TA-MMKELA-BiGRU-based discharged material water content prediction method and system

Abstract

The invention discloses a discharged material water content prediction method based on TA-MMKELA-BiGRU, and the method comprises the steps: sequentially carrying out the preprocessing and noise reduction of collected flow production line time sequence working condition data, and obtaining a denoised sample set; constructing a time sequence attention mechanism, and taking the de-noised sample set as the input of the time sequence attention mechanism; constructing a multi-level multi-kernel extreme learning machine encoder, and taking the output of the time sequence attention mechanism as the input of the multi-level multi-kernel extreme learning machine encoder; constructing a two-way gating cycle unit model, and taking the output of the multi-level multi-kernel extreme learning machine encoder as the input of the two-way gating cycle unit model; training a TA-MMKELA-BiGRU prediction model constructed by the time sequence attention mechanism, a multi-level multi-kernel extreme learning machine encoder and a bidirectional gating cycle unit model; and inputting the trained TA-MMKELA-BiGRU prediction model into the process parameter characteristic data of the flow production line for prediction, and outputting a prediction result. According to the invention, the prediction precision of the quality index of the flow production line is improved.

Description

Technical Field

[0001] The invention relates to a method and system for predicting moisture content of discharged materials based on TA-MMKELA-BiGRU, and belongs to the field of intelligent manufacturing. Background Art

[0002] With the rapid development of digitalization, data analysis and machine learning technologies, the integration of intelligent industrial process monitoring, quality prediction and predictive maintenance solutions has received widespread attention. Traditional production systems are gradually transforming into intelligent systems. This evolution highlights the increasing importance of intelligent process industries.

[0003] Although the quality prediction model based on actual industrial data has achieved certain results in the prediction of manufacturing product quality indicators and has produced many beneficial results. However, when faced with complex time series data in the production line, how to mine the deep correlation characteristics between data, thereby improving the prediction accuracy of quality indicators and effectively improving the stability of product quality still requires more effective methods and approaches. Therefore, providing high-quality methods and approaches for accurate prediction of multi-process quality in process production is still an urgent problem to be solved. Summary of the invention

[0004] The present invention provides a method for predicting the moisture content of discharged materials based on TA-MMKELA-BiGRU, so as to construct a TA-MMKELA-BiGRU prediction model to realize the prediction of the moisture content of discharged materials, which is a product quality indicator in a process manufacturing production line.

[0005] The technical scheme of the present invention is: according to the first aspect of the present invention, a method for predicting the moisture content of an outgoing material based on TA-MMKELA-BiGRU is provided, comprising the following steps: preprocessing and denoising the collected time series operating condition data of a process production line in turn to obtain a denoised sample set; the sample set includes process parameter feature data of the process and feature data of the moisture content of the outgoing material; constructing a temporal attention mechanism, and using the denoised sample set as the input of the temporal attention mechanism; constructing a multi-level multi-kernel extreme learning machine encoder, and using the output of the temporal attention mechanism as the input of the multi-level multi-kernel extreme learning machine encoder; constructing a bidirectional gated recurrent unit model, and using the output of the multi-level multi-kernel extreme learning machine encoder as the input of the bidirectional gated recurrent unit model; training a TA-MMKELA-BiGRU prediction model constructed by the temporal attention mechanism, the multi-level multi-kernel extreme learning machine encoder, and the bidirectional gated recurrent unit model; inputting the trained TA-MMKELA-BiGRU prediction model into the process parameter feature data of the process production line for prediction, and outputting a prediction result.

[0006] Furthermore, the collected time series operating condition data of the process production line are preprocessed and denoised in sequence, including: collecting the time series operating condition data of the process production line to construct a data set; preprocessing the data set to obtain data after removing null values, outliers and normalization as a preprocessed data set; using a Kalman filter algorithm to denoise the multi-noise situation in the preprocessed data set to obtain a denoised sample set.

[0007] Furthermore, the preprocessing is specifically as follows: firstly, the missing values ​​in the data set are deleted according to the time row; Method to eliminate abnormal data; The processed data set is then normalized as the preprocessed data set.

[0008] Furthermore, the specific process of the temporal attention mechanism is: the denoised sample set is used as the input of the temporal attention mechanism, the temporal attention mechanism identifies the time steps related to the prediction, and uses the probability weighting mechanism to extract information therefrom, assigns corresponding weights to the input features of each time step, obtains the feature dimension index through the assigned weights through the softmax function, and calculates the associated features by weighted calculation of the weights assigned to each process parameter and the corresponding feature dimension index as the input of the multi-level kernel extreme learning machine network.

