production process data, the box plot method is used to calculate the lower quartile Q1, median Q2 and upper quartile Q3 of the feature data, and the quartile distance IQR=Q3-Q1 is calculated; the upper limit value of the current feature is set to Q3+1.5IQR and the lower limit value is set to Q1-1.5IQR, and the feature data exceeding the upper limit value and below the lower limit value are judged as outliers.

[0027] Preferably, in step S1, a sliding average method is used to process the production process data noise.

[0028] Preferably, in step S2, the feature selection method is a feature ranking method based on the maximum information coefficient, wherein the calculation formula of the maximum information coefficient MIC[x:y] between the variable X and the variable Y is as follows:

[0029]

[0030] In the formula, is the mutual information between variables X and Y, p(X,Y) is the joint probability between variables X and Y, P(X) and P(Y) are the distribution probabilities of variables X and Y respectively; |X| and |Y| represent the number of grid divisions of the value ranges of variables X and Y respectively, and B represents the upper limit of the grid division |X|*|Y|.

[0031] Preferably, in step S3, the number of TCN residual network units is 3, and the number of GRU unit layers is 2.

[0032] Preferably, in step S4, the evaluation criteria of the deep learning model are the root mean square error RMSE and the goodness of fit R 2 , as shown below:

[0033]

[0034]

[0035] In the formula, n represents the number of prediction results; represents the intrinsic viscosity value of the i-th polyester fiber predicted by the deep learning model; y i For The corresponding intrinsic viscosity value detected in actual production; Indicates the mean value of intrinsic viscosity.

[0036] The present invention proposes a method for predicting the intrinsic viscosity of the polyester fiber polymerization process based on TCN-GRU of the Attention mechanism. The prediction of the intrinsic viscosity of the polyester fiber polymerization process is realized through a deep learning algorithm, so that workshop staff can adjust the process parameters in real time. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 The process of the intrinsic viscosity prediction method of polyester fiber polymerization process is illustrated;

[0038] Figure 2 and Figure 3 An example of data preprocessing is illustrated;

[0039] Figure 4 The polyester fiber polymerization production process is shown schematically;

[0040] Figure 5 This is the overall structure of the TCN-GRU model based on the Attention mechanism. DETAILED DESCRIPTION

[0041] The present invention is further described below in conjunction with specific examples. It should be understood that these examples are only to help those of ordinary skill in the art understand the principles and knowledge of the present invention, and are not intended to limit the scope of the present invention, and cannot be considered to limit the application scenarios 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 may make various changes or modifications to the present invention, but the deformation, changes and conversions made to the embodiments based on the principles and purposes of the present invention also fall within the scope defined by the claims attached to this application. And it is obvious that this specification is only exemplified by preferred embodiments, and there is no need to exhaust all embodiments.

[0042] This embodiment discloses a method for predicting intrinsic viscosity of a polyester fiber polymerization process, comprising the following steps:

[0043] S1 obtains the production process data of the polyester fiber polymerization workshop in the production data system, with a collection frequency of once every minute, and collects data for one month, totaling 43,200 data records. In the present invention, the production process data refers to the data of each production process parameter, and each production process parameter is a variable.

[0044] The collected production process data is preprocessed. The preprocessing methods mainly include: processing abnormal values ​​of production process data; processing noise of production process data; processing inconsistent dimensions of production process data.

[0045] Processing of abnormal values ​​of production process data: Use box plot method to calculate the lower quartile Q1, median Q2 and upper quartile Q3 of variable data, and calculate the quartile distance IQR = Q3-Q1. Set the upper limit of the current variable to Q3+1.5IQR and the lower limit to Q1-1.5IQR, then the variable data exceeding the upper limit and below the lower limit will be judged as abnormal values, such as Figure 2 shown.

