Virtual simulation and state prediction system for operation data of glass production line

Abstract

The invention relates to the field of intelligent manufacturing and industrial digital twinning, in particular to a virtual simulation and state prediction system for operation data of a glass production line. The system comprises a data acquisition module, a reduced-order fluid simulation module, a time sequence deep learning module and a reverse optimization module. The system drives a reduced-order physical model based on boundary and disturbance variables to generate virtual flow field features reflecting heat flow coupling, and spatial-temporal feature fusion is carried out to solve the defect occurrence probability; the core of the method is that an invisible high-temperature fluid state in a melting furnace is reconstructed at a second-level scale by utilizing a reduced-order model, and recommended operation parameters are output based on a reverse algorithm when the risk exceeds the limit; according to the method, the measurement blind area of the sensor is effectively filled, the calculation cost is greatly reduced, and real-time perspective and accurate regulation and control of the black box production environment are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and industrial digital twin technology, specifically a virtual simulation and status prediction system for glass production line operation data. Background Technology

[0002] In the current glass production line environment, the operation of the melting furnace is characterized by high temperature, high hysteresis and black box. During the production process, boundary condition data such as fuel flow rate and combustion air pressure, as well as disturbance variable data such as raw material composition and ambient temperature and humidity, are continuously generated. These data are usually scattered in heterogeneous systems such as DCS, LIMS and cold end detection.

[0003] For the aforementioned scenarios, existing quality control solutions generally rely on limited sensor monitoring or post-processing adjustments based on human experience. Because traditional sensors cannot penetrate deep into the liquid glass for measurement, blind spots exist in the observation of the high-temperature fluid state inside the melting furnace. Furthermore, existing quality inspections must wait until the glass melt has cooled and solidified; this delayed feedback mechanism struggles to capture the dynamic nonlinear effects of raw material composition fluctuations or environmental changes on the flow field evolution. This results in significant product loss by the time defects are detected, making it impossible to achieve refined and proactive closed-loop control of process parameters.

[0004] Therefore, how to effectively integrate multi-dimensional time-series data to gain real-time insight into the internal flow field state, and to achieve accurate prediction and reverse optimization control of future finished product quality within a second-level time scale, thereby reducing the scrap rate, has become an urgent technical problem to be solved. Summary of the Invention

[0005] The purpose of this invention is to provide a virtual simulation and status prediction system for glass production line operation data to solve the problems mentioned in the background art. Specifically, the technical solution of this invention includes:

[0006] The data acquisition module is used to acquire real-time operating data and historical quality data of the glass production line. The real-time operating data includes boundary condition data and disturbance variable data, and the historical quality data includes feedback variable data.

[0007] The first processing module is used to drive a preset reduced-order fluid physics simulation model based on the boundary condition data and the disturbance variable data to generate virtual flow field characteristic data that reflects the thermal-fluid coupling state inside the furnace.

[0008] The second processing module is used to perform spatiotemporal feature fusion of the virtual flow field feature data and the boundary condition data to construct a multidimensional feature vector, and input the multidimensional feature vector into a preset time-series deep learning model to calculate the probability of finished product defects occurring within a preset time period in the future.

[0009] The third processing module is configured to compare the probability of the finished product defect with a preset safety threshold; if the probability of the finished product defect is greater than the safety threshold, then based on a preset reverse optimization algorithm, a recommended combination of operating parameters for suppressing the occurrence of defects is determined.

[0010] The status output module is used to output a visualized image of the virtual flow field characteristics, a thermal distribution map of the probability of the finished product defects, and the recommended combination of operating parameters determined by the third processing module.

[0011] Preferably, the boundary condition data acquired by the data acquisition module includes fuel flow rate, combustion air pressure, traction speed, and set temperature for each zone;

[0012] The disturbance variable data includes raw material composition analysis data and environmental temperature and humidity data;

[0013] The feedback variable data includes defect coordinates, defect classifications, and defect image data collected by the cold-end online inspection equipment.

[0014] Preferably, the first processing module drives a preset reduced-order fluid physics simulation model based on the boundary condition data and the disturbance variable data, including:

[0015] The boundary condition data is used as the model input boundary, and the disturbance variable data is used as the model initial condition correction term.

[0016] Using the reduced-order fluid physics simulation model constructed based on the intrinsic orthogonal decomposition algorithm, the temperature field distribution and glass melt flow trajectory inside the melting furnace are rapidly iteratively calculated;

[0017] Extract the recirculation zone location parameters and maximum convection velocity parameters from the calculation results to generate the virtual flow field characteristic data.

[0018] Preferably, the second processing module performs spatiotemporal feature fusion of the virtual flow field feature data and the boundary condition data, including:

[0019] The virtual flow field feature data is used as a hidden layer feature to characterize the internal physical state;

[0020] The boundary condition data is used as an explicit feature characterizing the external operation state;

[0021] According to a preset time lag window, the hidden layer features and the explicit layer features are aligned and spliced ​​to generate the multidimensional feature vector containing physical mechanism information and operation timing information.

