A method for predicting moisture content of export materials using temporal convolutional neural network

Abstract

The present invention discloses a method for predicting the moisture content of outlet materials using a temporal convolutional neural network. The method comprises: collecting historical process data from a process manufacturing production line to construct a multivariate time series dataset; preprocessing the multivariate time series dataset to obtain preprocessed data, and dividing the preprocessed data into a training set and a test set according to the order of the acquisition cycle; introducing a dual attention mechanism into the temporal convolutional neural network to construct a DA-TCN prediction model; training the DA-TCN model using the training set to obtain a trained DA-TCN prediction model; and validating the test set using the trained DA-TCN prediction model to obtain a validated DA-TCN prediction model for use in predicting the moisture content of outlet materials. The present invention introduces a dual attention mechanism into the temporal convolutional neural network to construct a DA-TCN prediction model for predicting the moisture content of outlet materials, and further applies this model to predict the moisture content of outlet materials in the thin sheet drying process.

Description

Technical Field

[0001] The present invention relates to a method for predicting the moisture content of export materials by using a temporal convolutional neural network, and belongs to the technical field of time series prediction. Background Art

[0002] In modern complex process production, product quality prediction uses historical production data to mine the mapping relationship between various quality influencing factors and product performance. It is a reflection of technological innovation and an effective way to improve product quality, increase product stability, and reduce production costs. It plays a key role in process production.

[0003] Product quality in complex production processes is closely linked to process parameters, equipment operating parameters, and real-time operating condition changes. The complex, nonlinear relationship between the moisture content of the exiting material and process parameters in the thin-sheet drying process hinders the accurate prediction of the moisture content of the exiting material, hindering the improvement of the intelligent production system in the process workshop. Current product quality prediction methods for complex production processes are often complex, making it difficult to fully exploit the implicit and historical time series information in historical production indicator data. Therefore, accurately predicting product quality using production data generated by complex manufacturing processes remains a challenge that needs to be addressed. Summary of the Invention

[0004] The present invention provides a method for predicting the moisture content of outlet materials using a temporal convolutional neural network, which is used to introduce a dual attention mechanism into the temporal convolutional neural network to build a DA-TCN prediction model for predicting the moisture content of outlet materials, and further used for predicting the moisture content of outlet materials in the thin plate drying process.

[0005] The technical solution of the present invention is:

[0006] According to a first aspect of the present invention, a method for predicting the moisture content of export materials using a temporal convolutional neural network is provided, the method comprising the following steps:

[0007] S1. Collect historical process data of the process manufacturing production line and construct a multivariate time series data set; the multivariate time series data set includes process parameter index data and export material moisture index data;

[0008] S2. Preprocess the multivariate time series data set to obtain preprocessed data, and divide the preprocessed data into a training set and a test set according to the order of the acquisition cycle;

[0009] S3. Introducing the dual attention mechanism into the temporal convolutional neural network to build the DA-TCN prediction model;

[0010] S4. Use the training set to train the DA-TCN model to obtain a trained DA-TCN prediction model; use the trained prediction model DA-TCN to verify the test set to obtain a verified DA-TCN prediction model for use in predicting the moisture content of export materials.

[0011] Furthermore, the multivariate time series data set is preprocessed to obtain preprocessed data, and the preprocessed data is divided into a training set and a test set according to the order of the acquisition cycle, including:

[0012] According to the material head and tail interception rules of the process production process, the process parameter data and quality indicator data are judged to determine the critical position of the material head and tail, so as to retain the data between the critical positions of the material head and tail;

[0013] Then, the data between the critical positions of the material head and tail are retained and the abnormal data are eliminated using the 3σ method, and normalized.

[0014] The normalized data is divided into training set and test set according to the preset ratio.

