Preform rod processing process monitoring method and system combined with big data

Abstract

The present application belongs to the technical field of preform processing process monitoring method, especially to the preform processing process monitoring method and system combined with big data, a preform processing multi-physical field coupling mechanism model is constructed, the flow field, temperature field and chemical reaction kinetics characteristics are quantified based on the related basic equations of the multi-physical field, the mapping relationship between the key physical parameters and the quality index is established, and the mechanism parameter dictionary and the constraint rule library are formed; the original data of the sensor is preprocessed, the mechanism derived characteristics are deduced based on the multi-physical field coupling mechanism model, the traditional statistical characteristics and the mechanism derived characteristics are fused, and a multi-dimensional feature matrix is constructed; a mechanism-constrained data-driven fusion architecture is built, the theoretical parameter interval output by the mechanism model is taken as the regularization constraint term of the data-driven model, the multi-scene mechanism knowledge is integrated, and the virtual and real data fusion training is realized combined with the digital twin.

Description

Technical Field

[0001] This invention belongs to the technical field of preform processing monitoring methods, and particularly relates to a preform processing monitoring method and system that combines big data. Background Technology

[0002] As a core component in optical fiber manufacturing, the preform involves multiple critical processes, including deposition, sintering, and drawing. The flow fields, temperature fields, and chemical reaction fields within each process are strongly coupled. Even minor fluctuations in process parameters and differences in raw material batches directly affect core quality indicators such as preform diameter, refractive index distribution, and defect density. Therefore, precise and real-time monitoring of the processing is crucial to ensuring the stability of preform production quality. Currently, traditional monitoring methods in the industry rely heavily on manual inspections and parameter acquisition from single sensors. These methods can only achieve real-time monitoring of local process parameters, failing to quantify the inherent evolutionary laws of multi-physics coupling and establish a quantitative mapping relationship between key physical parameters and quality indicators. Furthermore, they exhibit significant lag in sensing process drift, making it difficult to meet the demands of large-scale production of high-precision, highly consistent preforms.

[0003] In recent years, although some technologies have attempted to apply mechanistic modeling or data-driven methods to the monitoring of preform manufacturing processes, existing technologies still have significant shortcomings and are difficult to adapt to actual production needs. While single-mechanistic modeling methods can reflect the physical essence of the manufacturing process, they significantly simplify complex production conditions during modeling, lack on-site data support for model parameter calibration, and have poor dynamic adaptability to process parameter adjustments and raw material batch changes. Pure data-driven models, on the other hand, rely excessively on a large number of labeled samples, easily becoming overfitted due to detachment from physical mechanisms. They lack the ability to perceive weak abnormal signals corresponding to early faults and minor defects, and cannot effectively distinguish between genuine process anomalies and false anomalies caused by environmental interference. Furthermore, existing technologies have not achieved effective integration of mechanistic constraints and digital twin data, and lack a feature system and dynamic threshold optimization mechanism for all types of anomalies. They cannot form a closed loop of mechanistic modeling, data training, on-site verification, and iterative optimization, making it difficult to achieve intelligent and high-precision monitoring of the entire preform manufacturing process. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for monitoring the preform manufacturing process using big data, the method comprising: A multi-physics coupling mechanism model for preform processing is constructed. Based on the relevant basic equations of multi-physics, the dynamic characteristics of flow field, temperature field and chemical reaction are quantified. The mapping relationship between key physical parameters and quality indicators is established, and a mechanism parameter dictionary and constraint rule base are formed. The raw sensor data is preprocessed, and mechanism-derived features are derived based on the multi-physics coupling mechanism model. Traditional statistical features and mechanism-derived features are then integrated to construct a multi-dimensional feature matrix. A data-driven fusion architecture constrained by mechanism is built, and the theoretical parameter range output by the mechanism model is used as the regularization constraint term of the data-driven model. Multi-scenario mechanism knowledge is integrated, and virtual and real data fusion training is achieved by combining digital twins. By capturing process drift signals through a sliding time window and combining the results of sensitivity analysis of the mechanism model, the data-driven model parameters and mechanism constraint strength are dynamically optimized to adapt to scenarios of process parameter adjustment and raw material batch change. Construct a multi-dimensional anomaly feature system covering all types of anomalies, and use signal frequency band decomposition and feature enhancement technology to enhance the ability to perceive weak signals corresponding to early faults and minor defects; A multimodal anomaly classification model with an embedded causal inference module is constructed. The model is trained using learning methods and loss functions to distinguish between true anomalies and false anomalies caused by environmental interference. An adaptive threshold dynamic optimization mechanism based on mechanistic constraints and real-time data is established, and the threshold is set and iteratively optimized according to the anomaly type. Complete the on-site deployment and generalization capability verification of the model, optimize the technical solution based on real-world application cases, construct a closed-loop system for mechanism modeling, data training, on-site verification, and iterative optimization, and continuously update the sample library and model parameters.

[0005] Furthermore, embodiments of the present invention also provide a preform processing monitoring system incorporating big data, characterized in that it includes: A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to perform the aforementioned method for monitoring the preform processing procedure using big data by executing the machine-executable instructions.

[0006] In another aspect, embodiments of the present invention also provide a computer program product, the computer program product including machine-executable instructions, the machine-executable instructions being stored in a computer-readable storage medium, a processor of a computer device reading the machine-executable instructions from the computer-readable storage medium, the processor executing the machine-executable instructions, causing the computer device to execute the above-described method for monitoring the preform processing of big data.

[0007] Based on the above, by constructing a multi-physics coupling mechanism model, the co-evolution laws of flow field, temperature field, and chemical reaction field are accurately quantified, and a quantitative mapping relationship between key physical parameters and quality indicators is established. Combined with multi-dimensional data preprocessing technology, the problems of insufficient data integrity and inaccurate drift correction in traditional monitoring are effectively solved, ensuring the reliability and effectiveness of input data. At the same time, an innovative data-driven fusion architecture with mechanism constraints is adopted, transforming mechanism rules into regularized constraint terms and integrating multi-scenario knowledge graphs and digital twin virtual and real data fusion training. This avoids the overfitting and physical decoupling problems of pure data-driven models and makes up for the poor adaptability of single mechanism models. It significantly improves the model's cross-scenario generalization ability in scenarios such as process parameter adjustment and raw material batch change, achieving a dual guarantee of physical mechanism conformity and data fitting accuracy.

[0008] This invention constructs a multi-dimensional feature system covering all types of anomalies and an anomaly classification model embedding a causal inference module. It enhances the perception of weak signals from early faults and minor defects through signal frequency band decomposition and feature enhancement techniques. Combined with causal contribution analysis, it accurately distinguishes between genuine anomalies and false anomalies caused by environmental interference, effectively reducing false alarm and false negative rates. Simultaneously, based on an adaptive threshold dynamic optimization mechanism using mechanistic constraints and real-time data, thresholds are set and iteratively optimized according to anomaly type. Coupled with a closed-loop system of mechanistic modeling, data training, on-site verification, and iterative optimization, it achieves intelligent and high-precision monitoring of the entire preform processing flow. This not only provides early warning of potential quality risks and ensures the consistency of product diameter and refractive index distribution indicators, but also continuously improves the stability and control efficiency of the production process, providing strong support for the large-scale, high-quality production of preforms. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the execution flow of the preform processing monitoring method combining big data provided in an embodiment of the present invention.

[0010] Figure 2 This is a schematic diagram of exemplary hardware and software components of the preform processing monitoring system that combines big data, provided in an embodiment of the present invention. Detailed Implementation

[0011] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating a method for monitoring the preform processing of a preform using big data, provided in one embodiment of the present invention. The following is a detailed description of this method for monitoring the preform processing of a preform using big data.

[0012] Step S110: Construct a multi-physics coupling mechanism model for preform processing, quantify the flow field, temperature field and chemical reaction kinetics based on the relevant basic equations of multi-physics, establish the mapping relationship between key physical parameters and quality indicators, and form a mechanism parameter dictionary and constraint rule base; The fundamental equations related to the flow field, temperature field, and chemical reaction field in the preform processing were retrieved. The Navier-Stokes equations were used for the flow field, the Fourier heat conduction equation for the temperature field, and the Arrhenius equation for the chemical reaction field. Key physical parameters such as vortex intensity, temperature gradient, and chemical reaction rate were quantified. Preform diameter, refractive index distribution, and defect density were selected as core quality indicators. Fifty typical working conditions were designed through orthogonal experiments, and physical parameters and quality inspection data were collected under each condition. A quantitative mapping model between physical parameters and quality indicators was established using multiple linear regression combined with mechanistic constraint correction, with a mapping error ≤3%. The parameters were categorized into physical field parameters, quality indicator parameters, and model parameters. Parameter definitions, units, value ranges, and detection methods were compiled. A mechanistic parameter dictionary was constructed in JSON format. Based on thermodynamic laws and process specifications, physical constraints, process constraints, and threshold constraints were extracted. A constraint rule base was written using logical expressions, with rule priority ordered as physical constraints > process constraints > threshold constraints.

[0013] Step S111: Clarify the boundary conditions and basic parameters of the preform deposition, sintering, and wire drawing processes; sort out the equipment structure parameters, raw material property parameters, and process operation parameters; and define the quantitative detection standards and accuracy requirements for the preform diameter, refractive index distribution, and defect density quality indicators. The inner wall of the reaction chamber in the deposition process is set as an adiabatic boundary, the heating zone in the sintering process is set as a constant temperature boundary, and the die outlet in the wire drawing process is set as a pressure boundary. Basic parameters include equipment structural parameters, raw material property parameters, and process operating parameters. Quality indicators are defined and quantified according to the following standards: diameter is measured using a laser diameter gauge with an accuracy of ±0.02mm; refractive index distribution is measured using a fiber optic interferometer with a spatial resolution of 10μm and an accuracy of 10... -6 Defect density was observed using a scanning electron microscope, and the number of defects in a 100μm×100μm area was counted, with a detection accuracy of 1 defect / μm².

[0014] Step S112: Establish a basic physical model for each field. Simplify the flow field control equations based on the Navier-Stokes equations and the geometric characteristics of the reaction chamber. Introduce a turbulence model and set the flow field boundary conditions. Discretize the equations and initialize the calculation parameters using the finite volume method. Integrate the radiation heat transfer model and the convective heat transfer coefficient calculation model based on the Fourier heat conduction equation to clarify the temperature field boundary conditions and quantify the spatiotemporal evolution of the temperature field. Based on the Arrhenius equation and the deposition / sintering reaction, determine the kinetic parameters and establish a quantitative relationship between the reaction rate and temperature and reactant concentration. Based on the simplified Navier-Stokes equations according to the geometric characteristics of the reaction chamber, neglecting the volume force term, a k-ε turbulence model is introduced. Inlet velocity and outlet pressure boundaries are defined, and the equations are discretized using the finite volume method. A second-order upwind scheme is used for solution. Initial pressure is 101325 Pa and temperature is 300 K. The convergence criterion is residual ≤10. -6 Temperature field modeling: Based on the Fourier anisotropic heat conduction equation, the Rosseland approximate radiative heat transfer model is incorporated, coupled with the Dittus-Boelter convective heat transfer coefficient calculation model. The heating element is set as the heat flux density boundary, and the external wall heat dissipation coefficient is 8-12 W / (m²·K). Using the flow field velocity distribution as input, the global temperature spatiotemporal distribution is numerically solved, with a temperature gradient calculation accuracy of ±0.1℃ / mm. Chemical reaction field modeling: The core reactions of silane oxidation and hydroxyl elimination are analyzed. Based on the Arrhenius equation, a reaction rate equation is constructed. The activation energy is experimentally calibrated to 80-120 kJ / mol, and the frequency factor is 10^6-10^8. A quantitative relationship between the reaction rate and temperature and reactant concentration is established, coupled with the component transport equation to achieve dynamic calculation.

