Heat treatment process optimization system based on multi-condition data evaluation

Abstract

The application is a heat treatment process optimization system based on multi-working condition data evaluation, and relates to the technical field of heat treatment process optimization, comprising: a dynamic weight distribution module for extracting current working condition information, matching with the working condition label library, generating an adaptive weight vector adapted to the current working condition, and performing online fine tuning of the weight when a working condition drift signal is detected.In the application, unsupervised clustering is performed on historical production data to generate a structured working condition label library, and data under different working conditions are given explicit classification identification, providing a working condition-aware basis for weight distribution; through matching of the working condition feature vector with the working condition label library and dynamic weight generation based on Bayesian optimization, the influence weight of each input feature can be adaptively adjusted according to the change of the working condition, overcoming the limitation that the static weight system cannot reflect the parameter influence difference under different working conditions.

Description

Technical Field

[0001] This invention relates to the field of heat treatment process optimization technology, and in particular to a heat treatment process optimization system based on multi-condition data evaluation. Background Technology

[0002] Optimal heat treatment process selection is a crucial step in manufacturing, significantly impacting the mechanical properties and microstructure of workpieces. In actual production, factors such as workpiece size variations, furnace loading changes, and equipment condition fluctuations create diverse operating conditions. Existing technologies typically employ static weight-based similarity matching methods, retrieving historical successful cases and recommending process parameters.

[0003] However, the weights of various process parameters on the final performance under different operating conditions exhibit dynamic changes, and static weighting systems cannot reflect these differences, leading to a decrease in the accuracy of process recommendations when operating conditions change. Furthermore, existing methods lack effective virtual verification methods before the recommended process is implemented, potentially resulting in significant discrepancies between the recommended results and the actual implementation, increasing production risks. Summary of the Invention

[0004] The purpose of this invention is to provide a heat treatment process optimization system based on multi-condition data evaluation in order to solve the above-mentioned problems.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A heat treatment process optimization system based on multi-condition data evaluation includes: The operating condition clustering module is used to collect historical production data of heat treatment, perform cleaning, standardization and unsupervised clustering, and generate an operating condition label library. The dynamic weight allocation module is used to extract current working condition information and match it with the working condition tag library to generate an adaptive weight vector that adapts to the current working condition, and to perform online fine-tuning of the weights when a working condition drift signal is detected. The process recommendation module is used to perform optimization search within the feasible domain of process parameters based on the adaptive weight vector and multi-objective constraints, and generate candidate recommended process schemes. The digital twin verification module is used to input the candidate recommended process scheme into the digital twin simulation model of the heat treatment process for virtual verification, and to perform parameter correction and re-verification when the indicators fail to meet the standards. The closed-loop feedback module is used to collect actual production data, perform three-way comparative analysis, trigger historical database updates, working condition label library updates, weight system re-optimization, and the calibration of the digital twin simulation model.

[0006] Preferably, generating an adaptive weight vector that fits the current operating condition includes: If the similarity between the current working condition feature vector and the center vector of a certain working condition cluster is higher than the preset matching threshold, then the current working condition is assigned to that working condition cluster, and the weight system obtained by statistical analysis of the working condition cluster on historical data is inherited as the initial weight vector. If the similarity between the current working condition feature vector and the center vectors of multiple working condition clusters is lower than the matching threshold, or if the current working condition does not belong to any existing working condition cluster, then based on the current working condition feature vector, a Bayesian optimization algorithm is used to interpolate or migrate in the historical weight space to generate a unique weight vector that is adapted to the current working condition.

[0007] Preferably, the statistical method of the weighting system is as follows: For the For each of the three operating condition clusters, the absolute value of the Pearson correlation coefficient between each input feature and the target performance index is used as the original importance score. The original importance scores of each input feature for: ; in For the first The first record Each input feature value For the corresponding target performance index value, The first in this operating condition cluster The mean of each input feature, This represents the average of the target performance indicators within this cluster of operating conditions. For the first The total number of historical records in each operating condition cluster; Normalize all the original importance scores to obtain the first... Weights of each input feature : ; in For the first The original importance scores of each input feature This represents the total number of input features.

[0008] Preferably, the step of using weighted Euclidean distance to retrieve similar historical cases in historical production data involves using weighted Euclidean distance. for: ; in The current working condition feature vector is the first... The values ​​of each feature, The first feature vector of the working condition in the historical record The values ​​of each feature, For the first The weights corresponding to each feature The total number of input features; weighted Euclidean distance The smaller the value, the more similar the historical record is to the current operating conditions.

