CN111738482A - A kind of adjustment method of process parameter in polyester fiber polymerization process - Google Patents

Abstract

The invention relates to a method for adjusting process parameters in a polyester fiber polymerization process, which comprises the steps of obtaining a predicted value of a polyester fiber performance index, comparing the predicted value with a set value, and adjusting the process parameters in the polyester fiber polymerization process according to a comparison result; the process of obtaining the predicted value is as follows: firstly, collecting m characteristic data from all the characteristic data collected by a sensor; then, performing feature relevance sequencing on the m feature data based on a grid search extreme gradient lifting tree algorithm; selecting the first n characteristic data with larger characteristic correlation from the sorted m characteristic data; then, all samples of the n characteristic data are subjected to normalization processing; and finally, processing the normalized n characteristic data by using a weighted double-flow bidirectional long-time and short-time memory attention network, and respectively outputting the predicted values of the 4 polyester fiber performance indexes. The method can obtain the predicted value of the performance index of the polyester fiber with high precision, and further can better guide industrial production.

Description

A kind of adjustment method of process parameter in polyester fiber polymerization process

technical field

The invention belongs to the technical field of chemical fiber production technology, and relates to a method for adjusting process parameters in a polyester fiber polymerization process.

Background technique

The polymerization process is the primary link in the research of the whole process of polyester fiber production. The process includes three stages of esterification, pre-polycondensation and final polycondensation. The polymerization process mainly includes four performance indicators, namely, the average molecular weight _Mn , the degree of polymerization _Pn , the intrinsic viscosity MIV and the esterification rate _Es . For example, the average molecular weight _Mn of the polymer determines the strength and impact resistance of the product. However, there are three performance indicators in actual industrial production that cannot be measured in real time by sensors, so the production process of polyester fiber cannot be directly controlled by controlling the changes of these four performance indicators in real time to ensure product quality. Specifically, the intrinsic viscosity MIV can be measured in real time by the sensor, while the remaining esterification rate _Es , the degree of polymerization P _n and the average molecular weight _Mn cannot be measured in real time by the sensor.

In the prior art, the three performance indicators of the esterification rate, the degree of polymerization and the average molecular weight of the polyester fiber during the polymerization process cannot be directly measured by sensors.

In the actual production process of polyester fiber polymerization, the operator always adjusts the process parameters according to the production experience. This cannot precisely adjust the process parameters that need to be adjusted, thereby improving the quality of polyester fiber products.

Therefore, in the prior art, the adjustment method of the process parameters in the polyester fiber polymerization process has problems such as imprecision and untimely that need to be solved urgently.

SUMMARY OF THE INVENTION

The invention aims to solve the problems existing in the prior art, and provides a method for adjusting process parameters in the polymerization process of polyester fibers. The object of the present invention is to provide a method for adjusting the process parameters according to the predicted results of the intrinsic viscosity MIV, the esterification rate _Es , the degree of polymerization P _n and the average molecular weight _Mn so as to obtain the ideal polymer in the polyester fiber polymerization process. Adjustment method of process parameters.

The present invention converts the value of intrinsic viscosity MIV to obtain three performance index values of esterification rate E _s , degree of polymerization P _n and average molecular weight _Mn , and the specific conversion equation is as follows:

P _n =187.95×(MIV ^{1.46-0.001718} );

K=2.1×10 ^-4 , α=0.82;

Among them, OHV represents the concentration of terminal hydroxyl groups, OHV% represents the percentage of terminal hydroxyl groups in the whole solution, AV represents the concentration of terminal carboxyl groups, and α and K are both constants.

That is, the present invention only needs to measure the value of the intrinsic viscosity MIV in real time through the sensor, and according to the current production process parameters, establish a grid search-based limit gradient boosting tree algorithm and a weighted dual-stream bidirectional long-term memory attention network. The prediction model of the current polymerization production line, and then in the actual polymerization production, the input process parameters are predicted in real time and compared with the set ideal values, and the appropriate process parameters are adjusted accordingly.

For achieving the above object, the technical scheme adopted in the present invention is as follows:

A method for adjusting process parameters in the polyester fiber polymerization process, after obtaining the predicted value of the polyester fiber performance index, comparing it with the set value, and adjusting the process parameters in the polyester fiber polymerization process according to the comparison result;

The performance indicators of polyester fiber are intrinsic viscosity MIV, esterification rate E _s , degree of polymerization P _n and average molecular weight _Mn ;

The process of obtaining the predicted value is as follows:

First, collect m feature data from all feature data collected by the sensor;

Then the limit gradient boosting tree algorithm based on grid search performs feature correlation sorting on the m feature data, that is, the value of the score function is calculated for each feature data. The higher the value of the score function, the greater the feature correlation; select the first n feature data with relatively large feature correlation, n∈m;

Then, normalize all the samples of the n feature data;

Finally, the weighted dual-stream bidirectional long-term and short-term memory attention network is used to process the normalized n feature data, and output the predicted values of the four polyester fiber performance indicators;

The weighted dual-stream bidirectional long-term and short-term memory attention network is a network in which the information obtained from the two tributaries is longitudinally spliced by the merging layer and input to the fully-connected layer through the attention layer;

The two tributaries are the first tributary and the second tributary in parallel; the first tributary consists of two identical weighted bidirectional long-short-term memory network units I, and a batch normalization layer is inserted between the two; the The second branch consists of two identical weighted bidirectional long-short-term memory network units II with a batch normalization layer inserted between them;

Described weighted two-way long-short-term memory network unit 1 is a two-way long-short-term memory network weighted to current moment information, is a two-way long-short-term memory network weighted to current moment information in memory cell unit state operation, described weighted two-way long-short-term memory network Memory cell unit state of memory network unit I

The weighted two-way long-short-term memory network unit II is a two-way long-short-term memory network that weights the information of the past time, and is a two-way long-short-term memory network that weights the information of the past time in the state operation of the memory cell unit. Memory cell unit states of memory network unit II

代表当前时刻的信息，c_t-1代表过去时刻的信息，λ₁为对保留下来的当前时刻信息进一步加权的权重，λ₂为对保留下来的过去时刻信息进一步加权的权重，i_t为记忆细胞单元状态对当前时刻信息保留的比重，f_t为记忆细胞单元状态对过去时刻信息保留的比重；in,

Represents the information of the current moment, c _t-1 represents the information of the past moment, λ ₁ is the weight that further weights the retained information of the current moment, λ ₂ is the weight that further weights the retained information of the past moment, i _t is the memory The proportion of the cell unit state to the information retained at the current moment, f _t is the proportion of the memory cell unit state to the information retained at the past moment;

After getting the predicted value, compare it with the set value:

If the error between the predicted value of the degree of polymerization, the average molecular weight and the process setting value is within plus or minus 1, and the error between the predicted value of intrinsic viscosity and esterification rate and the process setting value is within plus or minus 0.1, it is not necessary to adjust the process parameters;

If the above valid range is exceeded, the process parameters need to be adjusted;

The specific adjustment method is:

According to the feature correlation ranking and ranking obtained by the extreme gradient boosting tree algorithm of grid search, the process parameters with the highest correlation, that is, the highest ranking process parameters, are adjusted first. If the adjustment is still not satisfied to the limit, the process parameters that are ranked second will continue to be adjusted. parameters, and so on, until the predicted performance parameters meet the requirements;

The adjustment step is 1% of the specific value of the process parameter, and the adjustment limit is ±10% of the specific value of the process parameter;

The adjustment direction is based on the comparison between the performance prediction result and the set value. If the performance prediction result is higher than the set value, the adjustment direction is decrease, and if the performance prediction result is lower than the set value, the adjustment direction is increase.