[0015] Step S1121: Extract the geometric features of the actual structure of the reaction chamber, remove non-critical structures and retain the chamber, inlet nozzle, outlet and preformed rod base components, construct a simplified three-dimensional geometric model, divide the calculation grid according to the flow field change gradient and refine the grid in the flow field change region at the nozzle outlet and the preformed rod surface, establish the flow field control equation based on the incompressible fluid Navier-Stokes equation, ignore the secondary volume force terms and perform dimensionless simplification in combination with the gas phase flow condition of the reaction chamber, introduce the k-ε turbulence model closed equation set and define the turbulent kinetic energy and dissipation rate transport equation, match the preformed rod processing condition to calibrate the empirical coefficient of the turbulence model, set the flow field boundary conditions, adopt the velocity inlet boundary for the inlet and input the process gas velocity and component ratio, adopt the pressure outlet boundary for the outlet and match the process negative pressure of the chamber, set the inner wall of the reaction chamber and the surface of the preformed rod as a non-slip solid wall boundary and give the wall roughness parameter, use the finite volume method to spatially discretize the simplified flow field control equation, select the discretization format, initialize the flow field velocity, pressure and temperature calculation parameters and set the initial iteration convergence conditions; The actual geometric features of the reaction chamber were extracted, and non-critical structures such as chamfers and fillets were removed. The chamber, inlet nozzle, outlet, and preform substrate were retained, and a simplified three-dimensional geometric model was constructed using ANSYS ICEM. The computational mesh was divided according to the flow field variation gradient, with 500,000 to 2 million mesh elements worldwide. The mesh density was increased to three times the original density in regions with drastic flow field changes, such as the nozzle outlet and the preform surface. The flow field control equations were established based on the Navier-Stokes equations for incompressible fluids. The gravity term was neglected for dimensionless simplification, and a closed set of equations from the k-ε turbulence model was introduced. The turbulent kinetic energy and dissipation rate transport equations were defined, and empirical coefficients such as Cμ=0.09 and C1=1.44 were calibrated to match the preform processing conditions. The inlet is set as the velocity inlet boundary, with a process gas velocity of 0.5-2 m / s and a silane to oxygen ratio of 1:1.2-1:1.5. The outlet is set as the pressure outlet boundary, matched with a process negative pressure of -0.01 to -0.03 MPa. The inner wall of the reaction chamber and the surface of the preform are set as non-slip solid walls with a surface roughness of 0.8-1.2 μm. The finite volume method is used to spatially discretize the governing equations. A second-order upwind scheme is selected, and the initial flow field velocity is 0.5 m / s, pressure is 101325 Pa, and temperature is 300 K. The initial iteration convergence criterion is set as residual ≤ 10. -6 .

[0016] Step S1122: Based on the Fourier anisotropic heat conduction equation, a temperature field control equation is established. For high-temperature processing conditions, the Rosseland approximate radiation heat transfer model is incorporated, and a flow field-driven convective heat transfer coefficient model is coupled. The local flow velocity and fluid property parameters are substituted into the real-time solution of the heat transfer coefficient to form a temperature field control equation that couples conduction, convection, and radiation. Temperature field boundary conditions are set. The heating element region is set as a constant heat flux density or temperature boundary to match the heating power heat source input. The outer wall of the reaction chamber is set as a natural convection heat dissipation boundary and given the ambient temperature and comprehensive heat dissipation coefficient. The interface between the preform and the cavity is set as a coupled boundary with continuous temperature and heat flux density. The velocity distribution solved by the flow field is used as the temperature field input term. The influence of convective heat transfer is coupled, and the temperature field control equation is numerically solved to obtain the spatiotemporal temperature distribution of the entire reaction chamber and preform, and output the temperature value and temperature gradient at different times at each location. The temperature field control equation is established based on the Fourier anisotropic heat conduction equation. For high-temperature (300-1500K) processing conditions, the Rosseland approximate radiation heat transfer model is incorporated, with the radiation coefficient taken as 0.85-0.95. A flow field-driven convective heat transfer coefficient model is coupled in. The local flow velocity (0.5-2m / s) and fluid properties (thermal conductivity 0.02-0.05W / (m·K), specific heat capacity 1000-1200J / (kg·K)) are substituted into the Dittus-Boelter formula to solve for the convective heat transfer coefficient of 50-200W / (m²·K) in real time, forming a conduction-convection-radiation coupled temperature field control equation. Temperature field boundary conditions are set as follows: the heating element region is set as a constant heat flux density boundary of 100-300 kW / m², matched with a 5-15 kW heating power heat source input; the outer wall of the reaction chamber is set as a natural convection heat dissipation boundary, with an ambient temperature of 25-30℃ and a comprehensive heat dissipation coefficient of 8-12 W / (m²・K); the interface between the preform and the cavity is set as a coupled boundary with continuous temperature and heat flux density. The velocity distribution obtained from the flow field solution is used as the temperature field input term, and the SIMPLE algorithm is used to numerically solve the temperature field control equation with a time step of 0.01 s to obtain the spatiotemporal temperature distribution of the entire reaction chamber and preform, and output the temperature values ​​(accuracy ±0.5℃) and temperature gradient (accuracy ±0.1℃ / mm) at different times at each location.

[0017] Step S1123: Analyze the core gas-phase reactions in the preform deposition and sintering process, determine the reaction equations and the compositional concentration variation patterns of reactants and products; construct a chemical reaction rate equation based on the Arrhenius equation, determine the reaction order, obtain key kinetic parameters such as reaction activation energy and frequency factor through a combination of literature review and experimental calibration, and correct parameter deviations under high-temperature conditions; substitute the real-time temperature of the temperature field and the reactant concentration transported by the flow field as input variables into the Arrhenius equation to establish a quantitative relationship between reaction rate and temperature and reactant concentration; simultaneously, combine the dynamic changes in reactant consumption and product formation to construct a computational relationship coupling component transport and reaction kinetics.

[0018] The core gas-phase reactions in the preform deposition and sintering processes were analyzed. The core reaction in the deposition stage was identified as SiH4 + O2 → SiO2 + 2H2O, and the core reaction in the sintering stage was identified as Si(OH)4 → SiO2 + 2H2O. The concentration changes of reactants and products in each reaction were determined, showing a 1:1 stoichiometric ratio between the reactant consumption rate and the product formation rate. Based on the Arrhenius equation, a chemical reaction rate equation v = Aexp(-E) was constructed. a / (RT)), determining that the deposition reaction is a second-order reaction and the sintering reaction is a first-order reaction. Preliminary kinetic parameters were obtained through literature review and corrected by experimental calibration: the activation energy of the deposition reaction, E...a =90±10kJ / mol, frequency factor A=1.2×10 7 L / (mol・s); Activation energy of sintering reaction E a =110±15kJ / mol, frequency factor A=8.5×10 6 s -1 Correct parameter deviations to ≤5% under high-temperature conditions (>1200K). This includes real-time temperature of the temperature field (300-1500K) and reactant concentrations transported in the flow field (10...). -4 -10 -2 Substituting the concentration (mol / L) as an input variable into the Arrhenius equation, a quantitative relationship between reaction rate and temperature and reactant concentration is established. Combined with the component transport equation, the dynamic changes in reactant consumption and product formation are calculated in real time, constructing a computational relationship coupling component transport and reaction kinetics to ensure that the reaction rate calculation error is ≤3%.

[0019] Step S113: Construct a multi-field coupled solution system, determine the coupling order of the flow field, temperature field, and chemical reaction field, define multi-field coupling terms, use a partitioned iterative algorithm to solve the coupling equations and set the iterative convergence criterion, and optimize the calculation grid density by verifying grid independence; In this embodiment, a multi-field coupled solution system of flow field, temperature field, and chemical reaction field is constructed. The execution order of the three-field coupling is determined, inter-field coupling correlation terms are defined, and the mathematical expression and computational logic of each coupling term are clarified. The solution subdomain is divided according to the functional region of the reaction chamber, and the control equations of each field are solved region by region using a partitioned iterative algorithm. Parameter iterative transfer is achieved through the region boundaries. A dual convergence criterion of physical parameters and numerical residuals is set, and real-time monitoring is performed during the iteration process. The calculation terminates when either criterion is met. Multiple sets of computational grids with gradient distribution are generated, and grid independence verification is carried out. The relative error of key physical parameters is used as the verification standard. When the error is less than 3%, the optimal grid density is determined. While ensuring the accuracy of the multi-field coupled solution, the computational efficiency is improved, and the solution process is ensured to conform to the physical transfer law of preform processing.

[0020] Step S1131: Determine the closed-loop coupling order of the flow field, temperature field, and chemical reaction field. First, solve the flow field distribution. Then, solve the temperature field with the flow field velocity field as input. Calculate the chemical reaction rate based on the temperature field results. Finally, feed the reaction heat generated by the chemical reaction back to the temperature field for correction, forming a closed-loop coupling logic of the flow field, temperature field, chemical reaction field, and temperature field. In this embodiment, a closed-loop coupling logic is established for the flow field, temperature field, and chemical reaction field. First, the gas phase velocity distribution within the reaction chamber is solved based on the flow field control equations. The velocity field is then substituted into the temperature field equations as input to solve for the global temperature spatiotemporal distribution. The chemical reaction rate and heat release are calculated using the temperature field data as variables. The heat release is then fed back to the temperature field equations as an internal heat source, iteratively correcting the temperature field. This forms a closed-loop solution sequence: flow field → temperature field → chemical reaction field → temperature field correction. This sequence follows the multi-physics field co-evolution mechanism of preform fabrication, transferring parameters through field-by-field iterations to eliminate the mechanistic bias of independent single-field solutions, ensuring the physical realism and numerical stability of the coupled calculations.

[0021] Step S1132: Define multi-field coupling terms, where the coupling term between the flow field and the temperature field is the convective transport term of the flow field on temperature, the coupling term between the temperature field and the chemical reaction field is the influence term of temperature on reaction rate, the coupling term between the chemical reaction field and the temperature field is the source term of reaction heat on temperature field, and the coupling term between the flow field and the chemical reaction field is the transport term of the flow field on reactants / products. Clarify the mathematical expression and calculation logic of each coupling term. In this embodiment, four sets of inter-field coupling terms are defined: the flow field-temperature field coupling term is a velocity-driven temperature convection and transport term, characterized by the flow velocity and temperature gradient to represent the heat transfer intensity; the temperature field-chemical reaction field coupling term is an exponential effect of temperature on the reaction rate, constructed based on the Arrhenius equation; the chemical reaction field-temperature field coupling term is a reaction heat source term, included as an internal heat source in the temperature field equation; and the flow field-chemical reaction field coupling term is a reactant / product flow field transport term. The discretization format and computational logic of each coupling term are clearly defined, and they are embedded into the corresponding field's governing equations to achieve mutual constraint and transfer of the three field parameters, ensuring that the coupled solution fully conforms to the physical mechanism of preform processing.

[0022] Step S1133: Divide the solution region into the gas inlet zone, reaction zone, gas outlet zone and preformed rod matrix zone of the reaction chamber, and use the partition iterative algorithm to solve the field control equations of each field zone by zone. After the solution of each zone is completed, the calculation result of the zone is used as the input parameter of the adjacent zone and iteratively passed to achieve the coordinated solution of the field parameters of the whole flow domain. In this embodiment, the solution domain is divided into an inlet region, a reaction region, an outlet region, and a preformed rod matrix region, with the boundaries of each region serving as data transfer interfaces. A partitioned iterative algorithm is employed: first, the flow field and temperature field in the inlet region are solved, and the results are used as the initial boundary conditions for the reaction region. After completing the three-field coupled solution of the reaction region, the parameters are sequentially transferred to the outlet region and the preformed rod matrix region for calculation. After each round of global iteration, the boundary parameters of each region are updated as input for the next round, and the iteration continues until the global parameters stabilize. This partitioned solution method reduces the complexity of synchronous global computation, improves solution efficiency, and ensures continuous parameter transfer between regions, achieving accurate collaborative solution of multi-field parameters across the entire flow domain.

[0023] Step S1134: Set the iterative convergence criteria, which include physical parameter convergence criteria and residual convergence criteria. The physical parameter convergence criteria are that the relative changes of key parameters such as flow field velocity, temperature, and reaction rate in adjacent iteration steps are all less than 10. -5 The residual convergence criterion is that the numerical residuals of each field control equation decrease to 10. -6 The iterative calculation for the current working condition will stop once any convergence criterion is met. In this embodiment, a dual-iteration convergence criterion is set to ensure solution accuracy. The physical parameter convergence criterion is that the relative change of key parameters such as flow field velocity, temperature, and reaction rate in adjacent iteration steps is less than 10. -5 To avoid parameter fluctuation distortion; the numerical residual convergence criterion is that the discrete residuals of each field control equation are reduced to 10. -6 The following measures ensure the numerical accuracy of the equation solution. During the iteration process, two types of criteria are monitored in real time; if either criterion is met, the solution is considered converged, and the calculation for the current working condition is terminated. The combination of the two criteria balances solution efficiency and accuracy, avoids invalid iterations, and adapts to the high-precision, high-efficiency computational requirements of multi-field coupling in preform processing.

[0024] Step S1135: Conduct mesh independence verification, generate multiple sets of computational meshes with gradient mesh numbers, perform multi-field coupling solutions under the same process conditions, extract key physical parameters such as nozzle outlet velocity, preform surface temperature gradient, and peak reaction rate in the reaction zone, calculate the relative error of key parameters under different mesh densities, and determine the optimal computational mesh density when the relative error of parameters is less than 3% after the mesh density is increased.

[0025] In this embodiment, multiple sets of computational grids with progressively increasing grid numbers are generated. Multi-field coupled solutions are then performed under the same process conditions. The nozzle outlet velocity, the preform surface temperature gradient, and the peak reaction rate in the reaction zone are extracted as key verification parameters. The relative errors of the verification parameters under different grid densities are calculated. When the relative error of the parameters is less than 3% after increasing the grid density, the grid density is deemed to meet the accuracy requirements and is determined as the optimal computational grid density. This verification method eliminates the interference of grid density on the calculation results, maximizing the operational efficiency of multi-field coupled calculations while ensuring solution accuracy.