[0009] Preferably, the optimization objectives of the multi-objective optimization search include mechanical property compliance rate, microstructure uniformity, energy efficiency, and process operability; wherein: The mechanical performance compliance rate is based on the proportion of historical samples that are similar to the current working conditions, where each mechanical performance index reaches the target value under a given combination of process parameters. The uniformity of the structure is scored by the reciprocal of the range or standard deviation of the metallographic structure grade test values ​​in different areas of the same workpiece. The energy efficiency is scored as the reciprocal of the average actual energy consumption per unit of qualified output. The operability of the process is determined by checking whether the type of cooling medium in the candidate recommended process scheme is in the list of media available in the current production environment and whether the heating temperature does not exceed the temperature limit of the current equipment.

[0010] Preferably, the multi-objective optimization search adopts a method combining multi-objective weighted summation and constraint satisfaction. Performance objectives directly related to input features with higher weights in the adaptive weight vector are given higher optimization priority, while performance objectives related to input features with lower weights are given lower optimization priority.

[0011] Preferably, the step of inputting the candidate recommended process scheme into the digital twin simulation model of the heat treatment process for virtual verification includes: The digital twin simulation model is based on the thermal properties of the workpiece material and uses the finite element method to simulate the temperature field evolution, microstructure transformation process and stress-strain distribution of the workpiece under the conditions of the candidate recommended process scheme. The performance indicators predicted by the simulation are compared with the target performance one by one. If all indicators meet the standards, the candidate recommended process scheme is verified. If there are indicators that do not meet the standards, a Bayesian optimization method is used to perform a directional search in the process parameter space based on the direction and magnitude of the deviation between the simulation prediction value and the target value to generate a corrected combination of process parameters. The corrected combination of process parameters is then input into the digital twin simulation model for verification. The correction and verification process is repeated until all simulation prediction results meet the standards or the preset maximum number of iterations is reached.

[0012] Preferably, when performing the directional search, the deviation amplitude is used as the basis for adjusting the search step size; a larger search step size is used when the deviation amplitude is large, and a smaller search step size is used when the deviation amplitude is small.

[0013] Preferably, the method further includes: When a condition drift signal, such as equipment status fluctuation or ambient temperature change, is detected, the matching relationship between the current condition feature vector and the condition label library is recalculated, and the online fine-tuning of the adaptive weight vector is automatically triggered.

[0014] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention performs unsupervised clustering on historical production data and generates a structured working condition label library. Data under different working conditions obtains explicit classification labels, providing a basis for working condition awareness for weight allocation. Through matching working condition feature vectors with the working condition label library and generating dynamic weights through Bayesian optimization, the influence weights of each input feature can be adaptively adjusted with changes in working conditions, overcoming the limitation of static weight systems that cannot reflect the differences in parameter influence under different working conditions.

[0015] 2. This invention reduces the risk of significant deviations between the recommended results and the actual execution effect by virtually verifying the candidate recommended process schemes by inputting them into the digital twin simulation model of the heat treatment process before actual execution, and by performing parameter correction and re-verification when the indicators fail to meet the standards. Attached Figure Description

[0016] Further details, features, and advantages of this application are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which: Figure 1 This is a system structure diagram of the present invention. Detailed Implementation

[0017] Several embodiments of this application will now be described in more detail with reference to the accompanying drawings to enable those skilled in the art to implement this application. This application may be embodied in many different forms and for various purposes and should not be limited to the embodiments set forth herein. These embodiments are provided to make this application thorough and complete, and to fully convey the scope of this application to those skilled in the art. The embodiments described do not limit this application.

[0018] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It will be further understood that terms such as those defined in commonly used dictionaries shall be interpreted as having a meaning consistent with their meaning in the relevant field and / or the context of this specification, and shall not be interpreted in an idealized or overly formal sense unless expressly defined herein.

[0019] Example 1 Its specific implementation method is combined with the appendix Figure 1 Please provide a detailed explanation.

