A method and system for intelligent optimization of parameters in core coating equipment
By constructing a parameter coupling structure and a dynamic update model, the problem of unutilized parameter coupling relationships in the core coating process was solved, enabling continuous optimization of coating quality and energy consumption, and improving the adaptability and accuracy of the optimization algorithm.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- 宇鸿敏芯(山东)电子科技有限公司
- Filing Date
- 2026-02-06
- Publication Date
- 2026-05-26
AI Technical Summary
Existing intelligent optimization algorithms fail to effectively utilize the coupling relationship between parameters in core coating processes, resulting in infeasible optimization results or difficulty in finding Pareto optimal solutions with conflicting multi-objectives. Furthermore, they cannot adapt to the drift of process rules caused by changes in equipment status and environment.
By acquiring and cleaning historical process data, a predictive model is constructed, the parameter coupling structure is analyzed, and a structure-aware evolutionary algorithm is used for optimization. The model is dynamically updated during process drift to maintain the integrity of parameter coupling relationships and optimization capabilities.
It improves the efficiency and accuracy of parameter optimization, enabling it to adapt to long-term process changes and achieve continuous optimal synergy between coating quality and energy consumption.
Smart Images
Figure CN121683537B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing. More specifically, this invention relates to a method and system for intelligent optimization of parameters in core coating equipment. Background Technology
[0002] In complex processes such as core coating, optimizing process parameters is crucial for ensuring coating quality and controlling production costs. Currently, intelligent optimization algorithms, such as genetic algorithms, have been applied to parameter optimization in these processes. These algorithms, by simulating natural evolution, can search for optimal solutions in complex multi-parameter spaces, providing an effective alternative to traditional trial-and-error methods that rely on engineer experience. However, precisely because core coating processes are highly sensitive to parameter synergy, objective multidimensionality, and operating condition stability, existing general-purpose optimization algorithms exhibit significant limitations when directly applied.
[0003] First, traditional algorithms treat process parameters as independent variables, completely ignoring the strong physical and process coupling relationships that objectively exist between parameters. This often results in candidate solutions that are practically infeasible due to violations of inherent consistency. Second, when faced with the typical multi-objective conflict of improving bonding strength and reducing energy consumption, they lack an understanding of the complex mapping relationship between parameters and objectives, leading to a blind search process and difficulty in finding a truly process-feasible Pareto optimal solution. Finally, existing methods are usually static optimization models, which cannot adapt to the drift of process laws caused by changes in equipment status, environment, and other factors, making it difficult to maintain optimization results in the long term. Summary of the Invention
[0004] To address the aforementioned technical problems, the present invention provides solutions in the following aspects.
[0005] In the first aspect, a method for intelligent optimization of parameters in a core coating equipment includes:
[0006] Historical process data is acquired, outlier removal and constraint filtering are performed to obtain cleaned data;
[0007] An initial predictive model was constructed using the cleaned data to simultaneously predict the coating adhesion strength and process energy consumption.
[0008] The initial prediction model is used to select historical samples that have accurate predictions and meet the actual results, and a dynamically maintained reference sample library is constructed.
[0009] Based on the initial prediction model, the global correlation between process parameters is analyzed, and a parameter coupling structure is generated through adaptive clustering. The parameter coupling structure defines multiple strongly correlated parameter groups.
[0010] A structure-aware evolutionary algorithm is used for parameter optimization search. When performing parameter recombination, the integrity of each strongly correlated parameter group is maintained according to the parameter coupling structure. When performing parameter perturbation, a perturbation strategy that decays with the iteration process is adopted.
[0011] When the real-time prediction performance of the model degrades and exceeds the trigger threshold for dynamic adjustment, the adaptive model update process is initiated: the reference sample library is fused with recent process data to establish a distribution model representing the current operating conditions, and based on this, highly correlated samples are selected from historical data and assigned differentiated weights to retrain and update the prediction model.
[0012] Preferably, the data acquisition after cleaning includes: calculating the standardized offset value of the output parameters based on the principle of normal distribution, and removing data points whose standardized offset values exceed a preset confidence interval; and removing process records that do not conform to the hard constraint range of parameters defined by a preset process knowledge base.