[0022] Preferably, the second processing module inputs the multidimensional feature vector into a preset time-series deep learning model to calculate the probability of finished product defects occurring within a preset time period, including:

[0023] The temporal deep learning model is constructed using a long short-term memory network or a Transformer architecture.

[0024] The multidimensional feature vector is input into the time-series deep learning model to map the nonlinear influence of the current furnace flow field state on the quality of the finished glass product in the future.

[0025] Output the spatial distribution probability matrix for the defect types of bubbles, stones, and stripes, as the probability of the occurrence of defects in the finished product.

[0026] Preferably, in response to the probability of the finished product defect exceeding a preset safety threshold, the third processing module determines a recommended combination of operating parameters for suppressing defect occurrence based on a preset reverse optimization algorithm, including:

[0027] Identify the defect type with the highest probability value among the defects in the finished product and its corresponding spatial location;

[0028] The objective function is to reduce the probability of defects at the spatial location, and the adjustable range of the boundary condition data is used as the constraint.

[0029] The recommended combination of operating parameters is generated by iteratively optimizing the coupling space between the reduced-order fluid physics simulation model and the temporal deep learning model using a gradient descent search strategy.

[0030] Preferably, the system further includes a model correction module, used for:

[0031] Obtain the current feedback variable data, and calculate the numerical difference or loss function value between the actual defect status in the feedback variable data and the historically predicted probability of the occurrence of the finished product defect, and use the calculation result as the prediction deviation value.

[0032] If the prediction deviation value is greater than the preset calibration threshold, the weight parameters of the time series deep learning model are updated online using the current feedback variable data.

[0033] Preferably, the process of constructing the reduced-order fluid physics simulation model includes:

[0034] Acquire snapshot data of the full-order flow field generated by high-fidelity computational fluid dynamics simulation;

[0035] Modal decomposition is performed on the full-order flow field snapshot data to extract basis function modes containing a preset energy percentage;

[0036] The full-order flow field is projected into a low-dimensional subspace composed of the basis function modes to construct the reduced-order fluid physics simulation model.

[0037] Preferably, the status output module outputs a visualized image of the virtual flow field characteristics, including:

[0038] The virtual flow field feature data is mapped into a three-dimensional color field and streamline diagram;

[0039] The temperature gradient distribution and glass molten convection path inside the furnace are rendered in real time in three-dimensional space.

[0040] Compared with the prior art, the present invention has the following improvements and advantages:

[0041] 1. This invention utilizes a reduced-order fluid physics simulation model through a first processing module to reconstruct the invisible high-temperature fluid state inside the melting furnace within a second-level timescale based solely on boundary conditions. This effectively fills the blind spot where traditional sensors cannot penetrate deep into the liquid glass for measurement. This method not only significantly reduces the computational cost of full-order fluid dynamics simulation but also achieves real-time visibility into the black-box production environment by extracting key virtual flow field features such as the location of the reflux zone and the maximum convection velocity. This solves the technical problem that existing technologies cannot directly observe the thermal-fluid coupling state.

[0042] 2. This invention uses a second processing module to fuse the virtual flow field features representing the internal physical state with the boundary condition data representing the external operational state in a spatiotemporal manner, constructing a hybrid feature vector containing physical mechanism information and operational timing information. Combined with a temporal deep learning model, the system can uncover the deep nonlinear mapping relationship between operation, flow field, and quality, effectively overcoming the time lag of several hours from material feeding to forming in glass production, and achieving accurate prediction of the probability of future product defects, thereby avoiding the large amount of finished product loss caused by traditional lag inspection.

[0043] 3. When a high-risk defect is predicted, the present invention uses a third processing module to trigger a gradient descent-based reverse optimization algorithm, which can automatically search for the optimal control strategy in the coupled space of the simulation model and the deep learning model. Compared with the blindness of traditional methods that rely on human experience for post-event adjustments, the present invention can output a combination of recommended operating parameters that accurately suppress defects at specific locations and meet equipment constraints, thereby realizing refined closed-loop control of process parameters and significantly improving the intelligence level and yield of the production line.

[0044] 4. This invention not only incorporates fluctuations in raw material composition and disturbances such as ambient temperature and humidity into the monitoring scope to capture the continuous impact of uncontrollable external factors on the flow field, but also designs an online model correction module. The system can update the weight parameters of the deep learning model in real time based on the deviation between the actual detected feedback variable data and the prediction results, and even trigger the reconstruction of the physical model when necessary. This mechanism gives the system the ability to learn throughout its life, enabling it to adapt to the drift of physical properties caused by the erosion and aging of the refractory materials in the melting furnace, ensuring that the system maintains high-precision prediction and control capabilities throughout its entire life cycle. Attached Figure Description

[0045] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0046] Figure 1 This is a structural diagram of the system of the present invention; Detailed Implementation

[0047] The purpose of this invention is to provide a virtual simulation and status prediction system for glass production line operation data to solve the problems mentioned in the background art. Specifically, the technical solution of this invention includes:

[0048] Example 1:

[0049] Please see Figure 1 The virtual simulation and state prediction system for glass production line operation data includes: a data acquisition module, used to acquire real-time operation data and historical quality data of the glass production line. The real-time operation data includes boundary condition data and disturbance variable data, and the historical quality data includes feedback variable data.

