Vibration anomaly prediction method and system for continuous casting vibration table based on digital twin technology

Abstract

The present invention provides a method and system for predicting abnormal vibration of a continuous casting vibration table based on digital twin technology, including: constructing a high-fidelity continuous casting digital twin; real-time collection of characteristic parameters that affect the vibration of the continuous casting vibration table; preprocessing parameter data; constructing a VAE digital twin vibration abnormality algorithm model, and sending the processed data and labels into the model for training; comparing the reconstructed data in the model with the actual production performance, using the mean square error MSE and the maximum prediction category score as evaluation criteria, and when the MSE value and the score value are within the preset range, the vibration state of the vibration table is judged to be normal, otherwise it is considered abnormal; and the model output is fed back to the digital twin. The present invention can help inspection personnel to discover abnormal conditions of the vibration table during the casting process in advance, thereby preventing possible equipment failures and improving production efficiency and product quality.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial continuous casting, and in particular to a method and system for predicting abnormal vibration of a continuous casting vibration table based on digital twin technology. Background Art

[0002] Continuous casting is used to continuously cast liquid metal during the casting process. The core is to control the flow of liquid metal through vibration. Once the liquid metal passes through the mold, it begins to gradually solidify and take shape under the action of the cooling water wall. The role of the vibration table also plays a role in this stage, which helps to promote sufficient turbulence of the molten steel in the crystallizer, accelerate the cooling rate, reduce the adhesion between the crystal shell and the crystallizer, accelerate the growth of the crystal shell, help remove inclusions, and improve the quality of the ingot.

[0003] At present, there are still some difficulties and challenges in the prediction of vibration anomalies of the vibration table: the vibration data is complex and affected by many factors, such as liquid level, immersion depth, slag layer thickness, etc. There is a lack of standardized abnormal standards, and the normal vibration range under different working conditions is different; the harsh environment on site affects data collection, and there is a lot of noise interference; the effective abnormal mode is unclear, and there is a lack of corresponding fault sample data for annotation; there is a lack of environmental simulation verification, and it is impossible to arbitrarily simulate various abnormal scenarios in the actual production environment, resulting in huge difficulties in the verification and optimization of the anomaly detection model. Therefore, the integration of digital twin technology is of great significance to the online monitoring and abnormal prediction of the vibration status of the vibration table, which can timely discover hidden dangers of equipment failure, avoid accidents, and ensure production safety.

[0004] Patent document CN109029699A discloses a transformer vibration online abnormality detection method, which is to obtain new data samples about transformer operating state vibration signals within a preset time period, extract characteristic parameters of the new data samples based on wavelet packet analysis, complete the learning of the new data samples based on the fast convex hull algorithm, train and update the single-class abnormality detector model, call the updated single-class abnormality detector model, and perform online abnormality detection on the current transformer vibration. However, this patent cannot completely solve the current technical problems, nor can it meet the needs of the present invention. Summary of the invention

[0005] In view of the defects in the prior art, the purpose of the present invention is to provide a method and system for predicting vibration anomalies of a continuous casting vibration table based on digital twin technology.

[0006] The method for predicting abnormal vibration of a continuous casting vibration table based on digital twin technology provided by the present invention includes:

[0007] Step 1: Build a 3D model of key equipment components including the tundish car, crystallizer, vibrating table, fan segment, bracket, and roller table, import it into the 3D modeling and simulation engine, and build a continuous casting digital twin;

[0008] Step 2: Combined with the analysis of metallurgical principles and the experience of on-site operators, select and collect characteristic parameters that affect the vibration of the vibrating table during the casting process;

[0009] Step 3: Integrate the real-time collected equipment data with the digital twin, and use the real-time data to drive and control the twin to ensure that the dynamic behavior of the digital twin is consistent with the actual production environment;

[0010] Step 4: Preprocess the feature parameters in step 2;

[0011] Step 5: Based on deep learning, construct a VAE digital twin vibration anomaly algorithm model and train the preprocessed data in step 4; compare the reconstructed data in the model with the actual production performance to reflect the difference between the predicted value and the actual value, and use the mean square error MSE and the maximum prediction category score as evaluation criteria to measure the model fitting effect;

[0012] Step 6: Analyze the MSE and score of the prediction results of the VAE digital twin vibration anomaly algorithm model in step 5; when the MSE value and score value are within the preset range, the vibration state of the vibration table is determined to be normal, otherwise it is considered abnormal;

[0013] Step 7: Use the trained VAE digital twin vibration anomaly algorithm model to perform online real-time prediction of vibration table abnormalities, and feed the model output back to the digital twin.

[0014] Preferably, the characteristic parameters in step 2 are filtered out according to the constraints to obtain the vibration data of the casting process, wherein the constraints are that the tundish car is at the casting position, the liquid level is greater than 750 mm, the continuous temperature measurement temperature in the casting area is greater than 1500°C, and the slab section width and slab section thickness are greater than 0 mm.

[0015] Preferably, step 4 includes: performing forward filling of missing values ​​and z-score standardization preprocessing on the characteristic parameter data in step 2; using the k-means clustering method to divide the normal data into multiple clusters, characterizing them as multiple production states from the data level, and the production states are divided into three categories: casting steady state, non-casting steady state, and casting and non-casting switching process.

