A modeling method for industrial soft sensors based on coupled bidirectional information filtering long short-term memory network

Abstract

The present invention discloses an industrial soft sensor modeling method based on a coupled bidirectional information filtering long short-term memory network, which comprises at least the following steps: step S1: collecting a data set of a soft measurement task and performing preprocessing; step S2: creating a coupled bidirectional information filtering long short-term memory network model and performing supervised training on a training set; step S3: using the coupled bidirectional information filtering long short-term memory network model to test a test set. The technical solution of the present invention adopts a coupled bidirectional information filtering long short-term memory network (CB‑IF‑LSTM), which expands the depth of the neural network layer from input to output without increasing the parameter scale of the original recurrent neural network; an information filtering unit and a long short-term memory network unit are used on the basic unit, and when hidden state information enters the long short-term memory network unit, unimportant information is suppressed by the information filtering unit, and important hidden state information passes normally, so that the quality variable prediction accuracy is improved, and the ability to cope with unpredictable situations is improved.

Description

Technical Field

[0001] The invention belongs to the technical field of soft measurement modeling, and in particular relates to an industrial soft sensor modeling method based on a coupled bidirectional information filtering long short-term memory network. Background Art

[0002] In modern industrial processes, quality variables are key factors in industrial control and optimization, and these quality variables can usually be obtained through sensors. However, due to high costs, lags, or maintenance, it is not always convenient to directly obtain some quality variables. At the same time, some process variables are easy to obtain, and they are correlated with quality variables, so it is feasible to predict quality variables through process variables. This prediction method is called soft measurement, and the sensor using soft measurement is called soft sensor.

[0003] Generally, soft sensor models are divided into first principle models, data-driven models, and hybrid models. With the development of technology, on the one hand, a large amount of data collected from industrial sites is stored, and on the other hand, deep learning technology has made progress, which makes it more reasonable to use deep learning technology for soft sensor modeling. Generally, soft sensor components based on deep learning include autoencoders (AE), convolutional neural networks (CNN), recurrent neural networks (RNN), etc. Among them, recurrent neural networks can adapt well to the time series data of soft sensors. However, standard recurrent neural networks have the problem of gradient disappearance or gradient explosion. More commonly used are its variants such as long short-term memory networks (LSTM) and gated recurrent units (GRU). Standard recurrent neural networks and their variants are collectively referred to as recurrent neural networks. See Figure 1 , shown is a schematic diagram of a recurrent neural network.

[0004] The mechanism of action of LSTM and GRU is similar. The following uses LSTM as an example to describe the mechanism of action. Figure 2 , which is a schematic diagram of the internal structure of the long short-term memory network unit. The long short-term memory network is cyclically calculated in the time dimension. It is composed of long short-term memory network units at different time steps and shares parameters. The calculation process of one long short-term memory network unit is as follows:

[0005] f t =σ(w fh h t-1 +w fx x t +b f )

[0006] i t =σ(w ih h t-1 +w ix x t +b i)

[0007] o t =σ(w oh h t-1 +w ox x t +b o )

[0008]

[0009]

[0010] h t =o t ⊙tanh(c t )

[0011] where f t It is the forget gate, t is the input gate, o t is the output gate, c t is the cell state, h t is the hidden state, w and b are the corresponding weights and biases, x t is the component of the sequence data at time step t, ⊙ represents element-wise multiplication, σ represents the sigmoid function with an output range of [0,1], and tanh represents a nonlinear function with an output range of [-1,1]. The cell state is used to store long-term information, and the three gate structures are used to control the forgetting and updating of the cell state. These improvements are used to overcome the gradient explosion or gradient disappearance problem of the standard RNN. From the outside, the long short-term memory network receives x at each time step. t , c t-1 ,h t-1 As input, c t ,h t is the output.

[0012] Bidirectional recurrent neural network is a commonly used recurrent neural network structure, which consists of two long short-term memory networks in opposite directions, so that information from both directions can be used when outputting at each time. Figure 3 , shown is a schematic diagram of a bidirectional recurrent neural network.