[0050] The first processing module is used to drive a preset reduced-order fluid physics simulation model based on boundary condition data and disturbance variable data to generate virtual flow field characteristic data that reflects the thermal-fluid coupling state inside the furnace.

[0051] The second processing module is used to perform spatiotemporal feature fusion of virtual flow field feature data and boundary condition data, construct a multi-dimensional feature vector, and input the multi-dimensional feature vector into a preset time-series deep learning model to calculate the probability of finished product defects occurring within a preset time period in the future.

[0052] The third processing module is configured to compare the probability of finished product defects with a preset safety threshold; if the probability of finished product defects is greater than the safety threshold, a recommended combination of operating parameters for suppressing defects is determined based on a preset reverse optimization algorithm; the status output module is used to output a visualized image of the virtual flow field characteristics, a thermal distribution map of the probability of finished product defects, and the recommended combination of operating parameters determined by the third processing module.

[0053] This embodiment elaborates on the overall architecture and operation logic of the system. Its core lies in constructing a closed-loop control flow driven by physical mechanisms and data. The system establishes a digital connection with the physical production line through a data acquisition module, which acts as the system's sensory nerve and captures discrete signals from the underlying control system in real time.

[0054] The first processing module performs physical dimensionality reduction calculations, using a reduced-order fluid physics simulation model (ROM) to reconstruct the invisible high-temperature fluid state inside the melting furnace within a second-level timescale, filling the blind spot where traditional sensors cannot penetrate deep into the liquid glass for measurement. The second processing module deeply integrates the above physical state characteristics with the operation time series data, using a deep learning model to mine the nonlinear mapping relationship between operation, flow field, and quality, achieving advanced prediction of the finished product quality in the next 8 to 12 hours. Based on this, the third processing module responds to high-risk warning signals and automatically triggers a reverse optimization mechanism to calculate the optimal control strategy.

[0055] The status output module transforms abstract data streams into intuitive 3D visualizations to assist process engineers in decision-making.

[0056] The virtual simulation and state prediction system constructed in this embodiment, in the high-temperature, high-hysteresis, and black-box glass production scenario, uses a physical model to see through the internal flow field and an AI model to predict future quality. It not only solves the problem of finished product loss caused by traditional hysteresis inspection, but also achieves refined closed-loop control of process parameters through reverse optimization, significantly improving the intelligence level and yield of the production line.

[0057] Example 2:

[0058] The boundary condition data acquired by the data acquisition module includes fuel flow rate, combustion air pressure, drawing speed, and set temperature for each zone; the disturbance variable data includes raw material composition analysis data and ambient temperature and humidity data; the feedback variable data includes defect coordinates, defect classification, and defect image data acquired by the cold-end online detection equipment.

[0059] This embodiment defines the input dimensions of the data acquisition module in a refined manner to ensure the completeness of the inputs to the simulation model and the prediction model; the system reads boundary condition data from the DCS system via the OPC protocol. fuel flow The mass flow meters from each small furnace are physically represented by the heat power source input to the melting furnace, and the unit is... ;

[0060] Combustion wind pressure The pressure transmitter, originating from the bottom of the regenerator, physically represents the dynamic parameters controlling flame rigidity and atmosphere, measured in units of... Pulling speed Derived from the main drive encoder of the tin bath, its physical meaning is the kinematic parameter that determines the residence time of the molten glass, and the unit is... Specify the boundary condition data here. The vectorization of the definition is set. for 3D column vector, specifically in the form of ,in, A positive integer, representing the total number of independent temperature zones divided in the melting furnace, where each component corresponds to the aforementioned physical parameters, and Represents the total number of boundary condition parameters;

[0061] To facilitate numerical calculations in the subsequent model correction module, the system further incorporates feedback variable data. Defined as a structured vector, it contains normalized defect coordinates, classification confidence, and image feature vectors, i.e. ;

[0062] The system connects to the laboratory LIMS system to acquire disturbance variable data. Pay close attention to oxides in raw material components The fluctuation, in its physical sense, refers to the change in the high-temperature viscosity of the molten glass. The internal variables are expressed as percentages; the system synchronously collects feedback variable data from cold-end AOI equipment. Construct with spatial coordinates The defect dataset;

[0063] This embodiment effectively captures the nonlinear disturbances of the furnace flow field caused by uncontrollable external factors by including raw material composition fluctuations and ambient temperature and humidity in the monitoring range. In particular, by using cold end defect data as a monitoring signal, the system can establish a full-chain causal mapping from microscopic raw material changes to macroscopic flow field drift to final product quality, which significantly improves the model's generalization ability and robustness under complex working conditions.

