Continuous casting mold liquid level abnormality prediction method and system based on digital twin technology

Abstract

The present invention provides a method and system for predicting abnormal liquid level of continuous casting mold based on digital twin technology, including: constructing a high-fidelity continuous casting digital twin physical environment; collecting and fusing multi-source heterogeneous data that affect the liquid level of the crystallizer during the casting process; preprocessing the data; generating noise data by random noise; constructing a digital twin data-driven model of a GAN network; comparing the predicted value of the GAN generator with the actual value, using the WMSE weighted mean square error loss to evaluate the quality of the generator, and using the cross entropy loss to evaluate the quality of the discriminator; after the model is evaluated as qualified, the abnormal situation of the liquid level of the crystallizer is discriminated, and a confidence threshold is set. When the predicted score is greater than the threshold, it is judged as abnormal; and the model output is fed back to the digital twin environment. The present invention can predict the changing trend of the liquid level of the crystallizer online and predict whether the liquid level of the crystallizer is abnormal, which is beneficial to the application of practical engineering.

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 liquid level anomaly in a continuous casting crystallizer based on digital twin technology. Background Art

[0002] During the continuous casting process, abnormal fluctuations in the mold liquid level will have an adverse effect on the quality of the ingot. Reasonable control of the mold liquid level can not only optimize the formation of the primary shell and the metal solidification process, fundamentally ensure the surface quality and internal structure uniformity of the ingot, but also avoid serious accidents such as vibration distortion and jamming to the greatest extent.

[0003] At present, there are still some difficulties and challenges in predicting abnormal crystallizer liquid level: liquid level fluctuations are affected by many complex factors, involving multidisciplinary knowledge such as thermodynamics and fluid mechanics, and it is extremely challenging to build an accurate physical model; data-based black box modeling methods such as machine learning models have high requirements for large amounts of high-quality data, and some key process parameters are difficult to obtain. At the same time, there is a lack of accumulation of abnormal samples, which increases the difficulty of model training and prediction.

[0004] Patent document CN117386694A discloses a method for monitoring wear evolution and leakage of a hydraulic cylinder and its digital twin system. The method includes the following steps: 1. Constructing a leakage model that characterizes the mapping relationship between the leakage amount and contact pressure of the hydraulic cylinder. 2. Performing CFD simulation analysis on the three-dimensional model of the hydraulic cylinder to obtain a pressure distribution model under different wear depths. 3. Performing wavelet packet analysis on the pressure signal to determine the wear eigenvalue; and retaining the corresponding eigenvalue extraction model; 4. Collecting real data and creating a data-driven model from eigenvalue to wear depth. 5. Fusion of the basic model to obtain a wear evolution monitoring model and a leakage monitoring model; 6. Using the wear evolution monitoring model and the leakage monitoring model to predict the state of the collected pressure and displacement signals. 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 the abnormal liquid level of a continuous casting crystallizer based on digital twin technology.

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

[0007] Step 1: Build a high-fidelity continuous casting digital twin physical environment;

[0008] Construct 3D models of key equipment components including the tundish car, crystallizer, fan-shaped segment, bracket, and roller system, import 3D modeling and simulation engines, and assemble them into a complete continuous casting production line; use the particle system to simulate the pouring and flow of molten metal liquid to show the entire casting process;

[0009] Step 2: Collect and integrate multi-source heterogeneous data that affect the mold level during the casting process;

[0010] Connect with the monitoring system and data acquisition equipment on the industrial site to obtain real-time production data and control the movement trajectory and speed of the digital twin including the tundish, crystallizer, and roller system; combine the analysis of metallurgical principles and the experience of on-site operators to select the characteristic parameters that affect the crystallizer liquid level fluctuation;

[0011] Step 3: Preprocess the parameter data and divide the multivariate time series segments into segments with a fixed window length W as positive samples of the data-driven model, including input data and output data;

[0012] Step 4: Generate noise data from random noise as negative samples of the data-driven model. The dimension of the noise data is the same as the dimension of the input data after segmentation in step 3. Copy the output data after segmentation in step 3 as the output of the noise.

[0013] Step 5: Construct a sample set, take the input data in step 3 and the noise data in step 4 as input samples, record them as x1_true and x2_false respectively; take the output data in step 3 and the output of the noise in step 4 as output samples, record them as y1_true and y2_false respectively; at the same time, classify the real input data and the noise data, and mark them as 0 and 1 respectively;

[0014] Step 6: Build a digital twin data-driven model of the GAN network;

[0015] The GAN generator uses the transformer encoder-decoder architecture and introduces a cross-attention mechanism to capture dependencies across time and dimensions. The discriminator uses a multi-layer perceptron structure. The output of the encoder in the GAN generator is used as the input of the discriminator. The loss function of the discriminator uses a binary cross-entropy loss, and the output is a 0 or 1 label.

[0016] Step 7: Alternately train the GAN generator and discriminator;

[0017] The shuffle and batch functions of the python numpy library are used to randomly shuffle the sample data before training the GAN. When training the generator, the discriminator parameters are fixed and updated to minimize its loss function. When training the discriminator, the generator parameters are fixed and updated to minimize its loss function.

[0018] Step 8: After the model is evaluated as qualified, the abnormal situation of the crystallizer liquid level is judged, and the confidence threshold t is set. When the discriminator prediction score is greater than the threshold t, it is judged as abnormal; the prediction data output by the generator contains the change trend of all parameters in the future period of time, which is used to observe the change of the crystallizer liquid level in advance;

[0019] Step 9: Use the trained GAN data-driven model to perform online real-time prediction of abnormal crystallizer liquid level, and feed the model output back to the digital twin environment.

[0020] Preferably, characteristic parameters that affect the fluctuation of the crystallizer liquid level are selected, including 1 each of the net weight of molten steel in the tundish, the position of the stopper rod, the continuous temperature measurement in the casting area, the crystallizer liquid level, and the pulling speed, 4 crystallizer water outlet flow rates, 2 cylinder friction works, 11 drive roller torques, and 25 drawing roller currents, totaling 47 dimensions.