As a preferred solution:

The above-mentioned method for adjusting process parameters in a polyester fiber polymerization process, the specific modeling and training process of the weighted dual-stream bidirectional long-term memory attention network is:

(1) Modeling;

The weighted dual-stream bidirectional long-term and short-term memory attention network includes three parts: data preprocessing, information extraction and fusion, and regression output;

a) Data preprocessing part:

It consists of three parts: feature data collection, feature correlation sorting of feature data, and optimal processing;

First, collect m feature data from all the feature data collected by the sensor. The m feature data are respectively selected from all temperature sensors in the whole aggregation process with the same interval from the first temperature sensor. [(m-1) /5] The temperature of the temperature sensors, the same interval is selected from the first pressure sensor and rounded to the nearest [(m-1)/5] The pressure of the pressure sensors, 1 slurry ratio data of the slurry preparation tank , the liquid level of the first liquid level sensor with the same interval will be rounded to the nearest [(m-1)/5] liquid level sensors, and the liquid level of the first flow sensor with the same interval will be rounded to the nearest [(m The flow rate of -1)/5] flow sensors, and the rotation speed of [(m-1)/5] speed sensors is rounded up from the first speed sensor with the same interval;

The limit gradient boosting tree algorithm based on grid search performs feature correlation sorting on m feature data;

Optimal processing, select the top n feature data according to feature correlation sorting;

Normalize the selected n feature data to obtain new n feature data;

b) Extraction and fusion part:

Firstly, the past time information and current time information of the n characteristic data obtained by normalization are extracted respectively by using the double-stream framework of the first and second branches in parallel;

The information obtained from the two tributaries of the dual-stream framework is longitudinally spliced by using the merging layer, and then the input information is further weighted and summed by the attention layer according to the correlation between the input information and the output variable;

The two branches in the adopted dual-stream framework are: the first branch consists of two identical weighted bidirectional long-short-term memory network units I, and a batch normalization layer is inserted into them, that is, a weighted bidirectional long-short-term memory network unit. I is followed by a batch normalization layer followed by a weighted two-way long-short-term memory network unit I; Described weighted two-way long-short-term memory network unit I is a two-way long-short-term memory network weighted to the current moment information, which is a memory cell unit state calculation. A bidirectional long-term and short-term memory network for weighted operation of the information at the current moment, the memory cell unit state of the weighted bidirectional long-term and short-term memory network unit I

The second branch is composed of two identical weighted bidirectional long-short-term memory network units II, and a batch normalization layer is inserted into them, that is, a weighted bidirectional long-short-term memory network unit II is followed by a batch normalization layer followed by a weighted bidirectional long-term memory network unit II. Time memory network unit II; Described weighted bidirectional long and short-term memory network unit II is a bidirectional long and short-term memory network that weights the information of the past time, and is a bidirectional long and short-term memory network that weights the information of the past time in the state operation of the memory cell unit, The memory cell unit state of the weighted bidirectional long-term memory network unit II

Then use the merging layer to fuse the information extracted by the dual-stream framework, and input the fused information to the attention layer to further extract information;

c) regression output;

Finally, the information extracted by the attention layer is input to the fully connected layer for classification, and the prediction result is output;

The establishment of a weighted dual-stream bidirectional long-term and short-term memory attention network is completed;

(2) training;

The weighted dual-stream bidirectional long-short-term memory attention network is trained by the method of stochastic gradient descent; the training objects are all hyperparameters of the weighted dual-stream bidirectional long-short-term memory attention network;

The termination condition of training is that the accuracy is high enough or reaches the predefined maximum number of iterations; the accuracy is high enough means that MSE<0.005, and MAE<0.05, the calculation formula of MSE is as follows:

为预测值；In the formula, M is the number of samples, t is the sample number, y _t is the true value,

is the predicted value;

The formula for calculating MAE is as follows:

(3) Model calibration;

If the MSE ≥ 0.005 or MAE ≥ 0.05 after the training reaches the maximum number of iterations set in advance, choose to reduce the batch size and adjust the step size to 1, or increase the predefined number of generations and adjust the step size to 30 generations. The predefined maximum number of iterations is 100 generations, and the value of batch size is 32.

The adjustment method of the process parameter in a kind of polyester fiber polymerization process as above, the bidirectional long-short-term memory network unit of weighted bidirectional long-short-term memory network unit 1 is made up of input layer, output layer, forward layer and backward layer, Each layer is composed of multiple neurons, and each neuron is a long-short-term memory network unit. The backward layer is composed of multiple neurons, each of which is a long-short-term memory network unit. The batch normalization layer is a method that provides zero mean/unit variance for any layer in the neural network. input technology.

The above-mentioned method for adjusting process parameters in the polyester fiber polymerization process, i _t and f _t are both numbers between 0 and 1, and the values of i _t and f _t are obtained by stochastic gradient descent;

i _t =σ(W _i *[h _t-1 , x _t ]+ _bi );

f _t =σ(W _f *[h _t-1 , x _t ]+b _f );

In the formula, σ is the sigmoid function, h _t-1 is the information at time t-1, x _t represents the input feature at time t, and W _i , _bi , W _f and b _f are the parameters to be trained;

The specific training process is:

set objective function

Among them, J(θ) is the loss function, θ is the parameters to be trained, namely W _i , b _i , W _f and b _f , which are the values to be solved iteratively, and h(θ) is the function to be fitted, that is, for four For the prediction of performance indicators, y ⁱ is the real value of the four performance indicators, and m is the number of training sets. Because stochastic gradient descent is trained one by one, m=1 here; it is achieved by minimizing the loss function by gradient descent. Training with required parameters.

A method for adjusting the process parameters in the above-mentioned polyester fiber polymerization process, the specific process of the score function calculation is:

为第t个特征在树结构为q时的得分函数值，T为树结构为q时的树的叶子节点的个数， I_j为树结构为q时树的第j个节点，g_i为损失函数的一阶导数，w_j为树结构为q时的第j个叶子结点的权重值， h_i为损失函数的二阶导数，λ和γ为超参数；每棵树对于每个特征都有一个得分函数值，最后所有树关于该特征的得分函数值的和即为该特征关于输出变量相关性的得分值，得分函数的值越大，代表该特征与输出变量的相关性越大；In the formula,

is the score function value of the t-th feature when the tree structure is q, T is the number of leaf nodes of the tree when the tree structure is q, I _j is the j-th node of the tree when the tree structure is q, and g _i is The first derivative of the loss function, w _j is the weight value of the _jth leaf node when the tree structure is q, hi is the second derivative of the loss function, λ and γ are hyperparameters; each tree is for each feature. There is a score function value. Finally, the sum of the score function values of all trees about the feature is the score value of the feature’s correlation with the output variable. The larger the value of the score function, the greater the correlation between the feature and the output variable. Big;

The loss function is the square loss of the melt intrinsic viscosity MIV in the polyester fiber fiber polymerization process, and the square loss refers to the squared error between the actual value and the predicted value.

