Method for predicting key process parameters in variable working condition complex environment

Abstract

The invention relates to the technical field of computer deep learning, in particular to a method for predicting key process parameters in a variable working condition complex environment. The prediction method comprises the following steps: collecting process parameter data of a production process at different moments under a previous working condition to obtain a previous working condition process parameter data set, and training a pre-constructed key process parameter prediction MTEGAT model by adopting the previous working condition process parameter data set to obtain a trained key process parameter prediction MTEGAT model; s2, collecting technological parameter data of the production process at different moments under the current working condition to obtain a technological parameter data set of the current working condition, and training the trained key technological parameter prediction MTEGAT model obtained in the step S1 by adopting the technological parameter data set of the current working condition to obtain an updated key technological parameter prediction MTEGAT model; and inputting the process parameter data of the production process under the current working condition into the updated key process parameter prediction MTEGAT model for prediction to obtain a prediction result of the key process parameters of the production process under the current working condition. When the prediction method encounters a new operation condition, continuous learning can be carried out by consolidating knowledge of the previous condition, the prediction accuracy is high, and the robustness is good.

Description

Technical Field

[0001] The present application belongs to the field of computer deep learning technology, and specifically relates to a method for predicting key process parameters under complex and variable operating conditions. Background Art

[0002] In complex industrial production processes such as steel continuous casting, abnormal key process parameters can seriously affect the surface quality and mechanical properties of products. By predicting key process parameters, on-site operators can take necessary measures in advance to prevent potential accidents. In addition, analyzing the prediction results can provide valuable insights for optimizing production processes.

[0003] In recent years, data-driven methods based on machine learning and deep learning have become the mainstream of key process parameter prediction research due to their simple form and low degree of engineering. Taking the prediction of mold level fluctuation in the continuous casting process as an example, Lei et al. proposed a continuous casting mold level fluctuation prediction model based on empirical mode decomposition (EMD) and support vector regression (SVR). He et al. proposed a prediction method based on genetic algorithm (GA) and convolutional neural network (CNN). In order to reduce the impact of noise on the prediction of mold level fluctuation, Su et al. performed wavelet threshold denoising (WTD) on the high-frequency IMF after EMD decomposition of the continuous casting signal, and used long short-term memory network (LSTM) to predict the mold level fluctuation. Considering the frequency domain characteristics of continuous casting data, Meng et al. combined the time-frequency characteristics of different production process parameters to predict the mold level fluctuation.

[0004] The above models are usually based on a certain stable operating state of continuous casting, that is, the operating conditions remain constant. However, in the actual continuous casting process, products of different specifications are usually produced together, and operating parameters such as casting speed need to be constantly adjusted. In addition, it is also necessary to flexibly adjust various operating parameters according to the production speed, molten steel capacity and equipment process accident handling requirements. The frequent changes in operating conditions in the continuous casting process lead to a significant decrease in model accuracy. Therefore, it is necessary to establish a high-precision mold liquid level fluctuation prediction system that can adapt to the multi-condition characteristics of the continuous casting process.

[0005] At present, the main strategies for solving the multi-condition prediction problem include hybrid modeling methods, transfer learning methods, and continuous learning methods. The hybrid modeling method builds a model for each operating condition in the offline stage, and then uses the corresponding model for predictive analysis according to the operating condition in the online stage. The transfer learning method adjusts the model parameters by finding the similarities between the new and existing conditions to build a multi-condition prediction model. The operating conditions of complex industrial processes represented by continuous casting are not only changeable, but also new conditions will continue to appear. The hybrid modeling method requires a predefined number of conditions, and each condition requires a large amount of training data, which is difficult to implement in the actual production process. The transfer learning method helps to apply the trained model under various operating conditions, but it is usually suitable for situations with a small number of conditions and some prior knowledge. Unlike the above methods, the continuous learning method can adapt to changes in the surrounding environment, continuously learn new knowledge to update the model, and is suitable for the actual production process. However, the continuous learning method has a catastrophic forgetting problem, which causes the model to forget the knowledge of previous conditions when learning new condition data, thereby reducing the accuracy of the model.

[0006] In addition, the operating parameters and process parameters of complex industrial processes are coupled with each other, resulting in the prediction of key process parameters being affected by various direct and indirect factors. Therefore, at the production site, accurately understanding the main influencing factors of key process parameters under different working conditions and making targeted decisions is another major issue that engineers are interested in. However, current modeling methods are usually end-to-end models, which make it difficult to capture and express the complex dependencies between production process parameters, limiting their effectiveness in practical applications. Summary of the invention

[0007] In view of the problems and shortcomings in the prior art, the purpose of the present invention is to provide a method for predicting key process parameters under complex and variable operating conditions.

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

[0009] The first aspect of the present invention provides a method for predicting key process parameters under a complex environment with variable working conditions, comprising the following steps:

[0010] S1: Collect process parameter data of the production process at different times under previous working conditions to obtain a previous working condition process parameter data set, wherein the previous working condition process parameter data set includes multiple groups of process parameters at different times, and each group of process parameters includes key process parameters; divide the previous working condition process parameter data set into a training set and a test set in proportion, use the training set to train a pre-built key process parameter prediction MTEGAT model, update the network parameters of the key process parameter prediction MTEGAT model, and obtain the trained key process parameter prediction MTEGAT model; use the test set to test the trained key process parameter prediction MTEGAT model, evaluate the performance of the trained key process parameter prediction model MTEGAT model, until the accuracy of the trained key process parameter prediction model MTEGAT model for process parameter prediction is greater than or equal to a preset threshold, and obtain the trained key process parameter prediction MTEGAT model;

[0011] S2: Collecting process parameter data of the production process at different times under current working conditions to obtain a current working condition process parameter data set, wherein the current working condition process parameter data set includes multiple groups of process parameters at different times, and each group of process parameters includes key process parameters; using the current working condition process parameter data set to train the trained key process parameter prediction MTEGAT model obtained in step S1, and updating the network parameters of the trained key process parameter prediction MTEGAT model to obtain an updated key process parameter prediction MTEGAT model;

[0012] S3: Inputting the process parameter data of the production process under the current working conditions into the MTEGAT model for updating the key process parameter prediction to perform prediction, obtain the prediction results of the key process parameters of the production process under the current working conditions and output them.

[0013] According to the above prediction method, preferably, in step S1, the key process parameter prediction MTEGAT model consists of an MTE information layer, an embedded attention layer and a fully connected layer; the MTE information layer is used to calculate the causal dependency between different process parameters in the production process, and the input of the MTE information layer is the time series data of the process parameters at different times of the production process under the same working conditions collected over a period of time, and the output of the MTE information layer is the causal dependency between the process parameters; the embedded attention layer is used to calculate the input process parameter data and the causal dependency between the process parameters to obtain a feature representation of each input process parameter, and the input of the embedded attention layer is the causal dependency between the process parameter data and the process parameters, and the output of the embedded attention layer is the feature representation of the input process parameters; the fully connected layer is used to calculate the feature representation of the process parameters output by the embedded attention layer to obtain the prediction results of the key process parameters, wherein the input of the fully connected layer is the feature representation of the process parameters output by the embedded attention layer, and the output of the fully connected layer is the prediction results of the key process parameters.