[0026] Step S114: Design an orthogonal experimental scheme to cover the typical range of values ​​for process parameters and raw material properties. Obtain the spatiotemporal distribution data of key physical parameters under different working conditions through coupled model simulation. Simultaneously collect the detection data of preform quality indicators under the corresponding working conditions and establish a three-dimensional dataset containing working condition parameters, physical parameters, and quality indicators. In this embodiment, a multi-factor orthogonal experimental scheme was designed, selecting process parameters such as heating power, gas flow rate, and cavity pressure, as well as attribute parameters such as raw material purity and particle size, covering typical production value ranges to determine the experimental conditions. A multiphysics coupling model was used to simulate and calculate each condition, obtaining spatiotemporal distribution data of key physical parameters such as flow field vortex intensity, temperature gradient, and reaction rate. Simultaneously, equipment such as a laser diameter gauge and fiber optic interferometer were used to collect quality index detection data such as preform diameter, refractive index distribution, and defect density under the corresponding conditions. The three types of data were aligned according to the spatiotemporal dimensions to construct a three-dimensional dataset of operating condition parameters, physical parameters, and quality indicators, providing standardized samples for parameter mapping modeling.

[0027] Step S115: Use Pearson correlation coefficient analysis to screen physical parameters that are strongly correlated with quality indicators, construct a preliminary mapping model through multiple linear regression or support vector regression, and combine the physical mechanism of preform processing to constrain and correct the mapping model, clarify the influence weight and sensitivity coefficient of each physical parameter on quality indicators, and form a quantitative mapping relationship. In this embodiment, Pearson correlation coefficient is used to analyze the correlation between physical parameters and quality indicators, and parameters with strong correlation coefficients greater than 0.7 are selected. Based on the selection results, a preliminary mapping model for predicting quality indicators is constructed using multiple linear regression or support vector regression. Constraints are introduced by incorporating the preform processing fluid and thermodynamic mechanism to correct the model and prevent the output from deviating from physical laws. Through model training and testing, the influence weight and sensitivity coefficient of each physical parameter on the quality indicators are quantified, the core sensitive parameters are identified, and a quantitative mapping relationship between physical parameters and quality indicators is formed. The mapping error is controlled within a preset range to ensure that the prediction results are accurate and interpretable.

[0028] Step S116: Classify and organize the information of physical field parameters, mass index parameters and model parameters, store the parameter information in a structured format and establish a parameter retrieval index, and construct a mechanism parameter dictionary; In this embodiment, physical field parameters, quality index parameters, and model parameters are categorized and organized. Physical field parameters include three-field characteristic parameters; quality index parameters include detection indicators such as diameter and refractive index; and model parameters include calculation parameters such as equation coefficients and boundary conditions. The definitions, units, value ranges, and acquisition methods of each type of parameter are clearly defined. Parameter information is stored using an XML structured format, and a hierarchical retrieval index is established to enable rapid parameter lookup and retrieval. The constructed mechanistic parameter dictionary standardizes parameters for multi-field calculations and data processing, providing standardized parameter support for subsequent model training and anomaly identification, and improving the standardization of system parameter management.

[0029] Step S117: Based on the multiphysics coupling mechanism and production practice experience, extract physical constraint rules, process constraint rules and threshold constraint rules, use logical expressions to clarify the rule triggering conditions and constraint results, establish a rule priority sorting mechanism, and formulate a constraint rule library; In this embodiment, combining multiphysics coupling mechanisms with production practice, three types of rules are extracted: physical constraints, process constraints, and threshold constraints. Physical constraints are set based on the laws of materials and thermodynamics, process constraints are aligned with the specifications of each process step, and threshold constraints clearly define the safe range of key parameters. Logical expressions are used to write the rule triggering conditions and execution results, establishing a priority mechanism: physical constraints take precedence over process constraints, and process constraints take precedence over threshold constraints. The constraint rule library can impose mechanistic constraints on model prediction, data correction, and anomaly identification, preventing results from deviating from production realities and physical laws, thereby improving the reliability and rationality of system operation.

[0030] Step S118: Select normal and abnormal operating conditions, compare the predicted values ​​of physical parameters of the coupled model with the actual sensor detection values, calculate the relative error, adjust the model parameters when the error exceeds the limit, compare the predicted values ​​of the quality index of the mapping model with the actual detection results, optimize the mapping relationship through residual analysis, and correct the threshold parameters and logical expressions in the constraint rule base based on the verification results to complete the model verification and calibration.

[0031] In this embodiment, typical normal and abnormal operating conditions in production are selected for model validation. The predicted values ​​of the coupled model's physical parameters are compared with the measured values ​​from sensors to calculate the relative error. When the error exceeds the limit, the model's boundary conditions and dynamic parameters are adjusted to complete the calibration. The predicted values ​​of the quality indicators mapped by the model are compared with the measured results, and the fitting accuracy is optimized through residual analysis. Based on the validation results, the thresholds and logical expressions of the constraint rule base are corrected to ensure that the rules are adapted to actual operating conditions. After multiple rounds of validation and calibration, the model's prediction error and rule adaptability meet production requirements, completing the overall validation of the multiphysics coupling mechanism model and ensuring the model's practicality and accuracy.

[0032] Step S120: Preprocess the raw sensor data, derive mechanism-derived features based on the multi-physics coupling mechanism model, and integrate traditional statistical features with mechanism-derived features to construct a multi-dimensional feature matrix; In this embodiment, the raw time-series sensor data is first preprocessed by removing null values, repairing anomalies, and correcting drift to ensure data integrity and accuracy. Based on a multi-physics coupling mechanism model, mechanism-derived features such as the velocity field uniformity coefficient are derived from the flow field, temperature field, and chemical reaction field, respectively. Traditional statistical features in the time and frequency domains are extracted from the preprocessed data. Z-score standardization is used to process statistical features, and min-max normalization is used to process mechanism-derived features to eliminate dimensional differences. Redundant features with VIF > 10 are removed by variance inflation factor analysis. The two types of features after screening are aligned along the time dimension, and a multi-dimensional feature matrix is ​​constructed with 50 sampling points as the window and 25 steps as the step size. The physical meaning and source of each feature are labeled, providing standardized input that takes into account both data representation and mechanism support for subsequent model training.

[0033] Step S121: Collect raw time-series sensor data on temperature, pressure, gas flow, and transmission speed during the preform processing. Divide the data into segments according to the deposition, sintering, and wire drawing processes. Remove null data caused by sensor offline or communication interruption. Use the 3σ principle combined with the local anomaly factor algorithm to identify pulse-type and step-type anomalies. Use cubic spline interpolation to repair isolated anomalies. Constrain and fill the theoretical parameter range output by the multiphysics coupling mechanism model for continuous anomaly segments to ensure data integrity. In this embodiment, raw time-series sensor data for temperature (accuracy ±0.1℃), pressure (±0.01MPa), gas flow rate (±0.1L / min), and transmission speed (±0.01m / min) are collected and segmented into intervals according to the deposition, sintering, and wire drawing processes. Null values ​​caused by sensor offline are removed, pulse anomalies are identified using the 3σ principle, and step anomalies are detected using the LOF algorithm (20 neighborhood points). Isolated outliers are repaired using cubic spline interpolation, and continuous outlier segments are filled based on the theoretical parameter range (e.g., temperature 300-1500K) output by the multiphysics coupling mechanism model. The processed data has a missing rate ≤0.5% and an anomaly repair accuracy ≥95%, ensuring data integrity and reliability.

[0034] Step S122: Calculate the mean drift of the data using a sliding time window with a window length of 100 sampling points. Establish a drift correction model by combining the normal fluctuation range of each parameter determined by the multi-physics coupling mechanism model. For temperature data, use an adaptive Kalman filter algorithm to dynamically correct the drift deviation. For pressure and flow data, perform cross-correction based on the parameter correlation relationship derived from the mechanism model. In this embodiment, a sliding window with a window length of 100 sampling points and a step size of 50 sampling points is configured. The mean drift of the sensor data is calculated, and a drift correction model is established based on the normal fluctuation range (temperature ±5℃, pressure ±0.02MPa) determined by the multiphysics coupling mechanism model. Adaptive Kalman filtering is used for temperature data to dynamically adjust the covariance matrix and correct the drift. For pressure and flow data, cross-correction is performed based on the parameter correlation relationships derived from the mechanism model (e.g., P=kQ² / T). The relative error between the corrected data and the predicted values ​​of the mechanism model is ≤3%, and the drift is controlled within the normal fluctuation range, ensuring that the data accuracy meets the feature extraction requirements.

[0035] Step S1221: Configure a sliding time window with a window length of 100 sampling points and a step size of 50 sampling points, traverse the original time series data of temperature, pressure, gas flow, and transmission speed from the sensors, calculate the average value of the data window by window, accumulate the difference between the average values ​​of adjacent windows to calculate the overall average value drift, and record the fluctuation amplitude and trend change of the data in each window. In this embodiment, a sliding time window with a window length of 100 sampling points and a step size of 50 sampling points is configured to traverse the raw time-series data of temperature, pressure, gas flow rate, and transmission speed from the sensors. The arithmetic mean of the data is calculated for each window, and the overall mean drift is obtained by summing the differences between the means of adjacent windows. Simultaneously, the standard deviation (fluctuation amplitude) and the slope of the linear fit (trend change) of the data within each window are calculated and recorded. This method can accurately capture the drift trend and local fluctuation characteristics of the data over time, providing a quantitative basis for subsequent drift determination and adapting to the time-series characteristics of sensor data.

[0036] Step S1222: Retrieve the theoretical value range of each parameter output by the multiphysics coupling mechanism model, and determine the normal fluctuation range of temperature, pressure and gas flow parameters in combination with the preform processing specifications. Use the normal fluctuation range as the drift judgment threshold. When the calculated mean drift exceeds the drift judgment threshold, it is determined that the data has drifted. Establish a drift feature library containing drift amount, drift direction and parameter type. In this embodiment, the theoretical value ranges of each parameter output from the multiphysics coupling mechanism model are retrieved. Combined with the preform processing specifications, drift judgment thresholds are determined for normal temperature fluctuations of ±5℃, pressure of ±0.02MPa, and gas flow rate of ±5%. The mean drift calculated by the sliding window is compared with the thresholds; if it exceeds the threshold, the data is judged to have drifted. A drift feature library is established to store information such as drift amount (e.g., +8℃), drift direction (positive / negative), and parameter type (temperature / pressure), providing accurate input for subsequent targeted correction and ensuring the mechanistic consistency and accuracy of drift judgment.

[0037] Step S1223: Construct an adaptive Kalman filter model for temperature data, define the state equation as the dynamic change model of temperature drift, the observation equation as the difference equation between the sensor measured temperature and the mechanism model predicted temperature, initialize the process noise covariance and the observation noise covariance, dynamically adjust the covariance matrix based on the drift amount in the drift feature library, correct the drift deviation of temperature data in real time through the prediction-update iteration process of Kalman filtering, and output the corrected temperature time series data; In this embodiment, an adaptive Kalman filter model is constructed for temperature data. The state equation is defined as x(k) = Ax(k-1) + w(k) (where x is the temperature drift, A = 1.0, and w is the process noise), and the observation equation is z(k) = Hx(k) + v(k) (where z is the difference between the measured and predicted temperatures, H = 1.0, and v is the observation noise). The process noise covariance Q = 0.01 and the observation noise covariance R = 0.1 are initialized, and Q and R are dynamically adjusted based on the drift amount in the drift feature library. The temperature drift deviation is corrected in real time through prediction-update iteration (iteration step size 0.01s), and the corrected temperature time series data is output. The drift correction error is ≤ ±0.3℃.

[0038] Step S1224: Based on the multiphysics coupling mechanism model, combined with relevant physical models such as flow field transport and ideal gas state, derive the quantitative correlation between pressure and flow parameters. Through the mechanism model, calibrate the constant parameters in the correlation equation, extend the derivation of the correlation between other relevant pressure and flow parameters, and form a parameter correlation matrix. In this embodiment, based on a multiphysics coupling mechanism model, combined with a flow field transport model and the ideal gas law, a quantitative correlation equation for the gas flow rate Q and the reaction chamber pressure P is derived: P = kQ² / T (where k is a constant and T is the reaction chamber temperature). Ten typical operating conditions are simulated using the mechanism model, and k is calibrated to 0.002 MPa・s² / (L²・K). Similarly, correlation equations for parameters such as inlet flow rate and nozzle outlet pressure, and exhaust flow rate and chamber back pressure are derived. These correlations are organized in matrix form to form a parameter correlation matrix, providing a quantitative mechanistic basis for pressure and flow rate cross-correction.

