A coating process parameter self-optimizing system and method based on fusion model self-learning

Abstract

The present application relates to the field of coating parameter self-optimization, and discloses a coating process parameter self-optimization system and method fusing model self-learning, wherein a coating process parameter self-optimization method fusing model self-learning comprises the following steps: collecting process parameters, production conditions, equipment power states and detection result data of multiple batches of coating production in real time; aligning and correlating the data according to the process sequence to form a unique mapping relationship; reversely determining corresponding process parameter nodes and working condition node sets to generate a consistency index; jointly comparing the consistency index and the corresponding data chain distribution to output corresponding working condition drift identifiers; introducing the consistency index and the working condition drift identifiers as constraint conditions into a parameter update model to output a constrained parameter update result; and writing the corresponding new round of detection results into the unique mapping relationship to generate an incremental updated mapping relationship and a reachable state space. The present application has the advantage of improving the stability of the optimization result.

Description

Technical Field

[0001] This invention relates to the field of coating parameter self-optimization, specifically to a coating process parameter self-optimization system and method that integrates model self-learning. Background Technology

[0002] In the optimization of existing coating process parameters, parameter adjustments are usually made based on current test results. There is a lack of data chain analysis on the correlation between test results and historical production data, making it impossible to infer the rationality of the generation process from the test results, and thus calibrate the accuracy of the current data acquisition.

[0003] Existing solutions struggle to verify the consistency of detection results with real-time production conditions and power status. When process conditions or acquisition conditions drift, optimization strategies are still executed based on distorted data, which can easily lead to deviations in the optimization direction of process parameters that are difficult to identify in a timely manner, thereby affecting the stability and reliability of coating process parameter optimization results.

[0004] Therefore, it is necessary to design a self-optimization system and method for coating process parameters based on a fusion model that improves the stability of optimization results. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a coating process parameter self-optimization system and method that integrates model self-learning, which has the advantage of improving the stability of optimization results and solves the problems mentioned in the background technology.

[0006] To achieve the aforementioned goal of improving the stability of optimization results, this invention provides the following technical solution: a self-optimization method for coating process parameters that integrates model self-learning, comprising the following steps: The system collects process parameters, production conditions, equipment power status, and test results data for multiple batches of coating production in real time. The data is aligned and associated according to the process sequence to construct a directed data chain structure. A unique mapping relationship is formed by connecting the directed edges that meet the process reachability constraints and the minimum energy path calculation. Based on the unique mapping relationship, the corresponding process parameter nodes and operating condition node sets are determined in reverse. Within the reachable state space defined by the node state vector and directed edge constraints, the deviation of the current detection result from the reachable state space is calculated by combining historical data distribution weighting and dynamic operating condition association algorithms, and a consistency index characterizing the credibility of the detection result is generated. The consistency index is jointly compared with the data chain distribution corresponding to the current production conditions, real-time equipment power status and historical stable operating conditions. The drift probability of the current operating condition is calculated by combining the dynamic joint probability distribution model, and the corresponding operating condition drift identifier is output. The consistency index and operating condition drift identifier are introduced as constraints into the parameter update model. The process parameter update magnitude is calculated by combining reinforcement learning algorithm. The process parameters are updated only when the consistency index meets the threshold and is not marked as operating condition drift, and the constrained parameter update results are output. The updated results of the constrained process parameters and the corresponding new round of test results are written into the unique mapping relationship. The mapping relationship between parameter nodes, operating condition nodes and result nodes is incrementally corrected, and the reachable state space is updated synchronously to generate the incrementally updated mapping relationship and reachable state space.

[0007] Preferably, the process of forming a unique mapping relationship is as follows: The process parameters, production conditions, equipment power status and test results data collected during multiple batches of coating production are indexed in time sequence according to the process order, and a multi-dimensional state vector is generated for each node. Construct an reachability constraint matrix between nodes, and encode process feasibility rules, equipment capacity constraints, and dependencies between operating conditions as constraint edges; Based on the reachability constraint matrix, the minimum energy path algorithm is used to generate the optimal path set from the initial working condition node to each detection result node; The path set is recursively filtered and redundancy eliminated, removing paths that deviate significantly from the historical stable operating condition distribution. The path confidence is calculated by combining the node state vectors, and a unique mapping relationship is output, consisting of process parameter nodes, operating condition nodes, detection result nodes and constraint directed edges.

[0008] Preferably, the process of reversely determining the corresponding set of process parameter nodes and operating condition nodes is as follows: Based on the unique mapping relationship, reverse path backtracking is performed on each detection result node; During the backtracking process, the multi-dimensional state vector of each node is analyzed to extract parameter value range, equipment power characteristics, operating condition category and historical stability index. Simultaneously, a dynamic reachability matrix between nodes is constructed, encoding process reachability rules, equipment capability constraints, and temporal dependencies between nodes into matrix edge weights. On the backtracking path, the path confidence is calculated by combining the node state deviation, historical distribution consistency and dynamic correlation of operating conditions, and outputting a complete set of process parameter nodes and operating condition nodes.

[0009] Preferably, the process of generating a consistency index characterizing the reliability of the detection results is as follows: In the set of process parameter nodes and operating condition nodes, a state vector is generated for each detection result node to represent parameter values, operating condition categories, equipment power characteristics, environmental factors, and historical stability. Embed the state vector into the historical process data distribution space and calculate the deviation in the reachable state space; Based on the dynamic working condition association algorithm, the deviation of the current node is time-weighted with the state vectors of the previous and next time nodes to evaluate the abnormal spread trend and potential drift path, and form a deviation time series. The deviation time series, dynamic correlation of working conditions and historical stability indicators are weighted and integrated to construct a fusion vector, and finally generate a consistency index that represents the credibility of the test results.