[0009] Furthermore, the multi-level multi-kernel extreme learning machine encoder MMKELA is formed by stacking multiple multi-kernel extreme learning machine encoders MKELA, and the associated features output by the temporal attention mechanism are input into the multi-level multi-kernel extreme learning machine encoder, and the weights and biases from the input layer to the kernel hidden layer are randomly generated. Then, the kernel hidden layer output is calculated through the activation function, and the feature weights and outputs are obtained by training the kernel hidden layer layer by layer. The output of each kernel hidden layer is used as the input of the next layer, and the layers are stacked layer by layer to extract the multi-level features of the data as the input of the bidirectional gated recurrent unit model.

[0010] Furthermore, the multi-kernel extreme learning machine encoder MKELA adopts an RBF-Sig-Lap hybrid kernel function, which is expressed as follows:

[0011]

[0012] In the formula, K hybrid represents the hybrid kernel, K RBF represents the Gaussian kernel, K Sig represents the Sigmoid kernel, K Lap represents the Laplacian kernel, ω 1 ,ω 2 ,ω 3 is the weight ratio of each kernel function.

[0013] Furthermore, the bidirectional gated recurrent unit model is composed of two unidirectional gated recurrent units in opposite directions; the output of the multi-level multi-kernel extreme learning machine encoder is used as the input of the bidirectional gated recurrent unit model, and the bidirectional gated recurrent unit model uses the input at time t and the output at the previous time as input, and learns the complex correlation between the process data before and after the production line in the forward and reverse directions respectively, and finally merges the forward and reverse hidden states into a joint hidden state vector output, and then obtains the predicted value of the moisture content of the output material through the fully connected layer.

[0014] According to a second aspect of the present invention, a system for predicting the moisture content of an output material based on TA-MMKELA-BiGRU is provided, comprising a module for executing the method for predicting the moisture content of an output material based on TA-MMKELA-BiGRU as described in any one of the above.

[0015] According to a third aspect of the present invention, there is provided a terminal device, comprising a memory, a processor, and a program stored in the memory and executable by the processor, wherein when the processor executes the program, the method for predicting the moisture content of an output material based on TA-MMKELA-BiGRU as described in any one of the above is implemented.

[0016] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium comprising 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 discharged materials based on TA-MMKELA-BiGRU.

[0017] The beneficial effects of the present invention are:

[0018] 1. The present invention removes non-steady-state production data from the collected aromatherapy time series data and The method is used to detect and delete outliers, maintaining the rationality of the data and the integrity of the structure; through normalization, the data is converted to between 0 and 1, eliminating the impact of the dimension and improving the computing power.

[0019] 2. The present invention uses Kalman filtering to reduce the noise of time series data through two steps of prediction and update, effectively filtering out data that is not conducive to model establishment, making the data sample distribution more uniform, and solving the problem of preprocessing complex process data in process manufacturing.

[0020] 3. The present invention constructs a TA-MMKELA-BiGRU prediction model. Compared with the traditional GRU model, the temporal attention mechanism model and the combination model of BiGRU combined with the traditional attention mechanism, the prediction accuracy of the present invention is more accurate, providing a new method for the prediction model of production line process quality temporal data.

[0021] From the above, it can be seen that the present invention specifically solves the problem that complex relationships in data in the process manufacturing industry are difficult to predict. It can mine the deep-level correlation characteristics between data, thereby improving the prediction accuracy of quality indicators, effectively improving the stability of product quality, and providing an implementation method and approach for accurate prediction of multi-process quality in process production. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a schematic diagram of the process of the present invention;

[0023] Figure 2 This is a structural diagram of the temporal attention mechanism described in the present invention;

[0024] Figure 3 This is a structural diagram of the multi-level multi-kernel extreme learning machine encoder of the present invention;

[0025] Figure 4 This is the BiGRU structure diagram of the present invention;

[0026] Figure 5 Part of the data in the data set provided by the embodiment of the present invention;

[0027] Figure 6 A comparison diagram of the predicted values ​​and the actual values ​​of the present invention and the first deformation model;

[0028] Figure 7 It is a comparison chart of the predicted values ​​and the true values ​​of the present invention and the second deformation model. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical scheme and advantages of the embodiments of the present invention clearer, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with 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 the embodiments in this application and the features in the embodiments can be combined with each other arbitrarily without conflict.