A method for adjusting process parameters in a polyester fiber polymerization process as described above, the limit gradient boosting tree based on grid search selects the first n features from m features, and the specific number of n values is rounded up Round 9m/20.

The above-mentioned adjustment method of the technological parameter in a kind of polyester fiber polymerization process, the formula of normalization treatment is as follows:

In the formula, x _i is the result of the normalization of the ith sample, the set of all samples in a certain characteristic data is set X, X _i is the ith sample in the set X, and X _min is all the samples in the set X. The minimum value of X _max is the maximum value of all samples in the set X.

In the above-mentioned method for adjusting process parameters in the polyester fiber polymerization process, in step (2) training, stochastic gradient descent is used to train the hyperparameters in the prediction model; stochastic gradient descent is: each iteration Use a sample to update the parameters to speed up the training;

The objective function for one sample is:

Among them, J(θ) is the loss function, θ is the parameter, the value to be solved iteratively, h(θ) is the function to be fitted, y ⁱ is the true value, and m is the number of bars in the training set, because stochastic gradient descent is Training is performed one sample at a time, so m=1 here.

The above-mentioned method for adjusting the process parameters in the polyester fiber polymerization process, the value range of λ ₁ is 0-1, and the value range of λ ₂ is 0-1, when the maximum iteration set in advance is reached When MSE ≥ 0.005 or MAE ≥ 0.05 after the number of times, the two parameters are adjusted, and the adjustment step is 0.1.

The present invention can effectively solve the problems of uncertainty and lack of timeliness existing in the prior art due to the high prediction accuracy and the correlation sorting of the features.

Specifically, the present invention first sorts m features by feature correlation based on a grid search-based limit gradient boosting tree, and selects the top n features. The purpose of this is to remove redundant features and reduce After analyzing the influence of irrelevant features on the prediction results, normalize all the sample data of the first n features to reduce the influence of different dimensions on feature correlation, and then use weighted dual-stream bidirectional long-term memory attention. The network (TS-λBiLSTMs-attention) extracts and fuses information for n features, and finally classifies the obtained information through a fully connected layer, and outputs the predicted values of four performance indicators.

In the weighted dual-stream bidirectional long-short-term memory attention network (TS-λBiLSTMs-attention) algorithm framework, the present invention introduces parameters λ ₁ and λ ₂ to weight the information of the past moment and the information of the current moment, so that the first branch focuses on For the extraction of current time information, the second branch focuses on the extraction of past time information. And a bidirectional long-short-term memory network is used to further consider the next moment information. Therefore, the information extracted and fused by the weighted dual-stream bidirectional long-short-term memory attention network (TS-λBiLSTMs-attention) algorithm framework includes the information of the past moment, the current moment and the next moment.

In conclusion, the weighted dual-stream bidirectional long-short-term memory attention network (TS-λBiLSTMs-attention) proposed above has high accuracy.

The technical points of the present invention are as follows:

(1) Soft measurement

Currently, soft sensing builds predictive models based on the vast amounts of data being measured and stored in the process industry. Related work is reported to complement on-line instrumental measurements for process monitoring and control.

Soft-sensor modeling is mainly divided into three categories, namely first-principles models (FPM) (white-box models), data-driven models (black-box models), and hybrid models. As the name suggests, a hybrid model is a combination of a first-principles model and a data-driven model. The first-principles model is built on a deep understanding of the physical and chemical context of the process, which is often time-consuming and difficult to obtain. With the development of sensor technology, A large amount of data can be obtained through distributed control system (DCS), which makes data-driven modeling more attractive. The essence of data-driven model is to mine the hidden information in data. The typical algorithm of data-driven model is ordinary least squares, support Vector regression (SVR) and partial least squares (PLS). The invention uses the data collected by the factory DCS to establish a data-driven soft measurement prediction model, and then in industrial production, process parameters can be substituted into the prediction model of the invention to obtain the prediction results of four performance indicators, and then the prediction results of the four performance indicators can be used. Compare with its corresponding process setting value, and finally determine whether to adjust the process parameters according to the difference between the two, so as to achieve the purpose of producing polyester fiber with high quality.

(2) Grid-Xgboost algorithm (limit gradient boosting tree based on grid search)

The Xgboost algorithm has been widely used by data scientists to obtain state-of-the-art results on many machine learning applications, an efficient and scalable machine learning algorithm developed by Tianqi Chen et al. for tree augmentation.

Extreme Gradient Boosting Trees intelligently obtain feature scores by building boosted trees that indicate the importance of each feature to the model. The higher the score of a feature, the higher its correlation with the output variable.

The above equation can be used as a scoring function to measure the correlation between features and output variables.

Grid search is a typical method for parameter tuning, which methodically builds and evaluates a model for each parameter combination in a particular grid. Xgboost is a successful machine learning method based on the gradient boosting algorithm proposed by Chen Tianqi. The second-order Taylor expansion of the objective function enables predictions with high accuracy, which is also attributed to its high success rate in Kaggle competitions.

The present invention selects the features most relevant to the output variables by the Xgboost algorithm, since the quality of the polymer is determined by four main performance indicators. Therefore, there are four output variables. The other three performance indicators are determined by the melt intrinsic viscosity MIV. Then, the squared loss of melt intrinsic viscosity MIV is used as the loss function in the Xgboost algorithm to rank the feature importance. The squared loss is the sum of squared errors between the actual and predicted values.

(3) Bi-directional Long Short-term Memory Networks (BiLSTMs)

LSTM is proposed based on Recurrent Neural Network (RNN) to solve the long-term dependency problem of time series. It achieves state-of-the-art performance in various tasks such as industrial time series analysis and recognition applications. An LSTM unit consists of three multiplication gates, which are input gate i _t , output gate o _i and forget gate ft _. They respectively control the proportion of input, output or forgetting information to the next moment. The core of the long-short-term memory network is the memory cell unit state c _t , because it can store the information of the last moment, which effectively solves the long-term dependence problem in the industrial process. The structure of a unit of a specific LSTM is shown in Figure 1.

The principle of LSTM is described by the following equation:

代表着矩阵乘法，这代表着当前时刻的信息与过去时刻信息的融合。权重矩阵W_*和偏置矩阵b_*分别是模型的参数，他们的值在模型训练的过程中被决定。

代表当前时刻的信息，c_t-1代表上一时刻的信息。

代表保留i_t比例的当前时刻的信息，

代表保留f_t比例的过去时刻的信息。Here σ represents a logical sigmoid function, and the values obtained by it, ft and o _t are all between ₀ and ₁ , which indicates the current input, output or the proportion of information retained to the next moment, tanh is a double tangent function. symbol

Represents matrix multiplication, which represents the fusion of current moment information and past moment information. The weight matrix W _* and the bias matrix b _* are the parameters of the model, respectively, and their values are determined during the model training process.