[0021] Preferably, the step 3 comprises:

[0022] Step 3.1: Filter the parameter data in step 2 according to the constraint conditions to select the casting process data, where the constraint conditions are that the tundish is at the casting position, the liquid level of the crystallizer is greater than 750 mm, and the continuous temperature measurement temperature of the casting area is greater than 1500 °C;

[0023] Step 3.2: Use the fillna function in Python to forward fill in the empty values ​​and obtain multivariate time series data;

[0024] Step 3.3: Split the multivariate time series data into fixed window length W to obtain a series of data segments of length W; each segment contains the values ​​of multiple variables, forming a two-dimensional tensor, which is used as a sample data; among them, the first W1 are used as sample input data, and (W-W1) are used as sample output data;

[0025] Step 3.4: Prepare multiple batches of data, each batch containing multiple samples, for subsequent model training.

[0026] Preferably, step 6 comprises:

[0027] The GAN generator uses the transformer encoder-decoder architecture; the encoder maps the input data x1_true in step 3 and the noise data x2_false in step 4 to the latent space hidden vector z; the hidden vector z is used as the input of the decoder, and the decoder generates the output data y1_true in step 3 and the noise output y2_false according to the hidden vector z; at the same time, the hidden vector z output by the encoder is used as the input of the discriminator, and the discriminator classifies the hidden vector z and outputs 0 and 1 labels, 0 for normal and 1 for abnormal;

[0028] Both the encoder and decoder are composed of an embedding layer, a position encoding layer, and an attention layer; the embedding layer is a linear layer that maps the input feature vector to a higher-dimensional representation space; the position encoding layer is a learnable parameter matrix used to learn the relative or absolute position information of the high-dimensional sequence, and its shape is the same as the output of the embedding layer; the attention layer uses a scaled dot product attention mechanism, and on the basis of the dot product attention, the attention weight is scaled to improve the stability of the model. The specific formula is as follows:

[0029]

[0030] In the formula, Q·K T represents the dot product of the query vector Q and the key vector K, d k is the dimension scaling factor; V is the value vector;

[0031] The encoder of the GAN generator consists of 1 embedding layer, 1 position encoding layer, and 2 self-attention layers; the output of the embedding layer is added to the position encoding matrix, the added matrix vector is normalized, and sent to the first self-attention layer for information fusion; a learnable parameter matrix is ​​introduced as the query vector Q, the output of the first self-attention layer is used as the key vector K and the value vector V, and the three matrices are sent to the second self-attention layer. The parameter matrix is ​​used to control the intensity of information flow between different time segments; the output of the encoder is used as the latent vector in the latent space for the input of the decoder and the GAN discriminator;

[0032] The decoder of the GAN generator has the same structure as the encoder, with three additional linear layers for prediction. The output of the encoder is fed into the decoder embedding layer. The output of the second self-attention layer of the decoder is fed into the linear layer. The output of the linear layer is the prediction result.

[0033] The discriminator of GAN is a 3-layer perceptron, in which the 2 linear layers use the GELU activation function and the last output layer uses the Sigmoid activation function, so that the last layer of neurons outputs a decimal between 0 and 1 to determine the probability of normal or abnormal;

[0034] The encoder and decoder of the GAN generator use the GELU activation function; the Adam optimizer is used to train the GAN; the learning rate is set to 0.0001;

[0035] The loss function consists of two parts: the WMSE weighted mean square error is used to calculate the reconstruction loss of the GAN generator prediction data and the output time series segment; the cross entropy loss is used to calculate the classification result of the GAN discriminator. The specific formula is:

[0036] Loss = WMSE + L

[0037] In the formula, Loss is the total loss; WMSE is the weighted mean square error loss, and L is the cross entropy loss;

[0038]

[0039] Where n is the number of sample predictions, d is the sample parameter dimension; y′ ij is the model prediction value, y ij is the actual value; w j is the error loss of category j;

[0040]

[0041] Where: i is the observed sample, j is the category, M is the number of categories, p ij It represents the probability that the observed sample i belongs to category j.

[0042] Preferably, the confidence threshold t of the model discriminator is set in the range of 0.5 to 1. The larger the confidence threshold is set, the greater the probability of detecting an abnormality.

[0043] The continuous casting mold liquid level abnormality prediction system based on digital twin technology provided by the present invention includes:

[0044] Module M1: Building a high-fidelity continuous casting digital twin physical environment;

[0045] Construct 3D models of key equipment components including the tundish car, crystallizer, fan-shaped segment, bracket, and roller system, import 3D modeling and simulation engines, and assemble them into a complete continuous casting production line; use the particle system to simulate the pouring and flow of molten metal liquid to show the entire casting process;

[0046] Module M2: Collect and integrate multi-source heterogeneous data that affect the mold level during the casting process;

[0047] Connect with the monitoring system and data acquisition equipment on the industrial site to obtain real-time production data and control the movement trajectory and speed of the digital twin including the tundish, crystallizer, and roller system; combine the analysis of metallurgical principles and the experience of on-site operators to select the characteristic parameters that affect the crystallizer liquid level fluctuation;

[0048] Module M3: preprocesses the parameter data and divides the multivariate time series segments into fixed window length W as positive samples of the data-driven model, including input data and output data;

[0049] Module M4: Generates noise data from random noise as negative samples of the data-driven model. The dimension of the noise data is the same as the dimension of the input data after segmentation in module M3. The output data after segmentation in module M3 is copied as the output of the noise.

[0050] Module M5: Construct a sample set, take the input data in module M3 and the noise data in module M4 as input samples, record them as x1_true and x2_false respectively; take the output data in module M3 and the output of the noise in module M4 as output samples, record them as y1_true and y2_false respectively; at the same time, classify the real input data and the noise data, and mark them as 0 and 1 respectively;

[0051] Module M6: Building a digital twin data-driven model of the GAN network;

[0052] The GAN generator uses the transformer encoder-decoder architecture and introduces a cross-attention mechanism to capture dependencies across time and dimensions. The discriminator uses a multi-layer perceptron structure. The output of the encoder in the GAN generator is used as the input of the discriminator. The loss function of the discriminator uses a binary cross-entropy loss, and the output is a 0 or 1 label.

[0053] Module M7: Alternately train the GAN generator and discriminator;

[0054] The shuffle and batch functions of the python numpy library are used to randomly shuffle the sample data before training the GAN. When training the generator, the discriminator parameters are fixed and updated to minimize its loss function. When training the discriminator, the generator parameters are fixed and updated to minimize its loss function.

[0055] Module M8: After the model is evaluated as qualified, the abnormal situation of the crystallizer liquid level is judged, and the confidence threshold t is set. When the discriminator prediction score is greater than the threshold t, it is judged as abnormal; the prediction data output by the generator contains the change trend of all parameters in the future period of time, which is used to observe the change of the crystallizer liquid level in advance;

[0056] Module M9: Use the trained GAN data-driven model to perform online real-time prediction of abnormal crystallizer liquid level, and feed the model output back to the digital twin environment.