[0010] Preferably, the process of jointly comparing the consistency index with the data chain distribution corresponding to the current production conditions, real-time equipment power status, and historical stable operating conditions is as follows: The consistency index of the current detection results is matched with the data chain distribution of the corresponding production condition nodes, real-time equipment power status nodes and historical stable operating condition nodes to form a multi-dimensional comparison matrix. Based on the multidimensional data representation, the data distribution of each dimension is compared with the historical stable distribution; The matching degree is calculated by a joint probability model, and a joint matching vector reflecting the overall consistency of the operating conditions is generated. Based on the matching vector, according to the preset weights and operating condition sensitivity, the operating condition drift sensitivity score is calculated, and the joint comparison result of drift determination is output.

[0011] Preferably, the process of outputting the corresponding operating condition drift indicator is as follows: Based on the joint comparison results and joint matching vectors, a multidimensional state representation of the current working condition is constructed and input into the dynamic joint probability distribution model; The model is used to jointly model the distribution of historical data, temporal correlation, and dynamic dependencies between operating conditions, and to generate the drift probability distribution of the current operating condition in the reachable state space. Extract the mean and standard deviation of drift probabilities for each dimension from historical stable operating condition data to establish historical drift thresholds; The drift probability distribution generated by the model is compared with the historical drift threshold, and the operating condition drift index is calculated by combining short-term abnormal fluctuations and long-term trend changes. Based on the operating condition drift index and dynamic weighting rules, determine whether the current operating condition exceeds the historical acceptable drift range, and output the corresponding operating condition drift flag.

[0012] Preferably, the process of introducing consistency indicators and operating condition drift indicators as constraints into the parameter update model is as follows: For each process parameter node, extract the current consistency index and the corresponding operating condition drift identifier, and combine them with the historical operating condition stability characteristics to construct the parameter update input state vector; The state vector input parameters are used to update the model. Constraint logic is established within the model, allowing the corresponding process parameters to be updated only when the consistency index reaches the set threshold and the operating condition is not calibrated as drift. Based on the constraint logic, and combined with the state vectors of the preceding and following nodes and the dynamic working condition dependencies, the parameter update magnitude is weighted and adjusted, and the parameter update input vector after constraint and weighting processing is output.

[0013] Preferably, the process of outputting the constrained parameter update results is as follows: The parameter update input vector is used as the reinforcement learning state vector, the process parameter adjustment range is used as the action space, and a comprehensive reward function is defined, which includes the consistency index of detection results, the operating condition drift index, and the historical process stability index. During the reinforcement learning training process, the probability of action selection is dynamically adjusted according to different process parameter dimensions and operating conditions using a priority gating strategy. By incorporating an adaptive learning rate mechanism, the magnitude of parameter updates is constrained; The state vectors of the preceding and following nodes and the dynamic dependencies of the working conditions are incorporated into the training process to suppress local abnormal fluctuations or sudden drift behaviors with weights. Process parameter updates are performed only when the consistency index reaches a preset threshold and the current operating condition is not labeled as drift, outputting constrained, reinforcement learning-optimized process parameter update results.

[0014] Preferably, the process of generating the incrementally updated mapping relationship and reachable state space is as follows: Write the updated process parameters and the corresponding new test results into a unique mapping relationship; Incremental updates are performed based on the state vectors of parameter nodes, working condition nodes, and result nodes, while simultaneously recording the changes in path weights and constraint edges between nodes. Incremental correction of the mapping relationship is carried out by dynamically adjusting the path weight, optimizing the constraint edge, and combining historical working condition stability indicators. Based on the corrected mapping relationship, the reachable state space is updated synchronously, and the newly added or updated nodes and constraint edges are incorporated into the state space constraints. By combining historical stable operating conditions with current detection results, the state reachability boundary is recalculated, and the mapping relationship and reachability state space after incremental correction and reachability state update are output.

[0015] This invention also discloses another technical solution, a coating process parameter self-optimization system integrating model self-learning, comprising: Data construction module: Real-time collection of process parameters, production conditions, equipment power status and test results data of multiple batches of coating production, alignment and association of data according to process sequence and construction of a uniquely mapped directed data chain; Node backtracking module: Based on the unique mapping relationship, it reversely determines the set of process parameter nodes and operating condition nodes, and calculates the degree of deviation of the detection results by combining historical data weighting and dynamic operating condition association algorithms within the reachable state space, and generates consistency index; Drift assessment module: It compares the consistency index with the current production conditions, equipment power status and historical stable operating condition data, and calculates the operating condition drift probability through a dynamic joint probability model to generate the corresponding operating condition drift identifier. Parameter update module: The consistency index and operating condition drift identifier are used as constraints to input the parameter update model, and the process parameter update magnitude is calculated by combining reinforcement learning. The update is performed only when the consistency index meets the threshold and the drift is not calibrated, and the constrained parameter update result is output. Mapping Correction Module: Writes the updated results of constrained process parameters and new detection results into a unique mapping relationship, performs incremental correction on node mapping and synchronously updates the reachable state space, and generates the incrementally updated mapping relationship and reachable state space.