[0020] In this embodiment, it includes: The optimal heat treatment process is a crucial step in manufacturing, significantly impacting the final mechanical properties and microstructure of workpieces. In actual production, the influence weights of various process parameters on the final performance dynamically change under different operating conditions (such as variations in workpiece size, furnace loading, and equipment status fluctuations). Traditional similarity matching methods based on static weights struggle to adapt to these changes. Furthermore, the lack of effective virtual verification methods before actual execution of recommended processes leads to potential significant discrepancies between the recommended results and the actual performance. To address these issues, this embodiment proposes a heat treatment process optimization method. Through multi-condition automatic clustering, condition-aware dynamic weight adaptive allocation, multi-objective process parameter recommendation, digital twin reverse verification and parameter correction, and closed-loop feedback self-evolution, it achieves adaptive process recommendation for complex and variable operating conditions and ensures the reliability of the recommendation results.

[0021] The historical production data for heat treatment involved in this embodiment is stored in the enterprise's production database. The computing resources required for digital twin simulation are provided by industrial servers or cloud computing platforms.

[0022] A preferred method for heat treatment according to an embodiment of the present invention includes the following modules: Module 1 collects historical production data of heat treatment and performs automatic clustering of multiple operating conditions to generate an operating condition label library.

[0023] Historical production data is retrieved from the enterprise's heat treatment workshop production database. Each data record includes workpiece information, equipment information, process parameters, process monitoring data, and quality inspection results. Workpiece information includes material grade, chemical composition, geometric dimensions, and shape complexity. Equipment information includes furnace type, heating method, temperature control accuracy, and furnace chamber dimensions. Process parameters include heating temperature, holding time, heating rate, cooling medium and cooling method, and tempering parameters. Process monitoring data includes actual temperature profiles, furnace atmosphere, and workpiece placement. Quality inspection results include hardness, tensile strength, yield strength, elongation, impact toughness, and metallographic structure grade.

[0024] The acquired historical production data is cleaned to remove missing and outlier records. The cleaned data is then standardized to map all feature fields to a numerical space with the same dimensions. Continuous numerical features (such as heating temperature, holding time, and geometric dimensions) are standardized using Z-score; non-numerical features (such as material grade, furnace type, and cooling medium type) are converted to numerical representations using unique thermal encoding.

[0025] For standardized historical production data, feature fields corresponding to workpiece information, equipment information, and process constraints are extracted and concatenated into a working condition feature vector for each record. All working condition feature vectors are then input into an unsupervised clustering algorithm, which outputs multiple working condition clusters. Each working condition cluster corresponds to a typical working condition, i.e., a data set with similar workpiece characteristics, equipment conditions, and process constraints.

[0026] Structured work condition tags are generated for each work condition cluster. Each work condition tag contains descriptive information about the typical workpiece range, equipment type, and process constraint boundaries of that work condition cluster. All work condition tags are compiled into a work condition tag library for subsequent work condition matching and retrieval.

[0027] It should be noted that the unsupervised clustering algorithms mentioned above are either the DBSCAN algorithm or a Gaussian mixture model. When using the DBSCAN algorithm, the input is all operating condition feature vectors. Density reachability is determined based on the Euclidean distance between the feature vectors, grouping densely connected samples into the same operating condition cluster. Samples not assigned to any cluster are marked as outliers. The output is the sample set for each operating condition cluster and the set of outliers. When using a Gaussian mixture model, the input is all operating condition feature vectors. The mean and covariance matrices of each Gaussian component are iteratively fitted using the expectation-maximization algorithm. Each record is assigned to the operating condition cluster corresponding to the component with the highest posterior probability. The output is the sample set for each operating condition cluster and the corresponding Gaussian component parameters.

[0028] Module 2 generates an adaptive weight vector for the current working condition based on working condition label matching and real-time working condition perception.

[0029] When a new heat treatment process optimization request arrives, complete working condition information of the workpiece to be processed is extracted, including workpiece information, equipment information, and current environmental conditions. The current working condition information is processed using the same standardization and encoding method as in Module 1, encoded into a working condition feature vector, and then its similarity is calculated with each working condition tag in the working condition tag library to identify the working condition type to which the current request belongs. Cosine similarity is used for the similarity calculation.

[0030] If the similarity between the current working condition feature vector and the center vector of a certain working condition cluster is higher than the preset matching threshold, then the current working condition is assigned to that working condition cluster, and the weight system obtained from the historical data of that working condition cluster is inherited as the initial weight vector.

[0031] If the current operating condition feature vector lies at the boundary of multiple operating condition clusters (i.e., its similarity to the center vectors of multiple operating condition clusters is below the matching threshold), or if the current operating condition does not belong to any existing operating condition cluster, then the dynamic weight generation process is initiated. Based on the current operating condition feature vector, a Bayesian optimization algorithm is used to interpolate or migrate within the historical weight space to generate a unique weight vector adapted to the current operating condition. The input to the Bayesian optimization algorithm is the current operating condition feature vector and the weight system of each historical operating condition cluster; the output is the weight vector adapted to the current operating condition.

