A stamping process parameter optimization method and system based on rule-model double-layer decision

Abstract

The application discloses a stamping process parameter optimization method and system based on rule-model double-layer decision, and belongs to the technical field of manufacturing process control. The method collects stamping process historical data to construct a training sample set, and obtains an optimized feature subset through Pearson correlation coefficient, mutual information and variance inflation factor screening; a stacking fusion model of a random forest and a gradient boosting tree is trained as a parameter mapping model; a parameter self-selection rule library is constructed; in the recommendation stage, the rule library is preferentially queried, the mapping model is called for prediction when a hit is not made, and interpretability analysis is carried out based on SHAP values; finally, the model and the rule library are updated in a closed loop by using trial model result feedback. The application considers the certainty of known working condition recommendation and the generalization ability of unknown working condition, and can realize intelligent recommendation and closed-loop optimization of process parameters.

Description

Technical Field

[0001] This invention belongs to the field of manufacturing process control technology, and in particular relates to a method and system for optimizing stamping process parameters based on rule-model bi-level decision-making. Background Technology

[0002] In automotive sheet metal stamping production, the setting of stamping process parameters directly determines the forming quality and production efficiency of the sheet metal. Stamping process parameters mainly include blank holder force, die clearance, forming speed, and lubrication conditions. When introducing new workpieces or materials, a reasonable combination of process parameters needs to be determined to ensure that the formed parts are free from defects such as cracking and wrinkling, and that the dimensional accuracy meets design requirements.

[0003] Currently, the determination of stamping process parameters relies heavily on the experience and trial-and-error of process engineers. Specifically, process engineers typically set key process parameters initially based on historical experience, and then repeatedly test molds and make manual adjustments on-site until the forming quality meets the requirements. However, this traditional method of determining process parameters has the following significant drawbacks in actual production:

[0004] First, the trial molding and debugging cycle is long, and the production line resource occupancy rate is high. Under the traditional trial molding method, which relies mainly on manual experience, the exploration and verification of parameters often requires multiple rounds of repeated adjustments. This inefficient trial and error process not only significantly extends the introduction cycle of new products, but also occupies a large amount of valuable press and mold resources, increasing production costs.

[0005] Second, the setting of process parameters relies excessively on manual experience, resulting in low standardization. The stamping process exhibits highly non-linear characteristics, leading to individual differences among process engineers in parameter selection, adjustment strategies, and experience-based judgment. This results in inconsistent process execution for the same workpiece under different operators, making it difficult to guarantee process standardization and reproducibility.

[0006] Third, historical data utilization is low, and there is a lack of closed-loop self-learning and continuous optimization mechanisms. Although existing stamping workshops have accumulated a large amount of historical trial mold data and production records, they often lack systematic sorting and in-depth analysis of this data. When introducing new workpieces, it still relies heavily on manual on-site exploration, and a closed-loop iterative mechanism for data-driven process parameter recommendation, verification, and optimization has not yet been formed, making it difficult for enterprises to effectively accumulate and pass on their valuable process experience.

[0007] Therefore, how to provide a stamping process parameter optimization method that can balance the deterministic recommendation of known working conditions with the generalization ability of unknown working conditions, and realize continuous optimization of process parameters and closed-loop self-learning, is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0008] In view of this, the present invention aims to propose a method and system for optimizing stamping process parameters based on rule-model two-level decision-making, in order to solve the technical problems in the prior art, such as excessive reliance on manual experience in determining stamping process parameters, long trial and debugging cycles, poor consistency in process execution, low utilization of historical data, and lack of data-driven closed-loop self-learning and continuous optimization mechanisms.

[0009] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0010] This invention provides a method for optimizing stamping process parameters based on rule-model two-level decision-making, comprising the following steps:

[0011] Step S1: Collect process data and construct a training sample set; obtain historical data of stamping process, extract multiple process parameters as input features, use forming quality score as output label, and construct a training sample set after preprocessing the data;

[0012] Step S2: Correlation analysis and feature selection; perform correlation analysis and collinearity selection on the input features and output labels, remove features with strong collinearity, and obtain the preferred feature subset;

[0013] Step S3: Mapping Modeling; Based on the preferred feature subset, train a random forest model and a gradient boosting tree model, and fuse the random forest model and the gradient boosting tree model through a stacking fusion strategy to obtain a parameter mapping model;

[0014] Step S4: Rule base construction; Extract high-quality process records from the training sample set, cluster and summarize them according to the working conditions, encapsulate the optimal parameter combinations under similar working conditions into rule entries, and construct a parameter self-selection rule base;

[0015] Step S5: Two-layer decision recommendation; Receive the current operating condition parameters, first query the parameter self-selected rule base, if a high-confidence rule is hit, directly output the recommended parameter combination; if no rule is hit or only a low-confidence rule is hit, call the parameter mapping model to make a prediction, output the recommended parameter combination, and perform interpretability analysis on the recommendation result based on the SHAP value;

[0016] Step S6: Closed-loop self-learning; Obtain the actual forming results and number of molding trials after trial molding in the workshop according to the recommended parameters as feedback signals, perform incremental training and update of the parameter mapping model, and update the parameter self-selection rule base.

