An alloy process parameter optimization method and system based on transfer learning

Abstract

The application discloses an alloy process parameter optimization method and system based on transfer learning, belongs to the technical field of industrial intelligent control, and is used for solving the technical problems that current alloy process parameter optimization methods cannot adapt to new specification cold start, black box constraint processing capacity is limited, and single round batch output differentiated candidate process parameter points cannot be supported. The method comprises the following steps: training a historical prior model through a historical data set; completing the first round optimization of the current parameter optimization task according to the historical prior model; training a task adaptive model according to a measured performance data set obtained after the first round optimization; performing knowledge fusion on the historical prior model and the task adaptive model according to a self-defined historical confidence weight function; constructing a collection function with different search preferences in a process parameter search space and performing parameter optimization to obtain multiple groups of candidate process parameters and acquire measured performance data, retraining the task adaptive model until a preset convergence condition is met, and outputting optimal process parameters.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligent control technology, and in particular to a method and system for optimizing alloy process parameters based on transfer learning. Background Technology

[0002] In the manufacturing and heat treatment of special alloys such as high-strength steel and special structural components, it is usually necessary to jointly optimize multiple process parameters to obtain the best mechanical properties. In actual production, in addition to maximizing core performance indicators, it is also necessary to meet strict design baseline requirements, such as yield strength not being lower than the set standard and elongation after fracture not being lower than the minimum design requirement.

[0003] Currently, the industry commonly uses Bayesian optimization methods based on Gaussian process regression for black-box optimization of process parameters. However, existing solutions generally have the following shortcomings:

[0004] First, traditional Bayesian optimization methods suffer from a significant cold-start problem when dealing with new workpieces with changing feature dimensions. For example, when key feature dimensions such as workpiece thickness change, existing process experience is difficult to utilize effectively. If only a small number of random test points are used for initial exploration, the cost of blind trial and error in the early stages is high due to the high cost of sample processing and the reliance on third-party destructive testing.

[0005] Second, the output results of existing batch Bayesian optimization methods are often highly homogeneous. Although batch Bayesian optimization can output multiple candidate test points in a single round of testing to shorten the overall testing cycle, its candidate parameter points are often concentrated in local similar areas and lack sufficient differences, resulting in limited information gain for a single batch of testing.

[0006] Third, existing optimization methods have limited ability to handle black-box constraints. In actual process parameter optimization, auxiliary indicators such as yield strength and elongation often cannot be directly represented by analytical expressions, and their relationship with process parameters can often only be known after experiments. If the optimization algorithm only focuses on improving tensile strength, it may output invalid parameters with high tensile strength but insufficient yield strength or elongation, causing the entire long-cycle experiment to fail and seriously wasting R&D resources.

[0007] Therefore, there is currently a lack of an alloy process parameter optimization method that can adapt to cold starts of new specifications, take into account black box constraints, and support single-round batch output of differentiated candidate process parameter points. Summary of the Invention

[0008] This invention provides a method and system for optimizing alloy process parameters based on transfer learning, which addresses the following technical problems: current alloy process parameter optimization methods cannot adapt to cold starts of new specifications, have limited black-box constraint processing capabilities, cannot support single-round batch output of differentiated candidate process parameter points, and do not meet current alloy production needs.

[0009] The embodiments of the present invention adopt the following technical solutions:

[0010] On one hand, embodiments of the present invention provide a method for optimizing alloy process parameters based on transfer learning. The method includes: constructing a historical dataset based on historical process parameters and alloy performance data, and training a first Gaussian process regression model and a first black-box constraint model with augmented input parameters as input.

[0011] Based on the first Gaussian process regression model and the first black-box constraint model, complete the first round of optimization for the current parameter optimization task;

[0012] Based on the measured performance dataset obtained from alloy processing samples after the first round of optimization, a second Gaussian process regression model and a second black-box constraint model with process parameters as input are trained.