[0030] like Figure 1-7As shown, according to the first aspect of an embodiment of the present invention, a method for predicting the moisture content of an outgoing material based on TA-MMKELA-BiGRU is provided, comprising the following steps: preprocessing and denoising the collected time series operating condition data of the process production line in turn to obtain a denoised sample set; the sample set includes process parameter feature data of the process and feature data of the moisture content of the outgoing material; constructing a temporal attention mechanism, and using the denoised sample set as the input of the temporal attention mechanism; constructing a multi-level multi-kernel extreme learning machine encoder, and using the output of the temporal attention mechanism as the input of the multi-level multi-kernel extreme learning machine encoder; constructing a bidirectional gated recurrent unit model, and using the output of the multi-level multi-kernel extreme learning machine encoder as the input of the bidirectional gated recurrent unit model; training a TA-MMKELA-BiGRU prediction model constructed by the temporal attention mechanism, the multi-level multi-kernel extreme learning machine encoder, and the bidirectional gated recurrent unit model; inputting the trained TA-MMKELA-BiGRU prediction model into the process parameter feature data of the process production line for prediction, and outputting a prediction result.

[0031] Furthermore, the collected time series operating condition data of the process production line are preprocessed and denoised in sequence, specifically including: collecting the time series operating condition data of a single process in the process production line to construct a data set; wherein the data set includes process parameter characteristic data of the process and discharge moisture content characteristic data; preprocessing the data set to obtain data after removing null values, outliers and normalization as a preprocessed data set; using the Kalman filter algorithm to denoise the multi-noise situation in the preprocessed data set to obtain a denoised sample set.

[0032] Furthermore, the preprocessing is specifically as follows: firstly, the missing values ​​in the data set are deleted according to the time row; Method to eliminate abnormal data; The processed data set is then normalized as the preprocessed data set; The formula is as follows: x represents the data after removing outliers. represents the average value, Represents variance.

[0033] Specifically, the normalization process refers to performing maximum-minimum normalization on all process parameters and discharge moisture content data, mapping all data to the [0,1] interval, and completing dimensionless process parameters and discharge moisture content data. The formula is as follows:

[0034]

[0035] Where: is the normalized process parameter / discharge moisture content data; xi is the process parameters / discharge moisture content data before normalization; x max and x min x i The maximum and minimum values ​​of .

[0036] Process manufacturing is usually scheduled in a continuous manner or according to material batches. In this process, the moisture content of the discharge material and some key process parameter indicators are mainly collected through sensors or intelligent terminals. These devices have a short sampling cycle, so a large amount of data is generated. At the same time, due to the large number of production batches involved in process manufacturing, there are significant differences in the raw material composition between batches; furthermore, when the production equipment is in the start-up, stop and other states, a large amount of abnormal data will be generated, so the process data presents a multi-noise feature. These factors jointly affect the validity and reliability of the data. Therefore, it is necessary to perform data denoising. The present invention uses Kalman filtering to perform data denoising on the preprocessed data set through two steps of prediction and update, effectively filtering out data that is not conducive to model establishment, and making the data sample distribution more uniform. Through the prediction step of Kalman filtering, the future state can be predicted according to the system model, and the covariance of the state error can be calculated; then, the difference between the actual measured value and the predicted value is used to calculate the Kalman gain to adjust the predicted value, thereby updating the state estimate.

[0037] Specifically:

[0038] According to the system state transfer matrix A k and the control input matrix B k , using the current state estimate and control input u k Predict the state at the next moment The expression is as follows:

[0039]

[0040] Then calculate the Kalman gain K k , which weighs the reliability of the predicted value and the measured value. Meanwhile, the Kalman gain determines how to combine the measured value with the predicted value to minimize the estimation error.

[0041]

[0042] In the formula, is the observation matrix, R k is the measurement noise covariance matrix.

[0043] Furthermore, combined with the measured value z k and predicted values Update state estimate The expression is as follows:

[0044]

[0045] In the formula, is the measurement residual, which represents the difference between the actual measurement and the predicted measurement.

[0046] Furthermore, update the state error covariance matrix P k|k , the expression is as follows:

[0047] P k|k =(IK k H k ) k|k-1 ;

[0048] Where I is the identity matrix, K k H k is the factor that adjusts the prediction covariance. The updated covariance matrix reflects the degree of uncertainty reduction in the state estimate. These steps combined enable the Kalman filter to provide optimized state estimates in noisy environments, thereby effectively reducing noise.

