A method for dynamic prediction of rolling force based on feature enhancement and deep neural networks

Abstract

The present invention provides a method for dynamic prediction of rolling force based on feature dimension increase and deep neural network, and relates to the field of rolling technology. The method specifically comprises: collecting production data during the hot strip rolling process and performing data processing, constructing a sample data set, and dividing a training set from the sample data set according to a preset ratio; constructing a feature dimension increase convolutional deep neural network FACDNN model, and using the training set to train the feature dimension increase convolutional deep neural network FACDNN model to obtain a trained feature dimension increase convolutional deep neural network FACDNN model; by re-collecting production data during the hot strip rolling process and inputting the trained feature dimension increase convolutional deep neural network FACDNN model, a rolling force prediction result during the hot strip rolling process is obtained. The method fully considers the influence of changes in process parameters on the accuracy of the model, thereby realizing high-precision dynamic prediction of the rolling force of the downstream rack.

Description

Technical Field

[0001] The present invention relates to the field of rolling technology, and in particular to a rolling force dynamic prediction method based on feature dimension increase and deep neural network. Background Art

[0002] In hot rolling production, rolling force is one of the most important process parameters. The prediction accuracy of rolling force directly affects the final setting of the rolling procedure, thereby affecting the three-dimensional size and yield rate of hot-rolled products. Hot rolling production is a serial production process with multiple upstream and downstream processes, and generally multiple racks are arranged. Due to the limitations of the measurement environment and the installation position of the instrument, the process parameters directly related to the rolling force, such as the temperature and width of the rolled piece between the racks, are difficult to obtain directly through measurement. After the rolled piece is deformed in the upstream rack, some process parameters will change due to changes in working conditions, which will inevitably cause changes in the rolling force of the downstream rack. The rolling force predicted using the original process parameters is inaccurate, and the rolling force needs to be re-predicted to ensure the quality of the end product.

[0003] With the continuous development of artificial intelligence technology, the use of intelligent algorithms for rolling force prediction has been widely cited. The Chinese patent application with publication number CN111790762B, "A method for setting rolling force for hot-rolled strip steel based on random forests", discloses a method for setting rolling force for hot-rolled strip steel. By randomly combining the SIMS model with the base decision tree, the defect of insufficient generalization ability of the rolling force mechanism model is solved. The Chinese patent application with publication number CN112711867A, "A method for predicting rolling force by integrating theoretical models and big data models", discloses an integrated model, which takes actual production data as the research object, integrates theoretical models and backpropagation neural networks (BP neural networks), and realizes the prediction of rolling force through model integration. The Chinese patent application with publication number CN116689515A, "Rolling force prediction method for finishing rolling based on SA-SCSO-TSVR algorithm", discloses a hot rolling force prediction method based on the SA-SCSO-TSVR algorithm. The Sand Cat Swarm Optimization (SCSO) algorithm is used to optimize the Twin Support Vector Regression (TSVR) model parameters, and the Simulated Annealing (SA) algorithm is used to globally optimize the SCSO-TSVR model to determine the optimal parameters, thereby improving the rolling force prediction accuracy of the finishing rolling process.

[0004] However, the above research processes all predict the rolling force as a static process, assuming that the process parameters used for rolling force prediction are fixed values, and do not consider the impact of changes in process parameters caused by the deformation of the rolled piece through the upstream stand. The rolling forces of all stands are predicted simultaneously, resulting in limited accuracy in rolling force prediction. As a serial production process, the hot rolling process, as the rolling process proceeds, the rolled piece is deformed through the upstream stand, and the real-time data generated contains the latest rolling conditions. Therefore, it is necessary to consider the impact of the latest rolling conditions in the downstream rolling force prediction. Summary of the invention

[0005] In view of the above-mentioned deficiencies in the prior art, the present invention takes the changes in process parameters caused by the deformation of the upstream stand during the rolling process as dimensionality-added features, and uses them together with the measured data of the stand as the input of the rolling force prediction model of the downstream stand. A dynamic prediction method of rolling force based on feature dimensionality increase and deep neural network is proposed. The influence of changes in process parameters on the model accuracy is fully considered, thereby realizing high-precision dynamic prediction of the rolling force of the downstream stand.