[0039] Step S1225: Perform cross-calibration on pressure data, select flow parameters that are strongly correlated with the target pressure parameter, substitute the calibrated flow data into the corresponding correlation equation to calculate the theoretical value of the target pressure, and then weight and fuse the theoretical value with the sensor measured pressure value according to the set weight to output the calibrated pressure data. In this embodiment, when performing cross-calibration on pressure data, flow parameters strongly correlated with the target pressure parameter (such as the inlet flow rate corresponding to the reaction chamber pressure) are selected. The calibrated flow data is substituted into the corresponding equation in the parameter correlation matrix to calculate the theoretical value of the target pressure. Weighted fusion is performed with a theoretical value weight of 0.6 and a sensor-measured pressure value weight of 0.4, using the formula Pcalibrated = 0.6Ptheoretical + 0.4Pmeasured. The calibrated pressure data is output, and the relative error between the calibrated pressure data and the mechanism prediction value is ≤2%, ensuring the accuracy of the pressure data.

[0040] Step S1226: Perform cross-calibration on flow data, select pressure parameters that are strongly correlated with the target flow parameters, substitute them into the corresponding correlation equation to calculate the theoretical flow value, and then weight and fuse the theoretical flow value with the sensor measured flow value according to the set weight to complete the flow data drift correction. If the deviation between the measured value and the theoretical value exceeds the allowable range of the mechanism model during the correction process, an abnormal alarm is triggered and the correlation equation parameters are re-verified. In this embodiment, the cross-calibration of flow rate data adopts the same logic as that of pressure data, selecting pressure parameters strongly correlated with the target flow rate parameter (such as the inlet pressure of the reaction chamber corresponding to the intake flow rate). The theoretical flow rate value is calculated by substituting it into the parameter correlation equation, and then weighted and fused according to a weight of 0.6 for the theoretical value and 0.4 for the measured value to obtain the corrected flow rate data. During the calibration process, if the deviation between the measured value and the theoretical value exceeds the allowable range (±8%) of the mechanistic model, an anomaly alarm is immediately triggered, and the constant parameters (such as the k value) in the correlation equation are re-verified to ensure the reliability of the flow rate calibration.

[0041] Step S1227: Verify the validity of all corrected data, compare the error index of the data before and after correction with the predicted value of the mechanism model, and check whether the data drift after correction is within the preset normal fluctuation range. If the accuracy requirements are not met, iteratively execute the correction steps until the standard is met.

[0042] In this embodiment, the validity of all corrected data is verified: the mean square error (MSE) of the data before and after correction and the predicted values ​​of the multiphysics coupling mechanism model are calculated, requiring that the MSE after correction be reduced by more than 30% compared with that before correction; the mean drift of the data is detected again through the original sliding time window to ensure that the drift is within the preset normal fluctuation range (e.g., temperature ±5℃). If any indicator does not meet the requirements, the process returns to step S1222 to re-determine the drift and perform correction, iterating until both indicators meet the requirements, ensuring the accuracy and usability of the corrected data.

[0043] Step S123: Calculate the velocity field uniformity coefficient, vortex intensity variance, and streamline deflection angle based on the flow field transport model; derive the temperature gradient standard deviation, local heat flux density, and temperature field spatiotemporal change rate based on the temperature field model; and calculate the reaction rate fluctuation coefficient, reactant consumption rate gradient, and reaction heat release intensity by combining the chemical reaction kinetic model. In this embodiment, the mechanism-derived features are derived based on the multiphysics coupling mechanism model: according to the flow field transport model, the velocity field uniformity coefficient (the ratio of the standard deviation of the flow field velocity to the average velocity), vortex intensity variance (spatiotemporal sequence variance), and streamline deflection angle (mean angle with the central axis of the cavity) are calculated; based on the temperature field model, the temperature gradient standard deviation (global statistical standard deviation), local heat flux density (instantaneous value derived from the Fourier equation), and the spatiotemporal change rate of the temperature field (the change rate of cosine similarity between adjacent sampling points) are derived; combined with the chemical reaction kinetics model, the reaction rate fluctuation coefficient (coefficient of variation), reactant consumption rate gradient (difference between adjacent time points), and reaction heat release intensity (product of reaction rate and enthalpy change) are calculated. All features have clear physical mechanism support.

[0044] Step S124: Extract traditional statistical features from the preprocessed sensor raw data, calculate the time-domain features of mean, variance, peak value, valley value, median, range, skewness, kurtosis, moving average, and autocorrelation coefficient, and convert the time-series data to the frequency domain through fast Fourier transform to extract the frequency domain features of dominant frequency amplitude, spectral energy ratio, harmonic distortion rate, and frequency band center frequency to form a set of traditional statistical features covering the time and frequency domains; In this embodiment, traditional statistical features are extracted from the preprocessed raw sensor data: time-domain features include mean, variance, peak value, valley value, median, range, skewness, kurtosis, moving average (window length 50), and autocorrelation coefficient (lag order 10), which are obtained by traversing the time-series data; the time-series data are converted to the frequency domain using Fast Fourier Transform (FFT), and frequency-domain features such as dominant frequency amplitude, spectral energy proportion (0-10Hz band), harmonic distortion rate, and frequency band center frequency are extracted. The time-domain and frequency-domain features are integrated to form a set of traditional statistical features covering the statistical characteristics and frequency distribution of the data, totaling 20 feature parameters.

[0045] Step S125: Standardize the mechanism-derived features and traditional statistical features. Use the Z-score standardization method to normalize the traditional statistical features and use min-max normalization to map the mechanism-derived features to the [0,1] interval to ensure the consistency of the feature value range. In this embodiment, the mechanism-derived features and traditional statistical features are standardized respectively: for traditional statistical features such as mean, variance, and dominant frequency amplitude, the Z-score standardization method (x'=(x-μ) / σ, where μ is the mean and σ is the standard deviation) is used to eliminate dimensional differences; for mechanism-derived features such as velocity field uniformity coefficient and reaction rate fluctuation coefficient, the min-max normalization method (x'=(xx) / σ) is used. min ) / (x max -x min The feature values ​​are mapped to the [0,1] interval. After standardization, all feature values ​​have a uniform range, avoiding the impact of differences in units or numerical values ​​on subsequent feature selection and model training.

[0046] Step S126: Use variance inflation factor analysis to screen features, remove highly redundant features with VIF greater than 10, retain features that are strongly correlated with the quality indicators of the preform, align the screened traditional statistical features and mechanism-derived features according to the time dimension, and construct a multi-dimensional feature matrix with rows representing samples and columns representing features by using a sliding time window with a window length of 50 sampling points and a step size of 25 sampling points as a sample, and label the physical meaning and source of each feature.

[0047] In this embodiment, variance inflation factor (VIF) analysis is used to screen the standardized features. A VIF > 10 is set as the threshold for redundancy features, eliminating highly redundant features such as temperature mean and moving average. The screened traditional statistical features (such as peak temperature and pressure variance) and mechanism-derived features (such as temperature gradient standard deviation) are aligned by timestamps. A sliding time window with a window length of 50 sampling points and a step size of 25 sampling points is configured, with each window serving as a sample. A multi-dimensional feature matrix is ​​constructed, with rows representing samples and columns representing features. The matrix dimension is [number of samples × 30] (10 mechanistic features + 20 statistical features), and the physical meaning (such as reaction heat release intensity: product of reaction rate and enthalpy change) and source (mechanism / statistics) of each feature are labeled.

[0048] Step S130: Build a data-driven fusion architecture with mechanistic constraints, use the theoretical parameter range output by the mechanistic model as the regularization constraint term of the data-driven model, integrate multi-scenario mechanistic knowledge, and combine digital twins to achieve virtual and real data fusion training; In this embodiment, a fusion architecture of mechanism constraints and data-driven approaches is constructed, with the core being mechanism constraint regularization, multi-scenario knowledge integration, and virtual-real data fusion training. The theoretical parameter range output by the multi-physics coupling mechanism model is transformed into a weighted L2 regularization term and embedded into a CNN-LSTM data-driven model. A multi-scenario mechanism knowledge graph is constructed to provide prior support, and a 1:1 digital twin is built to generate high-fidelity virtual data. Virtual-real data fusion is achieved through spatiotemporal alignment and consistency verification. A two-stage pre-training-fine-tuning strategy is used to train the model, embedding a dynamic constraint adjustment mechanism to adapt to scene changes, and establishing a multi-indicator monitoring system to ensure training effectiveness. This architecture balances data fitting accuracy and physical mechanism conformity, improving the model's cross-scenario generalization ability and prediction reliability.

[0049] Step S131: Quantify the mechanism constraints and transform them into regularized forms. Retrieve the theoretical value ranges of key parameters such as flow field velocity, temperature gradient, and reaction rate output by the multi-physics coupling mechanism model. Combine the process specifications to determine the upper and lower boundary thresholds of each parameter. Transform the interval constraints into regularization terms of the data-driven model. Use weighted L2 regularization to construct the constraint function. The weight coefficients are set according to the sensitivity coefficients of the parameters to the preform quality indicators, so that the parameter prediction values ​​during model training remain within the allowable range of the mechanism. Construct a multi-scenario mechanism knowledge graph, sort out the mechanism rules corresponding to different combinations of process parameters and raw material batches, and store the reaction kinetic parameter range, flow field-temperature field coupling law, and quality indicator sensitive parameter set in a structured manner with rule nodes and associated edges to form a multi-scenario mechanism knowledge graph. In this embodiment, the flow field velocity (0.5-2 m / s), temperature gradient (10-50℃ / mm), and reaction rate (10) are adjusted. -6 -10 -4 The theoretical range of the mechanism (mol / (L・s)) is determined by combining the upper and lower boundary thresholds with the process specifications. A constraint function R(θ)=Σωᵢ(θᵢ-θ0ᵢ)² is constructed using weighted L2 regularization, with the weight ωᵢ set according to the sensitivity coefficients of the parameters to quality indicators (e.g., temperature gradient sensitivity coefficient 0.8, flow field velocity 0.5). The mechanistic rules corresponding to different combinations of process parameters (heating power 5-15kW) and raw material batches (purity 99.99%-99.999%) are analyzed. The reaction kinetic parameter ranges and flow field-temperature field coupling laws are structured and stored as rule nodes (e.g., deposition scenario-reaction activation energy 90kJ / mol) and associated edges, forming a multi-scenario mechanism knowledge graph.

[0050] Step S132: Construct a digital twin for preform processing. Based on the three-dimensional geometric model of the reaction chamber, the multi-physics coupling mechanism model, and the equipment operating parameters, construct a digital twin that is mapped 1:1 to the physical production system. Real-time data synchronization between the twin and the physical equipment is achieved through the Industrial Internet protocol. The spatiotemporal evolution of the flow field, temperature field, and chemical reaction field is dynamically replicated and high-fidelity virtual data is generated. In this embodiment, a digital twin is constructed based on the three-dimensional geometric model of the reaction chamber (diameter 50-80cm, height 100-150cm), the multiphysics coupling mechanism model, and equipment operating parameters (transmission speed 0.5-2m / min), mapping it 1:1 to the physical production system. Real-time data synchronization is achieved via the OPCUA industrial internet protocol at a frequency of 100Hz, dynamically replicating states such as flow field vortex distribution, spatiotemporal evolution of the temperature field, and changes in chemical reaction rates. High-fidelity virtual data is generated based on the twin simulation, covering three types of data: process parameters, physical field parameters, and quality indicators. The sampling frequency is 50Hz, and the data format is uniformly JSON. The relative error between the virtual data and the real data is ≤5%, providing sufficient samples for fusion training.

[0051] Step S133: Perform spatiotemporal alignment and consistency verification on the actual production data collected by physical sensors and the twin virtual data. Use a timestamp synchronization algorithm to complete the spatiotemporal dimension alignment and unify the sampling frequency and data format. Calculate the relative error between virtual and real data to achieve verification. For virtual data with excessive deviation, reverse the twin boundary conditions and model parameters based on physical sensor data. In this embodiment, a timestamp synchronization algorithm (accuracy ±1ms) is used to spatiotemporally align the physical sensor measured data and the twin virtual data, with a unified sampling frequency of 50Hz and data format of CSV. The relative error between the virtual and real data is calculated as (|Xreal - Xvirtual| / Xreal × 100%), and a consistency check is performed with an error threshold of ≤5%. For virtual data with deviations exceeding the limit (e.g., temperature deviation of 7%), the boundary conditions (e.g., heat flux density of heating elements) and model parameters (e.g., turbulence empirical coefficients) of the twin are reversed based on the physical sensor data at the corresponding time, and iteratively corrected until the relative error between the virtual and real data meets the threshold requirement, ensuring data homogeneity and reliability.