[0057] Preferably, characteristic parameters that affect the fluctuation of the crystallizer liquid level are selected, including 1 each of the net weight of molten steel in the tundish, the position of the stopper rod, the continuous temperature measurement in the casting area, the crystallizer liquid level, and the pulling speed, 4 crystallizer water outlet flow rates, 2 cylinder friction works, 11 drive roller torques, and 25 drawing roller currents, totaling 47 dimensions.

[0058] Preferably, the module M3 comprises:

[0059] Module M3.1: Filter the parameter data in module M2 according to the constraints to select the casting process data. The constraints are that the tundish is in the casting position, the liquid level of the crystallizer is greater than 750mm, and the continuous temperature measurement temperature in the casting area is greater than 1500℃.

[0060] Module M3.2: Use the fillna function in Python to forward fill in empty values ​​and obtain multivariate time series data;

[0061] Module M3.3: Split the multivariate time series data into fixed window length W to obtain a series of data segments of length W; each segment contains the values ​​of multiple variables, forming a two-dimensional tensor, which is used as a sample data; among them, the first W1 are used as sample input data, and (W-W1) are used as sample output data;

[0062] Module M3.4: Prepare multiple batches of data, each batch containing multiple samples, for subsequent model training.

[0063] Preferably, the module M6 comprises:

[0064] The GAN generator uses the transformer encoder-decoder architecture; the encoder maps the input data x1_true in module M3 and the noise data x2_false in module M4 to the latent space hidden vector z; the hidden vector z is used as the input of the decoder, and the decoder generates the output data y1_true in module M3 and the noise output y2_false according to the hidden vector z; at the same time, the hidden vector z output by the encoder is used as the input of the discriminator, and the discriminator classifies the hidden vector z and outputs 0 and 1 labels, 0 for normal and 1 for abnormal;

[0065] Both the encoder and decoder are composed of an embedding layer, a position encoding layer, and an attention layer; the embedding layer is a linear layer that maps the input feature vector to a higher-dimensional representation space; the position encoding layer is a learnable parameter matrix used to learn the relative or absolute position information of the high-dimensional sequence, and its shape is the same as the output of the embedding layer; the attention layer uses a scaled dot product attention mechanism, and on the basis of the dot product attention, the attention weight is scaled to improve the stability of the model. The specific formula is as follows:

[0066]

[0067] In the formula, Q·K T represents the dot product of the query vector Q and the key vector K, d k is the dimension scaling factor; V is the value vector;

[0068] The encoder of the GAN generator consists of 1 embedding layer, 1 position encoding layer, and 2 self-attention layers; the output of the embedding layer is added to the position encoding matrix, the added matrix vector is normalized, and sent to the first self-attention layer for information fusion; a learnable parameter matrix is ​​introduced as the query vector Q, the output of the first self-attention layer is used as the key vector K and the value vector V, and the three matrices are sent to the second self-attention layer. The parameter matrix is ​​used to control the intensity of information flow between different time segments; the output of the encoder is used as the latent vector in the latent space for the input of the decoder and the GAN discriminator;

[0069] The decoder of the GAN generator has the same structure as the encoder, with three additional linear layers for prediction. The output of the encoder is fed into the decoder embedding layer. The output of the second self-attention layer of the decoder is fed into the linear layer. The output of the linear layer is the prediction result.

[0070] The discriminator of GAN is a 3-layer perceptron, in which the 2 linear layers use the GELU activation function and the last output layer uses the Sigmoid activation function, so that the last layer of neurons outputs a decimal between 0 and 1 to determine the probability of normal or abnormal;

[0071] The encoder and decoder of the GAN generator use the GELU activation function; the Adam optimizer is used to train the GAN; the learning rate is set to 0.0001;

[0072] The loss function consists of two parts: the WMSE weighted mean square error is used to calculate the reconstruction loss of the GAN generator prediction data and the output time series segment; the cross entropy loss is used to calculate the classification result of the GAN discriminator. The specific formula is:

[0073] Loss = WMSE + L

[0074] In the formula, Loss is the total loss; WMSE is the weighted mean square error loss, and L is the cross entropy loss;

[0075]

[0076] Where n is the number of sample predictions, d is the sample parameter dimension; y′ ij is the model prediction value, y ij is the actual value; w j is the error loss of category j;

[0077]

[0078] Where: i is the observed sample, j is the category, M is the number of categories, p ij It represents the probability that the observed sample i belongs to category j.

[0079] Preferably, the confidence threshold t of the model discriminator is set in the range of 0.5 to 1. The larger the confidence threshold is set, the greater the probability of detecting an abnormality.

[0080] Compared with the prior art, the present invention has the following beneficial effects:

[0081] The digital twin technology is combined with the generative adversarial network (GAN) model, especially the introduction of the Transformer encoder-decoder architecture in the GAN generator for the prediction of abnormal crystallizer liquid level. The digital twin technology can build a high-fidelity continuous casting physical environment and provide a large amount of realistic simulation data for the GAN model, solving the problem of insufficient data caused by the lack of abnormal samples in traditional methods. At the same time, the Transformer structure is good at capturing the long-term dependencies of time series data and can better learn the inherent laws of liquid level fluctuations. It can predict the changing trend of the crystallizer liquid level online and predict whether the crystallizer liquid level is abnormal, which is more beneficial to the application of actual engineering. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] 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:

[0083] Figure 1 This is a flow chart of a method for predicting abnormal liquid level in a continuous casting mold based on digital twin technology in the present invention;

[0084] Figure 2 is a typical parameter curve diagram affecting the crystallizer liquid level in the present invention;

[0085] Figure 3 It is a thermal diagram of the correlation analysis of typical parameters affecting the liquid level of the crystallizer in the present invention;

[0086] Figure 4 It is a top graph of the correlation analysis of typical parameters affecting the crystallizer liquid level in the present invention;

[0087] Figure 5 This is a graph showing the loss curve of the model training in the present invention;

[0088] Figure 6 It is a segmentation diagram of normal data and noise data by the discriminator in the present invention;

[0089] Figure 7 It is the real curve diagram of some parameters in the present invention;

[0090] Figure 8 This is a prediction curve diagram of some parameters in the present invention. DETAILED DESCRIPTION

[0091] 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.