[0017] Furthermore, in step S1, the input features include 16 process parameters: material thickness, mold type, blank holder force, mold clearance, forming speed, number of strokes, lubrication conditions, mold fillet radius, mold surface roughness, sheet yield strength, sheet tensile strength, sheet elongation, billet size, blank holder resistance, friction coefficient, and mold temperature.

[0018] The output label is the forming quality score, which includes three levels: first-class part, second-class part, and scrap.

[0019] The data preprocessing includes: applying a sliding window to smooth the mean of continuous process parameters, wherein the width N of the sliding window is preferably 10 to 30, and performing one-hot encoding on discrete parameters.

[0020] Furthermore, step S2 specifically includes:

[0021] The Pearson correlation coefficient was used to calculate the linear correlation strength between each continuous input feature and the output label.

[0022] The mutual information method is used to calculate the nonlinear correlation strength between each input feature and the output label;

[0023] After normalizing the Pearson correlation coefficient and mutual information value, the combined correlation strength is obtained and sorted in descending order;

[0024] Calculate the variance inflation factor among each feature, and remove features whose variance inflation factor is greater than a preset threshold, preferably 5 to 15, to obtain the preferred feature subset.

[0025] Furthermore, in step S3, the Stacking fusion strategy specifically includes:

[0026] The random forest model and the gradient boosting tree model are used as the first-layer base models and cross-validation training is performed respectively. The out-of-sample prediction results of each base model on the validation fold are used as the input features of the second layer.

[0027] The logistic regression model is trained in the second layer as a meta-learner, and the prediction output of the first layer base model is fused to obtain the final parameter mapping model.

[0028] Furthermore, in step S4, the method for constructing the parameter self-selection rule base also includes:

[0029] Each rule is configured to include a condition field, a recommended parameter combination field, a confidence field, and a hit count field.

[0030] When the number of hits of a rule exceeds a preset threshold and the verification success rate exceeds a preset ratio, the rule is marked as a high-confidence rule; wherein, the preset threshold is preferably 10 to 50 times, and the preset ratio is preferably 80% to 95%.

[0031] Furthermore, in step S5, the interpretability analysis of the recommendation results based on the SHAP value specifically includes:

[0032] Calculate the SHAP value of each process parameter and output the weighted ranking of the influence of each process parameter on the forming quality;

[0033] The candidate parameter schemes are ranked according to the model prediction confidence level. When the confidence level is lower than a preset threshold, a review prompt message is output. The preset threshold is preferably 0.6 to 0.8.

[0034] Furthermore, in step S6, the closed-loop self-learning specifically includes:

[0035] When the trial molding result is a first-class part, the working condition and parameter combination is added to the training set as a positive sample and written into the candidate area of ​​the rule base. After the cumulative number of successful times reaches the target, it is promoted to a new rule entry.

[0036] When the trial molding result is a level 2 part or scrap, it is added to the training set as a negative sample. The failure reason is analyzed by SHAP value, and the confidence of the corresponding rule item is reduced. The confidence update formula is as follows: The attenuation coefficient Preferably, it is between 0.8 and 0.95;

[0037] When the number of consecutive trial failures under the same working condition exceeds a preset threshold, the corresponding entry in the rule base is downgraded or invalidated; wherein, the preset threshold is preferably 3 to 5 times.

[0038] Furthermore, it also includes step S7: tracing the source of forming quality anomalies;

[0039] The forming quality data output from the stamping process is correlated with the corresponding stamping process parameters to establish a traceability mapping relationship;

[0040] When a molding quality abnormality occurs, the contribution of each process parameter to the quality abnormality is analyzed based on the SHAP value of the parameter mapping model. The key process parameters that caused the quality deviation and their deviation direction are traced back in combination with historical trial mold records, and a root cause analysis report is output.

[0041] Based on the same concept, the present invention also provides a stamping process parameter optimization system based on rule-model two-level decision-making, applied to the aforementioned stamping process parameter optimization method based on rule-model two-level decision-making, characterized in that it includes:

[0042] The data acquisition module is used to collect historical data of the stamping process and real-time operating parameters to build a training sample set;

[0043] The feature engineering module is used to perform feature engineering processing and correlation analysis on the collected data, select the best feature subset, and train the fusion model as a parameter mapping model.

[0044] The rule base management module is used to store and manage the parameter-selected rule base, and supports the addition, updating, downgrading and invalidation of rules;

[0045] The two-layer decision architecture module is used to receive current operating condition parameters and execute a two-layer decision process. The first layer queries the rule base to match the hit items, and the second layer calls the mapping model to predict and complete.

[0046] The interpretability analysis module is used to perform feature importance analysis and influence weight ranking output on the recommendation results based on SHAP values;

[0047] The molding quality anomaly tracing module is used to correlate the process parameters of the stamping process with the final molding quality inspection data, so as to achieve accurate tracing and root cause analysis of molding quality.