[0015] Furthermore, the method of using a temporal convolutional neural network to predict the moisture content of the outlet material is applied to the thin plate drying process. In the thin plate drying process, the rules for cutting the head and tail of the material during the production process are as follows:

[0016] If the inlet material moisture reference value is ≤ n1%, the position of the material head and tail is determined as follows: according to the batch order of the collection cycle, the first data of the outlet material moisture greater than m1% in the current batch process data is used as the material head node, and the data of the outlet material moisture after a delay of t is used as the material head critical position; the first data of the outlet material moisture less than m1% in the current batch process data is used as the material tail node, and the data after the tail node is shifted forward by t is used as the material tail critical position;

[0017] If the inlet material moisture reference value is greater than n1% and ≤ n2%, the material head and tail positions are determined as follows: according to the batch order of the acquisition cycle, the first data in the current batch process data with the outlet material moisture greater than m2% is used as the material head node, and the data after a delay of t from the material head node is used as the material head critical position; the first data in the current batch process data with the outlet material moisture less than m2% is used as the material tail node, and the data after a delay of t from the material tail node is used as the material tail critical position;

[0018] If the reference value of moisture content of the inlet material is greater than n2%, the method for determining the position of the head and tail of the material is as follows: according to the batch order of the acquisition cycle, the first data in the process data of the current batch in which the moisture content of the outlet material is greater than m3% is used as the head node, and the data after a delay of t from the head node is used as the critical position of the head; the first data in the process data of the current batch in which the moisture content of the outlet material is less than m3% is used as the tail node, and the data after a delay of t from the tail node is used as the critical position of the tail.

[0019] Furthermore, the DA-TCN prediction model includes a feature attention mechanism FA, a temporal convolutional neural network TCN, a temporal attention mechanism TA, a ReLU layer, and a fully connected layer;

[0020] The pre-processed process production parameter index characteristic variables and the export material moisture index characteristic variables are used as the input of the feature attention mechanism FA, and the feature weighted matrix X' is obtained by weighted summation of the attention weights of each index feature and the input feature;

[0021] The obtained feature weight matrix X' and the hidden layer output of the temporal convolutional neural network (TCN) at the previous time step are used as the hidden layer input of the temporal convolutional neural network (TCN) at the next moment. The dilated causal convolution in the temporal convolutional neural network (TCN) is used to achieve long sequence interval sampling of all indicators. The PReLU function is used to capture the long-term temporal dependencies between indicators. The residual block of the temporal convolutional neural network (TCN) is used to extract features of all sequences in the input data.

[0022] The output of the TCN hidden layer after iterating the residual block of the temporal convolutional neural network (TCN) to time t is used as the input of the temporal attention mechanism TA. By comparing the time attention weights of each historical moment in the process production process, the multivariate key temporal features affecting the moisture index of the export material are determined. All weights are weighted and summed with the hidden layer output of each historical moment to obtain the comprehensive temporal information; then, through the ReLU layer and the fully connected layer, the moisture prediction value of the export material is obtained.

[0023] According to a second aspect of the present invention, a device for predicting the moisture content of outlet materials using a time convolutional neural network is provided, comprising a module for executing the method for predicting the moisture content of outlet materials using a time convolutional neural network taking into account the current fed out of the grid-side converter GSC as described in any of the above items.

[0024] According to a third aspect of the present invention, a terminal device is provided, comprising a memory, a processor, and a program stored on the memory and executable by the processor, wherein when the processor executes the program, the method for predicting the moisture content of export materials using a temporal convolutional neural network as described in any one of the above is implemented.

[0025] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, wherein the computer-readable storage medium includes a stored program, wherein when the program is running, the device where the computer-readable storage medium is located is controlled to execute any one of the above-mentioned methods for predicting the moisture content of export materials using a temporal convolutional neural network.

[0026] The beneficial effects of the present invention are:

[0027] 1. The present invention preprocesses the collected multivariate time series data, including screening the process data of the process production process and detecting and eliminating outliers using the 3σ method, so that the multivariate time series data is stable and reasonable; through normalization processing, all process data are mapped to the [0,1] interval to eliminate the influence of the characteristic dimension differences between the data and improve the calculation efficiency.