[0006] The present invention proposes a rolling force dynamic prediction method based on feature dimension increase and deep neural network, which comprises the following steps:

[0007] Step 1: Collect production data during the hot strip rolling process;

[0008] Step 2: Process the production data and construct a sample data set, and divide the training set from the sample data set according to a preset ratio;

[0009] Step 3: Construct a feature-enhanced convolutional deep neural network FACDNN model;

[0010] Step 4: Determine the loss function and optimizer, set the training parameters, and use the training set to train the feature-enhanced-dimensional convolutional deep neural network FACDNN model to obtain the trained feature-enhanced-dimensional convolutional deep neural network FACDNN model;

[0011] Step 5: re-collect the production data of the hot strip rolling process and input the trained feature-added-dimensional convolutional deep neural network FACDNN model to obtain the rolling force prediction results of the hot strip rolling process;

[0012] The step 1 further comprises:

[0013] Step 1.1: Collect PDA data during the hot strip rolling process, including: composition of rolled product, slab size, intermediate billet size, target size of finished product, roll speed, roll radius and reduction rate;

[0014] Step 1.2: Collect the measured data of the strip passing through each stand and calculate the average value; the measured data of the strip passing through each stand includes: rolling force, tension and deformation resistance;

[0015] Step 1.3: Calculate the unmeasurable data during the hot strip rolling process, including: strip temperature, strip width and rolling force of the downstream stand;

[0016] The step 1.3 further comprises:

[0017] Step 1.3.1: Use the temperature drop model to calculate the heat loss of the strip during the rolling process, and calculate the strip temperature based on the heat loss of the strip;

[0018] Step 1.3.2: Calculate the strip width during rolling using the width expansion model;

[0019] Step 1.3.3: Use SIMS rolling force model to calculate the rolling force of the downstream stand during rolling;

[0020] The temperature drop model described in step 1.3.1 includes: an air cooling temperature drop model, a water cooling temperature drop model, a plastic deformation heat model and a contact temperature drop model;

[0021] The step 2 further comprises:

[0022] Step 2.1: Use linear interpolation method to interpolate missing values ​​of the collected production data;

[0023] Step 2.2: Use the Gaussian mixture model to mark and clean the data in the production data that does not conform to the normal distribution to obtain the cleaned production data;

[0024] Step 2.3: Use the min-max normalization method to normalize the cleaned production data, and use the normalized production data to construct a sample data set;

[0025] Step 2.4: The PDA data and measured data in the sample data set are used as basic features, wherein the basic features are divided into basic features of upstream racks and basic features of downstream racks according to the positions of the racks; the unmeasurable data in the sample data set are used as dimension-increasing features;

[0026] Step 2.5: Divide the training set from the sample data set according to the preset ratio;

[0027] The feature-enhanced-dimensional convolutional deep neural network FACDNN model described in step 3 includes: a FACDNN shallow basic model and a FACDNN deep basic model;

[0028] The FACDNN shallow basic model is used to input the basic features of the upstream rack and the basic features of the downstream rack, and perform nonlinear processing on the input basic features of the upstream rack and the basic features of the downstream rack to obtain the basic feature sequence of the upstream rack and the basic feature sequence of the downstream rack and transmit them to the FACDNN deep basic model;

[0029] The FACDNN deep basic model is used to predict the rolling force of the downstream stand according to the input basic feature sequence of the upstream stand, the basic feature sequence of the downstream stand and the dimension-increased feature, so as to obtain the predicted value of the rolling force of the downstream stand;

[0030] The FACDNN shallow basic model is a dual 1D CNN network, including two sets of parallel 1D CNN structures, respectively denoted as 1D CNN1 and 1D CNN2;

[0031] The 1D CNN1 is used to perform nonlinear processing on the basic features of the input upstream rack, and transmit the basic feature sequence of the output upstream rack to the FACDNN deep basic model;

[0032] The 1D CNN1 includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer and a flattening layer connected in sequence; in the process of constructing the 1D CNN1, the convolution kernel size, the step size, the number of network hidden layer nodes and the activation function of the 1D CNN1 are set;

[0033] The 1D CNN2 is used to perform nonlinear processing on the basic features of the input downstream rack, and transmit the output basic feature sequence of the downstream rack to the FACDNN deep basic model;

[0034] The 1D CNN2 includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer and a fully connected layer connected in sequence; in the process of constructing the 1D CNN2, the convolution kernel size, the step size, the number of network hidden layer nodes and the activation function of the 1D CNN2 are set;

[0035] The FACDNN deep basic model is a deep neural network DNN model, comprising: a first fully connected layer, a first discard layer, a second fully connected layer, a second discard layer, a third fully connected layer, a fourth fully connected layer and a fifth fully connected layer connected in sequence, and a dimension-increasing feature input port parallel to the first discard layer; wherein the second fully connected layer and the fourth fully connected layer are skip-connected;

[0036] The step 4 further comprises:

[0037] Step 4.1: Determine the loss function and optimizer, and set the training parameters; wherein the training parameters include: learning rate, batch size and training cycle;

[0038] Step 4.2: Input the basic features of the upstream rack and the basic features of the downstream rack in the training set into 1D CNN1 and 1D CNN2 in the shallow basic model of FACDNN for nonlinear processing, and obtain the basic feature sequence of the upstream rack and the basic feature sequence of the downstream rack;

[0039] Step 4.3: Input the obtained basic feature sequence of the upstream rack and the basic feature sequence of the downstream rack into the FACDNN deep basic model, and use the first fully connected layer in the FACDNN deep basic model to merge and nonlinearly process the input basic feature sequence of the upstream rack and the basic feature sequence of the downstream rack to obtain the merged basic features;