[0046] Processing of production process data noise: Use the sliding average method to process data noise, as shown in Figure 3, expressed as the following formula:

[0047]

[0048] Where M is the smoothing coefficient, x t-n is the noisy data at the tnth moment, and y[t] represents the denoised data at the tth moment.

[0049] Processing of inconsistent data dimensions in the production process: Use maximum and minimum value normalization to process inconsistent data dimensions, as shown in the following formula:

[0050]

[0051] Where: x i and are the i-th variable data before and after normalization, x max With x min are the maximum and minimum values ​​of the i-th variable data before normalization, and the variable data of the original polyester fiber polymerization are quantized to the interval of 0 to 1 by linearization method.

[0052] S2 selects variables that have a greater impact on the characteristic viscosity of the polyester fiber polymerization process from all variables based on the production process data obtained in step S1 according to feature selection methods and manual experience. All selected variables constitute a feature subset, and each variable in the feature subset is defined as a feature.

[0053] In this embodiment, the feature selection method is a feature sorting method based on the maximum information coefficient. Figure 4 As shown in the figure, in this example, 35 basic variables such as temperature, pressure, additive flow rate and reaction liquid level at each stage of the polymerization process are selected based on artificial experience and chemical action mechanism. Then, the feature sorting method based on the maximum information coefficient is used to select the variables with the largest correlation with the target variable among all the variables as features, and all the obtained features constitute the best feature subset.

[0054] The calculation formula of the maximum information coefficient MIC[x:y] between variables X and Y is as follows:

[0055]

[0056] In the formula, is the mutual information between variables X and Y, p(X,Y) is the joint probability between variables X and Y, P(X) and P(Y) are the distribution probabilities of variables X and Y respectively; |X| and |Y| represent the number of grid divisions of the value ranges of variables X and Y respectively, and B represents the upper limit of the grid division |X|*|Y|.

[0057] S3 establishes a deep learning model for predicting the intrinsic viscosity of polyester fiber polymerization process.

[0058] The deep learning model established by the present invention includes an input layer, a TCN residual network, a GRU unit layer, an Attention layer and an output layer, wherein the number of TCN residual network units is 3 and the number of GRU unit layers is 2.

[0059] The feature data in the feature subset is imported into the input layer of the deep learning model. The input features are extracted through the TCN residual network for time series features, and the feature vector is output. The feature vector output by the TCN residual network is then input into the GRU unit layer and the Attention layer, and the change law of the aggregation process is learned from the feature vector to achieve model prediction. Finally, the prediction result is obtained through the output layer.

[0060] The input layer data format is:

[0061] The data input through the input layer is first subjected to a one-dimensional dilated convolution operation in the TCN residual network, as shown in the following formula:

[0062]

[0063] In the formula, x s-di is the input sequence, s represents the sth convolution object, di represents the expansion coefficient of the ith convolution kernel, s-di ensures that only the past input can be convolved; f(i) represents the filter of the ith convolution kernel; k represents the size of the convolution kernel.

[0064] After the data undergoes a one-dimensional dilated convolution operation, the weights are normalized, the ReLU function is used as the activation function, and finally the Dropout operation is used to randomly discard neurons.

[0065] The TCN residual network extracts the temporal features of the data and then inputs them into the GRU unit layer. The GRU unit layer consists of an update gate and a reset gate, as shown in the following formula:

[0066] r t =σ(W r [h t-1 ,u t ])

[0067] z t =σ(W z [h t-1 ,u t ])

[0068]

[0069]

[0070] Where: u t Indicates the input information at the current moment; h t-1 represents the hidden layer state at time t-1; h t represents the hidden layer state at the tth moment; r t Represents the reset gate value, which is used to control the degree of information forgetting at the tth moment; z t represents the update gate value, which indicates the updated value of the hidden layer information at time t-1 and the input information at time t; represents the alternative state; σ represents the sigmoid activation function; tanh represents the hyperbolic tangent function; W r is to reset the gate weight matrix; W z is the update gate weight matrix; W h is the candidate state weight matrix.