[0028] 2. In order to extract process parameter features more accurately, the present invention introduces a dual attention mechanism to determine the multivariate key time series features that affect product quality. Through the ability of adaptive feature extraction, the model helps to simultaneously extract the intrinsic connection between time series relationships and different feature variables, select key information for processing, strengthen the influence of process feature sequences and time series with higher correlation on the prediction results, and weaken the influence of sequences with lower correlation.

[0029] 3. The present invention constructs a DA-TCN prediction model. Compared with the traditional GRU model, TCN model, and LSTM-Attention model, the prediction accuracy of the present invention is more accurate, which realizes the accurate prediction of process production process quality and provides a reference for the accurate prediction of process manufacturing production line product quality.

[0030] From the above, it can be seen that the present invention addresses the problem that the complex nonlinear relationship between process parameters in the process production process makes it difficult to accurately predict the process quality. A process production quality prediction method based on DA-TCN has been established, which provides a reference for the accurate prediction of product quality in complex process production processes. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 Schematic diagram of the process of the present invention;

[0032] Figure 2 This is a structural diagram of the DA-TCN prediction model of the present invention;

[0033] Figure 3 This is a diagram of the internal unit structure of the temporal convolutional neural network TCN of the present invention;

[0034] Figure 4 This is the internal unit structure diagram of the feature attention mechanism FA of the present invention;

[0035] Figure 5 This is the internal unit structure diagram of the temporal attention mechanism TA of the present invention;

[0036] Figure 6 Part of the process data of the thin plate drying step of the silk thread production process in the embodiment of the present invention;

[0037] Figure 7 Schematic diagram of a fitting curve comparing predicted values ​​and actual values ​​of the quality indicators of the thin-plate tofu drying process predicted on the test set using the DA-TCN prediction model in an embodiment of the present invention;

[0038] Figure 8 Schematic diagram of fitting curves comparing predicted values ​​and actual values ​​of various comparative experiments in the embodiments of the present invention. DETAILED DESCRIPTION

[0039] To make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. It should be noted that, in the absence of conflict, the embodiments in this application and the features in the embodiments can be combined with each other in any way.

[0040] Example 1: Figure 1-8 As shown, according to a first aspect of an embodiment of the present invention, a method for predicting the moisture content of export materials using a temporal convolutional neural network is provided, comprising the following steps:

[0041] S1. Collect historical process data of the process manufacturing production line and construct a multivariate time series data set; the multivariate time series data set includes process parameter index data and export material moisture index data;

[0042] S2. Preprocess the multivariate time series data set to obtain preprocessed data, and divide the preprocessed data into a training set and a test set according to the order of the acquisition cycle;

[0043] S3. Introducing the dual attention mechanism into the temporal convolutional neural network to build the DA-TCN prediction model;

[0044] S4. Use the training set to train the DA-TCN model to obtain a trained DA-TCN prediction model; use the trained prediction model DA-TCN to verify the test set to obtain a verified DA-TCN prediction model;

[0045] S5. Based on the verified DA-TCN prediction model, the moisture content of the export material is predicted based on the input process data of the process manufacturing production line.

[0046] Furthermore, the multivariate time series data set is preprocessed to obtain preprocessed data, and the preprocessed data is divided into a training set and a test set according to the order of the acquisition cycle, including: judging the head and tail of the process parameter data and the quality indicator data according to the head and tail interception rules of the process production process, determining the critical positions of the head and tail of the material, and retaining the data between the critical positions of the head and tail of the material; then using the 3σ method to eliminate abnormal data from the data retained between the critical positions of the head and tail of the material, and performing normalization processing; dividing the normalized data into a training set and a test set according to a preset ratio.

[0047] Furthermore, the normalized data is divided into a training set and a test set according to a preset ratio, specifically: the process data is divided into a ratio of 7:3, and each set of data consists of The composition includes m-1 dimensional process parameters and 1 dimensional quality index, wherein the quality index is the moisture content of the outlet material.

