A rolling mill vibration time sequence prediction method based on multi-level network fusion

Abstract

The application provides a rolling mill vibration time sequence prediction method based on multi-level network fusion, and relates to the technical field of metal rolling mills. First, long and short period characteristics in multi-dimensional data are extracted by using a one-dimensional convolutional neural network, then the short period and long period data are trained by using a recurrent neural network and a time convolution network respectively, an attention mechanism is introduced to prevent the decline of calculation efficiency, and finally two kinds of modal characteristics are fused to output a prediction result. The application fully utilizes the characteristics that all rules of a rolling production process are contained in industrial big data and the advantages of convolutional neural network feature extraction, and compared with existing rolling mill vibration prediction models, the prediction accuracy of the fused prediction model is greatly improved, and the proposed model has good credibility.

Description

Technical Field

[0001] This invention relates to the field of metal rolling mill technology, and in particular to a rolling mill vibration timing prediction method based on multi-level network fusion. Background Technology

[0002] Stable operation of strip mills is a key technology for ensuring stable production processes and consistent strip quality. Mill vibration directly affects the thickness accuracy of strip steel, and severe vibration can even cause strip breakage accidents. Mill vibration has long been a major challenge restricting the improvement of steel production efficiency and product quality, and it remains a pressing issue that cold continuous rolling production lines urgently need to address.

[0003] Currently, significant progress has been made in the study of vibration mechanisms based on the rolling mill mass-spring-damping system. However, establishing a research model for the rolling mill vibration mechanism requires extensive expert experience and knowledge. The model-building process is complex, and the rolling process parameters exhibit strong coupling and nonlinearity, making model solving difficult. Numerous assumptions are required when establishing the mechanism model, limiting the accuracy of the calculation results. With advancements in modern information collection and intelligent technologies, many scholars have utilized data-driven methods combined with machine learning to study rolling mill vibration in rolling production. The rolling production process data is highly complex and strongly coupled. Previous studies that only performed noise reduction on the process data were insufficient to analyze the temporal characteristics of long-term and short-term modes, significantly impacting the accuracy of prediction models. Summary of the Invention

[0004] The technical problem this invention aims to solve is to address the shortcomings of existing technologies by providing a rolling mill vibration time series prediction method based on multi-level network fusion. This method fully leverages the inherent regularity of the entire rolling production process within industrial big data, utilizing convolutional neural networks for feature extraction. It extracts long-cycle and short-cycle data from the rolling data, trains them separately, and then combines the two modalities to output the final prediction result. This fusion produces a multi-dimensional, multi-modal rolling mill vibration time series prediction model based on a multi-level deep fusion method, significantly improving prediction accuracy compared to existing rolling mill vibration prediction models.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A rolling mill vibration time-series prediction method based on multi-level network fusion first uses a one-dimensional convolutional neural network to extract long and short-period features from multi-dimensional data. Then, a recurrent neural network and a temporal convolutional network are used to train the short-period and long-period data respectively. An attention mechanism is introduced to prevent a decrease in computational efficiency. Finally, the two modal features are fused to output the prediction result. Specific content includes: Step 1: Data acquisition and preprocessing during the rolling process; Step 2: Based on the multi-modal and multi-dimensional data, a multi-dimensional and multi-modal rolling mill vibration signal time series prediction model is established to address the two aspects of the influence of multi-dimensional vibration-induced process parameters on vibration state and the mixing of long and short period modes of vibration signal. The network fusion structure of this model includes a multi-dimensional long and short period feature extraction layer, a recurrent neural network layer, a temporal convolutional neural network layer, an attention mechanism layer, and a feature fusion layer. Step 3: Construct a dataset for the established model and optimize the model's hyperparameters to ensure the time series data's temporal sequence, thereby obtaining the results of the prediction model established in Step 2.

[0006] Furthermore, step 1 specifically includes: Step 1.1: During rolling production, the rolling process parameters and vibration signals at the same starting moment are collected using the Production Data Acquisition (PDA) system and the vibration data acquisition system.

[0007] Step 1.2: Preprocess the collected rolling data and vibration data, including data time matching, outlier removal, filling missing values, and data normalization.

[0008] Further, step 1.2 specifically includes: Step 1.2.1: Data timing matching. The timing of strip rolling and shearing in the PDA system and the timing of a sharp increase in vibration acceleration amplitude in the vibration signal acquisition system are used as the basis for timing matching between the rolling process parameters and the vibration acceleration signal, thus completing the data timing matching. Step 1.2.2: Outlier removal from time series data, using 3... σ The criterion for removing outliers from vibration data is as follows: (1) (2) In the formula, y i This represents the original vibration acceleration value; This represents the average vibration acceleration. σ y The standard deviation of the vibration acceleration; n Indicates the number of samples; Step 1.2.3: Fill in the missing values ​​in the time series data using linear interpolation. The calculation formula is as follows: (3) In the formula, x 1 and x2 represents two known data points before and after the missing value. x Indicates the interpolation position. y This indicates the numerical value corresponding to the interpolation position. y 1 and y 2 represents the numerical value corresponding to the two known data points before and after the missing value; Step 1.2.4: Data normalization. Select the maximum and minimum value normalization method. The calculation criteria are as follows: (4) In the formula, x It is the raw data. x ' is the normalized data, x max and x min These are the maximum and minimum values ​​of the original data, respectively.