[0016] Compared with existing technologies, this invention provides a coating process parameter self-optimization system and method that integrates model self-learning, which has the following beneficial effects: This invention constructs a directed data chain structure for multiple batches of coating production process parameters, production conditions, equipment power status, and test results. A unique mapping relationship is formed using process reachability constraints and minimum energy path calculations, achieving precise association and traceability management of production data. Based on this, by reversely determining process parameters and operating condition nodes, and combining historical data distribution and dynamic operating condition association algorithms to calculate the deviation of test results, a consistency index is generated. This index is then jointly judged with a dynamic joint probability distribution model to determine operating condition drift, achieving real-time reliable assessment of the process status. Furthermore, the consistency index and operating condition drift identifier are introduced into the parameter update model, and a reinforcement learning algorithm is used to optimize process parameters under constraints. This ensures both the safety of parameter updates and adaptive optimization of the process. Finally, by incrementally writing the update results into the mapping relationship and synchronously updating the reachable state space, dynamic maintenance and iterative optimization of the data chain are achieved. This tightly integrates process parameter optimization with data-driven state assessment, not only improving the stability and consistency of the coating process but also significantly increasing production efficiency, reducing reliance on manual adjustments, and providing dynamic identification and self-correction capabilities for abnormal operating conditions. This enables intelligent and highly reliable management of coating production. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the method of the present invention; Figure 2 This is a schematic diagram of the structure of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1: Please refer to Figure 1 As shown in the figure, a self-optimization method for coating process parameters based on fusion model self-learning in an embodiment of the present invention includes the following steps: S1: Real-time acquisition of process parameters, production conditions, equipment power status and test results data of multiple batches of coating production, alignment and association of data according to process sequence, construction of directed data chain structure, and formation of unique mapping relationship through directed edge connection that satisfies process reachability constraints and minimum energy path calculation.

[0020] The process of forming a unique mapping relationship in S1 is as follows: The process parameters, production conditions, equipment power status, and test results collected during multiple batches of coating production are indexed in time sequence according to the process order, and a multi-dimensional state vector is generated for each node. The process parameters, production conditions, equipment power status, and test results involved in multiple batches of coating production are collected in real time. The production data of each batch is uniformly connected to the data processing system through the process control terminal, sensor network, or data acquisition module. The collected data is indexed in time sequence according to the process execution order, and a corresponding multi-dimensional state vector is generated for the data at each time node. This vector includes process parameter values, equipment power indicators, production environment conditions, and preliminary test result information, ensuring that the data at each node is uniquely identified in path generation and mapping. A reachability constraint matrix is ​​constructed between nodes, and process feasibility rules, equipment capacity constraints, and dependencies between operating conditions are encoded as constraint edges. Based on the process feasibility rules, equipment capacity constraints, and dependencies between operating conditions, a reachability constraint matrix is ​​constructed between nodes. Each element in the matrix represents the reachability and constraints from one node to another. This can be understood as follows: if the process sequence, equipment capacity, and operational safety requirements are met between nodes, the matrix element value is reachable; if there is a process conflict or capacity exceedance, it is marked as unreachable. By combining process rules with historical data, penalty values ​​for deviations from historical stable operating conditions are encoded in the matrix, providing a quantitative basis for path selection. Based on the reachability constraint matrix, the minimum energy path algorithm is used to generate the optimal path set from the initial operating condition node to each detection result node. Based on the reachability constraint matrix, the dynamic programming algorithm based on weight optimization is used to calculate the optimal path set from the initial operating condition node to each detection result node. The weight of the path is determined by the node state vector and the constraint matrix. Taking into account process stability, energy consumption minimization and equipment utilization, multiple candidate paths that meet process feasibility and deviate from historical stable operating conditions are generated, forming a preliminary mapping structure. The path set is recursively filtered and redundancy eliminated to remove paths that deviate significantly from the historical stable operating condition distribution. Based on the generated path set, further recursive filtering is performed. The filtering rules include removing paths that deviate too much from the historical operating condition distribution, have abnormal node states, or have excessively high energy weights. At the same time, duplicate or highly similar paths are merged. The degree of deviation can be quantified by calculating the Mahalanobis distance or Euclidean distance between the path and the historical operating condition distribution. Paths that exceed the threshold are removed or their weights are adjusted to ensure that each path in the path set is representative and executable. The path confidence is calculated by combining the node state vectors, and a unique mapping relationship consisting of process parameter nodes, operating condition nodes, detection result nodes, and constrained directed edges is output. After filtering, the confidence of each path is calculated by combining the node state vectors. The confidence calculation can be based on the statistical distribution of historical process data, the consistency of detection results, and the equipment power stability index through weighted fusion. According to the path confidence, the optimal path is selected to form a unique mapping relationship. This mapping relationship consists of process parameter nodes, operating condition nodes, detection result nodes, and coded constrained directed edges, which can completely represent the correspondence from production parameters to detection results.

[0021] S2: Based on the unique mapping relationship, the corresponding process parameter nodes and operating condition node sets are determined in reverse. Within the reachable state space defined by the node state vector and directed edge constraints, the deviation of the current detection result from the reachable state space is calculated by combining historical data distribution weighting and dynamic operating condition association algorithms, and a consistency index characterizing the credibility of the detection result is generated.