[0048] Furthermore, the method of using a temporal convolutional neural network to predict the moisture content of the outlet material is applied to the thin plate drying process. In the thin plate drying process, the rules for cutting the head and tail of the material during the production process are as follows:

[0049] If the inlet material moisture reference value is ≤ n1%, the position of the material head and tail is determined as follows: according to the batch order of the collection cycle, the first data of the outlet material moisture greater than m1% in the current batch process data is used as the material head node, and the data of the outlet material moisture after a delay of t is used as the material head critical position; the first data of the outlet material moisture less than m1% in the current batch process data is used as the material tail node, and the data after the tail node is shifted forward by t is used as the material tail critical position;

[0050] If the inlet material moisture reference value is greater than n1% and ≤ n2%, the material head and tail positions are determined as follows: according to the batch order of the acquisition cycle, the first data in the current batch process data with the outlet material moisture greater than m2% is used as the material head node, and the data after a delay of t from the material head node is used as the material head critical position; the first data in the current batch process data with the outlet material moisture less than m2% is used as the material tail node, and the data after a delay of t from the material tail node is used as the material tail critical position;

[0051] If the reference value of moisture content of the inlet material is greater than n2%, the method for determining the position of the head and tail of the material is as follows: according to the batch order of the acquisition cycle, the first data in the process data of the current batch in which the moisture content of the outlet material is greater than m3% is used as the head node, and the data after a delay of t from the head node is used as the critical position of the head; the first data in the process data of the current batch in which the moisture content of the outlet material is less than m3% is used as the tail node, and the data after a delay of t from the tail node is used as the critical position of the tail.

[0052] Optionally, the parameters involved above may be as shown in Table 1 below:

[0053] Table 1

[0054]

[0055] n1: low-level process quality standard. When the process quality reference value is less than or equal to n1, it means that the batch of materials belongs to low-level materials;

[0056] n2: high-level process quality standard. When the quality index setting value is greater than or equal to n2, it means that the moisture content of this batch of materials belongs to high-level materials.

[0057] When the process quality reference value of process production is between n1 and n2, it means that the batch of materials is at a medium or high level.

[0058] m1: Standard value of quality index when the material is at a low level;

[0059] m2: Standard value of quality index when the material is at medium or high level;

[0060] m3: Standard value of quality index when the material is at a high level;

[0061] t: The buffer time from the shutdown of production equipment to the full recovery of normal working conditions.

[0062] Complex process production processes usually have multi-batch characteristics. The process data collected from multiple batches are complex, large, and have obvious differences. When the production equipment is in the start-up and shutdown states, critical values ​​of the process data at the beginning and end of the materials will be generated, thereby generating a large number of abnormal values. Therefore, before establishing a prediction model, it is necessary to analyze, screen, and eliminate the data at the beginning and end of the materials according to the beginning and end interception rules of the process production process to obtain an accurate data set.

[0063] Furthermore, if Figure 2-Figure 5 As shown, the DA-TCN prediction model includes a feature attention mechanism FA, a temporal convolutional neural network TCN, a temporal attention mechanism TA, a ReLU layer, and a fully connected layer;

[0064] The pre-processed process production parameter index characteristic variables and the export material moisture index characteristic variables are used as the input of the feature attention mechanism FA. The feature weighted matrix X' is obtained by weighted summing the attention weights of each indicator feature and the input feature. This feature weighted matrix reflects the coupling dependency relationship between different indicator feature variables.

[0065] The obtained feature weight matrix X' and the hidden layer output of the temporal convolutional neural network (TCN) at the previous time step are used as the hidden layer input of the temporal convolutional neural network (TCN) at the next moment. The dilated causal convolution in the temporal convolutional neural network (TCN) is used to achieve long sequence interval sampling of all indicators. The PReLU function is used to capture the long-term temporal dependencies between indicators. The residual block of the temporal convolutional neural network (TCN) is used to extract features of all sequences in the input data.