[0009] Furthermore, step 2 specifically includes: Step 2.1: The first layer of the network structure uses a one-dimensional convolutional neural network to extract long-period and short-period modes and local dependencies between variables in the time dimension. Convolution operations are performed on the short-period mode dataset and the long-period mode dataset respectively, and the extracted features are used as the input of the next layer of the network. A one-dimensional convolutional neural network consists of multiple convolutional layers and pooling layers, as well as a fully connected layer for classification or regression. In the convolutional layers, assuming the input data is a one-dimensional vector x of length n, a convolution kernel w of length k slides along the length of x to perform convolution calculations, resulting in a feature vector c of length n-k+1. i This indicates the result of the convolution kernel starting to act from the i-th position of x: (5) Then, in the pooling layer, max pooling or average pooling is performed on each feature to reduce the feature size and further extract features; the formula for max pooling is as follows: (6) In the formula, x is the input data, y is the result after pooling, s is the pooling step size, and p is the pooling window size; Finally, the features are passed to the fully connected layer to obtain the final output; Step 2.2: The second layer of the network structure is trained to predict short-period and long-period datasets respectively; the output of the short-period pattern convolutional layer is input into the Long Short-Term Memory Neural Network (LSTM) for training; Assume the input vector x t and the hidden state h from the previous moment t-1After linear transformation and activation function processing, the outputs of the input gate, forget gate, and output gate are represented as follows: (7) (8) (9) In the formula, σ It is the sigmoid function, x t h represents the input vector at time t. t Let i represent the hidden state at time t. t f t and o t W represents the outputs of the input gate, forget gate, and output gate, respectively. i W f W o These are the weight matrices for the input gate, forget gate, and output gate, respectively. i b f b o These are the bias vectors for the input gate, forget gate, and output gate, respectively. The formula for updating the memory state is as follows: (10) (11) in, This represents the vector dot product operation; tanh is the hyperbolic tangent function, which controls the memory and forgetting of information; the output i of the input gate... t The output f of the forget gate controls the input of new information into the memory state. t Controlling the forgetting of old information from memory; Finally, the update formula for the hidden state is as follows: (12) Output gate output o t Control the output of information retrieved from the memory state to the hidden state; Step 2.3: The output of the long-period convolutional layer is fed into the temporal convolutional network (TCN) for learning; In TCN, residual modules are used instead of convolutional layers. Let the input be x, the network mapping relationship be H(x), and the formula for calculating residual connections is: f(x)=H(x)-x (15) In the formula, f(x) represents the residual; Step 2.4: The third layer of the network structure is the attention mechanism layer. The attention mechanism allows for more direct dependencies between the states of the model at different time points. The feedforward attention mechanism model is as follows: Hidden state sequence h t The vector input in the middle is fed into the learnable function a(h) t In this process, a series of nonlinear transformations are used to map the hidden state sequence to an output vector e. t : e t= a(h t (16) The softmax function is used to transform the output vector into a probability distribution, which represents the attention weights of each element in the input vector to produce the probability vector. α t : (17) Calculate h t Weighted average: (18) By analyzing each time step T The hidden states are weighted to represent the importance of each time step; Step 2.5: The fourth layer of the network structure is the feature fusion layer, which fuses the long-period patterns of the time series extracted by the TCN part and the short-period patterns extracted by the LSTM part to output the results of the prediction model.

[0010] Furthermore, step 3 specifically includes: Step 3.1: Construct the dataset collected by the model. Time series prediction problems need to ensure the temporal sequence of the data; The input dataset is constructed using a sliding window transformation, which divides the time series dataset into a series of fixed-size sliding windows. Each sliding window is used as a sample to train the model, and then the model is used to predict the next one or more time points. Suppose that a series of time series data observed during the operation of the rolling mill is Y={y1,y2,…,y…} T},in , n The representative feature dimensions include a measured characteristic of mill vibration acceleration and nine variables closely related to mill vibration: tension between the 4th and 5th stands, tension before the 5th stand, rolling speed of the 5th stand, thickness of cold-rolled incoming material, strip thickness at the exit of the 4th stand, strip thickness at the exit of the 5th stand, rolling speed of the 4th stand, rolling force of the 5th stand, and roll gap value of the 5th stand. Let y t,1 Indicates the vibration amplitude of the rolling mill; Define a window size based on the length of the time series dataset. m and prediction step size iStarting from the beginning of the time series dataset, slide the data using the window size, with each slide typically having a step size of 1; the original dataset will then be constructed as follows: (19) In the formula, X This represents the historical data input matrix, with an input window length of m. Y This indicates the predicted vibration acceleration output matrix, with a prediction length of i; y m+1,1 express m The vibrational acceleration at time +1; Using the sliding window method described above, vibration process parameters and vibration acceleration time-series data are transformed into supervised learning datasets of historical data and future vibration acceleration signals; each row of matrix X serves as input, corresponding to the same row... Y As output, that is, when {y1,y2,…,y T Given the information, predict y. T+i,1 The amplitude of vibration at time {y1,y2,…,y T , y T+1 Given the information, predict y. T+i+1,1 The amplitude of vibration at any given moment; therefore, the time step is... T The input matrix at is represented as The output matrix of the vibration amplitude prediction result is represented as Y={ y T+1,1 , y T+2,1 ,…,y T+i,1}; Step 3.2: Optimize the hyperparameters of the model used in Step 2; the CNN feature extraction module, TCN long-term pattern extraction module, and attention mechanism module directly use fixed parameters; Hyperparameter optimization is performed on the LSTM short-period pattern extraction module. Parameters requiring optimization in the LSTM module include the input window time step. T Number of LSTM layers n and the number of hidden layer neurons p ; where the input window time step length T With the length of the sliding window m The same; the grid search method was used to optimize the three variables, and the formula determined the range of the three parameters; (20) The dataset was divided chronologically into a training:validation:test set ratio of 6:2:2. The Adam optimizer was used with dropout=0.2. During model training, the validation set was used to avoid overfitting. While training on the training set, the model was continuously validated using the validation set, and the validation error was calculated. If the validation error did not decrease by 0.00001 after 20 iterations, model training was forcibly stopped. K After training, the final test set RMSE is used as the evaluation criterion, and the obtained hyperparameters are input into the multi-level network model for prediction.