[0032] It should be noted that the statistical method for the weighting system of the above historical operating condition clusters is as follows: for the first... A set of operating conditions, assuming that the set of operating conditions contains historical records, For the first The total number of historical records in each operating condition cluster. The first record The input feature values ​​are The corresponding target performance index value is ,in For record indexing, This serves as the input feature index. The absolute value of the Pearson correlation coefficient between each input feature and the target performance indicator is used as the original importance score. The original importance scores of each input feature for: ; in The first in this operating condition cluster The mean of each input feature, This represents the average of the target performance indicators within this operating condition cluster. Because... Let be the absolute value of the Pearson correlation coefficient, and its range is . Since they have the same dimensions, normalization operations can be performed directly. Let there be a total of... Each input feature This represents the total number of input features. Normalizing all original importance scores yields the [n]th [value]. Weights of each input feature :

[0033] in For the first The original importance scores of each input feature For the first The original importance scores of each input feature The summation index is used to iterate through all input features. This represents the total number of input features. (All) Weights of each input feature The weighting system that constitutes this cluster of operating conditions.

[0034] It should be noted that the above input features include feature fields corresponding to workpiece information, equipment information, and process constraints, that is, all the feature fields used in module 1 to assemble the working condition feature vector. Each feature field. Weight vector. The dimension of the eigenvector is the same as the dimension of the eigenvector of the operating condition, both being 0. The dimension defines the importance of each input feature to the target performance index under the current operating conditions, and is used for the weighted Euclidean distance calculation in module 3.

[0035] It should be noted that the above target performance indicators In single-objective scenarios, a single performance index value (such as hardness) is used; in multi-objective scenarios, the weighted composite score of the performance indices of each objective is used as the composite score. The weights of each target performance indicator are preset by the process requirements.

[0036] In this embodiment, to address changes in operating conditions during process optimization, the weight allocation employs a real-time adaptive adjustment method. When the system detects operating condition drift signals such as equipment status fluctuations or changes in ambient temperature, it recalculates the matching relationship between the current operating condition feature vector and the operating condition label library, automatically triggering online fine-tuning of the weight vector and updating the importance of each input feature.

[0037] Module 3 generates candidate recommended process solutions based on adaptive weight vectors and multi-objective constraints.

[0038] Based on an adaptive weight vector, several historical successful cases most similar to the current operating condition are retrieved from historical production data. The process parameter combinations corresponding to these historical successful cases are extracted to form an initial recommendation candidate set. During retrieval, the similarity between historical records and the current operating condition is measured using a weighted Euclidean distance. Since module 1 has performed uniform standardization on each feature field, each feature component is a dimensionless value and can directly participate in the weighted summation operation. Let the feature vector of the current operating condition be... The historical working condition feature vector is The adaptive weight vector is , The weighted Euclidean distance is the total number of input features. for: ; in The current working condition feature vector is the first... The values ​​of each feature, The first feature vector of the working condition in the historical record The values ​​of each feature, For the first The weights corresponding to each feature For feature indexing. Weighted Euclidean distance. The smaller the value, the more similar the historical record is to the current operating conditions.

[0039] Starting with the combination of process parameters in the initial recommended candidate set, a multi-objective optimization search is performed within the feasible region of the process parameters. The feasible region of the process parameters is determined by equipment capability constraints, including boundary conditions such as the upper limit of equipment temperature and the availability of cooling medium.

[0040] The objectives of multi-objective optimization include: mechanical property compliance rate, microstructure uniformity, energy efficiency, and process operability. Mechanical property compliance rate refers to the probability that indicators such as hardness and strength reach target values. Microstructure uniformity refers to the consistency of metallographic microstructure grades across different regions of the workpiece. Energy efficiency refers to the energy consumption per unit of qualified output. Process operability refers to the degree to which the availability of cooling media and the upper limit constraints of equipment temperature are met.

[0041] A multi-objective weighted summation method combined with constraint satisfaction is employed for optimization. In the adaptive weight vector, performance metrics with higher weights have higher priority in the optimization process. Multiple candidate process solutions are generated through local search optimization. Each candidate solution includes a predicted performance metric, a confidence interval, and similarity to historical successful cases.