[0092] Example 1

[0093] like Figure 1 As shown, a method for predicting abnormal liquid level of a continuous casting mold based on digital twin technology includes the following steps:

[0094] Step 1: Build a high-fidelity continuous casting digital twin physical environment;

[0095] The three-dimensional models of key equipment components such as the car, crystallizer, fan-shaped segment, bracket, roller system, etc. are accurately modeled and imported into the 3D modeling and simulation engine to assemble a complete continuous casting production line; the particle system is used to simulate the pouring and flow of molten metal liquid to show the entire casting process.

[0096] Step 2: Collect and integrate multi-source heterogeneous data that affect the mold level during the casting process;

[0097] Connect with the monitoring system and data acquisition equipment at the industrial site to obtain real-time production data, control the movement trajectory and speed of digital twins such as the tundish car, crystallizer, and roller system; combine the analysis of metallurgical principles and the experience of on-site operators to select characteristic parameters that affect the crystallizer liquid level fluctuation; the parameters include tundish steel net weight, stopper rod position, continuous temperature measurement in the casting area, crystallizer liquid level, and drawing speed, 4 crystallizer water outlet flow rates, 2 cylinder friction works, 11 drive roller torques, and 25 drawing roller currents, totaling 47 dimensions.

[0098] In this embodiment, the actual performance data of continuous casting production of a steel plant is taken as an example, the sampling frequency is 1s / time, and the data of relevant parameters of the mold liquid level during the casting process of one week are collected; Figure 2 The following is a typical parameter curve that affects the crystallizer liquid level; Figure 3 Heat map for correlation analysis between parameters, Figure 4 This is the top diagram for correlation analysis between parameters. Due to the large number of parameters, typical parameters are selected to analyze their correlation. Among them, S1_STS_Car1_StopperPos is the stopper position; CMO1ST_Fast_uD_uCy1_uMsg_AvgFrictionWork is the friction work of the cylinder; STR_GERM_WS_M13_DRV_TRQ is the driving roller torque.

[0099] Step 3: Data preprocessing;

[0100] Step 3.1: Filter the parameter data in step 2 according to the constraint conditions to select the casting process data, where the constraint conditions are that the tundish is at the casting position, the liquid level of the crystallizer is greater than 750 mm, and the continuous temperature measurement temperature of the casting area is greater than 1500 °C;

[0101] Step 3.2: Use the fillna function in Python to forward fill in the empty values ​​and obtain multivariate time series data;

[0102] Step 3.3: Split the multivariate time series data into fixed window length W to obtain a series of data segments of length W; each segment contains the values ​​of multiple variables, forming a two-dimensional tensor, which is used as a sample data; the first W1 are used as sample input data, and (W-W1) are used as sample output data; in this embodiment, W is set to 360, W1 is 300, and (W-W1) is 60, which means that the historical data of the previous 5 minutes is used as input to predict the data change trend within the next 1 minute.

[0103] Step 3.4: Prepare multiple batches of data, each batch containing multiple samples, for subsequent model training.

[0104] Step 4: Generate noise data from random noise as negative samples of the GAN network. The dimension of the noise data is the same as the dimension of the input data after segmentation in step 3. Copy the output data after segmentation in step 3 as the output of the noise. The random noise is generated by Gaussian noise. The specific formula is:

[0105]

[0106] Where: d represents the dimension of sample data, R 1xd Indicates that the dimension of the vector generating noise is d, and the noise data is a d-dimensional vector obtained by random sampling from d independent Gaussian distributions, each dimension comes from the corresponding Gaussian distribution, μ is the mean of the Gaussian distribution, and σ is the standard deviation of the Gaussian distribution; in this embodiment, μ is set to 0 and σ is set to 1.

[0107] Step 5: Construct sample set;

[0108] The input data in step 3 and the noise data in step 4 are taken as input samples, denoted as x1_true and x2_false respectively; the output data in step 3 and the output of the noise in step 4 are taken as output samples, denoted as y1_true and y2_false respectively; at the same time, the real input data and the noise data are classified and marked as 0 and 1 respectively.

[0109] Step 6: Build a GAN to generate a discriminant network model:

[0110] Step 6.1: Model input and output;

[0111] The GAN generator adopts the encoder-decoder architecture of transformer; the encoder maps the input data x1_true in step 3 and the noise data x2_false in step 4 to the latent space latent vector z; the latent vector z is used as the input of the decoder, and the decoder generates the output data y1_true in step 3 and the noise output y2_false according to the latent vector z; at the same time, the latent vector z output by the encoder is used as the input of the discriminator, and the discriminator classifies the latent vector z and outputs 0 and 1 labels, 0 indicates normal and 1 indicates abnormal.

[0112] Step 6.2: Basic module;

[0113] Both the encoder and decoder are composed of basic transformer modules such as the embedding layer, position encoding layer, and attention layer. The embedding layer is a linear layer that maps the input feature vector to a higher-dimensional representation space. The position encoding layer is a learnable parameter matrix used to learn the relative or absolute position information of the high-dimensional sequence, and its shape is the same as the output of the embedding layer. The attention layer uses a scaled dot product attention mechanism, and scales the attention weights based on the dot product attention to improve the stability of the model. The specific formula is as follows:

[0114]

[0115] In the formula, Q·K T represents the dot product of the query vector Q and the key vector K, d k is the dimension scaling factor; V is the value vector;

[0116] Step 6.3: Encoder;

[0117] The encoder of the GAN generator consists of 1 embedding layer, 1 position encoding layer, and 2 self-attention layers; the output of the embedding layer is added to the position encoding matrix, the added matrix vector is normalized, and sent to the first self-attention layer for information fusion; a learnable parameter matrix is ​​introduced as the query vector Q, the output of the first self-attention layer is used as the key vector K and the value vector V, and the three matrices are sent to the second self-attention layer, and the parameter matrix is ​​used to control the intensity of information flow between different time segments; the output of the encoder is used as the latent vector in the latent space for the input of the decoder and the GAN discriminator. In this embodiment, the embedding layer is a linear layer with an input of 47 dimensions and an output of 128 dimensions, which converts sparse vectors into dense vectors and captures more information in the input data.

[0118] Step 6.4: Decoder;

[0119] The decoder and encoder structures of the GAN generator are the same, with three additional linear layers for prediction; the output of the encoder is fed into the decoder embedding layer; the output of the second self-attention layer of the decoder is fed into the linear layer; the output of the linear layer is the prediction result. In this embodiment, the output of the decoder embedding layer is 128 dimensions, and the three additional linear layers are 128, 47, and 47 respectively, which restore the dense vector to its original dimension for prediction.