[0022] The process of reversely determining the corresponding set of process parameter nodes and operating condition nodes in S2 is as follows: Based on the unique mapping relationship, reverse path backtracking is performed on each detection result node. Based on the unique mapping relationship, each detection result node is used as the backtracking starting point. By traversing the directed edges in reverse, the backtracking is gradually performed to the corresponding initial operating condition node. During the backtracking process, each process parameter node and operating condition node passed along the way is recorded to form a preliminary candidate node path set. This process can be implemented by depth-first search or reverse breadth-first search to ensure that all possible parameter and operating condition combinations are covered, while retaining the path time sequence information for confidence calculation and screening. During the backtracking process, the multidimensional state vector of each node is analyzed to extract parameter value ranges, equipment power characteristics, operating condition categories, and historical stability indicators. For each node on the backtracking path, its multidimensional state vector is extracted and analyzed, including process parameter value ranges, such as temperature, humidity, and spraying rate; equipment power characteristics, such as heating power and motor power; operating condition categories, such as preheating, spraying, and curing; and historical stability indicators, such as the frequency and fluctuation range of the state in historical production batches. The analysis results are used to evaluate the reliability and rationality of the path nodes. Simultaneously, a dynamic reachability matrix between nodes is constructed, encoding process reachability rules, equipment capability constraints, and temporal dependencies between nodes into matrix edge weights. Based on the backtracking path node set, a dynamic reachability matrix is ​​constructed, where the rows and columns of the matrix correspond to nodes, and the edge weights represent the strength of the reachability constraint from one node to another. The matrix edge weights are combined with process reachability rules, such as process sequence constraints and material compatibility, equipment capability constraints, such as maximum power and maximum speed limits, and temporal dependencies between nodes, such as the requirement that a preceding node must be completed before subsequent nodes can be executed. The matrix is ​​dynamically updated to reflect the executability under different batch operating conditions or equipment states. On the backtracking path, a weighted summation algorithm is applied to calculate the path confidence by combining node state deviation, historical distribution consistency, and dynamic correlation of operating conditions, outputting a complete set of process parameter nodes and operating condition nodes. For each path on the backtracking path, a weighted summation is performed to calculate the confidence of each path, combining node state deviation, deviation of the current node state from historical stable values, historical distribution consistency, frequency of node state occurrence in historical batches, dynamic correlation of operating conditions, and temporal and functional correlation of the node with other operating condition nodes. Paths with high confidence indicate that the combination of process parameters and operating condition nodes conforms to historical stable patterns, with low deviation risk, and can be used as reliable candidates. The weights of the weighted summation can be adaptively adjusted according to process importance indicators, equipment sensitivity, and operating condition safety. Backtracking paths are filtered based on confidence scores, retaining paths with high confidence and conforming to dynamic reachability constraints. Finally, a complete set of process parameter nodes and operating condition nodes is output, which includes the process parameter values ​​of the nodes, operating condition types, equipment states, and inter-node constraint information.

[0023] The process of generating a consistency index representing the reliability of the detection results in S2 is as follows: Within the set of process parameter nodes and operating condition nodes, a state vector is generated for each detection result node, representing parameter values, operating condition categories, equipment power characteristics, environmental factors, and historical stability. Based on the set of process parameter nodes and operating condition nodes, a state vector is generated for each detection result node. This vector comprehensively represents the node's process parameter values, such as temperature, humidity, and spraying rate; operating condition categories, such as preheating, spraying, and curing; equipment power characteristics, such as heating power and motor power; environmental factors, such as air humidity and pressure; and historical stability indicators, such as the frequency and fluctuation range of the node in historical batches. The multidimensional structure of the state vector ensures a comprehensive description of all process and environmental information on which the detection results depend. The state vector is embedded into the historical process data distribution space, and the deviation in the reachable state space is calculated. The state vector of each detection result node is embedded into the historical process data distribution space. By comparing the current node state with the historical data distribution, the deviation of the node in the reachable state space is calculated, including the degree of deviation of each parameter from the historical mean and variance, and the overall confidence decrease of the node combination state. The larger the deviation value, the more obvious the deviation of the current detection result from the historical stable process mode, thus initially reflecting potential process anomalies or equipment fluctuations. Based on the dynamic operating condition association algorithm, the deviation of the current node is weighted with the state vectors of the preceding and following time nodes to evaluate the anomaly propagation trend and potential drift path, forming a deviation time series. For example, by using exponential weighted moving average or time decay weighted algorithm, the influence of the state deviation of neighboring nodes on the current node is included in the calculation to form a deviation time series, which can reflect the propagation dynamics of deviation in the process, identify anomaly propagation or local drift phenomena, and provide time series information for credibility assessment. By weighted and integrated time-series deviations, dynamic correlations of operating conditions, and historical stability indicators, a fusion vector is constructed, ultimately generating a consistency index characterizing the reliability of the test results. The weighting coefficients can be adaptively adjusted based on process criticality, equipment sensitivity, and operating condition importance to balance the influence of historical stability and dynamic deviation information. The fusion vector, by integrating multi-dimensional information, achieves a comprehensive characterization of the reliability of the test results, enabling unified measurement of information from different sources and time points. Finally, a consistency index is calculated for each test result node based on the fusion vector to characterize its reliability. A high index value indicates a high degree of consistency between the test results and historical process patterns, low deviation, and low anomaly risk; a low index value suggests potential process deviations or anomalies. The consistency index is used for process optimization, parameter adjustment, and anomaly alarms.

[0024] S3: Jointly compare the consistency index with the data chain distribution corresponding to the current production conditions, real-time equipment power status, and historical stable operating conditions. Combine the dynamic joint probability distribution model to calculate the drift probability of the current operating condition and output the corresponding operating condition drift identifier.