[0049] Furthermore, if Figure 2-Figure 4 As shown, a TA-MMKELA-BiGRU prediction model is constructed, including:

[0050] The denoised sample set is used as the input of the temporal attention mechanism, which assigns corresponding weights to the input features of each time step by identifying the time steps related to the prediction and extracting information from them using a probability weighting mechanism. The assigned weights are passed through the softmax function to obtain the feature dimension index The weights assigned to each process parameter and the corresponding feature dimension index are weighted to obtain the associated feature x t ′ is used as the input of the multi-level kernel extreme learning machine network; by applying the temporal attention mechanism, the key features in the time series data can be effectively utilized, and the effective extraction of key features provides strong support for the subsequent modules to participate in the prediction of the moisture content of the discharge, thereby achieving the effect of significantly improving the prediction accuracy;

[0051] The multi-level multi-kernel extreme learning machine encoder MMKELA is composed of multiple multi-kernel extreme learning machine encoders MKELA stacked together to convert the associated features x output by the temporal attention mechanism into t ′ is input into the multi-level multi-kernel extreme learning machine encoder, and the weight w from the input layer to the kernel hidden layer is randomly generated j and bias b j , and then calculate the kernel hidden layer output H through the activation function j , the feature weight β is obtained by training the kernel hidden layer layer by layer 1 ,β 2 …,βJ and output X j , the output of each kernel hidden layer is used as the input of the next layer, stacked layer by layer to extract the multi-level features of the data as the input of the bidirectional gated recurrent unit model;

[0052] The Bidirectional Gated Recurrent Unit (BiGRU) model consists of two unidirectional GRUs in opposite directions. The output of the multi-level multi-kernel extreme learning machine encoder is used as the input of the Bidirectional Gated Recurrent Unit model. The Bidirectional Gated Recurrent Unit model uses the input x at time t. t and the output h at the previous moment t-1 As input, the complex correlation between the process data before and after the process production line is effectively learned in the forward and reverse directions, and finally the forward and reverse hidden states are merged into a joint hidden state vector h t Output, and then pass through the fully connected layer to obtain the predicted value of the moisture content of the output.

[0053] Specifically, the weight calculation formula of the temporal attention mechanism is as follows:

[0054]

[0055] Where, ν p is the temporal attention weight vector, W a is the weight matrix, f() is the activation function, and b is the bias vector.

[0056] Furthermore, the characteristic dimension index calculation formula is as follows:

[0057]

[0058] Furthermore, the associated feature x t The calculation formula is as follows:

[0059]

[0060] The multi-kernel extreme learning machine encoder MKELA adopts the RBF-Sig-Lap hybrid kernel function, which is expressed as follows:

[0061]

[0062] In the formula, K hybrid represents the hybrid kernel, K RBF represents the Gaussian kernel, K Sig represents the Sigmoid kernel, K Lap represents the Laplacian kernel, ω 1 ,ω 2 ,ω 3 is the weight ratio of each kernel function.

[0063] In the above, the hidden layer part is calculated:

[0064] H j =f(ω j x t ′+b j );

[0065] In the formula, H j represents the hidden layer output, f() is the activation function, ω j represents the distribution weight ratio, b j Indicates the offset.

[0066] The output weight of any multi-kernel extreme learning machine encoder MKELA in the multi-level kernel extreme learning machine encoder can be expressed as:

[0067] β j =(H j +Ω MKELM ) -1 X j ;

[0068] In the formula, Ω MKELM is the kernel function of the multi-kernel extreme learning machine encoder MKELA, i.e., the hybrid kernel; X j is the process input parameter vector of the j-th layer model.

[0069] The multi-level multi-kernel extreme learning machine encoder MMKELA mentioned above is based on the process parameter weights assigned by the temporal attention mechanism. Multiple MKELAs are stacked to form a multi-level multi-kernel extreme learning machine encoder MMKELA. Its structure is as follows: Figure 3 As shown in the figure, the multi-level multi-kernel extreme learning machine encoder MMKELA stacks the single-layer MKBELA into a multi-layer structure, which reduces the reconstruction error of the learning sample process layer by layer, and the generalization ability is further improved. It still has excellent performance in the face of high-dimensional complex data; furthermore, it ensures the accurate inverse solution of the output under the reversible kernel matrix and reduces the reconstruction error; the multi-layer stacking structure can be trained layer by layer to learn more abstract fusion features, and has a stronger feature extraction capability.

[0070] The RBF-Sig-Lap hybrid kernel function proposed in the present invention performs linear combination operations on Gaussian kernel function (RBF) with powerful feature representation capability, Sigmoid kernel function (Sig) and Laplacian kernel (Lap). Gaussian kernel function can effectively capture local structure and nonlinear relationship in multidimensional data, which is very useful for complex multidimensional features; Sigmoid kernel can simulate the activation function of neurons, and is suitable for capturing nonlinear coupling relationship between features, especially when processing features with complex interactions between multiple dimensions, it can effectively capture the coupling between these features; Laplace kernel is based on Manhattan distance, and can perform well in capturing local changes and outliers in time series data, better balance the global and local influences, and thus more accurately reflect the dynamic changes of time series data.