[0016] Preferably, step 4 comprises:

[0017] Step 4.1: Use the fillna function in pandas to forward fill the missing values ​​of the feature parameters; the current position of the middle car is a constant during casting and is not used for subsequent modeling, so the change rate v of the sample data of each parameter except the current position of the middle car within k seconds is extracted i , the expression is:

[0018]

[0019] Where: Q i is the parameter Q value of the i-th sample, and n is the total number of samples;

[0020] Step 4.2: Standardize the change rate data of each parameter using the z-score method;

[0021] Step 4.3: Use K-means clustering on the normalized data, randomly select 3 points as the initial cluster centers, represent them as 3 production states from the data level, calculate the distance between the sample and each cluster center, and then assign it to the category corresponding to the cluster center closest to it, marking them with label values ​​0, 1, and 2 respectively.

[0022] Preferably, step 5 comprises:

[0023] Step 5.1: Import and parse the preprocessed data set in step 4: Each set of input data contains 10 features. When training and evaluating the data, enter the label into the program; the data used is 11 columns, and the last column is the label; the data file is stored in csv format;

[0024] Step 5.2: Construct a VAE digital twin vibration anomaly algorithm model. The model includes an encoder and a decoder. The encoder maps the data to the normal distribution parameters of the latent space, and the decoder maps the latent space points back to the input space; the number of model layers is set to 9, including the input layer, three feature extraction layers, two encoding layers, a classification layer, and two decoding layers; the number of input layer nodes is 10, corresponding to the first 10 columns of input parameters in the above data set; the feature extraction layer nodes are 33, 33, and 10 respectively; the number of nodes in the two encoding layers is 3, corresponding to the three production states in step 4, used to characterize the distribution mean and standard deviation of the latent space; the number of classification layer nodes is 3, corresponding to the three label values ​​after clustering in step 4.3; the number of nodes in the two decoding layers is 7 and 10 respectively, which restores the data to the input space; the optimizer uses AdagradOptimizer, the activation function uses ReLU, Sigmoid and softmax, and the learning rate is 0.001;

[0025] Step 5.3: Use the shuffle and batch functions of the python numpy library to batch train the VAE digital twin vibration anomaly algorithm model with the data in the feature column; the loss function includes three parts: the MSE mean square error is used to calculate the reconstruction loss of the decoder restored data and the input data; the KL divergence loss is used to measure the difference between the distribution of the latent space and the standard normal distribution; the cross entropy loss calculates the difference between the label column in step 5.1 and the prediction result of the classification layer. The specific formula is:

[0026] Loss = MSE + KL + L

[0027] Where: Loss is the total loss; MSE is the mean square error of the model, KL is the model divergence loss, and L is the cross entropy loss;

[0028]

[0029] Where n is the number of sample predictions; y′ i is the model prediction value, y i is the actual value;

[0030]

[0031] Where: σ and μ are the mean and standard deviation of the latent space distribution of the original data after encoder training;

[0032]

[0033] Where: i is the observed sample, j is the category, M is the number of categories, y ij Indicates whether the category is the same as sample i, if they are the same, it is 1, if they are different, it is 0, p ij Represents the probability that the observed sample i belongs to category j;

[0034] Step 5.4: After analyzing all input parameters, the model outputs the predicted value of the decoder, compares the model reconstructed data with the actual value, reflects the difference between the predicted value and the actual value, and uses the mean square error MSE and the maximum prediction category score as evaluation criteria to measure the model fitting effect.

[0035] The continuous casting vibration table vibration abnormality prediction system based on digital twin technology provided by the present invention includes:

[0036] Module M1: Build 3D models of key equipment components including tundish car, crystallizer, vibrating table, fan segment, bracket, roller table, etc., import 3D modeling and simulation engine, and build continuous casting digital twin;

[0037] Module M2: Combined with the analysis of metallurgical principles and the experience of on-site operators, the characteristic parameters that affect the vibration of the vibrating table during the casting process are selected and collected;

[0038] Module M3: Integrate the real-time collected equipment data with the digital twin, drive and control the twin with real-time data, and ensure that the dynamic behavior of the digital twin is consistent with the actual production environment;

[0039] Module M4: preprocessing the characteristic parameters in module M2;

[0040] Module M5: Based on deep learning, a VAE digital twin vibration anomaly algorithm model is constructed to train the preprocessed data in module M4; the reconstructed data in the model is compared with the actual production performance to reflect the difference between the predicted value and the actual value, and the mean square error (MSE) and the maximum prediction category score (score) are used as evaluation criteria to measure the model fitting effect;

[0041] Module M6: Analyze the MSE and score of the prediction results of the VAE digital twin vibration anomaly algorithm model in module M5; when the MSE value and score value are within the preset range, the vibration state of the vibration table is determined to be normal, otherwise it is considered abnormal;

[0042] Module M7: Use the trained VAE digital twin vibration anomaly algorithm model to perform online real-time prediction of vibration table abnormalities, and feed the model output back to the digital twin.