[0025] In S3, the process of jointly comparing consistency indicators with the data chain distribution corresponding to current production conditions, real-time equipment power status, and historical stable operating conditions is as follows: The consistency index of the current test results is matched with the data chain distribution of the corresponding production condition nodes, real-time equipment power status nodes, and historical stable operating condition nodes to form a multi-dimensional comparison matrix. The consistency index of the current test results is matched with the data chain distribution of the corresponding production condition nodes, such as raw material batches and ambient temperature and humidity, real-time equipment power status nodes, such as heating power, motor power, and pump flow, and historical stable operating condition nodes. During the matching process, the data of each type of node is arranged according to time and process order to form a multi-dimensional comparison matrix containing the status, parameter values, and historical stability information of each node. This matrix can simultaneously represent the correspondence between the current process status and the historical stable mode in different dimensions. Based on the multidimensional data representation, the data distribution of each dimension is compared with the historical stable distribution; independent data distribution analysis is performed on each dimension in the multidimensional comparison matrix, and the current production data, including consistency indicators, process parameters, and equipment status, is compared with the historical stable distribution. The comparison method uses statistical analysis, such as mean, variance, quantiles, and distribution similarity calculation, to quantify the degree to which the current node deviates from the historical pattern. The matching degree is calculated by a joint probability model to generate a joint matching vector that reflects the overall consistency of the operating conditions. The deviation of each dimension is input into the joint probability model, such as a multidimensional Gaussian model, a Bayesian network, or a conditional random field, to calculate the overall matching degree between the current detection result and the historical stable operating conditions. The joint probability model can comprehensively consider the correlation and dependence between different dimensions and output a joint matching vector that reflects the overall consistency of the operating conditions, including the matching degree value and the weight distribution of each dimension, which can be directly used for operating condition drift assessment. Based on the matching vector, according to preset weights and operating condition sensitivity, the operating condition drift sensitivity score is calculated, and the joint comparison result of drift judgment is output. The deviation degree of each dimension is input into the joint probability model, such as a multidimensional Gaussian model, Bayesian network or conditional random field, to calculate the overall matching degree between the current detection result and the historical stable operating condition. The joint probability model can comprehensively consider the correlation and dependence between different dimensions and output a joint matching vector that reflects the overall operating condition consistency. This vector contains the matching degree value and the weight distribution of each dimension. Based on the joint matching vector, the joint comparison result of drift judgment is further generated, including the operating condition drift sensitivity score. The output result can be used for real-time process monitoring, abnormal alarm, parameter adjustment or intelligent control decision-making, providing the production system with reliable operating condition assessment based on multi-source information and historical stable patterns.

[0026] The process of outputting the corresponding operating condition drift flag in S3 is as follows: Based on the joint comparison results and joint matching vectors, a multidimensional state representation of the current operating condition is constructed and input into the dynamic joint probability distribution model. Based on the joint comparison results and joint matching vectors, the operating condition information, process parameters, consistency index, equipment power status and environmental factors of the current detection results are integrated into a multidimensional state representation. This state representation can comprehensively describe the actual situation of the current operating condition in each dimension and its deviation from the historical stable pattern. This multidimensional state is then input into the dynamic joint probability distribution model. By jointly modeling the distribution of historical data, temporal correlation, and dynamic dependencies between operating conditions, a drift probability distribution of the current operating condition within the reachable state space is generated. A dynamic joint probability distribution model is also used to jointly model the distribution of historical data, temporal correlation, and dynamic dependencies between operating conditions. The modeling can employ multidimensional Gaussian processes, Bayesian networks, or conditional random fields, which can capture the correlation between states in each dimension and their dynamic characteristics over time, thereby generating the drift probability distribution of the current operating condition within the reachable state space, reflecting the deviation risk of the current operating condition relative to historical stable operating condition patterns. Extract the mean and standard deviation of drift probability for each dimension from historical stable operating data to establish historical drift thresholds; extract the mean and standard deviation of drift probability for each dimension from historical stable operating data, and combine the fluctuation characteristics of historical data and process stability indicators to establish historical drift thresholds for each dimension. The thresholds can be established by using a statistical method of adding or subtracting several times the standard deviation from the mean, or by adaptively adjusting according to the process criticality and quality sensitivity, to determine the normal range of the current drift probability. The drift probability distribution generated by the model is compared with the historical drift threshold, and the operating condition drift index is calculated by combining short-term abnormal fluctuations and long-term trend changes. The drift probability distribution generated by the dynamic joint probability model is compared with the historical drift threshold, and the operating condition drift index is calculated by combining short-term abnormal fluctuations and long-term trend changes. The operating condition drift index comprehensively reflects the degree to which the current operating condition exceeds the historical stable range and the deviation trend in various dimensions. The higher the value, the more obvious the drift. The index calculation can use weighted linear combination, fuzzy comprehensive evaluation or exponential smoothing method to balance short-term fluctuations and long-term trends. The system determines whether the current operating condition exceeds the historical acceptable drift range based on the operating condition drift index and dynamic weighting rules, and outputs the corresponding operating condition drift flag. If the drift index exceeds the preset threshold or the adaptive historical limit, it is determined as an abnormal drift and the corresponding operating condition drift flag is output. If it does not exceed the limit, a normal flag is output. The drift flag can be used for real-time process monitoring, abnormal alarms, process parameter adjustment, or intelligent control decisions to realize the management of the production system and the evaluation of operating conditions.

[0027] S4: Introduce the consistency index and operating condition drift identifier as constraints into the parameter update model, and combine reinforcement learning algorithm to calculate the process parameter update magnitude. Update the process parameters only when the consistency index meets the threshold and is not marked as operating condition drift, and output the constrained parameter update results.