[0014] According to the above prediction method, preferably, the specific operation of the MTE information layer for calculating the causal dependency relationship between different process parameters in the production process is:

[0015] S01, let X and Y be two different process parameters of the production process, X = {x 1 ,x 2 ...x T},Y={y 1 ,y 2 ,...y T}, input the time series data of process parameters X and Y collected within a period of time into the MTE information layer, and use the MTE information layer to perform coarse-grained processing of the time series data of process parameters X and Y at different decomposition scales to obtain coarse-grained sequences P of process parameters X and Y at different decomposition scales m and Q m , P m and Q m The calculation formulas are shown in Formula 1 and Formula 2 respectively;

[0016]

[0017] Where m is the decomposition scale, m = 1, 2, ... s; s is the maximum decomposition scale; i and j are the time stamps of X and Y after coarse-graining, T is the length of the time series;

[0018] S02, coarse-grained sequence P of different decomposition scales according to process parameters X and Y m and Q m, the MTE information layer is used to calculate the causal dependency relationship MTE between process parameters X and Y with the causal relationship propagating in the direction of X→Y X→Y ;

[0019] S03, coarse-grained sequence P of different decomposition scales according to process parameters X and Y m and Q m , the MTE information layer is used to calculate the causal dependency relationship MTE between process parameters X and Y with the causal relationship propagating in the direction of Y→X Y→X ;

[0020] S04. According to MTE X→Y With MTE Y→X The magnitude relationship determines the direction of the causal relationship between process parameters X and Y. If MTE X→Y >MTE Y→X , then process parameter X is the neighbor process parameter of process parameter Y, and the causal relationship propagation direction between process parameters X and Y is X→Y, that is, the causal dependency relationship between process parameters X and Y is MTE X→Y ; If MTE X→Y <MTE Y→X , then process parameter Y is the neighbor process parameter of process parameter X, and the causal relationship propagation direction between process parameters X and Y is Y→X, that is, the causal dependency relationship between process parameters X and Y is MTE Y→X .

[0021] According to the above prediction method, preferably, in step S02, the causal dependency relationship MTE between the process parameters X and Y with the causal relationship propagating in the direction of X→Y is calculated. X→Y The specific operations are:

[0022] S021, MTE information layer calculates the coarse-grained sequence P of each decomposition scale according to formula 3 m and Q m The causal dependence between them;

[0023]

[0024] Among them, copulaTE X→Y(m) Represents the coarse-grained sequence P m and Q m The causal dependence between them; τ and t are P m and Q m The lag time, k and l are P m and Q m The embedding dimension of and The calculation formulas are shown in Equation 4 and Equation 5, respectively. C is the copula entropy, The calculation formulas are shown in Equation 6, Equation 7, and Equation 8;

[0025]

[0026] In formula 6, formula 7, and formula 8, for and The copula density function, Q m and The copula density function, for Q m and The copula density function, The calculation formulas are shown in Formula 9, Formula 10 and Formula 11 respectively;

[0027]

[0028] In formula 9, formula 10 and formula 11, yes and The copula density function, It's Q m and The copula density function, yes Q m and The copula density function;

[0029] S022. According to the weight w of each decomposition scale m The MTE information layer fuses the coarse-grained sequences P at different decomposition scales according to Equation 12. m and Q m The causal dependency between X→Y(m) , we get the causal dependency relationship MTE between process parameters X and Y, where the causal relationship propagates in the direction of X→Y X→Y ,

[0030]

[0031] According to the above prediction method, preferably, in step S03, the causal dependency relationship MTE between process parameters X and Y with the causal relationship propagation direction Y→X is calculated. Y→X The specific operations are:

[0032] S031, MTE information layer calculates the coarse-grained sequence Q of each decomposition scale according to formula 13 m and P m The causal dependence between them;

[0033]

[0034] Among them, copulaTE Y→X(m) Denotes the coarse-grained sequence Q m and P m The causal dependence between them; τ and t are P m and Q m The lag time, k and l are P m and Q m The embedding dimension of and The calculation formulas are shown in Equation 4 and Equation 5, respectively. C is the copula entropy, The calculation formulas are shown in Equation 14, Equation 15, and Equation 16;

[0035]

[0036] In formula 14, formula 15, and formula 16, yes and The copula density function, YesP m and The copula density function, yes Q m and The copula density function, The calculation formulas are shown in Equation 17, Equation 18, and Equation 19;

[0037]

[0038]

[0039] In formula 17, formula 18, and formula 19, They are and P m and and Q m as well as The copula density function;

[0040] S032. According to the weight w of each decomposition scale m The MTE information layer fuses the coarse-grained sequences Q at different scales according to Equation 20: m and P m The causal dependency between Y→X(m) , and obtain the causal dependency relationship MTE between process parameters X and Y Y→X ,

[0041]

[0042] According to the above prediction method, preferably, the embedding information generation module is used to establish a random embedding vector V for each process parameter input into the attention layer, and connect the random vector V with the process parameter; the embedding vector V is a one-dimensional matrix, through the embedding vector V, the key process parameter prediction MTEGAT model can capture the common characteristics of the industrial process under different working conditions during the training process, and then apply the learned knowledge to new operating conditions to improve the generalization ability of the model, and the embedding vector V is trained together with the key process parameter prediction MTEGAT model. The attention calculation module generates a feature representation of each input process parameter according to the aggregate representation of formula 21;

[0043]

[0044] Where Z is the feature representation of process parameter X; N(X) is the set of neighboring process parameters of process parameter X calculated by the MTE information layer, and Y is the process parameter in N(X); W E is the weight matrix; ReLU is the activation function; α X,Y is the attention coefficient after SoftMax normalization, α X,Y The calculation formula is shown in Formula 22.

[0045]

[0046] In Formula 22, LeakyReLU is a nonlinear activation function; q T is the learning coefficient vector of the attention mechanism; g X Indicates connecting the process parameter X with the embedding vector V, g X The calculation formula of g is shown in formula 23; Y Indicates connecting the process parameter Y with the embedding vector V, g Y The calculation formula is shown in Formula 24:

[0047]

[0048] In formula 23 and 24, represents the connection; V is the embedding vector, W E is the weight matrix.

[0049] According to the above prediction method, preferably, the fully connected layer calculates the feature representation of the process parameters output by the embedded attention layer using Formula 25 to obtain the prediction result of the key process parameters. Formula 25 is as follows:

[0050]

[0051] in, represents the predicted value of the key process parameters, W F is the weight matrix, b F is the bias term, and N(G) is the characteristic representation of all process parameters.

[0052] According to the above prediction method, preferably, in step S1, the pre-built key process parameter prediction MTEGAT model is trained using a training set, and the specific operation of updating the network parameters of the key process parameter prediction MTEGAT model is:

[0053] The process parameters at different times in the training set are input into the pre-built key process parameter prediction MTEGAT model for training to obtain the predicted values ​​of the key process parameters in each group of process parameters; the predicted values ​​of the key process parameters are compared with the true values ​​of the key process parameters, and the prediction loss L is calculated. k , to predict the loss L k As the loss function, the network parameters of the key process parameter prediction MTEGAT model are updated by using the back propagation method to reduce the loss function value until the loss function converges, thereby obtaining the trained key process parameter prediction MTEGAT model; wherein the prediction loss L k The calculation formula is shown in Equation 26:

[0054]

[0055] In Formula 26, n represents the number of training set samples. represents the predicted value of the key process parameter, and y represents the true value of the key process parameter.