Represents the information of the current moment, and c _t-1 represents the information of the previous moment.

represents the information of the current moment in which the proportion of it is _preserved ,

Represents information about past _moments in which the scale of ft is preserved.

Usually, the model prediction results are not only related to the current input, but also related to the input at the next moment. Based on this, bidirectional long short-term memory networks (BiLSTMs) are proposed to improve the prediction accuracy of LSTM models.

As shown in Figure 2, the bidirectional long short-term memory network (BiLSTMs) mainly includes input layer, output layer, forward layer and backward layer. The forward layer and the backward layer are jointly connected to the output layer, which contains 6 shared Weights w ₁ -w ₆ . First, forward computation from time 1 to time t in the forward layer, so that the output of the hidden layer can be obtained and saved each time. Next, the backward layer computes from time t to time 1 to obtain and save the output of the hidden layer after each time instant. Finally, at each moment, the final output is obtained by combining the output results of the forward layer and the backward layer at the corresponding moment. The mathematical expression is as follows:

h _t =f(w ₁ x _t +w ₂ h _t-1 )

h _t ′=f(w ₃ x _t +w ₀ h _t+1 ′)

o _t =g(w ₄ h _t +w ₆ h _t ′)

where f is usually a sigmoid function. Similarly, g is an activation function, and for different problems, g takes different functions. h _t-1 , h _t and h _t+1 represent the information of the hidden layer at the past moment, the current moment and the next moment, respectively, and x _t represents the current input information.

(4) attention (attention mechanism)

In the past two years, attention mechanisms have been widely used in natural language processing, image recognition, speech recognition and other different types of deep learning tasks. It is one of the core technologies of deep learning technology, which deserves attention and in-depth understanding. The essence of the attention mechanism is to get more detailed information about what needs attention and suppress other useless information. That is, its core operation is a set of weight parameters to learn the importance of each element from the sequence, and then combine those elements based on importance. The weight parameter is the attention coefficient distribution, reflecting how much attention is assigned to which element.

Specifically, in the traditional attention mechanism, first, the previous state H={h ₁ , h ₂ , ... h _t-1 } is defined. Then the weight sum of each column is obtained as v _t . The vector h _t-1 represents the information extracted from the past state. h _t represents the information extracted from the current state. Then assume a score function f: R ^m × R ^m → R to calculate the correlation between the input variables. Finally, the value of the vector v _t is calculated by the following equation, where α _i represents the weight of the correlation between the ith feature and the output variable.

Beneficial effects:

(1) The present invention proposes a new feature extraction and fusion method, and introduces a weighted dual-stream bidirectional long-short-term memory attention network (TS-λBiLSTMs-attention) algorithm framework that not only considers the information of the past and present moments, but also considers the upcoming moment at the next moment. The input information, in which, by introducing an attention mechanism to weight the features related to the output variables, the prediction accuracy is further improved;

(2) The present invention solves the multi-input multi-output soft-sensor modeling problem in the polyester fiber polymerization process, and establishes a soft-sensor model for the polyester fiber performance indicators, which is convenient for real-time monitoring of the changes of the four performance indicators in the industrial production process. To achieve the purpose of high-level production;

(3) In the present invention, as long as the value of the process parameter is given, the predicted value can be accurately obtained, and the producer can better guide the industrial production by judging whether the obtained predicted value is close to the set value.

Description of drawings

Figure 1 is a schematic diagram of the structure of a long short-term memory network (LSTM);

Figure 2 is a schematic diagram of the structure of a bidirectional long short-term memory network (BiLSTMs);

Fig. 3 is the algorithm framework adopted by the present invention to obtain the predicted value of the polyester fiber performance index;

4 is a schematic diagram of a polyester fiber polymerization process;

Figure 5 is a comparison chart of the predicted value and the real value of the four performance indicators based on the weighted dual-stream bidirectional long-short-term memory attention network (TS-λBiLSTMs-attention) algorithm;

FIG. 6 is a comparison diagram of absolute errors of MIV based on different soft sensing algorithms.

Detailed ways

The present invention will be further described below in conjunction with specific embodiments. It should be understood that these examples are only used to illustrate the present invention and not 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 also fall within the scope defined by the appended claims of the present application.

A method for adjusting process parameters in the polyester fiber polymerization process, after obtaining the predicted value of the polyester fiber performance index, comparing it with the set value, and adjusting the process parameters in the polyester fiber polymerization process according to the comparison result;

The performance indicators of polyester fiber are intrinsic viscosity MIV, esterification rate E _s , degree of polymerization P _n and average molecular weight _Mn ;

As shown in Figure 3, the acquisition process of the predicted value is as follows:

First, collect m feature data from all feature data collected by the sensor;

Then the limit gradient boosting tree algorithm based on grid search performs feature correlation sorting on the m feature data, that is, the value of the score function is calculated for each feature data. The higher the value of the score function, the greater the feature correlation; For the m feature data, select the first n feature data with high feature correlation, n∈m, and the specific number of n values is rounded to the nearest 9m/20;

The specific process of calculating the score function is as follows:

为第t个特征在树结构为q时的得分函数值，T为树结构为q时的树的叶子节点的个数， I_j为树结构为q时树的第j个节点，g_i为损失函数的一阶导数，w_j为树结构为q时的第j个叶子结点的权重值，h_i为损失函数的二阶导数，λ和γ为超参数；每棵树对于每个特征都有一个得分函数值，最后所有树关于该特征的得分函数值得和即为该特征关于输出变量相关性的得分值，得分函数的值得结果越大，代表该特征与输出变量的相关性越大；In the formula,

is the score function value of the t-th feature when the tree structure is q, T is the number of leaf nodes of the tree when the tree structure is q, I _j is the j-th node of the tree when the tree structure is q, and g _i is The first derivative of the loss function, w _j is the weight value of the _jth leaf node when the tree structure is q, hi is the second derivative of the loss function, λ and γ are hyperparameters; each tree is for each feature. There is a score function value. Finally, the sum of the score function values of all trees about the feature is the score value of the feature’s correlation with the output variable. The greater the value of the score function, the greater the correlation between the feature and the output variable. Big;

Then, normalize all samples of n feature data; the formula for normalization is as follows:

In the formula, x _i is the result of the normalization of the ith sample, the set of all samples in a certain characteristic data is set X, X _i is the ith sample in the set X, and X _min is all the samples in the set X. The minimum value of , X _max is the maximum value of all samples in the set X;

Finally, the weighted dual-stream bidirectional long-term and short-term memory attention network is used to process the normalized n feature data, and output the predicted values of the four polyester fiber performance indicators;

The weighted dual-stream bidirectional long-term and short-term memory attention network is a network in which the information obtained from the two tributaries is longitudinally spliced by the merging layer and input to the fully-connected layer through the attention layer;

The two tributaries are the first tributary and the second tributary in parallel; the first tributary consists of two identical weighted bidirectional long-short-term memory network units I, and a batch normalization layer is inserted between the two; the The second branch consists of two identical weighted bidirectional long-short-term memory network units II with a batch normalization layer inserted between them;

Described weighted two-way long-short-term memory network unit 1 is a two-way long-short-term memory network weighted to current moment information, is a two-way long-short-term memory network weighted to current moment information in memory cell unit state operation, described weighted two-way long-short-term memory network Memory cell unit state of memory network unit I