[0013] Preferably, the initial prediction model is constructed using a gradient boosting decision tree framework, and the model hyperparameters are adjusted through multi-fold cross-validation until the model's prediction determination coefficient on the validation set reaches a preset standard.
[0014] Preferably, the construction of the reference sample library specifically involves: calculating the prediction bias and the actual binding strength bias of historical samples, and storing only samples that simultaneously meet the condition that both biases are below their respective preset thresholds into a fixed-capacity first-in-first-out queue in chronological order.
[0015] Preferably, the step of generating parameter coupling structure through adaptive clustering includes: calculating the global correlation matrix between parameters; evaluating the index of intra-group aggregation tightness under different numbers of groupings within a preset range of candidate groupings; determining the number of groupings that maximizes the index as the optimal number of groupings, and clustering the parameters accordingly.
[0016] Preferably, maintaining the integrity of each strongly correlated parameter group in the parameter coupling structure means that when a randomly selected crossover point is located inside a strongly correlated parameter group, the crossover point is forcibly adjusted to the start or end boundary of the parameter group before gene exchange.
[0017] Preferably, the perturbation strategy that decays with the iteration process means that the perturbation amplitude of the parameter in each generation starts from an initial value and decays according to a nonlinear function as the number of generations increases.
[0018] Preferably, the dynamically adjusted trigger threshold is updated by weighted averaging based on the prediction error within historical periods, and a minimum threshold is set to prevent excessively frequent updates due to model overfitting.
[0019] Preferably, the allocation of differentiated weights means that the weights assigned to each sample in the training dataset are positively correlated with the probability density of that sample calculated based on the current working condition distribution model, and samples from the reference sample library are given higher base weight coefficients.
[0020] In a second aspect, a system for intelligent optimization of parameters of a core coating equipment includes: a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the system implements any of the methods for intelligent optimization of parameters of a core coating equipment as described above.
[0021] The beneficial effects of this invention are:
[0022] This invention automatically identifies and constructs parameter coupling structures by analyzing the global correlation matrix between process parameters, thereby guiding the optimization search. This transforms the algorithm from blindly adjusting independent parameters to collaboratively optimizing strongly correlated parameter groups, thus improving the optimization efficiency of parameter coupling. At the same time, through a model update mechanism with dynamic triggering and operating condition weighting, it ensures that the optimization capability can adapt to process drift in the long term and maintain stable optimization accuracy, ultimately reducing process energy consumption and achieving continuous synergistic optimization of quality and energy efficiency. Attached Figure Description
[0023] Figure 1 This is a flowchart of steps S1-S6 in an intelligent optimization method for core coating equipment parameters according to an embodiment of the present invention.
[0024] Figure 2 This is a structural block diagram of an intelligent optimization system for core coating equipment parameters according to an embodiment of the present invention. Detailed Implementation
[0025] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.
[0026] Reference Figure 1 A method for intelligent optimization of parameters in core coating equipment includes steps S1-S6, as detailed below:
[0027] S1: Obtain historical process data and perform dual cleaning of statistical outlier removal and constraint filtering.
[0028] There are risks in directly using the original historical process records to build a predictive model, because the data may contain outliers caused by instantaneous sensor failures, manual input errors, or extreme operating conditions. These outliers can severely distort the statistical regularity of the model and lead to the learning of incorrect parameter relationships. Therefore, statistical outlier removal is performed first.
[0029] For core output indicators of the process, such as bonding strength With energy consumption Their normal production data usually follow or approximately follow a normal distribution, so quantitative methods based on distribution assumptions can be used for identification.
[0030] With bonding strength For example, let's say the data is extracted from the database. The initial dataset consists of 10 historical records. , These values are derived from experience and can be adjusted by the implementer according to the specific implementation scenario. The output value for each sample... Calculate the mean of each indicator across all samples. and standard deviation Then, its standardized offset value is obtained. Set an experience threshold This value can be adjusted by the implementer according to the specific implementation scenario; it corresponds to approximately [value missing] under a normal distribution. The confidence interval is used to remove all values that meet the criteria. The data points are filtered out to eliminate statistically improbable outliers.