[0056] According to the above prediction method, preferably, in step S2, the trained key process parameter prediction MTEGAT model obtained in step S1 is trained using the current process parameter data set, and the specific operation of updating the network parameters of the trained key process parameter prediction MTEGAT model is:

[0057] Input the process parameters at different times in the current process parameter data set into the trained key process parameter prediction MTEGAT model obtained in step S1 for training, and obtain the predicted value of the key process parameter in each group of process parameters; compare the predicted value of the key process parameter with the true value of the key process parameter, and calculate the prediction loss L k ; Based on the prediction loss L k and the update loss L of the important network parameters in the MTEGAT model that limits the key process parameter prediction conConstruct a loss function L; then use the back propagation method to update the network parameters of the key process parameter prediction MTEGAT model, reduce the value of the loss function L, until the loss function L converges, and obtain the updated key process parameter prediction MTEGAT model; wherein the calculation formula of the loss function L is shown in Formula 27, and the updated loss L con The calculation formula is shown in formula 28;

[0058] L=L k +L con Formula 27

[0059]

[0060] In formula 28, λ is the importance coefficient of the previous operating condition; Predict MTEGAT for the critical process parameters under current conditions

[0061] The network parameters of the embedded attention layer and the fully connected layer in the model; Predict the network parameters of the embedded attention layer and the fully connected layer in the MTEGAT model for the key process parameters under the previous working conditions; F i (k-1) yes The importance of each network parameter, F i (k-1) The calculation formula is shown in Equation 29:

[0062] In formula 29, D k-1 is the previous working condition process parameter data set, and B is a data set batch.

[0063] According to the above prediction method, preferably, in step S1, before using the previous working condition process parameter data set to train the pre-built key process parameter prediction MTEGAT model, it is necessary to perform Min-Max normalization processing on each process parameter data in the previous working condition process parameter data set to facilitate the key process parameter prediction MTEGAT model training; in step S2, before using the current working condition process parameter data set to train the trained key process parameter prediction MTEGAT model obtained in step S1, it is necessary to perform Min-Max normalization processing on each process parameter data in the current working condition process parameter data set to facilitate the key process parameter prediction MTEGAT model training. More preferably, the calculation formula of the Min-Max normalization processing is shown in Formula 30:

[0064] X norm =(XX min ) / (X max -X min ) Formula 30

[0065] In formula 30, X norm is the normalized value, X is the value of a process parameter in the data set, and X min is the minimum value of a process parameter in the data set, X max It is the maximum value of a process parameter in the data set.

[0066] A second aspect of the present invention provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the prediction method as described in the first aspect when executing the computer program.

[0067] A third aspect of the present invention provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the prediction method as described in the first aspect is implemented.

[0068] Compared with the prior art, the present invention has the following positive and beneficial effects:

[0069] (1) Considering the multi-scale characteristics of the time series data of process parameters in complex production processes, the present invention designs a multi-scale transfer entropy graph attention neural network model (MTEGAT). The MTEGAT model consists of an MTE information layer, an embedded attention layer, and a fully connected layer. The MTE information layer captures the causal dependency between production process parameters related to key process parameters through a multi-scale transfer entropy method; the embedded attention layer establishes an embedding vector V for each process parameter to represent the semantic information of each process parameter. Through the embedding vector V, the MTEGAT model can capture the common characteristics of the industrial production process under different operating conditions during the training process, and then apply the learned knowledge to new operating conditions, thereby improving the generalization ability of the MTEGAT model.

[0070] (2) In view of the frequent changes in working conditions and the continuous emergence of new working conditions, the present invention proposes a key process parameter prediction method based on the MTEGAT model. The method first uses the production process process parameter data set of the previous working condition to train the MTEGAT model to obtain a trained MTEGAT model; then collects the process parameter data set of the production process under the new working condition, uses the new working condition process parameter data set and a continuous learning method to re-train the trained MTEGAT model, updates the trained MTEGAT model, obtains an updated MTEGAT model, and uses the updated MTEGAT model to predict the key process parameters of the new working condition; the prediction method uses an elastic weight consolidation method to train the MTEGAT model, which has ensured that the MTEGAT model can retain the previously learned knowledge. When encountering a new operating condition, it can continue to learn by consolidating the knowledge of the previous working condition to ensure generalization ability, thereby solving the problem of catastrophic forgetting when the model faces multiple working conditions.

[0071] (3) The present invention uses actual continuous casting process data of steel enterprises to strictly verify the effectiveness and practicality of the method of the present invention. The experimental results show that the prediction method of the present invention has better prediction accuracy and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 A schematic diagram of the network architecture of the MTEGAT model for predicting key process parameters in Example 1 of the present invention;

[0073] Figure 2 A schematic diagram of a process for calculating the causal dependency relationship between different process parameters in a production process of the MTE information layer in Embodiment 1 of the present invention;

[0074] Figure 3 A schematic diagram of a process in which the embedded attention layer calculates the input process parameter data and the causal dependency relationship between the process parameters to obtain a feature representation of each input process parameter in Embodiment 1 of the present invention;

[0075] Figure 4 The ablation experiment verifies the R of the prediction method of key process parameters under complex environment with variable working conditions. 2 and RMSE evaluation curves;

[0076] Figure 5 is the average R of the prediction method described in Example 1 of the present invention and the prediction methods of CNN-EWC, MLP-EWC, and Trans-EWC 2 and average RMSE evaluation curve;

[0077] Figure 6The prediction method recorded in Example 1 of the present invention predicts the continuous casting process variable dependency heat map under three working conditions in the steel continuous casting production process; wherein, (a) represents working condition 1; (b) represents working condition 2; (c) represents working condition 3; the numbers in the figure represent the continuous casting production process parameter information: 1 is the pulling speed, 2 is the fixed side pulling heat, 3 is the left side pulling heat, 4 is the right side pulling heat, 5 is the pulling speed fluctuation, 6 is the moving side pulling heat, 7 is the stopper rod position, and 8 is the crystallizer liquid level fluctuation. DETAILED DESCRIPTION

[0078] The present invention is further described in detail below through specific examples, but the scope of the present invention is not limited thereto.

[0079] Embodiment 1:

[0080] A method for predicting key process parameters under a complex environment with variable working conditions comprises the following steps:

[0081] S1: Collect process parameter data of the production process at different times under previous operating conditions to obtain a previous operating condition process parameter data set, wherein the previous operating condition process parameter data set includes multiple groups of process parameters at different times, and each group of process parameters includes key process parameters; divide the previous operating condition process parameter data set into a training set and a test set in proportion, use the training set to train a pre-built key process parameter prediction MTEGAT model, update the network parameters of the key process parameter prediction MTEGAT model, and obtain the trained key process parameter prediction MTEGAT model; use the test set to test the trained key process parameter prediction MTEGAT model, evaluate the performance of the trained key process parameter prediction model MTEGAT model, until the accuracy of the trained key process parameter prediction model MTEGAT model for process parameter prediction is greater than or equal to a preset threshold, and obtain the trained key process parameter prediction MTEGAT model.