[0071] The bidirectional gated recurrent unit obtains the feature information hidden in the time series and predicts the hidden layer state output at each moment. The expression is as follows:

[0072]

[0073] Furthermore, the number of neurons in the temporal attention mechanism TA is set to 32; the number of hidden neurons in the MBELA layer is set to 16; the number of neurons in the BiGRU layer is 64, and the activation function is Relu; the number of output neurons in the fully connected layer is 1.

[0074] By applying the above technical solutions, it can be known that the present invention uses a temporal attention mechanism to identify temporal dependencies in data, assigns weights to determine the parts that have the greatest impact on temporal features, and thus extracts key temporal features; the multi-level multi-kernel extreme learning machine encoder MMKELA introduces a hybrid kernel function into the autoencoder, constructs a multi-layer network structure, and maps the input data from the original space to a high-dimensional feature space by minimizing the reconstruction error layer by layer, thereby realizing nonlinear feature process parameter extraction; this method not only simplifies the model parameter configuration, but also can effectively extract complex correlations in process data, identify key feature information that best reflects product quality, and make the entire model prediction more accurate.

[0075] The denoised sample set is divided into training set and test set according to the proportion, and the TA-MMKELA-BiGRU prediction model is trained with the training set to obtain the trained TA-MMKELA-BiGRU prediction model; the test set is used to verify the trained TA-MMKELA-BiGRU prediction model, and a comparative experiment is designed to verify the effectiveness of this combined network model. The mean absolute error MAE, mean square error MSE, and goodness of fit R 2 As an evaluation index, the calculation formula of the evaluation index is as follows:

[0076]

[0077] In order to accurately evaluate the prediction performance of the model proposed in this paper, the mean absolute error (MAE) and root mean square error (RMSE) were selected as the main evaluation indicators, and the goodness of fit R 2 To judge the model's fitting effect. Generally speaking, R 2 The closer the value is to 1, the better the model fitting effect is; at the same time, the smaller the mean absolute error and root mean square error values ​​are, the lower the degree of difference in the regression analysis is, which reflects that the model fitting performance is better.

[0078] Illustratively, an optional implementation of the present invention is described in detail below, but the content of the present invention is not limited to the scope.

[0079] This implementation aims at the problem that in process production, due to complex variable coupling and significant time series characteristics, and the need to consider a large amount of historical data during analysis, it is impossible to quickly mine the knowledge information in real-time data, which leads to the difficulty in accurately predicting process quality indicators. A process quality indicator prediction method based on TA-MMKELA-BiGRU is proposed. The moisture content of the discharge material is further explained as follows: First, the pre-processed data set is denoised by Kalman filtering and then input into the time series attention mechanism-multi-level multi-kernel extreme learning machine encoder-BiGRU for training, and finally the moisture content quality indicator prediction of the discharge material is completed.

[0080] The data collected in this example is the flavoring process of a silk production line of a process manufacturing company. The data collection period is from March 31, 2022 to April 8, 2022. The data collection cycle is 8s. A total of 19,887 samples were collected. Some of the data are as follows: Figure 5 As shown. By preprocessing the data in the dataset, first replace the null value data in the dataset with NaN values, and then delete the NaN value rows to clean up the remaining 16625 data. Then use The abnormal data was deleted by the method. After data cleaning, there were 13233 data samples, which were normalized and used as the preprocessing data set. Then, the Kalman filter algorithm was used to reduce the data noise of the preprocessing data set through two steps of prediction and update. The denoised sample set was allocated to the training set and the test set in a ratio of 7:3. The trained TA-MMKELA-BiGRU prediction model was obtained by training the training set data. The test set was used to perform predictions under different models to verify the prediction performance of the TA-MMKELA-BiGRU prediction model proposed in the present invention. The parameter settings of the TA-MMKELA-BiGRU prediction model are shown in Table 1:

[0081] Table 1

[0082] parameter Parameter settings Number of attention mechanism neurons 32 Number of hidden neurons in MKELA 16 Number of neurons in the BiGRU hidden layer 64 Step Length 12 Iterations 100 Learning Rate 0.01

