Large model driven data governance decision support system

Abstract

The invention relates to a large-model-driven data governance decision support system, which belongs to the technical field of intelligent manufacturing and industrial automation control, and comprises a physical field evolution modeling module for establishing a real-time global stress field based on a collected multi-modal state data set and outputting a time sequence stress field data set; the dynamic drift evaluation module is used for receiving the time-sequenced stress field data set, comparing the time-sequenced stress field data set with a preset reference stress field data set and calculating a drift risk level; the defect evolution prediction module is used for inputting the time sequence stress field data set and the collected defect correlation characteristic data set into a fusion prediction model to generate a latent defect risk index; and the closed-loop self-adaptive control module is used for generating self-adaptive control gain by integrating the drift risk level and the latent defect risk index, calculating the deviation between the time sequence and the reference stress field data set, and generating a control instruction for adjusting production parameters based on the gain and the deviation. According to the method, the crossing from discrete sensing data to a continuous dynamic physical field model is realized, and a high-fidelity data basis is provided for decision making.

Description

Technical Field

[0001] The present invention relates to the fields of intelligent manufacturing and industrial automation control, and specifically to a large model-driven data governance decision support system. Background Art

[0002] In the production process of high-end manufacturing, traditional quality control and process monitoring methods mainly rely on discrete, single-point sensor data and offline material sampling and testing. They have problems with insufficient coverage and poor real-time performance in reflecting the overall picture of production. They are unable to construct a panoramic view of the internal stress state of the material during the production process, and it is difficult to capture its dynamic process of changes over time and space.

[0003] This situation results in quality control measures being inherently lagging and reactive. Managers are unable to effectively distinguish between temporary fluctuations in the production process and systematic drifts that may cause product failures, and they lack the ability to predict future defects. This means that production adjustments are often made only after problems occur, making it difficult to provide scientific preventive control strategies.

[0004] The reasons for the above status quo and shortcomings are the limitations of data collection methods and analysis models. The limited sensor points on the production line lead to insufficient spatial resolution of the data, and the offline and lagging detection methods cannot meet the needs of real-time monitoring. At the same time, the system lacks a comprehensive understanding and modeling of the coupling effects of multiple physical fields, and is unable to integrate scattered measurement data into a continuous and dynamic physical field.

[0005] As a result, when potential quality risks arise in the production process, the system is unable to provide early warning or accurately locate the root cause of the problem, resulting in unnecessary waste and unable to achieve forward-looking, refined closed-loop control of product quality and process stability.

[0006] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Summary of the Invention

[0007] The purpose of the present invention is to provide a large model-driven data governance decision support system to solve the problems raised in the above background technology.

[0008] The technical solution of the present invention comprises: a physical field evolution modeling module, which collects a multimodal state data set, establishes a real-time global stress field based on the multimodal state data set, and outputs a time-series stress field data set;

[0009] A dynamic drift assessment module receives the time-series stress field data set, compares it with a preset reference stress field data set, and calculates a drift risk level;

[0010] A defect evolution prediction module collects a defect-related feature dataset, uses the time-series stress field dataset and the defect-related feature dataset as inputs of a fusion prediction model, and generates a latent defect risk index;

[0011] The closed-loop adaptive control module integrates the drift risk level and the latent defect risk index to generate an adaptive control gain, calculates the deviation between the time-series stress field dataset and the reference stress field dataset, and generates a control instruction for adjusting the production parameters based on the adaptive control gain and the deviation.

[0012] Preferably, the multimodal state data set includes: polarization light delay data, material thickness data, surface temperature data and material transmission speed data;

[0013] The physical field evolution modeling module reconstructs the mechanical stress field based on the polarization light delay data, the material thickness data and the surface temperature data;

[0014] The physical field evolution modeling module calculates the thermal stress field based on the surface temperature data;

[0015] The real-time global stress field is formed by superposition of the mechanical stress field and the thermal stress field.

[0016] Preferably, the physical field evolution modeling module further includes:

[0017] Based on the material transport velocity data, calculating the advection term describing the transport effect of the stress field along with the material movement;

[0018] Simulate the internal stress relaxation of the material and obtain the diffusion term;

[0019] Based on the difference between the currently calculated new stress and the stress field at the previous moment, the source term used to drive the model update is set;

[0020] The physical field evolution modeling module establishes a stress field dynamic evolution equation based on the advection term, the diffusion term, and the source term;

[0021] The physical field evolution modeling module updates the real-time global stress field based on the stress field dynamic evolution equation.

[0022] Preferably, the dynamic drift assessment module includes:

[0023] Calculating the difference between the time-series stress field dataset and the reference stress field dataset to generate a real-time deviation;

[0024] calculating a relative drift index based on the real-time deviation and the reference stress field dataset;

[0025] Calculating an absolute overrun index based on the real-time deviation and a preset maximum allowable stress deviation;

[0026] The determination of the drift risk level depends on the relative drift index and the absolute overlimit index.