[0064] Example 3:

[0065] The first processing module, based on boundary condition data and disturbance variable data, drives a pre-defined reduced-order fluid physics simulation model, including:

[0066] Boundary condition data is used as the model input boundary, and disturbance variable data is used as the model initial condition correction term;

[0067] Using a reduced-order fluid physics simulation model based on the intrinsic orthogonal decomposition algorithm, the temperature field distribution and glass melt flow trajectory inside the melting furnace are rapidly calculated iteratively.

[0068] Extract the location parameters of the recirculation zone and the maximum convection velocity parameters from the calculation results to generate virtual flow field characteristic data.

[0069] This embodiment details the entire lifecycle of a reduced-order fluid physics simulation model, covering two stages: offline construction and online execution. In the offline construction stage, a high-fidelity CFD simulation is run using a high-performance computing cluster to generate a full-order flow field snapshot matrix covering various operating conditions. ;

[0070] right Perform eigenorthogonal decomposition to calculate the eigenvalues ​​of the snapshot matrix. And select the smallest integer according to the energy cutoff criterion. This results in a cumulative energy percentage Thus, the previous First-order basis function modes This compresses millions of degrees of freedom in the mesh into a low-dimensional modal coefficient space;

[0071] By projecting the Navier-Stokes equations onto this lower-dimensional subspace, we can construct a specific quadratic reduced-order dynamic equation:

[0072]

[0073] in: : Modal coefficient vector, derived from POD decomposition, physically represents the weights of the main features of the flow field; : Constant vector, derived from the average flow field projection; : Linear operator matrix, representing viscous diffusion and linear convection effects; where Representing a third-order tensor with vector The second contraction operation, its first... The components are:

[0074]

[0075] Quadratic nonlinear term The Each component The specific calculation formula is strictly defined as follows:

[0076]

[0077] in, Represents the elements of a third-order tensor, index and Corresponding to the modal coefficient vectors The The and the first Each component, traversing all Perform a contraction operation on the first mode;

[0078] This ensures that all terms on the right-hand side of the equation are... Dimension, satisfying the dimension consistency constraint of vector addition; Item Specifically, it characterizes the second-order nonlinear convection effect in fluid dynamics, namely This reflects the process of energy transfer between turbulent vortices of different scales within the melting furnace; , Input matrices, corresponding to the normalized boundary control inputs. With parameter correction terms Mapping relationship for disturbance variables; Regarding the physical mapping logic, this embodiment employs an operator sensitivity linearization method: Although the description of the embodiment mentions initial condition correction, in the mathematical coding implementation, in order to capture the continuous influence of raw material composition fluctuations on the flow field evolution, this embodiment uses viscosity... Diffusion matrix drift caused by change It is approximately an externally forced term;

[0079] Specifically, At reference viscosity Expand at the location, and include the first-order perturbation term. Mapped to addition terms To solve the problem of multidimensional component vectors , dimension The dimension matching problem with the terms of the dynamic equations is defined in this embodiment. The matrix is ​​constructed as follows:

[0080] For the sensitivity matrix The specific acquisition path, in this embodiment, is obtained using the central difference method for numerical calculation: setting a small perturbation amount. Construct the perturbed operator matrix and ,but

[0081]

[0082] This method avoids the complexity of analytical differentiation of implicit operators and ensures universality under different viscosity models;

[0083] Calculate the sensitivity vector of viscosity with respect to each oxide component. ( Its dimensions are Due to the sensitivity of linear operators The dimensions are To maintain dimensional consistency, we define:

[0084]

[0085] in, The reference viscosity has the following dimensions: , making The dimensions are , and dimensionless After multiplying, we get This embodiment abandons the traditional static averaging approach and instead... Defined as the vector of known instantaneous modal coefficients at the current time step of numerical integration. Therefore, the sensitivity matrix Updated dynamically over time:

[0086]

[0087] This dynamic definition ensures that operator drift caused by viscosity changes is minimized when the flow field is in a rapidly changing non-equilibrium state. It can be precisely mapped to a state-dependent coercive term; that is... Vector and The matrix multiplication of vectors yields for Matrix; this ensures The result of the calculation is The column vectors conform to the dimensional requirements of the equations; this approach physically transforms parameter perturbations into additive source terms, while in code logic, they are uniformly loaded through a dynamic input interface, thus solving the compatibility problem between parameter variation and model form.

[0088] To verify the dimensional consistency of the above dynamic equations, this embodiment specifically defines: modal coefficients. It is a dimensionless scalar, that is , representing the weights of the normalized flow field modes; therefore, the time derivative term on the left side of the equation... The dimensions are Accordingly, the dimensions of the operators and variables on the right-hand side of the equation are defined as follows: constant vector Dimensions are ;

[0089] Linear operator matrix The characteristic dimensions are , making Dimensions are Third-order tensor The dimensions are This makes the quadratic term Dimensions are Input control vector and parameter correction vector All inputs are normalized before being used in the model and defined as dimensionless quantities. ;

[0090] Input matrix and The dimensions are defined as follows: Thus ensuring and The dimensions of the items are all By defining it as described above, the dimensions on both sides of the equation are unified, which conforms to the physical consistency constraint.