[0120] Step 6.5: GAN discriminator;

[0121] The discriminator of GAN is a 3-layer perceptron, in which 2 linear layers use the GELU activation function, and the last output layer uses the Sigmoid activation function, so that the last layer of neurons outputs a decimal between 0 and 1, which is used to determine the probability of normal or abnormal. In this embodiment, the discriminator inputs a 47-dimensional feature vector, and the output dimensions of the 3-layer perceptron are 256, 33, and 1 respectively, which are finally used for classification.

[0122] Step 6.6: Model parameters;

[0123] The encoder and decoder of the GAN generator use the GELU activation function; the Adam optimizer is used to train the GAN; the learning rate is set to 0.0001;

[0124] The loss function consists of two parts: the WMSE weighted mean square error is used to calculate the reconstruction loss of the GAN generator prediction data and the output time series segment; the cross entropy loss is used to calculate the classification result of the GAN discriminator. The specific formula is:

[0125] Loss = WMSE + L

[0126] In the formula, Loss is the total loss; WMSE is the weighted mean square error loss, and L is the cross entropy loss;

[0127]

[0128] Where n is the number of sample predictions, d is the sample parameter dimension; y′ ij is the model prediction value, y ij is the actual value;

[0129]

[0130] Where: i is the observed sample, j is the category, M is the number of categories, y ij log(p 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 It represents the probability that the observed sample i belongs to category j.

[0131] In this embodiment, a total of 288,473 casting process parameter data were collected, and the data set was divided into training data and test data at a ratio of 70% and 30%; the model was trained in batches, with 64 samples in each batch and the dimension of each sample being (360, 47); the weight of the crystallizer liquid level in WMSE was 0.31, and the weights of other parameters were 0.015 respectively, focusing on learning the change trend of the crystallizer liquid level; Figure 5 It is the loss of the model training process. The horizontal axis represents the number of iterations (thousands). The model is trained until convergence.

[0132] Step 7: Alternately train the GAN generator and discriminator:

[0133] The shuffle and batch functions of the python numpy library are used to randomly shuffle the sample data and then train the GAN; when training the generator, the discriminator parameters are fixed and the generator parameters are updated to minimize its loss function; when training the discriminator, the generator parameters are fixed and the discriminator parameters are updated to minimize its loss function. In this embodiment, Figure 6 This is the segmentation map of the trained discriminator for normal data and noise data.

[0134] Step 8: After the model is evaluated as qualified, the abnormal situation of the crystallizer liquid level is judged, and the confidence threshold t is set. When the discriminator predicts that the score is greater than the threshold t, it is judged as abnormal; the prediction data output by the generator contains the change trend of all parameters in the future period of time, which can be used to observe the change of the crystallizer liquid level in advance. In this embodiment, Figure 7 and Figure 8 These are the true curves and predicted curves for some parameters. It can be seen from the figure that the predicted trend is roughly the same as the true trend, indicating that the model fits well.

[0135] Step 9: Use the trained GAN data-driven model to perform online real-time prediction of abnormal crystallizer liquid level, and feed the model output back to the digital twin environment.

[0136] Example 2

[0137] The present invention also provides a continuous casting mold liquid level anomaly prediction system based on digital twin technology. The continuous casting mold liquid level anomaly prediction system based on digital twin technology can be realized by executing the process steps of the continuous casting crystallizer liquid level anomaly prediction method based on digital twin technology, that is, technical personnel in this field can understand the continuous casting crystallizer liquid level anomaly prediction method based on digital twin technology as the preferred implementation mode of the continuous casting crystallizer liquid level anomaly prediction system based on digital twin technology.

[0138] The continuous casting mold liquid level abnormality prediction system based on digital twin technology provided by the present invention includes:

[0139] Module M1: Building a high-fidelity continuous casting digital twin physical environment;

[0140] Construct 3D models of key equipment components including the tundish car, crystallizer, fan-shaped segment, bracket, and roller system, import 3D modeling and simulation engines, and assemble them into a complete continuous casting production line; use the particle system to simulate the pouring and flow of molten metal liquid to show the entire casting process;

[0141] Module M2: Collect and integrate multi-source heterogeneous data that affect the mold level during the casting process;

[0142] Connect with the monitoring system and data acquisition equipment on the industrial site to obtain real-time production data and control the movement trajectory and speed of the digital twin including the tundish, crystallizer, and roller system; combine the analysis of metallurgical principles and the experience of on-site operators to select the characteristic parameters that affect the crystallizer liquid level fluctuation;

[0143] Module M3: preprocesses the parameter data and divides the multivariate time series segments into fixed window length W as positive samples of the data-driven model, including input data and output data;

[0144] Module M4: Generates noise data from random noise as negative samples of the data-driven model. The dimension of the noise data is the same as the dimension of the input data after segmentation in module M3. The output data after segmentation in module M3 is copied as the output of the noise.

[0145] Module M5: Construct a sample set, take the input data in module M3 and the noise data in module M4 as input samples, record them as x1_true and x2_false respectively; take the output data in module M3 and the output of the noise in module M4 as output samples, record them as y1_true and y2_false respectively; at the same time, classify the real input data and the noise data, and mark them as 0 and 1 respectively;

[0146] Module M6: Building a digital twin data-driven model of the GAN network;

[0147] The GAN generator uses the transformer encoder-decoder architecture and introduces a cross-attention mechanism to capture dependencies across time and dimensions. The discriminator uses a multi-layer perceptron structure. The output of the encoder in the GAN generator is used as the input of the discriminator. The loss function of the discriminator uses a binary cross-entropy loss, and the output is a 0 or 1 label.

[0148] Module M7: Alternately train the GAN generator and discriminator;

[0149] The shuffle and batch functions of the python numpy library are used to randomly shuffle the sample data before training the GAN. When training the generator, the discriminator parameters are fixed and updated to minimize its loss function. When training the discriminator, the generator parameters are fixed and updated to minimize its loss function.

[0150] Module M8: After the model is evaluated as qualified, the abnormal situation of the crystallizer liquid level is judged, and the confidence threshold t is set. When the discriminator prediction score is greater than the threshold t, it is judged as abnormal; the prediction data output by the generator contains the change trend of all parameters in the future period of time, which is used to observe the change of the crystallizer liquid level in advance;

[0151] Module M9: Use the trained GAN data-driven model to perform online real-time prediction of abnormal crystallizer liquid level, and feed the model output back to the digital twin environment.