[0040] Step 4.4: Input the dimension-increased features in the training set into the FACDNN deep basic model through the dimension-increased feature input port, and merge the merged basic features and the dimension-increased features to obtain the merged features;

[0041] Step 4.5: Use the first dropout layer, the second fully connected layer, the second dropout layer, the third fully connected layer and the fourth fully connected layer in the deep neural network DNN model to perform nonlinear processing on the merged features in sequence, and use jump connections to merge the features obtained after the nonlinear processing with the merged features, and use the fifth fully connected layer in the deep neural network DNN model to summarize the merged features to obtain the rolling force prediction value of the downstream stand;

[0042] Step 4.6: Use the loss function and optimizer to update the feature-enhanced-dimensional convolutional deep neural network FACDNN model, repeat steps 4.2-4.5 until the set training cycle is completed, and obtain the trained feature-enhanced-dimensional convolutional deep neural network FACDNN model.

[0043] The beneficial effects of adopting the above technical solution are:

[0044] The method of the present invention designs a feature augmented convolutional deep neural network (FACDNN) model for dynamic prediction of rolling force of hot rolling multi-stands. First, the feature extraction and nonlinear processing of input measured data are completed through two shallow 1D CNN models, and the extracted feature parameters are passed into the deep jump connection DNN network, that is, the FACDNN deep basic model. Then, the calculated change data between the stands, that is, the increased dimension features, are merged with the input feature parameters through the increased dimension port of the FACDNN deep basic model to achieve feature increased dimension. Finally, the rolling force prediction result of the downstream stand is obtained, and the dynamic prediction result of the rolling force of the hot rolling multi-stand is obtained through multiple repetitions.

[0045] The method of the present invention considers the influence of process parameter changes caused by rolling deformation of the rolled piece through the upstream stand on the rolling force of the downstream stand during hot rolling. When the strip is threaded, the actual measured values ​​of the upstream stand and the unmeasurable data values ​​calculated by the formula between the stands are obtained, and the data are input into the feature-increased convolutional deep neural network for prediction through feature-increased dimensionality.

[0046] The method of the present invention proposes a method for dynamic prediction of rolling force based on feature dimension increase and deep neural network, which takes into account the changes in process parameters caused by rolling deformation of the rolled piece through the upstream stand and the influence of unmeasurable parameters between stands, and solves the problem of unmeasurable parameters through mathematical model calculation. As the rolling process proceeds, the process parameters that are affected by the deformation of the upstream stand are used as dimension increase features, and together with the upstream measured data, are used as the input of the downstream rolling force prediction model. When the rolling parameters change, the dynamic prediction of the rolling force of the downstream stand is realized through the proposed specific dimension increase network structure. The final evaluation index shows that the prediction of the model has low error and good fitting effect, and fully considers the influence of changes in process parameters on the accuracy of the model. It has the characteristics of high accuracy and good generalization, and is used to realize the dynamic prediction of the rolling force of each finishing rolling stand in the hot rolling process. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a flow chart of a rolling force dynamic prediction method based on feature dimension increase and deep neural network in this embodiment;

[0048] Figure 2 The schematic diagram of constructing and training the feature-augmented-dimensional convolutional deep neural network FACDNN model in this embodiment;

[0049] Figure 3 It is a structural diagram of the feature dimension-enhanced convolutional deep neural network FACDNN model in this implementation mode;

[0050] Figure 4 This is a rolling force prediction result diagram taking the first stand result as an example in this embodiment. DETAILED DESCRIPTION

[0051] In order to facilitate the understanding of the present application, the specific embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings and embodiments. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present application more thoroughly understood.

[0052] In this embodiment, in order to verify the effectiveness of the Feature Augmentative Convolutional Deep Neural Network (FACDNN) in the dynamic prediction of rolling force during the finishing rolling process, the production data of a hot rolling production process is taken as an example to verify the method.

[0053] A rolling force dynamic prediction method based on feature dimension increase and deep neural network in this embodiment is as follows: Figure 1 As shown, the method comprises the following steps:

[0054] Step 1: Collect production data during the hot strip rolling process.

[0055] Step 1.1: Collect PDA data during the hot strip rolling process, including: composition of rolled product, slab size, intermediate billet size, target size of finished product, roll speed, roll radius and reduction rate.

[0056] In this embodiment, a process data acquisition system (PDA) is used to collect data from the hot strip rolling process. The PDA data acquisition system is often suitable for data acquisition of distributed multi-systems, processing large amounts of data in a unified and synchronous manner, and providing good solutions for different data formats and communication methods, thereby realizing real-time display of data and continuous storage of data.

[0057] Step 1.2: Collect the measured data of the steel strip passing through each stand and calculate the average value; the measured data of the steel strip passing through each stand includes: rolling force, tension and deformation resistance.

[0058] In this embodiment, the rolling force, tension and deformation resistance of the steel strip passing through each stand are collected, and the collected rolling force, tension and deformation resistance are respectively summarized and averaged.

