Intelligent metallurgical process virtual simulation method and system based on digital twinning

Abstract

The invention relates to the technical field of metallurgical industry simulation and intelligent control, and discloses an intelligent metallurgical process virtual simulation method and system based on digital twinning. The method comprises the following steps: acquiring multi-modal high-dimensional data in a metallurgical process and reducing dimensions to obtain a feature vector set of a hidden layer space; performing disturbance injection simulation by using the feature vector set of the hidden layer space to obtain a multi-path state sequence with time dependence; predicting the abnormal path by using a recurrent neural network to obtain a predicted state evolution trajectory; utilizing a preset inverse mapping function to obtain multi-path state representation in the physical space; screening to obtain a risk path set; key evolution nodes are extracted from the set to be processed, and a virtual simulation scene is obtained; and performing optimization simulation on preset process adjustment parameters according to the virtual simulation scene, and determining optimized parameter configuration. The method can solve the problem that it is difficult to construct a comprehensive virtual simulation scene which truly restores the physical production rule.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of metallurgical industry simulation and intelligent control technology, and in particular to an intelligent metallurgical process virtual simulation method and system based on digital twinning. BACKGROUND

[0002] At present, the metallurgical industrial production process involves high-temperature multiphase reaction, complex mass and heat transfer behavior and extreme working condition environment, which is directly related to the quality of finished products, energy consumption and production safety. With the deep integration of industrial sensor technology and industrial internet, digital twinning technology has been gradually applied to real-time monitoring and virtual simulation of the whole metallurgical process. By constructing a mirror image of the physical entity in the virtual space, process parameter optimization and potential risk prediction are supported.

[0003] In one prior art, the metallurgical state is usually simulated or single path deduced by using physical mechanism model or simple historical trajectory data. Such method has certain reference value under the condition that the working condition is relatively stable. However, the data collected by the metallurgical field sensor has the characteristics of high dimension, high coupling and strong nonlinearity, covering temperature gradient, instantaneous pressure, dynamic concentration of components and other heterogeneous variables. In actual production, the metallurgical state is significantly affected by random disturbances such as raw material composition fluctuation and equipment operation deviation, showing extremely complex state evolution characteristics. When dealing with such high-dimensional characteristics, the prior art often has difficulty in balancing the real-time performance and physical accuracy of the deduction under limited computing resources.

[0004] There is a difficulty in constructing a comprehensive virtual simulation scene that truly restores the physical production law in the prior art. SUMMARY

[0005] The present application provides an intelligent metallurgical process virtual simulation method and system based on digital twinning to solve the difficulty in constructing a comprehensive virtual simulation scene that truly restores the physical production law in the prior art.

[0006] In a first aspect, to solve the above technical problems, the present application provides an intelligent metallurgical process virtual simulation method based on digital twinning, comprising: Obtaining multi-modal high-dimensional data of the metallurgical process, using the encoder part in the pre-trained variational autoencoder to reduce the dimension of the high-dimensional data, and obtaining a feature vector set of the hidden layer space; According to the feature vector set of the hidden layer space, using a pre-trained time series generative adversarial model to simulate random disturbance injection, and obtaining a multi-path state sequence with time dependence; If there is an abnormal path deviating from the preset path deviation threshold in the multi-path state sequence, using a pre-trained recurrent neural network to predict the time series evolution trend of the abnormal path, and obtaining a predicted state evolution trajectory; Using the decoder part of the variational autoencoder, a preset inverse mapping function is constructed to map the predicted state evolution trajectory to the physical space, thereby obtaining a multipath state representation in the physical space. Based on the multi-path state representation in the physical space, potential risk paths are screened by integrating preset association constraints to obtain a set of risk paths; Key evolutionary nodes are extracted from the risk path set, and data augmentation mapping is performed on the key evolutionary nodes using a pre-trained node extension generative adversarial model to obtain a virtual simulation scenario. Based on the virtual simulation scenario, the preset process adjustment parameters are optimized and simulated to determine the optimized parameter configuration.

[0007] Secondly, the present invention provides a digital twin-based intelligent metallurgical process virtual simulation system, comprising: The data dimensionality reduction module is used to acquire multimodal high-dimensional data of the metallurgical process, and to perform dimensionality reduction processing on the high-dimensional data using the encoder part of the pre-trained variational autoencoder to obtain a set of feature vectors in the hidden layer space. The multi-path simulation module is used to simulate random perturbation injection using a pre-trained temporal generative adversarial model based on the feature vector set of the hidden layer space, to obtain a time-dependent multi-path state sequence. The state prediction module is used to predict the temporal evolution trend of the abnormal path if there is an abnormal path in the multi-path state sequence that deviates from a preset path deviation threshold, by using a pre-trained recurrent neural network to obtain the predicted state evolution trajectory. The inverse mapping module is used to map the predicted state evolution trajectory to the physical space using a preset inverse mapping function constructed by the decoder part of the variational autoencoder, so as to obtain a multipath state representation in the physical space. The risk screening module is used to screen potential risk paths based on the multi-path state representation in the physical space and by integrating preset association constraints, thereby obtaining a set of risk paths. The scene generation module is used to extract key evolution nodes from the risk path set and use a pre-trained node extension generative adversarial model to perform data augmentation mapping on the key evolution nodes to obtain a virtual simulation scene. The parameter optimization module is used to optimize and simulate preset process adjustment parameters based on the virtual simulation scenario, and determine the optimized parameter configuration.

[0008] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention achieves high-precision mapping between high-dimensional feature space and physical space through an encoder-decoder architecture of a variational autoencoder. This invention utilizes a pre-trained variational autoencoder to perform nonlinear projection on high-dimensional heterogeneous data of the metallurgical process, reducing the original data to a low-dimensional manifold space for deduction. In the dimension recovery stage, the hidden layer manifold distribution features pre-learned by the decoder are used in conjunction with physical constraints for correction. Compared to traditional linear dimensionality reduction algorithms (such as PCA), this method effectively captures the complex nonlinear coupling relationships between metallurgical variables, eliminates numerical distortion and logical offsets generated when high-dimensional features are inversely mapped back to physical coordinates, and ensures the physical consistency of the simulation scenario.

[0009] (2) This invention improves the temporal logic and risk coverage of multi-path inference by introducing a temporal generative adversarial model with a time-step attention mechanism. This invention injects random perturbations into the hidden feature space and uses a temporal generative adversarial model to simulate multiple potential paths of metallurgical state evolution over time. Through the built-in time-step attention mechanism of the model, it ensures that the generated simulated instantaneous states have strong causal dependencies on the time axis, overcoming the problem of sequence disorder or physical mutation that easily occurs in traditional methods during random inference. This process enables the digital twin system to predict more extreme and rare potential risk conditions, enhancing the system's perception depth of complex dynamic production environments.