[0071] The input of the Attention layer is the feature vector processed by the GRU unit layer. The Attention layer assigns different weights to the features in the feature vector, highlighting more critical influencing factors and helping the model improve prediction accuracy.

[0072] The weight coefficient calculation formula of the Attention layer is:

[0073] v t =∑a t h t

[0074] In the formula, v t Represents the output of the Attention layer at time t; e t =utanh(wh t +b), e t It is represented as the feature vector h output by the GRU unit layer at time t t The determined attention probability distribution value; u and w are weight coefficients, and b is the bias coefficient.

[0075] Finally, the output layer calculates the output of the Attention layer through the fully connected layer to obtain the prediction result.

[0076] S4 uses the feature subset as input to train the deep learning model for predicting the characteristic viscosity of the polyester fiber polymerization process established in step S3, optimizes the parameters of the deep learning model using the gradient descent method, and obtains the best prediction model for predicting the characteristic viscosity of the polyester fiber polymerization process through training.

[0077] When training the deep learning model, the model training batch_size is set to 128, the learning rate is set to 0.01, and the learning rate decay rate is set to 0.01. In the deep learning model: the number of neurons in each layer of the GRU unit layer is set to 256, and the probability of neuron loss is set to 0.6; the convolution kernel size of the TCN residual network is set to 3. The number of iterations of the deep learning model is set to 200 times.

[0078] The evaluation criteria for deep learning models are root mean square error RMSE and goodness of fit R 2 , as shown below:

[0079]

[0080]

[0081] In the formula, n represents the number of prediction results; represents the intrinsic viscosity value of the i-th polyester fiber predicted by the deep learning model; y i For The corresponding intrinsic viscosity value detected in actual production; Indicates the mean value of intrinsic viscosity.

[0082] S5 deployed the best prediction model in the polyester workshop and made predictions based on the real-time production data of the workshop. After deploying the model, the polyester workshop can greatly utilize and integrate the relationship between historical data, fully consider the time series characteristics of the data, and build a reasonable time series model, which can improve the prediction ability of the intrinsic viscosity of the polyester fiber polymerization process.