[0013] However, the existing recurrent neural networks, including long short-term memory networks, are designed based on standard recurrent neural networks for long-term dependency problems and lack adaptive adjustments for soft measurement problems. Therefore, the existing soft measurement methods based on recurrent neural networks have the following main problems:

[0014] (1) From the perspective of industrial application scenarios, when soft-measuring quality variables, soft sensors always first obtain those process variables that are easy to measure, and then predict the quality variables based on the process variables. Therefore, before predicting the quality variables, the process variables can be predicted first, and the prediction error of the process variables can be calculated. It is beneficial to incorporate the prediction error of the process variables into the hidden state conversion process. The existing soft-measuring methods based on recurrent neural networks do not explicitly apply this feature.

[0015] (2) From the perspective of deep learning development, deep neural networks have more advantages than shallow neural networks. However, at each time step, the existing soft measurement methods based on recurrent neural networks process the input data by shallow neural networks, which limits the performance of soft sensors. Although RNN variants such as LSTM and GRU have achieved performance improvements compared to standard RNNs, they all solve the problem of gradient vanishing or gradient exploding from the perspective of information updating and forgetting, and do not solve the problem of shallow neural network layers faced by standard recurrent neural networks.

[0016] Therefore, in view of the defects of the prior art, it is necessary to propose a technical solution to solve the technical problems existing in the prior art. Summary of the invention

[0017] In order to overcome the technical defects of the prior art, the invention proposes an industrial soft sensor modeling method based on a coupled bidirectional information filtering long short-term memory network. The method adopts a coupled bidirectional information filtering long short-term memory network (CB-IF-LSTM). The soft sensor based on CB-IF-LSTM takes process variables as input and quality variable prediction values ​​as output. A coupled bidirectional structure is adopted in the network structure. On the basis of not increasing the parameter scale of the original recurrent neural network, the mapping from input to hidden state is no longer completed by a fixed shallow neural network, but becomes a variable process determined by the training process, thereby expanding the depth of the neural network layer from input to output. Information filtering units and long short-term memory network units are adopted in the basic units. When the hidden state information enters the long short-term memory network unit, unimportant information is suppressed by the information filtering unit, and important hidden state information passes normally, so that the prediction accuracy of quality variables is improved, and the ability to cope with unpredictable situations is improved.

[0018] In order to solve the technical problems existing in the prior art, the technical solution of the present invention is as follows:

[0019] A method for modeling an industrial soft sensor based on a coupled bidirectional information filtering long short-term memory network includes at least the following steps:

[0020] Step S1: Collect the data set of the soft measurement task and perform preprocessing;

[0021] Step S2: Create a coupled bidirectional information filtering long short-term memory network model and use it in the training set Conducting supervised training;

[0022] Step S3: Testing the test set using the coupled bidirectional information filtering long short-term memory network model;

[0023] Wherein, step S2 comprises the following steps:

[0024] Step S21: The training set The process variable subsequence in is input into the coupled bidirectional information filtering long short-term memory network and outputs the predicted value of the corresponding quality variable; when predicting the sample of time step t, the network inputs the process variable subsequence {x t-k+1 ,x t-k+2 ,…,x t}, the network calculates and outputs the quality variable y of the corresponding subsequence t The predicted value y t ; The calculation is completed by two recurrent neural network layers: reverse loop and forward loop. The calculation method is as follows:

[0025]

[0026]

[0027]

[0028] Among them, y t is the quality variable y t The predicted value of represents the reverse long short-term memory network layer, represents the forward information filtering long short-term memory network layer, σ represents the sigmoid function and the output range is [0,1];

[0029] Step S22: Calculate the objective function loss, and use the optimization algorithm to solve the model parameters that minimize the objective function, and store the coupled bidirectional information filtering long short-term memory network model and the optimal parameters. The objective function calculation method is as follows:

[0030]

[0031] Where j∈Train represents the calculation of all samples in the training set, and m is the number of samples in the training set.