[0013] Based on the custom historical confidence weight function, the first Gaussian process regression model and the second Gaussian process regression model are combined to obtain a joint Gaussian process regression model; the first black-box constraint model and the second black-box constraint model are combined to obtain a joint black-box constraint model.

[0014] Based on the joint black-box constraint model, the joint feasibility probability of the process parameters satisfying each constraint is calculated, and the constraint penalty factor is obtained.

[0015] Within the process parameter search space, multiple acquisition functions with different search preferences are constructed based on the constraint penalty factor, the joint Gaussian process regression model, and the joint black-box constraint model. Parameter optimization is performed on each acquisition function to obtain multiple sets of candidate process parameters.

[0016] Based on the multiple sets of candidate process parameters, alloy processing samples are prepared and measured performance data is obtained. The measured performance data is then added to the measured performance dataset. The second Gaussian process regression model and the second black-box constraint model are retrained until the preset convergence condition is met, and the optimal process parameters are output.

[0017] In one feasible implementation, before constructing a historical dataset based on historical process parameters and alloy property data, the method further includes:

[0018] Define the feature dimensions of the workpiece to be processed as context variables; define the set of process parameters according to the type of process parameters to be optimized;

[0019] Based on the context variables and the set of process parameters, augmented input parameters are constructed;

[0020] Based on maximizing the tensile strength of the workpiece to be processed, an objective function and black-box constraints are defined; wherein, the black-box constraints include yield strength constraints and elongation after fracture constraints.

[0021] Based on the boundary values ​​of each process parameter in the set of process parameters, the process parameter search space is constructed, and the basic variables are defined.

[0022] In one feasible implementation, a historical dataset is constructed based on historical process parameters and alloy property data, and a first Gaussian process regression model and a first black-box constraint model are trained using augmented input parameters as input, specifically including:

[0023] Collect historical process parameters and corresponding alloy property data corresponding to different feature dimensions of the workpiece to be processed, and construct the historical dataset.

[0024] Using the augmented input parameters as model input data, and the objective function and black-box constraints as training constraints, a first Gaussian process regression model and a first black-box constraint model are constructed, and the models are trained using the historical dataset.

[0025] In one feasible implementation, the first round of optimization for the current parameter optimization task is completed based on the first Gaussian process regression model and the first black-box constraint model, specifically including:

[0026] Obtain the workpiece dimensions to be optimized corresponding to the current parameter optimization task;

[0027] The context variable is set as the size of the workpiece to be optimized, and substituted into the first Gaussian process regression model to obtain the historical prior prediction mean and historical prior prediction standard deviation, which are used as the first round prediction mean and first round prediction standard deviation; at the same time, they are substituted into the first black box constraint model to calculate the joint feasibility probability and obtain the first round constraint penalty factor.

[0028] Substitute the first-round prediction mean, the first-round prediction standard deviation, and the first-round constraint penalty factor into the acquisition function to perform parameter optimization and obtain the first-round optimized process parameters.

[0029] Alloy processing samples were prepared based on the first round of optimized process parameters, and measured performance data were obtained to obtain the measured performance dataset after the first round of optimization.

[0030] In one feasible implementation, based on the measured performance dataset obtained from alloy processing samples after the first round of optimization, a second Gaussian process regression model and a second black-box constraint model are trained, with process parameters as input. Specifically, this includes:

[0031] Based on the first round of process parameters obtained after the first round of optimization, alloy processing samples were prepared and measured performance datasets were obtained.

[0032] A second Gaussian process regression model and a second black-box constraint model are constructed with only process parameters as input, and the models are trained using the first round of process parameters and the measured performance dataset.