[0059] Step 1.3: Calculate the unmeasurable data during the hot strip rolling process, including: strip temperature, strip width and rolling force of the downstream stand.

[0060] Step 1.3.1: Use the temperature drop model to calculate the heat loss of the strip during rolling, and calculate the strip temperature based on the heat loss of the strip.

[0061] In this embodiment, if Figure 2 As shown, the hot strip rolling process involves air cooling temperature drop, water cooling temperature drop, plastic deformation heat and contact temperature drop. In the above process, heat loss will cause the temperature of the strip to gradually drop during the rolling process. It should be noted that the order of the above four temperature drop processes is not fixed and needs to be determined according to the actual situation of the hot strip rolling process.

[0062] The temperature drop model includes: an air cooling temperature drop model, a water cooling temperature drop model, a plastic deformation heat model and a contact temperature drop model.

[0063] The air cooling temperature drop model is:

[0064] (1)

[0065] in express Heat lost from the inside The heat dissipation time of air cooling temperature drop, in seconds or hours; is the emissivity, also known as blackness, and <1; is the Stefan-Boltzmann constant; is the strip temperature before air cooling, in °C; is the heat dissipation area, in m 2 .

[0066] The water-cooled temperature drop mold is:

[0067] (2)

[0068] in It indicates the amount of heat lost during convection heat transfer; represents the convective heat transfer coefficient; is the strip temperature before air cooling, in °C; is the water temperature, in °C; It is the heat dissipation time of water cooling temperature drop, in seconds or hours.

[0069] The plastic deformation thermal model is:

[0070] (3)

[0071] in Indicates the heat generated by the deformation of the strip; is the heat conversion factor, take 0.9; is the deformation resistance, unit is kg / mm 2 ; Indicates the inlet thickness of the strip, in mm; Indicates the export thickness of the strip, in mm; is the heat-work equivalent, in kcal / kgm; is the specific heat capacity, in kcal / kg℃; is the specific gravity, the unit is kg / m 3 .

[0072] The contact temperature drop model is:

[0073] (4)

[0074] in Indicates the heat loss due to contact between the strip and the roll; The heat conduction efficiency between the rolled piece and the roller is generally 0.48~0.55; Indicates the contact arc length of the deformation zone, in m; is the average thickness of the strip deformation zone, in m; is the roller linear speed, in m / s; is the strip temperature before contact temperature drop, in °C; is the roller temperature, in °C; Expressed as thermal conductivity, unit is m 2 / hr; and:

[0075] (5)

[0076] in is the contact heat transfer coefficient, and its unit is W / m℃.

[0077] Step 1.3.2: Use the width expansion model to calculate the strip width during rolling.

[0078] The broadening model is:

[0079] (6)

[0080] (7)

[0081] in Indicates The outlet width of each rack; Indicates Racks; Indicates The outlet width of each rack; is the reduction rate; is the unit post tension; It represents the total post tension; Indicates the average deformation resistance of the strip.

[0082] In this embodiment, since vibration will be generated when the roller is rolling, the contact arc length of the deformation zone after the roller is flattened is also called the flattening radius.

[0083] Step 1.3.3: Use the SIMS rolling force model to calculate the rolling force of the downstream stand during rolling.

[0084] In this embodiment, if Figure 2 As shown in the figure, in order to improve the prediction accuracy of rolling force prediction using the Feature Augmentative Convolutional Deep Neural Network (FACDNN) model, it is necessary to first use the Sims equation for hot rolling force (SIMS rolling force model) to calculate the rolling force of the downstream stand during the rolling process as part of the training sample.

[0085] The SIMS rolling force model is:

[0086] (8)

[0087] (9)

[0088] (10)

[0089] in is the calculated rolling force of the downstream stand, in kN; is the metal deformation resistance, unit is MPa; is the strip width, in mm; is the stress state coefficient; Indicates the roller flattening radius, in mm; Indicates the amount of compression, in mm; Indicates the roller radius in mm.

[0090] In this embodiment, based on the data obtained in step 1.1 and step 1.2, the temperature drop model, width expansion model and rolling force calculation model are used to calculate the temperature, width and rolling force of the downstream stand during the rolling process.

[0091] Step 2: Process the production data and construct a sample data set, and divide the training set from the sample data set according to a preset ratio.

[0092] Step 2.1: Use linear interpolation method to interpolate missing values ​​of the collected production data.

[0093] The process of using the linear interpolation method to interpolate missing values ​​of the collected production data is as follows: counting all missing data points in the production data, and for any missing data point , determine the two known data points adjacent to the data point and and by and Construct a straight line to calculate the data points The missing values ​​of all missing data points are supplemented into the production data.

[0094] (11)

[0095] in Represents data points The data value of Represents known data points The data value of Represents known data points The data value of .

[0096] Step 2.2: Use the Gaussian mixture model to mark and clean the data in the production data that does not conform to the normal distribution to obtain the cleaned production data.