[0031] Since some combinations of process parameters may not appear to be mathematically or statistically abnormal, but violate the physical limits of the equipment or the safety boundaries of the materials, and are not feasible in actual production, using them for training would lead to invalid or even dangerous configurations recommended by the model. Therefore, constraint filtering needs to be further implemented.
[0032] Obtain a pre-set process knowledge base, which defines the hard allowable range of each adjustable process parameter, such as the upper limit of RF power and the temperature safety threshold. For the dataset after statistical cleaning in the previous step, check all process parameter values one by one. If any parameter value exceeds its corresponding hard constraint range, the record is determined to be infeasible and the data is removed.
[0033] For example, a temperature safety threshold range is set. If the value of the adjustable process parameter is outside the temperature safety threshold range, the data entry can be removed. middle, The empirical values of 100°C and 300°C are taken respectively, and can be adjusted by the implementer according to the implementation scenario. Other process parameters can be derived from this.
[0034] S2: Use the post-cleaning data to build an initial predictive model that simultaneously predicts the coating adhesion strength and process energy consumption.
[0035] Other core process output metrics can be derived sequentially from the steps in S1, thus obtaining a cleaned dataset of all core process output metrics. Cleaned dataset If used directly for training, the model may fall into overfitting or underfitting due to improper hyperparameter settings, failing to obtain stable and reliable predictive capabilities. This would cause subsequent optimization steps to lose their accurate foundation. Therefore, it is necessary to construct a structured predictive model. And its parameters were optimized.
[0036] There is a complex nonlinear relationship between process parameters and output indicators. Tree-based algorithms can effectively fit this relationship. However, their performance is highly dependent on the combination of hyperparameters, so they must be evaluated and selected through an independent validation set.
[0037] In building the model At that time, it is clearly stated that its input is a feature vector of process parameters. The output consists of two target variables that need to be predicted simultaneously: the binding strength. With energy consumption This constitutes a multi-output regression problem.
[0038] Gradient Boosting Decision Tree (GBDT) framework was chosen because it can automatically handle feature interactions and has high prediction accuracy. The core structure of GBDT consists of a sequence of regression trees, and its initial hyperparameters can be preset with a set of empirical values, such as the initial learning rate. Maximum tree depth The number of trees The initial hyperparameters can be adjusted by the implementer according to the specific implementation scenario.
[0039] To ensure the model's prediction accuracy meets the preset standard, model parameters must be fine-tuned using a validation set, and a K-fold cross-validation strategy must be employed to fully utilize limited data and reliably evaluate performance. The specific implementation process is as follows: Randomly divided into A mutually exclusive subset, usually taken as an empirical value. The implementer can adjust the set according to the specific implementation scenario, and select one subset as the verification set in turn. ,the remaining A subset and used as the training set , conducted Each training and validation cycle is independent. After each validation, the computational model is... Coefficient of determination of the prediction results This metric measures the goodness of fit between predicted and actual values; the closer the predicted value is to the actual value, the better the result. The higher the precision, the better.
[0040] Finally, by iterating through the preset combinations of hyperparameters, the one that makes the parameter selection optimal is determined. Average of verification times The set of parameters with the highest value is used as The optimal hyperparameters are obtained, and the calculation of the coefficient of determination is a well-known technique, which will not be elaborated here.
[0041] The preset accuracy standard can be set to average. The value of 0.8 is an empirical value and can be adjusted by the implementer according to the specific implementation scenario. After optimization, use all... The data and the determined optimal hyperparameters are used to retrain the model, resulting in the initial prediction model that is finally put into use. This ensures that it has the basic ability to accurately predict process results.
[0042] S3: Use the initial prediction model to filter historical samples that have accurate predictions and meet the standards to build a dynamically maintained reference sample library.
[0043] The initial prediction model obtained after cleaning and training Although it has basic predictive capabilities, it is not equally reliable to predict all historical data. If all historical data is used indiscriminately for subsequent model updates, low-quality prediction data will pollute the training set, leading to a degradation in model performance.