[0091] Based on this, the system enters the online operation phase, where the first processing module calls the boundary condition data in real time. Drive the above equations and solve the current modal coefficients iteratively. Furthermore, through linear reconstruction Reconstruct the overall flow field and extract key hydraulic parameters, including the location of the recirculation zone. and maximum convection velocity ;

[0092] To ensure the mathematical differentiability of the reverse optimization process, this embodiment clarifies the calculation logic of the above indicators: the maximum convection velocity parameter is set as follows:

[0093]

[0094] The specific calculation definition of the smooth approximation maximum operator is as follows:

[0095]

[0096] This function effectively avoids numerical overflow and maintains differentiability. This is the reconstructed set of velocity vectors for all grid nodes, physically representing the velocity of fluid particles, in units of... ; Reflux zone location parameters The softened centroid formula based on vorticity weight is used for calculation:

[0097]

[0098] in, Indicates the first The spatial coordinate vector of each grid node, in units of , Indicates the first The vorticity modulus at each grid node, physically representing the rotational intensity of a fluid element, is measured in units of... , For the Sigmoid function, As a smoothing factor, this embodiment explicitly defines it. The value retrieval logic is adaptive scaling: Setting:

[0099]

[0100] in, This represents the standard deviation of the vorticity modulus of all grid nodes in a historical full-order flow field snapshot; this definition ensures that the width of the activation region of the Sigmoid function matches the fluctuation scale of the flow field vorticity, avoiding the influence of... The problem of gradient vanishing due to excessive size or boundary blurring due to excessive size; To determine the vorticity threshold in the recirculation region, this embodiment explicitly defines its value selection logic to eliminate ambiguity: It is not a fixed constant, but an adaptive threshold based on the energy distribution of the snapshot set, calculated as follows:

[0101]

[0102] in Represents the dataset The Percentiles are calculated using linear interpolation; specifically, the 85th percentile of the vorticity modulus in the full-field historical snapshot data is calculated using linear interpolation. This statistical threshold ensures the adaptability of feature extraction under different Reynolds number conditions. The generated virtual flow field feature data is defined as a composite vector.

[0103]

[0104] Its total dimension is ;in, It preserves the full-space modal information of the flow field, while and As an explicit feature, it enhances the model's sensitivity to physical critical states;

[0105] This embodiment uses the POD-Galerkin projection method to clarify the dynamic evolution function containing the quadratic nonlinear term. While preserving the complex thermal convection physical mechanism inside the melting furnace, the computational efficiency has been improved by several orders of magnitude. This approach of reducing the order without compromising quality allows fluid dynamics simulations that originally required several days of computation to run in real time on the industrial site, providing deep feature inputs containing physical mechanisms for subsequent quality prediction.

[0106] Example 4:

[0107] The second processing module performs spatiotemporal feature fusion of virtual flow field feature data and boundary condition data, including: using virtual flow field feature data as hidden layer features representing the internal physical state; using boundary condition data as explicit layer features representing the external operation state; and aligning and splicing the hidden layer features and explicit layer features according to a preset time lag window to generate a multidimensional feature vector containing physical mechanism information and operation timing information.

[0108] This embodiment addresses the large time lag characteristic in the glass production process by designing a specific spatiotemporal feature fusion strategy; the system defines virtual flow field feature data. The hidden layer features have a dimension of . This characteristic reflects the thermal inertia and fluid topology inside the melting furnace, including modal coefficients. and physical parameters Define boundary condition data For explicit features, the dimension is... This reflects the control exerted from the outside;

[0109] Considering that the material needs to undergo several hours of melting and clarification from the feeding port to the forming end, the system sets a time lag window, the specific time step of which is... Based on the effective length of the melting furnace With real-time pull speed Calculated;

[0110] in, Defined as the length of the mainstream fluid path from the center of the feed inlet to the bottleneck inlet, typically a value of Lagging steps The method for determining it is as follows:

[0111]

[0112] Typical value ,in, Defined as the minimum permissible drawing speed in furnace process design or the lower limit of historical operating data, its physical meaning is a boundary parameter that determines the longest residence time of molten glass in the furnace, and its unit is 1. The minimum permissible pulling speed is used here. The aim is to determine a fixed time window length that sufficiently covers the longest dwell time under all operating conditions. In subsequent calculations, the specific lag time for each step is determined. At that time, the real-time pull speed is used. Perform dynamic calculations;

[0113] Defined as the data sampling interval or the step size for discrete model calculations, typically a value of or This is to ensure that the time window fully covers the maximum theoretical residence time of the molten glass; that is: Perform feature alignment and splicing operations to construct... Multidimensional feature vector at time step :

[0114]

[0115] in, :Multidimensional feature vector, whose specific dimensions are:

[0116]

[0117] The physical meaning is the set of all-time and spacetime states that determine the current quality of the finished product; Defined as boundary condition data vector The feature dimension is the total number of input physical parameters, such as the number of variables like fuel, pressure, and temperature mentioned above. Vector concatenation operator; The dynamically calculated lag time step, in hours. ;

[0118] This embodiment successfully solves the problem that single-moment data cannot characterize cumulative effects by constructing a hybrid feature vector that includes the current internal state and historical operation sequences. This fusion strategy effectively combines physical mechanisms with statistical data, eliminates the risk of causal breakage in pure data-driven models when facing long-delay systems, and ensures the completeness of the prediction model input.