[0010] (3) This invention achieves a logical closed loop from risk perception to precise decision-making by combining a virtual scene with node-enhanced extension with a metallurgical kinetic evolution model. After identifying key abnormal evolution nodes, this invention utilizes a generative adversarial model to perform feature derivation and constructs comprehensive and formatted virtual scene data. By mapping scene components to a preset metallurgical kinetic evolution model, key process indicators such as reaction rate fluctuations and slag layer thickness distribution are simulated, and the process parameters are iteratively corrected using a multi-objective particle swarm optimization algorithm. This process achieves a deep transformation from abstract risk data to concrete kinetic simulation and then to executable parameter configuration, improving the targeting of process optimization and the steady-state control level of metallurgical production. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the digital twin intelligent metallurgical process virtual simulation method provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the digital twin intelligent metallurgical process virtual simulation system provided in the second embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] Reference Figure 1 The first embodiment of the present invention provides a digital twin-based intelligent metallurgical process virtual simulation method, comprising the following steps: S11, acquire multimodal high-dimensional data of the metallurgical process, and use the encoder part of the pre-trained variational autoencoder to perform dimensionality reduction processing on the high-dimensional data to obtain the feature vector set of the hidden layer space. S12, Based on the feature vector set of the hidden layer space, a random perturbation injection simulation is performed using a pre-trained temporal generative adversarial model to obtain a time-dependent multipath state sequence. S13, If there is an abnormal path in the multi-path state sequence that deviates from the preset path deviation threshold, the temporal evolution trend of the abnormal path is predicted by a pre-trained recurrent neural network to obtain the predicted state evolution trajectory. S14, using the preset inverse mapping function constructed by the decoder part of the variational autoencoder, the predicted state evolution trajectory is mapped to the physical space to obtain the multipath state representation in the physical space; S15, Based on the multi-path state representation in the physical space, potential risk paths are screened by integrating preset association constraints to obtain a set of risk paths; S16, extract key evolution nodes from the risk path set, and use a pre-trained node extension generative adversarial model to perform data augmentation mapping on the key evolution nodes to obtain a virtual simulation scenario; S17. Based on the virtual simulation scenario, optimize and simulate the preset process adjustment parameters to determine the optimized parameter configuration.

[0014] In step S11, multimodal high-dimensional data of the metallurgical process is acquired, and the high-dimensional data is reduced in dimensionality using the encoder part of a pre-trained variational autoencoder to obtain a set of feature vectors in the hidden layer space, including: Acquire multimodal high-dimensional data of the metallurgical process, and construct an initial high-dimensional matrix using the multimodal high-dimensional data; The initial high-dimensional matrix is ​​input into the encoder part, and the encoder part performs multimodal hybrid convolution mapping processing to project the initial high-dimensional matrix into a low-dimensional manifold space. The latent variable distribution parameters of the low-dimensional manifold space are extracted and determined as the feature vector set of the latent layer space.

[0015] In one implementation, this embodiment acquires the multimodal high-dimensional data through a multi-source sensor cluster. The multi-source sensor cluster includes, but is not limited to, platinum-rhodium thermocouples rigidly mounted at different depths within the furnace lining, a capacitive pressure transmitter installed in the dust removal duct at the furnace top, an electromagnetic flowmeter monitoring the oxygen lance system, and an infrared thermal imager installed in the observation window of the metallurgical furnace. The multimodal high-dimensional data includes real-time acquired smelting temperature data, furnace pressure data, oxygen supply flow rate data, and infrared thermal image data characterizing the spatial distribution of the temperature field on the surface of the molten pool within the furnace.

[0016] In one implementation, this embodiment utilizes the multimodal high-dimensional data to construct an initial high-dimensional matrix. First, this embodiment performs synchronous sampling processing on the raw electrical signals fed back from each sensor, setting the sampling frequency to 50Hz. This value is chosen to cover the characteristic frequencies of fluctuations in the main physical quantities during electric arc furnace steelmaking. Subsequently, this embodiment performs standardization processing on the synchronously acquired variable sequences. Specifically, for each dimension variable, its arithmetic mean and standard deviation are calculated within a preset historical observation period; the preset historical observation period is preferably real-time monitoring data from the most recent 24 hours or historical data from the most recent 100 complete metallurgical furnace cycles. This embodiment subtracts the arithmetic mean from the raw observation value and divides the resulting difference by the standard deviation to produce dimensionless standardized feature values. For infrared thermal image data, this embodiment converts it into corresponding grayscale matrix components and aligns and superimposes them with the standardized electrical signal features according to the sampling timestamp, thereby constructing the initial high-dimensional matrix containing time axis information, spatial field information, and multidimensional process components.

[0017] It should be noted that in this embodiment, the initial high-dimensional matrix is ​​input into the encoder part, which then performs multimodal hybrid convolutional mapping processing. The encoder part is constructed as a hybrid convolutional neural network architecture, which includes a first-order convolutional branch for extracting time-series correlation features of process variables (its kernel size is preset to be...). (with a step size of 1), and a second-order convolutional branch for extracting the spatial features of the furnace temperature field distribution in the infrared thermal image (its kernel size is preset to 1). (with a stride of 1). Rectified linear units (ReLUs) are nested between convolutional layers as nonlinear activation functions. In this embodiment, a multimodal feature fusion layer performs nonlinear projection on temporal and spatial features, eliminating data redundancy and thus compressing and projecting the high-dimensional original information onto a low-dimensional manifold space. This process effectively preserves the core manifold structure characterizing thermodynamic equilibrium and chemical reaction rates.

[0018] In the multimodal hybrid convolutional mapping process, after the temporal convolutional branch and the spatial convolutional branch extract features respectively, their output feature maps are spliced ​​together in the channel dimension to form a fused feature map. This fused feature map is then integrated and reduced in the channel dimension through a 1x1 convolutional layer before being input into the subsequent fully connected layer. This design ensures the deep fusion of temporal signals and spatial image information at the feature level.

[0019] It is worth noting that this embodiment extracts the latent variable distribution parameters of the low-dimensional manifold space. Specifically, the encoder part is connected to two parallel fully connected layers at its end: a mean mapping layer and a variance mapping layer. In this embodiment, the projected features are simultaneously input into these two layers, and the mean vector is output using a linear transformation matrix. Sum of logarithmic variance vector The set of feature vectors for the hidden layer space is determined. In this embodiment, the dimension of the feature vector set is preset to 256 dimensions. This dimension value is determined based on a compression ratio of 10% to 20% of the original high-dimensional data features, aiming to ensure that essential features are condensed while retaining sufficient expressive power, and to provide sufficient feature support for the reconstruction of the three-dimensional tensor in subsequent steps.

[0020] It should be noted that the pre-trained variational autoencoder was trained offline using image-electrical signal samples containing historical anomalous operating conditions. This embodiment constructs a composite loss function consisting of a reconstruction loss term and a divergence constraint term as the optimization objective. This embodiment employs the Adaptive Moment Estimation (Adam) algorithm for iterative solution, minimizing the reconstruction error to ensure sufficient model training without significant gradient vanishing risk, thus ensuring that the dimensionality reduction mapping process accurately preserves the metallurgical holographic state.