[0048] Based on the same concept, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method for optimizing stamping process parameters based on rule-model bi-level decision-making.

[0049] Compared with existing technologies, the stamping process parameter optimization method and system based on rule-model bi-level decision-making described in this invention has the following beneficial effects:

[0050] 1. Improve the accuracy, efficiency, and generalization ability of parameter recommendation. This invention constructs a two-layer decision architecture consisting of a parameter self-selection rule base and a parameter mapping model. Under known conditions, it prioritizes matching high-confidence rules, ensuring the determinism of recommendations and extremely fast response speed, significantly shortening the trial and debugging cycle. Under unknown conditions, it calls the Stacking fusion model for prediction completion, combining the advantages of random forest and gradient boosting tree algorithms, significantly improving the accuracy and robustness of parameter prediction, and effectively solving the problem of insufficient generalization ability of a single model.

[0051] 2. Enhancing Model Interpretability and Process Transparency. This invention introduces SHAP values ​​to analyze the interpretability of recommendation results, enabling a fair and quantitative calculation of the impact weights of each process parameter on forming quality. This mechanism breaks the black-box effect of traditional machine learning models, providing clear logical support for recommendation results and helping process engineers quickly understand the basis for recommendations, thereby effectively improving the adoption rate of system recommendations.

[0052] 3. Achieving closed-loop self-learning and continuous accumulation of process experience. This invention establishes a closed-loop self-learning mechanism based on actual trial mold feedback. By using trial mold results as positive and negative sample increments to train the model, and dynamically updating the confidence and status of rule base entries, it solves the technical defects of traditional stamping processes that rely excessively on manual experience, are difficult to standardize, and lack continuous accumulation. It enables virtuous cycle iteration of process parameter recommendation, verification, and optimization, allowing enterprises to effectively accumulate and inherit valuable process experience.

[0053] 4. Possesses precise anomaly tracing and root cause analysis capabilities. This invention establishes a tracing mapping relationship between stamping process parameters and forming defects. When batch forming quality anomalies occur during production, the system can quickly trace back the key process parameters causing the quality deviation and their deviation direction based on SHAP value analysis, outputting a root cause analysis report. This significantly shortens troubleshooting time and improves the stability and quality control level of stamping production. Attached Figure Description

[0054] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0055] Figure 1 This is a schematic diagram of the overall process of a stamping process parameter optimization method based on rule-model two-level decision-making provided by an embodiment of the present invention;

[0056] Figure 2 This is a flowchart illustrating the feature engineering and correlation analysis process provided in an embodiment of the present invention;

[0057] Figure 3 This is a schematic diagram of the Stacking fusion model provided in an embodiment of the present invention;

[0058] Figure 4 This is a schematic diagram of the workflow of the two-layer decision-making architecture provided in an embodiment of the present invention;

[0059] Figure 5 This is a schematic diagram of the workflow of the closed-loop self-learning mechanism provided in the embodiments of the present invention;

[0060] Figure 6 This is a schematic diagram of the module structure of a stamping process parameter optimization system based on rule-model two-level decision-making provided in an embodiment of the present invention. Detailed Implementation

[0061] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0062] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0063] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0064] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0065] like Figures 1 to 6 As shown, a method and system for optimizing stamping process parameters based on rule-model bi-level decision-making are disclosed. The method for optimizing stamping process parameters based on rule-model bi-level decision-making includes the following steps:

[0066] Step S1: Collect process data and construct a training sample set; obtain historical data of stamping process, extract multiple process parameters as input features, use forming quality score as output label, and construct a training sample set after preprocessing the data;

[0067] Step S2: Correlation analysis and feature selection; perform correlation analysis and collinearity selection on the input features and output labels, remove features with strong collinearity, and obtain the preferred feature subset;

[0068] Step S3: Mapping Modeling; Based on the preferred feature subset, train a random forest model and a gradient boosting tree model, and fuse the random forest model and the gradient boosting tree model through a stacking fusion strategy to obtain a parameter mapping model;

[0069] Step S4: Rule base construction; Extract high-quality process records from the training sample set, cluster and summarize them according to the working conditions, encapsulate the optimal parameter combinations under similar working conditions into rule entries, and construct a parameter self-selection rule base;

[0070] Step S5: Two-layer decision recommendation; Receive the current operating condition parameters, first query the parameter self-selected rule base, if a high-confidence rule is hit, directly output the recommended parameter combination; if no high-confidence rule is hit or only a low-confidence rule is hit, the hit rule that does not meet the high-confidence rule condition is a low-confidence rule, then call the parameter mapping model to make a prediction, output the recommended parameter combination, and perform interpretability analysis on the recommendation result based on the SHAP value;

[0071] Step S6: Closed-loop self-learning; Obtain the actual forming results and number of molding trials after trial molding in the workshop according to the recommended parameters as feedback signals, perform incremental training and update of the parameter mapping model, and update the parameter self-selection rule base;

[0072] Step S7: Tracing the source of forming quality anomalies; Associating the forming quality data output from the stamping process with the corresponding stamping process parameters to establish a traceability mapping relationship;

[0073] When a molding quality abnormality occurs, the contribution of each process parameter to the quality abnormality is analyzed based on the SHAP value of the parameter mapping model. The key process parameters that caused the quality deviation and their deviation direction are traced back in combination with historical trial mold records, and a root cause analysis report is output.