[0027] Preferably, the dynamic drift assessment module further includes:

[0028] comparing the absolute exceedance index to a quality compliance threshold;

[0029] comparing the relative drift index to a process stability warning threshold;

[0030] When the absolute over-limit index is greater than the quality compliance threshold, the drift risk level is determined to be a quality over-limit warning;

[0031] When the absolute over-limit index is not greater than the quality compliance threshold, and the relative drift index is greater than the process stability warning threshold, the drift risk level is determined to be a process instability warning;

[0032] When the absolute over-limit index is not greater than the quality compliance threshold, and the relative drift index is not greater than the process stability warning threshold, the drift risk level is determined to be qualified.

[0033] Preferably, the defect-related feature data set includes: acoustic emission signal feature data acquired by an acoustic sensor array, and surface visual feature data acquired by a line scan camera;

[0034] The latent defect risk indicators include: a stress gradient risk factor and a latent defect index.

[0035] Preferably, the defect evolution prediction module further includes:

[0036] Calculating the stress gradient risk factor based on the spatial gradient and temporal change rate of the time-series stress field data set;

[0037] The defect evolution prediction module uses the time-series stress field dataset, the acoustic emission signal feature data, the surface visual feature data, and the stress gradient risk factor as inputs of the fusion prediction model to output the latent defect index.

[0038] Preferably, the closed-loop adaptive control module includes:

[0039] Inputting the deviation between the time-series stress field data set and the reference stress field data set into an inverse model to generate a control target error vector;

[0040] Establishing a decision logic for mapping the drift risk level and the latent defect risk index to the adaptive control gain;

[0041] The closed-loop adaptive control module generates the adaptive control gain based on the decision logic.

[0042] Preferably, the closed-loop adaptive control module further includes:

[0043] multiplying the control target error vector by the adaptive control gain to generate the control instruction;

[0044] The control instruction is a final control adjustment vector.

[0045] Preferably, the closed-loop adaptive control module further includes:

[0046] Adding the final control adjustment vector to the set value vector of the current production parameter to generate an updated set value vector;

[0047] The closed-loop adaptive control module sends the updated set value vector to the production line execution unit to achieve a closed-loop control.

[0048] The present invention provides a large-model driven data governance decision support system through improvements. Compared with the existing technology, it has the following improvements and advantages:

[0049] 1. A significant leap forward has been achieved from discrete sensor data to continuous dynamic physical field models, enabling managers to gain a global and dynamic understanding of the internal state of materials during the production process for the first time. This provides an unprecedented high-dimensional, high-fidelity data foundation for all subsequent decision-making.

[0050] 2. This invention can output clear drift risk levels, such as process instability warnings or quality out-of-limit warnings. This differentiation enables the system to take appropriate measures for different risks: gentle preventative fine-tuning for the former and decisive corrective intervention for the latter, thereby greatly improving control accuracy and efficiency.

[0051] 3. It can issue early warnings at the incipient stage of defects or even before they occur, achieving true predictive quality assurance;

[0052] 4. The entire production system is intelligent and adaptive, and can respond to various production disturbances in the most efficient and robust manner, thereby improving the stability of product quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The present invention will be further explained below in conjunction with the accompanying drawings and examples:

[0054] Figure 1It is a flow chart of the system of the present invention. DETAILED DESCRIPTION

[0055] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to specific embodiments.

[0056] Example 1:

[0057] See also Figure 1 The present invention provides a technical solution for a data governance decision support system driven by a large model, including: a physical field evolution modeling module, which collects a multimodal state data set, establishes a real-time global stress field based on the multimodal state data set, and outputs a time-series stress field data set; a dynamic drift assessment module, which receives the time-series stress field data set, compares it with a preset reference stress field data set, and calculates the drift risk level; a defect evolution prediction module, which collects a defect-related feature data set, uses the time-series stress field data set and the defect-related feature data set as inputs of a fusion prediction model, and generates a latent defect risk index; a closed-loop adaptive control module, which integrates the drift risk level and the latent defect risk index to generate an adaptive control gain, calculates the deviation between the time-series stress field data set and the reference stress field data set, and generates a control instruction for adjusting the production parameters based on the adaptive control gain and the deviation;

[0058] In this embodiment, a large-model-driven data governance decision support system, through the serialized collaborative operation of its internal physical field evolution modeling module, dynamic drift assessment module, defect evolution prediction module, and closed-loop adaptive control module, builds a complete closed loop from data perception to decision execution. The system converts the discrete multimodal sensor data on the production line into a continuous, dynamic, and visual global stress field time series data stream, which not only accurately reflects the physical state of the material at any moment, but also deeply reveals the evolution trend of stress along the production line flow;

[0059] The system quantitatively compares the real-time stress field with the ideal state of the golden batch, achieving a two-dimensional dynamic assessment of production process stability and product quality compliance. The system integrates multi-source information such as internal stress, acoustic emission and surface vision, achieving a technological leap from passive detection to active prediction, and can predict hidden defects that may form at a certain point in the future. All evaluation and prediction analysis results are converted into precise, quantitative control instructions that fully match the risk level, automatically adjusting production parameters, thereby achieving forward-looking, refined and automated closed-loop control of product quality, significantly improving the yield and process stability in the high-end manufacturing field.