Claims

1. A method for predicting intrinsic viscosity of polyester fiber polymerization process, characterized in that: A method for predicting the intrinsic viscosity of polyester fiber polymerization process based on a temporal convolutional network TCN-gated neural network GRU with an attention mechanism is provided, comprising the following steps: S1 obtains process data of a polyester fiber polymerization production workshop, where the production process data is data of various production process parameters, and each production process parameter is a variable; preprocesses the obtained production process data, and the preprocessing method includes: processing abnormal values ​​of the production process data; processing noise of the production process data; and processing inconsistent dimensions of the production process data; S2, based on the production process data obtained in step S1, selects variables having a greater influence on the intrinsic viscosity of the polyester fiber polymerization process from all variables according to a feature selection method and manual experience, all the selected variables constitute a feature subset, and each variable in the feature subset is defined as a feature; S3 establishes a deep learning model for predicting the intrinsic viscosity of polyester fiber polymerization process, which includes an input layer, a TCN residual network, a GRU unit layer, an Attention layer, and an output layer; The data input through the input layer is first subjected to a one-dimensional dilated convolution operation in the TCN residual network, as shown in the following formula: In the formula, x s-di is the input sequence, s represents the sth convolution object, di represents the expansion coefficient of the i-th convolution kernel, s-di ensures that only past inputs can be convolved; f(i) represents the filter of the i-th convolution kernel; k represents the size of the convolution kernel; After the data is subjected to a one-dimensional dilated convolution operation, the weights are normalized, the ReLU function is used as the activation function, and finally the Dropout operation is used to randomly discard neurons; The TCN residual network is used to extract the temporal features of the data and then input them into the GRU unit layer. The GRU unit layer consists of an update gate and a reset gate, as shown in the following formula: r t =σ(W r [h t-1 ,u t ]) With t =σ(W z [h t-1 ,at t ]) Where: u t Indicates the input information at the current moment; h t-1 represents the hidden layer state at time t-1; h t represents the hidden layer state at the tth moment; r t Represents the reset gate value, which is used to control the degree of information forgetting at the tth moment; z t represents the update gate value, which indicates the updated value of the hidden layer information at time t-1 and the input information at time t; represents the alternative state; σ represents the sigmoid activation function; tanh represents the hyperbolic tangent function; W r is to reset the gate weight matrix; W z is the update gate weight matrix; W h is the alternative state weight matrix; The input of the Attention layer is the feature vector processed by the GRU unit layer. The Attention layer assigns different weights to the features in the feature vector to highlight more critical influencing factors and help the model improve prediction accuracy. The weight coefficient calculation formula of the Attention layer is: v t =Σa t h t In the formula, v t Represents the output of the Attention layer at time t; e t =utanh(wh t +b), e t It is represented as the feature vector h output by the GRU unit layer at time t t The determined attention probability distribution value; u and w are weight coefficients, and b is the bias coefficient; Finally, the output layer calculates the output of the Attention layer through the fully connected layer to obtain the prediction result; S4 uses the feature subset as input to train the deep learning model for predicting the intrinsic viscosity of the polyester fiber polymerization process established in step S3, optimizes the parameters of the deep learning model using the gradient descent method, and obtains the best prediction model for predicting the intrinsic viscosity of the polyester fiber polymerization process through training, wherein the feature data in the feature subset is imported into the input layer of the deep learning model, the input feature is extracted through the TCN residual network for time series features, and a feature vector is output, and then the feature vector output by the TCN residual network is input into the GRU unit layer and the Attention layer, the change law of the polymerization process is learned from the feature vector, the model prediction is realized, and finally the prediction result is obtained through the output layer; S5 deployed the optimal prediction model in the polyester workshop and made predictions based on the workshop's real-time production data.

2. A method for predicting intrinsic viscosity of polyester fiber polymerization process according to claim 1, characterized in that: When processing outliers in production process data, the box plot method is used to calculate the lower quartile Q1, median Q2 and upper quartile Q3 of the feature data, and the quartile distance IQR = Q3-Q1; the upper limit value of the current feature is set to Q3+1.5IQR and the lower limit value is Q1-1.5IQR, and the feature data exceeding the upper limit value and below the lower limit value are judged as outliers.

3. The method for predicting intrinsic viscosity of polyester fiber polymerization process according to claim 1, characterized in that: In step S1, the sliding average method is used to process the noise of the production process data.

4. The method for predicting intrinsic viscosity of polyester fiber polymerization process according to claim 1, characterized in that: In step S2, the feature selection method is a feature ranking method based on the maximum information coefficient, wherein the calculation formula of the maximum information coefficient MIC[x:y] between the variable X and the variable Y is as follows: In the formula, is the mutual information between variables X and Y, p(X,Y) is the joint probability between variables X and Y, P(X) and P(Y) are the distribution probabilities of variables X and Y respectively; |X| and |Y| represent the number of grid divisions of the value ranges of variables X and Y respectively, and B represents the upper limit of the grid division |X|*|Y|.

5. The method for predicting intrinsic viscosity of polyester fiber polymerization process according to claim 1, characterized in that: In step S3, the number of TCN residual network units is 3 and the number of GRU unit layers is 2.

6. The method for predicting intrinsic viscosity of polyester fiber polymerization process according to claim 1, characterized in that: In step S4, the evaluation criteria of the deep learning model are the root mean square error RMSE and the goodness of fit R 2 , as shown below: In the formula, n represents the number of prediction results; represents the intrinsic viscosity value of the i-th polyester fiber predicted by the deep learning model; y i For The corresponding intrinsic viscosity value detected in actual production; Indicates the mean value of intrinsic viscosity.

Images