[0119] Example 5:

[0120] The second processing module inputs multi-dimensional feature vectors into a preset temporal deep learning model to calculate the probability of finished product defects occurring within a preset time period. This includes: constructing a temporal deep learning model using a long short-term memory network or a Transformer architecture; inputting multi-dimensional feature vectors into the temporal deep learning model to map the nonlinear influence of the current furnace flow field state on the quality of the finished glass product in the future; and outputting a spatial distribution probability matrix for bubble, stone, and stripe defect types as the probability of finished product defects occurring.

[0121] This embodiment details the quality prediction process based on a deep attention mechanism; the system initializes a pre-trained Transformer model, which uses a multi-head self-attention mechanism to capture multi-dimensional feature vectors. Long-range dependencies in the data; will be built in real time. The input model encoder performs a weighted mapping between the input physical field features and the operation time sequence features using an internal model weight matrix; the decoder outputs a probability matrix of product defects occurring within a preset future time period. :

[0122]

[0123] in, The defect probability matrix, derived from the model output, physically represents the probability of a specific type of defect occurring at different plate width locations; it is dimensionless. Hidden state vectors, derived from encoder output; Model weights and biases are derived from historical data training.

[0124] The matrix The distribution probability of the three main defect types—bubbles, stones, and streaks—in the width direction of the glass plate was clearly distinguished.

[0125] This embodiment utilizes the superior sequence modeling capabilities of the Transformer architecture to accurately capture the complex butterfly effect within the melting furnace, i.e., how minute flow field distortions evolve into final quality defects. By outputting a fine-grained spatial distribution probability matrix, the system not only answers whether there are defects, but also where the defects are and what kind of defects they are, providing a high-resolution decision-making basis for subsequent precise control.

[0126] Example 6:

[0127] The third processing module, in response to the probability of finished product defects exceeding a preset safety threshold, determines a recommended combination of operating parameters to suppress defect occurrence based on a preset reverse optimization algorithm. This includes: identifying the defect type with the highest probability among finished product defects and its corresponding spatial location; using the reduction of the defect probability at the spatial location as the objective function and the adjustable range of boundary condition data as the constraint; and using a gradient descent search strategy to iteratively optimize in the coupling space of the reduced-order fluid physics simulation model and the time-series deep learning model to generate the recommended combination of operating parameters.

[0128] This embodiment illustrates the core decision-making logic of the system, namely, closed-loop control from prediction to intervention; the third processing module monitors the probability matrix in real time. The probability of responding to any element exceeds the safety threshold. Immediately identify the highest-risk defect type and its spatial coordinates; construct a reverse optimization objective function. In order to solve the boundary condition data Different physical quantities, such as flow rate ,pressure ,speed To address the issues of gradient calculation errors and lack of physical meaning in regularization caused by inconsistent dimensions, this embodiment explicitly performs the optimization process in a dimensionless normalized space; the normalization mapping is defined as follows:

[0129]

[0130] in, This represents element-wise division. For the range of values ​​within The dimensionless control vector between them; the objective function is defined in this normalized space:

[0131]

[0132] in, Defined as the target defect location, i.e., the highest-risk defect identified, within the optimized control parameters. The probability values ​​are obtained by predicting them using a time-series deep learning model.

[0133] At this point, both terms on the right-hand side of the formula are dimensionless scalars, satisfying the physical addition rule; weighting coefficients Set as , Using the L-curve method The selection is within a certain range; correspondingly, the gradient calculation formula is modified to apply to dimensionless variables. Differentiate:

[0134]

[0135] in, The sensitivity of the composite model is represented by the chain rule, where, This represents the sensitivity of the modal coefficients of the reduced-order model to the normalized input. Since the reduced-order model is solved using discrete-time step integration, according to the explicit Euler discretization formula:

[0136]

[0137] The sensitivity term is specifically calculated as follows:

[0138]

[0139] in, The time integration step size for the model, in seconds; introduced here. This ensures the consistency of gradient dimensions, enabling the gradient of the physical control quantity to be correctly backpropagated to the dimensionless modal coefficient space.