[0042] It should be noted that the above mechanical performance compliance rate is calculated as follows: based on samples from historical production data that are similar to the current working conditions, the proportion of samples that achieve the target value for each mechanical performance index under a given combination of process parameters is statistically analyzed, and this proportion is used as an estimate of the compliance probability.

[0043] It should be noted that the above method for quantifying the uniformity of the structure is as follows: extract samples from historical production data that are similar to the current working conditions, and statistically analyze the range or standard deviation of the metallographic structure grade test values ​​in different areas of the same workpiece. The reciprocal of the range or standard deviation is used as the score for the uniformity of the structure. The smaller the range or standard deviation, the higher the score for the uniformity of the structure, indicating that the uniformity of the structure is better.

[0044] It should be noted that the above energy efficiency is calculated as follows: based on samples from historical production data that are similar to the current operating conditions, the average actual energy consumption corresponding to the unit qualified output under a given combination of process parameters is statistically analyzed, and the reciprocal of the average actual energy consumption is used as the energy efficiency score. The smaller the average actual energy consumption, the higher the energy efficiency score.

[0045] It should be noted that the above calculation method for process operability is as follows: It checks whether the cooling medium type in the candidate recommended process scheme is in the list of media available in the current production environment, and whether the heating temperature does not exceed the temperature limit of the current equipment. When both constraints are satisfied, the process operability score is the maximum value; when either constraint is not satisfied, the process operability score is the minimum value, and the candidate recommended process scheme is assigned the lowest priority in the optimization process.

[0046] It should be noted that the weights of each optimization objective in the above multi-objective weighted summation are the same as the adaptive weight vector in module 2. The correspondence is as follows: Adaptive weight vector This reflects the importance of each input feature to the target performance index. In multi-objective optimization, performance objectives directly related to input features with higher weights (such as the mechanical performance compliance rate) are given higher optimization priority, while performance objectives related to input features with lower weights are given lower optimization priority, so that the optimization search direction is consistent with the feature importance distribution of the current working condition.

[0047] Module 4 inputs the candidate recommended process schemes into the digital twin simulation model for virtual verification, and outputs the verified process scheme or the revised process scheme.

[0048] The candidate recommended process schemes are input into the digital twin simulation model of the heat treatment process. Based on the thermal properties of the workpiece material, the digital twin simulation model uses the finite element method to simulate the temperature field evolution, microstructure transformation, and stress-strain distribution of the workpiece under the conditions of the candidate recommended process schemes. The thermal properties include thermal conductivity, specific heat capacity, and latent heat of phase change.

[0049] The simulation output includes: temperature-time curves at various points inside the workpiece, degree of austenitization, microstructure composition after quenching, predicted hardness distribution, predicted deformation, and predicted residual stress.

[0050] The simulated performance indicators are compared one by one with the target performance in Module 3. If all indicators meet the standards, the candidate recommended process scheme is validated.

[0051] If any indicators fail to meet the standards, a parameter correction process is initiated. Based on the direction and magnitude of the deviation between the simulation predictions and the target values, a Bayesian optimization method is used to perform a targeted search in the process parameter space to generate a corrected combination of process parameters. The corrected combination of process parameters is then input into the digital twin simulation model of the heat treatment process for verification. The above correction and verification process is repeated until all simulation prediction results meet the standards or the preset maximum number of iterations is reached.

[0052] It should be noted that the above-mentioned unmet standards include situations such as insufficient core hardness or excessive deformation. When core hardness is insufficient, the deviation direction points to adjusting parameters by increasing the heating temperature or extending the holding time. When deformation exceeds the limit, the deviation direction points to adjusting parameters by decreasing the cooling rate or adjusting the cooling method.

[0053] In this embodiment, to improve the convergence efficiency of parameter correction, the deviation magnitude is used as the basis for adjusting the search step size during directional search. A larger search step size is used when the deviation magnitude is large to accelerate convergence, and a smaller search step size is used when the deviation magnitude is small to improve accuracy.

[0054] Module 5 collects actual production data and compares it with predicted performance indicators and simulation predictions, triggering model calibration and updating of the operating condition label library.

[0055] After the recommended process, validated by digital twins, is put into actual production, process data such as temperature curves and actual cooling rates during the actual heat treatment process are collected in real time. After the workpiece completes heat treatment, comprehensive quality inspection is performed to obtain actual hardness, tensile strength, yield strength, elongation, impact toughness, and metallographic structure grade.

[0056] The actual quality test results are compared with the predicted performance indicators in Module 3 and the simulated predicted values ​​in Module 4, and the prediction deviation and simulation deviation are calculated.

