Galvanized sheet production defect early warning method and system based on cluster analysis

Abstract

This invention discloses a defect early warning method and system for galvanized steel sheet production based on cluster analysis, belonging to the field of industrial automation control. It includes: constructing a quantitative causal model between processes and deriving a simulation evolution sample library; integrating real-time production data with the sample library and inputting it into a two-layer clustering model to obtain online clustering status; when the status deviates from the boundary of a high-quality cluster, using the causal model to calculate the shortest adjustment path and generate a collaborative control instruction set, thereby adjusting equipment parameters. This invention employs a deep integration of mechanistic causal model and two-layer clustering algorithm, enabling accurate identification and closed-loop correction of quality trends across processes, improving the defect early warning accuracy and process adaptive control capability of the galvanized steel production line.

Description

Technical Field

[0001] This invention relates to the field of industrial automation control, and in particular to a method and system for early warning of defects in galvanized steel sheet production based on cluster analysis. Background Technology

[0002] In continuous industrial production processes, such as galvanized sheet production lines, multiple complex processes are involved in their sequential and parallel execution, including cold rolling, annealing, galvanizing, and post-processing. Each process contains numerous controllable process parameters and corresponding quality indicators. These processes involve complex physicochemical reactions and dynamic time lags; even minor fluctuations or deviations in upstream processes, accumulated and transmitted over time, can lead to serious quality defects in downstream products. Therefore, establishing an effective production process monitoring, defect early warning, and quality control mechanism is crucial for ensuring product quality and improving production efficiency.

[0003] Existing technologies typically rely on statistical process control charts, historical data analysis, or rule-based systems based on expert experience for production process monitoring and fault diagnosis. For example, by setting upper and lower limits for various process parameters, an alarm is triggered when real-time data exceeds these limits. Some systems also attempt to use machine learning models such as neural networks and support vector machines to perform pattern recognition based on historical process and quality inspection data to predict or classify potential quality problems. These methods can, to some extent, detect known and simple process anomalies.

[0004] However, the aforementioned existing technologies have many shortcomings in practical applications. First, statistical control and rule-based systems can often only identify obvious, single-parameter anomalies. Their early warning capabilities are limited when dealing with complex fluctuations of multiple parameters, dynamic correlations across processes, and emerging defect patterns not fully reflected in historical data. Second, while machine learning models can handle complex data, their performance heavily relies on the completeness and diversity of historical data. They lack a deep understanding of the physical mechanisms, making it difficult to explain the root causes of predictions and cope with model drift caused by changes in production conditions. When abnormal operating conditions occur that the model has not learned, its diagnostic and early warning capabilities drop significantly, making it even more difficult to provide specific and effective adjustment suggestions. This results in the need for manual investigation and adjustment after quality problems occur, leading to low production efficiency and product losses. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a method and system for early warning of defects in galvanized steel sheet production based on cluster analysis. It employs a quantitative causal model constructed between processes, combined with a simulation sample library, and utilizes a two-layer clustering model to achieve real-time data fusion and online status assessment. This provides accurate defect warnings and calculates the shortest adjustment path for regressing high-quality clusters, generating a collaborative control instruction set to complete defect warning and quality control.

[0006] To achieve the above objectives, this application adopts the following technical solution: Firstly, a method for early warning of defects in galvanized steel sheet production based on cluster analysis is provided, including: acquiring historical process data and quality data of the production line; constructing an inter-process quantitative causal model describing the inter-process influence relationship based on the physicochemical mechanism of the entire galvanizing process; using the inter-process quantitative causal model to simulate and extrapolate the disturbance scenarios of process parameters, generating a simulation evolution sample library containing virtual defect features and corresponding defect root cause labels; fusing real-time production data with the simulation evolution sample library and inputting it into a pre-constructed two-layer clustering model for calculation to obtain the current online clustering state, wherein the two-layer clustering model includes a first-layer local clustering for identifying parameter fluctuations within a process and a second-layer global clustering for identifying cross-process quality trends; when the online clustering state indicates that the current production process deviates from the preset high-quality cluster boundary, calculating the shortest adjustment path to regress the high-quality cluster based on the inter-process quantitative causal model, generating a collaborative control instruction set, wherein the high-quality cluster is a cluster center and influence range constructed based on quality-compliant samples in historical data; and adjusting the process parameters of the corresponding production equipment according to the collaborative control instruction set to complete defect early warning and quality control.

[0007] Based on the above technical solution, the galvanized sheet production defect early warning method based on cluster analysis provided in this application adopts the construction of a quantitative causal model between processes and combined with a simulation sample library. It also achieves real-time data fusion and online status assessment through a two-layer clustering model, which can provide accurate defect early warning. Furthermore, it calculates the shortest adjustment path of the regression high-quality cluster and generates a collaborative control instruction set to complete defect early warning and quality control.

[0008] In conjunction with the first aspect above, in one possible implementation, the construction of the inter-process quantitative causal model describing the inter-process influence relationship includes: pre-setting a causal operator library covering cold rolling, galvanizing, and aluminized zinc plating processes, wherein the causal operators include nonlinear transfer functions used to characterize the mapping relationship between upstream process variables and downstream quality status; using the historical process data and quality data, offline training and calibration are performed on the weight parameters and time delay parameters in the causal operator library to obtain trained causal operators; according to the pre-set galvanizing production process flow, the trained causal operators are logically connected according to the material flow direction to dynamically generate a multidimensional directed causal graph characterizing the cross-process influence path; the multidimensional directed causal graph is defined as the inter-process quantitative causal model, and by introducing a state-space equation into the model, quantitative deduction calculation from process parameter disturbances to quality defect evolution results is realized; and using the differentiability of the nonlinear transfer function, the inverse mapping operator of the inter-process quantitative causal model is constructed to support sensitivity analysis of production status and inverse parameter solving.

[0009] In conjunction with the first aspect above, in one possible implementation, generating a simulation evolution sample library containing virtual defect features and corresponding defect root cause labels includes: pre-setting a combination of adversarial perturbation parameters covering cold rolling process fluctuations, coating thickness anomalies, and annealing temperature deviations, and injecting the adversarial perturbation parameter combination as an initial input variable into the inter-process quantitative causal model; using the multidimensional directed causal graph in the inter-process quantitative causal model, performing forward inference calculations based on physical mechanisms to simulate the transmission and accumulation of the adversarial perturbation parameters between processes, and obtaining the final virtual defect features; mapping and associating the generated virtual defect features with the adversarial perturbation parameter combination as the trigger source, and assigning each virtual defect feature a clear defect root cause label; generating several sets of simulation evolution samples with root cause labels by traversing the value space of the adversarial perturbation parameter combination, and assembling them into the simulation evolution sample library to assist the two-layer clustering model in delineating defect boundaries.

[0010] In conjunction with the first aspect above, in one possible implementation, the fusion of real-time production data with the simulation evolution sample library and input into a pre-constructed two-layer clustering model for calculation includes: constructing a first-layer local clustering of the two-layer clustering model to independently cluster real-time parameters within each process of cold rolling, annealing, and coating, identifying and outputting process-specific state feature vectors; constructing a second-layer global clustering of the two-layer clustering model, which receives the process-specific state feature vectors and introduces cross-process time-delay correlation operators to identify the overall process quality evolution trend; using sample points in the simulation evolution sample library as reference nodes and projecting them into the high-dimensional feature space constructed by the two-layer clustering model, and achieving data fusion by calculating the distribution density and positional offset of the real-time production data relative to the reference nodes in the high-dimensional feature space; and calculating an online clustering state including process-specific fluctuation patterns, overall process quality trends, and risk level scores based on the depth of the real-time production data falling into the preset sensitive risk area of ​​the simulation evolution sample library.