[0032] As a further improved technical solution, step S1 further includes the following steps:

[0033] Step S11: Collect data D = {X, Y} = {(x i ,y i)}, where i∈1,2,…,n, n is the number of samples collected, x i is the process variable at time step i, y i is the mass variable at time step i;

[0034] Step S12: preprocessing the collected data set; wherein step S12 includes:

[0035] Step S121: Window processing is performed on the collected sequence data to obtain a series of shorter subsequences. in, is the subsequence of process variables corresponding to the jth time step, is the quality variable corresponding to the jth time step, j∈k,k+1,…,n, k is the window size;

[0036] Step S122: randomly divide the windowed sequence into 1:1 proportions to obtain a training set With the test set

[0037] As a further improved technical solution, step S21 further includes:

[0038] Step S211: Substituting the process variable subsequence {x t-k+1 ,x t-k+2 ,…,x t} Input the reverse recurrent neural network layer. The reverse recurrent calculation process is a reverse cycle along the time dimension. The calculation process is as follows:

[0039]

[0040] in is the hidden state of the reverse LSTM network layer output, symbol Indicates that the sequence is reversely calculated, and the reverse cycle calculation is completed by a long short-term memory network layer;

[0041] Step S212: The hidden state output by the reverse long short-term memory network layer As the input of the forward LSTM network layer, the forward loop calculation process is a forward loop along the time dimension, and its calculation process is as follows:

[0042]

[0043] in, represents the hidden state of the output of the positive information filtering long short-term memory network layer, It indicates that the sequence is calculated in a forward cycle. The forward calculation process is carried out along the time dimension and completed by the information filtering unit and the long short-term memory network unit.

[0044] Step S213: The hidden state of the last time step of the forward layer And predict the quality variable through a single-layer neural network. The calculation method is as follows:

[0045]

[0046] Where σ represents the sigmoid function and the output range is [0,1], y t is the quality variable y t The predicted value of

[0047] Step S212 further includes:

[0048] Step S2121: At each time step, before the information filtering unit acts on the long short-term memory network unit, it filters the hidden state entering the long short-term memory network unit. The calculation process is as follows:

[0049]

[0050]

[0051]

[0052] in, and are the hidden states of the forward layer before and after filtering, respectively, t-k+l is the filter attenuation coefficient of the current time step, z t-k+l Represents the mapped l = 1, 2…, k, k is the length of the windowed subsequence, σ represents the sigmoid function with an output range of [0, 1], and tanh represents a nonlinear function with an output range of [-1, 1]. The process is described as follows:

[0053] (1) The information filtering unit receives input at time step t-k+l Then, a single-layer neural network is used to Map to and Similar space, the mapping result is z t-k+l ;

[0054] (2) and z t-k+l Subtract the difference, and then perform nonlinear mapping. The number of nodes is the same as Same, f t-k+l is the result of nonlinear mapping;

[0055] (3) f t-k+l As a filter coefficient, it filters and updates the latent variables. represents the hidden state after filtering;

[0056] Step S2122: Use replace The long short-term memory network unit is passed in, and the long short-term memory network unit outputs the hidden state h of the t time step through its calculation method. t :

[0057]

[0058] Among them, LSTMcell(·) represents the working mode of the long short-term memory network unit.

[0059] As a further improved technical solution, step S3 includes:

[0060] Step S31: creating a stored coupled bidirectional information filtering long short-term memory network model and loading the optimal parameters;

[0061] Step S32: Test set Input coupled bidirectional information filtering long short-term memory network model, the model outputs the test quality variable prediction value in the manner described in step S21;

[0062] Step S33: Calculate the root mean square error (RMSE) of the test set, and the calculation method is as follows:

[0063]

[0064] Where j∈Test means calculation for all samples in the test set, and m is the number of samples in the test set.