[0057] When a systematic deviation occurs between the actual quality inspection results and the predicted performance indicators, the complete data from this test is added as a new sample to the historical production database. The complete data includes current operating condition information, recommended process parameters, actual process data, and actual quality inspection results. Based on the updated historical production database, the operating condition clustering process in Module 1 is re-executed to update the operating condition label library. Simultaneously, the weight optimization process in Module 2 is re-executed to update the weight system corresponding to each operating condition cluster.

[0058] The simulation deviation data is fed back to the digital twin simulation model of the heat treatment process to calibrate the thermophysical parameters or phase transformation kinetic model coefficients in the digital twin simulation model of the heat treatment process, so that the simulation prediction accuracy can be continuously improved with the accumulation of actual data.

[0059] It should be noted that the aforementioned systematic bias refers to the tendency of actual results to deviate from predicted values ​​in the same direction multiple times under the same material or working conditions. For example, if the actual hardness of a certain material is repeatedly lower than the predicted value, it is considered that a systematic bias exists.

[0060] Example 2 The heat treatment workshop of a mechanical parts manufacturing company is undertaking the task of optimizing the quenching and tempering process for bearing rings. The workshop is equipped with both box-type resistance furnaces and salt bath furnaces, and commonly uses quenching oil and water-soluble quenching fluid as cooling media. Currently, they have received a batch of 120 bearing rings made of GCr15 bearing steel, with an outer diameter of 68 mm and a wall thickness of 8 mm. The required surface hardness is 61 to 65 HRC, and the metallographic structure grade is no lower than level 2. The workshop's historical database contains nearly three years of heat treatment production records, covering various bearing steel materials and specifications of bearing rings and shafts.

[0061] Step 1: Historical production data collection and automatic clustering under multiple operating conditions Valid records after cleaning were extracted from the historical production database, and fields such as workpiece size, material grade, furnace type, furnace load, heating temperature, holding time, and cooling medium were standardized. Continuous fields such as heating temperature and holding time were standardized using Z-score, while categorical fields such as material grade, furnace type, and cooling medium type were categorized using unique thermal encoding. After concatenating the fields to form a working condition feature vector, the DBSCAN algorithm was used to perform unsupervised clustering on all historical records, outputting four working condition clusters and a small number of outliers, as shown in Table 1 below.

[0062] Table 1. Clustering results and working condition labels for each working condition.

[0063] Condition cluster 1 closely matches the workpiece characteristics and equipment conditions of the current batch to be processed (GCr15 ring, outer diameter 68 mm, box-type resistance furnace, quenching oil), and is marked as the candidate matching condition cluster for the current request.

[0064] Step 2: Working Condition Matching and Adaptive Weight Vector Generation After processing the current workpiece's working condition information using the same standardization and one-hot encoding method as in step 1, it is encoded into a working condition feature vector. The cosine similarity is then calculated between this vector and the center vector of each working condition cluster in the working condition tag library. The cosine similarity between the current working condition feature vector and the center vector of working condition cluster 1 is 0.93, which is higher than the preset matching threshold of 0.80. Therefore, the current request belongs to working condition cluster 1, and its historical weight system is inherited as the initial adaptive weight vector.

[0065] Based on 312 historical records of working condition cluster 1, the original importance score between each input feature and the target performance index (surface hardness) was calculated according to the Pearson correlation coefficient formula. Then according to Normalization yields the weights of each feature. As shown in Table 2 below.

[0066] Table 2 Adaptive weights of the main features of operating condition cluster 1

[0067] Heating temperature and holding time have the highest weighting on surface hardness, which is consistent with the fact that the degree of austenitization of GCr15 bearing steel has a significant impact on quenching hardness.

[0068] Step 3: Generation of multi-objective process schemes based on adaptive weight vectors Based on the current working condition feature vector and adaptive weight vector Based on this, in the 312 historical records of condition cluster 1, according to the weighted Euclidean distance... The nearest neighbor historical records are retrieved, and the process parameter combinations of the top 5 records with the smallest weighted distance are extracted to form an initial recommended candidate set. Using the process parameters in the candidate set as the starting point, a multi-objective optimization search is performed under the constraints of the upper limit of equipment temperature (1050 degrees Celsius for box-type resistance furnace) and the availability of quenching oil, generating three candidate recommended process schemes, as shown in Table 3 below.