[0011] In conjunction with the first aspect above, in one possible implementation, running the two-layer clustering model includes: mapping real-time production data to a high-dimensional feature space using a kernel principal component analysis algorithm through the first-layer local clustering, and performing step-by-step dimensionality reduction by combining the process delay operator determined by the inter-process quantitative causal model, to calculate local offset vectors characterizing the fluctuations of single processes such as cold rolling, annealing, and coating; inputting the local offset vectors into the second-layer global clustering, and calculating the online clustering state characterizing the overall process quality risk level by combining the risk boundary defined by the defect root cause label in the simulation evolution sample library; calculating the overlap matching degree of the online clustering state with known samples in the simulation evolution sample library in real time through the two-layer clustering model, and evaluating the capture accuracy value of the two-layer clustering model for known risk features; when the capture accuracy value deviates from the performance benchmark determined by historical quality data, triggering the inter-process quantitative causal model to re-execute the disturbance scenario inference according to the current production conditions, generating supplementary simulation evolution samples, and incrementally reconstructing the two-layer clustering model.

[0012] In conjunction with the first aspect above, in one possible implementation, calculating the shortest adjustment path for regressing the high-quality cluster using the inter-process quantitative causal model includes: calculating the Euclidean distance vector between the feature point representing the current online clustering state and the center of the high-quality cluster within the high-dimensional feature space constructed by the two-layer clustering model, as the initial deviation vector; calling the inter-process quantitative causal model to calculate the sensitivity matrix of the influence of each process parameter on the initial deviation vector, and determining the search space of the regression path by combining the physical constraints of the equipment actuator; based on the sensitivity matrix, using the optimization of process adjustment cost as the objective function, inversely solving to obtain the parameter adjustment vector that causes the feature point to regress to the boundary of the high-quality cluster, generating the shortest adjustment path; decomposing the parameter adjustment vector into a feedforward compensation component for the preceding process and a feedback correction component for the current process, and generating a collaborative control instruction set through temporal coupling mapping between the components.

[0013] In conjunction with the first aspect above, in one possible implementation, decomposing the parameter adjustment vector into feedforward compensation components for preceding processes includes: mapping the high-quality clusters to target quality index ranges in the inter-process quantitative causal model, and setting target state trajectories for future production cycles according to the production plan; constructing a multi-objective optimization cost function with the objectives of optimizing the deviation between the predicted quality and the target state trajectory, and optimizing the adjustment range of process parameters; performing backpropagation calculation based on the chain rule on the multi-dimensional directed causal graph of the inter-process quantitative causal model to solve for the preceding process parameter adjustment amount that optimizes the cost function; defining the preceding process parameter adjustment amount as a feedforward compensation component, used for predictive intervention in upstream cold rolling or annealing processes before feature points deviate from the boundary.

[0014] In conjunction with the first aspect above, in one possible implementation, triggering the inter-process quantitative causal model to re-execute the disturbance scenario inference based on the current production conditions, generating supplementary simulation evolution samples, and incrementally reconstructing the two-layer clustering model includes: when the capture accuracy value deviates from the performance benchmark, performing uncertainty analysis on the clustering boundary of the two-layer clustering model in the high-dimensional feature space to identify weak boundary regions with sparse sample distribution; obtaining the spatial feature vector corresponding to the weak boundary region, using the inter-process quantitative causal model to reverse-infer the initial disturbance parameter range that leads to the generation of the spatial feature vector, defining it as a directional disturbance mode as a supplement to the adversarial disturbance parameter combination; performing forward inference calculation around the directional disturbance mode to generate supplementary simulation evolution samples with root cause labels, and injecting the supplementary simulation evolution samples into the simulation evolution sample library to achieve precise reinforcement of the identification boundary of the two-layer clustering model.

[0015] In conjunction with the first aspect above, in one possible implementation, adjusting the process parameters of the corresponding production equipment according to the collaborative control instruction set to complete defect early warning and quality control includes: mapping the feedforward compensation component and feedback correction component in the collaborative control instruction set to action adjustment amounts for specific actuators on the production line; calculating the advance amount of the feedforward compensation component using the time delay parameters in the inter-process quantitative causal model, and sending the corresponding instruction to the upstream cold rolling or annealing process equipment prior to the production cycle time according to the advance amount; and adjusting the action adjustment amount corresponding to the feedback correction component... The system sends data in real time to the galvanizing or aluminized zinc plating equipment in the current process to achieve immediate correction of the current production status. It also monitors the online clustering status after the action adjustment is executed in real time and determines the regression trend of the feature points relative to the boundary of the high-quality cluster. If the feature points regress to the boundary or interior of the high-quality cluster, the current production defect risk is determined to be eliminated, and the system switches to continuous monitoring mode. If the feature points do not enter the high-quality cluster and show a deviation trend, the system triggers the inter-process quantitative causal model to recalculate the sensitivity matrix and updates the collaborative control instruction set until the feature points regress.

[0016] Secondly, a clustering analysis-based early warning system for defects in galvanized steel sheet production is provided, comprising: a causal modeling module for acquiring historical process and quality data of the production line, and constructing a quantitative causal model describing the inter-process influence relationship based on the physicochemical mechanism of the entire galvanizing process; a simulation evolution module for simulating and extrapolating perturbation scenarios of process parameters using the quantitative causal model, generating a simulation evolution sample library containing virtual defect features and corresponding defect root cause labels; and a two-layer clustering analysis module for fusing real-time production data with the simulation evolution sample library, inputting it into a pre-constructed two-layer clustering model for calculation, and obtaining the current online clustering state. The two-layer clustering model includes a first-layer local clustering for identifying parameter fluctuations within a process and a second-layer global clustering for identifying quality trends across processes. A path decision module is used to calculate the shortest adjustment path to regress the high-quality cluster when the online clustering status indicates that the current production process deviates from the preset high-quality cluster boundary, combining the inter-process quantitative causal model, and generating a collaborative control instruction set. The high-quality cluster is a cluster center and its influence range constructed based on quality-compliant samples from historical data. An execution and control module is used to adjust the process parameters of the corresponding production equipment according to the collaborative control instruction set, completing defect early warning and quality control.

[0017] Compared with the prior art, the present invention has the following advantages: This invention, by constructing a quantitative causal model between processes based on physicochemical mechanisms, enables the quantitative deduction of the evolution of quality defects from process parameter disturbances in galvanized sheet production. This allows the system not only to simulate defect formation processes under known abnormal conditions, but also to pre-set adversarial disturbances based on professional experience and historical data, generating a simulation evolution sample library containing virtual defect features and their root cause labels. This enriches the understanding of potential defect patterns and provides a solid theoretical and data foundation for early warning and control.

[0018] This invention proposes and implements a two-layer clustering model that can integrate and analyze real-time production data with a simulation sample library. The first layer, local clustering, focuses on identifying fluctuations in internal parameters of each process, including cold rolling, annealing, and coating. The second layer, global clustering, effectively identifies the overall quality evolution trend by introducing cross-process time-delay correlation operators. This hierarchical and integrated clustering analysis capability ensures that the system can comprehensively grasp the production status, avoiding misjudgments or omissions caused by missing local information or complex inter-process coupling in traditional methods.

[0019] This invention, upon detecting a deviation from the boundary of a high-quality cluster in the production process, can calculate the shortest adjustment path back to a high-quality production state by combining an inter-process quantitative causal model and generate a collaborative control instruction set. This not only provides defect early warning but also offers an operable and optimized solution. The path considers the sensitivity of process parameters to deviations and equipment physical constraints. By decomposing the adjustment amount into feedforward compensation components for preceding processes and feedback correction components for the current process, it achieves collaborative optimization control throughout the entire process, reducing quality risks and process adjustment costs.

[0020] This invention possesses adaptive and iterative optimization capabilities. When the capture accuracy value of the two-layer clustering model deviates from the preset performance benchmark, the system can trigger the inter-process quantified causal model to re-execute the perturbation scenario simulation, generate supplementary simulation evolution samples, and incrementally reconstruct the two-layer clustering model. This dynamic learning and model update mechanism enables the early warning system to continuously adapt to changes in the production environment, accumulate new defect knowledge, and continuously improve its early warning accuracy and robustness, thereby maintaining its advanced nature and effectiveness in the long term.