[0043] Preferably, the characteristic parameters in module M2 are filtered out according to the constraints to obtain the casting process vibration data, wherein the constraints are that the tundish car is at the casting position, the liquid level is greater than 750 mm, the continuous temperature measurement temperature in the casting area is greater than 1500°C, and the slab section width and slab section thickness are greater than 0 mm.

[0044] Preferably, the module M4 includes: performing forward filling of missing values ​​and standardization preprocessing of the characteristic parameter data in the module M2 by the z-score method; dividing the normal data into multiple clusters by the k-means clustering method, characterizing them as multiple production states from the data level, and the production states are divided into three categories: casting steady state, non-casting steady state, and casting and non-casting switching process.

[0045] Preferably, the module M4 comprises:

[0046] Module M4.1: Use the fillna function in pandas to fill missing values ​​of feature parameters. The current position of the middle car is a constant during casting and is not used for subsequent modeling. Therefore, the change rate v of each parameter sample data within k seconds except the current position of the middle car is extracted. i , the expression is:

[0047]

[0048] Where: Q i is the parameter Q value of the i-th sample, and n is the total number of samples;

[0049] Module M4.2: Standardize the change rate data of each parameter using the z-score method;

[0050] Module M4.3: K-means clustering is used on the normalized data. Three points are randomly selected as the initial cluster centers. They are represented as three production states at the data level. The distance between the sample and each cluster center is calculated, and then it is assigned to the category corresponding to the cluster center closest to it, marked with label values ​​of 0, 1, and 2 respectively.

[0051] Preferably, the module M5 comprises:

[0052] Module M5.1: Import and parse the preprocessed data set in module M4: Each set of input data contains 10 features. When training and evaluating the data, the labels are input into the program. The data used is 11 columns, and the last column is the label. The data file is stored in csv format.

[0053] Module M5.2: Construct a VAE digital twin vibration anomaly algorithm model. The model includes an encoder and a decoder. The encoder maps the data to the normal distribution parameters of the latent space, and the decoder maps the latent space points back to the input space. The number of model layers is set to 9, including an input layer, three feature extraction layers, two encoding layers, a classification layer, and two decoding layers. The number of input layer nodes is 10, corresponding to the first 10 columns of input parameters in the above data set. The feature extraction layer nodes are 33, 33, and 10 respectively. The number of nodes in the two encoding layers is 3, corresponding to the three production states in module M4, which are used to characterize the distribution mean and standard deviation of the latent space. The number of classification layer nodes is 3, corresponding to the three label values ​​after clustering in module M4.3. The number of nodes in the two decoding layers is 7 and 10 respectively, which restores the data to the input space. The optimizer uses AdagradOptimizer, and the activation functions use ReLU, Sigmoid, and softmax, with a learning rate of 0.001.

[0054] Module M5.3: The shuffle and batch functions of the python numpy library are used to batch train the VAE digital twin vibration anomaly algorithm model with the data in the feature column; the loss function includes three parts: the MSE mean square error is used to calculate the reconstruction loss of the decoder restored data and the input data; the KL divergence loss is used to measure the difference between the distribution of the latent space and the standard normal distribution; the cross entropy loss is used to calculate the difference between the label column in module M5.1 and the prediction result of the classification layer. The specific formula is:

[0055] Loss = MSE + KL + L

[0056] Where: Loss is the total loss; MSE is the mean square error of the model, KL is the model divergence loss, and L is the cross entropy loss;

[0057]

[0058] Where n is the number of sample predictions; y′ i is the model prediction value, y i is the actual value;

[0059]

[0060] Where: σ and μ are the mean and standard deviation of the latent space distribution of the original data after encoder training;

[0061]

[0062] Where: i is the observed sample, j is the category, M is the number of categories, y ij Indicates whether the category is the same as sample i, if they are the same, it is 1, if they are different, it is 0, p ij Represents the probability that the observed sample i belongs to category j;

[0063] Module M5.4: After analyzing all input parameters, the model outputs the decoder's prediction value, compares the model's reconstructed data with the actual value, reflects the difference between the predicted value and the actual value, and uses the mean square error (MSE) and the maximum prediction category score (score) as evaluation criteria to measure the model fitting effect.

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] The present invention combines digital twin technology with a variational autoencoder (VAE) model for prediction of continuous casting vibration anomalies, overcoming the problem of difficult labeling. The VAE model can autonomously learn the data distribution under normal conditions. Any abnormal state will deviate from the distribution under normal conditions, and can fully mine the intrinsic statistical characteristics and patterns of the data without too many prior assumptions. The digital twin virtually reproduces the entire continuous casting vibration process, and the output of the model is fed back to the digital twin environment, which is convenient for real-time monitoring and virtual verification, and is more suitable for practical engineering applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0067] Figure 1 This is a flow chart of a method for predicting abnormal vibration of a continuous casting vibration table based on digital twin technology in the present invention;

[0068] Figure 2 It is a characteristic parameter curve diagram of the vibration of the vibration table in the present invention;

[0069] Figure 3 is a graph of other parameters affecting the casting process in the present invention;