[0082] Among them, the key process parameter prediction MTEGAT model (such as Figure 1The method (shown in Figure 1) consists of an MTE information layer, an embedded attention layer and a fully connected layer. The MTE information layer is used to calculate the causal dependency between different process parameters in the production process. The input of the MTE information layer is the time series data of process parameters at different times of the production process under the same working conditions collected over a period of time, and the output of the MTE information layer is the causal dependency between the process parameters; the embedded attention layer is used to calculate the input process parameter data and the causal dependency between the process parameters to obtain the feature representation of each input process parameter. The input of the embedded attention layer is the causal dependency between the process parameter data and the process parameters, and the output of the embedded attention layer is the feature representation of the input process parameters; the fully connected layer is used to calculate the feature representation of the process parameters output by the embedded attention layer to obtain the prediction results of the key process parameters, wherein the input of the fully connected layer is the feature representation of the process parameters output by the embedded attention layer, and the output of the fully connected layer is the prediction results of the key process parameters.

[0083] S2: Collect process parameter data of the production process at different times under current operating conditions to obtain a current operating condition process parameter data set, wherein the current operating condition process parameter data set includes multiple groups of process parameters at different times, and each group of process parameters includes key process parameters; use the current operating condition process parameter data set to train the trained key process parameter prediction MTEGAT model obtained in step S1, update the network parameters of the trained key process parameter prediction MTEGAT model, and obtain an updated key process parameter prediction MTEGAT model.

[0084] S3: Inputting the process parameter data of the production process under the current working conditions into the MTEGAT model for updating the key process parameter prediction to perform prediction, obtain the prediction results of the key process parameters of the production process under the current working conditions and output them.

[0085] Furthermore, the specific steps of calculating the causal dependency between different process parameters in the production process by the MTE information layer in the key process parameter prediction MTEGAT model are as follows:

[0086] S01. Let X and Y be two different process parameters of the production process (taking the steel continuous casting production process as an example, X represents the process parameter in the steel continuous casting production process - continuous casting speed, and Y represents the process parameter in the steel continuous casting production process - continuous casting mold liquid level fluctuation), X = {x 1 ,x 2 ...x T},Y={y 1 ,y 2 ,...y T}, input the time series data of process parameters X and Y collected within a period of time into the MTE information layer, and use the MTE information layer to perform coarse-grained processing of the time series data of process parameters X and Y at different decomposition scales to obtain coarse-grained sequences P of process parameters X and Y at different decomposition scales m and Q m , P m and Q m The calculation formulas are shown in Formula 1 and Formula 2 respectively;

[0087]

[0088] Where m is the decomposition scale, m = 1, 2, ... s; s is the maximum decomposition scale; i and j are the time stamps of X and Y after coarse-graining, T is the length of the time series;

[0089] S02, coarse-grained sequence P of different decomposition scales according to process parameters X and Y m and Q m , the MTE information layer is used to calculate the causal dependency relationship MTE between process parameters X and Y with the causal relationship propagation direction X→Y X→Y Among them, the causal dependency relationship MTE between the process parameters X and Y is calculated, and the causal relationship propagation direction is X→Y X→Y Specific operations (such as Figure 2 shown) is:

[0090] S021, MTE information layer calculates the coarse-grained sequence P of each decomposition scale according to formula 3 m and Q m The causal dependence between them;

[0091]

[0092] Among them, copulaTE X→Y(m) Represents the coarse-grained sequence P m and Q m The causal dependence between them; τ and t are P m and Q m The lag time, k and l are P m and Q m The embedding dimension of and The calculation formulas are shown in Equation 4 and Equation 5, respectively. C is the copula entropy, The calculation formulas are shown in Equation 6, Equation 7, and Equation 8;

[0093]

[0094]

[0095] In formula 6, formula 7, and formula 8, for and The copula density function, Q m and The copula density function, for Q m and The copula density function, The calculation formulas are shown in Formula 9, Formula 10 and Formula 11 respectively;

[0096]

[0097] In formula 9, formula 10 and formula 11, yes and The copula density function, It's Q m and The copula density function yes Q m and The copula density function;

[0098] S022. According to the weight w of each decomposition scale m The MTE information layer fuses the coarse-grained sequences P at different decomposition scales according to Equation 12. m and Q m The causal dependency between X→Y(m) , we get the causal dependency relationship MTE between process parameters X and Y, where the causal relationship propagates in the direction of X→Y X→Y ,

[0099]

[0100] S03, coarse-grained sequence P of different decomposition scales according to process parameters X and Y m and Q m , the MTE information layer is used to calculate the causal dependency relationship MTE between process parameters X and Y with the causal relationship propagation direction Y→X Y→X Among them, the causal dependency relationship MTE between the process parameters X and Y is calculated, and the causal relationship propagation direction is Y→X Y→X The specific operations are:

[0101] S031, MTE information layer calculates the coarse-grained sequence Q of each decomposition scale according to formula 13 m and Pm The causal dependence between them;

[0102]

[0103] Among them, copulaTE Y→X(m) Denotes the coarse-grained sequence Q m and P m The causal dependence between them; τ and t are P m and Q m The lag time, k and l are P m and Q m The embedding dimension of and The calculation formulas are shown in Equation 4 and Equation 5, respectively. C is the copula entropy, The calculation formulas are shown in Equation 14, Equation 15, and Equation 16;

[0104]

[0105] In formula 14, formula 15, and formula 16, yes and The copula density function, YesP m and The copula density function, yes Q m and The copula density function, The calculation formulas are shown in Equation 17, Equation 18, and Equation 19;

[0106]

[0107] In formula 17, formula 18, and formula 19, They are and P m and and Q m as well as The copula density function;

[0108] S032. According to the weight w of each decomposition scale m The MTE information layer fuses the coarse-grained sequences Q at different scales according to Equation 20: m and P m The causal dependency between Y→X(m) , and obtain the causal dependency relationship MTE between process parameters X and Y Y→X,

[0109]

[0110] S04. According to MTE X→Y With MTE Y→X The magnitude relationship determines the direction of the causal relationship between process parameters X and Y. If MTE X→Y >MTE Y→X , then process parameter X is the neighbor process parameter of process parameter Y, and the causal relationship propagation direction between process parameters X and Y is X→Y, that is, the causal dependency relationship between process parameters X and Y is MTE X→Y ; If MTE X→Y <MTE Y→X , then process parameter Y is the neighbor process parameter of process parameter X, and the causal relationship propagation direction between process parameters X and Y is Y→X, that is, the causal dependency relationship between process parameters X and Y is MTE Y→X .