[0074] The aforementioned stamping process parameter optimization system based on rule-model bi-level decision-making, applied to the aforementioned stamping process parameter optimization method based on rule-model bi-level decision-making, includes:

[0075] The data acquisition module is used to collect historical data of the stamping process and real-time operating parameters to build a training sample set;

[0076] The feature engineering module is used to perform feature engineering processing and correlation analysis on the collected data, select the best feature subset, and train the fusion model as a parameter mapping model.

[0077] The rule base management module is used to store and manage the parameter-selected rule base, and supports the addition, updating, downgrading and invalidation of rules;

[0078] The two-layer decision architecture module is used to receive current operating condition parameters and execute a two-layer decision process. The first layer queries the rule base to match the hit items, and the second layer calls the mapping model to predict and complete.

[0079] The interpretability analysis module is used to perform feature importance analysis and influence weight ranking output on the recommendation results based on SHAP values;

[0080] The molding quality anomaly tracing module is used to correlate the process parameters of the stamping process with the final molding quality inspection data, so as to achieve accurate tracing and root cause analysis of molding quality.

[0081] The specific implementation method is as follows:

[0082] Example 1: A method for optimizing stamping process parameters based on rule-model bi-level decision-making

[0083] like Figure 1 As shown, this embodiment provides a method for optimizing stamping process parameters based on a rule-model two-layer decision-making approach, applied to the process parameter recommendation scenario in automotive sheet metal stamping workshops. This method covers the entire process from data acquisition, feature engineering, model training, rule base construction to two-layer decision-making recommendation and closed-loop self-learning. Through a two-layer decision-making architecture that prioritizes rule bases and completes model predictions, it balances the determinism of recommendations under known conditions with the generalization ability under unknown conditions.

[0084] Specifically, the method provided in this embodiment includes the following steps:

[0085] Step S1: Collect process data and construct a training sample set

[0086] like Figure 2 As shown, this step first collects and preprocesses process data from multiple sources. Historical stamping process data from the workshop Manufacturing Execution System (MES) or field control equipment is obtained. The data in this embodiment covers multiple typical workpieces and multiple mainstream dies. Several process parameters are extracted from the valid process records as input features. Specifically, the input features include 16 process parameters: material thickness, die type, blank holder force, die clearance, forming speed, number of strokes, lubrication conditions, die corner radius, die surface roughness, sheet yield strength, sheet tensile strength, sheet elongation, billet size, blank holder resistance, friction coefficient, and die temperature.

[0087] These 16 parameters comprehensively reflect the physical mechanisms of the stamping process: material thickness, yield strength, tensile strength, and elongation determine the deformation resistance and forming limit of the sheet metal; die type, fillet radius, surface roughness, and clearance determine the geometric constraints and friction state of material flow; blank holder force, blank holder resistance, and friction coefficient directly affect the resistance of material flowing into the die; forming speed, stroke count, lubrication conditions, and die temperature reflect the influence of the dynamic processing environment on material properties. Let the input feature vector be... .

[0088] Simultaneously, the output label is set to the forming quality rating. The forming quality rating includes three levels: Level 1, Level 2, and Scrap. Level 1 represents a qualified product that fully meets design requirements; Level 2 represents a semi-finished product with minor defects but acceptable or requiring rework; Scrap represents a defective product with serious defects that renders it unusable. Let the output label be... .

[0089] In the data preprocessing stage, in order to eliminate data noise and standardize the data format, the following operations are performed:

[0090] First, invalid samples with missing records, abnormal fields, or duplicate entries are removed. For missing values, the median of similar working conditions is used for filling; for outliers, the median is used for filling. The principles are used to identify and eliminate [illegible items].

[0091] Secondly, a sliding window is used to smooth the continuous process parameters to eliminate noise fluctuations. Let the original continuous parameter sequence be... Set the window width to , The preferred value range is 10 to 30. In this embodiment... The smoothed parameter values The calculation formula is as follows:

[0092] ;

[0093] in, This represents the floor function. For points at sequence boundaries, a truncated window or mirror fill method is used. This formula effectively filters out high-frequency noise while preserving the trend characteristics of the parameters.

[0094] Finally, the discrete parameters are one-hot encoded and converted into binary vector representations. A unified index is established by number, and the records are uniformly converted into CSV format. A globally unique index is created in the form of workpiece number, process, and timestamp to construct a training sample set. ,in For the first The input feature vector of each sample, For the corresponding output labels, The total number of samples.