[0052] Step S134: Construct a multi-scenario mechanism knowledge transfer mechanism. Based on domain adaptive transfer learning, transform the multi-scenario mechanism rules in the mechanism knowledge graph into prior knowledge for model training. Introduce a knowledge attention mechanism in the embedding layer of the CNN-LSTM data-driven model to assign weights to the mechanism rules of different scenarios. In this embodiment, based on the domain-adaptive transfer learning method, the rules in the mechanistic knowledge graph are transformed into prior knowledge for model training. A knowledge attention mechanism is introduced into the embedding layer of the CNN-LSTM model, constructing an attention weight matrix W∈R.mxn (where m is the number of scenes and n is the number of rules), the weights Wᵢⱼ=exp(sᵢⱼ) / Σexp(sᵢⱼ) are calculated using the Softmax function, where sᵢⱼ is the matching degree between the current scene and the j-th rule. Differentiated weights are assigned to the mechanism rules of different scenes such as deposition, sintering, and wire drawing (e.g., the reaction kinetics rule in the deposition scene has a weight of 0.7, and the temperature field coupling rule in the sintering scene has a weight of 0.8), so that the model prioritizes learning the mechanism knowledge that is suitable for the current scene and improves the cross-scene transfer capability.

[0053] Step S135: Design a virtual-real data fusion training strategy, divide the training dataset into a virtual data subset, a real data subset and a virtual-real fusion data subset, and adopt a two-stage training mode of pre-training-fine-tuning. In the first stage, pre-training is based on the virtual data subset and the mechanism knowledge graph. In the second stage, the virtual-real fusion data subset is divided into a training set and a validation set according to the proportion to fine-tune the model and dynamically adjust the weight coefficients of the regularization constraint term. In this embodiment, the dataset is divided into a virtual data subset, a real data subset, and a virtual-real fusion data subset in a 4:3:3 ratio. A two-stage training process of pre-training and fine-tuning is employed: the first stage involves 100 rounds of pre-training based on the virtual data subset and the mechanistic knowledge graph to learn fundamental physical laws, with an initial learning rate of 0.001; the second stage involves dividing the virtual-real fusion data subset into a training set and a validation set in a 7:3 ratio, and fine-tuning for 200 rounds, dynamically adjusting the weight coefficients of the regularization constraint terms (adjustment range 0.1-0.9). When the prediction error of the validation set quality index increases, the constraint weights are increased; when the error stabilizes, the weights are decreased to enhance data fitting, balancing mechanistic conformity and fitting accuracy.

[0054] Step S136: Embed a dynamic constraint adjustment mechanism, calculate the degree of deviation between the model prediction value and the mechanism interval in real time through a sliding window, adaptively adjust the regularization constraint weight according to the magnitude of the deviation, and call the corresponding mechanism constraint parameters in combination with the scene recognition results of the multi-scene knowledge graph; In this embodiment, a sliding window with a window length of 100 sampling points is configured to calculate the deviation between the model's predicted value and the mechanism interval in real time: D = |ypredicted - ymechanism| / (yupper - ylower). When D < 1%, the regularization constraint weight is reduced to 0.1-0.3 to enhance data fitting; when D > 3%, the weight is increased to 0.7-0.9 to force the model to regress to the mechanism; when 1% ≤ D ≤ 3%, the weight remains at 0.4-0.6. Combining the scene recognition results from the multi-scene knowledge graph (e.g., identified as a high-power sintering scene), the mechanism constraint parameters corresponding to the scene are automatically called (e.g., temperature gradient threshold of 40℃ / mm) to achieve dynamic adaptation of constraint strength to the scene.

[0055] Step S137: Establish a training monitoring and evaluation system for the fusion model, set key evaluation indicators such as prediction accuracy, cross-scenario generalization error, and mechanism constraint satisfaction rate, monitor the changes of indicators in real time, and automatically trigger multi-scenario knowledge graph rule updates and virtual and real data supplementary training when the generalization error exceeds the standard.

[0056] In this embodiment, three key evaluation indicators are set: the accuracy of quality indicator prediction (diameter deviation ≤ ±0.3%, refractive index distribution error ≤ 10). -6 The training process includes monitoring cross-scenario generalization error (error fluctuation ≤5% under different process scenarios) and mechanism constraint satisfaction rate (≥98%). Changes in these metrics are monitored every 10 rounds during training, with data recorded and visualized in real time. If the generalization error exceeds the limit (e.g., fluctuation reaches 7%), a multi-scenario knowledge graph rule update is automatically triggered (supplementing the mechanism rules for that scenario). Additionally, 20% of the corresponding scenario's virtual and real data is supplemented from the digital twin and physical system, and fine-tuning is performed again to ensure the stability and effectiveness of the fusion model.

[0057] Step S140: Capture process drift signals through a sliding time window, and dynamically optimize data-driven model parameters and mechanism constraint strength by combining the results of sensitivity analysis of the mechanism model, so as to adapt to the scenarios of process parameter adjustment and raw material batch change. In this embodiment, a sliding time window with a window length of 100 sampling points and a step size of 25 sampling points is configured to traverse the multi-dimensional feature matrix of preform processing, extract the time-series change features of process parameters in real time, and accurately capture process drift signals such as temperature drift, flow fluctuation, and transmission speed deviation. Based on a multi-physics coupling mechanism model, parameter sensitivity analysis is conducted to calculate the sensitivity coefficients of key parameters such as flow field velocity, temperature gradient, and reaction rate to process changes and raw material batch changes, classifying parameters into high and low sensitivity levels. Based on the sensitivity analysis results, the learning rate, batch size, and network weights of the data-driven model are dynamically adjusted, and the mechanism constraint strength is simultaneously and adaptively optimized. Constraint weights are increased for high-sensitivity parameters and reduced for low-sensitivity parameters, quickly adapting to scenarios of process parameter adjustments, raw material purity changes, and batch changes, ensuring the prediction accuracy and mechanism consistency of the model when operating conditions change.

[0058] Step S150: Construct a multi-dimensional anomaly feature system covering all types of anomalies, and use signal frequency band decomposition and feature enhancement technology to enhance the ability to perceive weak signals corresponding to early faults and minor defects. In this embodiment, a multi-dimensional anomaly feature system covering all types of anomalies in preform processing is constructed, with the core objective of enhancing weak signals of early faults and minor defects. Wavelet packet transform is used to decompose the time-series signals from sensors such as temperature, pressure, and gas flow rate into frequency bands. The number of decomposition layers is set according to the frequency characteristics of the processing anomaly signals, and sensitive frequency bands highly correlated with weak flow field disturbances, slowly changing parameters, and micron-level bright spot defects are selected, while environmental noise and equipment interference frequency bands are eliminated. The sensitive frequency band features are enhanced through a channel attention mechanism and a temporal differential amplification algorithm, amplifying weak anomaly signals and improving the ability to detect early latent faults. By integrating statistical, temporal, spatial, and multi-physics coupling features, a full-dimensional anomaly characterization system is formed. Validated and iterated using actual production samples, this system can comprehensively cover explicit anomalies and early minor defects, providing highly recognizable input features for anomaly classification.

[0059] Step S151: Review the types of anomalies throughout the preform processing, identify the temporal manifestations of pulse-type, step-type, and pattern-type anomalies, define the characteristic manifestations of micron-level bright spot defects, weak flow field disturbances, and early faults with slowly changing parameters, and establish the correspondence between anomaly types and characteristic manifestations. In this embodiment, anomaly types throughout the entire process of preform deposition, sintering, and wire drawing are comprehensively reviewed, and classified into three types of explicit anomalies: pulse-type, step-type, and pattern-type. Pulse-type anomalies are defined as instantaneous data spikes, step-type anomalies as parameters jumping to a steady state, and pattern-type anomalies as regular signal distortions. Three types of early faults are identified: micron-level bright spots (micro-defects), weak flow field disturbances, and slowly changing parameters. Micron-level bright spots are local pixel anomalies in visual images, weak flow field disturbances are small periodic fluctuations in flow velocity, and slowly changing parameters represent long-term, gradual shifts in process parameters. Actual samples of each anomaly and early fault are collected, and their temporal, numerical, and morphological characteristics are labeled. A one-to-one correspondence between anomaly types and characteristic manifestations is established, forming an anomaly feature labeling library, providing precise target guidance for the construction of the feature framework.

[0060] Step S152: Based on the correspondence between the anomaly types and feature manifestations, a multi-dimensional anomaly feature framework is constructed, and statistical features, temporal features, spatial features and multi-physics coupling features are extracted. At the same time, features derived from the flow field uniformity and temperature gradient variance mechanism are incorporated to form a basic feature set for the full-dimensional representation of anomalies. In this embodiment, a four-layer, multi-dimensional anomaly feature framework is constructed based on the correspondence between anomaly types and feature manifestations. Statistical features of sensor data are extracted, including mean, variance, and coefficient of variation; temporal features are extracted, including trend slope, frequency of abrupt changes, and duration; spatial features of visual inspection are extracted, including defect pixel density, location correlation, and geometric morphology parameters; and multi-physics coupling features are extracted, including the coordinated changes in temperature and flow field, and the linkage deviation between reaction rate and pressure. Simultaneously, mechanism-derived features such as flow field uniformity coefficient, temperature gradient variance, and reaction rate fluctuation coefficient are incorporated, forming a comprehensive set of basic features for anomaly characterization covering numerical, temporal, spatial, and physical coupling aspects. This ensures that the features can fully characterize the essential attributes of various anomalies and early faults.

[0061] Step S153: Wavelet packet transform is used to decompose the time-series signals of temperature, pressure and gas flow sensors into frequency bands. The number of decomposition layers is set according to the frequency characteristics of abnormal signals. The differences in energy distribution of each frequency band under normal and abnormal conditions are compared. Sensitive frequency bands corresponding to early faults and minor defects are screened and noise interference frequency bands are removed. In this embodiment, wavelet packet transform is used to perform frequency band decomposition on the time-series signals of temperature, pressure, and gas flow sensors. Based on the frequency characteristics of abnormal signals from preform processing concentrated in the 0-50Hz range, the decomposition layer is set to 4 layers, decomposing the original signal into 16 frequency band sub-signals. Energy values ​​of each frequency band are extracted under normal and abnormal operating conditions, and the energy distribution difference is calculated. Frequency bands with a difference greater than 0.6 are selected as sensitive bands for early faults and minor defects, while noise interference bands with a difference less than 0.2 are removed. This frequency band selection achieves preliminary purification of weak abnormal signals, reduces noise interference on subsequent feature extraction and anomaly identification, and significantly improves the sensitivity of the feature system to early anomalies.

[0062] Step S154: Feature enhancement is performed on the sensitive frequency band sub-signals. Adaptive weights are assigned to the sensitive frequency bands through the channel attention mechanism. The time-series differential amplification algorithm is used to extract subtle changes in the signal and to perform local enhancement processing on the spatial features of micron-level defects. In this embodiment, multi-level feature enhancement is performed on the sensitive frequency band sub-signals. A channel attention mechanism is introduced to construct a frequency band weight allocation network, which assigns adaptive weights according to the correlation between the sensitive frequency band and the anomaly, thereby enhancing the features of highly correlated frequency bands and suppressing weakly correlated components. To address the issue of low amplitude and easy masking of early fault and minute defect signals, a temporal differential amplification algorithm is used to calculate the difference between adjacent sampling points and amplify it by 10-20 times to extract subtle signal change features. For the spatial image features of micron-level bright spot defects, a local contrast enhancement algorithm is used to improve the pixel difference between the defect and normal areas, highlighting the morphology of minute defects. Through multi-dimensional enhancement, the model's ability to perceive and distinguish weak anomalous signals is significantly improved.

[0063] Step S155: Perform standardization processing on the enhanced multidimensional anomaly features, remove redundant features through variance inflation factor analysis and Pearson correlation analysis, and retain features that are highly sensitive to various anomalies, early failures, and minor defects. In this embodiment, the enhanced multi-dimensional anomaly features are standardized. Z-score standardization is used to eliminate differences in statistical and temporal feature dimensions, and min-max normalization is used to map spatial and coupled features to the [0,1] interval, unifying the feature value range. Variance inflation factor and Pearson correlation analysis are used to screen features, setting VIF > 10 as highly redundant and correlation coefficient absolute value < 0.3 as weakly correlated, and corresponding redundant and invalid features are eliminated. After screening, the number of features is reduced by more than 30%, with no loss of anomaly identification information, retaining core features that are highly sensitive to explicit anomalies, early faults, and minor defects, reducing the computational complexity of the model, and improving feature effectiveness and identification efficiency.

[0064] Step S156: Classify and integrate the filtered features according to the anomaly type, match feature subsets for each anomaly, and form a multi-dimensional anomaly feature system covering all types of anomalies; In this embodiment, the highly sensitive core features selected are categorized and integrated according to anomaly type, and a dedicated feature subset is matched for each type of anomaly. Pulse-type anomalies are assigned a feature subset based on amplitude and variance; step-type anomalies are assigned a feature subset based on trend slope and abrupt change points; weak flow field disturbances are assigned a feature subset based on frequency domain energy and uniformity; and micrometer-level defects are assigned a feature subset based on spatial morphology and local gradients. Integrating these dedicated subsets forms a multi-dimensional feature system covering all types of anomalies, including overt anomalies, early faults, and minute defects. This system can specifically characterize the core features of different anomalies, avoid feature confusion, and significantly improve the recognition accuracy and response speed of subsequent anomaly classification models.