[0097] The Gaussian mixture model is:

[0098] (12)

[0099] in Represents data points The probability density function of Indicates shared Gaussian distribution; Indicates The weights of the Gaussian distribution; is the probability density function of the multivariate Gaussian distribution, where Indicates The mean vector of the Gaussian distribution; For the The covariance matrix of a Gaussian distribution.

[0100] In this embodiment, a Gaussian mixture model is used to mark data in the production data that does not conform to the normal distribution, that is, to mark the abnormal, erroneous, or special data in the production data, and then clean up the marked data.

[0101] Step 2.3: Use the min-max normalization method to normalize the cleaned production data, and use the normalized production data to construct a sample data set.

[0102] The formula for the normalization process is:

[0103] (13)

[0104] in For sample data Normalized dimensionless data; The sample data set is Sample data; is the minimum value in the original data; is the maximum value in the original data.

[0105] This implementation adopts the min-max normalization method to process data to eliminate the influence of data dimension on the model.

[0106] Step 2.4: The PDA data and measured data in the sample data set are used as basic features, wherein the basic features are divided into basic features of upstream racks and basic features of downstream racks according to rack positions; and the unmeasurable data in the sample data set are used as dimension-increasing features.

[0107] In this implementation, the PDA data and measured data in the sample data set are divided into basic features, as shown in Table 1.

[0108] Table 1 Basic characteristic parameters

[0109] Serial number Input Features Serial number Input Features Serial number Input Features Serial number Input Features Serial number Input Features 1 Inlet thickness 7 Deformation resistance 13 Friction coefficient 19 Roller speed 25 Roller diameter 2 tension 8 Reduction rate 14 Si 20 Nb 26 Sn 3 Cr 9 N 15 Cu 21 Ni 27 Ti 4 As 10 B 16 Fe 22 P 28 V 5 Al 11 C 17 Mn 23 Pb 29 W 6 Ce 12 Ca 18 Mo 24 S 30 Zr

[0110] In this implementation, the inter-rack unmeasurable data is divided into increased-dimensional features, as shown in Table 2.

[0111] Table 2 Dimensional feature parameter table

[0112] Serial number Model input features Serial number Model input features 1 Temperature formula calculated value 3 Width formula calculated value 2 Downstream SIMS formula calculation value

[0113] Step 2.5: Divide the training set from the sample data set according to the preset ratio.

[0114] In this embodiment, the processed data set is divided into a training set and a test set in a ratio of 7:3.

[0115] Step 3: Construct a Feature Augmentative Convolutional Deep Neural Network (FACDNN) model.

[0116] In this embodiment, if Figure 3 As shown in the figure, the Convolutional Neural Networks (CNN) is used as the shallow basic model of FACDNN, specifically a double one-dimensional convolutional neural network (1D Convolutional Neural Network, 1D CNN). 1D CNN automatically extracts local features in the data by using convolution operations to capture the temporal or spatial dependencies in the input basic features. The Deep Neural Network (DNN) is used as the deep basic model of FACDNN. The FACDNN network is deeply connected to the DNN model, and the nonlinear relationship in the input data is captured through a multi-layer structure and nonlinear activation function.

[0117] The feature-dimensionality-increased convolutional deep neural network FACDNN model includes: a FACDNN shallow basic model and a FACDNN deep basic model.

[0118] The FACDNN shallow basic model is used to input the basic features of the upstream rack and the basic features of the downstream rack, and perform nonlinear processing on the input basic features of the upstream rack and the basic features of the downstream rack to obtain the basic feature sequence of the upstream rack and the basic feature sequence of the downstream rack and transmit them to the FACDNN deep basic model.

[0119] The FACDNN shallow basic model is a dual 1D CNN network, including two sets of parallel 1D CNN structures, which are respectively denoted as 1D CNN1 and 1D CNN2.

[0120] The 1D CNN1 is used to perform nonlinear processing on the basic features of the input upstream rack, and transmit the output basic feature sequence of the upstream rack to the FACDNN deep basic model.

[0121] The 1D CNN1 includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer and a flattening layer which are connected in sequence; in the process of constructing the 1D CNN1, the convolution kernel size, step size, number of network hidden layer nodes and activation function of the 1D CNN1 are set.

[0122] In this embodiment, a flatten layer is usually added before the fully connected layer in the convolutional neural network. The function of the flattening layer is to convert the multi-dimensional feature map output by the previous convolutional layer or pooling layer into a one-dimensional vector so as to connect to the fully connected layer. In this embodiment, the convolution kernel size of 1D CNN1 is set to 3×3, the step size is 1, the padding is 0, the number of neuron nodes or channels is 64, and the activation function is Relu.

[0123] The 1D CNN2 is used to perform nonlinear processing on the basic features of the input downstream rack, and transmit the output basic feature sequence of the downstream rack to the FACDNN deep basic model.