[0095] Step S2: Correlation Analysis and Feature Screening

[0096] like Figure 2 As shown, this step performs a deep correlation analysis between the 16 input features and the forming quality label to extract a preferred feature subset, thereby reducing the dimensionality of subsequent model training and improving generalization ability. The specific process includes:

[0097] (1) The linear correlation strength between each continuous input feature and the output label is calculated using the Pearson correlation coefficient. Let the input feature vector be... The output label vector is Pearson correlation coefficient The calculation formula is:

[0098] ;

[0099] in, and They are respectively and The mean of the formula, , Refers to the first The specific values ​​for each sample.

[0100] (2) The mutual information method is used to calculate the nonlinear correlation strength between each input feature and the output label. Mutual information It measures the amount of information shared between two variables, and the formula is as follows:

[0101] ;

[0102] in, It is a joint probability distribution. and It is a marginal probability distribution. It represents the logarithm to the base 2.

[0103] (3) The Pearson correlation coefficient and mutual information value are fused to obtain the comprehensive correlation strength. First, normalization is performed, and the absolute value of the normalized Pearson correlation coefficient is set to 1. The normalized mutual information value is Overall correlation strength The calculation formula is:

[0104] ;

[0105] In this embodiment, weights are taken. In practical applications, and The value can be adjusted between 0.3 and 0.7. (According to...) Sort in descending order to initially identify features with higher importance.

[0106] (4) Calculate the variance inflation factor (VIF) among the features to eliminate the effects of multicollinearity. For features Perform a linear regression on it against all other features to obtain the coefficient of determination. The VIF calculation formula is:

[0107] ;

[0108] Eliminate Redundant features, where the threshold Preferably 5 to 15, in this embodiment we take 5 to 15. The optimal feature subset is finally obtained.

[0109] Step S3: Mapping Modeling

[0110] like Figure 3 As shown, this step constructs and trains a parameter mapping model based on the preferred feature subset. This embodiment employs a stacking fusion strategy to fuse a Random Forest (RF) model with a Gradient Boosting Decision Tree (GBDT) model.

[0111] The specific training process is as follows:

[0112] (1) Train a random forest model and a gradient boosting tree model as the first-layer base models respectively. For the random forest model, set the number of decision trees to 200, the maximum depth to 15, and the minimum number of sample splits to 5; for the gradient boosting tree model, set the number of iterations to 300, the learning rate to 0.1, and the maximum depth to 8. The above hyperparameters can be adjusted according to the actual amount of data. Adjust within a certain range.

[0113] (2) A stacking fusion strategy is used for fusion. In the first layer, five-fold cross-validation is performed on both the trained random forest model and the gradient boosting tree model. The out-of-sample prediction results of each base model at the validation fold are used as the input features of the second layer. Let the prediction output of the random forest model be... The predicted output of the gradient boosting tree model is Then the input feature vector of the second layer for:

[0114] ;

[0115] In the second layer, a logistic regression model is trained as a meta-learner. This logistic regression model fuses the prediction outputs of the two base models from the first layer. Using the cross-validation average as the evaluation metric, this stacking fusion model is selected as the final parameter mapping model. .

[0116] Step S4: Rule base construction

[0117] A parameter-selectable rule base is constructed by combining historical experience and historical records. Specific construction methods include:

[0118] (1) Extract high-quality process records of parts with a forming quality score of Grade 1 from the training sample set.

[0119] (2) Clustering and summarizing according to working conditions. After standardizing the data, clustering algorithms such as K-Means, DBSCAN or hierarchical clustering are used to encapsulate the optimal parameter combination under similar working conditions into rule entries.

[0120] (3) Set each rule to include a condition field, a recommended parameter combination field, a confidence field, and a hit count field.

[0121] (4) When the number of hits of a certain rule exceeds the preset threshold The preset threshold Preferably, the number of tests is 10 to 50; in this embodiment, 20 tests are used, and the success rate of verification exceeds a preset ratio. The preset ratio Preferably, it is 80% to 95%. In this embodiment, when it is 90%, it is marked as a high confidence rule.

[0122] Step S5: Two-layer decision recommendation

[0123] like Figure 4 As shown, this step obtains the current operating condition parameters and performs a two-layer decision recommendation through the rule base and parameter mapping model.

[0124] Receive current operating parameters Then, the first level of decision-making is performed by querying the parameter self-selection rule base to determine whether the current working condition matches an existing rule.

[0125] If a high-confidence rule is hit, the recommended parameter combination corresponding to the rule will be output directly. And mark the recommended source as a rule base hit.

[0126] If no rule is hit, or only a low-confidence rule is hit, then proceed to the second-level decision-making stage and invoke the parameter mapping model. Predict the current operating conditions and output recommended parameter combinations. With confidence level .

[0127] Simultaneously, interpretability analysis of the recommendation results is performed based on the SHAP value. Let the model prediction result be... The input feature dimension is The To optimize the number of features in a feature subset, the SHAP value is used. satisfy:

[0128] ;

[0129] in As the baseline value, For the first The contribution value of each feature is calculated. The weighted ranking of the influence of each process parameter on the forming quality is output by calculating the SHAP value of each process parameter. Based on the model prediction confidence level, the priority ranking of alternative parameter schemes is output. When the confidence level is lower than a preset threshold, a review prompt message is output. That is, when the confidence level... When the value is below a preset threshold, preferably between 0.6 and 0.8, the system outputs a review prompt message.