[0111] Furthermore, the embedded attention layer in the key process parameter prediction MTEGAT model includes an embedded information generation module and an attention calculation module. The embedded information generation module is used to establish a random embedding vector V for each process parameter input into the embedded attention layer, and connect the random vector V with the process parameter; the embedding vector V is a one-dimensional matrix. Through the embedding vector V, the key process parameter prediction MTEGAT model can capture the common characteristics of the industrial process under different operating conditions during the training process, and then apply the learned knowledge to new operating conditions to improve the generalization ability of the model. The embedding vector V is trained together with the key process parameter prediction MTEGAT model. The attention calculation module generates a feature representation of each input process parameter according to the aggregate representation of Formula 21 (the schematic diagram of the process of calculating the feature representation of each input process parameter by the embedded attention layer is shown in Figure 3 shown);

[0112]

[0113] In Equation 21, Z is the feature representation of process parameter X; N(X) is the set of neighboring process parameters of process parameter X calculated by the MTE information layer, and Y is the process parameter in N(X); W E is the weight matrix; ReLU is the activation function; α X,Y is the attention coefficient after SoftMax normalization, α X,Y The calculation formula is shown in Formula 22.

[0114]

[0115] In Formula 22, LeakyReLU is a nonlinear activation function; qT is the learning coefficient vector of the attention mechanism; g X Indicates connecting the process parameter X with the embedding vector V, g X The calculation formula of g is shown in formula 23; Y Indicates connecting the process parameter Y with the embedding vector V, g Y The calculation formula is shown in Formula 24:

[0116]

[0117] In formula 23 and 24, represents the connection; V is the embedding vector, W E is the weight matrix.

[0118] Furthermore, the fully connected layer calculates the feature representation of the process parameters output by the embedded attention layer using Formula 25 to obtain the prediction results of the key process parameters. Formula 17 is as follows:

[0119]

[0120] in, represents the predicted value of the key process parameters, W F is the weight matrix, b F is the bias term, and N(G) is the characteristic representation of all process parameters.

[0121] Furthermore, in step S1, the pre-built key process parameter prediction MTEGAT model is trained using the training set, and the specific operation of updating the network parameters of the key process parameter prediction MTEGAT model is:

[0122] The process parameters at different times in the training set are input into the pre-built key process parameter prediction MTEGAT model for training to obtain the predicted values ​​of the key process parameters in each group of process parameters; the predicted values ​​of the key process parameters are compared with the true values ​​of the key process parameters, and the prediction loss L is calculated. k , to predict the loss L k As the loss function, the network parameters of the key process parameter prediction MTEGAT model are updated by using the back propagation method to reduce the loss function value until the loss function converges, thereby obtaining the trained key process parameter prediction MTEGAT model; wherein the prediction loss L k The calculation formula is shown in Equation 26:

[0123]

[0124] In Formula 26, n represents the number of training set samples. represents the predicted value of the key process parameter, and y represents the true value of the key process parameter.

[0125] Furthermore, in step S2, the trained key process parameter prediction MTEGAT model obtained in step S1 is trained using the current process parameter data set, and the specific operation of updating the network parameters of the trained key process parameter prediction MTEGAT model is:

[0126] Input the process parameters at different times in the current process parameter data set into the trained key process parameter prediction MTEGAT model obtained in step S1 for training, and obtain the predicted value of the key process parameter in each group of process parameters; compare the predicted value of the key process parameter with the true value of the key process parameter, and calculate the prediction loss L k ; Based on the prediction loss L k and the update loss L of the important network parameters in the MTEGAT model that limits the key process parameter prediction con Construct a loss function L; then use the back propagation method to update the network parameters of the key process parameter prediction MTEGAT model, reduce the value of the loss function L, until the loss function L converges, and obtain the updated key process parameter prediction MTEGAT model; wherein the calculation formula of the loss function L is shown in Formula 27, and the updated loss L con The calculation formula is shown in formula 28;

[0127] L=L k +L con Formula 27

[0128]

[0129] In formula 28, λ is the importance coefficient of the previous operating condition; Predict MTEGAT for the critical process parameters under current conditions

[0130] The network parameters of the embedded attention layer and the fully connected layer in the model; Predict the network parameters of the embedded attention layer and the fully connected layer in the MTEGAT model for the key process parameters under the previous working conditions; F i (k-1) yes The importance of each network parameter, F i (k-1) The calculation formula is shown in Equation 29:

[0131] In formula 29, D k-1 is the previous working condition process parameter data set, and B is a data set batch.

[0132] Furthermore, in step S1, before using the previous working condition process parameter data set to train the pre-built key process parameter prediction MTEGAT model, it is necessary to perform Min-Max normalization processing on each process parameter data in the previous working condition process parameter data set to facilitate the key process parameter prediction MTEGAT model training; in step S2, before using the current working condition process parameter data set to train the trained key process parameter prediction MTEGAT model obtained in step S1, it is necessary to perform Min-Max normalization processing on each process parameter data in the current working condition process parameter data set to facilitate the key process parameter prediction MTEGAT model training. The calculation formula for the Min-Max normalization processing is shown in Formula 30:

[0133] X norm =(XX min ) / (X max -X min ) Formula 30.

[0134] In formula 30, X norm is the normalized value, X is the value of a process parameter in the data set, and X min is the minimum value of a process parameter in the data set, X max It is the maximum value of a process parameter in the data set.

[0135] Embodiment 2:

[0136] An electronic device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the prediction method as described in the above-mentioned embodiment 1 when executing the computer program.

[0137] Embodiment 3:

[0138] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the prediction method as described in the above-mentioned embodiment 1.

[0139] Verification of the prediction effect of the prediction method for key process parameters under complex environment with variable working conditions recorded in Example 1 of the present invention:

[0140] In order to verify the effect of the method for predicting key process parameters under complex and variable operating conditions on the prediction of key process parameters in complex industrial production processes, the actual continuous casting production process of a steel company was selected for experimental testing.

[0141] Sensors are used to collect data on eight process parameters at different production moments in the steel continuous casting production process, including continuous casting speed, speed fluctuation, stopper rod position, continuous casting mold level fluctuation, and the amount of heat to be drawn in four directions of the mold (the amount of heat to be drawn in one direction is one process parameter). Among them, the continuous casting mold level fluctuation is the key process parameter, and the data on continuous casting speed, speed fluctuation, stopper rod position, continuous casting mold level fluctuation, and the amount of heat to be drawn in four directions of the mold at the same moment are a set of process parameter samples. Based on the importance and representativeness of the actual production process, three operating conditions with continuous casting speeds of 1.2 m / min, 1.0 m / min, and 1.4 m / min were selected for analysis. According to the occurrence time of the three conditions, the condition with a casting speed of 1.2 m / min was set as the existing condition, and the other two conditions were set as new conditions. The data sample size of the process parameters collected under each operating condition is shown in Table 1. Since the data of different process parameters of continuous casting have different amplitudes, the Min-Max normalization method (the calculation formula is shown in Formula 22 in Example 1) is used to normalize the collected process parameter data to facilitate model training.

[0142] Table 1 Number of samples under different operating conditions

[0143] Working conditions 1(1.2m / min) 2(1.0m / min) 3(1.4m / min) Sample size 2704 1226 861

[0144] The prediction method for key process parameters under complex variable working conditions described in Example 1 of the present invention is used to predict the key process parameters in the steel continuous casting production process under the new working conditions, namely, the liquid level fluctuation of the continuous casting mold. In order to evaluate the effectiveness of the prediction method, two widely recognized indicators, root mean square error (RMSE) and determination coefficient (R 2 ) is used as an indicator to evaluate the effectiveness of the prediction method, among which RMSE and R 2 The calculation formulas are shown in Equation 31 and Equation 32 respectively.