[0070] Figure 4It is the k-means cluster diagram in the present invention;

[0071] Figure 5 This is a structural diagram of the VAE digital twin vibration anomaly algorithm constructed in the present invention;

[0072] Figure 6a and Figure 6b They are respectively the loss curves of the model training and verification data in the present invention. DETAILED DESCRIPTION

[0073] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0074] Example 1

[0075] like Figure 1 As shown, the present invention provides a method for predicting abnormal vibration of a continuous casting vibration table based on digital twin technology, comprising the following steps:

[0076] Step 1: Build a high-fidelity continuous casting digital twin;

[0077] Accurately model the three-dimensional models of key equipment components such as the car, crystallizer, vibration table, fan segment, bracket, roller table, etc., import them into the 3D modeling and simulation engine, and build a digital twin of the continuous casting.

[0078] Step 2: Feature parameter selection and data collection;

[0079] Combined with the analysis of metallurgical principles and the experience of on-site operators, characteristic parameter data that affect the vibration of the vibration table during the casting process are selected and collected; the parameters include continuous temperature measurement in the casting area of ​​the tundish car, liquid level in the crystallizer, pressure of vibration table cylinder 1, pressure of vibration table cylinder 2, actual amplitude value, set amplitude value, vibration frequency, casting speed, slab cross-section width, slab cross-section thickness, and current position of the tundish car, totaling 11 dimensions.

[0080] In this example, the actual performance data of continuous casting in a steel plant is taken as an example, and the sampling frequency is 1s / time, and the vibration-related characteristic parameter data of the casting process for one week is collected. Figure 2 , Figure 3 As shown, Figure 2 is the characteristic parameter curve of the vibration table. Figure 3 It is a curve diagram of other parameters that affect the casting process, where CSR1_SegAll_uOp_SlabThick is the slab section thickness in mm.

[0081] Step 3: Device data drives the digital twin;

[0082] The characteristic parameters in step 2 are filtered out according to the constraints to obtain the casting process vibration data, wherein the constraints are that the trolley is at the casting position, the liquid level is greater than 750 mm, the continuous temperature measurement in the casting area is greater than 1500 °C, and the slab section width and slab section thickness are greater than 0 mm.

[0083] Integrate the real-time collected equipment data with the digital twin, and use the real-time data to drive and control the twin to ensure that the dynamic behavior of the digital twin is consistent with the actual production environment.

[0084] Step 4: Preprocess parameter data;

[0085] Use the fillna function in pandas to forward fill missing values ​​for feature parameters; the current position of the middle car during casting is a constant and is not used for subsequent modeling, so the change rate of each parameter sample data within k seconds except the current position of the middle car is extracted, that is:

[0086]

[0087] Where: Q i is the parameter Q value of the i-th sample, and n is the total number of samples.

[0088] After the z-score method was used to standardize the data of the change rate of each parameter, K-means clustering was used to divide the data into multiple clusters so that they could be represented as multiple production states at the data level. The production states were divided into three categories: casting steady state, non-casting steady state, and switching process between casting and non-casting.

[0089] Specifically, we randomly select three points as the initial cluster centers, calculate the distance between the sample and each cluster center, and then assign it to the category corresponding to the cluster center closest to it, marking the label values ​​0, 1, and 2 respectively. In order to observe the results after clustering, Figure 4 As shown in the figure, the 10 parameter data related to casting vibration are reduced in dimension by T-SNE, and a cluster diagram is obtained after k-means clustering. The data points form a tight cluster structure in the diagram, and there are obvious boundaries or intervals between data points of different categories, indicating that the clustering effect is good.

[0090] Step 5: Build a VAE digital twin vibration anomaly algorithm model;

[0091] Based on deep learning, a VAE variational autoencoder model is constructed and trained on the data preprocessed in step 4. The model reconstructed data is compared with the actual production performance to reflect the difference between the predicted value and the actual value. The mean square error (MSE) and the maximum prediction category score are used as the main evaluation criteria to measure the model fitting effect.

[0092] As a specific implementation method, the preprocessed data set in step 34 is imported and parsed, each data set contains 10 features and 1 column label; therefore, the data used is 11 columns, and the data file is stored in csv format; the VAE model is constructed, and the specific parameters of the model are set as follows:

[0093] The number of model layers is set to 9, including input layer, three feature extraction layers, two encoding layers, classification layer, and two decoding layers; the number of input layer nodes is 10, corresponding to the first 10 columns of input parameters in the above data set; the number of feature extraction layer nodes is 33, 33, and 10 respectively; the number of nodes in the two encoding layers is 3, corresponding to the three working states in step 3, used to characterize the distribution mean and standard deviation of the latent space; the number of classification layer nodes is 3, corresponding to the three label values ​​after clustering in step 3; the number of nodes in the two decoding layers is 7 and 10 respectively, restoring the data to the input space; the optimizer uses AdagradOptimizer, the activation function uses ReLU, Sigmoid and softmax, and the learning rate is 0.001;

[0094] In an embodiment, the VAE model network structure diagram constructed visually is as follows: Figure 5 As shown, fc11, fc12, and fc13 represent three feature extraction layers, f21 and f22 represent two encoding layers that output the mean and standard deviation of the potential distribution space, f23 represents the classification layer, which is used to calculate the classification cross entropy loss, and f31 and f32 represent two decoding layers that restore the encoded data to the input space.