[0033] In one feasible implementation, the first Gaussian process regression model and the second Gaussian process regression model are combined according to a custom historical confidence weight function to obtain a joint Gaussian process regression model; the first black-box constraint model and the second black-box constraint model are combined to obtain a joint black-box constraint model, specifically including:

[0034] Based on the number of trials n and the preset decay factor Define the historical confidence weight function. Among them, historical confidence weight As the number of trials n increases, it gradually decreases;

[0035] Based on the historical confidence weight function, the first Gaussian process regression model and the second Gaussian process regression model are weighted and fused to obtain the joint Gaussian process regression model; wherein, the joint Gaussian process regression model is used to calculate the joint predicted mean and joint standard deviation of the objective function;

[0036] Based on the historical confidence weight function, the first black-box constraint model and the second black-box constraint model are weighted and fused to obtain the joint black-box constraint model; wherein, the joint black-box constraint model is used to calculate the joint prediction mean and joint prediction standard deviation of the black-box constraint function.

[0037] In one feasible implementation, based on the joint black-box constraint model, the joint feasibility probability of process parameters satisfying each constraint is calculated to obtain the constraint penalty factor, specifically including:

[0038] Based on the joint prediction mean and joint prediction standard deviation of the black box constraint function obtained by the joint black box constraint model, the first feasibility probability of the process parameters of the current round satisfying the yield strength constraint condition and the second feasibility probability of satisfying the elongation after fracture constraint condition are calculated respectively.

[0039] The first feasibility probability is multiplied by the second feasibility probability to obtain the joint feasibility probability, and the joint feasibility probability is used as the constraint penalty factor of the acquisition function in the subsequent optimization process.

[0040] In one feasible implementation, within the process parameter search space, multiple acquisition functions with different search preferences are constructed based on the constraint penalty factor, the joint Gaussian process regression model, and the joint black-box constraint model. Parameter optimization is then performed on each acquisition function to obtain multiple sets of candidate process parameters, specifically including:

[0041] Within the process parameter search space, based on the constraint penalty factor, the joint Gaussian process regression model, and the joint black-box constraint model, at least three types of acquisition functions with search preferences—robust, compromise, and exploratory—are constructed.

[0042] The differential evolution algorithm is used to perform a global search for each acquisition function, and the quasi-Newton method is used for local refinement to obtain the optimal process parameters for each acquisition function in the process parameter search space, resulting in at least three sets of candidate process parameters.

[0043] In one feasible implementation, alloy processing samples are prepared based on the multiple sets of candidate process parameters, and measured performance data is obtained. The measured performance data is then added to the measured performance dataset. The second Gaussian process regression model and the second black-box constraint model are retrained until a preset convergence condition is met, and the optimal process parameters are output. Specifically, this includes:

[0044] Based on the aforementioned multiple candidate processes, single-batch processing is performed to obtain various alloy processing samples, and performance testing experiments are conducted to obtain the measured performance data; wherein, the measured performance data includes at least tensile strength, yield strength, and elongation after fracture;

[0045] The measured performance data is added to the measured performance dataset, and the number of trials n is updated. The second Gaussian process regression model and the second black-box constraint model are retrained. The subsequent steps are executed iteratively until the preset convergence condition is met, and the multiple sets of process parameters output at this time are obtained as the optimal process parameters. The preset convergence condition is: the measured performance data reaches the preset target value, or the improvement value of the measured performance data after a preset number of consecutive iterations is lower than the preset threshold.

[0046] On the other hand, embodiments of the present invention also provide an alloy process parameter optimization system based on transfer learning, the system comprising:

[0047] The historical prior model pre-training module is used to construct a historical dataset based on historical process parameters and alloy performance data, and to train the first Gaussian process regression model and the first black-box constraint model with augmented input parameters as input; based on the first Gaussian process regression model and the first black-box constraint model, the first round of optimization for the current parameter optimization task is completed;

[0048] The task adaptive model training module is used to train the second Gaussian process regression model and the second black-box constraint model with process parameters as input, based on the measured performance dataset obtained from the alloy processing samples after the first round of optimization.

[0049] The model fusion module is used to combine the first Gaussian process regression model and the second Gaussian process regression model according to a custom historical confidence weight function to obtain a joint Gaussian process regression model; and to combine the first black-box constraint model and the second black-box constraint model to obtain a joint black-box constraint model.