[0152] The characteristic parameters that affect the fluctuation of the crystallizer liquid level are selected, including the net weight of molten steel in the tundish, the position of the stopper rod, the continuous temperature measurement in the casting area, the crystallizer liquid level, and the drawing speed, 4 crystallizer water outlet flow rates, 2 cylinder friction works, 11 driving roller torques, and 25 drawing roller currents, totaling 47 dimensions.

[0153] The module M3 comprises:

[0154] Module M3.1: Filter the parameter data in module M2 according to the constraints to select the casting process data. The constraints are that the tundish is in the casting position, the liquid level of the crystallizer is greater than 750mm, and the continuous temperature measurement temperature in the casting area is greater than 1500℃.

[0155] Module M3.2: Use the fillna function in Python to forward fill in empty values ​​and obtain multivariate time series data;

[0156] Module M3.3: Split the multivariate time series data into fixed window length W to obtain a series of data segments of length W; each segment contains the values ​​of multiple variables, forming a two-dimensional tensor, which is used as a sample data; among them, the first W1 are used as sample input data, and (W-W1) are used as sample output data;

[0157] Module M3.4: Prepare multiple batches of data, each batch containing multiple samples, for subsequent model training.

[0158] The module M6 comprises:

[0159] The GAN generator uses the transformer encoder-decoder architecture; the encoder maps the input data x1_true in module M3 and the noise data x2_false in module M4 to the latent space hidden vector z; the hidden vector z is used as the input of the decoder, and the decoder generates the output data y1_true in module M3 and the noise output y2_false according to the hidden vector z; at the same time, the hidden vector z output by the encoder is used as the input of the discriminator, and the discriminator classifies the hidden vector z and outputs 0 and 1 labels, 0 for normal and 1 for abnormal;

[0160] Both the encoder and decoder are composed of an embedding layer, a position encoding layer, and an attention layer; the embedding layer is a linear layer that maps the input feature vector to a higher-dimensional representation space; the position encoding layer is a learnable parameter matrix used to learn the relative or absolute position information of the high-dimensional sequence, and its shape is the same as the output of the embedding layer; the attention layer uses a scaled dot product attention mechanism, and on the basis of the dot product attention, the attention weight is scaled to improve the stability of the model. The specific formula is as follows:

[0161]

[0162] In the formula, Q·K T represents the dot product of the query vector Q and the key vector K, d k is the dimension scaling factor; V is the value vector;

[0163] The encoder of the GAN generator consists of 1 embedding layer, 1 position encoding layer, and 2 self-attention layers; the output of the embedding layer is added to the position encoding matrix, the added matrix vector is normalized, and sent to the first self-attention layer for information fusion; a learnable parameter matrix is ​​introduced as the query vector Q, the output of the first self-attention layer is used as the key vector K and the value vector V, and the three matrices are sent to the second self-attention layer. The parameter matrix is ​​used to control the intensity of information flow between different time segments; the output of the encoder is used as the latent vector in the latent space for the input of the decoder and the GAN discriminator;

[0164] The decoder of the GAN generator has the same structure as the encoder, with three additional linear layers for prediction. The output of the encoder is fed into the decoder embedding layer. The output of the second self-attention layer of the decoder is fed into the linear layer. The output of the linear layer is the prediction result.

[0165] The discriminator of GAN is a 3-layer perceptron, in which the 2 linear layers use the GELU activation function and the last output layer uses the Sigmoid activation function, so that the last layer of neurons outputs a decimal between 0 and 1 to determine the probability of normal or abnormal;

[0166] The encoder and decoder of the GAN generator use the GELU activation function; the Adam optimizer is used to train the GAN; the learning rate is set to 0.0001;

[0167] The loss function consists of two parts: the WMSE weighted mean square error is used to calculate the reconstruction loss of the GAN generator prediction data and the output time series segment; the cross entropy loss is used to calculate the classification result of the GAN discriminator. The specific formula is:

[0168] Loss = WMSE + L

[0169] In the formula, Loss is the total loss; WMSE is the weighted mean square error loss, and L is the cross entropy loss;

[0170]

[0171] Where n is the number of sample predictions, d is the sample parameter dimension; y′ ij is the model prediction value, y ij is the actual value; w j is the error loss of category j;

[0172]

[0173] Where: i is the observed sample, j is the category, M is the number of categories, p ij It represents the probability that the observed sample i belongs to category j.

[0174] The setting range of the model discriminator confidence threshold t is 0.5 to 1. The larger the confidence threshold is set, the greater the probability of detecting an anomaly.

[0175] 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.