[0021] In step S12, based on the feature vector set of the hidden layer space, a pre-trained temporal generative adversarial model is used to simulate random perturbation injection, resulting in a time-dependent multipath state sequence, including: A preset random disturbance factor and a preset random noise vector are invoked, and the random disturbance factor is superimposed on the random noise vector to synthesize a disturbance input vector set; The set of feature vectors in the hidden layer space and the set of perturbation input vectors are synchronously input into the temporal generative adversarial model; The temporal dependency mapping is performed using the built-in time step attention mechanism of the temporal generative adversarial model to obtain a set of interrelated simulated instantaneous states; According to the generated time order, the data points in the simulated instantaneous state set are subjected to time sequence connection processing to construct the multipath state sequence.

[0022] In one implementation, this embodiment utilizes a pre-trained temporal generative adversarial model to simulate random perturbation injection. The temporal generative adversarial model is constructed as a sequence reasoning architecture based on a conditional generative adversarial network (CGAN), comprising a generator network and a discriminator network. The generator network employs a gated recurrent unit (GRU) as its core temporal processing unit, and nests an attention computation layer at each output time step to capture long-range temporal dependencies of the metallurgical state.

[0023] In one implementation, this embodiment calls a preset random noise vector. The random noise vector Follows a standard normal distribution with a mean of 0 and a variance of 1. The dimension is preset to 100, chosen to provide sufficient random entropy space to cover complex process fluctuations. The preset random disturbance factor characterizes uncontrollable physical fluctuations in the metallurgical process, such as batch differences in raw materials or equipment wear, and its value range is determined through statistical analysis. Deviation sequences related to oxygen lance height, bottom blowing flow rate, and raw material composition are collected from historical process logs at the metallurgical site; the standard deviation multiple of each deviation sequence relative to the mean is calculated. In this embodiment, the value range of the random disturbance factor is set between -0.15 and 0.15, which is consistent with the standardized data scale in S11, ensuring the physical rationality of the disturbance injection. This embodiment performs algebraic addition to superimpose the random disturbance factor onto the random noise vector. The perturbation input vector set is obtained by synthesizing the components of the perturbation input vector set.

[0024] It should be noted that in this embodiment, the feature vector set of the hidden layer space is used as a condition input to limit the generated evolution trajectory to always be within the neighborhood of the current metallurgical thermodynamic state, so as to prevent the generation of invalid paths that are detached from the actual production background.

[0025] It is worth noting that this embodiment utilizes the time-step attention mechanism built into the temporal generative adversarial model to perform temporal dependency mapping, obtaining a set of interrelated simulated instantaneous states. The time-step attention mechanism achieves cross-time-step information fusion by calculating the correlation weight between the hidden state at the current inference moment and the hidden state sequence at historical moments. Specifically, this embodiment uses a scaled dot product function to calculate attention scores and obtains a weight distribution through normalization, guiding the generator to produce state points that are logically causally continuous with historical trends. The time-step attention mechanism employs scaled dot product attention; at each GRU time step of the generator, the hidden state of the current GRU is used as the query vector, and the sequence of GRU hidden states from all previous time steps is used as the key and value vectors to calculate attention weights; the weighted summed context vector is fused with the input of the current GRU to jointly determine the output of that time step; this mechanism enables the generator to dynamically focus on the most critical information in historical states for the current inference.

[0026] In one implementation, this embodiment performs time-series sequence connection processing on the data points in the simulated instantaneous state set according to the generation time order to construct the multipath state sequence. This embodiment will continuously generate... The instantaneous state points are arranged vertically according to their timestamp indices, forming multiple trajectory sequences representing different directions of perturbation evolution. The preset time step sequence length.

[0027] It should be noted that the temporal generative adversarial model is trained offline. This embodiment constructs a composite loss function consisting of an adversarial loss term, a temporal consistency loss term, and a physical manifold constraint term. The physical manifold constraint term is implemented by calculating the Euclidean distance between the simulated instantaneous state produced by the generator and its projection onto a low-dimensional feature manifold, aiming to ensure that the generated trajectory strictly follows the laws of metallurgical physics. This embodiment determines the training convergence criterion using a performance-driven method. The discriminator loss volatility of 100 consecutive batches is statistically analyzed. If the volatility is less than a preset stability threshold (e.g., ...), ... If the discriminator and generator reach Nash equilibrium, the model converges and training terminates. This process ensures that the virtual trajectories generated by the system possess both statistical diversity and the ability to cover more extreme and rare potential risk conditions.

[0028] In step S13, if there are abnormal paths in the multi-path state sequence that deviate from a preset path deviation threshold, a pre-trained recurrent neural network is used to predict the temporal evolution trend of the abnormal paths to obtain the predicted state evolution trajectory, including: Extract the state vector from the multipath state sequence and calculate the Euclidean distance between the state vector and the preset standard reference vector to obtain the path deviation numerical set. If the value in the set of path deviation values ​​exceeds the preset path deviation threshold, the corresponding continuous abnormal node segments in the multi-path state sequence are extracted and determined as abnormal path data to be processed. The abnormal path data to be processed is input into the pre-trained recurrent neural network to perform feature hidden layer extraction processing to obtain the hidden layer state vector; The hidden layer state vector is used to perform multiple rounds of recursive iterative calculations to obtain the continuous evolution prediction value; Sequence recombination processing is performed based on the continuous evolution prediction values ​​to obtain the predicted state evolution trajectory.

[0029] In one implementation, this embodiment extracts state vectors from the multipath state sequence. The state vectors are hidden layer features generated in S12 and located in a low-dimensional manifold space. This embodiment calculates the Euclidean distance between the state vectors and a preset standard reference vector.

[0030] In one implementation, the preset standard benchmark vector is determined through statistical analysis. Specifically, the determination method involves retrieving a preset historical production database, selecting high-quality metallurgical batches with a yield greater than 98.5% and energy consumption within a preset low-energy consumption range; extracting the historical hidden layer feature sequence corresponding to the high-quality metallurgical batches, calculating the arithmetic mean of all sample points in each dimension, and constructing the preset standard benchmark vector. In this embodiment, the square root of the sum of squared differences between the current state vector and the preset standard benchmark vector in each dimension is calculated to produce a scalar value representing the degree to which the current evolutionary state deviates from the ideal operating condition. In this embodiment, all time steps in the multi-path state sequence are traversed to obtain the set of path deviation values ​​composed of distance values ​​at each time point.

[0031] It should be noted that this embodiment determines whether the values ​​in the path deviation value set exceed the preset path deviation threshold. The preset path deviation threshold is determined using a performance-driven method. During the model validation phase, the distance fluctuation distribution under normal metallurgical conditions is statistically analyzed, and its mean and standard deviation are calculated. Based on... In principle, the preset path deviation threshold is set as the sum of the mean and three times the standard deviation. If the distance value continuously exceeds this threshold within a certain time period, this embodiment extracts the corresponding continuous abnormal node segments in the multi-path state sequence and identifies them as abnormal path data to be processed. In this way, the system can automatically eliminate normal random noise and accurately capture abnormal disturbances with unstable trends.