[0011] The beneficial effects of adopting the above technical solution are as follows: The rolling mill vibration time series prediction method based on multi-level network fusion provided by this invention fully utilizes the characteristics of industrial big data containing all the laws of the rolling production process and the advantages of convolutional neural network feature extraction. It establishes a multi-dimensional and multi-modal rolling mill vibration time series prediction model based on convolutional neural network. The model network structure can be divided into four layers: multi-dimensional long and short period modal feature extraction layer, recurrent neural network layer, temporal convolutional neural network layer, attention mechanism layer, and feature fusion layer. Compared with the existing rolling mill vibration prediction model, the accuracy of the prediction results of the fused prediction model is greatly improved, and the proposed model has good credibility. Attached Figure Description

[0012] Figure 1 A schematic diagram of a vibration acceleration sensor provided in an embodiment of the present invention; Figure 2 A schematic diagram of the sensor magnetic base provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the network structure of the multi-dimensional, multi-modal vibration signal time series prediction model provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of a long short-term memory neural network structure provided in an embodiment of the present invention; Figure 5 A schematic diagram of a temporal convolutional neural network structure provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the feedforward attention mechanism network structure provided in an embodiment of the present invention; Figure 7 A schematic diagram illustrating the change of RMSE with neurons and time steps when an LSTM layer is stacked in an embodiment of the present invention. Figure 8 A schematic diagram illustrating the variation of RMSE with neurons and time steps in the case of two stacked LSTM layers provided in an embodiment of the present invention. Figure 9 This is a schematic diagram illustrating how RMSE changes with training batch size, provided in an embodiment of the present invention. Figure 10 A schematic diagram illustrating the change of the evaluation index RMSE with time step length as the prediction step length increases under the ten-step prediction condition provided in this embodiment of the invention. Figure 11 A schematic diagram illustrating the change of the evaluation index MAPE with time step length under the ten-step prediction condition provided in this embodiment of the invention. Figure 12 A schematic diagram illustrating the change of the evaluation index MAE with time step length under the ten-step prediction condition provided in this embodiment of the invention. Figure 13 The evaluation index R increases with the prediction step size under the ten-step prediction condition provided in the embodiments of the present invention. 2 A schematic diagram showing the change over the time step; Figure 14 A scatter plot of the vibration acceleration prediction effect of the MDMRVM model provided in this embodiment of the invention; Figure 15 A line graph showing the predicted and actual vibration acceleration values ​​provided in this embodiment of the invention; Figure 16 This is a normal distribution diagram of the model residuals provided in an embodiment of the present invention; Figure 17 The residual quantile diagram of the model provided in the embodiment of the present invention; Figure 18 The image shows the autocorrelation function of the model residuals provided in this embodiment of the invention. Detailed Implementation

[0013] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0014] like Figure 1 As shown, the method of this embodiment is described below.

[0015] Step 1: Data acquisition and preprocessing during the rolling process, specifically including: Step 1.1: During the rolling production process, data from the cold rolling process at the same moment are collected using a Production Data Acquisition (PDA) system and a vibration data acquisition system. This embodiment mainly focuses on the vertical vibration acceleration signal of the cold rolling mill. The vibration sensor used in this embodiment is as follows: Figure 1 and Figure 2As shown in Table 1, the specific parameters are as follows. In this embodiment, the vibration sensor is mounted on the outer end of the work roll bearing housing on the operating side via a magnetic base, avoiding collisions between the sensor and the bearing housing, while ensuring the accuracy and stability of the collected vibration signals. Real-time data from the rolling process, such as rolling speed, front tension, back tension, rolling force, and stand entry thickness, are selected as input features for the model. The collected data is shown in Table 2.

[0016] Table 1 Vibration sensor parameters Value 100 0~10000 ±50 -50~121

[0017] Table 2 Input characteristics of vibration time series prediction model 1 Rolling speed of the 5th stand m / s 2 Tension between frames 4 and 5 kN 3 5th frame front tension kN 4 Incoming material thickness mm 5 Thickness of strip steel at the exit of frame 4 mm 6 Thickness of strip steel at the exit of frame 5 mm 7 Rolling speed of the 4th stand m / s 8 Rolling force of the 5th stand kN 9 Roll gap value of the 5th frame mm 10 Vibration acceleration of the working rolls on the 5th frame g Step 1.2: Since the data information of the rolling mill real-time monitoring system mainly comes from different measurement systems, the original data may have problems such as missing data, outliers, and large time delays between data. Therefore, it is necessary to preprocess the collected rolling process data and vibration data.