[0065] Compared with the prior art, the present invention has the following technical effects:

[0066] (1) The present invention adopts a coupled bidirectional structure. On the basis of not increasing the parameter scale of the original recurrent neural network, the mapping from input to hidden state is no longer completed by a fixed shallow neural network, but becomes a variable process determined by the training process, which expands the depth of the neural network layer from input to output. That is, from input x t To hidden state The mapping is completed by multiple neural networks with different depths, which effectively expands the depth of the neural network.

[0067] (2) Using information filtering recurrent neural network, when the industrial process changes within a predictable range, the hidden state h t-1 For input x t Make a prediction, that is, t To make a prediction, then h t-1 -z t To some extent, it shows the predictability of time series. IFU can be calculated based on h t-1 -zt Learning about changes in industrial processes. When the process changes drastically, the hidden state h t-1 The impact on the prediction of the current quality variable is weakened, and IFU adjusts h according to the learning results. t-1 Attenuation occurs. When the process changes relatively smoothly, h t-1 The impact on the prediction of the current quality variable is more important, and IFU does not need to t-1 Therefore, through IFU, important hidden state information enters the recurrent neural network unit normally, unimportant hidden information is suppressed, and the prediction accuracy of quality variables is improved; at the same time, the recurrent neural network benefits from the filtering mechanism of IFU, and its ability to cope with unpredictable situations is improved.

[0068] (3) The technical solution of the present invention can be applied to the debutanizer soft sensor to achieve accurate prediction of butane concentration. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 Schematic diagram of the recurrent neural network structure.

[0070] Figure 2 Schematic diagram of the internal structure of a long short-term memory network unit.

[0071] Figure 3 Schematic diagram of the bidirectional recurrent neural network structure.

[0072] Figure 4 This is a structural block diagram of the coupled bidirectional information filtering long short-term memory network of the present invention.

[0073] Figure 5 It is a structural block diagram of the coupled bidirectional structure long short-term memory network of the present invention.

[0074] Figure 6 It is a schematic diagram of the principle of the information filtering unit of the present invention.

[0075] Figure 7 It is a combination of the information filtering unit and the long short-term memory unit of the present invention.

[0076] Figure 8 This is the windowing processing of the data set of the present invention (taking k=3 as an example).

[0077] Fig. 9 This is a schematic diagram of the effect of the coupled bidirectional long short-term memory network technology of the present invention.

[0078] Fig.10 It is a schematic diagram of the technical effect of the information filtering unit of the present invention.

[0079] The following specific embodiments will further illustrate the present invention in conjunction with the above-mentioned drawings. DETAILED DESCRIPTION

[0080] The technical solution provided by the present invention will be further described below in conjunction with the accompanying drawings.

[0081] Example 1

[0082] The present invention proposes an industrial soft sensor modeling method based on a coupled bidirectional information filtering long short-term memory network, and applies it to a debutanizer soft sensor to achieve accurate prediction of butane concentration. The debutanizer is an important component of the desulfurization and naphtha separation unit in the oil production process, and is used to separate C5, C3 and C4. Since butane concentration (quality variable) has a great impact on the tower and is difficult to measure directly, it requires strict detection and control, so a soft sensor is used to predict it. The input of the soft sensor is a variable related to the butane concentration (process variable).

[0083] See also Figure 4 , which is a network structure diagram of the industrial soft sensor modeling method based on the coupled bidirectional information filtering recurrent neural network of the present invention, wherein a coupled bidirectional information filtering long short-term memory network (CB-IF-LSTM) is adopted. The soft sensor based on CB-IF-LSTM takes the process variable as input and the quality variable prediction value as output. CB-IF-LSTM adopts a coupled bidirectional structure in the network structure, with the first layer being the reverse layer and the second layer being the forward layer. Figure 5 As shown in the figure, the reverse layer unit uses LSTM unit, and the hidden state is not filtered and transmitted to the long short-term memory network unit; the forward layer uses information filtering unit (IFU) and LSTM combination unit (IFU-LSTM), and the hidden state is filtered and transmitted to the long short-term memory network unit by the information filtering unit, where the information filtering unit is as shown in the figure. Figure 6 , the combination of information filtering unit and long short-term memory unit is as follows Figure 7 CB-IF-LSTM outputs the hidden state of the last time step And the quality variables are predicted through a single-layer neural network.