[0065] Step S157: Verify the feature system using actual production anomaly samples, test the feature recognition accuracy of early faults and minor defects, optimize the wavelet packet decomposition layer and attention weight parameters based on the verification results, and iteratively update the feature system.

[0066] In this embodiment, 500 sets of explicit anomaly samples and 300 sets of early fault and minor defect samples were collected from the production site to construct an anomaly test sample library. The feature system was validated using these test samples, with an accuracy rate of ≥95% and a recall rate of ≥92% for early fault and minor defect identification set as the benchmarks. If these benchmarks were not met, the wavelet packet decomposition layer was optimized to 3 or 5 layers based on the results, and sensitive frequency bands were re-selected; the channel attention weight parameters were adjusted to strengthen the weight of weak signal features. Through multiple rounds of iterative validation and parameter optimization, the feature system was continuously updated to ensure that the system's representation accuracy and stability for various anomalies met the requirements of field applications.

[0067] Step S160: Construct a multimodal anomaly classification model with an embedded causal inference module, use learning methods and loss functions to complete model training, distinguish between true anomalies and false anomalies caused by environmental interference, establish an adaptive threshold dynamic optimization mechanism based on mechanism constraints and real-time data, and set and iteratively optimize the threshold according to the anomaly type. In this embodiment, a multimodal anomaly classification model with an embedded causal inference module is constructed to accurately distinguish between genuine anomalies and false anomalies caused by environmental interference, and an adaptive threshold dynamic optimization mechanism is established. The model takes a multi-dimensional anomaly feature system as input, integrates multimodal features, and embeds a causal inference module to separate the correlation between interference and the driving factors of genuine anomalies. A two-stage training strategy and a composite loss function are used to train the model, balancing data fitting and mechanistic constraints. Based on the multi-physics coupling mechanism and real-time production data, thresholds are set differently according to the degree of anomaly's impact on quality. Dynamic iteration of the threshold is achieved through sliding window and Bayesian optimization, realizing dynamic matching between the threshold and operating conditions and mechanisms, improving the accuracy of anomaly identification, and reducing false alarms and false negatives.

[0068] Step S161: Design the basic architecture of the multimodal anomaly classification model, clarify the input forms of statistical features, temporal features, spatial features and multi-physics coupling features, use CNN to extract spatial features, LSTM to extract temporal features, fully connected layers to process coupling and statistical features, introduce a modal attention mechanism to weight and fuse the features of each modality, and generate a multimodal fusion feature vector; In this embodiment, a multimodal anomaly classification model architecture is designed, clearly defining the vector input forms for statistical, temporal, spatial, and coupled features. A CNN network is used for convolutional pooling of spatial features to extract defect morphology and location vectors; an LSTM network is used to model temporal feature sequences to capture the temporal evolution patterns of anomalies; and a fully connected layer is used for nonlinear mapping of statistical and coupled features to generate corresponding feature vectors. A modal attention mechanism is introduced to calculate the contribution weights of each modality, and the four types of features are weighted and fused to generate a multimodal fused feature vector with unified dimensions. This architecture can fully exploit the anomaly representation information of different modalities, improving the model's ability to learn and distinguish features of complex anomalies.

[0069] Step S162: Embed the causal inference module, construct the anomaly causal graph based on the multi-physics coupling mechanism and production practice, clarify the causal relationship between the true anomaly driving factor, environmental interference factor and the abnormal result, learn the causal structure parameters by adopting the structural causal model, separate the causal features and interference features in the fusion features by adjusting the formula through the backdoor, and calculate the causal contribution of each feature. In this embodiment, a causal inference module is embedded in the model. Based on the multiphysics coupling mechanism and production practice, an anomaly causal graph is constructed to clarify the causal paths between true anomaly driving factors, environmental interference factors, and anomaly results. A structural causal model is used to learn causal parameters and determine the factor causal strength coefficients. A backdoor adjustment formula is used to eliminate confounding factor interference, separating causal and interference features from the fused features, and calculating the causal contribution of each feature to the anomaly. Features with a contribution greater than 0.7 are considered true anomaly driving features, while those less than 0.3 are considered environmental interference features, providing a quantitative basis for distinguishing between true and false anomalies and preventing the model from being misled by interference signals.

[0070] Step S163: The model is trained using a two-stage learning strategy of pre-training and fine-tuning. In the pre-training stage, the parameters are initialized with labeled normal working conditions, true anomalies, and false anomalies. In the fine-tuning stage, causal constraint optimization is introduced, and a composite loss function is designed, which includes weighted cross-entropy loss, causal consistency loss, and mechanism constraint loss. The weights of the three are determined by grid search, and the training is iteratively trained with the goal of minimizing the total loss. In this embodiment, a two-stage strategy of pre-training and fine-tuning is employed to train the model. Pre-training uses 1000 labeled normal, true anomaly, and false anomaly samples to initialize model parameters and learn basic anomaly patterns. Fine-tuning introduces causal constraint optimization, designing a composite loss function that includes weighted cross-entropy, causal consistency, and mechanistic constraints. The optimal weight ratio is determined through grid search, and the model is iteratively trained for 200 epochs with the goal of minimizing the total loss, with the learning rate dynamically decaying to 0.0001. The composite loss function balances the model's classification accuracy, causal consistency, and mechanistic compliance, effectively improving the model's generalization ability and anomaly recognition stability.

[0071] Step S164: Construct a mechanism to distinguish between true and false anomalies, set a threshold for causal contribution, and determine a true anomaly when the sum of the causal contributions of the true anomaly driving factors exceeds the threshold and the contribution of the interference factors is lower than the set value. If the contribution of the interference factors is dominant, it is determined to be a false anomaly caused by environmental interference. Record the distinction results for subsequent optimization. In this embodiment, a mechanism for distinguishing between true and false anomalies is constructed based on causal contribution. A threshold of 0.7 is set for the sum of the contributions of true anomaly driving factors, and a threshold of 0.3 is set for interference factors. During model inference, factor contributions are calculated in real time. If the sum of driving factors > 0.7 and the interference factor < 0.3, it is determined to be a true anomaly in the equipment or process; if the interference factor > 0.7 and the driving factor < 0.3, it is determined to be a false anomaly due to environmental interference. The distinction results, contributions, and anomaly types are recorded in real time and stored in an anomaly database, providing accurate data support for subsequent threshold optimization, model fine-tuning, and mechanism rule updates.

[0072] Step S165: Retrieve the theoretical range of key features of each anomaly type output by the multiphysics coupling mechanism model, determine the initial threshold range in combination with the process specifications, and adjust the priority of the threshold according to the sensitivity coefficient of the anomaly's impact on quality, with high-sensitivity anomalies having higher priority than low-sensitivity anomalies. In this embodiment, the theoretical range of key anomaly features output by the multiphysics coupling mechanism model is retrieved, and the initial threshold range for various anomalies is determined in conjunction with process specifications. Sensitivity coefficients are calculated based on the impact of anomalies on diameter, refractive index, and defect density, and threshold adjustment priorities are assigned. High-sensitivity anomalies such as cracks and diameter deviations are classified as Level 1, while low-sensitivity anomalies such as weak flow field disturbances and small parameter drifts are classified as Level 2. Level 1 anomalies are initially set with narrow threshold ranges, while Level 2 anomalies are set with wider ranges. High-priority anomalies are optimized and verified first to ensure timely identification of high-risk anomalies, balancing identification accuracy and on-site response efficiency.

[0073] Step S166: Use a sliding time window to statistically analyze the distribution parameters of real-time multimodal features. With the anomaly identification accuracy, false alarm rate, and false negative rate as optimization objectives, dynamically adjust the thresholds for each anomaly type using a Bayesian optimization algorithm. Set a narrow threshold range for high-sensitivity anomalies and a wide threshold range for low-sensitivity anomalies to achieve differentiated settings. In this embodiment, a sliding time window with 50 sampling points is configured to statistically analyze the distribution parameters such as mean, variance, and quantiles of multimodal anomaly features in real time. Accuracy, false alarm rate, and missed alarm rate are used as multi-objective optimization functions, and a Bayesian optimization algorithm is employed to dynamically adjust the thresholds for each anomaly. For first-level high-sensitivity anomalies, the interval is narrowed to reduce missed alarms; for second-level low-sensitivity anomalies, the interval is expanded to reduce false alarms. The optimization process simultaneously follows the mechanism-constrained intervals, achieving differentiated setting and dynamic optimization of anomaly thresholds, thereby improving the targeting and on-site adaptability of anomaly identification.

[0074] Step S167: Collect the anomaly identification results during model operation, calculate the threshold fit score for each anomaly type, expand the threshold range if the false positive rate exceeds the standard, and narrow the threshold range if the false negative rate exceeds the standard. Combine the mechanism model parameter update to synchronously correct the threshold range and ensure that the threshold is dynamically matched with the working conditions and mechanism constraints. In this embodiment, the model's on-site identification results are collected in real time, and the accuracy, false positive rate, and false negative rate are statistically analyzed. A threshold fit score of ≥0.9 is considered satisfactory. If the false positive rate exceeds the standard, the threshold is widened by a step of 5%, and if the false negative rate exceeds the standard, the threshold is narrowed by a step of 5%. When the mechanistic model parameters or process specifications are updated, the theoretical lower and lower limits of the threshold are simultaneously corrected to ensure that the threshold dynamically matches the real-time operating conditions and mechanistic constraints. A threshold iteration log is established to record the adjusted parameters and optimization effects, forming an adaptive optimization closed loop to ensure long-term stable and reliable anomaly identification.

[0075] Step S168: Use actual production samples containing true anomalies, false anomalies, and early weak anomalies to test the model, evaluate the accuracy of distinguishing between true and false anomalies and the accuracy of each anomaly identification, adjust the cause-effect graph structure, loss function weights and threshold optimization hyperparameters, and iteratively optimize the model and threshold mechanism until the preset accuracy requirements are met.

[0076] In this embodiment, 800 sets of samples of true anomalies, false anomalies, and early weak anomalies were collected from the field to construct a test set. The accuracy of the model in distinguishing between true and false anomalies, the accuracy of anomaly identification, and the early fault detection rate were tested. A distinction accuracy of ≥96% and an early detection rate of ≥93% were set as the benchmarks. If the benchmarks were not met, the anomaly cause-effect graph structure was adjusted, causal paths were supplemented, the weights of the composite loss function were optimized, and the Bayesian optimization hyperparameters were adjusted. After multiple rounds of iterative testing and parameter optimization, the indicators were optimized until the requirements were met, ensuring that the model could adapt to the complex and ever-changing working environment of the precast bar processing site.

[0077] Step S170: Complete the on-site deployment and generalization capability verification of the model, optimize the technical solution based on actual application cases, build a closed-loop system of mechanism modeling, data training, on-site verification, and iterative optimization, and continuously update the sample library and model parameters.

[0078] In this embodiment, a multiphysics coupling model, a data-driven fusion model, and anomaly classification model are deployed to the on-site industrial control system. Real-time data from sensors and visual inspections are integrated to conduct on-site generalization verification. Case studies involving different production lines, processes, and raw material batches are selected to test the model's prediction accuracy, recognition rate, and response speed across multiple scenarios. Based on application feedback, the mechanism parameters, feature structure, and training strategy are optimized in reverse, constructing a closed-loop system for mechanism modeling, data training, on-site verification, and iterative optimization. On-site samples are continuously collected to expand the sample library, and model parameters, mechanism rules, and thresholds are regularly updated to ensure the system adapts to long-term changes in operating conditions and continuously improves the intelligent control level of preform processing quality.

[0079] Based on the same inventive concept, please refer to Figure 2 The diagram shows a schematic block diagram of a preform processing monitoring system 100 based on big data, provided in an embodiment of this application, for performing the above-described monitoring method for preform processing based on big data. The preform processing monitoring system 100 based on big data may include a communication unit 110, a machine-readable storage medium 120, and a processor 130.

[0080] In this embodiment, both the machine-readable storage medium 120 and the processor 130 are located within the preform processing monitoring system 100 integrating big data and are separately configured. However, it should be understood that the machine-readable storage medium 120 may also be independent of the preform processing monitoring system 100 integrating big data and may be accessed by the processor 130 via a bus interface. Alternatively, the machine-readable storage medium 120 may also be integrated into the processor 130 and may communicate and interact with external systems via the communication unit 110.

[0081] The processor 130 is the control center of the preform processing monitoring system 100, which integrates big data. It connects to various parts of the system via various interfaces and lines. By running or executing software programs and / or modules stored in the machine-readable storage medium 120, and by calling data stored in the machine-readable storage medium 120, it performs various functions and processes data of the preform processing monitoring system 100, thereby providing overall monitoring of the system. Optionally, the processor 130 may include one or more processing cores; for example, the processor 130 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may also not be integrated into the processor. The machine-readable storage medium 120 is used to store machine-executable instructions for executing the scheme of this application, and the processor 130 is used to execute the machine-executable instructions stored in the machine-readable storage medium 120 to implement the preform processing monitoring method combined with big data provided in the aforementioned method embodiments.