[0145]

[0146] Where m is the number of continuous casting mold level fluctuation test sets, is the predicted value of the continuous casting mold liquid level fluctuation, y i is the true value of the continuous casting mold liquid level fluctuation, is the average value.

[0147] In order to compare with the prediction method of key process parameters under complex environment with variable working conditions recorded in Example 1 of the present invention, three comparative prediction methods are set up, namely, MTEGAT prediction method, GAT-EWC prediction method and TEGAT-EWC prediction method.

[0148] Among them, the MTEGAT prediction method is basically the same as the prediction method recorded in Example 1 of the present invention, and the difference is that the MTEGAT prediction method omits step S2 in the prediction method recorded in Example 1 of the present invention, and in step S3, the process parameter data of the production process under the current operating conditions is directly input into the trained key process parameter prediction MTEGAT model obtained in step S1 for prediction, and the prediction results of the key process parameters of the production process under the current operating conditions are obtained and output.

[0149] The GAT-EWC prediction method is basically the same as the prediction method described in Example 1 of the present invention, except that: the key process parameter prediction MTEGAT model in the prediction method described in Example 1 of the present invention is replaced by the GAT model; the GAT model is the model obtained by deleting the MTE information layer from the key process parameter prediction MTEGAT model described in Example 1 of the present invention, that is, the GAT model consists of an embedded attention layer and a fully connected layer; the process parameters of the production process are input into the embedded attention layer of the GAT model, and the feature representation of the process parameters is directly calculated according to formula (13), and when calculating the feature representation of the process parameters according to formula (13), except for the process parameters themselves, all other process parameters are regarded as neighbor process parameters of the process parameters.

[0150] The TEGAT-EWC prediction method is basically the same as the prediction method recorded in Example 1 of the present invention, except that: the key process parameter prediction MTEGAT model in the prediction method recorded in Example 1 of the present invention is replaced by the TEGAT model; the TEGAT model replaces the MTE information layer in the key process parameter prediction MTEGAT model recorded in Example 1 of the present invention with the TE information layer, that is, the TEGAT model consists of a TE information layer, an embedded attention layer, and a fully connected layer; when the TE information layer in the TEGAT model calculates the causal dependency between different process parameters in the production process, the process parameter data X, Y are directly used to calculate the causal dependency. The specific operation of the TEGAT model using the TE information layer to calculate the causal dependency between different process parameters in the production process is:

[0151] 1) Calculate the causal dependency copulaTE between process parameters X and Y with the causal relationship propagating in the direction of X→Y X→Y The specific operations are:

[0152] S021. Calculate the causal dependency relationship between X and Y according to Formula 33;

[0153]

[0154] Among them, copulaTE X→Y represents the causal dependency between X and Y; u and h are the lag times of X and Y, r and v are the embedding dimensions of X and Y, q and pv are the timestamps of X and Y, and The calculation formulas are shown in Equation 34 and Equation 35 respectively. C is the copula entropy, H C The calculation formula of (X,Y) is shown in Equation 36;

[0155]

[0156]

[0157] In Equation 36, c(X,Y) is the copula density function, and the calculation formula of c(X,Y) is shown in Equation 37;

[0158]

[0159] In Equation 37, C(X,Y) is the copula function of X and Y;

[0160] 2) Calculate the causal dependency relationship copulaTE between process parameters X and Y with the causal relationship propagating in the direction of Y→X Y→X The specific operations are:

[0161] Calculate the causal dependency between Y and X according to Equation 38

[0162]

[0163] Among them, copulaTE Y→X represents the causal dependency between Y and X; u and h are the lag times of X and Y, r and v are the embedding dimensions of X and Y, q and pv are the timestamps of X and Y, and The calculation formulas are shown in Equation 34 and Equation 35, respectively. C (Y,X) is the copula entropy, H C The calculation formula is shown in formula 39;

[0164]

[0165] In formula 39, c(Y,X) is the copula density function, and the calculation formula of c(Y,X) is shown in formula 40;

[0166]

[0167] In Equation 40, C(Y,X) is the copula function of Y and X;

[0168] 3) Compare copulaTE X→Y With copulaTE Y→X If the size relationship of copulaTE X→Y >copulaTE Y→X , then process parameter X is the neighbor process parameter of process parameter Y, and the causal relationship propagation direction between process parameters X and Y is X→Y, that is, the causal dependency relationship between process parameters X and Y is copulaTE X→Y ; If copulaTE X→Y <copulaTE Y→X , then process parameter Y is the neighbor process parameter of process parameter X, and the causal relationship propagation direction between process parameters X and Y is Y→X, that is, the causal dependency relationship between process parameters X and Y is copulaTE Y→X .

[0169] The prediction method described in Example 1 of the present invention, the MTEGAT prediction method, the GAT-EWC prediction method, and the TEGAT-EWC prediction method were subjected to ablation experiments to compare the prediction results of the four prediction methods. The results are shown in Table 2 and Figure 4 shown.

[0170] Table 2 Ablation experiment results

[0171]

[0172] From Table 2 and Figure 4 It can be seen that the prediction method described in Example 1 of the present invention is effective in R 2 The values ​​are 71.35%, 71.04% and 81.65% respectively, which are higher than the R of GAT-EWC under the same operating conditions. 2 In working condition 2, the R of the prediction method described in Example 1 of the present invention is 2It increased by about 15.6%. In addition, under all three operating conditions, the RMSE of the prediction method recorded in Example 1 of the present invention is lower than the GAT-EWC value under the same operating conditions. The experimental results compared with the GAT-EWC method show that the use of the MTE information layer in the MTEGAT model to predict key process parameters to calculate the causal dependencies between different process parameters in the production process can improve the accuracy of the model, mainly because the MTE information layer enables the model to extract higher quality information from adjacent nodes during information transmission. In addition, under the same operating conditions, the RMSE values ​​of the prediction method recorded in Example 1 of the present invention are lower than the RMSE of TEGAT-EWC, indicating that considering the multi-scale features present in the data can better capture the dependencies between the continuous casting process parameters, thereby improving the accuracy of the model. Compared with the MTEGAT method, the RMSE of the prediction method recorded in Example 1 of the present invention is lower under the other two conditions except for condition 1, and the RMSE and R 2 The reason for the sameness under working condition 1 is that the data under working condition 1 is regarded as the existing working condition data to train the basic model. Under working conditions 2 and 3, the prediction method R described in Example 1 of the present invention 2 The above experimental results show that after adding the continuous learning module, by retaining the knowledge of the existing working conditions, better accuracy can be obtained when training new working conditions. At the same time, it can be seen that although the performance of the MTEGAT method under working condition 3 is slightly worse than that of the GAT-EWC method, the performance of MTEGAT under all three conditions is more stable, which indirectly proves that accurate knowledge can make the prediction results more stable.