[0069] Table 3. Candidate Recommended Process Schemes and Multi-Objective Prediction Indicators

[0070] Option A achieved the best mechanical performance compliance rate and comprehensive multi-objective score, and thus entered the digital twin verification stage.

[0071] Step 4: Digital Twin Simulation Verification and Parameter Correction Scheme A (heating temperature 845 degrees Celsius, holding time 45 minutes, heating rate 8 degrees Celsius / minute, quenching oil cooling) was input into the digital twin simulation model of the heat treatment process. The simulation model, based on the thermophysical parameters (thermal conductivity, specific heat capacity, latent heat of phase transformation) of GCr15 steel, used the finite element method to simulate the temperature field evolution and microstructure transformation process in the thickness direction of the ring wall.

[0072] The simulation output shows that the predicted surface hardness is 63.2 HRC, reaching the target range of 61 to 65 HRC; the predicted core hardness is 60.8 HRC, slightly lower than the target lower limit of 61 HRC, indicating insufficient core hardness. The predicted deformation is 0.03 mm, which is within the limit, as shown in Table 4 below.

[0073] Table 4 Comparison of Simulation Predictions for Scheme A and the Corrected Scheme

[0074] To address the insufficient core hardness, the deviation was directed towards increasing the heating temperature or extending the holding time. A Bayesian optimization-oriented search was employed, adjusting parameters with a small search step size corresponding to the deviation amplitude (0.2 HRC) to generate a corrective solution: the heating temperature was adjusted to 850 degrees Celsius, the holding time to 48 minutes, and all other parameters remained unchanged. The corrected solution was then input into the simulation model again, and all indicators met the requirements, thus the corrected solution was validated.

[0075] Step 5: Actual production data collection and closed-loop feedback After the revised plan was implemented in the workshop, the temperature curve and actual cooling rate were collected in real time. After the workpiece completed the heat treatment, a comprehensive quality inspection was carried out, as shown in Table 5 below.

[0076] Table 5. Comparison of Actual Detection Results, Predicted Values, and Simulated Predicted Values

[0077] The actual surface hardness and core hardness were both lower than the simulation predictions, with the deviations occurring in the same direction, initially showing a unidirectional shift. The complete data (working condition information, recommended process parameters, actual process data, and actual test results) was added to the historical production database as a new sample. If subsequent batches of the same material continue to show a trend of actual hardness values ​​lower than simulation predictions, it will be determined as a systematic deviation, triggering re-clustering of the working condition label library and re-optimization of the working condition cluster 1 weight system. The simulation deviation data will then be fed back to the digital twin simulation model to calibrate the phase transformation kinetics model coefficients of GCr15 steel.

[0078] The data flow throughout the entire implementation process exhibits a clear hierarchical transmission relationship. After standardization, 312 records from the historical production database (Cluster 1) are used to generate adaptive weight vectors through Pearson correlation coefficient calculation. These weight vectors drive weighted Euclidean distance retrieval, providing initial points for multi-objective optimization. The optimized candidate solutions are then validated in digital twin simulation. Simulation biases drive Bayesian parameter correction, and the actual production data from the corrected solutions ultimately feeds back into the historical database, forming a complete closed loop. The output of each step serves as the direct input for the next step. The data's operating condition identification, feature encoding method, and dimensional system remain consistent throughout the entire process, ensuring the continuity and logical consistency of data flow.

[0079] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0080] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0081] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A heat treatment process optimization system based on multi-condition data evaluation, characterized in that, include: The operating condition clustering module is used to collect historical production data of heat treatment, perform cleaning, standardization and unsupervised clustering, and generate an operating condition label library. The dynamic weight allocation module is used to extract current working condition information and match it with the working condition tag library to generate an adaptive weight vector that adapts to the current working condition, and to perform online fine-tuning of the weights when a working condition drift signal is detected. The process recommendation module is used to perform optimization search within the feasible domain of process parameters based on the adaptive weight vector and multi-objective constraints, and generate candidate recommended process schemes. The digital twin verification module is used to input the candidate recommended process scheme into the digital twin simulation model of the heat treatment process for virtual verification, and to perform parameter correction and re-verification when the indicators fail to meet the standards. The closed-loop feedback module is used to collect actual production data, perform three-way comparative analysis, trigger historical database updates, working condition label library updates, weight system re-optimization, and the calibration of the digital twin simulation model.