[0018] Step 1.2.1: Data Timing Matching. The timing of strip rolling and shearing in the PDA system and the moment when the vibration acceleration amplitude in the vibration signal acquisition system increases sharply are used as the basis for timing matching between the rolling process parameters and the vibration acceleration signal, thus completing the data timing matching.

[0019] Step 1.2.2: Outlier Removal. In this embodiment, 3... σ The criterion is to remove outliers from the vibration data. The formula for this method is as follows: (1) (2) In the formula, y i This represents the original vibration acceleration value; This represents the average vibration acceleration. σ y The standard deviation of the vibration acceleration; n Indicates the number of samples.

[0020] Step 1.2.3: Filling in missing values. Removing outliers can lead to missing data, affecting the continuity of the time series and interfering with data analysis. Therefore, this embodiment uses linear interpolation to fill in missing values. The calculation formula is as follows: (3) In the formula, x 1 and x 2 represents two known data points before and after the missing value. x Indicates the interpolation position. yThis indicates the numerical value corresponding to the interpolation position. y 1 and y 2 represents the numerical values ​​corresponding to the two data points before and after the missing value.

[0021] Step 1.2.4: Data Normalization. The data obtained from the rolling mill vibration signal acquisition system and the rolling process data acquisition system have inconsistent dimensions. Therefore, data normalization is necessary to eliminate the influence between data with different dimensions and value ranges, making it more suitable for data analysis and modeling. The data normalization method selects maximum and minimum value normalization, and the calculation criteria are as follows: (4) In the formula, x It is the raw data. x ' is the normalized data, x max and x min These are the maximum and minimum values ​​of the original data, respectively.

[0022] Step 2: Addressing the dual challenges of the influence of multi-dimensional vibration-induced process parameters on vibration states and the mixing of long and short period modes in vibration signals, this embodiment establishes a multi-dimensional multi-modal rolling mill vibration time series prediction model (MDMRVM). This network fusion structure includes a multi-dimensional long and short period feature extraction layer, a recurrent neural network layer, a temporal convolutional neural network layer, an attention mechanism layer, and a feature fusion layer. The network structure is as follows: Figure 3 As shown. Specific content includes: Step 2.1: The first layer of the network structure uses a one-dimensional convolutional neural network to extract long-period and short-period modes in the time dimension, as well as the local dependencies between variables. Figure 3 The box with a solid line represents the extraction of short-period pattern data with a short time step span. Figure 3 The boxes marked with short dashed lines "----" indicate the extraction of long-period pattern data spanning a relatively long time step. Convolution operations are performed on the two datasets respectively, and the extracted features are used as inputs to the next layer of the network.

[0023] One-dimensional convolutional neural networks (1D CNNs) are a type of convolutional neural network (CNN) primarily used to process one-dimensional sequential data. Compared to traditional fully connected neural networks, 1D CNNs can better handle temporal and spatial local correlations. A 1D CNN consists of multiple convolutional and pooling layers, as well as a fully connected layer for classification or regression. In the convolutional layers, assuming the input data is a one-dimensional vector x of length n, a convolution kernel w of length k slides along the length of x to perform convolution calculations, resulting in a feature vector c of length n-k+1; c i This indicates the result of the convolution kernel starting to act from the i-th position of x: (5) Then, in the pooling layer, max pooling or average pooling is performed on each feature to reduce the feature size and further extract features; the formula for max pooling is as follows: (6) In the formula, x is the input data, y is the pooled result, s is the pooling stride, and p is the pooling window size. Pooling reduces the size of the feature map, thereby reducing the number of model parameters and lowering the risk of overfitting. Finally, the features are passed to a fully connected layer to obtain the final output.

[0024] Step 2.2: The second layer of the network structure is trained to predict short-period and long-period datasets respectively: the output of the short-period pattern convolutional layer is input into the Long Short-Term Memory Neural Network (LSTM) for training.

[0025] The LSTM network, used for processing short-period modal data, is a special type of recurrent neural network (RNN) primarily designed to address problems such as vanishing and exploding gradients in traditional RNNs. The main idea behind LSTM is the introduction of memory units and three gating units. A schematic diagram of the network structure is shown below. Figure 4 As shown. Figure 4 In the middle, x t h represents the input vector at time t. t Let C represent the hidden state at time t. t This represents the memory state at time t. t f t and o t Let represent the outputs of the input gate, forget gate, and output gate, respectively. W is the weight matrix, and b is the bias vector. Assume the input vector x... t and the hidden state h from the previous moment t-1 After linear transformation and activation function processing, the outputs of the input gate, forget gate, and output gate are represented as follows. (7) (8) (9) In the formula, σ It is the sigmoid function. The formula for updating the memory state is as follows: (10) (11) in, This represents the vector dot product operation. tanh is the hyperbolic tangent function, which controls the memory and forgetting of information. The output i of the input gate... tControlling the input of new information into the memory state, the output f of the forget gate t Controlling the forgetting of old information from the memory state.

[0026] Finally, the update formula for the hidden state is as follows: (12) The output o of the output gate t Controlling the output of the information extracted from the memory state into the hidden state.

[0027] Step 2.3: The output of the long-term pattern convolutional layer is input into the temporal convolutional network layer (TCN) for learning.