[0032] It is worth noting that in this embodiment, the abnormal path data to be processed is input into the pre-trained recurrent neural network to perform feature hidden layer extraction processing to obtain the hidden layer state vector. The pre-trained recurrent neural network adopts a long short-term memory (LSTM) network architecture, which contains an input gate, a forget gate, and an output gate, and can effectively handle metallurgical evolution features with long-range dependencies.

[0033] In one implementation, this embodiment utilizes the hidden layer state vector to perform multiple rounds of recursive iterative calculations to obtain continuously evolving predicted values. This embodiment uses the initial offset time of the abnormal path data to be processed as the initial value for iteration. Combined with the perturbation characteristics of the current time, it performs the following trend deduction operation: nonlinear mapping is performed between the hidden layer state vector of the previous time step and the predicted output to produce the predicted value for the current time step and the updated hidden layer state vector. This embodiment performs multiple iterative cycles until the predicted time step covers the length of the continuous abnormal node segment. This embodiment, through this deduction rather than forced correction method, fully preserves the subsequent development trend of the abnormal offset under the inertia of physical laws.

[0034] It should be noted that in this embodiment, sequence recombination processing is performed based on the continuous evolution prediction values ​​to obtain the predicted state evolution trajectory. This embodiment replaces the anomalous parts in the original sequence with the generated continuous evolution prediction values ​​and uses a third-order spline interpolation algorithm to perform smooth transition processing on the boundaries at both ends of the replacement points, ensuring that the generated trajectory has temporal continuity in the hidden layer space, providing a highly reliable basis for the reconstruction of the physical space in subsequent steps.

[0035] It is worth noting that the pre-trained recurrent neural network was trained offline. This embodiment uses over 5000 hours of historical metallurgical process records as the training dataset, covering various typical operating condition fluctuation scenarios. This embodiment uses the root mean square error (RMSE) as the optimization objective, minimizing the residual between the predicted feature vector and the true historical feature sequence, and employs the root mean square propagation (RMSprop) optimizer for parameter fitting. This embodiment continuously updates the weights until the prediction accuracy tolerance on the validation set reaches a preset accuracy standard, ensuring that the trend prediction results have extremely high logical reliability.

[0036] In step S14, a preset inverse mapping function constructed using the decoder portion of the variational autoencoder is used to map the predicted state evolution trajectory to the physical space, obtaining a multipath state representation in the physical space, including: The predicted state evolution trajectory is decomposed into component dimensions to extract the trajectory feature vector sequence; The trajectory feature vector sequence is input into the preset inverse mapping function constructed by the decoder part, and the dimension recovery operation is performed using the hidden layer spatial manifold distribution features learned in advance by the decoder part to obtain the initial physical space coordinate set; The initial physical space coordinate set is iteratively corrected using preset physical constraints to obtain corrected physical space coordinates; By integrating the time step information and physical components in the corrected physical space coordinates, a multipath state representation in the physical space is mapped and synthesized.

[0037] In one implementation, this embodiment decomposes the predicted state evolution trajectory into component dimensions. The predicted state evolution trajectory is a low-dimensional feature sequence generated by the preceding steps, carrying the future instability trend. This embodiment divides the predicted state evolution trajectory equally along the time axis according to a preset time sampling step (e.g., 10 seconds) to extract the trajectory feature vector sequence. The selection of the time sampling step is based on balancing the smoothness of the deduction with computational efficiency, ensuring that key state inflection points are not lost during the reconstruction process. Each vector in the trajectory feature vector sequence represents the state representation in the hidden space at a specific moment.

[0038] In one implementation, this embodiment inputs the trajectory feature vector sequence to a preset inverse mapping function constructed by the decoder. The decoder is constructed as a deep neural network symmetrical to the encoder, containing three one-dimensional transposed convolutional layers. This embodiment utilizes the decoder to perform dimensionality restoration operations. Specifically, the model uses weight parameters generated during the pre-training phase by learning the distribution of massive physical space data, and performs transposed convolution operations to project the 256-dimensional latent variables back to the original high-dimensional feature space, obtaining the initial physical space coordinate set. Since infrared image features are introduced at the input, this coordinate set not only includes the initially restored melting temperature and furnace pressure values, but also feature components characterizing the spatial distribution.

[0039] It should be noted that this embodiment utilizes preset physical constraints to perform iterative correction processing on the initial physical space coordinate set. These preset physical constraints include mass conservation constraints and energy balance constraints. Specifically, the mass conservation constraint requires that the absolute value of the difference between the total input and output of each key element (such as carbon and oxygen) in the metallurgical reaction and the residual amount be within a preset residual range. The energy balance constraint specifically requires that heat input (such as arc heating and chemical reaction exothermics) and heat expenditure (such as heat carried away by cooling water, furnace heat dissipation, and sensible heat of molten steel) conform to a thermodynamic equilibrium. The significance of this step lies in correcting non-physical distortions that may occur in a purely data-driven neural network, ensuring that the simulation results conform to the metallurgical mechanism.

[0040] It is worth noting that the specific method for performing iterative correction processing in this embodiment is to construct a penalty function containing multiple physical mechanism deviations. The penalty function is constructed by weighted summation of the absolute values ​​of the deviations from the aforementioned physical constraints. This embodiment uses gradient descent to fine-tune the initial set of physical space coordinates. By minimizing the penalty function and performing recursive iterative operations, the coordinate value change between two consecutive iterations is less than a preset convergence threshold, thereby obtaining the corrected physical space coordinates.

[0041] In one implementation, the convergence threshold is set to 0.15. This value is chosen based on the static measurement accuracy (e.g., 1.5 degrees Celsius) of the core hardware sensors such as platinum-rhodium thermocouples in the multi-source sensor cluster. The convergence threshold is set to 0.1 times the lowest accuracy to ensure that the corrected value has physical significance and will not fall below the hardware measurement noise level due to excessive iteration.

[0042] For example, if the initial melting temperature value restored by the decoder at a certain moment is 1605 degrees Celsius, but the energy balance equation calculated based on the current power supply and cooling intensity shows a large heat surplus, this embodiment will finely adjust the temperature value to 1598 degrees Celsius to meet the heat balance through iterative correction processing, and use this as the corrected physical space coordinate at that moment.

[0043] It is worth noting that this embodiment integrates the time step information and physical components in the corrected physical space coordinates, mapping and synthesizing the multipath state representation in the physical space. This embodiment associates the corrected physical space coordinates at each time step with its corresponding timestamp, and follows the time axis dimension. The data is then reorganized. Finally, a complete trajectory containing multi-dimensional physical attribute values ​​and conforming to temporal logic is constructed, which is determined as the multi-path state representation in the physical space. This embodiment eliminates the numerical offset problem common in traditional simulations through the above recursive restoration and physical correction logic, achieving accurate regression of the virtual space to physical laws.

[0044] In step S15, based on the multi-path state representation in the physical space, potential risk paths are screened by fusing preset association constraints to obtain a set of risk paths, including: Retrieve the multipath state representation in the physical space, and separate the temperature gradient value, pressure fluctuation value and component concentration value from it to obtain the coupled feature vector; Matrix mapping operations are performed using the coupled feature vectors to determine the state evolution manifold under normal operating conditions; Calculate the geodesic distance of the multipath state representation in the physical space relative to the state evolution manifold to obtain the anomalous evolution metric. If the abnormal evolution metric exceeds the preset critical threshold boundary, then path filtering is performed according to the pre-calculated cumulative risk weight to obtain the risk path set.