[0124] The 1D CNN2 includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer and a fully connected layer connected in sequence; in the process of constructing the 1D CNN2, the convolution kernel size, the step size, the number of network hidden layer nodes and the activation function of the 1D CNN2 are set.

[0125] In this embodiment, the convolution kernel size of 1D CNN2 is 3×3, the step size is 1, the padding is 0, the number of neuron nodes or channels is 64, and the activation function is Relu.

[0126] In this embodiment, the convolution operation performed in the 1D CNN can be expressed as:

[0127] (14)

[0128] in Indicates the first elements; is the first character sequence of the input elements; is the convolution kernel weights, is the size of the convolution kernel.

[0129] The FACDNN deep basic model is a deep neural network (DNN) model, which is used to predict the rolling force of the downstream stand according to the input basic feature sequence of the upstream stand, the basic feature sequence of the downstream stand and the dimension-increased feature, and obtain the predicted rolling force value of the downstream stand.

[0130] The deep neural network DNN model includes: a first fully connected layer, a first discard layer, a second fully connected layer, a second discard layer, a third fully connected layer, a fourth fully connected layer and a fifth fully connected layer connected in sequence, and a dimension-increasing feature input port parallel to the first discard layer; wherein the second fully connected layer and the fourth fully connected layer have a jump connection.

[0131] In this embodiment, the dimension-increasing feature is received by creating a new dimension-increasing feature input port in the deep DNN structure. In order to avoid the problem of error back propagation caused by the network structure being too long, such as gradient disappearance and gradient explosion, the generalization of the deep neural network DNN model is increased by adding a jump connection module, i.e., a jump connection. The jump connection module is used to transmit the input parameters to the output node through two paths. The specific mathematical expression is as follows.

[0132] (15)

[0133] in is the residual mapping; For jump connection; is the output vector after skip connection.

[0134] In this embodiment, the specific network layer structure in the FACDNN model is shown in Table 3.

[0135] Table 3 Network structure parameter table

[0136] 1D CNN1 1D CNN2 DNN Convolutional Layer Convolutional Layer Fully connected layer Pooling Layer Pooling Layer Dropout Layer Convolutional Layer Convolutional Layer Fully connected layer Pooling Layer Pooling Layer Dropout Layer Flatten Layer Fully connected layer Fully connected layer - - Fully connected layer - - Fully connected layer

[0137] In this embodiment, the number of network channels, the number of fully connected layer nodes and the activation function of each network layer in the FACDNN model are shown in Table 4.

[0138] Table 4 Network layer parameter table

[0139] Number of neuron nodes / channels Activation Function 1D CNN1 64 Relu 1D CNN2 64 Relu DNN 120 Relu

[0140] Step 4: Determine the loss function and optimizer, set the training parameters, and use the training set to train the feature-enhanced-dimensional convolutional deep neural network FACDNN model to obtain the trained feature-enhanced-dimensional convolutional deep neural network FACDNN model.

[0141] Step 4.1: Determine the loss function and optimizer, and set the training parameters; the training parameters include: learning rate, batch size, and training cycle.

[0142] In this embodiment, the loss function is determined to be the mean square error (MSE) loss, and the optimizer is determined to be Adam (Adaptive Moment Estimation). The final training parameters are determined based on multiple experiments, the learning rate is 0.01, the number of training batches is 128, and the training cycle is set to 300 times.

[0143] Step 4.2: Input the basic features of the upstream rack and the basic features of the downstream rack in the training set into 1D CNN1 and 1D CNN2 in the shallow basic model of FACDNN for nonlinear processing, and obtain the basic feature sequence of the upstream rack and the basic feature sequence of the downstream rack.

[0144] In this embodiment, when the strip is threaded, the processed basic features of the rack being threaded and the basic features of the downstream rack are respectively input into the two 1D CNN structures in the FACDNN model for feature nonlinear processing, and then passed into the DNN model.

[0145] Step 4.3: Input the obtained basic feature sequence of the upstream rack and the basic feature sequence of the downstream rack into the FACDNN deep basic model, and use the first fully connected layer in the FACDNN deep basic model to merge and nonlinearly process the input basic feature sequence of the upstream rack and the basic feature sequence of the downstream rack to obtain the merged basic features.

[0146] Step 4.4: Input the increased-dimensional features in the training set into the FACDNN deep basic model through the increased-dimensional feature input port, and merge the merged basic features and the increased-dimensional features to obtain the merged features.

[0147] In this embodiment, if Figure 3 As shown in the figure, T represents the calculated value of the strip temperature, F cal represents the calculated value of rolling force, and W represents the calculated value of strip width.

[0148] The process of merging the basic feature sequence of the upstream rack, the basic feature sequence of the downstream rack and the dimension-added features by the FACDNN deep basic model is expressed as:

[0149] (16)

[0150] in Represents the dimension-increasing feature; Represents the combined basic features; Represents the basic feature sequence after being processed by the 1D CNN structure; Represents the merged features.