[0050] The black-box constraint processing module is used to calculate the joint feasibility probability of process parameters satisfying each constraint based on the joint black-box constraint model, and obtain the constraint penalty factor.

[0051] The process parameter Bayesian optimization module is used to construct multiple acquisition functions with different search preferences within the process parameter search space, based on the constraint penalty factor, the joint Gaussian process regression model, and the joint black-box constraint model. It then optimizes the parameters of each acquisition function to obtain multiple sets of candidate process parameters. Based on these candidate process parameters, alloy processing samples are prepared and measured performance data is obtained. This measured performance data is then added to the measured performance dataset. The second Gaussian process regression model and the second black-box constraint model are retrained until a preset convergence condition is met, at which point the optimal process parameters are output.

[0052] Compared with the prior art, the alloy process parameter optimization method and system based on transfer learning provided in this invention has the following beneficial effects:

[0053] (1) This invention constructs a historical global model containing context variables and fixes the context variables to the current size when facing a new size task, thereby realizing the transfer of historical multi-specification process knowledge to the current task. Compared with the scheme that relies entirely on random initial experiments, this method can provide more targeted parameter prior guidance in the early stage and can reduce the cold start cost of optimizing process parameters for new specification workpieces.

[0054] (2) The present invention adopts historical confidence weights that decay dynamically with the number of trials, and integrates the historical context prior model with the current task adaptive model, so that the optimization process gradually transitions from history-driven to current task-driven, thereby improving the adaptability of the method when there are differences in actual working conditions, and can reduce the negative migration effect while utilizing historical experience.

[0055] (3) This invention establishes a black-box constraint model for constraint indicators such as yield strength and elongation, and embeds the joint feasibility probability into the acquisition function, so that the optimization results not only focus on the improvement of the target strength, but also take into account multiple performance bottom line requirements, thereby reducing the probability of outputting invalid parameters and effectively handling black-box physical constraints.

[0056] (4) This invention designs three types of heterogeneous acquisition functions: robust, compromise, and exploratory. Numerical optimization is then performed on each function to obtain three sets of candidate process parameter points with different physical exploration preferences. Compared with traditional homogeneous batch optimization methods, this method is more conducive to simultaneously acquiring conservative, safe, potential, and exploratory information in a single batch experiment, thereby improving the differentiation of single-round batch optimization. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0058] Figure 1 A flowchart of an alloy process parameter optimization method based on transfer learning is provided in an embodiment of the present invention.

[0059] Figure 2 A flowchart illustrating a specific implementation of an alloy process parameter optimization method based on transfer learning, provided in this embodiment of the invention.

[0060] Figure 3 A flowchart of dynamic knowledge fusion of a model is provided in an embodiment of the present invention;

[0061] Figure 4 A schematic diagram of batch Bayesian optimization logic provided in an embodiment of the present invention;

[0062] Figure 5 This is a schematic diagram of an alloy process parameter optimization system based on transfer learning, provided in an embodiment of the present invention. Detailed Implementation

[0063] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0064] This invention provides a method for optimizing alloy process parameters based on transfer learning, such as... Figure 1 As shown, the alloy process parameter optimization method based on transfer learning specifically includes steps S101-S106:

[0065] S101. Based on historical process parameters and alloy performance data, construct a historical dataset and train the first Gaussian process regression model and the first black-box constraint model with augmented input parameters as input; based on the first Gaussian process regression model and the first black-box constraint model, complete the first round of optimization for the current parameter optimization task.

[0066] Specifically, the basic variables required for the optimization process are first defined, including:

[0067] Define the feature dimensions of the workpiece to be processed as context variables; then define a set of process parameters based on the type of process parameters to be optimized. Construct augmented input parameters based on the context variables and the set of process parameters.

[0068] Furthermore, based on maximizing the tensile strength of the workpiece to be processed, an objective function and black-box constraints are defined; these black-box constraints include yield strength constraints and elongation after fracture constraints. Then, based on the boundary values ​​of each process parameter in the process parameter set, a process parameter search space is constructed, completing the definition of the basic variables.