[0173] In order to further evaluate the prediction performance of the prediction method recorded in Example 1 of the present invention, the present invention compares the prediction method recorded in Example 1 with the three existing deep learning prediction methods, and the three deep learning prediction methods all add the same continuous learning mechanism. The three deep learning prediction methods are: CNN-EWC prediction method, MLP-EWC prediction method, and Trans-EWC prediction method. The three prediction methods of CNN-EWC, MLP-EWC, and Trans-EWC are basically the same as the method recorded in Example 1 of the present invention. The difference is that the key process parameter prediction MTEGAT model in the prediction method recorded in Example 1 of the present invention is replaced by a convolutional neural network model (CNN), a multi-layer perceptron model (MLP), and a Transformer model, respectively. Based on the samples collected in Table 1, the CNN-EWC prediction method, the MLP-EWC prediction method, the Trans-EWC prediction method, and the prediction method recorded in Example 1 of the present invention are respectively used to predict the process parameter under the new working condition-the liquid level fluctuation of the continuous casting crystallizer. Each prediction method carried out 10 groups of experiments, and the average of the 10 groups of experiments was taken as the experimental result. The RMSE and R of the four prediction methods2 The statistical results are shown in Table 3.

[0174] Table 3 RMSE and R of the prediction method described in Example 1 of the present invention and the prediction methods of CNN-EWC, MLP-EWC, and Trans-EWC 2 Comparison results

[0175]

[0176] It can be seen from Table 3 that under the three operating conditions, the RMSE of the prediction method recorded in Example 1 of the present invention is 15.53% to 34.9%, 5.68% to 18.18% and 3.43% to 15.26% lower than other methods, respectively, indicating that the prediction method recorded in Example 1 of the present invention can achieve better results in predicting the crystallizer liquid level fluctuation. In addition, it can be seen from Table 3 that when the CNN-EWC, MLP-EWC or Trans-EWC method is used directly, the results under the three operating conditions are very poor, which may be due to the high quality requirements of the deep learning model for data. In addition, under the three operating conditions, the R 2 The values ​​are all greater than 70%, which can meet the production requirements of the continuous casting process.

[0177] In order to compare the generalization ability and robustness of the prediction method described in Example 1 of the present invention with the CNN-EWC, MLP-EWC, and Trans-EWC prediction methods, R under three working conditions was calculated in 10 experiments. 2 and the median of RMSE. The experimental results are as follows Figure 5 As shown. Figure 5 It can be seen that the prediction method described in Example 1 of the present invention not only has high accuracy, but also has good stability.

[0178] In addition, the prediction method described in Example 1 of the present invention can use the attention coefficient embedded in the attention layer in the key process parameter prediction MTEGAT model to explain the prediction results. Figure 6 The heat map of the attention coefficient of the prediction method described in Example 1 of the present invention under three operating conditions is shown; Figure 6It can be seen that when the pulling speed is 1m / min, the main factors affecting the fluctuation of the crystallizer liquid level are the fixed side pulling heat and the left side pulling heat. In addition, the moving side pulling heat has a significant effect on the fixed side pulling heat and the left side pulling heat, which has an indirect effect on the fluctuation of the crystallizer liquid level. Therefore, when the pulling speed is 1m / min, if the crystallizer liquid level fluctuation is abnormal, it is necessary to focus on the above three variables. When the pulling speed increases to 1.2m / min, the fixed side pulling heat becomes the main factor affecting the crystallizer liquid level fluctuation. As the pulling speed further increases to 1.4m / min, the main factor affecting the crystallizer liquid level fluctuation becomes the moving side pulling heat. Through the analysis of the above results, some insights can be provided for actual continuous casting production, helping engineers make more targeted decisions, thereby improving production efficiency and billet quality.

Claims

1. A method for predicting key process parameters under complex and variable working conditions, characterized in that: The following steps are involved: S1: collecting process parameter data of the production process at different times under previous working conditions to obtain a previous working condition process parameter data set, wherein the previous working condition process parameter data set includes multiple groups of process parameters at different times, and each group of process parameters includes key process parameters; The previous working condition process parameter data set is divided into a training set and a test set in proportion, and the pre-built key process parameter prediction MTEGAT model is trained using the training set, and the network parameters of the key process parameter prediction MTEGAT model are updated to obtain the trained key process parameter prediction MTEGAT model; the trained key process parameter prediction MTEGAT model is tested using the test set to evaluate the performance of the trained key process parameter prediction model MTEGAT model, until the accuracy of the trained key process parameter prediction model MTEGAT model for process parameter prediction is greater than or equal to a preset threshold, thereby obtaining a trained key process parameter prediction MTEGAT model; S2: collecting process parameter data of the production process at different times under current working conditions to obtain a current working condition process parameter data set, wherein the current working condition process parameter data set includes multiple groups of process parameters at different times, and each group of process parameters includes key process parameters; The trained key process parameter prediction MTEGAT model obtained in step S1 is trained using the current operating condition process parameter data set, and the network parameters of the trained key process parameter prediction MTEGAT model are updated to obtain an updated key process parameter prediction MTEGAT model; S3: Inputting the process parameter data of the production process under the current working conditions into the MTEGAT model for updating the key process parameter prediction to perform prediction, obtain the prediction results of the key process parameters of the production process under the current working conditions and output them.

2. The prediction method according to claim 1, characterized in that: In step S1, the key process parameter prediction MTEGAT model consists of an MTE information layer, an embedded attention layer and a fully connected layer; the MTE information layer is used to calculate the causal dependency between different process parameters in the production process, the input of the MTE information layer is the time series data of process parameters of the production process at different times under the same working conditions collected over a period of time, and the output of the MTE information layer is the causal dependency between the process parameters; The embedded attention layer is used to calculate the causal dependency relationship between the input process parameter data and the process parameters to obtain the feature representation of each input process parameter, the input of the embedded attention layer is the causal dependency relationship between the process parameter data and the process parameters, and the output of the embedded attention layer is the feature representation of the input process parameters; the fully connected layer is used to calculate the feature representation of the process parameters output by the embedded attention layer to obtain the prediction results of the key process parameters, wherein the input of the fully connected layer is the feature representation of the process parameters output by the embedded attention layer, and the output of the fully connected layer is the prediction results of the key process parameters.

3. The prediction method according to claim 2, characterized in that: The steps of calculating the causal dependency between different process parameters in the production process at the MTE information layer are as follows: S01, let X and Y be two different process parameters of the production process, X={x1,x2...x T }, Y={y1,y2,...y T }, input the time series data of process parameters X and Y collected within a period of time into the MTE information layer, and use the MTE information layer to perform coarse-grained processing of the time series data of process parameters X and Y at different decomposition scales to obtain coarse-grained sequences P of process parameters X and Y at different decomposition scales m and Q m , P m and Q m The calculation formulas are shown in Formula 1 and Formula 2 respectively; Where m is the decomposition scale, m = 1, 2, ... s; s is the maximum decomposition scale; i and j are the time stamps of X and Y after coarse-graining, T is the length of the time series; S02, coarse-grained sequence P of different decomposition scales according to process parameters X and Y m and Q m , the MTE information layer is used to calculate the causal dependency relationship MTE between process parameters X and Y with the causal relationship propagation direction X→Y X→Y ; S03, coarse-grained sequence P of different decomposition scales according to process parameters X and Y m and Q m , the MTE information layer is used to calculate the causal dependency relationship MTE between process parameters X and Y with the causal relationship propagation direction Y→X Y→X ; S04. According to MTE X→Y With MTE Y→X The magnitude relationship determines the direction of the causal relationship between process parameters X and Y. If MTE X→Y >MTE Y→X , then process parameter X is the neighbor process parameter of process parameter Y, and the causal relationship propagation direction between process parameters X and Y is X→Y, that is, the causal dependency relationship between process parameters X and Y is MTE X→Y ; If MTE X→Y <MTE Y→X , then process parameter Y is the neighbor process parameter of process parameter X, and the causal relationship propagation direction between process parameters X and Y is Y→X, that is, the causal dependency relationship between process parameters X and Y is MTE Y→X .