[0021] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims, and drawings. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This application provides a structural architecture diagram of a galvanized steel sheet production defect early warning system based on cluster analysis, as shown in the embodiments of this application. Figure 2 A flowchart illustrating the method for early warning of defects in galvanized steel sheet production based on cluster analysis, provided in an embodiment of this application; Figure 3 This is a virtual defect evolution trajectory diagram based on adversarial perturbation provided in the embodiments of this application; Figure 4 This is a cluster boundary uncertainty score distribution cloud map provided in the embodiments of this application. Detailed Implementation

[0024] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0025] The cluster analysis-based early warning method for galvanized steel sheet production provided in this application can be applied to, for example... Figure 1 In the cluster analysis-based early warning system for galvanized steel sheet production defects shown in the figure 100, such as... Figure 1 As shown, the system includes: The causal modeling module is used to acquire historical process data and quality data of the production line, and to construct a quantitative causal model between processes based on the physicochemical mechanism of the entire galvanizing process. The simulation evolution module is used to simulate and extrapolate the disturbance scenarios of process parameters using the inter-process quantitative causal model, and generate a simulation evolution sample library containing virtual defect features and corresponding defect root cause labels. The two-layer clustering analysis module is used to integrate real-time production data with the simulation evolution sample library and input it into a pre-built two-layer clustering model for calculation to obtain the current online clustering status. The two-layer clustering model includes a first-layer local clustering for identifying parameter fluctuations within the process and a second-layer global clustering for identifying quality trends across processes. The path decision module is used to calculate the shortest adjustment path to regress the high-quality cluster when the online clustering status indicates that the current production process deviates from the preset high-quality cluster boundary, in combination with the inter-process quantitative causal model, and generate a collaborative control instruction set. The high-quality cluster is a cluster center and influence range constructed based on quality-compliant samples in historical data. The execution and control module is used to adjust the process parameters of the corresponding production equipment according to the collaborative control instruction set, and to complete defect early warning and quality control.

[0026] like Figure 2 As shown in the embodiments of this application, a method for early warning of defects in galvanized steel sheet production based on cluster analysis is provided, including: Historical process and quality data of the production line are obtained, and based on the physicochemical mechanism of the entire galvanizing process, a quantitative causal model describing the inter-process influence relationship is constructed. The inter-process quantitative causal model is used to simulate and extrapolate the disturbance scenarios of process parameters, generating a simulation evolution sample library containing virtual defect features and corresponding defect root cause labels. Real-time production data is fused with the simulation evolution sample library and input into a pre-built two-layer clustering model for calculation to obtain the current online clustering state. The two-layer clustering model includes a first-layer local clustering for identifying parameter fluctuations within a process and a second-layer global clustering for identifying quality trends across processes. When the online clustering status indicates that the current production process deviates from the preset high-quality cluster boundary, the shortest adjustment path for regressing the high-quality cluster is calculated by combining the inter-process quantitative causal model, and a collaborative control instruction set is generated. The high-quality cluster is a cluster center and influence range constructed based on quality-compliant samples in historical data. Based on the collaborative control instruction set, the process parameters of the corresponding production equipment are adjusted to complete defect early warning and quality control.

[0027] It should be noted that the technical principle of this invention lies in firstly constructing a model that can quantitatively describe the causal relationship between various production processes by deeply integrating the physicochemical mechanisms and historical data of the entire galvanizing process. Based on this causal model, a virtual sample library containing a large number of potential defect patterns and their root cause labels is pre-generated through systematic simulation, thereby expanding the cognitive boundaries of abnormal operating conditions. In actual production, a two-layer clustering model integrates real-time production data with this virtual sample library for analysis. This model can not only identify local parameter fluctuations within a single process but also capture the global quality evolution trend across processes, achieving a precise profile of the production process status. Once it is identified that the current production status deviates from the preset high-quality production range, the causal model can be used again to calculate the shortest adjustment path to return to the optimal state and generate executable collaborative control instructions, forming a complete technical closed loop from mechanism modeling, simulation enhancement, real-time clustering to optimization decision-making.

[0028] In one possible implementation of the embodiments of this application, combined with Figure 2 The construction of the inter-process quantitative causal model describing the inter-process influence relationship includes: A pre-defined library of causal operators covering cold rolling, galvanizing, and aluminized zinc plating processes is provided. The causal operators include nonlinear transfer functions used to characterize the mapping relationship between upstream process variables and downstream quality status. Using the historical process data and quality data, the weight parameters and time delay parameters in the causal operator library are trained and calibrated offline to obtain trained causal operators; According to the preset galvanizing production process, the trained causal operators are logically connected in the direction of material flow to dynamically generate a multi-dimensional directed causal graph representing the cross-process influence path. The multidimensional directed causal graph is defined as the inter-process quantitative causal model. By introducing state-space equations into this model, the quantitative deduction and calculation of the evolution results from process parameter disturbances to quality defects are realized. The differentiability of the nonlinear transfer function is used to construct the inverse mapping operator of the inter-process quantitative causal model, which supports sensitivity analysis of production status and inverse parameter solving.

[0029] In some implementations, inter-process quantitative causal models are achieved by integrating thermodynamic mechanisms with data-driven operators. Taking the construction of a causal operator for the influence of the annealing process on the thickness of the subsequent zinc coating as an example, its nonlinear transfer function... The following Physical Information Neural Network (PINN) structure is used for mapping: ; in, This indicates the calculated expected oxide film thickness on the strip surface, in nm. This parameter directly determines the wettability of the zinc bath and the final coating adhesion quality. The data comes from the feedback of the laser thickness gauge at the production line exit. The real-time measured temperature of the soaking zone of the annealing furnace is expressed in Kelvin (K) and is collected by a bicolor infrared pyrometer located inside the furnace area. This indicates the real-time linear velocity of the strip, expressed in meters per second (m / s), which is calculated from the encoder pulse frequency of the tension roller at the unit's inlet. This represents the effective physical length of the heat spreader, in meters, and is a design parameter for the equipment. Let be the molar gas constant, and take the value of . ; The activation energy for the oxidation reaction is expressed in J / mol. This is a proportionality coefficient related to the reaction kinetics of the strip surface, and its dimensions depend on the values ​​of the reaction order $n$ and $m$. and These four parameters are the reaction order exponent. These are the weight parameters that require offline training and calibration. During the calibration process, data from the historical database of strip steel of the same specification are retrieved. , Sequence, combined with laboratory physicochemical analysis to obtain oxide film thickness Least squares fitting is performed. To handle the dynamic effects between processes, a time delay parameter is introduced. Align the action time of causal operators, where This represents the physical distance between the annealing furnace outlet and the zinc pot inlet rollers, in meters (m). By chaining the calibrated operators together according to the production process logic, each node in the resulting graph structure represents a state-space equation: ; in, This is a process state vector, such as the strip temperature drop rate. For process control variables, such as cooling fan frequency, this enables quantitative deduction from process disturbances to defect evolution. Utilizing the differentiability of the exponential function, the partial derivative matrix is ​​obtained. Construct an inverse mapping operator when it is found When the value deviates from the optimal range, it can be solved in reverse. The correction amount.

[0030] For example, in the annealing process of a galvanizing production line, the specific calculation process for constructing a quantitative causal model between processes is as follows: First, retrieve a set of measured process data for strip steel of the same specification from the historical database, assuming that the temperature is measured in real time. Infrared temperature measurement value of the soaking zone, and the running linear speed of the strip. The weight parameters in the preset operator library are calibrated offline as follows: activation energy Frequency factor Reaction order , Given the molar gas constant Substitute the parameters into the nonlinear transfer function. Set the effective physical length reactant concentration The first step is to calculate the exponent: The second step is to calculate the time effect term: The third step is to calculate the expected oxide film thickness: If the actual value fed back by the laser thickness gauge deviates from the calculated value, the sensitivity can be calculated using the differentiability of the transfer function. The sensitivity coefficient at the current temperature is calculated by differentiation, for example, as follows: If the current oxide film thickness If the deviation from the target value in the optimal range is 5nm, the temperature correction amount is solved in reverse using the inverse mapping operator. This enables precise reverse solution and parameter control of the production status.