[0069] In one embodiment, the feature size of the workpiece is defined as a context variable. For example, thickness.

[0070] Define the set of process parameters as follows: ;in, , , These represent the process parameters to be optimized, such as heating temperature, holding time, and cooling rate.

[0071] Construct augmented input as: The objective function is defined as maximizing tensile strength. Define the black-box constraints as follows: , ;in, Indicates yield strength. Indicates elongation after fracture. , These are the preset performance baselines.

[0072] Define the process parameter search space as: ;in, and The first The lower and upper bounds of each process parameter are defined. Subsequent target search, constraint evaluation, and acquisition function optimization are all performed within this bounded search space. It will be carried out internally.

[0073] Furthermore, historical process parameters and corresponding alloy property data corresponding to different feature dimensions of the workpiece to be processed are collected to construct a historical dataset. Using the augmented input parameters as model input data, and the objective function and black-box constraints as training constraints, a first Gaussian process regression model and a first black-box constraint model are constructed, and the models are trained using the historical dataset.

[0074] Further, the workpiece size to be optimized corresponding to the current parameter optimization task is obtained. The context variable is set to the workpiece size to be optimized and substituted into the first Gaussian process regression model to obtain the historical prior prediction mean and standard deviation, which are used as the first-round prediction mean and standard deviation. Simultaneously, these are substituted into the first black-box constraint model to calculate the joint feasibility probability, obtaining the first-round constraint penalty factor. The first-round prediction mean, first-round prediction standard deviation, and first-round constraint penalty factor are substituted into the acquisition function for parameter optimization, obtaining the first-round optimized process parameters. Alloy processing samples are prepared based on the first-round optimized process parameters, and measured performance data is obtained, resulting in the measured performance dataset after the first round of optimization.

[0075] As a feasible implementation method, Figure 2 A flowchart illustrating a specific implementation method for optimizing alloy process parameters based on transfer learning, as provided in this embodiment of the invention, is shown below. Figure 2 As shown, different feature sizes in history were collected in advance. Based on the corresponding process parameters and alloy property data, a historical dataset is constructed. Then, augmented input is used... As input, for the objective function and black-box constraint functions Independently train target Gaussian process regression models that include context variables. and constrained Gaussian process model The first Gaussian process regression model and the first black-box constraint model are obtained. The first Gaussian process regression model is used to learn the cross-size shared mapping relationship between process parameters and performance indicators under different size conditions. The initial target Gaussian process regression model and the constrained Gaussian process model are both existing Gaussian optimization models, and the existing model structures will not be elaborated upon in this invention.

[0076] When facing the required new workpiece size When performing process parameter optimization tasks, the context variable h is fixed as follows: Thus, the historical prior prediction function for the current size task is obtained: ;in, This indicates that the first Gaussian process regression model is based on the input... The predicted mean at that location. Similarly, the predicted mean at that location. Inputting the first black-box constraint model yields the constraint prediction mean. .

[0077] like Figure 2 As shown, in the first iteration, since there is no actual performance data for the current task, the number of trials n=0. At this time, the historical prior prediction mean is directly calculated using the first Gaussian process regression model based on historical priors, according to the process parameters of the historical dimensions. Then, the dynamic weights of the historical confidence are calculated. At this point, the historical prior prediction mean is directly used as the joint prediction mean. The parameters are then optimized using three types of constrained acquisition functions to calculate the joint feasibility probability, resulting in the first round of optimized process parameters. Alloy processing samples are then tested to obtain the first round of measured performance data. This step enables the transfer of historical multi-size process knowledge to the current task of optimizing process parameters for new sizes. This allows for prior evaluation of candidate process parameters based on historical experience, even when there are very few or no samples available for the current task. After obtaining the first batch of experimental results, an adaptive model for the current task is then constructed based on these results.