[0066] The output of the TCN hidden layer after iterating the residual block of the temporal convolutional neural network (TCN) to time t is used as the input of the temporal attention mechanism TA. By comparing the time attention weights of each historical moment in the process production process, the multivariate key temporal features affecting the moisture index of the export material are determined. All weights are weighted and summed with the hidden layer output of each historical moment to obtain the comprehensive temporal information; then, through the ReLU layer and the fully connected layer, the moisture prediction value of the export material is obtained.

[0067] Furthermore, the preprocessed data is normalized, wherein the normalization formula is: Where X * is the process data after normalization; X is the original process data; X max and X min are the maximum and minimum values ​​of the original data respectively.

[0068] Specifically:

[0069] The feature attention weight The calculation formula is as follows:

[0070]

[0071]

[0072] Where, is the value of the mth (1≤m≤24) indicator feature at time t (1≤t≤T); is the feature attention vector of the mth index; W e is the weight coefficient; b is the bias coefficient; σ(·) is the Sigmoid activation function, which is used here to stabilize the output range between 0 and 1.

[0073] The feature weight matrix calculation formula is as follows:

[0074]

[0075] Where, Represents the attention weight value of the mth indicator; Represents the value of the mth indicator feature at time t (1≤t≤T).

[0076] The residual block calculation formula is as follows:

[0077] X h =δ(G(X h-1 )+X h-1 )

[0078] Where: δ(·) is the activation function; G(·) represents the transformation of DCC, weight normalization, activation function, and random inactivation; X h-1 is the output of the h-1th residual block; X h is the output of the hth residual block.

[0079] The temporal attention weight calculation formula is as follows:

[0080]

[0081]

[0082] Where, is the temporal attention vector of each feature vector; is the output of the hidden layer of the TCN network at time t; W d is the weight coefficient; b d is the bias coefficient; ReLU(·) represents the activation function.

[0083] The calculation formula of the time series feature weight matrix is ​​as follows:

[0084]

[0085] In the formula, in the formula, represents the temporal attention weight at time t; Represents the output of the hidden layer of the TCN network at time t.

[0086] Furthermore, the mean absolute error MAE, root mean square error RMSE, mean square error MSE, goodness of fit R 2 As an evaluation indicator, the calculation formula is as follows:

[0087]

[0088] Where n is the number of samples; y i is the true value of the process quality indicator; is the mean value of process quality index; is the predicted value of the process quality index; reference goodness of fit R 2 To judge the fitting effect of the model, R 2 The closer it is to 0, the worse the model fitting effect is.

[0089] According to a second aspect of an embodiment of the present invention, a device for predicting the moisture content of export materials using a time convolutional neural network is provided, including a module for executing the method for predicting the moisture content of export materials using a time convolutional neural network taking into account the feed-out current of the grid-side converter GSC as described in any of the above items, specifically including: a first module for executing S1: collecting historical process data of a process manufacturing production line and constructing a multivariate time series data set; wherein the multivariate time series data set includes process parameter index data and export material moisture index data; a second module for executing S2: preprocessing the multivariate time series data set to obtain preprocessed data, and The preprocessed data is divided into a training set and a test set in a periodic order; the third module is used to execute S3: introduce a dual attention mechanism into the temporal convolutional neural network to build a DA-TCN prediction model; the fourth module is used to execute S4: use the training set to train the DA-TCN model to obtain a trained DA-TCN prediction model; use the trained prediction model DA-TCN to verify the test set to obtain a verified DA-TCN prediction model; the fifth module is used to execute S5: predict the moisture content of export materials based on the process data of the input process manufacturing production line according to the verified DA-TCN prediction model.

[0090] According to a third aspect of an embodiment of the present invention, a terminal device is provided, comprising a memory, a processor, and a program stored on the memory and executable by the processor, wherein when the processor executes the program, the method for predicting the moisture content of export materials using a temporal convolutional neural network as described in any one of the above is implemented.