[0031] In one possible implementation, combining Figure 2 The generation of the simulation evolution sample library containing virtual defect features and corresponding defect root cause labels includes: A set of counteracting disturbance parameters covering cold rolling process fluctuations, coating thickness anomalies, and annealing temperature deviations is preset, and the set of counteracting disturbance parameters is injected into the inter-process quantitative causal model as an initial input variable; By utilizing the multidimensional directed causal graph in the inter-process quantitative causal model, forward inference calculations based on physical mechanisms are performed to simulate the transmission and accumulation of the adversarial disturbance parameters between processes, thereby obtaining the virtual defect characteristics of the final output. The generated virtual defect features are mapped and associated with the adversarial perturbation parameter combination that serves as the triggering source, and each virtual defect feature is assigned a clear defect root cause label. By traversing the value space of the adversarial perturbation parameter combination, several sets of simulation evolution samples with root cause labels are generated and collected into the simulation evolution sample library to assist the two-layer clustering model in delineating defect boundaries.

[0032] In some implementations, the generation process of the simulation evolution sample library deeply mines production conditions through an adversarial perturbation mechanism. First, based on the extreme value distribution in the historical fault database, upper search bounds are determined for cold rolling process fluctuations, coating thickness anomalies, and annealing temperature deviations. and the lower world Within this interval, adversarial perturbation parameter combinations are generated using Latin hypercube sampling. The combination As input vectors, they are injected into the inter-process quantification causal model to perform forward inference calculations based on physical mechanisms, thereby predicting virtual defect feature values ​​in the final output. For example, the calculation formula is as follows: ; in, To quantify the virtual defects generated in the simulation, such as the diameter or thickness difference of the plating defects, the unit is... Its value is generated by model deduction and used for subsequent sample labeling; Indicates the first One of the counteracting disturbance process parameters, such as air knife pressure fluctuation value Strip running speed deviation The initial values ​​are derived from a preset combination of disturbance parameters; The weight contribution operator for the corresponding process is obtained from the correlation analysis of historical quality data and process parameters through offline training; This refers to the physical bias of the corresponding process, which originates from the basic error specified by the equipment at the factory. For strip steel in the first The cumulative running time of each process, in units of Recorded in real time by a linear velocity encoder; This represents the physicochemical reaction characteristic time constant for the corresponding process, in units of... The parameters are determined by the mechanistic constants given in the process manual. This formula is used to determine the perturbation parameters. The cumulative effect between various processes is transformed into specific virtual defect characteristics. Then, the calculated eigenvalues ​​for each group... Mapped to the corresponding root cause label ,like Exceeding the preset quality and safety threshold Generally take Then mark The process is designated as "abnormal-root cause process number" or otherwise marked as "critical-fluctuation". Finally, this is achieved by traversing the combinations of adversarial perturbation parameters. Generate a collection of multiple sets of sampled values ​​from the space of all samples. The simulated evolution samples of the data are compiled into a simulated evolution sample library. The data points in this library form clear defect evolution trajectories in the feature space, providing high-gain data support for the two-layer clustering model to delineate precise defect edge risk regions. This solves the problem of insufficient training of early warning models caused by the extreme scarcity of defect samples in actual production. Figure 3As shown, the virtual defect feature values ​​obtained by the simulation evolution module through formula derivation are illustrated. The dynamic evolution process; the solid line in the figure reflects the adversarial perturbation parameters. The overall trend of defect features as they increase; the dashed line represents the preset safety threshold. The two-layer clustering model is used to delineate precise defect boundaries by identifying the characteristic regions where samples cross the threshold.

[0033] For example, predicting the virtual defect feature values ​​of the final output. The specific calculation process for constructing the simulation evolution sample library is as follows: First, determine the combination of adversarial perturbation parameters. Assuming the currently sampled disturbance parameter is the air knife pressure fluctuation value and strip running speed deviation The weight contribution operator for the corresponding process is known. , Physical bias term , The cumulative running time of the strip steel in each process. , ; Physicochemical reaction characteristic time constant , Substitute into the formula Calculation: Step 1: Calculate the contribution of the first process: The second step is to calculate the contribution of the second process: The third step is to accumulate the virtual defect feature values: .because The preset quality and safety threshold was not exceeded. Root cause label of the sample This is labeled "critical-fluctuation". This process is repeated by traversing the sampling space, and finally compiled into a simulation evolution sample library.

[0034] In one possible implementation, combining Figure 2 The step of fusing real-time production data with the simulation evolution sample library and inputting it into a pre-constructed two-layer clustering model for calculation includes: The first layer of local clustering in the two-layer clustering model is used to independently cluster the real-time parameters within each process of cold rolling, annealing and coating, and to identify and output the state feature vector within the process. The second layer of global clustering in the two-layer clustering model is constructed. The second layer of global clustering receives the state feature vector within the process and introduces a time delay correlation operator across processes to identify the quality evolution trend of the entire process. The sample points in the simulation evolution sample library are used as reference nodes and projected into the high-dimensional feature space constructed by the two-layer clustering model. Data fusion is achieved by calculating the distribution density and position offset of the real-time production data relative to the reference nodes in the high-dimensional feature space. Based on the depth of the real-time production data falling into the preset sensitive risk area of ​​the simulation evolution sample library, the online clustering state, which includes the fluctuation pattern within the process, the overall quality trend, and the risk level score, is calculated.

[0035] In some implementations, the process of fusing real-time production data with a simulation evolution sample library and inputting it into a two-layer clustering model essentially maps multi-source heterogeneous data to a unified high-dimensional space. First, the pre-stored set of baseline node coordinates in the simulation evolution sample library is retrieved. The Gaussian radial basis kernel function was used to calculate real-time production data points. Relative distribution density between each reference node The specific calculation formula is as follows: ; in, This represents the fusion density score of real-time data in the feature space. The higher the value, the higher the overlap between the current production state and the known patterns in the sample library. This parameter is generated through real-time calculation. For the first in the simulation evolution sample library The feature vectors of each baseline node are derived from the virtual defect features generated by the simulation evolution module. This represents the total number of baseline nodes participating in the calculation. Represents the Euclidean norm; The spatial smoothing bandwidth parameter is obtained by statistically analyzing the standard deviation of each parameter in historical process data. For the first in real-time production data The physical offset distance of each process parameter, such as the inlet tension of the cold rolling mill stand and the hydrogen concentration of the annealing furnace, relative to the center of its high-quality cluster is obtained by the on-site sensors in real time and then processed by standardization. These are the parameter weighting factors determined by the inter-process quantitative causal model; This represents the total number of dimensions of process parameters collected in real time. This is a time lag correction factor used to compensate for the time difference in data acquisition across processes. Based on this, calculations are performed... The depth at which a sample falls into the sensitive risk region defined by root cause labels in the simulation evolution sample library: ; And combined with the process state feature vector output by the first layer of local clustering And the quality evolution trend of second-layer global clustering recognition Finally, the online clustering state is synthesized. .Should It includes a risk level score, which not only quantitatively describes the degree of deviation of the real-time production process from the scope of high-quality production, but also intuitively reflects the evolution path of potential defects through spatial projection position, thereby ensuring that the system can accurately determine the current production safety boundary when facing complex fluctuations.

[0036] For example, calculating the fusion density score of real-time production data in the feature space. The specific process of inputting data into a two-layer clustering model for fusion calculation is as follows: First, retrieve the pre-stored data from the simulation evolution sample library. coordinates of each reference node , Assuming real-time production data points The Euclidean distances from the reference node are respectively Preset space smoothing bandwidth Time delay correction factor Real-time data collection A process parameter, such as the hydrogen concentration in the annealing furnace, and its physical offset distance. Weighting factors Substitute into the formula Calculation: The first step is to calculate the contribution of the baseline node: The second step is to calculate the real-time parameter offset: The third step is to accumulate the fusion density score: Then the depth of the sensitive risk area was calculated. By combining this score with the feature vector output from the first-level local clustering, an online clustering state is synthesized. This allows us to determine the risk level of the current production process.