[0078] S102. Based on the measured performance dataset obtained from the alloy processing samples after the first round of optimization, train the second Gaussian process regression model and the second black-box constraint model with process parameters as input.

[0079] Specifically, alloy processing samples were prepared based on the first-round process parameters obtained after the first round of optimization, and measured performance datasets were acquired. A second Gaussian process regression model and a second black-box constraint model were constructed, using only process parameters as input, and the models were trained using the first-round process parameters and the measured performance datasets.

[0080] As a possible implementation method, such as Figure 2 As shown, the required size for the current task After obtaining the first round of measured performance data, the number of trials is n=1, at which point the condition of n>0 is met. Then, for the objective function... and constraint functions , Training a second Gaussian process regression model and a second black-box constraint model with only process parameter x as input, i.e. Figure 2 The text discusses the current task adaptive model and constraint model. It states that the second Gaussian process regression model can output the target prediction mean under any process parameter x. and the predicted standard deviation The second black-box constraint model can output the predicted mean and predicted standard deviation corresponding to the two black-box constraints: ; .

[0081] S103. Based on the custom historical confidence weight function, combine the first Gaussian process regression model and the second Gaussian process regression model to obtain a joint Gaussian process regression model; combine the first black-box constraint model and the second black-box constraint model to obtain a joint black-box constraint model.

[0082] Specifically, based on the number of trials n and the preset decay factor Define the historical confidence weight function Among them, historical confidence weight The number of trials, n, gradually decreases.

[0083] Furthermore, based on the historical confidence weight function, the first Gaussian process regression model and the second Gaussian process regression model are weighted and fused to obtain a joint Gaussian process regression model; wherein, the joint Gaussian process regression model is used to calculate the joint predicted mean and joint standard deviation of the objective function. Based on the historical confidence weight function, the first black-box constraint model and the second black-box constraint model are weighted and fused to obtain a joint black-box constraint model; wherein, the joint black-box constraint model is used to calculate the joint predicted mean and joint predicted standard deviation of the black-box constraint function.

[0084] As a feasible implementation method, Figure 3 A flowchart of dynamic knowledge fusion of a model is provided as an embodiment of the present invention, such as Figure 3 As shown, in order to fully utilize historical experience in the early stages of process parameter optimization, solve the zero-sample cold start problem, and prevent high standard deviations due to data scarcity in the early stages of the new model and negative historical transfer in the later stages, a method based on the number of trials is adopted. The dynamic confidence fusion mechanism.

[0085] First, define the historical confidence weight function as follows: ;in, This refers to the attenuation factor. In practical special alloy process optimization engineering, the attenuation factor... The preferred range is This ensures that historical knowledge plays a dominant role in the initial 3 to 5 rounds of iterative trials, and then smoothly transitions to being dominated by measured data and the current model.

[0086] when When smaller, A large number of trials indicates that there is currently limited measured performance data, and the optimization process relies more on prior knowledge of historical context; as the number of trials increases, As the dependency on the adaptive model for the current task gradually decreases, the optimization process becomes increasingly reliant on it. This approach enables a smooth transition from historical knowledge to current task knowledge, reducing the risk of negative transfer caused by differences in historical operating conditions.

[0087] Then, for the objective function and black-box constraint functions , The predicted mean of the output of the first Gaussian process regression model The predicted mean of the output of the second Gaussian process regression model We perform a weighted calculation to obtain the joint prediction mean of this optimization iteration. with joint forecast standard deviation .

[0088] With objective function For example, the fusion formula is as follows:

[0089] Joint forecast mean: ;

[0090] Joint forecast standard deviation: .

[0091] Similarly, the joint mean of the black-box constraint functions can be calculated using the same weighting method. and joint standard deviation .

[0092] As mentioned earlier, during the first round of optimization, i.e., the number of trials... At that time, the historical confidence weight was adjusted. At this point, the joint forecast mean is the first-round forecast mean calculated by the first Gaussian process regression model, i.e. The same applies to other predicted values. That is, the first round of optimization is entirely dominated by the historical prior model.