[0028] The temporal convolutional neural network (TCN) for processing long-term data is a network model based on the convolutional neural network (CNN). Its main feature is the introduction of a one-dimensional time dimension into the CNN, enabling the model to effectively learn long-term dependencies in the sequence. At the same time, techniques such as residual connections and dilated convolutions are used to improve the efficiency and accuracy of the model. The basic structure of the TCN is as Figure 5 shown. Figure 5 In it, x t represents the input vector at time t, k is the size of the convolutional kernel, and d is the dilation factor of the convolutional kernel. Causal convolution in the temporal convolutional neural network is a key technology, which can ensure that only the current and past information is used at each time step, avoiding the leakage of future information. Let the input time series be x=(x1,x2,…,x T ), the output time series be y=(y1,y2,…,y T ), and the size of the convolutional kernel be k, then the calculation formula for y i is: (13) In the formula, w j is the weight of the convolutional kernel. Obviously, this formula can only be calculated when i is greater than or equal to k. For the case of i < k, negative subscripts will appear in x i-j . To avoid this problem, the TCN uses the "causal padding" technique, adding k - 1 zeros at both ends of the input sequence, so that the calculation of x i only depends on x i-k+1 , x i-k+2 ,…,x i , thus avoiding the problem of future information leakage. When directly applying causal convolution to handle long time series problems, only by increasing the network depth can the increase in the view of historical data be achieved, which is somewhat difficult to implement. To eliminate this problem, dilated convolutional networks are adopted in the TCN, which can achieve an exponentially large receptive field. Assuming a one-dimensional sequence, the filter f→{0,…,k - 1}→ , sequence elements s dilated convolution operation F Defined as: (14) In the formula, d It is the expansion factor. k X is the filter size. s−d·i The past direction.

[0029] In this embodiment, residual modules are used instead of convolutional layers in TCN to avoid the gradient vanishing problem and improve the model's generalization ability. Specifically, let the input be x, the network mapping relationship be H(x), and the calculation formula for residual connections is: f(x)=H(x)-x (15) In the formula, f(x) represents the residual. By connecting the residuals, the model can directly learn the residual part, thereby avoiding the gradient vanishing problem and improving the model's generalization ability.

[0030] Step 2.4: The third layer of the network structure is the attention mechanism layer. This layer adaptively captures dependencies across time steps, improving computational efficiency during training and preventing sequence prediction algorithms from exhibiting low computational efficiency and information loss over long prediction times. The attention mechanism allows for more direct dependencies between the model's states at different time points. The feedforward attention mechanism model structure is as follows: Figure 6 As shown. Hidden state sequence h t The vector input in the middle is fed into the learnable function a(h) t In this process, a series of nonlinear transformations are used to map the hidden state sequence to an output vector e. t : e t= a(h t (16) The softmax function is used to transform the output vector into a probability distribution, which represents the attention weights of each element in the input vector to produce the probability vector. α t : (17) Calculate h t Weighted average: (18) The importance of each time step is represented by weighting the hidden state at each time step.

[0031] Step 2.5: The fourth layer of the network structure is the feature fusion layer, which fuses the long-term patterns extracted by the TCN part and the short-term patterns extracted by the LSTM part. During prediction, the prediction results for the next time step or multiple time steps will be influenced by both the long-term and short-term patterns. This layer effectively combines the two patterns, improving the model's prediction accuracy for subsequent time steps.

[0032] Step 3: Construct the dataset for the model and optimize the hyperparameters of the model. This includes: Step 3.1: Construct the dataset collected by the model. Time series prediction problems require ensuring the temporal sequence of the data. Therefore, this embodiment uses a sliding window transformation to construct the input dataset, dividing the time series dataset into a series of fixed-size sliding windows, and using each sliding window as a sample to train the model. Then, the model is used to predict the next one or more time points.

[0033] Suppose that a series of time series data observed during the operation of the rolling mill is Y={y1,y2,…,y…} T},in , n The representative feature dimensions include a measured characteristic of mill vibration acceleration and nine variables closely related to mill vibration: tension between the 4th and 5th stands, tension before the 5th stand, rolling speed of the 5th stand, thickness of cold-rolled incoming material, strip thickness at the exit of the 4th stand, strip thickness at the exit of the 5th stand, rolling speed of the 4th stand, rolling force of the 5th stand, and roll gap value of the 5th stand. Let y t,1 This indicates the vibration amplitude of the rolling mill.

[0034] Define a window size based on the length of the time series dataset. m and prediction step size i Typically, this is a fixed time interval. Starting from the beginning of the time series dataset, a sliding window is used, with each slide typically having a step size of 1. This results in the original dataset being constructed as follows: (19) In the formula, X This represents the historical data input matrix, with an input window length of m. Y This indicates the predicted vibration acceleration output matrix, with a prediction length of i; y m+1,1 express m Vibrational acceleration at time +1.