[0037] In one possible implementation, combining Figure 2 Running the two-layer clustering model includes: Through the first layer of local clustering, the kernel principal component analysis algorithm is used to map real-time production data to a high-dimensional feature space, and the process delay operator determined by the inter-process quantitative causal model is combined to perform step-by-step dimensionality reduction, and the local offset vector representing the fluctuation of single processes of cold rolling, annealing and coating is calculated. The local offset vector is input into the second-layer global clustering, and combined with the risk boundary defined by the defect root cause label in the simulation evolution sample library, the online clustering state characterizing the quality risk level of the entire process is calculated. The overlap and matching degree of the online clustering state with known samples in the simulation evolution sample library is calculated in real time using a two-layer clustering model, and the accuracy of the two-layer clustering model in capturing known risk features is evaluated. When the capture accuracy value deviates from the performance benchmark determined by historical quality data, the inter-process quantitative causal model is triggered to re-execute the disturbance scenario simulation based on the current production conditions, generate supplementary simulation evolution samples, and incrementally reconstruct the two-layer clustering model.

[0038] In some implementations, the process of running a two-layer clustering model achieves closed-loop optimization of the model through high-dimensional space transformation and accuracy feedback mechanisms. First, the first-layer local clustering is used to perform step-by-step dimensionality reduction on the real-time production data. Then, kernel principal component analysis is employed to transform the original data... Mapping a 3D process vector to a high-dimensional feature space, the mapping operator It is implemented using a radial basis function (RBF), and its expression is: ; in, To calculate the high-dimensional mapping feature value of the output, which is used to transform nonlinear production fluctuations into a linearly separable spatial distribution, the parameters are derived from the real-time data stream of the field sensor array; The vector of process parameters is collected in real time, and its data comes from the field sensor array, such as cold rolling mill stand pressure, annealing furnace tension, etc. A matrix of historical high-quality process samples pre-stored in the system memory; The preset spatial smoothing parameter, with units consistent with the process parameter, is obtained by statistically analyzing the sample variance of each process parameter in historical data. Based on this, the local offset vector characterizing the fluctuation of a single process is calculated: ; in, To calculate the single-process local offset vector of the output, used to quantitatively describe the running deviation of cold rolling, annealing or coating monomers; This is the principal component feature vector matrix obtained through offline training using historical process data; The time delay compensation coefficient is preset by the inter-process quantitative causal model; This represents the time difference between the current moment and the previous sampling period. Then, this... The second-level global clustering is input, and combined with the risk boundary defined by root cause labels in the simulation evolution sample library, the online clustering state is calculated. To evaluate the accuracy of the early warning, the current online clustering state point is calculated. Center of known defect samples in the simulation evolution sample library Overlapping matching degree between them: ; in, The capture accuracy value is generated in real time, and its value is between 0 and 1. It is used to characterize the accuracy of the model in identifying the current risk. These are the online clustering coordinates in the high-dimensional space at the current moment; The center coordinates of the defect patterns are preset in the sample library and are derived from the simulation evolution module. This is the spatial diffusion radius parameter determined by the distribution of historical quality compliance data. When this capture accuracy value... Below the performance benchmark threshold determined by historical quality data When it is determined that the current model is insufficient to identify new types of fluctuations, it is generally taken as follows: This triggers the inter-process quantitative causal model to re-execute the perturbation simulation based on the current process parameters, generating supplementary simulation evolution samples and incrementally reconstructing the two-layer clustering model. This process ensures that the early warning system can adaptively adjust to the constantly evolving production environment, thereby eliminating doubts about the functional description and improving the system's robustness under complex operating conditions.

[0039] For example, the specific calculation process of running a two-layer clustering model and triggering incremental reconstruction: First stage: Local offset vector Dimensionality reduction calculation to obtain real-time process parameter vectors High-dimensional eigenvalues ​​are obtained by radial basis kernel function mapping. Retrieve the pre-stored principal component eigenvector matrix. Delay compensation coefficient The time difference between the current moment and the previous cycle Calculation process: .Should This represents the physical offset of the current single process relative to the steady state. Second stage: Global clustering state points. The local offset vectors of each process, such as cold rolling, annealing, and coating, are input into the second-layer global clustering. Combined with cross-process time delay operators, the current online clustering state point is determined in the feature space. Third stage: Capture precision value The evaluation retrieves preset defect mode centers from the simulation sample library. Set the current Spatial diffusion radius The first step is to calculate the exponent: The second step is to calculate the precision value: Phase Four: Decision-making and Restructuring Triggers, Comparison Compared with performance benchmark threshold .because The system determines that the current model capture accuracy meets the standard and maintains real-time monitoring of the mode. If subsequent changes occur due to operating condition drift... If increased to 0.5, then .because This will trigger the inter-process quantitative causal model to re-execute the disturbance scenario simulation, generate supplementary simulation samples, and incrementally reconstruct the two-layer clustering model.

[0040] In one possible implementation, combining Figure 2 The shortest adjustment path for regressing the high-quality clusters, calculated using the inter-process quantitative causal model, includes: Within the high-dimensional feature space constructed by the two-layer clustering model, the Euclidean distance vector between the feature points representing the current online clustering state and the center of the high-quality cluster is calculated as the initial bias vector; The inter-process quantitative causal model is invoked to calculate the sensitivity matrix of the influence of each process parameter on the initial deviation vector, and the search space of the regression path is determined by combining the physical constraints of the equipment actuator. Based on the aforementioned sensitivity matrix, and with the optimization of process adjustment cost as the objective function, the parameter adjustment vector that causes the feature points to regress to the boundary of the high-quality cluster is obtained by reverse calculation, thereby generating the shortest adjustment path. The parameter adjustment vector is decomposed into a feedforward compensation component for the preceding process and a feedback correction component for the current process. A collaborative control instruction set is generated through the temporal coupling mapping between the components.

[0041] In some implementations, calculating the shortest adjustment path for regressing high-quality clusters using an inter-process quantitative causal model is achieved by performing gradient backpropagation in a high-dimensional feature space. First, the feature coordinates of the current online clustering state are acquired in real time. And locate the preset high-quality cluster center point. Then, the initial deviation vector between the two is calculated: ; Subsequently, the inter-process quantitative causal model was invoked, and the influence sensitivity matrix was constructed by calculating the partial derivatives of each process parameter with respect to the deviation vector. For each process parameter to be adjusted Its adjustment increment The calculation follows the cost function minimization criterion: ; in, The optimal parameter adjustment vector for calculating the output, i.e. the shortest adjustment path, is used to guide the equipment actuator to perform precise compensation. The elements of the sensitivity matrix are obtained by differentiating the nonlinear transfer function in the quantification causal model, and represent the characteristic space displacement caused by a unit parameter change. The Euclidean distance vector between the current state point and the high-quality center point is derived from the real-time output of the two-layer clustering module. The preset regularization coefficient is used to balance the adjustment accuracy and adjustment cost. The total dimension of the process parameters to be adjusted; For the first The cost weighting of adjusting each process parameter, such as the ratio of energy consumption cost of changing furnace temperature to output loss of changing drawing speed, is derived from the equipment operating parameters preset in the production management system. In order to target the The specific adjustment increment of a process parameter, such as increasing the air knife pressure by 0.2 kPa, is considered during the solution process, taking into account the physical constraints of the equipment's actuators. For parameters such as the upper limit of motor speed and the range of valve opening, the parameter adjustment vector that minimizes the above cost function is solved in reverse within a limited search space. Finally, the vector is decomposed into a feedforward compensation component for the preceding process and a feedback correction component for the current process, thereby achieving quality regression control under full-process collaboration and ensuring that the production status can quickly return to the range of high-quality clusters with minimal process fluctuation costs.