The weighted two-way long-short-term memory network unit II is a two-way long-short-term memory network that weights the information of the past time, and is a two-way long-short-term memory network that weights the information of the past time in the state operation of the memory cell unit. Memory cell unit states of memory network unit II

代表当前时刻的信息，c_t-1代表过去时刻的信息，λ₁为对保留下来的当前时刻信息进一步加权的权重，λ₂为对保留下来的过去时刻信息进一步加权的权重，i_t为记忆细胞单元状态对当前时刻信息保留的比重，f_t为记忆细胞单元状态对过去时刻信息保留的比重；in,

Represents the information of the current moment, c _t-1 represents the information of the past moment, λ ₁ is the weight that further weights the retained information of the current moment, λ ₂ is the weight that further weights the retained information of the past moment, i _t is the memory The proportion of the cell unit state to the information retained at the current moment, f _t is the proportion of the memory cell unit state to the information retained at the past moment;

Both i _t and f _t are numbers between 0 and 1, and the values of i _t and f _t are obtained by training through stochastic gradient descent;

l _t =σ(W _i *[h _t-1 , x _t ]+ _bi );

f _t =σ(W _f *[h _t-1 , x _t ]+b _f );

In the formula, σ is the sigmoid function, h _t-1 is the information at time t-1, x _t represents the input feature at time t, and W _i , _bi , W _f and b _f are the parameters to be trained;

The specific training process is:

set objective function

Among them, J(θ) is the loss function, the loss function is the squared loss of the melt intrinsic viscosity MIV, the squared loss refers to the squared error between the actual value and the predicted value, and θ is the parameter to be trained, namely W _i , b _i , W _f and b _f , are the values to be solved iteratively, h(θ) is the function to be fitted, that is, the prediction of the four performance indicators, y ⁱ is the real value of the four performance indicators, and m is the bar of the training set. number, because the stochastic gradient descent is trained one by one, so here m=1; the training of the required parameters is achieved by minimizing the loss function through the gradient descent method;

The specific modeling and training process of the weighted dual-stream bidirectional long-term memory attention network is as follows:

(1) Modeling;

The weighted dual-stream bidirectional long-term and short-term memory attention network includes three parts: data preprocessing, information extraction and fusion, and regression output;

a) Data preprocessing part:

It consists of three parts: feature data collection, feature correlation sorting of feature data, and optimal processing;

First, collect m feature data from all the feature data collected by the sensor. The m feature data are respectively selected from all temperature sensors in the whole aggregation process with the same interval from the first temperature sensor. [(m-1) /5] The temperature of the temperature sensors, the same interval is selected from the first pressure sensor and rounded to the nearest [(m-1)/5] The pressure of the pressure sensors, 1 slurry ratio data of the slurry preparation tank , the liquid level of the first liquid level sensor with the same interval will be rounded to the nearest [(m-1)/5] liquid level sensors, and the liquid level of the first flow sensor with the same interval will be rounded to the nearest [(m The flow rate of -1)/5] flow sensors, and the rotation speed of [(m-1)/5] speed sensors is rounded up from the first speed sensor with the same interval;

The limit gradient boosting tree algorithm based on grid search performs feature correlation sorting on m feature data;

Optimal processing, select the top n feature data according to feature correlation sorting;

Normalize the selected n feature data to obtain new n feature data;

b) Extraction and fusion part:

Firstly, the past time information and current time information of the n characteristic data obtained by normalization are extracted respectively by using the double-stream framework of the first and second branches in parallel;

The information obtained from the two tributaries of the dual-stream framework is longitudinally spliced by using the merging layer, and then the input information is further weighted and summed by the attention layer according to the correlation between the input information and the output variable;

The two branches in the adopted dual-stream framework are: the first branch consists of two identical weighted bidirectional long-short-term memory network units I, and a batch normalization layer is inserted into them, that is, a weighted bidirectional long-short-term memory network unit. I is followed by a batch normalization layer followed by a weighted two-way long-short-term memory network unit I; Described weighted two-way long-short-term memory network unit I is a two-way long-short-term memory network weighted to the current moment information, which is a memory cell unit state calculation. A bidirectional long-term and short-term memory network for weighted operation of the information at the current moment, the memory cell unit state of the weighted bidirectional long-term and short-term memory network unit I

The weighted bidirectional long and short-term memory network unit I of the bidirectional long-term and short-term memory network unit consists of an input layer, an output layer, a forward layer and a backward layer, each layer is composed of multiple neurons, and each neuron is a long and short Time memory network unit, weighted bidirectional long and short-term memory network unit II The bidirectional long and short-term memory network unit is composed of an input layer, an output layer, a forward layer and a backward layer, each layer is composed of multiple neurons, each of which is composed of neurons. The element is a long-short-term memory network unit, and the batch normalization layer is a technique that provides zero mean/unit variance input to any layer in the neural network;

The second branch is composed of two identical weighted bidirectional long-short-term memory network units II, and a batch normalization layer is inserted into them, that is, a weighted bidirectional long-short-term memory network unit II is followed by a batch normalization layer followed by a weighted bidirectional long-term memory network unit II. Time memory network unit II; Described weighted bidirectional long and short-term memory network unit II is a bidirectional long and short-term memory network that weights the information of the past time, and is a bidirectional long and short-term memory network that weights the information of the past time in the state operation of the memory cell unit, The memory cell unit state of the weighted bidirectional long-term memory network unit II

Then use the merging layer to fuse the information extracted by the dual-stream framework, and input the fused information to the attention layer to further extract information;

c) regression output;

Finally, the information extracted by the attention layer is input to the fully connected layer for classification, and the prediction result is output;

The establishment of a weighted dual-stream bidirectional long-term and short-term memory attention network is completed;

(2) training;

The weighted dual-stream bidirectional long-short-term memory attention network is trained by the method of stochastic gradient descent; the training objects are all hyperparameters of the weighted dual-stream bidirectional long-short-term memory attention network;

The termination condition of training is that the accuracy is high enough or reaches the predefined maximum number of iterations; the accuracy is high enough means that MSE<0.005, and MAE<0.05, the calculation formula of MSE is as follows:

为预测值；In the formula, M is the number of samples, t is the sample number, y _t is the true value,

is the predicted value;

The formula for calculating MAE is as follows:

The hyperparameters in the prediction model are trained by means of stochastic gradient descent; stochastic gradient descent means that each iteration uses a sample to update the parameters, which speeds up the training speed;

The objective function for one sample is:

Among them, J(θ) is the loss function, θ is the parameter, the value to be solved iteratively, h(θ) is the function to be fitted, y ⁱ is the true value, and m is the number of bars in the training set, because stochastic gradient descent is One by one for training, so here m=1;

(3) Model calibration;

If the MSE ≥ 0.005 or MAE ≥ 0.05 after the training reaches the maximum number of iterations set in advance, choose to reduce the batch size and adjust the step size to 1, or increase the predefined number of generations and adjust the step size to 30 generations. The predefined maximum number of iterations is 100 generations, and the value of batch size is 32;