[0028] In S4, the process of introducing consistency indicators and operating condition drift indicators as constraints into the parameter update model is as follows: For each process parameter node, extract the current consistency index and the corresponding operating condition drift identifier, and combine them with historical operating condition stability characteristics to construct a parameter update input state vector; for each process parameter node, extract the current consistency index and the corresponding operating condition drift identifier, and combine them with the stability characteristics of the node in historical operating conditions, such as historical deviation magnitude, frequency of occurrence, and fluctuation range of key process parameters, to construct a parameter update input state vector, which includes the current value of the process parameter, consistency index, drift identifier, historical stability data, and the time series information of the node; The state vector input parameters are used to update the model. Constraint logic is established within the model to allow the corresponding process parameters to be updated only when the consistency index reaches a set threshold and the operating condition is not marked as drifting. The constructed state vector input parameters are used to update the model. Constraint logic rules are established within the model to stipulate that the corresponding process parameters are only allowed to be updated when the consistency index reaches a set threshold and the corresponding operating condition is not marked as drifting. If the consistency index is lower than the threshold or the operating condition is judged to be drifting, the parameter update is prohibited, ensuring that parameter updates are only performed when the process is stable and the test results are reliable. Based on the constraint logic, and combining the state vectors of preceding and following nodes with dynamic operating condition dependencies, the parameter update magnitude is weighted and adjusted, outputting a parameter update input vector after constraint and weighting processing. Based on the constraint logic, and combining the state vectors of preceding and following nodes with dynamic operating condition dependencies, the magnitude of the allowed update process parameters is weighted and adjusted. The weighting strategy can adjust the parameter update magnitude based on historical stability, temporal dependencies between nodes, and operating condition sensitivity. For example, for critical process nodes or parameters that are highly sensitive to the final detection results, a smaller step size can be used to reduce risk, while for non-critical nodes or parameters with high stability, a larger step size can be used to speed up optimization. Based on the constraint logic and weighting adjustment results, the final processed parameter update input vector is generated.

[0029] The process of outputting the constrained parameter update results in S4 is as follows: The parameter update input vector is used as the reinforcement learning state vector, and the process parameter adjustment range is used as the action space. A comprehensive reward function is defined, which includes the detection result consistency index, the operating condition drift index, and the historical process stability index. The parameter update input vector is used as the reinforcement learning state vector, which contains the current value of the process parameter, the consistency index, the operating condition drift indicator, the historical stability information, and the node temporal features. The action space is defined as the adjustment range of each process parameter, that is, the range of update actions to be performed on each parameter. The reinforcement learning training process uses a comprehensive reward function, which simultaneously considers the detection result consistency index, the operating condition drift index, and the historical process stability index, to ensure that the training objective is to optimize detection consistency. During reinforcement learning training, a priority gating strategy is used to dynamically adjust the action selection probability based on different process parameter dimensions and operating conditions. For example, for critical process parameters or parameters that are highly sensitive to the final quality, the priority selection probability is increased during training, while the action frequency is reduced for non-critical parameters, so as to achieve a balance between key optimization and risk control. The adaptive learning rate mechanism constrains the parameter update magnitude. The learning rate can be adaptively adjusted according to the process stability index, the historical fluctuation magnitude of the node, and the current consistency index. For example, when the consistency index is low or the operating condition is close to the drift boundary, the learning rate is automatically reduced to reduce the parameter update magnitude, thereby avoiding process instability caused by large adjustments. The state vectors of preceding and following nodes and the dynamic dependencies of operating conditions are incorporated into the training process to suppress local abnormal fluctuations or sudden drift behaviors with weights. During the training process, the state vectors of preceding and following nodes and the dynamic dependencies of operating conditions are incorporated into the calculation to suppress local abnormal fluctuations or sudden drift behaviors with weights. For example, when the current node experiences a short-term abnormal deviation, but the preceding and following nodes are in stable operating conditions, the contribution weight of the abnormal node's action to parameter updates is reduced to prevent single-point anomalies from having too much impact on the overall update strategy. Process parameter updates are performed only when the consistency index reaches a preset threshold and the current operating condition is not marked as drift. The output is a constrained process parameter update result optimized by reinforcement learning. When performing parameter updates, the constraint logic is strictly followed. Process parameter updates are only allowed when the consistency index reaches a preset threshold and the current operating condition is not marked as drift. The output is a constrained process parameter update vector optimized by reinforcement learning, which includes the update magnitude of each parameter and the corresponding state information. It can be directly used by the process control module to ensure that no risky operations are performed under abnormal or drift conditions.

[0030] S5: Write the updated results of the constrained process parameters and the corresponding new round of test results into the unique mapping relationship, perform incremental correction on the mapping relationship between parameter nodes, operating condition nodes and result nodes, and update the reachable state space synchronously to generate the incrementally updated mapping relationship and reachable state space.

[0031] The process of generating the incrementally updated mapping relationship and reachable state space in S5 is as follows: Write the updated process parameters and the corresponding new test results into the unique mapping relationship. The writing process includes aligning the updated parameter nodes, operating condition nodes, and test result nodes with the existing mapping relationship, while keeping the temporal relationship and path identifier between nodes unchanged, to ensure that the new data can be seamlessly integrated into the existing mapping structure. Incremental updates are performed based on the state vectors of parameter nodes, operating condition nodes, and result nodes, while recording changes in path weights and constraint edges between nodes. Incremental updates are also performed based on the state vectors of newly added or updated parameter nodes, operating condition nodes, and result nodes, while recording changes in path weights and constraint edges between nodes. The state vectors include process parameter values, operating condition categories, equipment power characteristics, environmental factors, and historical stability indicators. Path weights reflect the reachability strength or energy consumption between nodes, and constraint edges represent process sequence, equipment capacity, and operating condition dependencies. Incremental updates can quickly reflect the latest process status without rebuilding the entire mapping relationship. Incremental correction of the mapping relationship is performed by dynamically adjusting path weights, optimizing constraint edges, and combining historical operating condition stability indicators. This corrects path deviations or constraint conflicts that may occur due to the introduction of new data. For example, for paths between historically stable nodes and newly added nodes, path weights can be adjusted according to the deviation of node status. For constraint edges, if a new node is detected to cause an reachability conflict, the constraint relationship is automatically optimized to ensure the overall rationality of the mapping relationship. Based on the corrected mapping relationship, the reachable state space is updated synchronously, and the newly added or updated nodes and constraint edges are included in the state space constraints. The state space is used to describe the range of states that can be achieved under different combinations of process parameters and operating conditions. During the state space update process, the reachability information of historical nodes is not lost, and the reachable boundary between the newly added nodes and historical nodes is calculated according to the latest mapping relationship. By combining historical stable operating conditions with current detection results, the reachability boundary of the state is recalculated, and the mapping relationship and reachable state space after incremental correction and reachable state update are output. By combining historical stable operating condition data with current detection results, the reachability boundary of each node in the state space is recalculated, including minimum energy path, process feasibility constraints and dynamic operating condition dependencies. Finally, the mapping relationship and reachable state space after incremental correction and reachable state update are output, providing a reliable, continuous and operable data foundation for the next round of process parameter optimization, path analysis and operating condition control.