[0042] For example, the specific calculation process for calculating the shortest adjustment path for regressing high-quality clusters using an inter-process quantitative causal model is as follows: First, the feature coordinates of the current online clustering state are obtained in real time. And locate the preset high-quality cluster center point. Assuming in a high-dimensional feature space, The coordinates are (1.2, 0.8). The coordinates are (0.2, 0.1). Calculate the initial deviation vector. Construct the sensitivity influence matrix By invoking the inter-process quantitative causal model and differentiating the nonlinear transfer function, the influence coefficients of each process parameter on the characteristic space displacement are calculated. Assuming the current process involves... The sensitivity matrix obtained by adjusting parameters such as annealing temperature and air knife pressure is as follows: Calculating the adjustment path based on cost function optimization. Set regularization coefficient Parameter adjustment cost weight , By solving in reverse, the cost function is made easier to obtain. The minimized adjustment vector. Assuming that the optimal adjustment increment is obtained through gradient optimization iterations, the solution is... If the temperature is increased by 1.1K, For example, the pressure is increased by 0.8 kPa. Verify the deviation compensation effect: The result is very close to the initial deviation. Finally, adjust the vector. It is decomposed into feedforward compensation components and feedback correction components to generate a collaborative control instruction set, ensuring that the production status quickly returns to the range of high-quality clusters with minimal fluctuations.

[0043] In one possible implementation, combining Figure 2 The step of decomposing the parameter adjustment vector into feedforward compensation components for the preceding process includes: The high-quality clusters are mapped to target quality index ranges in the inter-process quantitative causal model, and the target state trajectory of future production cycles is set according to the production plan. A multi-objective optimization cost function is constructed with the objectives of optimizing the deviation between the predicted quality and the target state trajectory, and optimizing the adjustment range of process parameters. In the multidimensional directed causal graph of the inter-process quantitative causal model, backpropagation calculation based on the chain rule is performed on the multi-objective optimization cost function to obtain the adjustment amount of the preceding process parameters that optimizes the cost function; The adjustment amount of the preceding process parameters is defined as a feedforward compensation component, which is used to proactively intervene in the upstream cold rolling or annealing process before the feature point deviates from the boundary.

[0044] In some implementations, decomposing the parameter adjustment vector into feedforward compensation components for preceding processes is achieved by performing predictive inverse solving based on the target state trajectory in a multidimensional directed causal graph. First, pre-defined high-quality cluster centers and their influence ranges are retrieved and mapped to target quality index intervals in the high-dimensional feature space of the inter-process quantified causal model. Simultaneously, combined with the current production plan, such as the target specifications for strip steel and the unit operating speed, the target state trajectory for the future production cycle is set. To find the optimal process parameter adjustment, a multi-objective optimization cost function is constructed. Its specific mathematical expression is: ; in, To calculate the total value of output, it is used to quantify the combined level of forecast bias and adjustment costs; To predict future quality values, such as the predicted zinc layer thickness, obtained through forward extrapolation using an inter-process quantitative causal model, in units of... ; The preset target state trajectory reference value, in units of It originates from production planning instructions; and These are preset weighting coefficients used to balance deviation control and parameter fluctuations; The total dimension of the preceding process parameters to be adjusted; For the first The adjustment increments of several preceding process parameters, such as the cold rolling five-stand pressing force and the hydrogen flow rate of the annealing furnace, need to be solved; The damping factor corresponding to the parameter is derived from the stability constraints in the equipment operation manual. During the solution process, the differentiability of the nonlinear transfer function in the inter-process quantitative causal model is utilized to refine the cost function in the multidimensional directed causal graph. Perform backpropagation computation based on the chain rule, that is, iteratively optimize by calculating gradients: ; Finally, the cost function is obtained. The process parameter adjustment amount that reaches the global minimum. This adjustment amount is defined as the feedforward compensation component, which can issue intervention commands to the equipment actuators of the upstream cold rolling or annealing process in advance before the current online clustering state feature points actually deviate from the quality boundary, thereby achieving proactive suppression of quality fluctuations throughout the entire process.

[0045] For example, the specific calculation process of decomposing the parameter adjustment vector into feedforward compensation components for the preceding process is as follows: First, the target state trajectory for the future production cycle is set based on the current strip steel specifications. For example, the target zinc layer thickness. Feedforward inference is performed using an inter-process quantitative causal model to obtain future quality predictions. Constructing a multi-objective optimization cost function Set weight coefficients , Assume the dimensions of the preceding process parameters to be adjusted are... The adjustment increment to be solved is Its regulating damping factor The cost function expression is: The system performs backpropagation calculations based on the chain rule to compute the gradient. The feedforward compensation component is obtained through iterative optimization, and then iteratively... To reach the global minimum, and considering the sensitivity mapping of the nonlinear transfer function in the causal graph, we assume that the optimal adjustment amount of the preceding process parameters that brings the prediction bias back to zero can be obtained by solving the problem. For example, the flow rate is reduced by 2.5 units. This adjustment is defined as the feedforward compensation component, which is used to proactively intervene in the upstream annealing process before the characteristic point actually deviates from the boundary, thereby achieving advance suppression of quality fluctuations.

[0046] In one possible implementation, combining Figure 2 The process triggers the inter-process quantitative causal model to re-execute the disturbance scenario simulation based on the current production conditions, generate supplementary simulation evolution samples, and incrementally reconstruct the two-layer clustering model, including: When the capture accuracy value deviates from the performance benchmark, uncertainty analysis is performed on the clustering boundary of the two-layer clustering model in the high-dimensional feature space to identify weak boundary regions with sparse sample distribution. Obtain the spatial feature vector corresponding to the weak boundary region, and use the inter-process quantitative causal model to reverse deduce the range of initial perturbation parameters that cause the spatial feature vector to be generated, which is defined as a directional perturbation mode as a supplement to the adversarial perturbation parameter combination. Forward inference calculations are performed around the directional perturbation mode to generate supplementary simulation evolution samples with root cause labels, and the supplementary simulation evolution samples are injected into the simulation evolution sample library to achieve accurate reinforcement of the identification boundary of the two-layer clustering model.

[0047] In some implementations, triggering the inter-process quantified causal model to re-execute the perturbation scenario deduction and generate supplementary simulation evolution samples is a targeted reinforcement mechanism for blind spots in model cognition. By performing uncertainty analysis on the clustering boundaries of the two-layer clustering model in the high-dimensional feature space, the sample sparsity of the current spatial region is calculated using the information entropy metric. The calculation formula is: ; in, The uncertainty score for the boundary region is used to quantitatively identify weak locations with sparse sample distribution. The higher the score, the lower the reliability of the model's warning in that region. The score is calculated and generated in real time by the system backend. The total dimension of the parameters within the selected boundary neighborhood; For the first The standardized offset distance of each process feature dimension is derived from the comparison results between real-time production data and high-quality cluster centers. The local sample point density at this location in the feature space is obtained by statistically analyzing the number of existing benchmark nodes near this coordinate point in the simulation evolution sample library. This is a preset spatial smoothing constant, whose value is taken with reference to the global variance of historical process parameters. When Exceeding the preset reconstruction threshold If the current two-layer clustering model is deemed insufficient to capture known risks in the region, an incremental learning process needs to be initiated; this reconstruction threshold... The specific values ​​are obtained by performing leave-one-out cross-validation on historical quality-compliant data, and are typically set to the critical information entropy level that ensures the overall capture accuracy of the model is not lower than the performance baseline of 0.85. Then, the central feature vector of this weak region is extracted. Furthermore, by utilizing the inverse mapping operator of the inter-process quantization causal model, the corresponding initial perturbation parameter range is obtained through reverse solution. This is defined as a directional perturbation mode. The core performs forward inference calculations, generating supplementary simulation evolution samples with clearly defined defect root cause labels, and incrementally injects them into the simulation evolution sample library. This targeted reinforcement method ensures that the two-layer clustering model, without disrupting the existing knowledge structure, can accurately reshape the defect warning boundary by selectively adding edge feature samples, thereby effectively addressing the risk of missed detection of new defects caused by production environment drift. Figure 4 As shown, the contour gradient illustrates the sample sparsity distribution in the high-dimensional feature space; the "x" in the figure represents the known high-quality cluster centers, and the contour values ​​represent the uncertainty score. The areas with higher scores are the weak boundary areas, which trigger a directional perturbation mode to generate supplementary samples for incremental reconstruction.