[0060] Example 2

[0061] The multimodal state data set includes polarization delay data, material thickness data, surface temperature data, and material transmission speed data. The physical field evolution modeling module reconstructs the mechanical stress field based on the polarization delay data, material thickness data, and surface temperature data. The physical field evolution modeling module calculates the thermal stress field based on the surface temperature data. The real-time global stress field is the superposition of the mechanical stress field and the thermal stress field.

[0062] The physical field evolution modeling module also includes: calculating the advection term used to describe the transport effect of the stress field with material movement based on material transmission velocity data; simulating the internal stress relaxation of the material to obtain the diffusion term; setting the source term used to drive the model update based on the difference between the current calculated new stress and the stress field at the previous moment; the physical field evolution modeling module establishes the dynamic evolution equation of the stress field based on the advection term, diffusion term and source term; the physical field evolution modeling module updates the real-time global stress field based on the dynamic evolution equation of the stress field;

[0063] In a specific implementation, the physical field evolution modeling module serves as the perception and modeling core of the system. It constructs a dynamic stress field evolution model coupled with thermal effects, decomposes the total stress field into the sum of mechanical stress and thermal stress, and accurately describes the dynamic changes of the stress field with material movement through a comprehensive evolution equation.

[0064] Calculation of thermal stress fields, aimed at quantifying internal stresses caused by temperature changes; ;

[0065] This formula is derived from the theory of thermoelasticity. Its technical motivation is that temperature changes during the material production process are a key physical factor that leads to the generation and uneven distribution of internal stress. To accurately model the total stress field, the passively measured temperature field data is converted into stress field components with clear physical meaning, thereby incorporating thermal effects as calculable variables into the overall model.

[0066] in, Indicates the spatial position at time t Thermal stress tensor at ; Young's modulus represents the material's core intrinsic property, which characterizes the material's ability to resist elastic deformation. Its value is determined through preliminary material mechanical properties experiments such as tensile testing, or directly obtained from the material specifications provided by the supplier; The thermal expansion coefficient of the material is an inherent thermophysical property that measures the degree to which the material's dimensions change with temperature. Its value is determined by experimental methods such as thermomechanical analysis (TMA), or can be obtained by consulting authoritative material properties manuals. It is the stress-free reference temperature, which usually refers to the curing temperature of the material during molding or a set zero-stress reference process temperature. This is a key process setting parameter; It is collected in real time by the non-contact infrared sensor array and is located at the spatial position at time t. The surface temperature of the material;

[0067] This formula is applied to the physical field evolution modeling module to convert the real-time collected surface temperature data Combined with known material constants, thermal stress components are directly calculated. This allows for precise quantification of thermally induced stresses, enabling the model to clearly distinguish and evaluate stress problems caused by uneven cooling or ambient temperature fluctuations, providing a key physical basis for subsequent stress analysis and defect tracing.

[0068] The core of tomographic reconstruction of mechanical stress fields lies in using optical measurement data to invert the mechanical stress inside the material. This technical solution is based on the principle of photoelastic effect and the idea of ​​computed tomography. Its technical motivation is that it is extremely difficult to directly measure the internal stress field, but the birefringence field, which is proportional to the stress field, can be measured non-contactly along a specific path (i.e., the polarization light delay) by optical means. Through the tomography reconstruction algorithm, the internal birefringence field distribution is inverted using the measured multi-angle polarization light delay data. According to the law of photoelasticity, the reconstructed birefringence field is converted into a mechanical stress field.

[0069] ; ;

[0070] in, is the birefringence at time t and spatial position (x,y); represents the inverse operation operator of tomographic reconstruction such as filtered back projection; At time t, along the detection angle The original polarized light retardation data measured in the direction; the technical principle of this formula is that the collected multi-angle polarized light retardation data is regarded as the birefringence field inside the material Along different angles By applying the filtered back-projection algorithm to the projection data at each angle, the two-dimensional birefringence field can be reconstructed. After obtaining the birefringence field, the mechanical stress can be calculated according to the law of photoelasticity .in, is the temperature-dependent stress-optical coefficient, which characterizes the material's ability to convert mechanical stress into optical anisotropy and is related to temperature. The dependence of is determined by preliminary calibration experiments;