[0091] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium including a stored program. When the program is executed, the device containing the computer-readable storage medium is controlled to execute any of the aforementioned methods for predicting export material moisture using a temporal convolutional neural network. In an exemplary embodiment, the computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard drive, a magnetic disk, or an optical disk.

[0092] Example 2: An optional implementation of the present invention is described in detail below, but the content of the invention is not limited to the scope of the embodiment.

[0093] The experiment was conducted on the thin plate drying process in a certain process of silk production to predict the moisture content of the outlet material, a quality indicator of this process. Figure 2The process production quality prediction method based on DA-TCN of the present invention comprises the following steps:

[0094] Step (1): Use a high-frequency sensor to collect and compile the original process data every 1s and store it in the OPC server. Extract the historical process data of the thin plate drying process between 16:17:00 on October 17, 2022 and 14:40:00 on October 18, 2022 from the server as a data set. By evaluating the correlation between process parameters and quality indicators, select the 23-dimensional feature variables with the strongest correlation with the moisture content of the outlet material as process parameters, including the actual value of the moisture content of the inlet material, the humidity of the dehumidification wind, and the set value of the moisture content of the outlet material. , HT inlet material temperature, exhaust air temperature, exhaust air door opening setting value, exhaust air door start and stop times, discharge hood negative pressure, hot air temperature setting value, front room air door actual value, ambient temperature, HT inlet steam flow, inlet thin plate steam pressure actual value, steam quality temperature, cylinder wall temperature setting value, cylinder wall temperature measurement average value, cylinder wall temperature pressure conversion value, thin plate condensate temperature, radiator condensate temperature, cylinder wall temperature (front) measurement value, cylinder wall temperature (middle) measurement value, cylinder wall temperature (back) measurement value, outlet material temperature, quality index is the actual value of outlet material moisture. There are 20107 groups of extracted raw data, some of which are as follows Figure 6 shown.

[0095] Step (2): When the tow-drying equipment is in the start-up and stop states, it will generate critical values ​​of the process data for the head and tail of the material, thereby generating a large number of abnormal values. According to the head and tail interception rules of the process production process, the head and tail characteristics of the material in each batch of data are analyzed to determine the critical position of the head and tail of the material, and the data at the head and tail of the material are screened and eliminated. The interception rules for the head and tail data of the moisture content of the export material are shown in Table 2.

[0096] Table 2

[0097]

[0098] According to Table 2, the material head and tail are intercepted, and the data between the critical position of the material head and the critical position of the material tail are retained, and then the abnormal data are eliminated using the 3σ method.

[0099] Step (3): Apply the formula to the preprocessed data Perform normalization processing, where X * is the process data after normalization; X is the original process data of the drying process; X max and X min are the maximum and minimum values ​​of the original data respectively.

[0100] Step (4): Divide the preprocessed data into a training data set and a test data set in a ratio of 7:3, where the training set has 14,074 data items and the test set has 6,033 data items.

[0101] Step (5): Build a DA-TCN prediction model consisting of a feature attention mechanism FA, a temporal convolutional neural network TCN, a temporal attention mechanism TA, a ReLU layer, and a fully connected layer, and set the model parameters. The basic parameter settings are shown in Table 3:

[0102] Table 3

[0103] Training parameters Parameter settings Number of features 24 Number of iterations 100 Time step 5 Activation Function ReLU Learning rate 0.001 batch_size 128

[0104] Step (6): Based on the divided data set, the DA-TCN prediction model is trained with the training set data as input. The TCN network in the model is used to capture the long-term temporal dependency between indicators, and the feature extraction of all sequences in the multivariate temporal data set is realized. The feature attention mechanism FA is introduced to enhance the expression of relevant features according to the adaptive weights, and reduce the influence of minor features on the model prediction results; the temporal attention mechanism TA is introduced to further determine the multivariate key temporal features that affect product quality; the activation function and the fully connected layer are used in the output layer to integrate the information of each layer, so that the network can better learn the nonlinear relationship of the feature vector.