4. The prediction method according to claim 3, characterized in that: In step S02, the causal dependency relationship MTE between process parameters X and Y with the causal relationship propagating in the direction of X→Y is calculated. X→Y The specific operations are: S021, MTE information layer calculates the coarse-grained sequence P of each decomposition scale according to formula 3 m and Q m The causal dependence between them; Among them, copulaTE X→Y(m) Represents the coarse-grained sequence P m and Q m The causal dependence between them; τ and t are P m and Q m The lag time, k and l are P m and Q m The embedding dimension of and The calculation formulas are shown in Equation 4 and Equation 5, respectively. C is the copula entropy, The calculation formulas are shown in Equation 6, Equation 7, and Equation 8; In formula 6, formula 7, and formula 8, for and The copula density function, Q m and The copula density function, for Q m and The copula density function, The calculation formulas are shown in Formula 9, Formula 10 and Formula 11 respectively; In formula 9, formula 10 and formula 11, yes and The copula density function, It's Q m and The copula density function, yes Q m and The copula density function; S022. According to the weight w of each decomposition scale m The MTE information layer fuses the coarse-grained sequences P at different decomposition scales according to Equation 12. m and Q m The causal dependency between X→Y(m) , we get the causal dependency relationship MTE between process parameters X and Y, where the causal relationship propagates in the direction of X→Y X→Y , In step S03, the causal dependency relationship MTE between process parameters X and Y with the causal relationship propagating in the direction of Y→X is calculated. Y→X The specific operations are: S031, MTE information layer calculates the coarse-grained sequence Q of each decomposition scale according to formula 13 m and P m The causal dependence between them; Among them, copulaTE Y→X(m) Denotes the coarse-grained sequence Q m and P m The causal dependence between them; τ and t are P m and Q m The lag time, k and l are P m and Q m The embedding dimension of and The calculation formulas are shown in Equation 4 and Equation 5, respectively. C is the copula entropy, The calculation formulas are shown in Equation 14, Equation 15, and Equation 16; In formula 14, formula 15, and formula 16, yes and The copula density function, YesP m and The copula density function, yes Q m and The copula density function, The calculation formulas are shown in Equation 17, Equation 18, and Equation 19; In formula 17, formula 18, and formula 19, They are and P m and and Q m as well as The copula density function; S032. According to the weight w of each decomposition scale m The MTE information layer fuses the coarse-grained sequences Q at different scales according to Equation 20: m and P m The causal dependency between Y→X(m) , we can get the causal dependency relationship MTE between process parameters X and Y Y→X , 5. The prediction method according to any one of claims 1 to 4, characterized in that: The embedded attention layer includes an embedded information generation module and an attention calculation module; the embedded information generation module is used to establish a random embedding vector V for each process parameter input into the embedded attention layer, and connect the random vector V with the process parameter; the attention calculation module generates a feature representation of each input process parameter according to the aggregate representation of formula 21; Where Z is the feature representation of process parameter X; N(X) is the set of neighboring process parameters of process parameter X calculated by the MTE information layer, and Y is the process parameter in N(X); W E is the weight matrix; ReLU is the activation function; α X,Y is the attention coefficient after SoftMax normalization, α X,Y The calculation formula is shown in Formula 22. In Formula 22, LeakyReLU is a nonlinear activation function; q T is the learning coefficient vector of the attention mechanism; g X Indicates connecting the process parameter X with the embedding vector V, g X The calculation formula of g is shown in formula 23; Y Indicates connecting the process parameter Y with the embedding vector V, g Y The calculation formula is shown in Equation 16 In formula 23 and 24, represents the connection; V is the embedding vector, W E is the weight matrix.

6. The prediction method according to claim 5, characterized in that: The fully connected layer calculates the feature representation of the process parameters output by the embedded attention layer using Formula 25 to obtain the prediction results of the key process parameters. Formula 25 is as follows: in, represents the predicted value of the key process parameters, W F is the weight matrix, b F is the bias term, and N(G) is the characteristic representation of all process parameters.

7. The prediction method according to claim 6, characterized in that: In step S1, the pre-built key process parameter prediction MTEGAT model is trained using the training set, and the specific operation of updating the network parameters of the key process parameter prediction MTEGAT model is as follows: The process parameters at different times in the training set are input into the pre-built key process parameter prediction MTEGAT model for training to obtain the predicted values ​​of the key process parameters in each group of process parameters; the predicted values ​​of the key process parameters are compared with the true values ​​of the key process parameters, and the prediction loss L is calculated. k , to predict the loss L k As the loss function, the network parameters of the key process parameter prediction MTEGAT model are updated by using the back propagation method, and the loss function value is reduced until the loss function converges, thereby obtaining the trained key process parameter prediction MTEGAT model; Among them, the prediction loss L k The calculation formula is shown in Equation 26: In formula 18, n represents the number of training set samples. represents the predicted value of the key process parameter, and y represents the true value of the key process parameter.

8. The prediction method according to claim 7, characterized in that: In step S2, the trained key process parameter prediction MTEGAT model obtained in step S1 is trained using the current process parameter data set, and the specific operation of updating the network parameters of the trained key process parameter prediction MTEGAT model is as follows: Input the process parameters at different times in the current process parameter data set into the trained key process parameter prediction MTEGAT model obtained in step S1 for training, and obtain the predicted value of the key process parameter in each group of process parameters; compare the predicted value of the key process parameter with the true value of the key process parameter, and calculate the prediction loss L k ; Based on the prediction loss L k and the update loss L of the important network parameters in the MTEGAT model that limits the key process parameter prediction con Construct a loss function L; then use the back propagation method to update the network parameters of the key process parameter prediction MTEGAT model, reduce the value of the loss function L, until the loss function L converges, and obtain the updated key process parameter prediction MTEGAT model; wherein the calculation formula of the loss function L is shown in Formula 27, and the updated loss L con The calculation formula is shown in formula 28; L=L k +L con Formula 27 In formula 20, λ is the importance coefficient of the previous operating condition; Predicting the network parameters of the embedded attention layer and the fully connected layer in the MTEGAT model for the key process parameters under the current working conditions; Predict the network parameters of the embedded attention layer and the fully connected layer in the MTEGAT model for the key process parameters under the previous working conditions; F i (k-1) yes The importance of each network parameter in F i (k-1) The calculation formula is shown in Equation 29: In formula 29, D k-1 is the previous working condition process parameter data set, and B is a data set batch.

9. An electronic device, comprising a memory and a processor, characterized in that: The memory stores a computer program, and when the processor executes the computer program, the prediction method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the prediction method according to any one of claims 1 to 7 is implemented.

Images