[0071] To determine the stress-optical coefficient and temperature The following calibration method can be used to determine the dependence of the material on the stress. The material sample is placed in a heating chamber with a transparent window and precise temperature control. A known, uniform uniaxial tensile stress is applied to the sample through an external device. , using the same light source and polarization detector as this system, measure the optical path delay through the sample , at multiple different temperature points Repeat this process, according to the basic formula of photoelasticity: ; ;

[0072] Calculate the coefficient values ​​at different temperatures, and finally get the function by curve fitting ;

[0073] In the system, this formula fuses multimodal data from the optical sensing system—polarization delay, thickness, and temperature—to reconstruct a high-resolution, two-dimensional mechanical stress field. This significant technical benefit is that it enables real-time, global, and visual monitoring of internal mechanical stresses in flexible material production lines. This transcends the limitations of traditional single-point or offline measurements and provides an unprecedentedly powerful tool for understanding and controlling stresses introduced by process steps like mechanical tension and rolling.

[0074] The stress field dynamic evolution equation, which is the culmination of the modules and describes the complete dynamics of the total stress field in time and space; ;

[0075] This formula is a physics-based transport equation, whose structure draws on the convection-diffusion equations in fluid mechanics and heat transfer. The technical motivation lies in the fact that static snapshots of stress fields are insufficient to fully describe the continuous production process. As materials flow through the production line, the internal stress field is both transported and relaxed due to changes in the internal structure, while new stresses are constantly generated. This equation aims to establish a mathematical framework that can uniformly describe these complex dynamic processes, thereby achieving a leap from stress field snapshots to dynamic representations of stress fields.

[0076] in, Represents the total stress field (Right now ) rate of change over time; is the advection term, which is used to describe the stress field as the material moves at a speed The transport effect caused by movement, Monitored by laser Doppler velocimetry, is the gradient operator; is a diffusion term, which is used to simulate the stress relaxation and diffusion phenomenon caused by viscoelasticity inside the material. is the Laplace operator, and D is the stress diffusion coefficient, which is a physical quantity that characterizes the speed of stress relaxation of a material. Its unit is m² / s. It is calibrated through targeted material experiments such as stress relaxation experiments, or is set empirically based on the viscoelastic model of the material.

[0077] Stress diffusion coefficient The calibration can be determined by the following experiment: a test area is selected on the production line, and a local, instantaneous stress concentration is created in the center of the area using external excitation; then, the physical field evolution modeling module of the present invention is used to continuously monitor the stress peak value over time. and space The dissipation process of the collected time-series stress field data Fitting with the analytical solution of the two-dimensional diffusion equation, the expression of the two-dimensional diffusion equation is:

[0078] Represents the change of stress field σ with time t, so as to solve the optimal diffusion coefficient D value;

[0079] is the source term, which is defined as: ;

[0080] Since σ_total(t) = σ_mech(t) + σ_thermal(t), this formula is equivalent to: ;

[0081] Represents the newly calculated stress field at the current moment Compared with the stress field at the previous moment The difference is the effective new information that drives the model to continuously update to the real state in a physical sense, ensuring that the model can keep up with the changes in production status in real time;

[0082] This equation is the computational core of the physical field evolution modeling module. It is based on the advection term, diffusion term, and source term, and dynamically updates the real-time global stress field of the entire production area through numerical solution, such as the finite element or finite difference method. The core technical effect is that it truly converts discrete multimodal sensor data into a continuous, dynamic, and visual dynamic representation of the global stress field. This dynamic representation not only reflects the current state, but also reflects the complete history and future trends of stress evolution along the production line, providing the highest-dimensional and most insightful data input for all subsequent evaluation and prediction modules.

[0083] Example 3

[0084] The dynamic drift assessment module includes: calculating the difference between the time-series stress field dataset and the reference stress field dataset to generate a real-time deviation; calculating the relative drift index based on the real-time deviation and the reference stress field dataset; calculating the absolute limit deviation index based on the real-time deviation and the preset maximum allowable stress deviation; determining the drift risk level based on the relative drift index and the absolute limit deviation index;

[0085] The dynamic drift assessment module further includes: comparing the absolute over-limit index with the quality compliance threshold; comparing the relative drift index with the process stability warning threshold; when the absolute over-limit index is greater than the quality compliance threshold, determining the drift risk level as a quality over-limit warning; when the absolute over-limit index is not greater than the quality compliance threshold and the relative drift index is greater than the process stability warning threshold, determining the drift risk level as a process instability warning; when the absolute over-limit index is not greater than the quality compliance threshold and the relative drift index is not greater than the process stability warning threshold, determining the drift risk level as qualified;

[0086] The dynamic drift assessment module of the technical solution in this embodiment is designed to conduct a comprehensive real-time assessment of the production status. It achieves dual monitoring of process stability and product quality compliance by calculating two core indicators: relative drift index and absolute overrun index. In this embodiment, in order to conduct a refined assessment of the material, the entire global stress field is discretized into multiple sub-regions, and the subscript i represents a specific sub-region.