The value range of λ ₁ is 0-1, and the value range of λ ₂ is 0-1. When the maximum number of iterations set in advance is reached, MSE≥0.005 or MAE≥0.05, these two parameters are adjusted, Adjust the step size to 0.1;

After getting the predicted value, compare it with the set value:

If the error between the predicted value of the degree of polymerization, the average molecular weight and the process setting value is within plus or minus 1, and the error between the predicted value of intrinsic viscosity and esterification rate and the process setting value is within plus or minus 0.1, it is not necessary to adjust the process parameters;

If the above valid range is exceeded, the process parameters need to be adjusted;

The specific adjustment method is:

According to the feature correlation ranking and ranking obtained by the extreme gradient boosting tree algorithm of grid search, the process parameters with the highest correlation, that is, the highest ranking process parameters, are adjusted first. If the adjustment is still not satisfied to the limit, the process parameters that are ranked second will continue to be adjusted. parameters, and so on, until the predicted performance parameters meet the requirements;

The adjustment step is 1% of the specific value of the process parameter, and the adjustment limit is ±10% of the specific value of the process parameter;

The adjustment direction is based on the comparison between the performance prediction result and the set value. If the performance prediction result is higher than the set value, the adjustment direction is decrease, and if the performance prediction result is lower than the set value, the adjustment direction is increase.

Example 1

A method for adjusting process parameters in a polyester fiber polymerization process, the steps are as follows:

(1) Collect m characteristic data, and the characteristic data is the process parameter data related to the polyester fiber performance index;

As shown in Figure 4, the polyester fiber polymerization process mainly includes three stages, namely esterification process, pre-polycondensation process and final polycondensation process. The research of the present invention is mainly aimed at the currently widely used DuPont three-pot process. Polyterephthalic acid (PTA) and ethylene glycol (EG) are mixed in a certain proportion in the slurry mixing tank, and then the mixed slurry is continuously fed into the slurry feeding tank for esterification. The reaction kettle, under the action of a certain temperature and pressure, the raw materials are chemically reacted to form oligomers, and then the oligomers are continuously fed into the pre-polycondensation reactor and the final polycondensation reactor to form high polymers, which are produced in various stages. The water vapor and ethylene glycol vapor obtained from the vacuum system enter the vacuum system for removal or recycling, and the production process has the characteristics of high nonlinearity, slow time change, and many parameter distributions;

The characteristic data is collected from a factory in China, and the frequency is per second. According to the process knowledge and operator experience of the factory, the present invention finally selects 67 process parameters as the characteristic data, that is, m=67;

(2) The extreme gradient boosting tree (Grid-Xgboost) algorithm based on grid search selects the top n characteristic data with the highest correlation with the polyester fiber performance index from m characteristic data;

Because the extreme gradient boosting tree can only sort the feature importance of one output at a time, and the present invention has four outputs, the present invention adopts the grid search-based extreme gradient boosting tree, namely the Grid-Xgboost algorithm, to characterize the four outputs respectively. order of importance;

The present invention needs to adjust the parameters to obtain the best performance of the extreme gradient boosting tree, similar to the random forest, the extreme gradient boosting tree is adjusted using hyperparameters, and all the parameters of the extreme gradient boosting tree are divided into three categories: the first category includes control The general parameters of the macro function; the second category includes the booster parameters used to control the booster (tree/regression) at each step; the last category has the learning target parameter, which is used to control the performance of the training target;

The present invention focuses on the two most important parameters of grid search, namely max_depth and n_estimators, in particular, max_depth is used to control the depth of the tree structure, and the parameter of n_estimators is an important tuning parameter because it is related to the complexity of the limit gradient boosting tree model In addition, it also represents the number of weak learners in the decision tree, and Table 1 lists the ranges and optimal values of the two parameters:

Table 1

Then, by applying grid-search-based extreme gradient boosting tree to each feature data, the feature importance ranks of the four performance indicators are ranked separately, and the results are shown in Table 2. Due to space constraints, only the same output as the output is shown. The 40 most relevant features of the variable, the present invention found that for four different performance indicators, the importance order of most features is the same:

Table 2

And because MIV is dominant among the four performance indicators, the application has selected 30 characteristic data most relevant to the intrinsic viscosity MIV. The statistical MSE results of the test set with different number of features are shown in Table 3:

table 3

As can be seen from Table 3, when the number of characteristic data is 30, the value of MSE in the test set tends to converge. To sum up, this application selects the 30 characteristic data most relevant to the intrinsic viscosity MIV as the weighted dual-stream bidirectional length The input of the temporal memory attention network (TS-λBiLSTMs-attention), that is, n=30, the 30 feature data are highlighted in Figure 4;

(3) The 30 feature data contains a total of 10,000 samples, 8,000 samples are selected to form the training set, and the rest form the test set;

(4) Construct and train a weighted dual-stream bidirectional long-short-term memory attention network (TS-λBiLSTMs-attention);

Using the training set to train a weighted two-stream bidirectional long-short-term memory attention network (TS-λBiLSTMs-attention) through stochastic gradient descent, the training objects are all parameters of the weighted two-stream bidirectional long-short-term memory attention network (TS-λBiLSTMs-attention) , in order to obtain the optimal values of the parameters λ ₁ and λ ₂ , this application adjusts them from 0 to 1, sets the parameter step size to 0.1, and finally sets λ ₁ to the optimal value of 0.9, the value of λ ₁ is the same as λ ₂ is the same, then the same method is used for the learning rate (dropout rate), and the optimal value of the final dropout rate is 0.7. There is a trade-off between batch size and convergence. The larger the batch size, the faster the training speed. Finally, the batch size The size is set to 36, and the maximum number of iterations is set to 100; the termination condition of training is that the accuracy is high enough or reaches a predefined algebra; the accuracy is high enough means that MSE<0.005, and MAE<0.05, the calculation formula of MSE is as follows:

为预测值；In the formula, M is the number of samples, t is the sample number, y _t is the true value,

is the predicted value;

The formula for calculating MAE is as follows:

(5) Input the test value into the weighted dual-stream bidirectional long-short-term memory attention network (TS-λBiLSTMs-attention), which outputs the predicted value of the polyester fiber performance index, as shown in Figure 5, the predicted value is basically the same as the actual value. Consistent, compare the predicted value with the set value, the set value is: intrinsic viscosity MIV 0.7, esterification rate E _s 0.95, polymerization degree P _n 125, average molecular weight M _n 20000, adjust polyester fiber polymerization process according to the comparison results Process parameters in;

To demonstrate the effectiveness of the proposed algorithm framework, the present invention performs two sets of experiments:

为预测值，AE的值越接近0，代表预测结果越好；In the first group, weighted two-stream bidirectional long-short-term memory attention network (TS-λBiLSTMs-attention), weighted two-stream bidirectional long-short-term memory network (TS-λBiLSTMs), two-stream bidirectional long-short-term memory network (TS-BiLSTMs) , the absolute error of the MIV performance index of the bidirectional long-short-term memory attention network (BiLSTMs-attention) and the bidirectional long-short-term memory network (BiLSTMs) are shown in Figure 6. These soft-sensor models are not in the multi-output soft-sensor model. As seen in the article, this set of experiments is performed only to prove that the proposed dual-stream framework, the weighted BiLSTM and its algorithmic framework combined with the attention mechanism are very effective, where the absolute error is defined as the difference between the predicted value and the The absolute value of the true value error, the specific formula is as follows: m refers to the mth test sample, y _m is the true value,

is the predicted value, the closer the value of AE is to 0, the better the prediction result;