[0044] This will lead to the construction of a high-quality, high-confidence reference sample library. ;use My own historical data Backtesting was conducted to select the best samples that showed both accurate model predictions and good actual results, which were then used as reliable data for subsequent model optimization and working condition matching.
[0045] If only the prediction bias is low, it may include anomalous samples that are actually poor but were predicted correctly by the model by chance. Such samples have no optimization value. On the other hand, if only the actual quality is good, it may include samples that the model cannot understand the underlying rules. Forcing learning will increase the uncertainty of the model.
[0046] Therefore, two conditions must be met simultaneously: first, the model must have high confidence in its predictions; and second, the model itself must be a successful technological practice. Only in this way can we ensure that the samples included in the database are both learnable and imitable.
[0047] In the specific implementation of the screening process, the prediction bias for each historical sample is first calculated. The normalized root mean square error (RMSE) is used as the bias metric. .in, It is the root mean square error between the model-predicted binding strength value and the actual value; and These are the maximum and minimum values of the historical binding strength, used for normalization to eliminate dimensions.
[0048] Set a preset deviation threshold. Take experience points It can be adjusted by the implementer according to the specific implementation scenario, requiring... At the same time, the actual bonding strength is calculated. With target value relative deviation And set a quality deviation threshold. Take experience points It can be adjusted by the implementer according to the specific implementation scenario, requiring... .
[0049] The samples that pass the above dual screening will be filled into a fixed capacity according to the timestamp of their process completion. First-in, first-out queue These values are derived from experience and can be adjusted by the implementer according to the specific implementation scenario, thus forming a reference sample library. .
[0050] When a new sample is added to the database, if the database is full, the oldest record is automatically discarded to ensure... We continuously and dynamically maintain a batch of high-quality process data examples with reliable model predictions.
[0051] S4: Analyze the global correlation between process parameters based on the initial prediction model and generate parameter coupling structures through adaptive clustering.
[0052] When using evolutionary algorithms to optimize process parameters, treating all parameters as independent variables and randomly rearranging and perturbing them ignores the objectively existing physical or process coupling relationships between the parameters. For example, temperature and pressure may need to be adjusted synergistically to achieve a specific effect. Blindly severing these strong correlations can lead to inconsistent candidate process configurations, or even violate process principles, thereby significantly reducing the efficiency of the optimization search and the practical feasibility of the output solution. Therefore, it is essential to identify and construct the coupling structure between parameters in advance.
[0053] In the initial prediction model In particular, tree-based models encode the complex interactions between features (i.e., process parameters) through their internal decision paths. To calculate the nonlinear interaction strength between any two process parameters, a global correlation matrix between parameters can be calculated by analyzing the marginal impact of each parameter on the SHAP value of another parameter in the model's prediction results. This matrix can effectively characterize the tightness of coupling between parameter pairs.
[0054] Extract from the cleaned data One sample, These values are based on experience and can be adjusted by the implementer according to the specific implementation scenario for each pair of process parameters. Calculate the SHAP interaction value, and use the mean of the absolute values of the SHAP interaction values across all samples as an element in the global association matrix. The global correlation matrix is obtained. ,Right now Global association matrix It is A symmetric matrix whose diagonal elements are usually set to numerical values The larger the value, the higher the parameter. and The stronger the interaction when influencing model predictions, the higher the coupling.
[0055] Obtain the correlation matrix Next, it needs to be transformed into a clear grouping structure, that is, to determine which parameters should be assigned to the same group; an adaptive grouping algorithm is then used to determine the optimal number of groups. Agglomerative hierarchical clustering algorithm was used to analyze the association matrix. Perform clustering.
[0056] Number of groups The search scope is set to start from arrive ,in Set an upper limit for the total number of parameters. Based on empirical balance, 2 and The values are based on experience and can be adjusted by the implementer according to the specific implementation scenario. This can prevent the optimization from becoming fragmented due to overly fine grouping, or the internal coupling from being too coarse. Among them, the cohesive hierarchical clustering algorithm can use the Ward method, which can be adjusted by the implementer according to their specific implementation scenario.