[0035] Using the sliding window method described above, the vibration-induced process parameters and vibration acceleration time-series data are transformed into a supervised learning dataset of historical data and future vibration acceleration signals. Each row of matrix X serves as input, corresponding to the data in the same row.Y As output, it can be briefly summarized as {y1, y2, ..., y...} T Given the information, predict y. T+i,1 The amplitude of vibration at time {y1,y2,…,y T , y T+1 Given the information, predict y. T+i+1,1 The amplitude of the vibration at any given moment. Therefore, the time step... T The input matrix at is represented as The output matrix of the vibration amplitude prediction result is represented as Y={ y T+1,1 , y T+2,1 ,…,y T+i,1}

[0036] Step 3.2: Optimize the hyperparameters of the model used in Step 2. Using a sliding window model can effectively utilize the autocorrelation of time series data, improving prediction accuracy. To select a suitable window size, hyperparameter optimization is performed in conjunction with the proposed network model. First, some key parameters that need optimization in the model include batch size, CNN module hyperparameters, LSTM short-term pattern extraction module hyperparameters, TCN module parameters, and attention mechanism parameters. The CNN feature extraction, TCN long-term pattern extraction, and attention mechanism modules used in this embodiment have relatively simple structures, and their parameters are directly fixed as shown in Table 3. Without hyperparameter tuning, the impact of each module on prediction accuracy will be verified in subsequent experiments.

[0037] Table 3 Hyperparameters of other module networks TCN 18 5 1、2、4、8、16、32 1 1 Attention mechanism 128 - - - - CNN 24 5 - 1 - The main focus is on hyperparameter optimization of the LSTM short-period pattern extraction module. Some parameters in the LSTM module that need optimization include the input window time step. T Number of LSTM layers n and the number of hidden layer neurons p The input window time step length T With the length of the sliding window m The same. A grid search method was used to optimize the three variables, and the formula determined the range of the three parameters.

[0038] (20) The dataset was divided chronologically into a training:validation:test set ratio of 6:2:2. The Adam optimizer was used with dropout=0.2. During model training, the validation set was used to avoid overfitting. While training on the training set, the model was continuously validated using the validation set, and the validation error was calculated. If the validation error did not decrease by 0.00001 for 20 consecutive iterations, model training was forcibly stopped, i.e., "early-stopping". The model was then... K The training was repeated several times, and the final RMSE on the test set was used as the evaluation criterion.

[0039] Figure 7 , Figure 8 The effects of LSTM time step T, number of neurons p, and number of LSTM stacked layers n on the RMSE metric of the test set are shown. Figure 7 This represents the optimization results for a short time frame with n=1 stacked layers and a relatively small number of neurons. With 100 neurons and a time step of 40, the RMES reaches a minimum value of 0.00169. Figure 8 When the number of stacked layers is 2, the number of neurons is 80, and the time step is 50, the minimum RMSE value is 0.00161.

[0040] Table 4 shows the training set loss, validation set loss, and test set RMSE values ​​during network training with optimal parameters when stacking 1 and 2 layers.

[0041] Table 4 Training metrics for LSTM models with one and two stacked layers. Number of neurons 100 80 Time step 40 50 training set loss 0.00028 0.000046 validation set loss 0.0000649 0.000152 Test set RMSE 0.00169 0.00161 Table 4 clearly shows that with the optimal parameters for a single stack, the training set loss is 0.00028, while the validation set error is 0.0000649, significantly lower than the training set error. This indicates that the model has already overfitted. With the optimal parameters for a double stack, the network training loss metrics perform normally. The optimal parameters for the LSTM module were ultimately determined to be: 2 stacked layers, 80 neurons, and a time step of 50.

[0042] from Figure 9 As can be seen, the root mean square error (RMSE) first decreases and then increases with the increase of the training batch size, reaching its minimum when the training batch size is 128 (log2Batch_Size=7). In summary, the optimal training effect is achieved when the training batch size is 128, at which point the RMSE between the predicted and actual values ​​of the vibration data on the test set is 0.00318, indicating good model parameter optimization.

[0043] Based on the establishment of the above model and the optimization of hyperparameters, the vibration signal of the rolling mill is predicted, and the prediction results are obtained.

[0044] The following section evaluates the results and performance of the established prediction model, including: Step 4.1: To comprehensively compare algorithm performance and determine evaluation metrics, this embodiment selects the correlation coefficient (R²). 2 The root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) are used to evaluate the prediction accuracy of different models. The calculation formulas are as follows: (twenty one) (twenty two) (twenty three) (twenty four) In the formula, N y represents the number of test samples. i and These represent the actual value and the predicted value, respectively. This represents the average value of the test set data.

[0045] Step 4.2: Compare and analyze the prediction results of the established model.

[0046] In this embodiment, the prediction length is set to ten samples after each window. That is, after the model predicts 10 samples, the window moves forward one step to predict the next 10 samples. The model is compared with GRU, LSTM, CNN-LSTM, TCN, the No-Attention Mechanism (NO-ATT) model, and the network model proposed in this embodiment (MDMRVM). The four evaluation metrics are plotted as follows, increasing according to the prediction step size: Figure 10 , Figure 11 , Figure 12 and Figure 13 The line graph shown indicates that the MDMRVM model performs best, with the correlation coefficient R from the first prediction point to the fifth prediction point being the highest. 2 The values ​​all remained above 97%, and the prediction accuracy of the sixth to tenth prediction points was also stable at above 92%. It can be seen that the prediction model proposed in this embodiment is superior to other methods.

[0047] Figure 14 The scatter plot shows the prediction effect of the established prediction model on the mill vibration acceleration in the tenth prediction step. Figure 14 The solid line diagonally represents the error between the predicted and actual vibration acceleration amplitude, which is ±0.005 g. Different colors represent the specific distribution of the error magnitude in the sample data prediction results. Figure 14The study found that the MDMRVM model had the fewest total number of sample points with prediction errors exceeding ±0.005g on the test set, the smallest model prediction error, and a more concentrated sample distribution.