[0151] In this embodiment, when the strip is threaded, the dimension-increasing features obtained by formula calculation in the training set are directly input into the DNN model through the dimension-increasing input port, and are combined with the basic features obtained after nonlinear processing. Merge, the input received by the fully connected layer in the DNN model is .

[0152] Step 4.5: Use the first dropout layer, the second fully connected layer, the second dropout layer, the third fully connected layer and the fourth fully connected layer in the deep neural network DNN model to perform nonlinear processing on the merged features in turn, and use jump connections to merge the features obtained after the nonlinear processing with the merged features, and use the fifth fully connected layer in the deep neural network DNN model to summarize the merged features to obtain the rolling force prediction value of the downstream stand.

[0153] In this embodiment, the merged features Through two paths, one path is through the skip connection, and the other path is through the DNN hidden layer to the output node in the deep neural network DNN model. The DNN hidden layer is implemented through the activation functions and bias variables of multiple fully connected layers. The nonlinear processing of The output results are directly combined with the output results of the DNN hidden layer, and the combined output results are activated using the activation function of the last fully connected layer in the deep neural network DNN model to obtain the final rolling force prediction value of the downstream stand.

[0154] Step 4.6: Use the loss function and optimizer to update the feature-enhanced-dimensional convolutional deep neural network FACDNN model, repeat steps 4.2-4.5 until the set training cycle is completed, and obtain the trained feature-enhanced-dimensional convolutional deep neural network FACDNN model.

[0155] Step 5: Re-collect the production data during the hot strip rolling process and input it into the trained feature-added dimension convolutional deep neural network FACDNN model to obtain the rolling force prediction results during the hot strip rolling process.

[0156] In this embodiment, if Figure 4 As shown in the figure, taking the results of the first stand as an example, the FACDNN network is used to predict the rolling force, and the mean square error (MSE), root mean square error (RMSE) and determination coefficient are selected. (R-Square) is used as the result evaluation index. The specific mathematical expression is:

[0157] (1) Coefficient of determination:

[0158] (17)

[0159] in For the The predicted value of samples; For the The actual value of samples; is the total number of samples.

[0160] The closer the value is to 1, the stronger the explanatory power of the model is and the better the fitting effect is.

[0161] (2) Root mean square error:

[0162] (18)

[0163] One of the regression model evaluation indicators. The smaller the value, the higher the model prediction performance.

[0164] (3) Mean absolute error:

[0165] (19)

[0166] One of the regression model evaluation indicators. The smaller the value, the higher the model prediction performance.

[0167] The test set obtained in step 2.4 is input into the trained model for prediction, and the model prediction results and actual values ​​are evaluated by indicators. The prediction results and indicator evaluation are shown in Table 5.

[0168] Table 5 Test data Rolling force prediction performance evaluation

[0169]

[0170] The continuous prediction model proposed in this embodiment, namely the feature-added-dimensional convolutional deep neural network FACDNN model, performs better in MAE, RMSE and All of them maintained good prediction effects, indicating that in the hot rolling process, fully considering the inter-stand factors and the influence of upstream rolling parameters can improve the accuracy of predicting rolling force.

[0171] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A rolling force dynamic prediction method based on feature dimension increase and deep neural network, characterized in that: The method comprises the following steps: Step 1: Collect production data during the hot strip rolling process; Step 2: Process the production data and construct a sample data set, and divide the training set from the sample data set according to a preset ratio; Step 3: Construct a feature-enhanced convolutional deep neural network FACDNN model; The feature-enhanced-dimensional convolutional deep neural network FACDNN model includes: a FACDNN shallow basic model and a FACDNN deep basic model; The FACDNN shallow basic model is used to input the basic features of the upstream rack and the basic features of the downstream rack in the sample data set, and perform nonlinear processing on the input basic features of the upstream rack and the basic features of the downstream rack to obtain the basic feature sequence of the upstream rack and the basic feature sequence of the downstream rack and transmit them to the FACDNN deep basic model; The FACDNN shallow basic model is a dual 1D CNN network, including two sets of parallel 1D CNN structures, respectively denoted as 1D CNN1 and 1D CNN2; The 1D CNN1 is used to perform nonlinear processing on the basic features of the input upstream rack, and transmit the basic feature sequence of the output upstream rack to the FACDNN deep basic model; The 1D CNN1 includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer and a flattening layer connected in sequence; in the process of constructing the 1D CNN1, the convolution kernel size, the step size, the number of network hidden layer nodes and the activation function of the 1D CNN1 are set; The 1D CNN2 is used to perform nonlinear processing on the basic features of the input downstream rack, and transmit the output basic feature sequence of the downstream rack to the FACDNN deep basic model; The 1D CNN2 includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer and a fully connected layer connected in sequence; in the process of constructing the 1D CNN2, the convolution kernel size, the step size, the number of network hidden layer nodes and the activation function of the 1D CNN2 are set; The FACDNN deep basic model is used to predict the rolling force of the downstream stand according to the input basic feature sequence of the upstream stand, the basic feature sequence of the downstream stand and the dimension-increased feature, so as to obtain the predicted value of the rolling force of the downstream stand; The FACDNN deep basic model is a deep neural network DNN model, comprising: a first fully connected layer, a first discard layer, a second fully connected layer, a second discard layer, a third fully connected layer, a fourth fully connected layer and a fifth fully connected layer connected in sequence, and a dimension-increasing feature input port parallel to the first discard layer; wherein the second fully connected layer and the fourth fully connected layer are skip-connected; Step 4: Determine the loss function and optimizer, set the training parameters, and use the training set to train the feature-enhanced-dimensional convolutional deep neural network FACDNN model to obtain the trained feature-enhanced-dimensional convolutional deep neural network FACDNN model; Step 5: Re-collect the production data during the hot strip rolling process and input it into the trained feature-added dimension convolutional deep neural network FACDNN model to obtain the rolling force prediction results during the hot strip rolling process.