[0048] For example, the specific calculation process for triggering the inter-process quantitative causal model to re-execute the perturbation scenario simulation and generate supplementary simulation evolution samples is as follows: First, perform uncertainty analysis on the cluster boundaries of the two-layer clustering model in the high-dimensional feature space. Assume the total dimension of the parameters in the selected boundary neighborhood is... The standardized offset distance of real-time production data relative to the center of high-quality clusters The statistical simulation evolution sample library shows the density of benchmark nodes near this location. Preset spatial smoothing constant Substitute into the formula Assume the reconstruction threshold is determined through cross-validation of historical data. Due to the current situation The model was determined to be insufficient in capturing risk in this area, and incremental learning was initiated. Feature vectors of this weak area were extracted. The range of initial perturbation parameters is calculated using the inverse mapping operator. Assuming the calculated air knife pressure fluctuation range is... .around Perform feedforward inference to generate root cause labels.

[0049] In one possible implementation, combining Figure 2 According to the collaborative control instruction set, the process parameters of the corresponding production equipment are adjusted to complete defect early warning and quality control, including: The feedforward compensation component and feedback correction component in the collaborative control instruction set are mapped to the action adjustment amount for the specific actuators of the production line. Using the time delay parameters in the inter-process quantitative causal model, the advance amount of the feedforward compensation component is calculated, and the corresponding instruction is sent to the upstream cold rolling or annealing process equipment with priority over the production cycle based on the advance amount. The action adjustment amount corresponding to the feedback correction component is sent to the galvanizing or aluminized zinc plating equipment of the current process in real time to realize the immediate correction of the current production status. The system monitors the online clustering status after the action adjustment is executed in real time, and determines the regression trend of the feature points relative to the boundary of the high-quality cluster. If the feature points regress to the boundary or interior of the high-quality cluster, the current production defect risk is determined to be eliminated, and the system switches to continuous monitoring mode. If the feature points do not enter the high-quality cluster and show a deviation trend, the system triggers the inter-process quantitative causal model to recalculate the sensitivity matrix and updates the collaborative control instruction set until the feature points regress.

[0050] In some implementations, the process of adjusting process parameters and executing closed-loop control based on the collaborative control instruction set is achieved through spatiotemporally aligned action issuance and dynamic status monitoring. First, the parameter adjustment vector in the collaborative control instruction set is decomposed into feedforward compensation components. With feedback correction component And the lead time for issuance is calculated using the time delay parameter in the inter-process quantitative causal model: ; in, The advance issuance time of the feedforward command, measured in seconds, is used to ensure that the adjustment action is synchronized with the running cycle of the strip steel. The physical distance from the exit of the preceding process to the current galvanizing point is measured in meters and is determined by the equipment installation drawings. The real-time linear velocity of the strip steel, in m / s, is calculated from the encoder pulse frequency of the tension roller at the unit's inlet. During execution, the regression trend of feature points relative to the boundaries of high-quality clusters is monitored in real time, and the regression rate index is calculated to determine the control effect. ; in, The regression rate of the current output is scored, and a value greater than 0 indicates that the feature point is moving towards a high-quality region; For the current moment Online clustering coordinates in high-dimensional space; The coordinates of the point at the previous sampling time; The preset coordinates of the high-quality cluster centers are derived from cluster analysis of historical data that meet the standards. The system sampling period is measured in seconds (s). If a feature point successfully regresses to the boundary of a high-quality cluster, the current production defect risk is considered eliminated, and the system switches to continuous monitoring mode. If the feature point does not enter a high-quality cluster and its regression rate... Continuously below the preset safety threshold Generally take If the current adjustment is deemed insufficient, the inter-process quantitative causal model will be triggered to recalculate the influence sensitivity matrix. The collaborative control instruction set is updated in real time according to the formula until the feature points are fully regressed. ; This closed-loop control mechanism based on regression trend perception ensures that the system can cope with complex nonlinear disturbances and achieves self-healing quality assurance in the production process.

[0051] For example, the specific calculation process for adjusting process parameters and executing closed-loop control based on the collaborative control instruction set is as follows: First, the collaborative control instruction set is decomposed. Assume the physical distance from the exit of the preceding process to the current galvanizing point... Real-time running linear speed of strip steel Calculate the lead time for issuance: Send the feedforward compensation component to the upstream equipment 40 seconds in advance. During the implementation of the control measures, the regression trend is monitored in real time. A sampling period is set. Preset high-quality cluster centers Assume the Euclidean distance from the feature point to the center at the previous time step is... The distance from the current moment Calculate the regression rate: .because The feature points are determined to be moving towards high-quality areas. If subsequent monitoring reveals... Reduced to 0.005, below the safety threshold. Furthermore, the feature points did not enter the optimal region, indicating insufficient adjustment. In this case, the sensitivity influence matrix is ​​retrieved. Assume its inverse matrix Current deviation . The action adjustment amount is updated accordingly until the feature points successfully regress to the boundary of the high-quality cluster.

[0052] It should be noted that all equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A method for early warning of defects in galvanized sheet production based on cluster analysis, characterized in that, The method includes: Historical process and quality data of the production line are obtained, and based on the physicochemical mechanism of the entire galvanizing process, a quantitative causal model describing the inter-process influence relationship is constructed. The inter-process quantitative causal model is used to simulate and extrapolate the disturbance scenarios of process parameters, generating a simulation evolution sample library containing virtual defect features and corresponding defect root cause labels. Real-time production data is fused with the simulation evolution sample library and input into a pre-built two-layer clustering model for calculation to obtain the current online clustering state. The two-layer clustering model includes a first-layer local clustering for identifying parameter fluctuations within a process and a second-layer global clustering for identifying quality trends across processes. When the online clustering status indicates that the current production process deviates from the preset high-quality cluster boundary, the shortest adjustment path for regressing the high-quality cluster is calculated by combining the inter-process quantitative causal model, and a collaborative control instruction set is generated. The high-quality cluster is a cluster center and influence range constructed based on quality-compliant samples in historical data. Based on the collaborative control instruction set, the process parameters of the corresponding production equipment are adjusted to complete defect early warning and quality control.

2. The method for early warning of defects in galvanized steel sheet production based on cluster analysis according to claim 1, characterized in that, The construction of the inter-process quantitative causal model describing the inter-process influence relationship includes: A pre-defined library of causal operators covering cold rolling, galvanizing, and aluminized zinc plating processes is provided. The causal operators include nonlinear transfer functions used to characterize the mapping relationship between upstream process variables and downstream quality status. Using the historical process data and quality data, the weight parameters and time delay parameters in the causal operator library are trained and calibrated offline to obtain trained causal operators; According to the preset galvanizing production process, the trained causal operators are logically connected in the direction of material flow to dynamically generate a multi-dimensional directed causal graph representing the cross-process influence path. The multidimensional directed causal graph is defined as the inter-process quantitative causal model. By introducing state-space equations into this model, the quantitative deduction and calculation of the evolution results from process parameter disturbances to quality defects are realized. The differentiability of the nonlinear transfer function is used to construct the inverse mapping operator of the inter-process quantitative causal model, which supports sensitivity analysis of production status and inverse parameter solving.

3. The method for early warning of defects in galvanized steel sheet production based on cluster analysis according to claim 2, characterized in that, The simulation evolution sample library generated, which includes virtual defect features and corresponding defect root cause labels, includes: A set of counteracting disturbance parameters covering cold rolling process fluctuations, coating thickness anomalies, and annealing temperature deviations is preset, and the set of counteracting disturbance parameters is injected into the inter-process quantitative causal model as an initial input variable; By utilizing the multidimensional directed causal graph in the inter-process quantitative causal model, forward inference calculations based on physical mechanisms are performed to simulate the transmission and accumulation of the adversarial disturbance parameters between processes, thereby obtaining the virtual defect characteristics of the final output. The generated virtual defect features are mapped and associated with the adversarial perturbation parameter combination that serves as the triggering source, and each virtual defect feature is assigned a clear defect root cause label. By traversing the value space of the adversarial perturbation parameter combination, several sets of simulation evolution samples with root cause labels are generated and collected into the simulation evolution sample library to assist the two-layer clustering model in delineating defect boundaries.