[0140] This represents the direct sensitivity of the input layer of a deep learning model to a normalized input; all terms in the above formula are dimensionless, eliminating the dimensionlessness of the original physical space. and The dimensional fallacy of addition;

[0141] Perform gradient descent search:

[0142]

[0143] To strictly implement the technical feature of using the adjustable range of boundary condition data as a constraint in the embodiments, this embodiment performs hard constraint projection in the normalized space:

[0144]

[0145] in, The truncation function ensures that all optimized parameters fall within the range of Within the effective normalization interval, it is equivalent to the physical parameters not exceeding the equipment limits;

[0146] When the objective function converges or reaches the maximum number of iterations, an inverse normalization transformation is performed:

[0147]

[0148] Output the final recommended combination of operating parameters. ,in, This is element-wise multiplication; The recommended operating parameters for the final output are derived from iterative updates and inverse normalization, and their physical meaning is the optimal control quantity for suppressing defects. Learning rate, selection range is An exponential decay strategy is employed to ensure convergence accuracy near extreme points; The gradient of the objective function with respect to the normalized control variables is obtained through automatic differentiation calculation;

[0149] The reverse optimization algorithm proposed in this embodiment utilizes the differentiability of the surrogate model to clarify the optimization objective and gradient calculation path, realizing zero-cost trial and error in virtual space. Compared with the blindness of traditional manual experience adjustment, this method can quickly find the fine operation parameters that can accurately suppress defects in specific locations, such as eliminating bubbles on the left side, and have the least impact on other areas, based on mathematical gradient guidance, which greatly reduces the risk of process adjustment.

[0150] Example 7:

[0151] The system also includes a model correction module, which is used to: acquire the current feedback variable data, and calculate the numerical difference or loss function value between the actual defect status in the feedback variable data and the historical predicted probability of finished product defects, and use the calculation result as the prediction deviation value; if the prediction deviation value is greater than the preset calibration threshold, the weight parameters of the time series deep learning model are updated online using the current feedback variable data.

[0152] This embodiment introduces an online adaptive correction mechanism to address characteristic drift throughout the entire lifecycle of the melting furnace; at the end of each production shift, the model correction module retrieves the actual defect data detected by the cold end AOI. ; Retrospectively analyze the historical predicted probabilities for the corresponding time window Using Gaussian kernel function to analyze discrete real-world defect data Mapping to AND True defect probability density map of the same dimension The prediction deviation between the two is calculated using the cross-entropy function. ;

[0153] The system executes a hierarchical correction strategy: using the current feedback variable data to adjust the weight parameters of the time-series deep learning model. Perform online backpropagation updates; if continuous After each production shift is updated, the predicted deviation value The sliding average is still greater than the preset structural drift threshold. If this occurs, it is determined that the physical boundary conditions of the melting furnace have undergone structural drift, such as refractory material erosion. At this point, the system locks the weights of the deep learning model and triggers the physical model reconstruction procedure: calling the offline CFD server to regenerate a snapshot set based on the latest kiln body scan geometry data. And update the POD base functions. and order reduction operators This allows for fundamental calibration of the physical mechanism layer;

[0154] The system executes logical judgments and responds to... Greater than the preset calibration threshold But less than the structural drift threshold The system determines that the current model has local biases; based on this, it triggers an online learning process, utilizing the current... Data pairs are used to fine-tune the weight parameters of a deep learning model using the backpropagation algorithm. ;

[0155] This embodiment endows the system with lifelong learning capabilities through this continuous prediction-verification-correction cycle. In scenarios where the physical properties of the melting furnace change slowly over time due to the gradual erosion of refractory materials and the decay of thermal insulation performance, this mechanism ensures that the prediction model can automatically follow and adapt to these changes, avoiding the common problem of model aging as soon as it goes online, and ensuring high-precision operation of the system throughout the entire kiln life cycle.

[0156] Example 8:

[0157] The status output module outputs a visualized image of the virtual flow field characteristics, including: mapping the virtual flow field characteristic data into a three-dimensional color field and streamline diagram; and rendering the temperature gradient distribution and glass melt convection path inside the melting furnace in real time in three-dimensional space.

[0158] This embodiment focuses on the intuitive presentation of complex physical data, aiming to achieve transparency in the dark; the status output module receives virtual flow field feature data from the first processing module. The modal coefficients containing the weights of the spatial basis functions are extracted from them. And utilize the POD basis functions stored in the offline stage. Perform reverse reconstruction operation , here This is specifically defined as single-frame 3D flow field data reconstructed at the current moment, distinct from historical snapshot matrices. This allows the low-dimensional data to be restored to full-order field data containing temperature scalars and velocity vectors on three-dimensional grid nodes;

[0159] Using a graphics rendering engine, the temperature scalar is mapped into a three-dimensional color field with varying warm and cool tones, which intuitively displays the temperature gradient distribution deep within the melting furnace and identifies potential stagnant zones or local hot spots. Based on the velocity vector field, a dynamic streamline diagram is generated to depict the convection path and reflux vortex structure of the molten glass at the bottom of the pool. The above three-dimensional rendering objects are superimposed and displayed on the human-machine interface (HMI) for operators to observe from all angles.

[0160] This embodiment uses advanced visualization technology to transform abstract mathematical simulation results into visual images that conform to human cognition. This not only helps process engineers intuitively understand the current melting state, such as whether the backflow at the feeding port is normal, but also provides visualized physical attribution analysis when the system issues a defect warning, such as bubbles caused by flow line disturbance, which greatly enhances the mutual trust and decision-making efficiency of human-machine collaboration.