[0082] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A method for monitoring a preform processing procedure in connection with big data, characterized by: Includes the following steps: A multi-physics coupling mechanism model for preform processing is constructed. Based on the relevant basic equations of multi-physics, the dynamic characteristics of flow field, temperature field and chemical reaction are quantified. The mapping relationship between key physical parameters and quality indicators is established, and a mechanism parameter dictionary and constraint rule base are formed. The raw sensor data is preprocessed, and mechanism-derived features are derived based on the multi-physics coupling mechanism model. Traditional statistical features and mechanism-derived features are then integrated to construct a multi-dimensional feature matrix. A data-driven fusion architecture constrained by mechanism is built, and the theoretical parameter range output by the mechanism model is used as the regularization constraint term of the data-driven model. Multi-scenario mechanism knowledge is integrated, and virtual and real data fusion training is achieved by combining digital twins. By capturing process drift signals through a sliding time window and combining the results of sensitivity analysis of the mechanism model, the data-driven model parameters and mechanism constraint strength are dynamically optimized to adapt to scenarios of process parameter adjustment and raw material batch change. Construct a multi-dimensional anomaly feature system covering all types of anomalies, and use signal frequency band decomposition and feature enhancement technology to enhance the ability to perceive weak signals corresponding to early faults and minor defects; A multimodal anomaly classification model with an embedded causal inference module is constructed. The model is trained using learning methods and loss functions to distinguish between true anomalies and false anomalies caused by environmental interference. An adaptive threshold dynamic optimization mechanism based on mechanistic constraints and real-time data is established, and the threshold is set and iteratively optimized according to the anomaly type. Complete the on-site deployment and generalization capability verification of the model, optimize the technical solution based on real-world application cases, construct a closed-loop system for mechanism modeling, data training, on-site verification, and iterative optimization, and continuously update the sample library and model parameters.

2. The method for monitoring the preform processing of rods using big data as described in claim 1, characterized in that: The aforementioned construction of a multi-physics coupling mechanism model for preform processing quantifies the flow field, temperature field, and chemical reaction kinetics based on multi-physics related fundamental equations, establishes the mapping relationship between key physical parameters and quality indicators, and forms a mechanism parameter dictionary and constraint rule base, including: Clarify the boundary conditions and basic parameters of the preform deposition, sintering and wire drawing processes, sort out the equipment structure parameters, raw material property parameters and process operation parameters, and define the quantitative detection standards and accuracy requirements for the preform diameter, refractive index distribution and defect density quality indicators. A basic physical model is established for each field. The flow field control equations are simplified based on the Navier-Stokes equations and the geometric characteristics of the reaction chamber. A turbulence model is introduced and the flow field boundary conditions are set. The equations are discretized and the calculation parameters are initialized using the finite volume method. The radiation heat transfer model and the convective heat transfer coefficient calculation model are integrated based on the Fourier heat conduction equation. The temperature field boundary conditions are clarified to quantify the spatiotemporal evolution of the temperature field. Based on the Arrhenius equation and the deposition / sintering reaction, the kinetic parameters are determined and a quantitative relationship between the reaction rate and temperature and reactant concentration is established. A multi-field coupled solution system is constructed, the coupling order of flow field, temperature field and chemical reaction field is determined, multi-field coupling terms are defined, the coupling equation is solved by partitioned iterative algorithm and the iterative convergence criterion is set, and the calculation grid density is optimized by verifying grid independence. Design an orthogonal experimental scheme to cover the typical range of values ​​for process parameters and raw material properties. Obtain the spatiotemporal distribution data of key physical parameters under different working conditions through coupled model simulation. Simultaneously collect the detection data of preform quality indicators under the corresponding working conditions and establish a three-dimensional dataset containing working condition parameters, physical parameters, and quality indicators. Pearson correlation coefficient analysis was used to screen physical parameters that are strongly correlated with quality indicators. A preliminary mapping model was constructed by multiple linear regression or support vector regression. The mapping model was then constrained and corrected by combining the physical mechanism of preform processing. The influence weight and sensitivity coefficient of each physical parameter on the quality indicators were clarified, and a quantitative mapping relationship was formed. Information on physical field parameters, mass index parameters, and model parameters is categorized and organized. Parameter information is stored in a structured format, a parameter retrieval index is established, and a mechanistic parameter dictionary is constructed. Based on the multiphysics coupling mechanism and production practice experience, physical constraint rules, process constraint rules and threshold constraint rules are extracted, logical expressions are used to clarify the rule triggering conditions and constraint results, a rule priority ranking mechanism is established, and a constraint rule library is formulated. Under normal and abnormal operating conditions, the predicted values ​​of physical parameters of the coupled model are compared with the actual sensor detection values. The relative error is calculated and the model parameters are adjusted when the error exceeds the limit. The predicted values ​​of the quality indicators of the mapped model are compared with the actual detection results. The mapping relationship is optimized through residual analysis. Based on the verification results, the threshold parameters and logical expressions in the constraint rule base are corrected to complete the model verification and calibration.

3. The method for monitoring the preform processing of rods using big data as described in claim 2, characterized in that: The aforementioned subfield establishes a basic physical model, simplifies the flow field control equations based on the Navier-Stokes equations and the geometric characteristics of the reaction chamber, introduces a turbulence model and sets flow field boundary conditions, discretizes the equations using the finite volume method and initializes the calculation parameters, incorporates a radiation heat transfer model and a convective heat transfer coefficient calculation model based on the Fourier heat conduction equation, clarifies the temperature field boundary conditions to quantify the spatiotemporal evolution of the temperature field, and determines the kinetic parameters and establishes a quantitative relationship between the reaction rate and temperature and reactant concentration based on the Arrhenius equation and the deposition / sintering reaction, including: Geometric features of the actual structure of the reaction chamber were extracted. After removing non-critical structures, the chamber, inlet nozzle, outlet, and preform substrate were retained to construct a simplified three-dimensional geometric model. The computational grid was divided according to the flow field change gradient, and the grid was refined in the flow field change regions at the nozzle outlet and the preform surface. The flow field control equations were established based on the Navier-Stokes equations for incompressible fluids. Secondary volume force terms were ignored and dimensionless simplification was performed in combination with the gas phase flow conditions of the reaction chamber. The k-ε turbulence model closed equation set was introduced, and the turbulent kinetic energy and dissipation rate transport equations were defined. The empirical coefficients of the turbulence model were calibrated to match the preform processing conditions. The flow field boundary conditions were set. The inlet was set as a velocity inlet boundary and the process gas velocity and component ratio were input. The outlet was set as a pressure outlet boundary and matched with the process negative pressure of the chamber. The inner wall of the reaction chamber and the surface of the preform were set as non-slip solid wall boundaries and the wall roughness parameters were given. The simplified flow field control equations were spatially discretized using the finite volume method. The discretization format was selected, the flow field velocity, pressure, and temperature calculation parameters were initialized, and the initial iteration convergence conditions were set. The temperature field control equation is established based on the Fourier anisotropic heat conduction equation. For high-temperature processing conditions, the Rosseland approximate radiation heat transfer model is incorporated, coupled with the flow field-driven convective heat transfer coefficient model, and the local flow velocity and fluid property parameters are substituted into the real-time solution of the heat transfer coefficient to form the temperature field control equation coupled with conduction, convection and radiation. The temperature field boundary conditions are set, the heating element area is set as a constant heat flux density or temperature boundary to match the heating power heat source input, the outer wall of the reaction chamber is set as a natural convection heat dissipation boundary and given the ambient temperature and comprehensive heat dissipation coefficient, and the interface between the preform and the cavity is set as a coupled boundary with continuous temperature and heat flux density. The velocity distribution solved by the flow field is used as the temperature field input, coupled with the convective heat transfer effect, and the temperature field control equation is solved numerically to obtain the spatiotemporal temperature distribution of the entire reaction chamber and preform, and output the temperature value and temperature gradient at different times at each location. The core gas-phase reactions in the preform deposition and sintering process were analyzed to determine the reaction equations and the variation patterns of reactant and product component concentrations. Based on the Arrhenius equation, a chemical reaction rate equation was constructed to determine the reaction order. Key kinetic parameters such as reaction activation energy and frequency factor were obtained through a combination of literature review and experimental calibration, and parameter deviations under high-temperature conditions were corrected. Real-time temperature and reactant concentration transported by the flow field were substituted into the Arrhenius equation as input variables to establish a quantitative relationship between reaction rate and temperature and reactant concentration. At the same time, the dynamic changes in reactant consumption and product formation were combined to construct a computational relationship coupling component transport and reaction kinetics.

4. The method for monitoring the preform processing of rods using big data as described in claim 2, characterized in that: The construction of a multi-field coupled solution system involves determining the coupling order of the flow field, temperature field, and chemical reaction field, defining multi-field coupling terms, using a partitioned iterative algorithm to solve the coupling equations and setting an iterative convergence criterion, and optimizing the computational grid density through grid independence verification, including: The closed-loop coupling order of the flow field, temperature field, and chemical reaction field is determined by first solving the flow field distribution, then solving the temperature field with the flow field velocity field as input, calculating the chemical reaction rate based on the temperature field results, and then feeding back the reaction heat generated by the chemical reaction to the temperature field for correction, thus forming a closed-loop coupling logic of the flow field, temperature field, chemical reaction field, and temperature field. Define multiple field coupling terms, where the coupling term between the flow field and the temperature field is the convective transport term of the flow field on temperature, the coupling term between the temperature field and the chemical reaction field is the effect term of temperature on reaction rate, the coupling term between the chemical reaction field and the temperature field is the source term of reaction heat on temperature field, and the coupling term between the flow field and the chemical reaction field is the transport term of the flow field on reactants / products. Clarify the mathematical expressions and calculation logic of each coupling term. The solution domain is divided into the inlet zone, reaction zone, outlet zone and preformed rod matrix zone of the reaction chamber. The control equations of each field are solved zone by zone using a partitioned iterative algorithm. After the solution of each region is completed, the calculation result of that region is used as the input parameter of the adjacent region. The results are iteratively passed on to achieve the coordinated solution of the field parameters of the whole watershed. Set iteration convergence criterion, the convergence criterion includes physical parameter convergence criterion and residual convergence criterion, the physical parameter convergence criterion is that the relative variation of flow field velocity, temperature, reaction rate key parameter of adjacent iteration step is all less than 10 -5 , the residual convergence criterion is that the numerical residual of each field control equation is reduced to 10 -6 Below, any one convergence criterion is met, and the iteration calculation of the current working condition is stopped; Mesh independence verification was carried out, and multiple sets of computational meshes with gradient distribution of mesh number were generated. Multi-field coupling solutions were performed under the same process conditions to extract key physical parameters such as nozzle outlet velocity, preform surface temperature gradient, and peak reaction rate in the reaction zone. The relative error of key parameters under different mesh densities was calculated. When the relative error of parameters is less than 3% after increasing the mesh density, the mesh density is determined to be the optimal computational mesh density.

5. The method for monitoring the preform processing of rods using big data as described in claim 1, characterized in that: The process involves preprocessing the raw sensor data, deriving mechanism-derived features based on the multi-physics coupling mechanism model, and fusing traditional statistical features with mechanism-derived features to construct a multi-dimensional feature matrix, including: Raw time-series sensor data on temperature, pressure, gas flow, and transmission speed during the preform processing are collected. Data intervals are divided into stages according to the deposition, sintering, and wire drawing processes. Null data caused by sensor offline or communication interruption are removed. Pulse-type and step-type outliers are identified using the 3σ principle combined with the local anomaly factor algorithm. Isolated outliers are repaired using cubic spline interpolation. The theoretical parameter intervals output by the multiphysics coupling mechanism model are constrained and filled for continuous outlier segments to ensure data integrity. The mean drift of the data was calculated using a sliding time window with a window length of 100 sampling points. A drift correction model was established by combining the normal fluctuation range of each parameter determined by the multiphysics coupling mechanism model. For temperature data, an adaptive Kalman filter algorithm was used to dynamically correct the drift deviation. For pressure and flow data, cross-correction was performed based on the parameter correlation derived from the mechanism model. Based on the flow field transport model, the velocity field uniformity coefficient, vortex intensity variance, and streamline deflection angle are calculated. Based on the temperature field model, the temperature gradient standard deviation, local heat flux density, and temperature field spatiotemporal change rate are derived. Combined with the chemical reaction kinetics model, the reaction rate fluctuation coefficient, reactant consumption rate gradient, and reaction heat release intensity are calculated. Traditional statistical features are extracted from the preprocessed raw sensor data. Time-domain features such as mean, variance, peak value, valley value, median, range, skewness, kurtosis, moving average, and autocorrelation coefficient are calculated. The time-series data is converted to the frequency domain through fast Fourier transform. Frequency-domain features such as dominant frequency amplitude, spectral energy ratio, harmonic distortion rate, and frequency band center frequency are extracted to form a set of traditional statistical features covering the time and frequency domains. The mechanism-derived features and traditional statistical features are standardized. The traditional statistical features are normalized using the Z-score standardization method, and the mechanism-derived features are mapped to the [0,1] interval using min-max normalization to ensure the consistency of the feature value range. Variance inflation factor analysis was used to screen features, eliminating highly redundant features with a VIF greater than 10, and retaining features that are strongly correlated with the quality indicators of the preform. The screened traditional statistical features and mechanism-derived features were aligned along the time dimension. A multi-dimensional feature matrix was constructed with a sliding time window of 50 sampling points and a step size of 25 sampling points as a sample, and rows representing samples and columns representing features. The physical meaning and source of each feature were labeled.