It is worth mentioning that the five algorithms in Figure 6 have the same training set, test set, number of iterations and batch size. It can be seen that the absolute MIV performance index of the weighted dual-stream bidirectional long-short-term memory attention network (TS-λBiLSTMs-attention), bidirectional long-short-term memory attention network (BiLSTMs-attention) and bidirectional long-short-term memory network (BiLSTMs) The error fluctuates between 0-0.2, while the absolute error of the MIV performance metrics based on two-stream bidirectional long-short-term memory networks (TS-BiLSTMs) and weighted two-stream bidirectional long-short-term memory networks (TS-λBiLSTMs) fluctuates between 0-0.4 , specifically, the MIV absolute error of the weighted dual-stream bidirectional long-short-term memory attention network (TS-λBiLSTMs-attention) algorithm is mainly concentrated between 0 and 0.05, and the MIV of the bidirectional long-short-term memory attention network (BiLSTMs-attention) The absolute error changes from 0 to 0.1, and the weighted dual-stream bidirectional long-short-term memory network (TS-λBiLSTMs) is slightly better than the dual-stream bidirectional long-short-term memory network (TS-BiLSTMs), and its absolute error mainly fluctuates between 0.2 and 0.3, and it can be obtained that It is concluded that the introduction of the parameter λ is effective, and the multi-output dual-stream λBiLSTM and dual-stream BiLSTM are not as effective as BiLSTMs. The main reason is that the multiple outputs confuse the information and cannot effectively learn the weight of each feature, which also proves the introduction of the attention mechanism. Necessity and effectiveness, in conclusion, the proposed weighted two-stream bidirectional long-short-term memory attention network (TS-λBiLSTMs-attention) algorithm is superior to other methods;

In the second group, the proposed weighted two-stream bidirectional long-short-term memory attention network (TS-λBiLSTMs-attention) algorithm framework is compared with several widely used soft-sensing modeling methods, which are LSTMs (Long Short-term Memory Networks), GRU (Gate Recurrent Unit), PLS (PartialLeast Squares) and SVR (Support Vactor Regression), the 10 statistical average results are shown in Table 5. It can be seen from Table 5 that the algorithm proposed by the present invention The MSE values of the four performance indicators are all concentrated at about 0.0013, while the best MSE value of the current relatively advanced centralized soft-sensor prediction method is about 0.003, which shows that the algorithm proposed by the present invention is obviously better than other algorithms.

table 5

Claims (9)

1. a method for adjusting a process parameter in a polyester fiber polymerization process, characterized in that: after obtaining the predicted value of the polyester fiber performance index, compare it with a set value, and adjust the polyester fiber polymerization process according to the comparison result Process parameters in; The performance indicators of polyester fiber are intrinsic viscosity MIV, esterification rate E _s , degree of polymerization P _n and average molecular weight _Mn ; The process of obtaining the predicted value is as follows: First, collect m feature data from all feature data collected by the sensor; Then the limit gradient boosting tree algorithm based on grid search performs feature correlation sorting on the m feature data, that is, the value of the score function is calculated for each feature data. The higher the value of the score function, the greater the feature correlation; select the first n feature data with relatively large feature correlation, n∈m; Then, normalize all the samples of the n feature data; Finally, the weighted dual-stream bidirectional long-term and short-term memory attention network is used to process the normalized n feature data, and output the predicted values of the four polyester fiber performance indicators; The weighted dual-stream bidirectional long-short-term memory attention network is a network in which the information obtained from the two tributaries is longitudinally spliced by the merging layer and input to the fully-connected layer through the attention layer; The two tributaries are the first tributary and the second tributary in parallel; the first tributary consists of two identical weighted bidirectional long-short-term memory network units I, and a batch normalization layer is inserted between the two; the The second branch consists of two identical weighted bidirectional long-short-term memory network units II with a batch normalization layer inserted between them; Described weighted two-way long-short-term memory network unit 1 is a two-way long-short-term memory network weighted to current moment information, is a two-way long-short-term memory network weighted to current moment information in memory cell unit state operation, described weighted two-way long-short-term memory network Memory cell unit state of memory network unit I

The weighted two-way long-short-term memory network unit II is a two-way long-short-term memory network that weights the information of the past time, and is a two-way long-short-term memory network that weights the information of the past time in the state operation of the memory cell unit. Memory cell unit states of memory network unit II

in,

Represents the information of the current moment, c _t-1 represents the information of the past moment, λ ₁ is the weight that further weights the retained information of the current moment, λ ₂ is the weight that further weights the retained information of the past moment, i _t is the memory The proportion of the cell unit state to the information retained at the current moment, f _t is the proportion of the memory cell unit state to the information retained at the past moment; After getting the predicted value, compare it with the set value: If the error between the predicted value of the degree of polymerization, the average molecular weight and the process setting value is within plus or minus 1, and the error between the predicted value of intrinsic viscosity and esterification rate and the process setting value is within plus or minus 0.1, it is not necessary to adjust the process parameters; If the above valid range is exceeded, the process parameters need to be adjusted; The specific adjustment method is: According to the feature correlation ranking and ranking obtained by the extreme gradient boosting tree algorithm of grid search, the process parameters with the highest correlation, that is, the highest ranking process parameters, are adjusted first. If the adjustment is still not satisfied to the limit, the process parameters that are ranked second will continue to be adjusted. parameters, and so on, until the predicted performance parameters meet the requirements; The adjustment step is 1% of the specific value of the process parameter, and the adjustment limit is ±10% of the specific value of the process parameter; The adjustment direction is based on the comparison between the performance prediction result and the set value. If the performance prediction result is higher than the set value, the adjustment direction is decrease, and if the performance prediction result is lower than the set value, the adjustment direction is increase. 2. the adjustment method of the technological parameter in a kind of polyester fiber polymerization process according to claim 1, is characterized in that, the concrete modeling and training process of described weighted dual-stream bidirectional long-short-term memory attention network are: (1) Modeling; The weighted dual-stream bidirectional long-term and short-term memory attention network includes three parts: data preprocessing, information extraction and fusion, and regression output; a) Data preprocessing part: It consists of three parts: feature data collection, feature correlation sorting of feature data, and optimal processing; First, collect m feature data from all the feature data collected by the sensor. The m feature data are respectively selected from all temperature sensors in the whole aggregation process with the same interval from the first temperature sensor. [(m-1) /5] The temperature of the temperature sensors, the same interval is selected from the first pressure sensor and rounded to the nearest [(m-1)/5] The pressure of the pressure sensors, 1 slurry ratio data of the slurry preparation tank , the liquid level of the first liquid level sensor with the same interval will be rounded to the nearest [(m-1)/5] liquid level sensors, and the liquid level of the first flow sensor with the same interval will be rounded to the nearest [(m The flow rate of -1)/5] flow sensors, and the rotation speed of [(m-1)/5] speed sensors is rounded up from the first speed sensor with the same interval; The limit gradient boosting tree algorithm based on grid search performs feature correlation sorting on m feature data; Optimal processing, select the top n feature data according to feature correlation sorting; Normalize the selected n feature data to obtain new n feature data; b) Extraction and fusion part: Firstly, the past time information and current time information of the n characteristic data obtained by normalization are extracted respectively by using the double-stream framework of the first and second branches in parallel; The information obtained from the two tributaries of the dual-stream framework is longitudinally spliced by using the merging layer, and then the input information is further weighted and summed by the attention layer according to the correlation between the input information and the output variable; The two branches in the adopted dual-stream framework are: the first branch consists of two identical weighted bidirectional long-short-term memory network units I, and a batch normalization layer is inserted into them, that is, a weighted bidirectional long-short-term memory network unit. I is followed by a batch normalization layer followed by a weighted two-way long-short-term memory network unit I; Described weighted two-way long-short-term memory network unit I is a two-way long-short-term memory network weighted to the current moment information, which is a memory cell unit state calculation. A bidirectional long-term and short-term memory network for weighted operation of the information at the current moment, the memory cell unit state of the weighted bidirectional long-term and short-term memory network unit I