[0105] Step (7): Use the divided test set data to verify the prediction performance of the DA-TCN prediction model, and use the mean absolute error (MAE), root mean square error (RMSE), and mean square error (MSE) as evaluation indicators to evaluate the prediction performance of the model, and refer to the goodness of fit R 2 To judge the fitting effect of the model, R 2 The closer it is to 0, the worse the model fitting effect is.

[0106] The export material moisture prediction result obtained by the process production quality prediction method based on DA-TCN of the present invention is compared with the actual value of process quality. The comparison results are as follows: Figure 7 As shown in the figure, it can be seen that the process quality prediction method of the present invention has high accuracy.

[0107] In this embodiment, in order to verify that the DA-TCN prediction model has advantages in predicting the process quality of the silk thread thin plate drying process, a comparative experiment is designed to verify the prediction effect of the model. The time series data of the same drying process are used for comparative analysis under the same experimental conditions. The comparative models are: GRU (gated recurrent unit), LSTM-Attention (prediction method of introducing attention mechanism in long short-term memory network), Attention-Seq2seq (Seq2seq network combined with attention mechanism method) and DA-LSTM network (prediction method combining long short-term memory network with the dual attention mechanism proposed in the present invention, i.e., feature attention mechanism FA, LSTM, temporal attention mechanism TA, ReLU layer, fully connected layer). The comparison results are shown in Table 4, and the fitting results are shown in Table 4. Figure 8 shown.

[0108] Table 4

[0109] Prediction Model MSE RMSE MAE <![CDATA[R 2 ]]> GRU 0.0038 0.0613 0.0479 0.901 LSTM-Attention 0.0014 0.0378 0.0293 0.960 Attention-Seq2seq 0.0013 0.0362 0.0280 0.962 DA-LSTM 0.0007 0.0272 0.0185 0.979 DA-TCN 0.0004 0.0212 0.0168 0.988

[0110] Combined with Table 4 Figure 8 It can be seen that compared with the single GRU network, the DA-TCN model uses the TCN network to predict the quality indicators of the tofu drying process. It can model the long-term dependency relationship between the process parameters and quality indicators of the tofu drying process in a shorter time. 2 The prediction effect was improved from 0.901 to 0.988. At the same time, compared with the LSTM-Attention, Attention-Seq2seq and DA-LSTM combination models, the DA-TCN prediction model combines the temporal convolutional neural network with the dual attention mechanism, which has certain advantages, and the prediction accuracy and fitting effect are better. 2 The mean square error (MSE), root mean square error (RMSE), and mean absolute error (MAE) are 0.0004, 0.0212, and 0.0168, respectively. The results show that the process production quality prediction method based on DA-TCN has a better prediction effect.

[0111] The specific embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the knowledge of ordinary technicians in this field without departing from the scope of the present invention.