[0095] Training model: Use the shuffle and batch functions of the python numpy library to train the VAE model in batches with the data in the feature column; the loss function includes three parts: MSE mean square error is used to calculate the reconstruction loss of the decoder restored data and the input data; KL divergence loss is used to measure the difference between the distribution of the latent space and the standard normal distribution; cross entropy loss, calculates the difference between the label column and the classification layer prediction result. Specific formula:

[0096] Loss = MSE + KL + L

[0097] Where: Loss is the total loss; MSE is the mean square error of the model, KL is the model divergence loss, and L is the cross entropy loss;

[0098]

[0099] Where n is the number of sample predictions; y′ i is the model prediction value, y i is the actual value;

[0100]

[0101] Where: σ and μ are the mean and standard deviation of the latent space distribution of the original data after encoder training;

[0102]

[0103] Where: i is the observed sample, j is the category, M is the number of categories, y ij Indicates whether the category is the same as sample i, if they are the same, it is 1, if they are different, it is 0, p ij Represents the probability that the observed sample i belongs to category j;

[0104] In this embodiment, after the screening of the constraint conditions in step 3, a total of 138,663 samples of characteristic parameters related to the casting process vibration were collected, and the data set was divided into training data and verification data in a ratio of 80% and 20% to train and optimize the VAE model. Observe the loss curves of the model training and verification data, as shown in Figure 6a and Figure 6b As shown in the figure, both training loss and validation loss have converged and the difference between them is very small. After 150 iterations, both loss curves begin to converge, and there is no noticeable difference between them, indicating that the generalization ability of the training model is strong, and there is no overfitting or underfitting phenomenon.

[0105] Model prediction: After analyzing all input parameters, the model will output the predicted value of the decoder, compare the predicted value with the actual value, and reflect the difference between the predicted value and the actual value. The mean square error (MSE) and the maximum prediction category score are used as the main evaluation criteria to measure the model fitting effect.

[0106] In this embodiment, the MES errors of the model after the final iteration in the training data set and the validation data set are shown in Table 1. The error results obtained in the training data set and the validation set are similar, indicating that the model performs well in both training and validation processes. The closer the maximum prediction category score is to 1, the greater the probability that the data is normal, and the closer it is to 0, the greater the probability that it is abnormal.

[0107] MSE Training set 0.900080291101117 Validation set 0.9082864078258003

[0108] Step 6: Analyze the MSE and maximum prediction category score of the VAE digital twin vibration anomaly model prediction results in step 5; when the MSE value and score value are within the preset range, the vibration state of the vibration table is judged to be normal, otherwise it is considered abnormal.

[0109] Step 7: Use the trained VAE model to perform online real-time prediction of abnormal conditions of the vibration table and feed the model output back to the digital twin.

[0110] Example 2

[0111] The present invention also provides a continuous casting vibration table vibration anomaly prediction system based on digital twin technology. The continuous casting vibration table vibration anomaly prediction system based on digital twin technology can be realized by executing the process steps of the continuous casting vibration table vibration anomaly prediction method based on digital twin technology, that is, technical personnel in this field can understand the continuous casting vibration table vibration anomaly prediction method based on digital twin technology as the preferred implementation mode of the continuous casting vibration table vibration anomaly prediction system based on digital twin technology.

[0112] The continuous casting vibration table vibration abnormality prediction system based on digital twin technology provided by the present invention includes:

[0113] Module M1: Build 3D models of key equipment components including tundish car, crystallizer, vibrating table, fan segment, bracket, roller table, etc., import 3D modeling and simulation engine, and build continuous casting digital twin;

[0114] Module M2: Combined with the analysis of metallurgical principles and the experience of on-site operators, the characteristic parameters that affect the vibration of the vibrating table during the casting process are selected and collected;

[0115] Module M3: Integrate the real-time collected equipment data with the digital twin, drive and control the twin with real-time data, and ensure that the dynamic behavior of the digital twin is consistent with the actual production environment;

[0116] Module M4: preprocessing the characteristic parameters in module M2;

[0117] Module M5: Based on deep learning, a VAE digital twin vibration anomaly algorithm model is constructed to train the preprocessed data in module M4; the reconstructed data in the model is compared with the actual production performance to reflect the difference between the predicted value and the actual value, and the mean square error (MSE) and the maximum prediction category score (score) are used as evaluation criteria to measure the model fitting effect;

[0118] Module M6: Analyze the MSE and score of the prediction results of the VAE digital twin vibration anomaly algorithm model in module M5; when the MSE value and score value are within the preset range, the vibration state of the vibration table is determined to be normal, otherwise it is considered abnormal;

[0119] Module M7: Use the trained VAE digital twin vibration anomaly algorithm model to perform online real-time prediction of vibration table abnormalities, and feed the model output back to the digital twin.