[0093] S104. Based on the joint black-box constraint model, calculate the joint feasibility probability of the process parameters satisfying each constraint, and obtain the constraint penalty factor.

[0094] Specifically, based on the joint prediction mean and joint prediction standard deviation of the black-box constraint function obtained by the joint black-box constraint model, the first feasibility probability of the process parameters in the current round satisfying the yield strength constraint condition and the second feasibility probability of satisfying the elongation after fracture constraint condition are calculated respectively.

[0095] Furthermore, the first feasibility probability is multiplied by the second feasibility probability to obtain the joint feasibility probability, which is then used as the constraint penalty factor for the acquisition function in the subsequent optimization process.

[0096] As a feasible implementation method, to avoid outputting invalid process parameters that do not meet the yield strength or elongation threshold, the joint prediction mean and joint prediction standard deviation of the obtained constraint functions are used to calculate arbitrary process parameters respectively. The probability of satisfying all constraints.

[0097] For yield strength constraint conditions Its first feasibility probability is: Under the assumption of a Gaussian predictive distribution, the above equation can be written as: Where y is the yield strength constraint condition. .

[0098] Similarly, for the post-fracture elongation constraint condition Its second feasibility probability is: .Right now: Where y is the elongation constraint after fracture. .

[0099] In a preferred embodiment, assuming the two constraints are independent, the joint feasibility probability is: Joint feasibility probability It is used to characterize the possibility of candidate process parameters simultaneously meeting multiple performance baseline requirements, and serves as a constraint penalty factor in subsequent acquisition functions.

[0100] S105. Within the process parameter search space, construct multiple acquisition functions with different search preferences based on constraint penalty factors, joint Gaussian process regression models, and joint black-box constraint models, and optimize the parameters of each acquisition function to obtain multiple sets of candidate process parameters.

[0101] Specifically, within the process parameter search space, based on constraint penalty factors, joint Gaussian process regression models, and joint black-box constraint models, at least three types of acquisition functions with search preferences—robust, compromise, and exploratory—are constructed.

[0102] Furthermore, a global search is performed on each acquisition function using a differential evolution algorithm, and local refinement is performed using a quasi-Newton method to obtain the optimal process parameters for each acquisition function in the process parameter search space, resulting in at least three sets of candidate process parameters.

[0103] As a feasible implementation method, Figure 4 A schematic diagram of batch Bayesian optimization logic provided in an embodiment of the present invention, such as... Figure 4 As shown, in order to shorten the single-round test cycle and improve the information acquisition efficiency of a single batch test, this invention expands the process parameter search space. Within the framework, three constrained acquisition functions representing different search preferences were constructed, and numerical optimization was performed on each acquisition function to output three sets of candidate process parameter points with differentiated characteristics at once.

[0104] For each acquisition function optimization, the preferred approach is to use the differential evolution algorithm for global search, followed by a quasi-Newton method for local refinement, to obtain the corresponding acquisition function within the bounded search space. The optimal parameter point within.

[0105] (1) For robust candidate process parameter points Search strategy: Robust candidate process parameter points focus on finding conservative regions with high target prediction mean and low uncertainty. Their constrained acquisition function is constructed as follows: ;in, This is a robustness adjustment coefficient. In a preferred embodiment, The range of values ​​is set to This is used to control the penalty applied to the prediction variance, prompting the model to select points within a known high-confidence safety region.

[0106] Optimizing the acquisition function yields robust candidate points: Robust candidate process parameters are suitable for process scenarios that place greater emphasis on parameter safety and stability in achieving targets.

[0107] (2) For compromise candidate process parameter points Search strategy:

[0108] Compromise-type candidate process parameters focus on finding the potential region of merit that yields the greatest expected improvement. The core idea is to evaluate the potential improvement that candidate points could bring to the target based on the currently known best feasible outcome.