[0161] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A virtual simulation and status prediction system for glass production line operation data, characterized in that, include: The data acquisition module is used to acquire real-time operating data and historical quality data of the glass production line. The real-time operating data includes boundary condition data and disturbance variable data, and the historical quality data includes feedback variable data. The first processing module is used to drive a preset reduced-order fluid physics simulation model based on the boundary condition data and the disturbance variable data to generate virtual flow field characteristic data that reflects the thermal-fluid coupling state inside the furnace. The second processing module is used to perform spatiotemporal feature fusion of the virtual flow field feature data and the boundary condition data to construct a multidimensional feature vector, and input the multidimensional feature vector into a preset time-series deep learning model to calculate the probability of finished product defects occurring within a preset time period in the future. The third processing module is configured to compare the probability of the finished product defect with a preset safety threshold; if the probability of the finished product defect is greater than the safety threshold, then based on a preset reverse optimization algorithm, a recommended combination of operating parameters for suppressing the occurrence of defects is determined. The status output module is used to output a visualized image of the virtual flow field characteristics, a thermal distribution map of the probability of the finished product defects, and the recommended combination of operating parameters determined by the third processing module.

2. The virtual simulation and status prediction system for glass production line operation data according to claim 1, characterized in that: The boundary condition data acquired by the data acquisition module includes fuel flow rate, combustion air pressure, traction speed, and set temperature for each zone. The disturbance variable data includes raw material composition analysis data and environmental temperature and humidity data; The feedback variable data includes defect coordinates, defect classifications, and defect image data collected by the cold-end online inspection equipment.

3. The virtual simulation and status prediction system for glass production line operation data according to claim 1, characterized in that, The first processing module, based on the boundary condition data and the disturbance variable data, drives a preset reduced-order fluid physics simulation model, including: The boundary condition data is used as the model input boundary, and the disturbance variable data is used as the model initial condition correction term. Using the reduced-order fluid physics simulation model constructed based on the intrinsic orthogonal decomposition algorithm, the temperature field distribution and glass melt flow trajectory inside the melting furnace are rapidly iteratively calculated; Extract the recirculation zone location parameters and maximum convection velocity parameters from the calculation results to generate the virtual flow field characteristic data.

4. The virtual simulation and status prediction system for glass production line operation data according to claim 3, characterized in that, The second processing module performs spatiotemporal feature fusion of the virtual flow field feature data and the boundary condition data, including: The virtual flow field feature data is used as a hidden layer feature to characterize the internal physical state; The boundary condition data is used as an explicit feature characterizing the external operation state; According to a preset time lag window, the hidden layer features and the explicit layer features are aligned and spliced ​​to generate the multidimensional feature vector containing physical mechanism information and operation timing information.

5. The virtual simulation and status prediction system for glass production line operation data according to claim 4, characterized in that, The second processing module inputs the multidimensional feature vector into a preset time-series deep learning model to calculate the probability of product defects occurring within a preset future time period, including: The temporal deep learning model is constructed using a long short-term memory network or a Transformer architecture. The multidimensional feature vector is input into the time-series deep learning model to map the nonlinear influence of the current furnace flow field state on the quality of the finished glass product in the future. Output the spatial distribution probability matrix for the defect types of bubbles, stones, and stripes, as the probability of the occurrence of defects in the finished product.

6. The virtual simulation and status prediction system for glass production line operation data according to claim 1, characterized in that, In response to the probability of the finished product defect exceeding a preset safety threshold, the third processing module determines a recommended combination of operating parameters for suppressing defect occurrence based on a preset reverse optimization algorithm, including: Identify the defect type with the highest probability value among the defects in the finished product and its corresponding spatial location; The objective function is to reduce the probability of defects at the spatial location, and the adjustable range of the boundary condition data is used as the constraint. The recommended combination of operating parameters is generated by iteratively optimizing the coupling space between the reduced-order fluid physics simulation model and the temporal deep learning model using a gradient descent search strategy.

7. The virtual simulation and status prediction system for glass production line operation data according to claim 1, characterized in that, The system also includes a model correction module, used for: Obtain the current feedback variable data, and calculate the numerical difference or loss function value between the actual defect status in the feedback variable data and the historically predicted probability of the occurrence of the finished product defect, and use the calculation result as the prediction deviation value. If the prediction deviation value is greater than the preset calibration threshold, the weight parameters of the time series deep learning model are updated online using the current feedback variable data.

8. The virtual simulation and status prediction system for glass production line operation data according to claim 3, characterized in that, The process of constructing the reduced-order fluid physics simulation model includes: Acquire snapshot data of the full-order flow field generated by high-fidelity computational fluid dynamics simulation; Modal decomposition is performed on the full-order flow field snapshot data to extract basis function modes containing a preset energy percentage; The full-order flow field is projected into a low-dimensional subspace composed of the basis function modes to construct the reduced-order fluid physics simulation model.

9. The virtual simulation and status prediction system for glass production line operation data according to claim 1, characterized in that, The status output module outputs a visualized image of the virtual flow field characteristics, including: The virtual flow field feature data is mapped into a three-dimensional color field and streamline diagram; The temperature gradient distribution and glass molten convection path inside the furnace are rendered in real time in three-dimensional space.

Images