[0045] It should be noted that the preset associated constraints include process coupling constraints and safety boundary constraints. Process coupling constraints are based on metallurgical mechanisms, such as empirical correlations between gas flow rate and top pressure in blast furnaces, and quantitative relationship models between oxygen blowing intensity and decarburization rate in converters. Safety boundary constraints are hard limits set according to equipment manuals and production procedures, such as the furnace lining hot spot temperature must not exceed the maximum allowable temperature of its refractory material, and the carbon monoxide concentration in flue gas must not exceed 50% of the lower explosive limit. When screening risky paths, the system will check whether the multipath state representation violates these constraints, and any violation at any point in time will cause the path to be marked as a potential risk.

[0046] In one implementation, this embodiment retrieves the multipath state representation in the physical space. The multipath state representation in the physical space is a time-series trajectory with definite physical dimensions produced by the preceding steps. This embodiment separates the temperature gradient value, pressure fluctuation value, and component concentration value from it to obtain a coupled feature vector sequence.

[0047] In one implementation, this embodiment performs spatial feature extraction on the multipath state representation in the physical space. For the physical state at each time step, the temperature gradient value is obtained by performing first-order partial differential operations on adjacent sampling points; high-frequency components are extracted from the total pressure signal using a high-pass filtering algorithm to determine the pressure fluctuation value; and time-varying data of key chemical components (such as the percentage of carbon content in molten steel) are extracted as the component concentration value. This embodiment concatenates the above three types of values ​​to construct a multi-dimensional coupled feature vector sequence.

[0048] It should be noted that this embodiment utilizes the coupled feature vector sequence to perform a Locally Linear Embedding (LLE) mapping operation to determine the state evolution manifold under normal operating conditions. The state evolution manifold represents the essential topological structure of metallurgical production within its stable operating range. This embodiment calls a pre-trained LLE mapping matrix to project the coupled feature vector sequence onto a low-dimensional manifold coordinate system.

[0049] It is worth noting that the mapping matrix used in the Locally Linear Embedded (LLE) mapping operation is obtained by minimizing the manifold reconstruction residual fitting during the pre-training stage. In this embodiment, the manifold nearest neighbor parameter k is determined to be 15 using an engineering calibration method. Specifically, the determination method involves executing a grid search algorithm in the offline stage to find the number of nearest neighbors that minimizes the variance of the reconstruction residual, ensuring that the evolving manifold can both filter out random noise from the sensor and accurately preserve the dynamic continuity characteristics of the metallurgical reaction.

[0050] It should be noted that in this embodiment, the geodesic distance of the multipath state representation in the physical space relative to the state evolution manifold is calculated to obtain the anomalous evolution metric. Compared with the Euclidean distance in linear space, the geodesic distance can measure the actual evolution length of the path along the surface of the nonlinear manifold, and more accurately reflect the degree of metallurgical state deviation. First, a k-nearest neighbor graph is constructed using the low-dimensional manifold coordinate point set obtained by the Locally Linear Embedding (LLE) mapping. Each manifold coordinate point is connected to its k nearest Euclidean points (k=15), and the weight of the edge is the Euclidean distance between the two points. Then, the mean point of all points in the normal operating condition manifold is used as the source point, and the projection point of the simulated state point to be evaluated on the manifold (determined by inverse LLE transformation or nearest neighbor search) is used as the target point. The Dijkstra algorithm is used to calculate the shortest path length on the k-nearest neighbor graph, and this length is the geodesic distance, which is used as the anomalous evolution metric.

[0051] It is worth noting that this embodiment determines whether the abnormal evolution metric exceeds a preset critical threshold boundary. The preset critical threshold boundary is determined through statistical analysis. The geodesic distance distribution corresponding to a preset historical stable operating dataset is retrieved, and the 99.5th percentile of this distribution is extracted as the abnormal trigger threshold. If the current abnormal evolution metric exceeds this threshold, the corresponding path is determined to have potential risk.

[0052] In one implementation, this embodiment performs path filtering based on pre-calculated cumulative risk weights to obtain the risk path set. The pre-calculated cumulative risk weights are determined through engineering calibration. Specifically, this embodiment uses the partial derivative matrix method to perform sensitivity analysis. By calculating the partial derivatives of the final material cost function with respect to physical quantities such as temperature gradient, pressure fluctuation, and concentration anomalies, the risk coefficients for each dimension are determined, and a linear weighted sum is performed based on the duration of the anomaly. This embodiment accumulates risk scores for paths that detect anomalies, retaining paths with a total score higher than a preset safety threshold, thereby constructing the risk path set.

[0053] For example, if a path exhibits an abnormal temperature gradient within 30 consecutive seconds and the geodesic distance remains at more than 1.2 times the critical threshold boundary, and the risk score corresponding to the path reaches a preset high-risk level after cumulative calculation (e.g., the score is higher than the set threshold of 4.0), then this embodiment determines the path as a risk path and stores it in the risk path set.

[0054] It should be noted that, through the above-mentioned filtering mechanism based on local linear embedding manifold learning and geodesic distance, this embodiment achieves automated high-precision filtering of massive multi-path evolution results while ensuring the rigor of physical logic, which greatly reduces the computational overhead of subsequent virtual simulation scene construction.

[0055] In step S16, key evolutionary nodes are extracted from the risk path set, and a pre-trained node-extended generative adversarial model is used to perform data augmentation mapping on the key evolutionary nodes to obtain a virtual simulation scenario, including: Identify the locations of state mutations in the set of risk paths and determine the key evolutionary nodes; The key evolution node is input into the node extension generative adversarial model, feature derivation mapping processing is performed, and an extended node data set topologically associated with the key evolution node is output. The extended node data set and the key evolution nodes are formatted and recombined to obtain virtual scene data, and the scene integrity metric is determined based on the preset data coverage rate of the virtual scene data in the feature space. If the scene integrity metric reaches the preset scene integrity threshold, the virtual simulation scene is output.

[0056] In one implementation, this embodiment identifies the state abrupt change locations in the risk path set and determines the key evolution nodes. The state abrupt change locations characterize the moments when physical characteristics undergo nonlinear and dramatic changes in the metallurgical evolution path. This embodiment performs second-order time derivative operations on each physical quantity in the risk path set to calculate the state evolution acceleration sequence.

[0057] In one implementation, this embodiment determines the mutation threshold through statistical analysis. The state evolution acceleration sequence under normal operating conditions is extracted from historical production data, and its standard deviation in each dimension is calculated. In this embodiment, the mutation threshold is set to a certain value. If the absolute value of the evolutionary acceleration of a node exceeds the mutation threshold, the position is determined to be the state mutation position, and the physical state data corresponding to the position is identified as the key evolutionary node. By introducing mutation determination logic based on standard deviation multiples, this embodiment can automatically and objectively capture evolutionary inflection points with high-risk characterization significance.