[0048] Step 4.3: Analyze the generalization ability and correlation of the prediction model.

[0049] The generalization effect of the model can be observed by plotting a line graph of the test set. Based on the above analysis, the prediction accuracy decreases as the prediction step size increases. A scatter plot is then created using the tenth prediction step value, which has the lowest prediction accuracy. Since the test set has a large number of samples, 100 samples are selected for easy viewing when plotting the scatter plot. Figure 15 As shown in the figure, the test set line graph shows that the model performs well in step 10, and can predict the magnitude of mill vibration relatively accurately overall.

[0050] The following results were obtained through residual analysis: Figure 16 , Figure 17 , Figure 18 The residual distribution plot and residual autocorrelation function plot are shown. Observe. Figure 17 It can be observed that the predicted residuals of the model exhibit a normal distribution. Figure 17 The quantiles of the model residuals and the quantiles of the standard normal distribution show a linear relationship. The autocorrelation function is plotted by calculating the autocorrelation of the model residuals. Figure 18 In the model, the autocorrelation coefficient of the residuals is generally less than 0.2, and it approaches 0 as the delay increases, indicating that the residuals exhibit random characteristics.

[0051] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions 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 mill vibration time series prediction method based on multi-level network fusion, characterized in that: The method first uses a one-dimensional convolutional neural network to extract long and short period features from multi-dimensional data, then uses a recurrent neural network and a temporal convolutional network to train the short period and long period data respectively, introduces an attention mechanism to prevent a decrease in computational efficiency, and finally fuses the two modal features to output the prediction result. The specific content includes: Step 1: Data acquisition and preprocessing during the rolling process; Step 2: Based on the multi-modal and multi-dimensional data, a multi-dimensional and multi-modal rolling mill vibration signal time series prediction model is established to address the two aspects of the influence of multi-dimensional vibration-induced process parameters on vibration state and the mixing of long and short period modes of vibration signal. The network fusion structure of this model includes a multi-dimensional long and short period feature extraction layer, a recurrent neural network layer, a temporal convolutional neural network layer, an attention mechanism layer, and a feature fusion layer. Step 3: Construct a dataset for the established model and optimize the model's hyperparameters to ensure the time series data's temporal sequence, thereby obtaining the results of the prediction model established in Step 2.

2. The rolling mill vibration time series prediction method based on multi-level network fusion according to claim 1, characterized in that: Step 1 specifically includes: Step 1.1: During the rolling production, the rolling process parameters and vibration signals at the same starting moment are collected using the production data acquisition system and the vibration data acquisition system; Step 1.2: Preprocess the collected rolling data and vibration data, including data time matching, outlier removal, filling missing values, and data normalization.

3. The rolling mill vibration time series prediction method based on multi-level network fusion according to claim 2, characterized in that: The data timing matching in step 1.2 uses the time of strip rolling and shearing in the production data acquisition system and the time when the vibration acceleration amplitude increases sharply in the vibration signal acquisition system as the basis for timing matching between the rolling process parameters and the vibration acceleration signal, thus completing the data timing matching.

4. The rolling mill vibration time series prediction method based on multi-level network fusion according to claim 2, characterized in that: The outlier removal in step 1.2 uses a 3 σ The criterion for removing outliers from vibration data is as follows: （1）； （2）； In the formula, y i This represents the original vibration acceleration value; This represents the average vibration acceleration. σ y The standard deviation of the vibration acceleration; n Indicates the number of samples.

5. The rolling mill vibration time series prediction method based on multi-level network fusion according to claim 2, characterized in that: The missing value in step 1.2 is filled using a linear interpolation method, and the calculation formula is as follows: （3）； In the formula, x 1 and x 2 represents two known data points before and after the missing value. x Indicates the interpolation position. y This indicates the numerical value corresponding to the interpolation position. y 1 and y 2 represents the numerical values ​​corresponding to the two known data points before and after the missing value.

6. The rolling mill vibration time series prediction method based on multi-level network fusion according to claim 2, characterized in that: The data normalization in step 1.2 uses the maximum and minimum value normalization method, and the calculation criteria are as follows: （4）； In the formula, x It is the raw data. x ' is the normalized data, x max and x min These are the maximum and minimum values ​​of the original data, respectively.