[0130] Step S6: Closed-loop self-learning

[0131] like Figure 5 As shown, this step performs closed-loop updates and self-learning on the rule base and model based on actual trial molding feedback. The actual forming results and number of trials after trial molding according to recommended parameters in the workshop are obtained and used as feedback signals to perform the following updates:

[0132] (1) When the trial molding result is a first-class part, the working condition and parameter combination are added to the training set as positive samples and written into the candidate area of ​​the rule base. The target is reached when the cumulative number of successful tests for the same working condition reaches the required level. Subsequently, it was promoted to a new rule entry. Simultaneously, this positive sample was used for incremental training of the parameter mapping model.

[0133] (2) When the trial molding result is a secondary part or scrap, it is added to the training set as a negative sample. The abnormal contribution parameters are analyzed using SHAP values ​​to clarify the cause of failure and reduce the confidence of the corresponding rule entries. The preferred confidence update formula is... The attenuation coefficient The preferred value is 0.8 to 0.95.

[0134] (3) When the number of consecutive trial failures under the same working condition exceeds a preset threshold, preferably 3 to 5 times, the corresponding entry in the rule base is downgraded or invalidated.

[0135] Step S7: Tracing the source of molding quality anomalies

[0136] Furthermore, the method of the present invention also includes a step for tracing the source of forming quality anomalies. The forming quality data output from the stamping process is correlated with the corresponding stamping process parameters to establish a source-tracing mapping relationship between the stamping process parameters and forming defects. When a forming quality anomaly occurs, the contribution of each process parameter to the quality anomaly is analyzed based on the SHAP value of the parameter mapping model. Combined with successful parameter combinations under similar working conditions in historical mold trial records, the key process parameters causing the quality deviation and their deviation direction are traced back, and a root cause analysis report is output.

[0137] Example 2: A Stamping Process Parameter Optimization System Based on Rule-Model Two-Level Decision Making

[0138] like Figure 6 As shown, this embodiment provides a system implementation of the above method. The system runs on a computer device including a processor and memory, such as a server or industrial control computer. The processor executes the computer program stored in the memory to implement the above method. The system includes the following functional modules:

[0139] (1) Data acquisition module: used to collect historical data and real-time operating parameters of the stamping process. This module connects to the workshop MES system and on-site industrial control equipment. Data storage adopts a hybrid storage scheme of structured process data and file object storage.

[0140] (2) Feature Engineering Module: This module is used to perform feature engineering and correlation analysis on the collected data and train the fusion model as a parameter mapping model. This module includes a feature preprocessing submodule, a correlation analysis submodule, and a model training submodule.

[0141] (3) Rule base management module: used to store and manage the parameter selection rule base. This module supports the addition, updating, downgrading and invalidation of rules.

[0142] (4) Two-layer decision architecture module: used to receive current operating condition parameters and execute a two-layer decision process. The first layer queries the rule base to match the hit items, and the second layer calls the mapping model to predict and complete.

[0143] (5) Interpretability analysis module: Based on SHAP value, the recommendation results are analyzed for feature importance and ranked by influence weight.

[0144] (6) Molding quality anomaly traceability module: used to associate the process parameters of the stamping process with the final molding quality inspection data to achieve accurate traceability and root cause analysis of molding quality.

[0145] The two-layer decision architecture provides parameter recommendation capabilities to the outside world through a unified service interface, and supports multi-terminal access from process engineers' workbench, workshop dashboard and alarm screen.

[0146] In summary, compared with the prior art, the present invention has the following significant advantages:

[0147] 1. Balancing recommendation efficiency and generalization ability, significantly shortening the trial modeling cycle.

[0148] This invention cleverly solves the challenge of recommending parameters for known and unknown working conditions in stamping processes through a two-layer decision-making architecture consisting of a parameter self-selection rule base and a parameter mapping model constructed in steps S4 and S5. Under known working conditions, the system prioritizes querying and matching high-confidence rules, directly outputting recommended parameters. This helps avoid the inference delays of complex machine learning models, achieving millisecond-level response and ensuring the determinism of the recommendations. In cases of unknown working conditions or rule mismatch, the system automatically calls a Stacking fusion model for prediction completion. This two-layer mechanism utilizes both the reliability of historical experience and the generalization ability of the data model, effectively reducing the number of repeated trial and error attempts during new product introduction, significantly shortening the trial molding and debugging cycle, and reducing production line resource usage costs.

[0149] 2. Feature selection and fusion modeling significantly improve prediction accuracy and robustness.