[0083] Combined with the evaluation index, the TA-MMKELA-BiGRU prediction model proposed in the present invention and the first deformation model of the present invention (the first deformation model is the GRU model, the BiGRU model, the TA-BiGRU model, the MMKELA-BiGRU model) are used to predict the moisture content of the discharge material of the quality index. The evaluation index is shown in Table 2. By comparing and verifying the quality index prediction of the TA-MMKELA-BiGRU prediction model with the GRU model, BiGRU model, TA-BiGRU model, and MMKELA-BiGRU model on the test set of the same data, it can be seen that the TA-MMKELA-BiGRU prediction model proposed in the present invention has a MAE index of 60%, 51.85%, 42.22%, and 26.76% lower than that of the GRU, BiGRU, TA-BiGRU, and MMKELA-BiGRU models, and a RMSE index of 60.82%, 44.63%, 41.22%, and 24.72% lower, respectively. This shows that the TA-MMKELA-BiGRU prediction model has a better prediction effect in dealing with complex time series problems, reflecting the necessity of introducing the time series attention mechanism and MKELA, and effectively improving the model's ability to extract complex features of production process parameters. Finally, the effects of the five models are quantitatively analyzed, and the determination coefficient R 2 The calculation results show that the R of the TA-MMKELA-BiGRU prediction model proposed in the present invention is 2 The value is 0.9833, which is closest to 1, that is, the prediction effect of the model of the present invention is the best.

[0084] Table 2

[0085] Model MAE RMSE <![CDATA[R 2 ]]> TA-MMKELA-BiGRU 0.0052 0.0067 0.9833 GRU 0.0130 0.0171 0.8909 BiGRU 0.0108 0.0121 0.9460 TA-BiGRU 0.0090 0.0114 0.9513 MMKELA-BiGRU 0.0071 0.0089 0.9708

[0086] In order to highlight the advantages of the present invention, the TA-MMKELA-BiGRU prediction model proposed in the present invention and the 100 predicted discharge moisture content quality indicators obtained by each first deformation model are compared and displayed. The display results are as follows: Figure 6 As shown in the figure, it can be seen that the TA-MMKELA-BiGRU prediction model proposed in the present invention has the highest prediction accuracy, has a better fitting ability for the actual discharge moisture content quality index, and is basically consistent with the change trend of the actual discharge moisture content.

[0087] In summary, the TA-MMKELA-BiGRU prediction model proposed for the present invention is used to predict quality indicators on the same data test set with the GRU model, BiGRU model, TA-BiGRU model, and MMKELA-BiGRU model respectively. It can be seen that the prediction effect of the bidirectional gated recurrent unit (BiGRU) is significantly better than that of the unidirectional gated recurrent unit (GRU), which shows that the bidirectional GRU structure has more advantages in effectively learning the correlation between the front and back processes of the process manufacturing production line. In addition, compared with the BiGRU model, the temporal attention mechanism adopted by the TA-BiGRU model can more effectively extract the internal features of the process data by dynamically allocating different parameter weights, thereby mining the potential laws between the temporal data; the MMKELA-BiGRU model introduces a multi-level multi-kernel extreme learning machine encoder to simultaneously mine the deep-level correlation features of high-dimensional and complex data, and more comprehensively learns the information in the hidden state output by the BiGRU network, which greatly enhances the prediction accuracy of the BiGRU network in the process manufacturing production line scenario. Therefore, the TA-MMKELA-BiGRU prediction model can maximize the advantages of the temporal attention mechanism TA, MMKELA model and BiGRU model, effectively handle the complex relationship between process parameters and quality indicators, and is of great significance for optimizing model parameters and realizing quality control in multi-process production. The main goal of the process production line is to improve the prediction accuracy of the process quality index parameters of the production line, so as to ensure the high accuracy of the quality index prediction results. The present invention, through the proposed prediction method, can deeply explore the characteristics of variables and solve the challenges of multi-process coupling in the process manufacturing process in the intelligent manufacturing environment, especially the bottleneck problems in the self-perception, self-calculation and self-prediction of the whole process elements. Therefore, the method of the present invention significantly improves the prediction accuracy of quality indicators and effectively reduces the operating costs of enterprises, providing a new method and path for realizing accurate prediction and optimal regulation of process production processes.

[0088] Furthermore, in order to illustrate the advantages of introducing the improved kernel function in the model of the present invention, the model of the present invention and the second deformation model using the same architecture as the present invention are compared and analyzed. The difference between the second deformation model and the present invention is that the MMKELA part uses a kernel function different from that of the present invention (Sigmoid kernel function, Laplace kernel function, Gaussian kernel function, Gaussian kernel-inverse cosine kernel function). The expressions of each kernel function are shown below. The evaluation index results of the quality indicators of the moisture content of the discharge predicted by different models and the actual values ​​are shown in Table 3.