[0057] In the preset Within the value range, there needs to be an indicator to evaluate each The clustering quality corresponding to the value. To this end, the clustering quality for each candidate value is calculated. Average cohesive density coefficient at value Average polymer density coefficient This is the ratio of the average intra-cluster association strength to the average inter-cluster association strength. Specifically, for each cluster in the clustering results... Calculate the average of all its internal parameters with respect to the associated values. Simultaneously calculate the average of the correlation values of all parameters between this cluster and other clusters. . That is, all clusters mean and The quotient of the mean. Optimal number of groups. Is to make The one that gets the maximum value value.
[0058] Based on the determined optimal number of groups Perform hierarchical clustering segmentation to obtain the final parameter group labels. Each group... ( It contains a set of process parameter indexes, and the parameters within the set have relatively strong correlations, ensuring the stability of the inherent process coupling relationship during the optimization process.
[0059] S5: A structure-aware evolutionary algorithm is used to optimize parameters, maintain the integrity of recombination, and employ a time-varying decay perturbation strategy.
[0060] Genetic algorithms typically treat each process parameter as an independent variable during recombination and mutation, which has significant drawbacks when solving for process parameters with strong coupling relationships. Random recombination may break up closely related parameter combinations, producing inconsistent or even contradictory offspring solution sets, leading to searches in invalid regions, slow convergence, and difficulty in finding the optimal solution.
[0061] To ensure that the structure is coupled with the constructed parameters during evolutionary operations. In collaborative work, during parameter recombination (or crossover) operations, when selecting a recombination point between two parent individuals, it is necessary to determine whether that point will split a strongly correlated parameter group. If so, the recombination point position needs to be adjusted to maintain the integrity of the group, thereby ensuring that the generated child individuals inherit meaningful and coordinated parameter modules. This is called structure-aware recombination.
[0062] When implementing structure-aware reorganization, the first step is to start with the current population. Two parent individuals are randomly selected. and Next, a recombination point is randomly pre-selected on the parameter index sequence. Then, query the parameter coupling structure. ,judge Does it fall into a certain group? The interior (i.e., not the boundary of the group).
[0063] If it is internal, then further calculate the parameter pairs within that group in the current global association matrix. Average correlation strength .
[0064] like Greater than the preset strong correlation threshold Strong correlation threshold If the average of all associated values is taken, the group is determined to be a strongly associated group and its integrity should be preserved. At this point, the regrouping points... Adjusted to this strongly associated group The starting or ending index position is determined. Finally, based on the adjusted recombination point, the parameter values of corresponding segments of the two parent individuals are swapped to generate child individuals, which serve as the child solution set.
[0065] In parameter perturbation (or mutation) operations, using random offsets with fixed amplitudes can easily lead to oscillations around the optimal solution in the later stages of optimization, making precise positioning difficult. Furthermore, different parameters have varying importance, and applying equal perturbations to them is unreasonable. To dynamically adjust the perturbation intensity as the search progresses and to differentiate the time-varying decay of parameter importance, a random offset strategy is employed.
[0066] The random offset strategy is implemented by offsetting according to the random offset formula, which is: .
[0067] in, Indicates the parameter The applied perturbation value; This is the importance weight of the parameter, which can be obtained from the initial prediction model. The feature importance is obtained from the statistical parameters, where the feature importance of the GBDT model is a well-known technique. The ratio of the total number of times a node is selected as a split node across all decision trees in the entire GBDT model to the total number of split nodes is used as the initial prediction model. The feature importance is given by the parameter. Importance weights; It is a parameter The domain range is used for normalization; The mean is Standard deviation is The normally distributed random numbers ensure that the perturbation revolves around the current value. The values are based on experience and can be adjusted by the implementer according to the specific implementation scenario; It is the core time-varying decay factor, and its calculation formula is: ,here, This is the initial disturbance amplitude, taken as an empirical value. The implementer can adjust it according to the specific implementation scenario. It is the current evolutionary generation. It is the preset maximum number of generations, taken from empirical values. It can be adjusted by the implementer according to the specific implementation scenario. It is the decay exponent, taken as an empirical value. This can be adjusted by the implementer according to the specific implementation scenario. With... Increase, Nonlinear decay enables the algorithm to explore a wide range in the early stages and converge finely in the later stages.