[0120] The characteristic parameters in module M2 are screened out according to the constraints to obtain the casting process vibration data. The constraints are that the trolley is at the casting position, the liquid level is greater than 750mm, the continuous temperature measurement in the casting area is greater than 1500℃, and the slab section width and slab section thickness are greater than 0mm.

[0121] The module M4 includes: performing forward filling of missing values ​​and z-score standardization preprocessing on the characteristic parameter data in the module M2; using the k-means clustering method to divide the normal data into multiple clusters, which are characterized as multiple production states from the data level. The production states are divided into three categories: casting steady state, non-casting steady state, and casting and non-casting switching process.

[0122] The module M4 comprises:

[0123] Module M4.1: Use the fillna function in pandas to fill missing values ​​of feature parameters. The current position of the middle car is a constant during casting and is not used for subsequent modeling. Therefore, the change rate v of each parameter sample data within k seconds except the current position of the middle car is extracted. i , the expression is:

[0124]

[0125] Where: Q i is the parameter Q value of the i-th sample, and n is the total number of samples;

[0126] Module M4.2: Standardize the change rate data of each parameter using the z-score method;

[0127] Module M4.3: K-means clustering is used on the normalized data. Three points are randomly selected as the initial cluster centers. They are represented as three production states at the data level. The distance between the sample and each cluster center is calculated, and then it is assigned to the category corresponding to the cluster center closest to it, marked with label values ​​of 0, 1, and 2 respectively.

[0128] The module M5 comprises:

[0129] Module M5.1: Import and parse the preprocessed data set in module M4: Each set of input data contains 10 features. When training and evaluating the data, the labels are input into the program. The data used is 11 columns, and the last column is the label. The data file is stored in csv format.

[0130] Module M5.2: Construct a VAE digital twin vibration anomaly algorithm model. The model includes an encoder and a decoder. The encoder maps the data to the normal distribution parameters of the latent space, and the decoder maps the latent space points back to the input space. The number of model layers is set to 9, including an input layer, three feature extraction layers, two encoding layers, a classification layer, and two decoding layers. The number of input layer nodes is 10, corresponding to the first 10 columns of input parameters in the above data set. The feature extraction layer nodes are 33, 33, and 10 respectively. The number of nodes in the two encoding layers is 3, corresponding to the three production states in module M4, which are used to characterize the distribution mean and standard deviation of the latent space. The number of classification layer nodes is 3, corresponding to the three label values ​​after clustering in module M4.3. The number of nodes in the two decoding layers is 7 and 10 respectively, which restores the data to the input space. The optimizer uses AdagradOptimizer, and the activation functions use ReLU, Sigmoid, and softmax, with a learning rate of 0.001.

[0131] Module M5.3: The shuffle and batch functions of the python numpy library are used to batch train the VAE digital twin vibration anomaly algorithm model with the data in the feature column; the loss function includes three parts: the MSE mean square error is used to calculate the reconstruction loss of the decoder restored data and the input data; the KL divergence loss is used to measure the difference between the distribution of the latent space and the standard normal distribution; the cross entropy loss is used to calculate the difference between the label column in module M5.1 and the prediction result of the classification layer. The specific formula is:

[0132] Loss = MSE + KL + L

[0133] Where: Loss is the total loss; MSE is the mean square error of the model, KL is the model divergence loss, and L is the cross entropy loss;

[0134]

[0135] Where n is the number of sample predictions; y′ i is the model prediction value, y i is the actual value;

[0136]

[0137] Where: σ and μ are the mean and standard deviation of the latent space distribution of the original data after encoder training;

[0138]

[0139] Where: i is the observed sample, j is the category, M is the number of categories, y ij Indicates whether the category is the same as sample i, if they are the same, it is 1, if they are different, it is 0, p ijRepresents the probability that the observed sample i belongs to category j;

[0140] Module M5.4: After analyzing all input parameters, the model outputs the decoder's prediction value, compares the model's reconstructed data with the actual value, reflects the difference between the predicted value and the actual value, and uses the mean square error (MSE) and the maximum prediction category score (score) as evaluation criteria to measure the model fitting effect.

[0141] Those skilled in the art know that, in addition to implementing the system, device and its various modules provided by the present invention in a purely computer-readable program code, it is entirely possible to implement the same program in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers and embedded microcontrollers by logically programming the method steps. Therefore, the system, device and its various modules provided by the present invention can be considered as a hardware component, and the modules included therein for implementing various programs can also be considered as structures within the hardware component; the modules for implementing various functions can also be considered as both software programs for implementing the method and structures within the hardware component.