[0032] Example 2: Please refer to Figure 2 As shown, a coating process parameter self-optimization system integrating model self-learning includes: Data construction module: Real-time collection of process parameters, production conditions, equipment power status and test results data of multiple batches of coating production, alignment and association of data according to process sequence and construction of a uniquely mapped directed data chain; Node backtracking module: Based on the unique mapping relationship, it reversely determines the set of process parameter nodes and operating condition nodes, and calculates the degree of deviation of the detection results by combining historical data weighting and dynamic operating condition association algorithms within the reachable state space, and generates consistency index; Drift assessment module: It compares the consistency index with the current production conditions, equipment power status and historical stable operating condition data, and calculates the operating condition drift probability through a dynamic joint probability model to generate the corresponding operating condition drift identifier. Parameter update module: The consistency index and operating condition drift identifier are used as constraints to input the parameter update model, and the process parameter update magnitude is calculated by combining reinforcement learning. The update is performed only when the consistency index meets the threshold and the drift is not calibrated, and the constrained parameter update result is output. Mapping Correction Module: Writes the updated results of constrained process parameters and new detection results into a unique mapping relationship, performs incremental correction on node mapping and synchronously updates the reachable state space, and generates the incrementally updated mapping relationship and reachable state space.

[0033] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0034] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for self-optimization of coating process parameters based on integrated model self-learning, characterized in that, Includes the following steps: The system collects process parameters, production conditions, equipment power status, and test results data for multiple batches of coating production in real time. The data is aligned and associated according to the process sequence to construct a directed data chain structure. A unique mapping relationship is formed by connecting the directed edges that meet the process reachability constraints and the minimum energy path calculation. Based on the unique mapping relationship, the corresponding process parameter nodes and operating condition node sets are determined in reverse. Within the reachable state space defined by the node state vector and directed edge constraints, the deviation of the current detection result from the reachable state space is calculated by combining historical data distribution weighting and dynamic operating condition association algorithms, and a consistency index characterizing the credibility of the detection result is generated. The consistency index is jointly compared with the data chain distribution corresponding to the current production conditions, real-time equipment power status and historical stable operating conditions. The drift probability of the current operating condition is calculated by combining the dynamic joint probability distribution model, and the corresponding operating condition drift identifier is output. The consistency index and operating condition drift identifier are introduced as constraints into the parameter update model. The process parameter update magnitude is calculated by combining reinforcement learning algorithm. The process parameters are updated only when the consistency index meets the threshold and is not marked as operating condition drift, and the constrained parameter update results are output. The updated results of the constrained process parameters and the corresponding new round of test results are written into the unique mapping relationship. The mapping relationship between parameter nodes, operating condition nodes and result nodes is incrementally corrected, and the reachable state space is updated synchronously to generate the incrementally updated mapping relationship and reachable state space.

2. The self-optimization method for coating process parameters based on fusion model self-learning according to claim 1, characterized in that, The process of forming a unique mapping relationship is as follows: The process parameters, production conditions, equipment power status and test results data collected during multiple batches of coating production are indexed in time sequence according to the process order, and a multi-dimensional state vector is generated for each node. Construct an reachability constraint matrix between nodes, and encode process feasibility rules, equipment capacity constraints, and dependencies between operating conditions as constraint edges; Based on the reachability constraint matrix, the minimum energy path algorithm is used to generate the optimal path set from the initial working condition node to each detection result node; The path set is recursively filtered and redundancy eliminated, removing paths that deviate significantly from the historical stable operating condition distribution. The path confidence is calculated by combining the node state vectors, and a unique mapping relationship is output, consisting of process parameter nodes, operating condition nodes, detection result nodes and constraint directed edges.

3. The self-optimization method for coating process parameters based on fusion model self-learning according to claim 2, characterized in that, The process of reversely determining the corresponding set of process parameter nodes and operating condition nodes is as follows: Based on the unique mapping relationship, reverse path backtracking is performed on each detection result node; During the backtracking process, the multi-dimensional state vector of each node is analyzed to extract parameter value range, equipment power characteristics, operating condition category and historical stability index. Simultaneously, a dynamic reachability matrix between nodes is constructed, encoding process reachability rules, equipment capability constraints, and temporal dependencies between nodes into matrix edge weights. On the backtracking path, the path confidence is calculated by combining the node state deviation, historical distribution consistency and dynamic correlation of operating conditions, and outputting a complete set of process parameter nodes and operating condition nodes.

4. The self-optimization method for coating process parameters based on fusion model self-learning according to claim 3, characterized in that, The process of generating a consistency index to characterize the reliability of the detection results is as follows: In the set of process parameter nodes and operating condition nodes, a state vector is generated for each detection result node to represent parameter values, operating condition categories, equipment power characteristics, environmental factors, and historical stability. Embed the state vector into the historical process data distribution space and calculate the deviation in the reachable state space; Based on the dynamic working condition association algorithm, the deviation of the current node is subjected to time-series weighting with the state vectors of the previous and subsequent time nodes to evaluate the abnormal diffusion trend and potential drift path, and form a deviation time series. The deviation time series, dynamic correlation of working conditions and historical stability indicators are weighted and integrated to construct a fusion vector, and finally a consistency index representing the credibility of the test results is generated.