[0084] More specifically, the industrial soft sensor modeling method based on the coupled bidirectional information filtering long short-term memory network of the present invention comprises the following steps:

[0085] Step S1: collect the data set for the soft measurement task of butane concentration in debutanizer and perform preprocessing;

[0086] Step S1 includes:

[0087] Step S11: Collect data D = {X, Y} = {(x i ,y i )}, where i∈1,2,…,n, n is the number of samples collected, x iis the process variable at time step i, y i is the mass variable (butane concentration) at time step i;

[0088] Step S12: preprocessing the collected data set; Step S12 includes:

[0089] Step S121: Figure 8 As shown (the figure takes k=3 as an example), the window size k=10 is selected, and the collected sequence data is windowed to obtain a series of shorter subsequences. where j∈k,k+1,…,n, is the subsequence of process variables corresponding to the jth time step, is the mass variable corresponding to time step j;

[0090] Step S122: randomly divide the windowed sequence into 1:1 proportions to obtain a training set With the test set

[0091] Step S2: Create a coupled bidirectional information filtering long short-term memory network model for soft measurement and use it in the training set Conducting supervised training;

[0092] Step S2 includes:

[0093] Step S21: The training set The process variable subsequence in is input into the coupled bidirectional information filtering long short-term memory network and outputs the predicted value of the corresponding quality variable; when predicting the sample of time step t, the network inputs the process variable subsequence {x t-k+1 ,x t-k+2 ,…,x t}, the network calculates and outputs the quality variable y of the corresponding subsequence t The predicted value y t ; The calculation is completed by two recurrent neural network layers: reverse loop and forward loop. The calculation method is as follows:

[0094]

[0095]

[0096]

[0097] where y t is the quality variable y t The predicted value of represents the reverse long short-term memory network layer, represents the forward information filtering long short-term memory network layer, and the other symbols are similar to the previous ones;

[0098] Step S21 includes:

[0099] Step S211: Substituting the process variable subsequence {x t-k+1 ,x t-k+2 ,…,x t} Input the reverse recurrent neural network layer. The reverse recurrent calculation process is a reverse cycle along the time dimension. The calculation process is as follows:

[0100]

[0101] in is the hidden state of the reverse LSTM network layer output, symbol Indicates that the sequence is reversely calculated, and the reverse cycle calculation is completed by a long short-term memory network layer;

[0102] Step S212: The hidden state output by the reverse long short-term memory network layer As the input of the forward LSTM network layer, the forward loop calculation process is a forward loop along the time dimension, and its calculation process is as follows:

[0103]

[0104] in Represents the hidden state of the output of the positive information filtering long short-term memory network layer, symbol It indicates that the sequence is calculated in a forward cycle. Here, the forward calculation process is carried out along the time dimension and completed by the information filtering unit (IFU) and the long short-term memory network unit (LSTM cell);

[0105] Step S212 includes:

[0106] Step S2121: At each time step, before the information filtering unit acts on the long short-term memory network unit, it filters the hidden state entering the long short-term memory network unit. The calculation process is as follows:

[0107]

[0108]

[0109]

[0110] The notation here is similar to that of long short-term memory networks, where and are the hidden states of the forward layer before and after filtering, respectively, t-k+l is the filter attenuation coefficient of the current time step, where z t-k+l Represents the mapped l = 1, 2…, k, k is the length of the windowed subsequence, and the text description of this process is as follows:

[0111] 1) The information filtering unit receives input of t-k+l time steps Then, a single-layer neural network is used to Map to and Similar spaces (with the same number of nodes), the mapping result is z t-k+l ;

[0112] 2) and z t-k+l Subtract them (i.e., compare them), and then perform nonlinear mapping on their differences. The number of nodes is Same, here f t-k+l is the result of nonlinear mapping;