4. The method for early warning of defects in galvanized steel sheet production based on cluster analysis according to claim 1, characterized in that, The step of fusing real-time production data with the simulation evolution sample library and inputting it into a pre-constructed two-layer clustering model for calculation includes: The first layer of local clustering in the two-layer clustering model is used to independently cluster the real-time parameters within each process of cold rolling, annealing and coating, and to identify and output the state feature vector within the process. The second layer of global clustering in the two-layer clustering model is constructed. The second layer of global clustering receives the state feature vector within the process and introduces a time delay correlation operator across processes to identify the quality evolution trend of the entire process. The sample points in the simulation evolution sample library are used as reference nodes and projected into the high-dimensional feature space constructed by the two-layer clustering model. Data fusion is achieved by calculating the distribution density and position offset of the real-time production data relative to the reference nodes in the high-dimensional feature space. Based on the depth of the real-time production data falling into the preset sensitive risk area of ​​the simulation evolution sample library, the online clustering state, which includes the fluctuation pattern within the process, the overall quality trend, and the risk level score, is calculated.

5. The method for early warning of defects in galvanized steel sheet production based on cluster analysis according to claim 4, characterized in that, Running the two-layer clustering model includes: Through the first layer of local clustering, the kernel principal component analysis algorithm is used to map real-time production data to a high-dimensional feature space, and the process delay operator determined by the inter-process quantitative causal model is combined to perform step-by-step dimensionality reduction, and the local offset vector representing the fluctuation of single processes of cold rolling, annealing and coating is calculated. The local offset vector is input into the second-layer global clustering, and combined with the risk boundary defined by the defect root cause label in the simulation evolution sample library, the online clustering state characterizing the quality risk level of the entire process is calculated. The two-layer clustering model is used to calculate in real time the overlap and matching degree of the online clustering state with known samples in the simulation evolution sample library, and to evaluate the accuracy of the two-layer clustering model in capturing known risk features. When the capture accuracy value deviates from the performance benchmark determined by historical quality data, the inter-process quantitative causal model is triggered to re-execute the disturbance scenario simulation based on the current production conditions, generate supplementary simulation evolution samples, and incrementally reconstruct the two-layer clustering model.

6. The method for early warning of defects in galvanized steel sheet production based on cluster analysis according to claim 5, characterized in that, The shortest adjustment path for regressing the high-quality cluster, calculated using the inter-process quantitative causal model, includes: Within the high-dimensional feature space constructed by the two-layer clustering model, the Euclidean distance vector between the feature points representing the current online clustering state and the center of the high-quality cluster is calculated as the initial bias vector; The inter-process quantitative causal model is invoked to calculate the sensitivity matrix of the influence of each process parameter on the initial deviation vector, and the search space of the regression path is determined by combining the physical constraints of the equipment actuator. Based on the aforementioned sensitivity matrix, and with the optimization of process adjustment cost as the objective function, the parameter adjustment vector that causes the feature points to regress to the boundary of the high-quality cluster is obtained by reverse calculation, thereby generating the shortest adjustment path. The parameter adjustment vector is decomposed into a feedforward compensation component for the preceding process and a feedback correction component for the current process. A collaborative control instruction set is generated through the temporal coupling mapping between the components.

7. The method for early warning of defects in galvanized steel sheet production based on cluster analysis according to claim 6, characterized in that, The step of decomposing the parameter adjustment vector into feedforward compensation components for the preceding process includes: The high-quality clusters are mapped to target quality index ranges in the inter-process quantitative causal model, and the target state trajectory of future production cycles is set according to the production plan. A multi-objective optimization cost function is constructed with the objectives of optimizing the deviation between the predicted quality and the target state trajectory, and optimizing the adjustment range of process parameters. In the multidimensional directed causal graph of the inter-process quantitative causal model, backpropagation calculation based on the chain rule is performed on the multi-objective optimization cost function to obtain the adjustment amount of the preceding process parameters that optimizes the cost function; The adjustment amount of the preceding process parameters is defined as a feedforward compensation component, which is used to proactively intervene in the upstream cold rolling or annealing process before the feature point deviates from the boundary.

8. The method for early warning of defects in galvanized steel sheet production based on cluster analysis according to claim 5, characterized in that, Triggering the inter-process quantitative causal model to re-execute the disturbance scenario simulation based on the current production conditions, generating supplementary simulation evolution samples, and incrementally reconstructing the two-layer clustering model includes: When the capture accuracy value deviates from the performance benchmark, uncertainty analysis is performed on the clustering boundary of the two-layer clustering model in the high-dimensional feature space to identify weak boundary regions with sparse sample distribution. Obtain the spatial feature vector corresponding to the weak boundary region, and use the inter-process quantitative causal model to reverse deduce the range of initial perturbation parameters that cause the spatial feature vector to be generated, which is defined as a directional perturbation mode as a supplement to the adversarial perturbation parameter combination. Forward inference calculations are performed around the directional perturbation mode to generate supplementary simulation evolution samples with root cause labels, and the supplementary simulation evolution samples are injected into the simulation evolution sample library to achieve accurate reinforcement of the identification boundary of the two-layer clustering model.

9. The method for early warning of defects in galvanized steel sheet production based on cluster analysis according to claim 6, characterized in that, Based on the aforementioned collaborative control instruction set, the process parameters of the corresponding production equipment are adjusted to complete defect early warning and quality control, including: The feedforward compensation component and feedback correction component in the collaborative control instruction set are mapped to the action adjustment amount for the specific actuators of the production line. Using the time delay parameters in the inter-process quantitative causal model, the advance amount of the feedforward compensation component is calculated, and the corresponding instruction is sent to the upstream cold rolling or annealing process equipment with priority over the production cycle based on the advance amount. The action adjustment amount corresponding to the feedback correction component is sent to the galvanizing or aluminized zinc plating equipment of the current process in real time to realize the immediate correction of the current production status. The system monitors the online clustering status after the action adjustment is executed in real time, and determines the regression trend of the feature points relative to the boundary of the high-quality cluster. If the feature points regress to the boundary or interior of the high-quality cluster, the current production defect risk is determined to be eliminated, and the system switches to continuous monitoring mode. If the feature points do not enter the high-quality cluster and show a deviation trend, the system triggers the inter-process quantitative causal model to recalculate the sensitivity matrix and updates the collaborative control instruction set until the feature points regress.

10. A defect early warning system for galvanized steel sheet production based on cluster analysis, characterized in that, The system is used in the cluster analysis-based early warning method for galvanized steel sheet production defects as described in any one of claims 1-9, and the system comprises: The causal modeling module is used to acquire historical process data and quality data of the production line, and to construct a quantitative causal model between processes based on the physicochemical mechanism of the entire galvanizing process. The simulation evolution module is used to simulate and extrapolate the disturbance scenarios of process parameters using the inter-process quantitative causal model, and generate a simulation evolution sample library containing virtual defect features and corresponding defect root cause labels. The two-layer clustering analysis module is used to integrate real-time production data with the simulation evolution sample library and input it into a pre-built two-layer clustering model for calculation to obtain the current online clustering status. The two-layer clustering model includes a first-layer local clustering for identifying parameter fluctuations within the process and a second-layer global clustering for identifying quality trends across processes. The path decision module is used to calculate the shortest adjustment path to regress the high-quality cluster when the online clustering status indicates that the current production process deviates from the preset high-quality cluster boundary, in combination with the inter-process quantitative causal model, and generate a collaborative control instruction set. The high-quality cluster is a cluster center and influence range constructed based on quality-compliant samples in historical data. The execution and control module is used to adjust the process parameters of the corresponding production equipment according to the collaborative control instruction set, and to complete defect early warning and quality control.

Images