[0068] Finally, the evolutionary algorithm iteratively executes the aforementioned structure-aware recombination and time-varying decay perturbation to generate a new population of candidate process configurations. Each individual is then processed by the updated prediction model. By evaluating the predicted binding strength and energy consumption, and through multiple generations of selection, recombination, and perturbation, the genetic algorithm can efficiently search for the optimal process parameter configuration with comprehensive performance while ensuring the constraints of process coupling relationships. This outputs a recommended solution.
[0069] S6: When the prediction performance degrades beyond the dynamic threshold, the data is fused to establish the current operating condition distribution model and the weighted data is filtered for retraining and model update.
[0070] Slow changes in raw materials, equipment conditions, or environmental factors can cause process drift, affecting the initial prediction model. The accuracy of predictions gradually decreased.
[0071] If a fixed error threshold is used to determine whether the model needs to be updated, either the threshold will be too wide to respond to the actual drift in a timely manner, or the threshold will be too strict to cause unnecessary frequent updates.
[0072] Introduce a dynamic trigger threshold When real-time production data is processed by the model After prediction, the actual prediction deviation is calculated. Exceeding the current threshold When this happens, a model update is triggered. Deviation. The threshold can be calculated using root mean square error or mean absolute percentage error. It is not fixed; it will adaptively adjust itself based on the model's performance before and after each successful update, forming a closed-loop feedback that allows it to track the drift trend of model performance.
[0073] Dynamic threshold The specific update rules are as follows: Let the threshold before the update be... The updated model prediction error, evaluated on the new training set, is... Then the new threshold From the formula The calculation yielded the following result. It is the memory decay factor, and an empirical value can be taken. It can be adjusted by the implementer according to the specific implementation scenario to smooth out threshold fluctuations, so that it reflects the recent error level without being overly affected by a single update. It is a preset absolute lower limit, and the empirical value is... It can be adjusted by the implementer according to the specific implementation scenario to prevent the threshold from being excessively lowered under abnormally good data, thereby avoiding a state of oversensitivity and frequent updates later.
[0074] Once an update is triggered, a statistical distribution model that can characterize the current production conditions needs to be built. Training directly using all historical data would introduce a large amount of outdated or irrelevant information, interfering with the model's learning of current patterns. Therefore, it is necessary to integrate data from a reference sample library. High-quality samples and recently generated process data This is used to fit the data distribution under the current operating conditions. recent The hourly production data is the closest to the current actual production status, while the 24-hour data is based on experience and can be adjusted by the implementer according to the specific implementation scenario.
[0075] In specific implementation, and Merge to form the current working condition dataset Assuming the process parameters follow a multivariate normal distribution, then by calculating... The mean vector of each process parameter for all samples Covariance Matrix This allows the establishment of a current operating condition distribution model. Covariance matrix Capturing the interrelationships between various process parameters under the current operating conditions is key to measuring the similarity between data points and the current distribution.
[0076] Obtain the current working condition distribution model Then, from the massive historical dataset In this process, data that is highly correlated with the current working conditions is selected to expand the training set.
[0077] The degree of correlation can be measured by calculating each historical sample. Belongs to the distribution probability density To achieve this, a correlation threshold is set. Only retain The samples constitute a highly correlated historical subset. Among them, the correlation threshold To select the former The density quantile values of the samples are used as the correlation threshold.
[0078] In order to differentiate between samples of different sources and qualities during model update training, it is necessary to provide a final training set with... Each sample in the dataset is assigned a priority weight. The specific calculation process for priority weights is as follows: (The text abruptly ends here, likely due to an incomplete sentence or a formatting error.) High-quality samples should be assigned higher base weights; samples more similar to the current working conditions have higher learning value. The priority weight calculation formula can be: .in, It is a source factor. Sample taking , Sample taking This value is based on experience and can be adjusted by the implementer according to the specific implementation scenario. The cumulative number of model updates serves as a bias factor. As the model is updated multiple times, the exponential term accelerates the reduction of the weight of old historical samples, making the model more focused on the latest and most relevant data patterns. The dataset, weighted by priority weights, is then used to train a new prediction model, completing this update.