[0089] The RBF-Sig-Lap kernel function (ω 1 =0.6,ω 2 =0.2,ω 3 =0.2), the specific expression is:

[0090]

[0091] The Sigmoid kernel function expression is: K(x,y) = tanh(αx T y+β);

[0092] The Laplace kernel function expression is:

[0093] The Gaussian kernel function expression is:

[0094] The Gaussian kernel-inverse cosine kernel function expression is:

[0095] Table 3

[0096] Kernel Function MAE RMSE <![CDATA[R 2 ]]> RBF-Sig-Lap kernel function 0.0052 0.0067 0.9833 Sigmoid kernel function 0.0191 0.0226 0.8109 Laplace kernel function 0.0112 0.0144 0.9226 Gaussian kernel function 0.0072 0.0101 0.9629 Gaussian kernel-inverse cosine kernel function 0.0065 0.0091 0.9706

[0097] The quality index prediction comparison and verification of the RBF-Sig-Lap kernel function proposed in the present invention and the Sigmoid kernel function, Laplace kernel function, Gaussian kernel function and Gaussian kernel-inverse cosine kernel function given in Table 3 on the test set of the same data shows that: the RBF-Sig-Lap kernel function used in the MMKELA model proposed in the present invention is compared with the Sigmoid kernel function, Laplace kernel function, Gaussian kernel function and Gaussian kernel-inverse cosine kernel function, and the MAE index decreases by 72.77%, 53.57%, 27.78% and 20.97% respectively, and the RMSE index decreases by 70.35%, 53.47%, 33.66% and 26.37% respectively, which shows that the RBF-Sig-Lap kernel function has better performance in processing complex process data, reflects the necessity of the RBF-Sig-Lap kernel function in the MMKELA model of the present invention, and effectively improves the model's ability to extract complex features of production process parameters. Finally, the effects of the five models were quantitatively analyzed by the determination coefficient R2 The calculation results show that the R of the RBF-Sig-Lap kernel function proposed in this invention is 2 The value is 0.9833, which is closest to 1, that is, the RBF-Sig-Lap kernel function of the present invention has the best prediction effect.

[0098] The TA-MMKELA-BiGRU prediction model proposed in the present invention obtains 100 predicted discharge moisture content quality indicators obtained by applying the above-mentioned different kernel functions and the change curves of the actual values. The display results are shown as follows: Figure 7 As shown in the figure, it can be seen that the RBF-Sig-Lap kernel function proposed in the present invention has the highest prediction accuracy, has a better fitting ability for the actual discharge moisture content quality index, and is more consistent with the actual discharge moisture content change trend.

[0099] According to a second aspect of an embodiment of the present invention, a system for predicting moisture content of discharged materials based on TA-MMKELA-BiGRU is provided, comprising a module for executing a method for predicting moisture content of discharged materials based on TA-MMKELA-BiGRU as described in any one of the above. Specifically, it includes: a first module, which is used to pre-process and de-noise the collected process production line time series working condition data in sequence to obtain a de-noised sample set; the sample set includes process parameter feature data of the process and discharge moisture content feature data; a second module, which is used to construct a temporal attention mechanism, and use the de-noised sample set as the input of the temporal attention mechanism; a third module, which is used to construct a multi-level multi-kernel extreme learning machine encoder, and use the output of the temporal attention mechanism as the input of the multi-level multi-kernel extreme learning machine encoder; a fourth module, which is used to construct a bidirectional gated recurrent unit model, and use the output of the multi-level multi-kernel extreme learning machine encoder as the input of the bidirectional gated recurrent unit model; a fifth module, which is used to train the TA-MMKELA-BiGRU prediction model constructed by the temporal attention mechanism, the multi-level multi-kernel extreme learning machine encoder, and the bidirectional gated recurrent unit model; a sixth module, which is used to input the trained TA-MMKELA-BiGRU prediction model into the process production line process parameter feature data for prediction, and output the prediction result. As used above, the term "module" can be a combination of software and / or hardware to implement a predetermined function. Although the system described in the above embodiment is preferably implemented by software, implementation by hardware, or a combination of software and hardware is also possible and conceivable. For parts that are not described in detail in each module, please refer to the relevant description of the embodiment.

[0100] 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 in the memory and executable by the processor, wherein when the processor executes the program, the method for predicting the moisture content of an output material based on TA-MMKELA-BiGRU as described in any one of the above is implemented.

[0101] According to a fourth aspect of an embodiment 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 discharged materials based on TA-MMKELA-BiGRU.

[0102] In an exemplary embodiment, the computer-readable storage medium may include, but is not limited to, various media that can store computer programs, such as a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk.

[0103] Obviously, those skilled in the art should understand that the above modules or steps of the present invention can be implemented by a general computing device, they can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, and in some cases, the steps shown or described can be executed in a different order than here, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. Thus, the present invention is not limited to any specific combination of hardware and software.

[0104] The specific implementation modes of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above implementation modes, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.