[0176] 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 liquid level in a continuous casting mold based on digital twin technology, characterized in that: include: Step 1: Build a high-fidelity continuous casting digital twin physical environment; Construct 3D models of key equipment components including the tundish car, crystallizer, fan-shaped segment, bracket, and roller system, import 3D modeling and simulation engines, and assemble them into a complete continuous casting production line; use the particle system to simulate the pouring and flow of molten metal liquid to show the entire casting process; Step 2: Collect and integrate multi-source heterogeneous data that affect the mold level during the casting process; Connect with the monitoring system and data acquisition equipment on the industrial site to obtain real-time production data and control the movement trajectory and speed of the digital twin including the tundish, crystallizer, and roller system; combine the analysis of metallurgical principles and the experience of on-site operators to select the characteristic parameters that affect the crystallizer liquid level fluctuation; Step 3: Preprocess the parameter data and divide the multivariate time series segments into segments with a fixed window length W as positive samples of the data-driven model, including input data and output data; Step 4: Generate noise data from random noise as negative samples of the data-driven model. The dimension of the noise data is the same as the dimension of the input data after segmentation in step 3. Copy the output data after segmentation in step 3 as the output of the noise. Step 5: Construct a sample set, take the input data in step 3 and the noise data in step 4 as input samples, record them as x1_true and x2_false respectively; take the output data in step 3 and the output of the noise in step 4 as output samples, record them as y1_true and y2_false respectively; at the same time, classify the real input data and the noise data, and mark them as 0 and 1 respectively; Step 6: Build a digital twin data-driven model of the GAN network; The GAN generator uses the transformer encoder-decoder architecture and introduces a cross-attention mechanism to capture dependencies across time and dimensions. The discriminator uses a multi-layer perceptron structure. The output of the encoder in the GAN generator is used as the input of the discriminator. The loss function of the discriminator uses a binary cross-entropy loss, and the output is a 0 or 1 label. Specifically, the GAN generator uses the transformer encoder-decoder architecture; the encoder maps the input data x1_true in step 3 and the noise data x2_false in step 4 to the latent space hidden vector z; the hidden vector z is used as the input of the decoder, and the decoder generates the output data y1_true in step 3 and the noise output y2_false in step 4 according to the hidden vector z; at the same time, the hidden vector z output by the encoder is used as the input of the discriminator, and the discriminator classifies the hidden vector z and outputs 0 and 1 labels, 0 for normal and 1 for abnormal; Both the encoder and decoder consist of an embedding layer, a position encoding layer, and an attention layer; the embedding layer is a linear layer that maps the input feature vector to a higher-dimensional representation space; The position encoding layer is a learnable parameter matrix used to learn the relative or absolute position information of high-dimensional sequences. Its shape is the same as the output of the embedding layer. The attention layer uses a scaled dot product attention mechanism. On the basis of the dot product attention, the attention weight is scaled to improve the stability of the model. The specific formula is as follows: In the formula, Q·K T represents the dot product of the query vector Q and the key vector K, d k is the dimension scaling factor; V is the value vector; The encoder of the GAN generator consists of 1 embedding layer, 1 position encoding layer, and 2 self-attention layers; the output of the embedding layer is added to the position encoding matrix, the added matrix vector is normalized, and sent to the first self-attention layer for information fusion; a learnable parameter matrix is ​​introduced as the query vector Q, the output of the first self-attention layer is used as the key vector K and the value vector V, and the three matrices are sent to the second self-attention layer. The parameter matrix is ​​used to control the intensity of information flow between different time segments; the output of the encoder is used as the latent vector in the latent space for the input of the decoder and the GAN discriminator; The decoder of the GAN generator has the same structure as the encoder, plus 3 linear layers for prediction; The output of the encoder is sent to the decoder embedding layer; the output of the second self-attention layer of the decoder is sent to the linear layer; the output of the linear layer is the prediction result; The discriminator of GAN is a 3-layer perceptron, in which the 2 linear layers use the GELU activation function and the last output layer uses the Sigmoid activation function, so that the last layer of neurons outputs a decimal between 0 and 1 to determine the probability of normal or abnormal; Step 7: Alternately train the GAN generator and discriminator; The shuffle and batch functions of the python numpy library are used to randomly shuffle the sample data before training the GAN. When training the generator, the discriminator parameters are fixed and updated to minimize its loss function. When training the discriminator, the generator parameters are fixed and updated to minimize its loss function. Step 8: After the model is evaluated as qualified, the abnormal situation of the crystallizer liquid level is judged, and the confidence threshold t is set. When the discriminator prediction score is greater than the threshold t, it is judged as abnormal; The prediction data output by the generator contains the change trends of all parameters in the future, which is used to observe the changes of the crystallizer liquid level in advance; Step 9: Use the trained GAN data-driven model to perform online real-time prediction of abnormal crystallizer liquid level, and feed the model output back to the digital twin environment.

2. The method for predicting abnormal liquid level of a continuous casting mold based on digital twin technology according to claim 1, characterized in that: The characteristic parameters that affect the fluctuation of the crystallizer liquid level are selected, including the net weight of molten steel in the tundish, the position of the stopper rod, the continuous temperature measurement in the casting area, the crystallizer liquid level, and the drawing speed, 4 crystallizer water outlet flow rates, 2 cylinder friction works, 11 driving roller torques, and 25 drawing roller currents, totaling 47 dimensions.

3. The method for predicting abnormal liquid level of a continuous casting mold based on digital twin technology according to claim 1, characterized in that: The step 3 comprises: Step 3.1: Filter the parameter data in step 2 according to the constraint conditions to select the casting process data, where the constraint conditions are that the tundish is at the casting position, the liquid level of the crystallizer is greater than 750 mm, and the continuous temperature measurement temperature of the casting area is greater than 1500 °C; Step 3.2: Use the fillna function in Python to forward fill in the null values ​​and obtain multivariate time series data; Step 3.3: Split the multivariate time series data into fixed window length W to obtain a series of data segments of length W; each segment contains the values ​​of multiple variables, forming a two-dimensional tensor, which is used as a sample data; among them, the first W1 are used as sample input data, and (W-W1) are used as sample output data; Step 3.4: Prepare multiple batches of data, each batch containing multiple samples, for subsequent model training.

4. The method for predicting abnormal liquid level of a continuous casting mold based on digital twin technology according to claim 1, characterized in that: The step 6 comprises: The encoder and decoder of the GAN generator use the GELU activation function; the Adam optimizer is used to train the GAN; the learning rate is set to 0.0001; The loss function consists of two parts: the WMSE weighted mean square error is used to calculate the reconstruction loss of the GAN generator prediction data and the output time series segment; the cross entropy loss is used to calculate the classification result of the GAN discriminator. The specific formula is: Loss = WMSE + L In the formula, Loss is the total loss; WMSE is the weighted mean square error loss, and L is the cross entropy loss; Where n is the number of sample predictions, d is the sample parameter dimension; y′ ij is the model prediction value, y ij is the actual value; w j is the error loss of category j; Where: i is the observed sample, j is the category, M is the number of categories, p ij It represents the probability that the observed sample i belongs to category j.

5. The method for predicting abnormal liquid level of a continuous casting mold based on digital twin technology according to claim 1, characterized in that: The setting range of the model discriminator confidence threshold t is 0.5 to 1. The larger the confidence threshold is set, the greater the probability of detecting an anomaly.