2. The heat treatment process optimization system based on multi-condition data evaluation according to claim 1, characterized in that, Generating an adaptive weight vector that fits the current operating conditions includes: If the similarity between the current working condition feature vector and the center vector of a certain working condition cluster is higher than the preset matching threshold, then the current working condition is assigned to that working condition cluster, and the weight system obtained by statistical analysis of the working condition cluster on historical data is inherited as the initial weight vector. If the similarity between the current working condition feature vector and the center vectors of multiple working condition clusters is lower than the matching threshold, or if the current working condition does not belong to any existing working condition cluster, then based on the current working condition feature vector, a Bayesian optimization algorithm is used to interpolate or migrate in the historical weight space to generate a unique weight vector that is adapted to the current working condition.

3. The heat treatment process optimization system based on multi-condition data evaluation according to claim 2, characterized in that, The statistical method for the weighting system is as follows: For the For each of the three operating condition clusters, the absolute value of the Pearson correlation coefficient between each input feature and the target performance index is used as the original importance score. The original importance scores of each input feature for: ； in For the first The first record Each input feature value For the corresponding target performance index value, The first in this operating condition cluster The mean of each input feature, This represents the average of the target performance indicators within this cluster of operating conditions. For the first The total number of historical records in each operating condition cluster; Normalize all the original importance scores to obtain the first... Weights of each input feature : ； in For the first The original importance scores of each input feature This represents the total number of input features.

4. The heat treatment process optimization system based on multi-condition data evaluation according to claim 1, characterized in that, Using weighted Euclidean distance to retrieve similar historical cases from historical production data. for: ； in The current working condition feature vector is the first... The values ​​of each feature, The first feature vector of the working condition in the historical record The values ​​of each feature, For the first The weights corresponding to each feature The total number of input features; Weighted Euclidean distance The smaller the value, the more similar the historical record is to the current operating conditions.

5. The heat treatment process optimization system based on multi-condition data evaluation according to claim 1, characterized in that, The optimization objectives of the multi-objective optimization search include mechanical property compliance rate, microstructure uniformity, energy efficiency, and process operability; among which: The mechanical performance compliance rate is based on the proportion of historical samples that are similar to the current working conditions, where each mechanical performance index reaches the target value under a given combination of process parameters. The uniformity of the structure is scored by the reciprocal of the range or standard deviation of the metallographic structure grade test values ​​in different areas of the same workpiece. The energy efficiency is scored as the reciprocal of the average actual energy consumption per unit of qualified output. The operability of the process is determined by checking whether the type of cooling medium in the candidate recommended process scheme is in the list of media available in the current production environment and whether the heating temperature does not exceed the temperature limit of the current equipment.

6. The heat treatment process optimization system based on multi-condition data evaluation according to claim 5, characterized in that, The multi-objective optimization search adopts a method that combines multi-objective weighted summation with constraint satisfaction. Performance objectives that are directly related to input features with higher weights in the adaptive weight vector are given higher optimization priority, while performance objectives that are related to input features with lower weights are given lower optimization priority.

7. The heat treatment process optimization system based on multi-condition data evaluation according to claim 1, characterized in that, The process of inputting the candidate recommended process scheme into the digital twin simulation model of the heat treatment process for virtual verification includes: The digital twin simulation model is based on the thermal properties of the workpiece material and uses the finite element method to simulate the temperature field evolution, microstructure transformation process and stress-strain distribution of the workpiece under the conditions of the candidate recommended process scheme. The performance indicators predicted by the simulation are compared with the target performance one by one. If all indicators meet the standards, the candidate recommended process scheme is verified. If there are indicators that do not meet the standards, a Bayesian optimization method is used to perform a directional search in the process parameter space based on the direction and magnitude of the deviation between the simulation prediction value and the target value to generate a corrected combination of process parameters. The corrected combination of process parameters is then input into the digital twin simulation model for verification. The correction and verification process is repeated until all simulation prediction results meet the standards or the preset maximum number of iterations is reached.

8. The heat treatment process optimization system based on multi-condition data evaluation according to claim 7, characterized in that, When performing the directional search, the deviation amplitude is used as the basis for adjusting the search step size. When the deviation amplitude is large, a larger search step size is used, and when the deviation amplitude is small, a smaller search step size is used.

9. The heat treatment process optimization system based on multi-condition data evaluation according to claim 1, characterized in that, Also includes: When a condition drift signal, such as equipment status fluctuation or ambient temperature change, is detected, the matching relationship between the current condition feature vector and the condition label library is recalculated, and the online fine-tuning of the adaptive weight vector is automatically triggered.

Images