Claims

1. A method for predicting the moisture content of export materials using a temporal convolutional neural network, characterized in that: The method comprises the following steps: S1. Collect historical process data of the process manufacturing production line and construct a multivariate time series data set; the multivariate time series data set includes process parameter index data and export material moisture index data; S2. Preprocess the multivariate time series data set to obtain preprocessed data, and divide the preprocessed data into a training set and a test set according to the order of the acquisition cycle; S3. Introducing the dual attention mechanism into the temporal convolutional neural network to build the DA-TCN prediction model; S4. Using the training set to train the DA-TCN model, a trained DA-TCN prediction model is obtained; using the trained prediction model DA-TCN to verify the test set, a verified DA-TCN prediction model is obtained for use in predicting the moisture content of export materials; The preprocessing of the multivariate time series data set to obtain preprocessed data and dividing the preprocessed data into a training set and a test set according to the order of the acquisition cycle includes: According to the material head and tail interception rules of the process production process, the process parameter data and quality indicator data are judged to determine the critical position of the material head and tail, so as to retain the data between the critical positions of the material head and tail; Then, the data between the critical positions of the head and tail of the material are retained using 3 σ The method is to remove abnormal data and perform normalization; The normalized data is divided into training set and test set according to the preset ratio; The DA-TCN prediction model includes a feature attention mechanism FA, a temporal convolutional neural network TCN, a temporal attention mechanism TA, a ReLU layer, and a fully connected layer; The pre-processed process production parameter index characteristic variables and the export material moisture index characteristic variables are used as the input of the feature attention mechanism FA, and the feature weighted matrix is ​​obtained by weighting the attention weights of each indicator feature and the input feature. X '; The obtained feature weight matrix X The hidden layer output of the previous time step of the temporal convolutional neural network (TCN) is used as the hidden layer input of the next time step of the temporal convolutional neural network (TCN). The dilated causal convolution in the temporal convolutional neural network (TCN) is used to achieve long sequence interval sampling of all indicators. The PReLU function is used to capture the long time series dependency between indicators. The residual block of the temporal convolutional neural network (TCN) is used to extract features of all sequences in the input data. The output of the TCN hidden layer after iterating the residual block of the temporal convolutional neural network (TCN) to time t is used as the input of the temporal attention mechanism TA. By comparing the time attention weights of each historical moment in the process production process, the multivariate key temporal features affecting the moisture index of the export material are determined. All weights are weighted and summed with the hidden layer output of each historical moment to obtain the comprehensive temporal information; then, through the ReLU layer and the fully connected layer, the moisture prediction value of the export material is obtained.

2. The method for predicting the moisture content of export materials using a temporal convolutional neural network according to claim 1, characterized in that: The method of using a temporal convolutional neural network to predict the moisture content of the outlet material is applied to the thin plate drying process. In the thin plate drying process, the specific rules for cutting the head and tail of the material during the production process are as follows: If the inlet material moisture reference value is ≤ n1%, the material head and tail positions are determined as follows: according to the batch order of the acquisition cycle, the first data in the current batch process data with the outlet material moisture greater than m1% is used as the material head node, and the data after a delay of t time for the outlet material moisture is used as the material head critical position; the first data in the current batch process data with the outlet material moisture less than m1% is used as the material tail node, and the data after a delay of t time for the material tail node is used as the material tail critical position; If the inlet material moisture reference value is greater than n1% and ≤ n2%, the material head and tail positions are determined as follows: according to the batch order of the acquisition cycle, the first data in the current batch process data with the outlet material moisture greater than m2% is used as the material head node, and the data after the material head node is delayed by time t is used as the material head critical position; the first data in the current batch process data with the outlet material moisture less than m2% is used as the material tail node, and the data after the material tail node is moved forward by time t is used as the material tail critical position; If the inlet material moisture reference value is greater than n2%, the method for determining the material head and tail positions is as follows: according to the batch order of the acquisition cycle, the first data in the current batch process data with the outlet material moisture greater than m3% is used as the material head node, and the data after the material head node is delayed by time t is used as the material head critical position; the first data in the current batch process data with the outlet material moisture less than m3% is used as the material tail node, and the data after the material tail node is moved forward by time t is used as the material tail critical position; Among them, m1 <m2<m3。 3. A device for predicting the moisture content of export materials using a temporal convolutional neural network, characterized in that: The method comprises a module for executing the method for predicting the moisture content of export materials using a temporal convolutional neural network as described in any one of claims 1 to 2.

4. A terminal device, characterized in that: It includes a memory, a processor, and a program stored in the memory and executable by the processor. When the processor executes the program, it implements the method for predicting the moisture content of export materials using a temporal convolutional neural network as described in any one of claims 1 to 2.

5. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored program, wherein when the program is running, the device where the computer-readable storage medium is located is controlled to execute the method for predicting the moisture content of export materials using a temporal convolutional neural network as described in any one of claims 1-2.

Images