Claims

1. A method for predicting moisture content of discharge material based on TA-MMKELA-BiGRU, characterized in that: The following steps are involved: The collected time series operating condition data of the process production line are preprocessed and denoised in turn to obtain a denoised sample set; The sample set includes process parameter characteristic data of the process and characteristic data of the moisture content of the output material; Construct a temporal attention mechanism and use the denoised sample set as the input of the temporal attention mechanism; Construct a multi-level multi-kernel extreme learning machine encoder and use the output of the temporal attention mechanism as the input of the multi-level multi-kernel extreme learning machine encoder; A bidirectional gated recurrent unit model is constructed, and the output of the multi-level multi-kernel extreme learning machine encoder is used as the input of the bidirectional gated recurrent unit model; Training the TA-MMKELA-BiGRU prediction model constructed by the temporal attention mechanism, the multi-level multi-kernel extreme learning machine encoder, and the bidirectional gated recurrent unit model; The trained TA-MMKELA-BiGRU prediction model is input into the process parameter characteristic data of the process production line for prediction and the prediction results are output.

2. The method for predicting the moisture content of a discharge material based on TA-MMKELA-BiGRU according to claim 1, characterized in that: The collected time series operating condition data of the process production line are preprocessed and denoised in sequence, including: The collected process production line time series operating data is used to construct a data set; Preprocess the data set to obtain data after removing null values, outliers and normalization as the preprocessed data set; The Kalman filter algorithm is used to reduce the noise in the preprocessed data set to obtain the denoised sample set.

3. The method for predicting the moisture content of a discharge material based on TA-MMKELA-BiGRU according to claim 2, characterized in that: The preprocessing is specifically as follows: firstly, the missing values ​​in the data set are deleted according to the time row; Method to eliminate abnormal data; go through The processed data set is then normalized as the preprocessed data set.

4. The method for predicting the moisture content of a discharge material based on TA-MMKELA-BiGRU according to claim 1, characterized in that: The specific process of the temporal attention mechanism is as follows: the denoised sample set is used as the input of the temporal attention mechanism. The temporal attention mechanism identifies the time steps related to the prediction and extracts information from them using a probability weighting mechanism. A corresponding weight is assigned to the input features of each time step. The assigned weight is passed through a softmax function to obtain a feature dimension index. The weight assigned to each process parameter is weighted with the corresponding feature dimension index to obtain the associated feature as the input of the multi-level kernel extreme learning machine network.

5. The method for predicting the moisture content of a discharge material based on TA-MMKELA-BiGRU according to claim 1, characterized in that: The multi-level multi-kernel extreme learning machine encoder MMKELA is formed by stacking multiple multi-kernel extreme learning machine encoders MKELA. The associated features output by the temporal attention mechanism are input into the multi-level multi-kernel extreme learning machine encoder, and the weights and biases from the input layer to the kernel hidden layer are randomly generated. Then, the kernel hidden layer output is calculated through the activation function, and the feature weights and outputs are obtained by training the kernel hidden layer layer by layer. The output of each kernel hidden layer is used as the input of the next layer, and the layers are stacked layer by layer to extract the multi-level features of the data as the input of the bidirectional gated recurrent unit model.

6. The method for predicting the moisture content of discharged materials based on TA-MMKELA-BiGRU according to claim 5, characterized in that: The multi-kernel extreme learning machine encoder MKELA adopts the RBF-Sig-Lap hybrid kernel function, which is expressed as follows: In the formula, K hybrid represents the hybrid kernel, K RBF represents the Gaussian kernel, K Sig represents the Sigmoid kernel, K Lap represents the Laplacian kernel, ω1, ω2, and ω3 are the weight ratios of each kernel function.

7. The method for predicting the moisture content of discharged materials based on TA-MMKELA-BiGRU according to claim 1, characterized in that: The bidirectional gated recurrent unit model is composed of two unidirectional gated recurrent units in opposite directions; the output of the multi-level multi-kernel extreme learning machine encoder is used as the input of the bidirectional gated recurrent unit model, and the bidirectional gated recurrent unit model uses the input at time t and the output at the previous time as input, and learns the complex correlation between the process data before and after the process production line in the forward and reverse directions respectively, and finally merges the forward and reverse hidden states into a joint hidden state vector output, and then obtains the predicted value of the moisture content of the output material through the fully connected layer.

8. A discharge moisture content prediction system based on TA-MMKELA-BiGRU, characterized by: The method comprises a module for executing the method for predicting the moisture content of discharged materials based on TA-MMKELA-BiGRU as described in any one of claims 1 to 7.

9. 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, the method for predicting the moisture content of the discharge material based on TA-MMKELA-BiGRU as described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored program, wherein when the program is executed, the device where the computer-readable storage medium is located is controlled to execute the method for predicting the moisture content of discharged materials based on TA-MMKELA-BiGRU as described in any one of claims 1 to 7.

Images