[0087] Relative drift index Calculation of , used to evaluate the relative stability of the production process; ;

[0088] This formula originates from the idea of ​​measuring process variability in statistical process control. The technical motivation is that it is not enough to know the absolute value of the deviation. It is also necessary to understand the size of this deviation relative to the normal state. A small deviation for a high-stress area may be a significant fluctuation for a low-stress area. A standardized, dimensionless indicator is needed to objectively evaluate the relative degree of fluctuation of the process.

[0089] in, is the relative drift index of region i; The real-time total stress field Subtract the reference stress field The resulting real-time bias tensor; It is a preset reference stress field, which is set based on production data from a golden batch or ideal finite element simulation results; represents the Frobenius norm of the matrix, which calculates the square root of the sum of the squares of all elements of the matrix and can fully reflect the overall size of the stress tensor;

[0090] This index is continuously calculated in the dynamic drift assessment module to quantify the fluctuation of the current production status compared to the ideal state. The technical effect is to provide a sensitive quantitative indicator of process stability. In a stable process, this index should remain within a small range. A continued increase or sharp fluctuation in the index, even if it has not yet reached the quality red line, indicates that the process is becoming unstable. This provides a direct basis for the subsequent first-level warning - process instability, thereby achieving preventive early intervention.

[0091] Absolute Exceedance Index Calculation of the product quality control method to assess the absolute conformity of the product quality; ;

[0092] The design concept of this formula is directly derived from the concept of specification limits in quality management. The technical motivation is that any production process must meet the final product quality specifications. A direct and clear indicator is needed to measure whether the current stress deviation has approached or exceeded the maximum range allowed by the product specifications, thereby determining whether there is a quality risk in the product.

[0093] in, is the absolute overlimit index of region i; Same as the definition in the previous formula, it is the real-time bias tensor; It is the maximum stress deviation tensor allowed in region i according to the final product quality specification, which defines the allowable fluctuation boundary of the real-time stress relative to the reference datum;

[0094] This index is a decisive criterion for quality compliance. When the stress deviation in this area exceeds the product specification, it clearly means that the stress deviation in this area has exceeded the product specification and constitutes a quality defect. Its technical effect is to transform the abstract quality specification into a quantitative indicator that can be calculated and monitored in real time. It provides a decisive and unambiguous trigger condition for the second-level warning - quality limit violation, ensuring that any situation that exceeds the quality bottom line can be immediately identified and corrective intervention can be initiated;

[0095] The key tunable parameters in this module are , that is, the relative drift warning threshold. The setting logic of this threshold is clear and feasible: technicians in this field analyze the production data of batches with known qualified quality in history, especially the periods when normal process fluctuations occur in these batches, and calculate the relative drift warning threshold within these periods. The distribution of the exponential The threshold is set to a high percentile of the distribution, such as the 95th percentile. This ensures that it is sensitive enough to issue an early warning when abnormal fluctuations begin to appear in the process but are far from reaching the level that affects product quality, reflecting the preventive principle of risk management.

[0096] Example 4

[0097] The defect-related feature dataset includes: acoustic emission signal feature data acquired by the acoustic sensor array, and surface visual feature data acquired by the line scan camera; the latent defect risk indicators include: stress gradient risk factor and latent defect index;

[0098] The defect evolution prediction module also includes: calculating the stress gradient risk factor based on the spatial gradient and temporal change rate of the time-series stress field dataset; the defect evolution prediction module uses the time-series stress field dataset, acoustic emission signal feature data, surface visual feature data, and stress gradient risk factor as inputs to a fusion prediction model to output a latent defect index;

[0099] The core task of the defect evolution multimodal monitoring and prediction module in this embodiment is to upgrade from passive detection to active prediction. By integrating multi-source data and performing deep learning prediction, it can discover hidden defects that will form at a certain point in the future.

[0100] Stress gradient risk factor Calculations to quantify the risks posed by stress concentrations and their rate of deterioration; ;

[0101] This formula is based on the fundamental principle of fracture mechanics in material mechanics, which states that material failure often begins in areas of stress concentration, and that increasing stress concentration is a precursor to crack propagation. The technical motivation is that the stress field amplitude alone is insufficient to fully assess defect risk. The spatial gradient of the stress field represents the degree of stress concentration, while the time rate of change of the stress field amplitude reflects the dynamic trend of state deterioration. Combining the two allows for a more comprehensive risk assessment.