[0113] 3) f t-k+l As a filter coefficient to filter and update the latent variables, here represents the hidden state after filtering;

[0114] Step S2122: Use replace The long short-term memory network unit is passed in, and the long short-term memory network unit outputs the hidden state h of the t time step through its calculation method. t :

[0115]

[0116] Among them, LSTMcell(·) represents the action mode of long short-term memory network unit;

[0117] Step S213: The hidden state of the last time step of the forward layer And predict the quality variable through a single-layer neural network. The calculation method is as follows:

[0118]

[0119] where y t is the quality variable y t The predicted value of , σ represents the sigmoid function and the output range is [0,1];

[0120] Step S22: Calculate the objective function loss, and use the optimization algorithm to solve the model parameters that minimize the objective function (i.e., train the model to obtain the optimal parameters), and store the coupled bidirectional information filtering long short-term memory network model and the optimal parameters. The objective function calculation method is as follows:

[0121]

[0122] Where j∈Train represents the calculation of all samples in the training set, and m is the number of samples in the training set;

[0123] Step S3: using a coupled bidirectional information filtering long short-term memory network model to test the debutanizer butane concentration test set;

[0124] Step S3 includes:

[0125] Step S31: creating a stored coupled bidirectional information filtering long short-term memory network model and loading the optimal parameters;

[0126] Step S32: Test set Input coupled bidirectional information filtering long short-term memory network model, the model outputs the test quality variable prediction value in the manner described in step S21;

[0127] Step S33: Calculate the root mean square error (RMSE) of the test set, and the calculation method is as follows:

[0128]

[0129] Where j∈Test means calculation for all samples in the test set, and m is the number of samples in the test set.

[0130] In the above technical solution, the coupled bidirectional structure is adopted, which can make the mapping from input to hidden state no longer completed by a fixed shallow neural network, but become a variable process determined by the training process without increasing the parameter scale of the original recurrent neural network, thus expanding the depth of the neural network layer from input to output. The technical effect is shown as follows Fig. 9 As shown, from the input x t To hidden state The mapping is completed by multiple neural networks with different depths, which effectively expands the depth of the neural network.

[0131] In addition, the present invention adopts information filtering recurrent neural network, and its technical effect is shown as follows Fig.10 As shown, when the industrial process changes within a predictable range, the hidden state h t-1 For input x t Make a prediction, that is, t To make a prediction, then h t-1 -z t To some extent, it shows the predictability of time series. IFU can be calculated based on h t-1 -z t Learning about changes in industrial processes. When the process changes drastically, the hidden state h t-1 The impact on the prediction of the current quality variable is weakened, and IFU adjusts h according to the learning results. t-1Attenuation occurs. When the process changes relatively smoothly, h t-1 The impact on the prediction of the current quality variable is more important, and IFU does not need to t-1 Therefore, through IFU, important hidden state information enters the recurrent neural network unit normally, unimportant hidden information is suppressed, and the prediction accuracy of quality variables is improved; at the same time, the recurrent neural network benefits from the filtering mechanism of IFU, and its ability to cope with unpredictable situations is improved.

[0132] The above embodiments are only used to help understand the method and core idea of ​​the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made to the present invention without departing from the principles of the present invention, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