[0058] It should be noted that in this embodiment, the key evolutionary node is input into the pre-trained node extension generative adversarial model, feature derivation mapping processing is performed, and an extended node data set topologically associated with the key evolutionary node is output. Since spatial features containing infrared images are extracted through a hybrid convolutional neural network in S11, the key evolutionary node not only contains time-series information but also carries spatial field features. The node extension generative adversarial model is constructed as a conditional generative adversarial network, whose generator network adopts a multi-layer fully connected residual structure. In this embodiment, feature derivation mapping processing is performed, using the spatiotemporal physical components of the key evolutionary node as conditional vectors. Local manifold sampling is performed in the latent space of the model, and the generator network produces multiple derived state points that satisfy physical continuity within the topological domain of the key node, thereby constructing the extended node data set. The feature derivation process completes the evolutionary details missing in the risk path due to sampling frequency limitations. Specifically, the feature vector of the key evolution node is used as a condition and input into the generator network. At the same time, a random noise vector is sampled from the standard normal distribution and concatenated with the condition vector in the latent space. Based on this concatenated vector, the generator network generates a new feature vector that is close to the key node in the feature space (i.e., the Euclidean distance is less than the preset radius) and conforms to the physical distribution learned from the training data through forward propagation. This new feature vector is the extended node. By repeating this process and fine-tuning the noise input, a set of extended node data with topological association (i.e., within the local neighborhood of the feature space) can be generated.

[0059] It is worth noting that the node-extended generative adversarial model is generated by fitting a sample set simulated using historical sparse outlier samples combined with a mechanistic model during the offline training phase. In this embodiment, the extended node dataset and the key evolutionary nodes are formatted and recombined to obtain virtual scene data. This virtual scene data represents the multidimensional state distribution within the metallurgical furnace in the form of a structured tensor, with its dimension defined as... ,in The length of the time step sequence. and Represents the coordinates of the internal furnace space grid. Represents physical characteristic channels.

[0060] It should be noted that this embodiment determines the scene integrity metric based on the preset data coverage rate of the virtual scene data in the feature space. To avoid the computational obstacle (curse of dimensionality) caused by directly calculating the convex hull volume in a high-dimensional feature space, this embodiment utilizes the feature space coverage rate of the extended node dataset relative to the historical complete dataset for calculation. Specifically, this embodiment calculates the proportion of nodes in the extended node dataset that fall within the hypersphere boundary constructed from historical normal samples, and determines this proportion as the scene integrity metric. This metric method is not only more practical in engineering but also accurately reflects the degree to which the generated scene covers the known safety boundaries.

[0061] For example, in this embodiment, the scene integrity metric is calculated as the ratio of the number of covered grid points to the total number of grid points.

[0062] It is worth noting that this embodiment determines whether the scene integrity metric reaches a preset scene integrity threshold. The preset scene integrity threshold is determined using a performance-driven method. During the simulation experiment phase, the contribution rate of scene data at different integrity levels to the downstream optimization of simulation accuracy is statistically analyzed. The performance inflection point balancing sensitivity and specificity is found using detection performance analysis, and the value corresponding to this inflection point is set as the scene integrity threshold, for example, 0.88. If the scene integrity metric reaches this threshold, the virtual simulation scene is output.

[0063] In another implementation, if the scene integrity metric does not reach the preset scene integrity threshold, this embodiment adjusts the sampling parameters of the node-extended generative adversarial model, increases the search radius of the latent space, and re-executes the feature derivation mapping process until the generated scene data meets the integrity requirements. This embodiment achieves a deep transformation from discrete anomaly prediction to a complete simulation scene with spatial dimensional information through the above process.

[0064] In step S17, based on the virtual simulation scenario, the preset process adjustment parameters are optimized and simulated to determine the optimized parameter configuration, including: The virtual simulation scene is indexed and sliced ​​according to the feature channel dimension to extract the furnace temperature distribution feature vector and the time-varying sequence of component concentration; The furnace temperature distribution feature vector is subjected to inverse normalization and thermodynamic temperature conversion based on preset statistical parameters to obtain the furnace temperature physical distribution matrix; The physical distribution matrix of furnace temperature and the time-varying sequence of component concentration are mapped to a preset metallurgical kinetic evolution model to perform simulation calculations, thereby obtaining the reaction rate fluctuation curve and slag layer thickness distribution map. If the reaction rate fluctuation curve exceeds the preset steady-state operating range, then the set of key process adjustment parameters is located from the preset process adjustment parameters based on the slag layer thickness distribution map. The set of key process adjustment parameters is subjected to multi-objective particle swarm optimization calculation using a normalized multi-objective fitness function to obtain optimized parameter values. The optimized parameter values ​​are then used to update the preset process adjustment parameters to determine the optimized parameter configuration.

[0065] In one implementation, this embodiment indexes and slices the virtual simulation scene according to the feature channel dimension. The virtual simulation scene is a structured four-dimensional tensor produced in the preceding steps, and its dimension is defined as... ,in Represents the length of the time step sequence. and Represents the coordinates of the internal furnace space grid. This represents a physical characteristic channel. This embodiment uses a fixed... The temperature values ​​at all time and space points are extracted using the index number corresponding to the temperature field in the dimension, forming a furnace temperature distribution feature vector; similarly, by fixing... The time-varying sequence of component concentrations is obtained by extracting the index number corresponding to the concentration of each component in the dimension.

[0066] In one implementation, this embodiment performs denormalization and thermodynamic temperature conversion processing on the furnace temperature distribution feature vector based on preset statistical parameters. Since the original data underwent standardization processing during the input stage, the resulting feature vector is a dimensionless statistical value. This embodiment calls the arithmetic mean of a preset historical observation period (e.g., the most recent 24 hours) stored in S11. with standard deviation Perform destandardization calculation:

[0067] in, This represents the normalized element value in the furnace temperature distribution feature vector. Subsequently, this embodiment will restore the obtained Celsius value. Adding 273.15 converts it to Kelvin temperature, which conforms to the laws of thermodynamics, to obtain the thermodynamic temperature. Thus, the physical distribution matrix of the furnace temperature is constructed.

[0068] It should be noted that this embodiment utilizes a pre-defined metallurgical kinetic evolution model to perform simulation calculations. This model integrates the Arrhenius equation:

[0069] in, The reaction rate constant is... The pre-exponential factor is the preset factor. This is the preset activation energy for the reaction, and its value is determined based on the kinetic constants of the dominant chemical reactions (such as carbon oxidation or dephosphorization) at the current metallurgical stage. Let be the ideal gas constant. The temperature is the thermodynamic temperature obtained from the aforementioned calculation. In this embodiment, by solving this equation and combining it with the component mass conservation equation, the reaction rate fluctuation curve and the slag layer thickness distribution map are output.