6. A continuous casting mold liquid level abnormality prediction system based on digital twin technology, characterized in that: include: Module M1: Building a high-fidelity continuous casting digital twin physical environment; Construct 3D models of key equipment components including the tundish car, crystallizer, fan-shaped segment, bracket, and roller system, import 3D modeling and simulation engines, and assemble them into a complete continuous casting production line; use the particle system to simulate the pouring and flow of molten metal liquid to show the entire casting process; Module M2: Collect and integrate multi-source heterogeneous data that affect the mold level during the casting process; Connect with the monitoring system and data acquisition equipment on the industrial site to obtain real-time production data and control the movement trajectory and speed of the digital twin including the tundish, crystallizer, and roller system; combine the analysis of metallurgical principles and the experience of on-site operators to select the characteristic parameters that affect the crystallizer liquid level fluctuation; Module M3: preprocesses the parameter data and divides the multivariate time series segments into fixed window length W as positive samples of the data-driven model, including input data and output data; Module M4: Generates noise data from random noise as negative samples of the data-driven model. The dimension of the noise data is the same as the dimension of the input data after segmentation in module M3. The output data after segmentation in module M3 is copied as the output of the noise. Module M5: Construct a sample set, take the input data in module M3 and the noise data in module M4 as input samples, record them as x1_true and x2_false respectively; take the output data in module M3 and the output of the noise in module M4 as output samples, record them as y1_true and y2_false respectively; at the same time, classify the real input data and the noise data, and mark them as 0 and 1 respectively; Module M6: Building a digital twin data-driven model of the GAN network; The GAN generator uses the transformer encoder-decoder architecture and introduces a cross-attention mechanism to capture dependencies across time and dimensions. The discriminator uses a multi-layer perceptron structure. The output of the encoder in the GAN generator is used as the input of the discriminator. The loss function of the discriminator uses a binary cross-entropy loss, and the output is a 0 or 1 label. Specifically, the GAN generator adopts the transformer encoder-decoder architecture; the encoder maps the input data x1_true in module M3 and the noise data x2_false in module M4 to the latent space hidden vector z; the hidden vector z is used as the input of the decoder, and the decoder generates the output data y1_true in module M3 and the noise output y2_false in module M4 according to the hidden vector z; at the same time, the hidden vector z output by the encoder is used as the input of the discriminator, and the discriminator classifies the hidden vector z and outputs 0 and 1 labels, 0 for normal and 1 for abnormal; Both the encoder and decoder consist of an embedding layer, a position encoding layer, and an attention layer; the embedding layer is a linear layer that maps the input feature vector to a higher-dimensional representation space; The position encoding layer is a learnable parameter matrix used to learn the relative or absolute position information of high-dimensional sequences. Its shape is the same as the output of the embedding layer. The attention layer uses a scaled dot product attention mechanism. On the basis of the dot product attention, the attention weight is scaled to improve the stability of the model. The specific formula is as follows: In the formula, Q·K T represents the dot product of the query vector Q and the key vector K, d k is the dimension scaling factor; V is the value vector; The encoder of the GAN generator consists of 1 embedding layer, 1 position encoding layer, and 2 self-attention layers; the output of the embedding layer is added to the position encoding matrix, the added matrix vector is normalized, and sent to the first self-attention layer for information fusion; a learnable parameter matrix is ​​introduced as the query vector Q, the output of the first self-attention layer is used as the key vector K and the value vector V, and the three matrices are sent to the second self-attention layer. The parameter matrix is ​​used to control the intensity of information flow between different time segments; the output of the encoder is used as the latent vector in the latent space for the input of the decoder and the GAN discriminator; The decoder of the GAN generator has the same structure as the encoder, plus 3 linear layers for prediction; The output of the encoder is sent to the decoder embedding layer; the output of the second self-attention layer of the decoder is sent to the linear layer; the output of the linear layer is the prediction result; The discriminator of GAN is a 3-layer perceptron, in which the 2 linear layers use the GELU activation function and the last output layer uses the Sigmoid activation function, so that the last layer of neurons outputs a decimal between 0 and 1 to determine the probability of normal or abnormal; Module M7: Alternately train the GAN generator and discriminator; The shuffle and batch functions of the python numpy library are used to randomly shuffle the sample data before training the GAN. When training the generator, the discriminator parameters are fixed and updated to minimize its loss function. When training the discriminator, the generator parameters are fixed and updated to minimize its loss function. Module M8: After the model is evaluated as qualified, the abnormal situation of the crystallizer liquid level is judged, and the confidence threshold t is set. When the discriminator prediction score is greater than the threshold t, it is judged as abnormal; The prediction data output by the generator contains the change trends of all parameters in the future, which is used to observe the changes of the crystallizer liquid level in advance; Module M9: Use the trained GAN data-driven model to perform online real-time prediction of abnormal crystallizer liquid level, and feed the model output back to the digital twin environment.

7. The continuous casting mold liquid level abnormality prediction system based on digital twin technology according to claim 6 is characterized in that: The characteristic parameters that affect the fluctuation of the crystallizer liquid level are selected, including the net weight of molten steel in the tundish, the position of the stopper rod, the continuous temperature measurement in the casting area, the crystallizer liquid level, and the drawing speed, 4 crystallizer water outlet flow rates, 2 cylinder friction works, 11 driving roller torques, and 25 drawing roller currents, totaling 47 dimensions.

8. The continuous casting mold liquid level abnormality prediction system based on digital twin technology according to claim 6 is characterized in that: The module M3 comprises: Module M3.1: Filter the parameter data in module M2 according to the constraints to select the casting process data. The constraints are that the tundish is in the casting position, the liquid level of the crystallizer is greater than 750mm, and the continuous temperature measurement temperature in the casting area is greater than 1500℃. Module M3.2: Use the fillna function in Python to forward fill in empty values ​​and obtain multivariate time series data; Module M3.3: Split the multivariate time series data into fixed window length W to obtain a series of data segments of length W; each segment contains the values ​​of multiple variables, forming a two-dimensional tensor, which is used as a sample data; among them, the first W1 are used as sample input data, and (W-W1) are used as sample output data; Module M3.4: Prepare multiple batches of data, each batch containing multiple samples, for subsequent model training.

9. The continuous casting mold liquid level abnormality prediction system based on digital twin technology according to claim 6 is characterized in that: The module M6 comprises: The encoder and decoder of the GAN generator use the GELU activation function; the Adam optimizer is used to train the GAN; the learning rate is set to 0.0001; The loss function consists of two parts: the WMSE weighted mean square error is used to calculate the reconstruction loss of the GAN generator prediction data and the output time series segment; the cross entropy loss is used to calculate the classification result of the GAN discriminator. The specific formula is: Loss = WMSE + L In the formula, Loss is the total loss; WMSE is the weighted mean square error loss, and L is the cross entropy loss; Where n is the number of sample predictions, d is the sample parameter dimension; y′ ij is the model prediction value, y ij is the actual value; w j is the error loss of category j; Where: i is the observed sample, j is the category, M is the number of categories, p ij It represents the probability that the observed sample i belongs to category j.

10. The continuous casting mold liquid level abnormality prediction system based on digital twin technology according to claim 6, characterized in that: The setting range of the model discriminator confidence threshold t is 0.5 to 1. The larger the confidence threshold is set, the greater the probability of detecting an anomaly.

Images