[0133] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for modeling industrial soft sensors based on coupled bidirectional information filtering long short-term memory networks, characterized in that: At least the following steps are included: Step S1: Collect the data set of the soft measurement task and perform preprocessing; Step S2: Create a coupled bidirectional information filtering long short-term memory network model and use it in the training set Conducting supervised training; Step S3: Use the coupled bidirectional information filtering long short-term memory network model to test the test set; Wherein, step S2 includes the following steps: Step S21: The training set The process variable subsequence in the coupled bidirectional information filtering long short-term memory network is input and outputs the predicted value of the corresponding quality variable; when predicting the sample of time step t, the network inputs the process variable subsequence {x t-k+1 ,x t-k+2 ,…,x t }, the network calculates and outputs the quality variable y of the corresponding subsequence t The predicted value y t The calculation is completed by two recurrent neural network layers: reverse loop and forward loop. The calculation method is as follows: Among them, y t is the quality variable y t The predicted value of represents the reverse long short-term memory network layer, stands for the forward information filtering long short-term memory network layer; Step S22: Calculate the objective function loss, and use the optimization algorithm to solve the model parameters that minimize the objective function, and store the coupled bidirectional information filtering long short-term memory network model and the optimal parameters. The objective function is calculated as follows: Where j∈Train represents the calculation of all samples in the training set, and m is the number of samples in the training set; Step S1 includes the following steps: Step S11: Collect data D = {X, Y} = {(x i ,y i )}, where i∈1,2,…,n, n is the number of samples collected, x i is the process variable at time step i, y i is the mass variable at time step i; Step S12: pre-processing the collected data set; wherein step S12 includes: Step S121: Window processing is performed on the collected sequence data to obtain a series of shorter subsequences. in, is the subsequence of process variables corresponding to time step j, is the quality variable corresponding to time step j, j∈k,k+1,…,n, k is the window size; Step S122: randomly divide the windowed sequence into 1:1 ratios to obtain a training set With the test set Step S21 includes: Step S211: Sub-sequence {x t-k+1 ,x t-k+2 ,…,x t Input the reverse recurrent neural network layer. The reverse recurrent calculation process is a reverse cycle along the time dimension. The calculation process is as follows: in is the hidden state output by the reverse long short-term memory network layer, symbol Indicates that the sequence is subjected to reverse loop calculation, which is completed by a long short-term memory network layer; Step S212: The hidden state output by the reverse long short-term memory network layer As the input of the forward network layer, the forward cycle calculation process is a forward cycle along the time dimension, and its calculation process is as follows: in, represents the hidden state of the output of the positive information filtering long short-term memory network layer, It represents the forward cycle calculation of the sequence. The forward calculation process is carried out along the time dimension and completed by the information filtering unit and the long short-term memory network unit. Step S213: The hidden state of the last time step of the forward layer And the quality variable is predicted by a single-layer neural network; the calculation method is as follows: where y t is the quality variable y t The predicted value of , σ represents the sigmoid function and the output range is [0,1]; Step S212 further includes: Step S2121: At each time step, the information filtering unit filters the hidden state entering the long short-term memory network unit before acting on the long short-term memory network unit. The calculation process is as follows: in, and are the hidden states of the forward layer before and after filtering, f t-k+l is the filter attenuation coefficient of the current time step, z t-k+l Represents the mapped l = 1, 2…, k, where k is the length of the windowed subsequence, σ represents the sigmoid function with an output range of [0, 1], and tanh represents a nonlinear function with an output range of [-1, 1]. The process is described as follows: (1) The information filtering unit receives the input of t-k+l time steps Then, a single-layer neural network is used to Mapped to and Similar space, the mapping result is z t-k+l ; (2) and z t-k+l Perform subtraction operation, and then perform nonlinear mapping on the difference, and the number of nodes is the same as Same, f t-k+l is the result of nonlinear mapping; (3) f t-k+l As a filter coefficient to filter and update the latent variables, Represents the hidden state after filtering; Step S2122: Use replace The long short-term memory network unit is input, and the long short-term memory network unit outputs the hidden state h of the t time step through its calculation method. t : Where LSTMcell(·) represents the mode of action of the long short-term memory network unit; Step S3 includes: Step S31: creating a stored coupled bidirectional information filtering long short-term memory network model and loading the optimal parameters; Step S32: Test set Input the coupled bidirectional information filtering long short-term memory network model, and the model outputs the test quality variable prediction value in the manner described in step S21; Step S33: Calculate the root mean square error (RMSE) of the test set. The calculation method is as follows: Where j∈Test represents the calculation of all samples in the test set, and m is the number of samples in the test set.

Images