6. The method for monitoring the preform processing of rods using big data as described in claim 5, characterized in that: The method employs a sliding time window with a window length of 100 sampling points to calculate the mean drift of the data. A drift correction model is established based on the normal fluctuation range of each parameter determined by the multiphysics coupling mechanism model. For temperature data, an adaptive Kalman filter algorithm is used to dynamically correct the drift deviation. For pressure and flow data, cross-correction is performed based on the parameter correlations derived from the mechanism model, including: Configure a sliding time window with a window length of 100 sampling points and a step size of 50 sampling points. It iterates through the original time-series data of temperature, pressure, gas flow, and transmission speed from the sensors, calculates the average value of the data window by window, and accumulates the overall average value drift by the difference between the average values ​​of adjacent windows. At the same time, it records the fluctuation amplitude and trend change of the data in each window. The theoretical value range of each parameter output by the multiphysics coupling mechanism model is retrieved, and the normal fluctuation range of temperature, pressure and gas flow rate parameters is determined in combination with the preform processing specifications. The normal fluctuation range is used as the drift judgment threshold. When the calculated mean drift exceeds the drift judgment threshold, the data is judged to have drift. A drift feature library containing drift amount, drift direction and parameter type is established. An adaptive Kalman filter model is constructed for temperature data. The state equation is defined as the dynamic change model of temperature drift, and the observation equation is the difference equation between the sensor-measured temperature and the temperature predicted by the mechanism model. The process noise covariance and the observation noise covariance are initialized. The covariance matrix is ​​dynamically adjusted based on the drift amount in the drift feature library. The drift deviation of the temperature data is corrected in real time through the prediction-update iteration process of Kalman filtering, and the corrected temperature time series data is output. Based on the multiphysics coupling mechanism model, the quantitative correlation between pressure and flow parameters is derived by combining relevant physical models such as flow field transport and ideal gas state. The constant parameters in the correlation equation are calibrated by the mechanism model, and the correlation between other relevant pressure and flow parameters is extended to form a parameter correlation matrix. The pressure data is cross-calibrated, and the flow parameters that are strongly correlated with the target pressure parameter are selected. The calibrated flow data is substituted into the corresponding correlation equation to calculate the theoretical value of the target pressure. The theoretical value is then weighted and fused with the sensor measured pressure value according to the set weights, and the calibrated pressure data is output. Cross-calibration is performed on flow data. Pressure parameters that are strongly correlated with the target flow parameters are selected and substituted into the corresponding correlation equation to calculate the theoretical flow value. The theoretical flow value and the actual flow value measured by the sensor are weighted and fused according to the set weight to complete the flow data drift correction. If the deviation between the measured value and the theoretical value exceeds the allowable range of the mechanism model during the correction process, an abnormal alarm is triggered and the correlation equation parameters are re-verified. The validity of all corrected data is verified by comparing the error index of the data before and after correction with the predicted values ​​of the mechanism model, and verifying whether the data drift after correction is within the preset normal fluctuation range. If the accuracy requirements are not met, the correction steps are iteratively executed until the standard is met.

7. The method for monitoring the preform processing of rods using big data as described in claim 1, characterized in that: The aforementioned data-driven fusion architecture, constrained by mechanistic principles, uses the theoretical parameter range output by the mechanistic model as a regularization constraint term for the data-driven model. It incorporates multi-scenario mechanistic knowledge and combines it with a digital twin to achieve virtual-real data fusion training, including: The mechanistic constraints are quantified and transformed into regularized forms. The theoretical value ranges of key parameters such as flow field velocity, temperature gradient, and reaction rate output from the multi-physics coupling mechanism model are retrieved. The upper and lower boundary thresholds of each parameter are determined in conjunction with the process specifications. The interval constraints are transformed into regularization terms of the data-driven model. The constraint function is constructed using weighted L2 regularization. The weight coefficients are set according to the sensitivity coefficients of the parameters to the preform quality indicators, so that the parameter prediction values ​​during model training are kept within the allowable range of the mechanism. A multi-scenario mechanism knowledge graph is constructed, and the mechanism rules corresponding to different combinations of process parameters and raw material batches are sorted out. The reaction kinetic parameter range, flow field-temperature field coupling law, and quality indicator sensitive parameter set are stored in a structured manner with rule nodes and associated edges to form a multi-scenario mechanism knowledge graph. A digital twin for preform processing is constructed. Based on the three-dimensional geometric model of the reaction chamber, the multi-physics coupling mechanism model, and the equipment operating parameters, a digital twin is constructed that is mapped 1:1 to the physical production system. Through the Industrial Internet protocol, the twin and the physical equipment are synchronized in real time, dynamically replicating the spatiotemporal evolution of the flow field, temperature field, and chemical reaction field and generating high-fidelity virtual data. Spatiotemporal alignment and consistency verification are performed on the actual production data collected by physical sensors and the twin virtual data. A timestamp synchronization algorithm is used to complete the spatiotemporal dimension alignment and unify the sampling frequency and data format. The relative error between virtual and real data is calculated to achieve verification. For virtual data with excessive deviation, the twin boundary conditions and model parameters are corrected in reverse based on the physical sensor data. A multi-scenario mechanism for knowledge transfer is constructed. Based on domain adaptive transfer learning, the multi-scenario mechanism rules in the mechanism knowledge graph are transformed into prior knowledge for model training. A knowledge attention mechanism is introduced into the embedding layer of the CNN-LSTM data-driven model to assign weights to mechanism rules in different scenarios. The design of the virtual-real data fusion training strategy divides the training dataset into virtual data subset, real data subset and virtual-real fusion data subset. The two-stage training mode of pre-training-fine-tuning is adopted. The first stage is based on the virtual data subset and the mechanism knowledge graph for pre-training. The second stage divides the virtual-real fusion data subset into training set and validation set according to the proportion for model fine-tuning and dynamically adjusts the weight coefficient of the regularization constraint term. An embedded dynamic constraint adjustment mechanism is used to calculate the deviation between the model's predicted value and the mechanism interval in real time through a sliding window. The regularization constraint weights are adaptively adjusted according to the magnitude of the deviation, and the corresponding mechanism constraint parameters are called in combination with the scene recognition results of the multi-scene knowledge graph. Establish a fusion model training monitoring and evaluation system, set key evaluation indicators such as prediction accuracy, cross-scenario generalization error, and mechanism constraint satisfaction rate, monitor indicator changes in real time, and automatically trigger multi-scenario knowledge graph rule updates and virtual and real data supplementary training when the generalization error exceeds the standard.

8. The method for monitoring the preform processing of rods using big data as described in claim 1, characterized in that: The construction of a multi-dimensional anomaly feature system covering all types of anomalies employs signal frequency band decomposition and feature enhancement techniques to improve the ability to perceive weak signals corresponding to early faults and minor defects, including: The types of anomalies throughout the preform manufacturing process were analyzed, the temporal manifestations of pulse-type, step-type, and pattern-type anomalies were identified, the characteristic manifestations of micron-level bright spot defects, weak flow field disturbances, and early faults with slowly changing parameters were defined, and the correspondence between anomaly types and characteristic manifestations was established. Based on the correspondence between the anomaly types and their characteristics, a multi-dimensional anomaly feature framework is constructed, extracting statistical features, temporal features, spatial features, and multi-physics coupling features. At the same time, features derived from flow field uniformity and temperature gradient variance mechanism are incorporated to form a basic feature set for the full-dimensional representation of anomalies. Wavelet packet transform is used to decompose the time-series signals of temperature, pressure and gas flow sensors into frequency bands. The number of decomposition layers is set according to the frequency characteristics of abnormal signals. The differences in energy distribution of each frequency band under normal and abnormal operating conditions are compared to screen the sensitive frequency bands corresponding to early faults and minor defects and eliminate noise interference frequency bands. Feature enhancement is performed on the sensitive frequency band sub-signals. Adaptive weights are assigned to the sensitive frequency bands through the channel attention mechanism. The time-series differential amplification algorithm is used to extract subtle changes in the signal and to perform local enhancement processing on the spatial features of micron-level defects. The enhanced multidimensional anomaly features are standardized, and redundant features are removed by variance inflation factor analysis and Pearson correlation analysis, while features that are highly sensitive to various anomalies, early failures, and minor defects are retained. The filtered features are categorized and integrated according to anomaly type, and feature subsets are matched for each anomaly to form a multi-dimensional anomaly feature system covering all types of anomalies. The feature system was validated using actual production anomaly samples. The feature recognition accuracy of early faults and minor defects was tested. Based on the validation results, the number of wavelet packet decomposition layers and attention weight parameters were optimized, and the feature system was iteratively updated.

9. The method for monitoring the preform processing of rods using big data as described in claim 1, characterized in that: The multimodal anomaly classification model with an embedded causal inference module is constructed, and training is performed using learning methods and loss functions. It distinguishes between true anomalies and false anomalies caused by environmental interference. An adaptive threshold dynamic optimization mechanism based on mechanistic constraints and real-time data is established, and thresholds are set and iteratively optimized according to anomaly type, including: Design the basic architecture of a multimodal anomaly classification model, clarify the input forms of statistical features, temporal features, spatial features and multi-physics coupling features, use CNN to extract spatial features, LSTM to extract temporal features, fully connected layers to process coupling and statistical features, and introduce a modal attention mechanism to weight and fuse the features of each modality to generate a multimodal fusion feature vector; An embedded causal inference module is used to construct an anomaly causal graph based on the multi-physics coupling mechanism and production practice. The causal relationship between the true anomaly driving factors, environmental interference factors and anomaly results is clarified. The causal structure parameters are learned by using a structural causal model. The causal features and interference features in the fusion features are separated by adjusting the formula through a backdoor, and the causal contribution of each feature is calculated. The model is trained using a two-stage learning strategy of pre-training and fine-tuning. In the pre-training stage, the parameters are initialized with labeled normal working conditions, true anomalies, and false anomalies. In the fine-tuning stage, causal constraint optimization is introduced, and a composite loss function is designed, which includes weighted cross-entropy loss, causal consistency loss, and mechanism constraint loss. The weights of the three are determined by grid search, and the model is iteratively trained with the goal of minimizing the total loss. A mechanism for distinguishing between true and false anomalies is constructed. A threshold for causal contribution is set. When the sum of the causal contributions of the true anomaly driving factors exceeds the threshold and the contribution of the interference factors is lower than the set value, it is determined to be a true anomaly. If the contribution of the interference factors is dominant, it is determined to be a false anomaly caused by environmental interference. The distinction results are recorded for subsequent optimization. The theoretical range of key features of each anomaly type output by the multiphysics coupling mechanism model is retrieved, and the initial threshold range is determined in combination with the process specifications. The threshold is adjusted according to the sensitivity coefficient of the anomaly to the quality, with high-sensitivity anomalies having higher priority than low-sensitivity anomalies. The distribution parameters of real-time multimodal features are statistically analyzed using a sliding time window. The optimization targets are anomaly identification accuracy, false alarm rate, and false negative rate. The thresholds for each anomaly type are dynamically adjusted using a Bayesian optimization algorithm. A narrow threshold range is set for high-sensitivity anomalies, and a wide threshold range is set for low-sensitivity anomalies to achieve differentiated settings. Collect anomaly identification results during model operation, calculate threshold fit scores for each anomaly type, expand the threshold range if the false positive rate exceeds the standard, and narrow the threshold range if the false negative rate exceeds the standard. Combine the update of mechanism model parameters to synchronously correct the threshold range, and ensure that the threshold is dynamically matched with the working conditions and mechanism constraints. The model was tested using actual production samples containing true anomalies, false anomalies, and early weak anomalies. The accuracy of distinguishing between true and false anomalies and the precision of each anomaly identification were evaluated. The hyperparameters of the cause-effect graph structure, loss function weights, and thresholds were adjusted and the model and threshold mechanism were iteratively optimized until the preset accuracy requirements were met.

10. A preform processing monitoring system incorporating big data, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the preform processing monitoring method incorporating big data as described in any one of claims 1 to 9 by executing the machine-executable instructions.

Images