[0079] This invention also provides an intelligent optimization system for parameters of core coating equipment. For example... Figure 2 As shown, the system includes a processor and a memory, the memory storing computer program instructions, which, when executed by the processor, implement the intelligent optimization method for core coating equipment parameters according to the first aspect of the present invention.
[0080] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.
[0081] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.
Claims
1. A method for intelligent optimization of parameters in core coating equipment, characterized in that, include: Historical process data is acquired, outlier removal and constraint filtering are performed to obtain cleaned data; An initial predictive model was constructed using the cleaned data to simultaneously predict the coating adhesion strength and process energy consumption. The initial prediction model is used to select historical samples that have accurate predictions and meet the actual results, and a dynamically maintained reference sample library is constructed. Based on the initial prediction model, the global correlation between process parameters is analyzed, and a parameter coupling structure is generated through adaptive clustering. The parameter coupling structure defines multiple strongly correlated parameter groups. A structure-aware evolutionary algorithm is used for parameter optimization search. When performing parameter recombination, the integrity of each strongly correlated parameter group is maintained according to the parameter coupling structure. When performing parameter perturbation, a perturbation strategy that decays with the iteration process is adopted; When the real-time prediction performance of the model deteriorates and exceeds the trigger threshold for dynamic adjustment, the adaptive model update process is initiated: the reference sample library is fused with recent process data to establish a distribution model that characterizes the current working conditions, and high-correlation samples are selected from historical data and assigned differentiated weights to retrain and update the prediction model. The step of generating a parameter coupling structure through adaptive clustering includes: calculating the global correlation matrix between parameters; Within a preset range of candidate group numbers, an index is evaluated to assess the aggregation density within groups under different group numbers; the group number that maximizes the index is determined as the optimal group number, and the parameters are clustered accordingly. The parameter coupling structure maintains the integrity of each strongly correlated parameter group by: when a randomly selected crossover point is located inside a strongly correlated parameter group, the crossover point is forcibly adjusted to the start or end boundary of the parameter group before gene exchange is performed. The perturbation strategy that decays with the iteration process refers to the fact that the perturbation amplitude of the parameter in each generation starts from an initial value and decays according to a nonlinear function as the number of generations increases. The differentiated weight allocation refers to assigning a weight to each sample in the training dataset that is positively correlated with the probability density of that sample calculated based on the current working condition distribution model, and assigning a higher base weight coefficient to samples from the reference sample library.
2. The intelligent optimization method for parameters of a core coating equipment according to claim 1, characterized in that, The acquisition of the cleaned data includes: Based on the principle of normal distribution, the standardized offset value of the output parameters is calculated, and data points whose standardized offset value exceeds the preset confidence interval are removed; Based on the hard constraint range of parameters defined by the preset process knowledge base, process records that do not conform to the hard constraint range are removed.
3. The intelligent optimization method for parameters of a core coating equipment according to claim 1, characterized in that, The initial prediction model is constructed using a gradient boosting decision tree framework, and the model hyperparameters are adjusted through multi-fold cross-validation until the model’s prediction determination coefficient on the validation set reaches the preset standard.
4. The intelligent optimization method for parameters of a core coating equipment according to claim 1, characterized in that, The construction of the reference sample library involves calculating the prediction bias and the actual binding strength bias of historical samples, and storing only samples that simultaneously meet the condition that both biases are below their respective preset thresholds into a fixed-capacity first-in-first-out queue in chronological order.
5. The intelligent optimization method for parameters of a core coating equipment according to claim 1, characterized in that, The dynamically adjusted trigger threshold is updated by weighted averaging the prediction errors within historical periods, and a minimum threshold limit is set to prevent excessively frequent updates due to model overfitting.
6. A system for intelligent parameter optimization of core coating equipment, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement the intelligent optimization method for core coating equipment parameters according to any one of claims 1-5.