[0102] in, is the stress gradient risk factor of region i at time t; is the norm of the stress field gradient, which directly quantifies the stress concentration in the region; The time rate of change of the stress field norm (state deterioration rate) is calculated by the material transmission speed It is normalized and converted into the stress gradient along the direction of material movement, so that it can be coordinated with the spatial gradient of the first term in terms of physical dimension; and is the weight coefficient. This factor is a key input feature of the fusion prediction model. Its technical effect is to provide the model with deep insights into the deterioration trend of the stress state. It enables the model to not only see where the stress is high, but also to combine the areas with the most concentrated stress and the areas with the fastest stress growth, thereby more accurately locating the high-risk areas most likely to develop internal defects such as microcracks.

[0103] Weight coefficient and is a dimensionless empirical parameter whose determination depends on the specific material properties and failure mode; for brittle materials that are more susceptible to stress concentration under static loading, The weight of will be higher; for scenarios where dynamic fatigue accumulation is more critical, The weights of should be increased accordingly; the optimal values ​​of these weights are determined by hyperparameter optimization, specifically, to improve the fusion prediction model The final prediction accuracy on the historical defect sample data set is the optimization target, and grid search or Bayesian optimization methods are used to optimize the final prediction accuracy. and Optimization is performed to maximize the model's prediction accuracy. These two coefficients are also used to balance two items with different physical dimensions and combine them into a unified risk score.

[0104] Latent Defect Index The calculation of is the core output of this module and represents the comprehensive prediction confidence of future defects; ;

[0105] This formula represents a deep learning fusion prediction model. The technical motivation is that a single indicator is difficult to detect early defects caused by the combined effects of multiple factors. For example, internal microcracks may only manifest as weak acoustic emission signals and slight stress field changes. Any single signal may be overwhelmed by noise. A powerful nonlinear model must be used to integrate multiple dimensional evidence, such as internal stress, acoustics, and vision, to effectively identify such complex and latent defect patterns.

[0106] in, is the latent defect index of region i at time t, whose value range is normalized to between 0 and 1, representing the credibility of the defect occurrence; Represents a trained deep learning model, such as a temporal model based on the Transformer or Gated Recurrent Unit (GRU) architecture, which is trained to recognize complex correlation patterns in multimodal data.

[0107] A deep learning network including gated recurrent units and attention mechanism can be used; the specific architecture can be set as follows: the four input data sequences of the past N time steps: real-time stress field , Acoustic emission signal characteristics , surface visual features and stress gradient risk factors , respectively, are fed into four parallel GRU layers to extract their respective time-dependent features; these four feature streams are weightedly fused through a multi-head attention layer, so that the model can dynamically focus on the data source that contributes most to defect prediction at a specific moment, and the fused feature vector is passed through a fully connected layer and a Sigmoid activation function to output a latent defect index in the range of [0, 1]. The training data set of this model consists of historical production data. Each data sample contains a period of multimodal data of N consecutive time steps. The label is set according to whether a defect that has been manually verified appears within M time steps after this time period. For example, if a defect appears, it is 1, and if it does not appear, it is 0. Among them, N is the time step of the input sequence, which represents the length of the time window for the model to look back at historical data; M is the time step for predicting the future, which is used to define the training label. The values ​​of N and M can be determined through experimental optimization based on the characteristic time of the production process and the typical cycle of defect evolution.

[0108] The input to the model includes the real-time stress field from module 1 , the acoustic emission signal characteristics of the corresponding area i from the high-frequency acoustic sensor array , the surface micromorphology features of region i from the high-resolution line scan camera , and the stress gradient risk factor calculated by this module ;

[0109] The model analyzes the multimodal data sequence of the past N time steps and outputs the final judgment on the potential defect risk at the current moment. The technical effect is that it has achieved a leap from passive detection to active prediction. It can not only detect existing defects, but also predict hidden defects that will be formed at a certain point in the future, and quantify risk indicators. and As the core input, it is passed to the closed-loop adaptive control logic module, providing a solid foundation for triggering different levels of more targeted predictive control strategies.

[0110] Example 5

[0111] The closed-loop adaptive control module includes: inputting the deviation between the time-series stress field dataset and the reference stress field dataset into the inverse model to generate a control target error vector; establishing a decision logic that maps the drift risk level and the latent defect risk index into an adaptive control gain; and generating the adaptive control gain based on the decision logic.

[0112] The closed-loop adaptive control module further includes: multiplying the control target error vector by the adaptive control gain to generate a control instruction; wherein the control instruction is a final control adjustment vector;

[0113] The closed-loop adaptive control module further includes: adding the final control adjustment vector to the set value vector of the current production parameter to generate an updated set value vector; the closed-loop adaptive control module sends the updated set value vector to the production line execution unit to realize the control closed loop;

[0114] The production parameter closed-loop adaptive control logic module of the present invention serves as the decision-making and execution terminal of the entire system. Its core function is to integrate the evaluation and prediction results of all upstream modules, dynamically calculate the optimal control adjustment amount, and determine the adjustment intensity based on the risk level, ultimately achieving a truly adaptive closed-loop control.