The second branch is composed of two identical weighted bidirectional long-short-term memory network units II, and a batch normalization layer is inserted into them, that is, a weighted bidirectional long-short-term memory network unit II is followed by a batch normalization layer followed by a weighted bidirectional long-term memory network unit II. Time memory network unit II; Described weighted bidirectional long and short-term memory network unit II is a bidirectional long and short-term memory network that weights the information of the past time, and is a bidirectional long and short-term memory network that weights the information of the past time in the state operation of the memory cell unit, The memory cell unit state of the weighted bidirectional long-term memory network unit II

Then use the merging layer to fuse the information extracted by the dual-stream framework, and input the fused information to the attention layer to further extract information; c) regression output; Finally, the information extracted by the attention layer is input to the fully connected layer for classification, and the prediction result is output; The establishment of a weighted dual-stream bidirectional long-term and short-term memory attention network is completed; (2) training; The weighted dual-stream bidirectional long-short-term memory attention network is trained by the method of stochastic gradient descent; the training objects are all hyperparameters of the weighted dual-stream bidirectional long-short-term memory attention network; The termination condition of training is that the accuracy is high enough or reaches the predefined maximum number of iterations; the accuracy is high enough means that MSE<0.005, and MAE<0.05, the calculation formula of MSE is as follows:

In the formula, M is the number of samples, t is the sample number, y _t is the true value,

is the predicted value; The formula for calculating MAE is as follows:

(3) Model calibration; If the MSE ≥ 0.005 or MAE ≥ 0.05 after the training reaches the maximum number of iterations set in advance, choose to reduce the batch size and adjust the step size to 1, or increase the predefined number of generations and adjust the step size to 30 generations. The predefined maximum number of iterations is 100 generations, and the value of batch size is 32. 3. the adjustment method of the process parameter in a kind of polyester fiber polymerization process according to claim 1 and 2, is characterized in that, the bidirectional long-short-term memory network unit of weighted bidirectional long-short-term memory network unit 1 is composed of input layer, output Layer, forward layer and backward layer, each layer is composed of multiple neurons, each neuron is a long and short-term memory network unit, and the weighted bidirectional long and short-term memory network unit II is a bidirectional long and short-term memory network unit. It is composed of input layer, output layer, forward layer and backward layer, each layer is composed of multiple neurons, and each neuron is a long-short-term memory network unit. The batch normalization layer is a kind of neural network. A technique where any layer in the network provides zero mean/unit variance input. 4. the adjustment method of the technological parameter in a kind of polyester fiber polymerization process according to claim 1, is characterized in that, i _t and f _t are the numbers between 0 and 1, the mode training by stochastic gradient descent _Get the values of it and _ft ; _{i t} =σ(W ₁ *[h _t-1 , x _t ]+bi ); f _t =σ(W _f *[h _t-1 , x _t ]+b _f ); In the formula, σ is the sigmoid function, h _t1 is the information at time t-1, x _t represents the input feature at time t, and W _i , b _i , W _f and b _f are the parameters to be trained; The specific training process is: set objective function

Among them, J(θ) is the loss function, θ is the parameters to be trained, namely W _i , b _i , W _f and b _f , which are the values to be solved iteratively, and h(θ) is the function to be fitted, that is, for four For the prediction of performance indicators, y ⁱ is the real value of the four performance indicators, and m is the number of training sets. Because stochastic gradient descent is trained one by one, m=1 here; it is achieved by minimizing the loss function by gradient descent. Training with required parameters. 5. the adjustment method of the technological parameter in a kind of polyester fiber polymerization process according to claim 1, is characterized in that, the concrete process of described score function calculation is:

In the formula,

is the score function value of the t-th feature when the tree structure is q, T is the number of leaf nodes of the tree when the tree structure is q, I _j is the j-th node of the tree when the tree structure is q, and g _i is The first derivative of the loss function, w _j is the weight value of the _jth leaf node when the tree structure is q, hi is the second derivative of the loss function, λ and γ are hyperparameters; each tree is for each feature. There is a score function value. Finally, the sum of the score function values of all trees about the feature is the score value of the feature’s correlation with the output variable. The larger the value of the score function, the more the correlation between the feature and the output variable. Big; The loss function is the square loss of the melt intrinsic viscosity MIV in the polyester fiber fiber polymerization process, and the square loss refers to the squared error between the actual value and the predicted value. 6. the adjustment method of the technological parameter in a kind of polyester fiber polymerization process according to claim 1, is characterized in that, the limit gradient boosting tree based on grid search selects the first n features from m features, The specific value of n is rounded to the nearest 9m/20. 7. the adjustment method of the technological parameter in a kind of polyester fiber polymerization process according to claim 1, is characterized in that, the formula of normalization is as follows:

In the formula, x _i is the result of the normalization of the ith sample, the set of all samples in a certain feature data is set X, X _i is the ith sample in the set X, and X _min is all the samples in the set X. The minimum value of X _max is the maximum value of all samples in the set X. 8. the adjustment method of the technological parameter in a kind of polyester fiber polymerization process according to claim 2, is characterized in that, in step (2) training, adopts the mode of stochastic gradient descent to train the hyperparameter in the prediction model ; Stochastic gradient descent means: each iteration uses a sample to update the parameters, which speeds up the training speed; The objective function for one sample is:

Among them, J(θ) is the loss function, θ is the parameter, the value to be solved iteratively, h(θ) is the function to be fitted, y ⁱ is the true value, and m is the number of bars in the training set, because stochastic gradient descent is Training is performed one sample at a time, so m=1 here. 9. the adjustment method of the technological parameter in a kind of polyester fiber polymerization process according to claim 1, is characterized in that, the value range of λ ₁ is 0-1, and the value range of λ ₂ is 0-1 , when MSE ≥ 0.005 or MAE ≥ 0.05 after reaching the maximum number of iterations set in advance, adjust these two parameters, and the adjustment step is 0.1.

Images