[0150] In terms of feature engineering, step S2 of this invention uses a fusion of Pearson correlation coefficient and mutual information to calculate the comprehensive correlation strength, and introduces a variance inflation factor (VIF) to eliminate multicollinear features. This technique effectively overcomes the complex nonlinear relationships between stamping process parameters (such as the inverted U-shaped relationship where too small a die gap leads to cracking and too large a gap leads to wrinkling) and linear redundancy problems, providing the model with a high-quality, low-redundancy subset of preferred features. In terms of modeling, step S3 adopts a Stacking fusion strategy, combining the overfitting advantages of the Random Forest (RF) model with the strong fitting ability of the Gradient Boosting Tree (GBDT) model, and performs secondary fusion through a logistic regression meta-learner. Through the above fusion strategy, this fusion model can combine the advantages of the two models, theoretically significantly improving the accuracy and robustness of parameter mapping.

[0151] 3. Introduce SHAP interpretability analysis to break the black box effect and improve process transparency.

[0152] Traditional machine learning models are often difficult for process engineers to trust in industrial applications due to a lack of interpretability. This invention introduces SHAP value analysis in steps S5 and S6, fairly allocating the contribution of each process parameter to the model output based on game theory. By ranking the weights of each parameter's influence on forming quality, the system can clearly demonstrate the logical basis of the recommendation results (e.g., explicitly indicating that the blank holder force contributes the most to the current recommendation). This white-box analysis mechanism helps process engineers quickly understand the recommendation basis and provides verification prompts when the confidence level is below a threshold, effectively improving the transparency of the system's recommendation results and the on-site adoption rate.

[0153] 4. Establish a closed-loop self-learning mechanism to realize the dynamic accumulation and self-evolution of process experience.

[0154] The closed-loop self-learning mechanism designed in step S6 of this invention completely changes the traditional situation where process parameter setting relies on manual intervention and is difficult to continuously optimize. The system collects actual trial molding results as feedback signals: when the trial molding is successful (first-level part), the operating parameters are used as positive samples to incrementally train the model, and automatically promoted to new rule entries after cumulative success; when the trial molding fails (second-level part or scrap), it is used as a negative sample to train and the confidence of the corresponding rule is reduced; especially when consecutive failures exceed a preset threshold (e.g., 3 times), the rule is automatically downgraded or invalidated. This mechanism prevents erroneous rules from continuing to mislead production, enables dynamic updates and self-evolution of the rule base and model, and allows the company's valuable process experience to be standardized, accumulated, and continuously refined.

[0155] 5. Possesses accurate anomaly tracing capabilities, improving the efficiency of quality control and fault diagnosis.

[0156] The forming quality anomaly tracing module established in step S7 of this invention establishes a mapping relationship between process parameters and defect data. When batch cracking or wrinkling occurs during production, the system can quickly trace back the key process parameters causing the quality deviation and their deviation direction based on the SHAP value (such as indicating that the blank holder force is too high). This root cause analysis capability transforms traditional blind troubleshooting into precise location, significantly shortening the fault diagnosis time and improving the stability and quality control level of stamping production.

[0157] 6. The system architecture is modular and combines hardware and software to support multi-terminal collaboration and the construction of digital workshops.

[0158] The system implementation provided in Example 2 encapsulates the above methods into six modules: data acquisition, feature engineering, rule base management, two-layer decision-making, interpretability analysis, and anomaly tracing, and runs on a computer device containing a processor and memory. The system provides recommendation capabilities through a unified service interface, supporting multi-terminal access from process engineer workbenches, workshop dashboards, and alarm screens. This modular and standardized architecture not only facilitates system maintenance and upgrades but also perfectly meets the construction needs of modern intelligent manufacturing and digital workshops, possessing extremely high engineering application value.

[0159] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for optimizing stamping process parameters based on rule-model bi-level decision-making, characterized in that, Includes the following steps: Step S1: Collect process data and construct a training sample set; Historical stamping process data was obtained, multiple process parameters were extracted as input features, and forming quality scores were used as output labels. After data preprocessing, a training sample set was constructed. Step S2: Correlation analysis and feature selection; perform correlation analysis and collinearity selection on the input features and output labels, remove features with strong collinearity, and obtain the preferred feature subset; Step S3: Mapping Modeling; Based on the preferred feature subset, train a random forest model and a gradient boosting tree model, and fuse the random forest model and the gradient boosting tree model through a stacking fusion strategy to obtain a parameter mapping model; Step S4: Rule base construction; High-quality process records are extracted from the training sample set, clustered and summarized according to working conditions, and the optimal parameter combinations under similar working conditions are encapsulated as rule entries to build a parameter self-selection rule library. Step S5: Two-layer decision recommendation; Receive the current operating condition parameters, first query the parameter self-selected rule base, if a high-confidence rule is hit, directly output the recommended parameter combination; if no rule is hit or only a low-confidence rule is hit, call the parameter mapping model to make a prediction, output the recommended parameter combination, and perform interpretability analysis on the recommendation result based on the SHAP value; Step S6: Closed-loop self-learning; Obtain the actual forming results and number of molding trials after trial molding in the workshop according to the recommended parameters as feedback signals, perform incremental training and update of the parameter mapping model, and update the parameter self-selection rule base.