[0142] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A method for predicting abnormal vibration of a continuous casting vibration table based on digital twin technology, characterized in that: include: Step 1: Build a 3D model of key equipment components including the tundish car, crystallizer, vibrating table, fan segment, bracket, and roller table, import it into the 3D modeling and simulation engine, and build a continuous casting digital twin; Step 2: Combined with the analysis of metallurgical principles and the experience of on-site operators, select and collect characteristic parameters that affect the vibration of the vibrating table during the casting process; Step 3: Integrate the real-time collected equipment data with the digital twin, and use the real-time data to drive and control the twin to ensure that the dynamic behavior of the digital twin is consistent with the actual production environment; Step 4: Preprocess the feature parameters in step 2; Step 5: Based on deep learning, construct a VAE digital twin vibration anomaly algorithm model and train the preprocessed data in step 4; compare the reconstructed data in the model with the actual production performance to reflect the difference between the predicted value and the actual value, and use the mean square error MSE and the maximum prediction category score as evaluation criteria to measure the model fitting effect; Step 6: Analyze the MSE and score of the prediction results of the VAE digital twin vibration anomaly algorithm model in step 5; when the MSE value and score value are within the preset range, the vibration state of the vibration table is determined to be normal, otherwise it is considered abnormal; Step 7: Use the trained VAE digital twin vibration anomaly algorithm model to perform online real-time prediction of vibration table abnormalities, and feed the model output back to the digital twin; The step 4 includes: performing forward filling of missing values ​​and z-score standardization preprocessing on the characteristic parameter data in step 2; using the k-means clustering method to divide the normal data into multiple clusters, which are characterized as multiple production states from the data level, and the production states are divided into three categories: casting steady state, non-casting steady state, and casting and non-casting switching process; The step 4 comprises: Step 4.1: Use the fillna function in pandas to forward fill the missing values ​​of the feature parameters; the current position of the middle car is a constant during casting and is not used for subsequent modeling, so the change rate v of the sample data of each parameter except the current position of the middle car within k seconds is extracted i , the expression is: Where: Q i is the parameter Q value of the i-th sample, and n is the total number of samples; Step 4.2: Standardize the change rate data of each parameter using the z-score method; Step 4.3: K-means clustering is used on the normalized data. Three points are randomly selected as the initial cluster centers. From the data level, they are represented as three production states. The distance between the sample and each cluster center is calculated, and then it is assigned to the category corresponding to the cluster center closest to it, and the label values ​​are marked as 0, 1, and 2 respectively. The step 5 comprises: Step 5.1: Import and parse the preprocessed data set in step 4: Each set of input data contains 10 features. When training and evaluating the data, enter the label into the program; the data used is 11 columns, and the last column is the label; the data file is stored in csv format; Step 5.2: Construct a VAE digital twin vibration anomaly algorithm model. The model includes an encoder and a decoder. The encoder maps the data to the normal distribution parameters of the latent space, and the decoder maps the latent space points back to the input space; the number of model layers is set to 9, including the input layer, three feature extraction layers, two encoding layers, a classification layer, and two decoding layers; the number of input layer nodes is 10, corresponding to the first 10 columns of input parameters in the above data set; the feature extraction layer nodes are 33, 33, and 10 respectively; the number of nodes in the two encoding layers is 3, corresponding to the three production states in step 4, used to characterize the distribution mean and standard deviation of the latent space; the number of classification layer nodes is 3, corresponding to the three label values ​​after clustering in step 4.3; the number of nodes in the two decoding layers is 7 and 10 respectively, which restores the data to the input space; the optimizer uses AdagradOptimizer, the activation function uses ReLU, Sigmoid and softmax, and the learning rate is 0.001; Step 5.3: Use the shuffle and batch functions of the python numpy library to batch train the VAE digital twin vibration anomaly algorithm model with the data in the feature column; the loss function includes three parts: the MSE mean square error is used to calculate the reconstruction loss of the decoder restored data and the input data; the KL divergence loss is used to measure the difference between the distribution of the latent space and the standard normal distribution; the cross entropy loss calculates the difference between the label column in step 5.1 and the prediction result of the classification layer. The specific formula is: Loss = MSE + KL + L Where: Loss is the total loss; MSE is the mean square error of the model, KL is the model divergence loss, and L is the cross entropy loss; Where n is the number of sample predictions; y′ i is the model prediction value, y i is the actual value; Where: σ and μ are the mean and standard deviation of the latent space distribution of the original data after encoder training; Where: i is the observed sample, j is the category, M is the number of categories, y ij Indicates whether the category is the same as sample i, if they are the same, it is 1, if they are different, it is 0, p ij Represents the probability that the observed sample i belongs to category j; Step 5.4: After analyzing all input parameters, the model outputs the predicted value of the decoder, compares the model reconstructed data with the actual value, reflects the difference between the predicted value and the actual value, and uses the mean square error MSE and the maximum prediction category score as evaluation criteria to measure the model fitting effect.

2. The method for predicting abnormal vibration of a continuous casting vibration table based on digital twin technology according to claim 1 is characterized in that: The characteristic parameters in step 2 are filtered out according to the constraints to obtain the casting process vibration data, wherein the constraints are that the trolley is at the casting position, the liquid level is greater than 750 mm, the continuous temperature measurement in the casting area is greater than 1500 °C, and the slab section width and slab section thickness are greater than 0 mm.