7. The rolling mill vibration time series prediction method based on multi-level network fusion according to claim 2, characterized in that: Step 2 specifically includes: Step 2.1: The first layer of the network structure uses a one-dimensional convolutional neural network to extract long-period and short-period modes and local dependencies between variables in the time dimension. Convolution operations are performed on the short-period mode dataset and the long-period mode dataset respectively, and the extracted features are used as the input of the next layer of the network. A one-dimensional convolutional neural network consists of multiple convolutional layers and pooling layers, as well as a fully connected layer for classification or regression. In the convolutional layers, assuming the input data is a one-dimensional vector x of length n, a convolution kernel w of length k slides along the length of x to perform convolution calculations, resulting in a feature vector c of length n-k+1. i This indicates the result of the convolution kernel starting to act from the i-th position of x: （5）； Then, in the pooling layer, max pooling or average pooling is performed on each feature to reduce the feature size and further extract features; the formula for max pooling is as follows: （6）； In the formula, x is the input data, y is the result after pooling, s is the pooling step size, and p is the pooling window size; Finally, the features are passed to the fully connected layer to obtain the final output; Step 2.2: The second layer of the network structure is trained to predict short-period and long-period datasets respectively; the output of the short-period pattern convolutional layer is input into the long short-term memory neural network for training. Assume the input vector x t and the hidden state h from the previous moment t-1 After linear transformation and activation function processing, the outputs of the input gate, forget gate, and output gate are represented as follows: （7）； （8）； （9）； In the formula, σ It is the sigmoid function, x t h represents the input vector at time t. t Let i represent the hidden state at time t. t f t and o t W represents the outputs of the input gate, forget gate, and output gate, respectively. i W f W o These are the weight matrices for the input gate, forget gate, and output gate, respectively. i b f b o These are the bias vectors for the input gate, forget gate, and output gate, respectively. The formula for updating the memory state is as follows: （10）； （11）； in, This represents the vector dot product operation; tanh is the hyperbolic tangent function, which controls the memory and forgetting of information; the output i of the input gate... t The output f of the forget gate controls the input of new information into the memory state. t Controlling the forgetting of old information from memory; Finally, the update formula for the hidden state is as follows: （12）； Output gate output o t Control the output of information retrieved from the memory state to the hidden state; Step 2.3: The output of the long-period convolutional layer is fed into the temporal convolutional network layer for learning; In this network, residual modules are used instead of convolutional layers in the temporal convolutional network layer. Let the input be x, the network mapping relationship be H(x), and the formula for calculating the residual connections be: f(x) = H(x) - x (15); In the formula, f(x) represents the residual; Step 2.4: The third layer of the network structure is the attention mechanism layer. The attention mechanism allows for more direct dependencies between the states of the model at different time points. The feedforward attention mechanism model is as follows: Hidden state sequence h t The vector input in the middle is fed into the learnable function a(h) t In this process, a series of nonlinear transformations are used to map the hidden state sequence to an output vector e. t : e t= a(h t )(16); The softmax function is used to transform the output vector into a probability distribution, which represents the attention weights of each element in the input vector to produce the probability vector. α t : (17)； Calculate h t Weighted average: (18)； By analyzing each time step T The hidden states are weighted to represent the importance of each time step; Step 2.5: The fourth layer of the network structure is the feature fusion layer, which fuses the long-period patterns of the time series extracted by the TCN part and the short-period patterns extracted by the Long Short-Term Memory Neural Network part to output the result of the prediction model.

8. The rolling mill vibration time series prediction method based on multi-level network fusion according to claim 7, characterized in that: Step 3 specifically includes: Step 3.1: Construct the dataset collected by the model. Time series prediction problems need to ensure the temporal sequence of the data; The input dataset is constructed using a sliding window transformation, which divides the time series dataset into a series of fixed-size sliding windows. Each sliding window is used as a sample to train the model, and then the model is used to predict the next one or more time points. Suppose that a series of time series data observed during the operation of the rolling mill is Y={y1,y2,…,y…} T },in , n The representative feature dimensions include a measured characteristic of mill vibration acceleration and nine variables closely related to mill vibration: tension between the 4th and 5th stands, tension before the 5th stand, rolling speed of the 5th stand, thickness of cold-rolled incoming material, strip thickness at the exit of the 4th stand, strip thickness at the exit of the 5th stand, rolling speed of the 4th stand, rolling force of the 5th stand, and roll gap value of the 5th stand. Let y t,1 Indicates the vibration amplitude of the rolling mill; Define a window size based on the length of the time series dataset. m and prediction step size i Starting from the beginning of the time series dataset, slide the data using the window size, with each slide having a step size of 1; the original dataset will then be constructed as follows: (19)； In the formula, X This represents the historical data input matrix, with an input window length of m. Y This indicates the predicted vibration acceleration output matrix, with a prediction length of i; y m+1,1 express m The vibrational acceleration at time +1; Using the sliding window method described above, vibration process parameters and vibration acceleration time-series data are transformed into supervised learning datasets of historical data and future vibration acceleration signals; each row of matrix X serves as input, corresponding to the same row... Y As output, that is, when {y1,y2,…,y T Given the information, predict y. T+i,1 The amplitude of vibration at time {y1,y2,…,y T , y T+1 Given the information, predict y. T+i+1,1 The amplitude of vibration at any given moment; therefore, the time step is... T The input matrix at is represented as The output matrix of the vibration amplitude prediction result is represented as Y={ y T+1,1 , y T+2,1 ,…,y T+i,1 }; Step 3.2: Optimize the hyperparameters of the model used in Step 2; the CNN feature extraction module, TCN long-term pattern extraction module, and attention mechanism module directly use fixed parameters; Hyperparameter optimization is performed on the LSTM short-period pattern extraction module. Parameters requiring optimization in the LSTM module include the input window time step. T Number of LSTM layers n and the number of hidden layer neurons p ; where the input window time step length T With the length of the sliding window m The same; the grid search method is used to optimize the three variables, and equation (20) determines the range of the three parameters; (20)； The dataset was divided chronologically into a training:validation:test set ratio of 6:2:

2. The Adam optimizer was used with dropout=0.

2. During model training, the validation set was used to avoid overfitting. While training on the training set, the model was continuously validated using the validation set, and the validation error was calculated. If the validation error did not decrease by 0.00001 after 20 iterations, model training was forcibly stopped. K After training, the final test set RMSE is used as the evaluation criterion, and the obtained hyperparameters are input into the multi-level network model for prediction.

Images