[0070] It is worth noting that the preset metallurgical kinetic evolution model is a coupled equation solver; firstly, the furnace temperature physical distribution matrix is ​​input into the Arrhenius equation to calculate the reaction rate constants at different spatial locations at different times. Next, Along with the time-varying sequence of component concentrations, a set of mass conservation differential equations (describing the oxidation reactions of elements such as carbon, silicon, and manganese) are input. The Runge-Kutta method is used for numerical integration to solve for the rate of change of each component's concentration over time and space. Further integration yields a new concentration distribution and reaction rate fluctuation curve. Simultaneously, based on the amount of oxides generated, the slag basicity model, and a simplified model of the molten pool flow, a slag layer thickness distribution map is calculated and output. This model has been calibrated offline using a large amount of historical production data for key parameters (such as pre-exponential factor and activation energy).

[0071] It is worth noting that this embodiment determines whether the reaction rate fluctuation curve exceeds the preset steady-state operating range. If it does, this embodiment uses a sensitivity analysis matrix to locate the process variable that contributes the most to the current reaction fluctuation, based on the thickness gradient of each spatial grid point in the slag layer thickness distribution map. The sensitivity analysis matrix constructed in this embodiment is based on the perturbation method, which involves injecting small perturbations (e.g., set values) into each process adjustment parameter sequentially during simulation calculations. The feedback change rate of the reaction rate fluctuation curve is observed to quantify the contribution weight of each parameter (such as oxygen supply intensity, bottom blowing flow rate, etc.) and determine the set of key process adjustment parameters accordingly.

[0072] For example, this embodiment performs multi-objective particle swarm optimization on the set of key process adjustment parameters. This embodiment sets the particle swarm size to 50 and the maximum number of iterations to 100. These parameters are set to balance computational overhead and convergence accuracy. Experimental calibration shows that with a particle size of 50, the algorithm can converge to the optimal solution within 2 seconds, meeting the real-time interaction and high-frequency tuning requirements of the digital twin system. This embodiment constructs a normalized multi-objective fitness function. :

[0073] in, The target value for the reaction rate, The variance of the slag layer thickness distribution. This serves as a historical reference variance benchmark. By introducing a relative deviation form for normalization, this embodiment eliminates the dimensional and numerical magnitude gaps between different physical quantities, ensuring the balance of optimization directions. This embodiment uses particle swarm optimization iterative search to make the function... The minimized global optimal solution is determined as the value of the optimization parameter, and the process parameter configuration is updated accordingly.

[0074] It is worth noting that this embodiment solves the adaptation problem between high-dimensional statistical features and physical mechanism models through the above-mentioned anti-standardization conversion, sensitivity analysis based on perturbation method and normalization optimization logic, and realizes a precise decision-making closed loop from virtual risk extrapolation to physical process parameter adjustment.

[0075] In summary, this embodiment discloses a digital twin-based intelligent metallurgical process virtual simulation method. It achieves accurate reconstruction of complex metallurgical processes through deep integration of variational autoencoders and generative adversarial models, and realizes closed-loop decision-making from virtual risk prediction to physical process parameters through coupling of metallurgical dynamics models and optimization algorithms. This invention solves the technical problem in existing technologies of constructing comprehensive virtual simulation scenarios that realistically reproduce the physical laws of production.

[0076] Reference Figure 2 The second embodiment of the present invention provides a digital twin-based intelligent metallurgical process virtual simulation system, comprising: The data dimensionality reduction module is used to acquire multimodal high-dimensional data of the metallurgical process, and to perform dimensionality reduction processing on the high-dimensional data using the encoder part of the pre-trained variational autoencoder to obtain a set of feature vectors in the hidden layer space. The multi-path simulation module is used to simulate random perturbation injection using a pre-trained temporal generative adversarial model based on the feature vector set of the hidden layer space, to obtain a time-dependent multi-path state sequence. The state prediction module is used to predict the temporal evolution trend of the abnormal path if there is an abnormal path in the multi-path state sequence that deviates from a preset path deviation threshold, by using a pre-trained recurrent neural network to obtain the predicted state evolution trajectory. The inverse mapping module is used to map the predicted state evolution trajectory to the physical space using a preset inverse mapping function constructed by the decoder part of the variational autoencoder, so as to obtain a multipath state representation in the physical space. The risk screening module is used to screen potential risk paths based on the multi-path state representation in the physical space and by integrating preset association constraints, thereby obtaining a set of risk paths. The scene generation module is used to extract key evolution nodes from the risk path set and use a pre-trained node extension generative adversarial model to perform data augmentation mapping on the key evolution nodes to obtain a virtual simulation scene. The parameter optimization module is used to optimize and simulate preset process adjustment parameters based on the virtual simulation scenario, and determine the optimized parameter configuration.

[0077] It should be noted that the digital twin intelligent metallurgical process virtual simulation system provided in this embodiment of the invention is used to execute all the process steps of the digital twin intelligent metallurgical process virtual simulation method in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0078] This invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a digital twin intelligent metallurgical process virtual simulation program. When the processor executes the computer program, it implements the steps in the various embodiments of the digital twin intelligent metallurgical process virtual simulation methods described above, for example... Figure 1 The step S11 shown. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the above system embodiments, such as the data dimensionality reduction module.

[0079] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0080] The electronic device may be a desktop computer, laptop, handheld computer, or smart tablet, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. It may include more or fewer components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0081] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting all parts of the electronic device via various interfaces and lines.

[0082] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0083] Wherein, if the modules / units integrated in the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0084] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0085] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A virtual simulation method for intelligent metallurgical processes based on digital twins, characterized in that, include: Multimodal high-dimensional data of the metallurgical process is acquired, and the high-dimensional data is reduced in dimensionality using the encoder part of a pre-trained variational autoencoder to obtain a set of feature vectors in the hidden layer space. Based on the feature vector set of the hidden layer space, a random perturbation injection simulation is performed using a pre-trained temporal generative adversarial model to obtain a time-dependent multipath state sequence. If there is an abnormal path in the multi-path state sequence that deviates from the preset path deviation threshold, the temporal evolution trend of the abnormal path is predicted by a pre-trained recurrent neural network to obtain the predicted state evolution trajectory. Using the decoder part of the variational autoencoder, a preset inverse mapping function is constructed to map the predicted state evolution trajectory to the physical space, thereby obtaining a multipath state representation in the physical space. Based on the multi-path state representation in the physical space, potential risk paths are screened by integrating preset association constraints to obtain a set of risk paths; Key evolutionary nodes are extracted from the risk path set, and data augmentation mapping is performed on the key evolutionary nodes using a pre-trained node extension generative adversarial model to obtain a virtual simulation scenario. Based on the virtual simulation scenario, the preset process adjustment parameters are optimized and simulated to determine the optimized parameter configuration.

2. The intelligent metallurgical process virtual simulation method based on digital twins according to claim 1, characterized in that, The process of acquiring multimodal high-dimensional data of the metallurgical process involves using the encoder portion of a pre-trained variational autoencoder to perform dimensionality reduction on the high-dimensional data, resulting in a set of feature vectors in the hidden layer space, including: Acquire multimodal high-dimensional data of the metallurgical process, and construct an initial high-dimensional matrix using the multimodal high-dimensional data; The initial high-dimensional matrix is ​​input into the encoder part, and the encoder part performs multimodal hybrid convolution mapping processing to project the initial high-dimensional matrix into a low-dimensional manifold space. The latent variable distribution parameters of the low-dimensional manifold space are extracted and determined as the feature vector set of the latent layer space.