3. A continuous casting vibration table vibration abnormality prediction system based on digital twin technology, characterized in that: include: Module M1: Build 3D models of key equipment components including tundish car, crystallizer, vibrating table, fan segment, bracket, roller table, etc., import 3D modeling and simulation engine, and build continuous casting digital twin; Module M2: Combined with the analysis of metallurgical principles and the experience of on-site operators, the characteristic parameters that affect the vibration of the vibrating table during the casting process are selected and collected; Module M3: Integrate the real-time collected equipment data with the digital twin, drive and control the twin with real-time data, and ensure that the dynamic behavior of the digital twin is consistent with the actual production environment; Module M4: preprocessing the characteristic parameters in module M2; Module M5: Based on deep learning, a VAE digital twin vibration anomaly algorithm model is constructed to train the preprocessed data in module M4; the reconstructed data in the model is compared with the actual production performance to reflect the difference between the predicted value and the actual value, and the mean square error (MSE) and the maximum prediction category score (score) are used as evaluation criteria to measure the model fitting effect; Module M6: Analyze the MSE and score of the prediction results of the VAE digital twin vibration anomaly algorithm model in module M5; when the MSE value and score value are within the preset range, the vibration state of the vibration table is determined to be normal, otherwise it is considered abnormal; Module M7: Use the trained VAE digital twin vibration anomaly algorithm model to perform online real-time forecasting of vibration table anomalies, and feed the model output back to the digital twin; The module M4 includes: performing forward filling of missing values ​​and standardization preprocessing of the characteristic parameter data in the module M2 by using the z-score method; using the k-means clustering method to divide the normal data into multiple clusters, which are characterized as multiple production states from the data level, and the production states are divided into three categories: casting steady state, non-casting steady state, and casting and non-casting switching process; The module M4 comprises: Module M4.1: Use the fillna function in pandas to fill missing values ​​of feature parameters. The current position of the middle car is a constant during casting and is not used for subsequent modeling. Therefore, the change rate v of each parameter sample data within k seconds except the current position of the middle car is extracted. i , the expression is: Where: Q i is the parameter Q value of the i-th sample, and n is the total number of samples; Module M4.2: Standardize the change rate data of each parameter using the z-score method; Module M4.3: K-means clustering is used for the normalized data. Three points are randomly selected as the initial cluster centers. The data is represented as three production states. The distance between the sample and each cluster center is calculated. Then, the sample is assigned to the category corresponding to the cluster center closest to it, and the label values ​​are marked as 0, 1, and 2 respectively. The module M5 comprises: Module M5.1: Import and parse the preprocessed data set in module M4: Each set of input data contains 10 features. When training and evaluating the data, the labels are input into the program. The data used is 11 columns, and the last column is the label. The data file is stored in csv format. Module M5.2: Construct a VAE digital twin vibration anomaly algorithm model. The model includes an encoder and a decoder. The encoder maps the data to the normal distribution parameters of the latent space, and the decoder maps the latent space points back to the input space. The number of model layers is set to 9, including an input layer, three feature extraction layers, two encoding layers, a classification layer, and two decoding layers. The number of input layer nodes is 10, corresponding to the first 10 columns of input parameters in the above data set. The feature extraction layer nodes are 33, 33, and 10 respectively. The number of nodes in the two encoding layers is 3, corresponding to the three production states in module M4, which are used to characterize the distribution mean and standard deviation of the latent space. The number of classification layer nodes is 3, corresponding to the three label values ​​after clustering in module M4.

3. The number of nodes in the two decoding layers is 7 and 10 respectively, which restores the data to the input space. The optimizer uses AdagradOptimizer, and the activation functions use ReLU, Sigmoid, and softmax, with a learning rate of 0.

001. Module M5.3: The shuffle and batch functions of the python numpy library are used to batch train the VAE digital twin vibration anomaly algorithm model with the data in the feature column; the loss function includes three parts: the MSE mean square error is used to calculate the reconstruction loss of the decoder restored data and the input data; the KL divergence loss is used to measure the difference between the distribution of the latent space and the standard normal distribution; the cross entropy loss is used to calculate the difference between the label column in module M5.1 and the prediction result of the classification layer. The specific formula is: Loss = MSE + KL + L Where: Loss is the total loss; MSE is the mean square error of the model, KL is the model divergence loss, and L is the cross entropy loss; Where n is the number of sample predictions; y′ i is the model prediction value, y i is the actual value; Where: σ and μ are the mean and standard deviation of the latent space distribution of the original data after encoder training; Where: i is the observed sample, j is the category, M is the number of categories, y ij Indicates whether the category is the same as sample i, if they are the same, it is 1, if they are different, it is 0, p ij Represents the probability that the observed sample i belongs to category j; Module M5.4: After analyzing all input parameters, the model outputs the decoder's prediction value, compares the model's reconstructed data with the actual value, reflects the difference between the predicted value and the actual value, and uses the mean square error (MSE) and the maximum prediction category score (score) as evaluation criteria to measure the model fitting effect.

4. The continuous casting vibration table vibration anomaly prediction system based on digital twin technology according to claim 3 is characterized in that: The characteristic parameters in module M2 are screened out according to the constraints to obtain the casting process vibration data. The constraints are that the trolley is at the casting position, the liquid level is greater than 750mm, the continuous temperature measurement in the casting area is greater than 1500℃, and the slab section width and slab section thickness are greater than 0mm.

Images