[0115] Control target error vector The calculation is to determine the ideal adjustment direction and amplitude; ;

[0116] The core of this formula is the inverse model ; The technical motivation is to directly Deducing the production parameters that need to be adjusted is a complex, nonlinear inverse problem. By establishing an inverse physical model or training a neural network to fit this inverse relationship, the idealized production parameter adjustment required to completely eliminate the current stress deviation can be directly calculated.

[0117] in, It is a vector that contains the ideal adjustment values ​​for multiple production parameters such as tension, speed, temperature, etc. It is an inverse physical model or a pre-trained neural network whose input is the stress deviation and output is the parameter error that causes the deviation; is the real-time stress deviation at the current moment, provided by module 2;

[0118] Inverse Model It is preferably achieved by training a neural network; the method for generating training data is as follows: a forward physical simulation model consistent with the actual production line is constructed using finite element analysis software, which can be used to generate the actual production line according to the input production parameters. , accurate simulation output corresponding to the steady-state global stress field ; By using standard parameters in the gold batch A large number of systematic random perturbations are performed near the target, generating tens of thousands of parameter perturbation-stress deviation data pairs, i.e. ,in , ; Stress deviation As input to the neural network, the parameter perturbation that causes the deviation As output, train an inverse model such as a multilayer perceptron After training, the model can directly and quickly estimate the ideal parameter adjustment required to eliminate the real-time stress deviation. ;

[0119] This step is the first step in the control demand calculation. Its technical effect is that it directly translates the problem discovered by the upstream module, namely stress deviation, into a solution that the downstream module can understand, namely parameter adjustment, providing a clear goal and direction for the subsequent control instruction generation.

[0120] Adaptive gain matrix The generation of , whose purpose is to determine the aggressiveness of the control system response; ;

[0121] This formula originates from adaptive control theory. The technical motivation is that traditional fixed-gain controllers cannot cope with complex, time-varying operating conditions. An intelligent system should have a response strength that matches the risk level: for minor process fluctuations, smooth, small adjustments should be used to maintain stability; for serious quality violations or high-risk defect predictions, rapid and large-scale corrective interventions must be taken. It is the core of achieving this dynamic adjustment;

[0122] in, is a diagonal matrix with elements on the diagonal The adjustment gains corresponding to different production parameters; function is a decision logic, usually implemented by an implementable fuzzy logic rule set or rule engine, which combines the two drift indices from module 2 and two defect risk indicators from Module 3 These four risk indicators are used as input and mapped to specific gain values;

[0123] This function is the adaptive core of this module; the technical effect is that it builds a true adaptive closed-loop control system. The decision logic g operates based on a predefined set of rules. For example, if the input Greater than 1, corresponding to the second level warning, g is Assign higher values ​​to the diagonal elements of , such as in the range of 0.8 to 1.5; on the contrary, if Not greater than 1 but Exceed , corresponding to the first level warning, g is assigned a lower value, such as in the range of 0.2 to 0.5. This dynamic gain mechanism makes the control response both accurate and reasonable;

[0124] This logic adopts hierarchical priority and weighted combination, The maximum value of the four risk indicators in all regions i determines the gain matrix ;

[0125] Highest priority: If the latent defect index exceeds the critical value, e.g. Or the absolute overlimit index is greater than 1 ( ), the system determines that there is a serious quality risk. The diagonal elements of are set to a higher value range, such as [0.8, 1.5]; Second priority: If the first level condition is not met, but the stress gradient risk factor exceeds its warning threshold ( ) or the relative drift index exceeds its warning threshold ( ), the system determines that the process is unstable or there is a risk of potential defect evolution. The diagonal elements of are set to a lower value range, such as [0.2, 0.5]. Normal state: If none of the above conditions are met, a very small maintenance gain is set or no adjustment is made. Within the specified value range, the specific gain value can be proportional to the size of the risk indicator exceeding the threshold, for example: ;

[0126] in: Indicates that in all discrete sub-regions i, the absolute limit index is taken This means that the control system will determine the adjustment force based on the area with the most serious risk on the entire material, thereby achieving more refined adaptive control;

[0127] The setting can refer to the historical data, the batch that finally produces defects before the defects are formed The distribution of the factor, taking its specific quantile as the warning threshold;

[0128] Final control adjustment vector The calculation and execution of is the final formation and closed loop of control instructions;

[0129] ;

[0130] These two formulas are the basic execution logic of a standard digital control system. Their purpose is to combine the ideal adjustment amount with the adjustment strength determined by risk, generate the final executable instruction, update the system set value, and complete a control iteration.

[0131] in, It is the core output of this module, i.e. the final, executable control adjustment vector; is the set value vector of the current production parameters, is the new setting value after updating;

[0132] The calculation of the final control adjustment vector is the output link of this module, which combines the ideal adjustment direction and amplitude and the intensity of adjustments determined by the current risk level By multiplying them together, we get a final adjustment instruction that is both accurate and reasonable. This instruction is sent to each execution unit of the production line, such as motors and heaters, to update the system's set values. The completion of this process marks that the entire system has achieved a complete, intelligent, and adaptive control cycle from perception, modeling, evaluation, prediction to decision-making and execution.