5. The self-optimization method for coating process parameters based on fusion model self-learning according to claim 4, characterized in that, The process of jointly comparing consistency indicators with the data chain distribution corresponding to current production conditions, real-time equipment power status, and historical stable operating conditions is as follows: The consistency index of the current detection results is matched with the data chain distribution of the corresponding production condition nodes, real-time equipment power status nodes and historical stable operating condition nodes to form a multi-dimensional comparison matrix; Based on the multidimensional data representation, the data distribution of each dimension is compared with the historical stable distribution; The matching degree is calculated by a joint probability model, and a joint matching vector reflecting the overall consistency of the operating conditions is generated. Based on the matching vector, according to the preset weights and operating condition sensitivity, the operating condition drift sensitivity score is calculated, and the joint comparison result of drift determination is output.

6. The self-optimization method for coating process parameters based on fusion model self-learning according to claim 5, characterized in that, The process of outputting the corresponding operating condition drift indicator is as follows: Based on the joint comparison results and joint matching vectors, a multidimensional state representation of the current working condition is constructed and input into the dynamic joint probability distribution model; The model is used to jointly model the distribution of historical data, temporal correlation and dynamic dependence between operating conditions, and generate the drift probability distribution of the current operating condition in the reachable state space. Extract the mean and standard deviation of drift probabilities for each dimension from historical stable operating data to establish historical drift thresholds; The drift probability distribution generated by the model is compared with the historical drift threshold, and the operating condition drift index is calculated by combining short-term abnormal fluctuations and long-term trend changes. Based on the operating condition drift index and dynamic weighting rules, determine whether the current operating condition exceeds the historical acceptable drift range, and output the corresponding operating condition drift flag.

7. The self-optimization method for coating process parameters based on fusion model self-learning according to claim 6, characterized in that, The process of introducing consistency indicators and operating condition drift indicators as constraints into the parameter update model is as follows: For each process parameter node, extract the current consistency index and the corresponding operating condition drift identifier, and combine them with the historical operating condition stability characteristics to construct the parameter update input state vector; The state vector input parameters are used to update the model. Constraint logic is established within the model, allowing the corresponding process parameters to be updated only when the consistency index reaches the set threshold and the operating condition is not calibrated as drift. Based on the constraint logic, and combined with the state vectors of the preceding and following nodes and the dynamic working condition dependencies, the parameter update magnitude is weighted and adjusted, and the parameter update input vector after constraint and weighting processing is output.

8. The self-optimization method for coating process parameters based on fusion model self-learning according to claim 7, characterized in that, The process of outputting the constrained parameter update results is as follows: The parameter update input vector is used as the reinforcement learning state vector, the process parameter adjustment range is used as the action space, and a comprehensive reward function is defined, which includes the consistency index of detection results, the operating condition drift index, and the historical process stability index. During the reinforcement learning training process, the probability of action selection is dynamically adjusted according to different process parameter dimensions and operating conditions using a priority gating strategy. By incorporating an adaptive learning rate mechanism, the magnitude of parameter updates is constrained; The state vectors of the preceding and following nodes and the dynamic dependencies of the working conditions are incorporated into the training process to suppress local abnormal fluctuations or sudden drift behaviors with weights. Process parameter updates are performed only when the consistency index reaches a preset threshold and the current operating condition is not labeled as drift, outputting constrained, reinforcement learning-optimized process parameter update results.

9. The self-optimization method for coating process parameters based on fusion model self-learning according to claim 8, characterized in that, The process of generating the incrementally updated mapping and reachable state space is as follows: Write the updated process parameters and the corresponding new test results into a unique mapping relationship; Incremental updates are performed based on the state vectors of parameter nodes, working condition nodes, and result nodes, while simultaneously recording the changes in path weights and constraint edges between nodes. Incremental correction of the mapping relationship is carried out by dynamically adjusting the path weight, optimizing the constraint edge, and combining historical working condition stability indicators. Based on the corrected mapping relationship, the reachable state space is updated synchronously, and the newly added or updated nodes and constraint edges are incorporated into the state space constraints. By combining historical stable operating conditions with current detection results, the state reachability boundary is recalculated, and the mapping relationship and reachability state space after incremental correction and reachability state update are output.

10. A coating process parameter self-optimization system based on fusion model self-learning, applied to the coating process parameter self-optimization method based on fusion model self-learning as described in any one of claims 1-9, characterized in that, include: Data construction module: Real-time collection of process parameters, production conditions, equipment power status and test results data of multiple batches of coating production, alignment and association of data according to process sequence and construction of a uniquely mapped directed data chain; Node backtracking module: Based on the unique mapping relationship, it reversely determines the set of process parameter nodes and operating condition nodes, and calculates the degree of deviation of the detection results by combining historical data weighting and dynamic operating condition association algorithms within the reachable state space, and generates consistency index; Drift assessment module: It compares the consistency index with the current production conditions, equipment power status and historical stable operating condition data, and calculates the operating condition drift probability through a dynamic joint probability model to generate the corresponding operating condition drift identifier. Parameter update module: The consistency index and operating condition drift identifier are used as constraints to input the parameter update model, and the process parameter update magnitude is calculated by combining reinforcement learning. The update is performed only when the consistency index meets the threshold and the drift is not calibrated, and the constrained parameter update result is output. Mapping Correction Module: Writes the updated results of constrained process parameters and new detection results into a unique mapping relationship, performs incremental correction on node mapping and synchronously updates the reachable state space, and generates the incrementally updated mapping relationship and reachable state space.

Images