2. The method for optimizing stamping process parameters based on rule-model bi-level decision-making according to claim 1, characterized in that, In step S1, the input features include 16 process parameters: material thickness, mold type, blank holder force, mold clearance, forming speed, number of strokes, lubrication conditions, mold fillet radius, mold surface roughness, sheet yield strength, sheet tensile strength, sheet elongation, billet size, blank holder resistance, friction coefficient, and mold temperature. The output label is the forming quality score, which includes three levels: first-class part, second-class part, and scrap. The data preprocessing includes: applying a sliding window to smooth the mean of continuous process parameters and performing one-hot encoding on discrete parameters.

3. The method for optimizing stamping process parameters based on rule-model bi-level decision-making according to claim 1, characterized in that, Step S2 specifically includes: The Pearson correlation coefficient was used to calculate the linear correlation strength between each continuous input feature and the output label. The mutual information method is used to calculate the nonlinear correlation strength between each input feature and the output label; After normalizing the Pearson correlation coefficient and mutual information value, the combined correlation strength is obtained and sorted in descending order; Calculate the variance inflation factor among each feature, and remove features whose variance inflation factor is greater than a preset threshold to obtain the preferred feature subset.

4. The method for optimizing stamping process parameters based on rule-model bi-level decision-making according to claim 1, characterized in that, In step S3, the Stacking fusion strategy is specifically as follows: The random forest model and the gradient boosting tree model are used as the first-layer base models and cross-validation training is performed respectively. The out-of-sample prediction results of each base model on the validation fold are used as the input features of the second layer. The logistic regression model is trained in the second layer as a meta-learner, and the prediction output of the first layer base model is fused to obtain the final parameter mapping model.

5. The method for optimizing stamping process parameters based on rule-model bi-level decision-making according to claim 1, characterized in that, In step S4, the method for constructing the parameter self-selection rule base further includes: Each rule is configured to include a condition field, a recommended parameter combination field, a confidence field, and a hit count field. When the number of hits of a rule exceeds a preset threshold and the verification success rate exceeds a preset ratio, the rule is marked as a high-confidence rule.

6. The method for optimizing stamping process parameters based on rule-model bi-level decision-making according to claim 1, characterized in that, In step S5, the interpretability analysis of the recommendation results based on the SHAP value specifically includes: Calculate the SHAP value of each process parameter and output the weighted ranking of the influence of each process parameter on the forming quality; Based on the model's prediction confidence level, the alternative parameter schemes are ranked in order of priority. When the confidence level is lower than a preset threshold, a review prompt message is output.

7. The method for optimizing stamping process parameters based on rule-model bi-level decision-making according to claim 1, characterized in that, In step S6, the closed-loop self-learning specifically includes: When the trial molding result is a first-class part, the working condition and parameter combination is added to the training set as a positive sample and written into the candidate area of ​​the rule base. After the cumulative number of successful times reaches the target, it is promoted to a new rule entry. When the trial molding result is a level 2 part or scrap, it is added to the training set as a negative sample. The failure reason is analyzed by SHAP value, and the confidence of the corresponding rule item is reduced. The confidence update formula is as follows: ,in, The attenuation coefficient; When the number of consecutive trial failures under the same working condition exceeds a preset threshold, the corresponding entry in the rule base will be downgraded or invalidated.

8. The method for optimizing stamping process parameters based on rule-model bi-level decision-making according to claim 1, characterized in that, It also includes step S7: Tracing the source of forming quality anomalies; The forming quality data output from the stamping process is correlated with the corresponding stamping process parameters to establish a traceability mapping relationship; When a molding quality abnormality occurs, the contribution of each process parameter to the quality abnormality is analyzed based on the SHAP value of the parameter mapping model. The key process parameters that caused the quality deviation and their deviation direction are traced back in combination with historical trial mold records, and a root cause analysis report is output.

9. A stamping process parameter optimization system based on rule-model bi-level decision-making, applied to the stamping process parameter optimization method based on rule-model bi-level decision-making as described in any one of claims 1 to 8, characterized in that, include: The data acquisition module is used to collect historical data of the stamping process and real-time operating parameters to build a training sample set; The feature engineering module is used to perform feature engineering processing and correlation analysis on the collected data, select the best feature subset, and train the fusion model as a parameter mapping model. The rule base management module is used to store and manage the parameter-selected rule base, and supports the addition, updating, downgrading and invalidation of rules; The two-layer decision architecture module is used to receive current operating condition parameters and execute a two-layer decision process. The first layer queries the rule base to match the hit items, and the second layer calls the mapping model to predict and complete. The interpretability analysis module is used to perform feature importance analysis and influence weight ranking output on the recommendation results based on SHAP values; The molding quality anomaly tracing module is used to correlate the process parameters of the stamping process with the final molding quality inspection data, so as to achieve accurate tracing and root cause analysis of molding quality.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the stamping process parameter optimization method based on rule-model bi-level decision-making as described in any one of claims 1 to 8.

Images