[0133] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A large-model-driven data governance decision support system, characterized by: include: A physical field evolution modeling module collects a multimodal state data set, establishes a real-time global stress field based on the multimodal state data set, and outputs a time-series stress field data set; A dynamic drift assessment module receives the time-series stress field data set, compares it with a preset reference stress field data set, and calculates a drift risk level; A defect evolution prediction module collects a defect-related feature dataset, uses the time-series stress field dataset and the defect-related feature dataset as inputs of a fusion prediction model, and generates a latent defect risk index; The closed-loop adaptive control module integrates the drift risk level and the latent defect risk index to generate an adaptive control gain, calculates the deviation between the time-series stress field dataset and the reference stress field dataset, and generates a control instruction for adjusting the production parameters based on the adaptive control gain and the deviation.

2. The large model-driven data governance decision support system according to claim 1, characterized in that: The multimodal state data set includes: polarization light delay data, material thickness data, surface temperature data and material transmission speed data; The physical field evolution modeling module reconstructs the mechanical stress field based on the polarization light delay data, the material thickness data and the surface temperature data; The physical field evolution modeling module calculates the thermal stress field based on the surface temperature data; The real-time global stress field is formed by superposition of the mechanical stress field and the thermal stress field.

3. The large model-driven data governance decision support system according to claim 2, characterized in that: The physical field evolution modeling module also includes: Based on the material transport velocity data, calculating the advection term describing the transport effect of the stress field along with the material movement; Simulate the internal stress relaxation of the material and obtain the diffusion term; Based on the difference between the currently calculated new stress and the stress field at the previous moment, the source term used to drive the model update is set; The physical field evolution modeling module establishes a stress field dynamic evolution equation based on the advection term, the diffusion term, and the source term; The physical field evolution modeling module updates the real-time global stress field based on the stress field dynamic evolution equation.

4. The large model-driven data governance decision support system according to claim 1, characterized in that: The dynamic drift assessment module includes: Calculating the difference between the time-series stress field dataset and the reference stress field dataset to generate a real-time deviation; calculating a relative drift index based on the real-time deviation and the reference stress field dataset; Calculating an absolute overrun index based on the real-time deviation and a preset maximum allowable stress deviation; The determination of the drift risk level depends on the relative drift index and the absolute overlimit index.

5. The large model-driven data governance decision support system according to claim 4, characterized in that: The dynamic drift assessment module further includes: comparing the absolute exceedance index to a quality compliance threshold; comparing the relative drift index to a process stability warning threshold; When the absolute over-limit index is greater than the quality compliance threshold, the drift risk level is determined to be a quality over-limit warning; When the absolute over-limit index is not greater than the quality compliance threshold, and the relative drift index is greater than the process stability warning threshold, the drift risk level is determined to be a process instability warning; When the absolute over-limit index is not greater than the quality compliance threshold, and the relative drift index is not greater than the process stability warning threshold, the drift risk level is determined to be qualified.

6. The large model-driven data governance decision support system according to claim 1, characterized in that: The defect-related feature data set includes: acoustic emission signal feature data acquired by an acoustic sensor array, and surface visual feature data acquired by a line scan camera; The latent defect risk indicators include: a stress gradient risk factor and a latent defect index.

7. The large model-driven data governance decision support system according to claim 6, characterized in that: The defect evolution prediction module also includes: Calculating the stress gradient risk factor based on the spatial gradient and temporal change rate of the time-series stress field data set; The defect evolution prediction module uses the time-series stress field dataset, the acoustic emission signal feature data, the surface visual feature data, and the stress gradient risk factor as inputs of the fusion prediction model to output the latent defect index.

8. The large model-driven data governance decision support system according to claim 1, characterized in that: The closed-loop adaptive control module includes: Inputting the deviation between the time-series stress field data set and the reference stress field data set into an inverse model to generate a control target error vector; Establishing a decision logic for mapping the drift risk level and the latent defect risk index to the adaptive control gain; The closed-loop adaptive control module generates the adaptive control gain based on the decision logic.

9. The large model-driven data governance decision support system according to claim 8, characterized in that: The closed-loop adaptive control module also includes: multiplying the control target error vector by the adaptive control gain to generate the control instruction; The control instruction is a final control adjustment vector.

10. The large model-driven data governance decision support system according to claim 9, characterized in that: The closed-loop adaptive control module also includes: Adding the final control adjustment vector to the set value vector of the current production parameter to generate an updated set value vector; The closed-loop adaptive control module sends the updated set value vector to the production line execution unit to achieve a closed-loop control.

Images