2. According to claim 1, a rolling force dynamic prediction method based on feature dimension increase and deep neural network is characterized in that: The step 1 further comprises: Step 1.1: Collect PDA data during the hot strip rolling process, including: composition of rolled product, slab size, intermediate billet size, target size of finished product, roll speed, roll radius and reduction rate; Step 1.2: Collect the measured data of the strip passing through each stand and calculate the average value; the measured data of the strip passing through each stand includes: rolling force, tension and deformation resistance; Step 1.3: Calculate the unmeasurable data during the hot strip rolling process, including: strip temperature, strip width and rolling force of the downstream stand.

3. According to claim 2, a rolling force dynamic prediction method based on feature dimension increase and deep neural network is characterized in that: The step 1.3 further comprises: Step 1.3.1: Use the temperature drop model to calculate the heat loss of the strip during the rolling process, and calculate the strip temperature based on the heat loss of the strip; Step 1.3.2: Calculate the strip width during rolling using the width expansion model; Step 1.3.3: Use the SIMS rolling force model to calculate the rolling force of the downstream stand during rolling.

4. According to claim 3, a rolling force dynamic prediction method based on feature dimension increase and deep neural network is characterized in that: The temperature drop model described in step 1.3.1 includes: an air cooling temperature drop model, a water cooling temperature drop model, a plastic deformation heat model and a contact temperature drop model.

5. According to claim 3, a rolling force dynamic prediction method based on feature dimension increase and deep neural network is characterized in that: The step 2 further comprises: Step 2.1: Use linear interpolation method to interpolate missing values ​​of the collected production data; Step 2.2: Use the Gaussian mixture model to mark and clean the data in the production data that does not conform to the normal distribution to obtain the cleaned production data; Step 2.3: Use the min-max normalization method to normalize the cleaned production data, and use the normalized production data to construct a sample data set; Step 2.4: The PDA data and measured data in the sample data set are used as basic features, wherein the basic features are divided into basic features of upstream racks and basic features of downstream racks according to the positions of the racks; the unmeasurable data in the sample data set are used as dimension-increasing features; Step 2.5: Divide the training set from the sample data set according to the preset ratio.

6. The rolling force dynamic prediction method based on feature dimension increase and deep neural network according to claim 5 is characterized in that: The step 4 further comprises: Step 4.1: Determine the loss function and optimizer, and set the training parameters; wherein the training parameters include: learning rate, batch size and training cycle; Step 4.2: Input the basic features of the upstream rack and the basic features of the downstream rack in the training set into 1D CNN1 and 1D CNN2 in the shallow basic model of FACDNN for nonlinear processing, and obtain the basic feature sequence of the upstream rack and the basic feature sequence of the downstream rack; Step 4.3: Input the obtained basic feature sequence of the upstream rack and the basic feature sequence of the downstream rack into the FACDNN deep basic model, and use the first fully connected layer in the FACDNN deep basic model to merge and nonlinearly process the input basic feature sequence of the upstream rack and the basic feature sequence of the downstream rack to obtain the merged basic features; Step 4.4: Input the dimension-increased features in the training set into the FACDNN deep basic model through the dimension-increased feature input port, and merge the merged basic features and the dimension-increased features to obtain the merged features; Step 4.5: Use the first dropout layer, the second fully connected layer, the second dropout layer, the third fully connected layer and the fourth fully connected layer in the deep neural network DNN model to perform nonlinear processing on the merged features in sequence, and use jump connections to merge the features obtained after the nonlinear processing with the merged features, and use the fifth fully connected layer in the deep neural network DNN model to summarize the merged features to obtain the rolling force prediction value of the downstream stand; Step 4.6: Use the loss function and optimizer to update the feature-enhanced-dimensional convolutional deep neural network FACDNN model, repeat steps 4.2-4.5 until the set training cycle is completed, and obtain the trained feature-enhanced-dimensional convolutional deep neural network FACDNN model.

Images