3. The intelligent metallurgical process virtual simulation method based on digital twins according to claim 1, characterized in that, The step of using a pre-trained temporal generative adversarial model to simulate random perturbation injection based on the feature vector set of the hidden layer space to obtain a time-dependent multipath state sequence includes: A preset random disturbance factor and a preset random noise vector are invoked, and the random disturbance factor is superimposed on the random noise vector to synthesize a disturbance input vector set; The set of feature vectors in the hidden layer space and the set of perturbation input vectors are synchronously input into the temporal generative adversarial model; The temporal dependency mapping is performed using the built-in time step attention mechanism of the temporal generative adversarial model to obtain a set of interrelated simulated instantaneous states; According to the generated time order, the data points in the simulated instantaneous state set are subjected to time sequence connection processing to construct the multipath state sequence.

4. The intelligent metallurgical process virtual simulation method based on digital twins according to claim 1, characterized in that, If there are abnormal paths in the multi-path state sequence that deviate from a preset path deviation threshold, a pre-trained recurrent neural network is used to predict the temporal evolution trend of the abnormal paths to obtain the predicted state evolution trajectory, including: Extract the state vector from the multipath state sequence and calculate the Euclidean distance between the state vector and the preset standard reference vector to obtain the path deviation numerical set. If the value in the set of path deviation values ​​exceeds the preset path deviation threshold, the corresponding continuous abnormal node segments in the multi-path state sequence are extracted and determined as abnormal path data to be processed. The abnormal path data to be processed is input into the pre-trained recurrent neural network to perform feature hidden layer extraction processing to obtain the hidden layer state vector; The hidden layer state vector is used to perform multiple rounds of recursive iterative calculations to obtain the continuous evolution prediction value; Sequence recombination processing is performed based on the continuous evolution prediction values ​​to obtain the predicted state evolution trajectory.

5. The intelligent metallurgical process virtual simulation method based on digital twins according to claim 1, characterized in that, The preset inverse mapping function constructed using the decoder part of the variational autoencoder maps the predicted state evolution trajectory to the physical space, obtaining a multipath state representation in the physical space, including: The predicted state evolution trajectory is decomposed into component dimensions to extract the trajectory feature vector sequence; The trajectory feature vector sequence is input into the preset inverse mapping function constructed by the decoder part, and the dimension recovery operation is performed using the hidden layer spatial manifold distribution features learned in advance by the decoder part to obtain the initial physical space coordinate set; The initial physical space coordinate set is iteratively corrected using preset physical constraints to obtain corrected physical space coordinates; By integrating the time step information and physical components in the corrected physical space coordinates, a multipath state representation in the physical space is mapped and synthesized.

6. The intelligent metallurgical process virtual simulation method based on digital twins according to claim 1, characterized in that, The step involves filtering potential risk paths based on the multi-path state representation in the physical space and integrating preset association constraints to obtain a set of risk paths, including: Retrieve the multipath state representation in the physical space, and separate the temperature gradient value, pressure fluctuation value and component concentration value from it to obtain the coupled feature vector; Matrix mapping operations are performed using the coupled feature vectors to determine the state evolution manifold under normal operating conditions; Calculate the geodesic distance of the multipath state representation in the physical space relative to the state evolution manifold to obtain the anomalous evolution metric. If the abnormal evolution metric exceeds the preset critical threshold boundary, then path filtering is performed according to the pre-calculated cumulative risk weight to obtain the risk path set.

7. The intelligent metallurgical process virtual simulation method based on digital twins according to claim 1, characterized in that, The process of extracting key evolutionary nodes from the risk path set and using a pre-trained node-extended generative adversarial model to perform data augmentation mapping on the key evolutionary nodes to obtain a virtual simulation scenario includes: Identify the locations of state mutations in the set of risk paths and determine the key evolutionary nodes; The key evolution node is input into the node extension generative adversarial model, feature derivation mapping processing is performed, and an extended node data set topologically associated with the key evolution node is output. The extended node data set and the key evolution nodes are formatted and recombined to obtain virtual scene data, and the scene integrity metric is determined based on the preset data coverage rate of the virtual scene data in the feature space. If the scene integrity metric reaches the preset scene integrity threshold, the virtual simulation scene is output.

8. The intelligent metallurgical process virtual simulation method based on digital twins according to claim 1, characterized in that, The step of optimizing and simulating preset process adjustment parameters based on the virtual simulation scenario to determine the optimized parameter configuration includes: The virtual simulation scene is indexed and sliced ​​according to the feature channel dimension to extract the furnace temperature distribution feature vector and the time-varying sequence of component concentration; The furnace temperature distribution feature vector is subjected to inverse normalization and thermodynamic temperature conversion based on preset statistical parameters to obtain the furnace temperature physical distribution matrix; The physical distribution matrix of furnace temperature and the time-varying sequence of component concentration are mapped to a preset metallurgical kinetic evolution model to perform simulation calculations, thereby obtaining the reaction rate fluctuation curve and slag layer thickness distribution map. If the reaction rate fluctuation curve exceeds the preset steady-state operating range, then the set of key process adjustment parameters is located from the preset process adjustment parameters based on the slag layer thickness distribution map. The set of key process adjustment parameters is subjected to multi-objective particle swarm optimization calculation using a normalized multi-objective fitness function to obtain optimized parameter values. The optimized parameter values ​​are then used to update the preset process adjustment parameters to determine the optimized parameter configuration.

9. A digital twin-based intelligent metallurgical process virtual simulation system, characterized in that, include: The data dimensionality reduction module is used to acquire multimodal high-dimensional data of the metallurgical process, and to perform dimensionality reduction processing on the high-dimensional data using the encoder part of the pre-trained variational autoencoder to obtain a set of feature vectors in the hidden layer space. The multi-path simulation module is used to simulate random perturbation injection using a pre-trained temporal generative adversarial model based on the feature vector set of the hidden layer space, to obtain a time-dependent multi-path state sequence. The state prediction module is used to predict the temporal evolution trend of the abnormal path if there is an abnormal path in the multi-path state sequence that deviates from a preset path deviation threshold, by using a pre-trained recurrent neural network to obtain the predicted state evolution trajectory. The inverse mapping module is used to map the predicted state evolution trajectory to the physical space using a preset inverse mapping function constructed by the decoder part of the variational autoencoder, so as to obtain a multipath state representation in the physical space. The risk screening module is used to screen potential risk paths based on the multi-path state representation in the physical space and by integrating preset association constraints, thereby obtaining a set of risk paths. The scene generation module is used to extract key evolution nodes from the risk path set and use a pre-trained node extension generative adversarial model to perform data augmentation mapping on the key evolution nodes to obtain a virtual simulation scene. The parameter optimization module is used to optimize and simulate preset process adjustment parameters based on the virtual simulation scenario, and determine the optimized parameter configuration.

Images