[0109] Let the currently observed optimal feasible objective value be... In a preferred embodiment, the improvement relative to the optimal feasible target value is defined as: ;in, To improve the margin parameters, in a preferred embodiment, The value of is set to 1% to 5% of the expected range of the objective function, in order to avoid the model getting trapped in local optima of small improvements.

[0110] The corresponding expected improvement function is: ;in, The cumulative distribution function represents the standard normal distribution. This represents the probability density function of the standard normal distribution.

[0111] Further construct a constrained compromise acquisition function: By optimizing the acquisition function, compromise candidate points are obtained: Compromise-type candidate process parameters are used to balance the potential for improvement in objectives with the probability of satisfying constraints.

[0112] (3) For exploratory candidate process parameters Search strategy:

[0113] Exploratory candidate process parameters focus on finding high-uncertainty regions where current models have insufficient understanding but potential benefits.

[0114] Its acquisition function is constructed as follows: ;in, This is an exploratory adjustment coefficient. In a preferred embodiment, The range of values ​​is set to And usually set This encourages models to break out of their existing cognitive domains and boldly experiment with unknown boundaries where the standard deviation is large.

[0115] Optimize the acquisition function to obtain exploratory candidate points: Exploratory candidate process parameters are used to proactively enter regions of high uncertainty in the model, thereby enhancing the model's overall understanding of the parameter space.

[0116] S106. Based on multiple sets of candidate process parameters, alloy processing samples are prepared and measured performance data is obtained. The measured performance data is added to the measured performance dataset. The second Gaussian process regression model and the second black box constraint model are retrained until the preset convergence condition is met, and the optimal process parameters are output.

[0117] Specifically, single-batch processing is carried out based on multiple candidate processes to obtain various alloy processing samples, and performance testing is conducted to obtain measured performance data; among them, the measured performance data includes at least tensile strength, yield strength and elongation after fracture.

[0118] Furthermore, the measured performance data is added to the measured performance dataset, and the number of trials n is updated. The second Gaussian process regression model and the second black-box constraint model are retrained. The subsequent steps are executed iteratively until the preset convergence condition is met, and then multiple sets of process parameters are output as the optimal process parameters. The preset convergence condition is: the measured performance data reaches the preset target value, or the improvement value of the measured performance data after a preset number of consecutive iterations is lower than the preset threshold.

[0119] As a feasible implementation method, by optimizing the above three types of acquisition functions respectively, three sets of candidate process parameter points with different search preferences are obtained simultaneously in a single iteration: These three approaches correspond to robust, compromise, and exploratory optimization strategies, respectively, thereby avoiding the problem of multiple candidate process parameter points being overly concentrated in similar regions in traditional batch Bayesian optimization and improving the information diversity and effectiveness of single-batch experiments.

[0120] Based on the three sets of candidate process parameters mentioned above, three samples were processed in a single batch and sent for testing. After obtaining the actual tensile strength, yield strength, and elongation data, the new data was added to the measured performance dataset for the current task, and the number of tests was updated. Retrain the second Gaussian process regression model and the second black-box constraint model, and repeat the above steps.

[0121] The above closed-loop process continues until one of the following conditions is met:

[0122] 1. Target performance indicators such as tensile strength meet the preset convergence conditions;

[0123] 2. The optimal performance improvement becomes insignificant after several consecutive iterations.

[0124] After the stopping conditions are met, the final optimal combination of process parameters is output, completing the current optimization task.

[0125] In one embodiment, to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to a specific optimized embodiment of a high-strength steel heat treatment process.

[0126] In the manufacturing of a certain special high-strength steel, the goal is to maximize its tensile strength while simultaneously satisfying two stringent physical baselines (black box constraints): yield strength. Elongation after fracture The process parameter space to be optimized is defined as: heating temperature. Insulation time Cooling rate time The thickness feature dimension of the new batch of workpieces to be optimized is currently: .

[0127] Historical Prior Acquisition and Cold Start: A historical database contains heat treatment data from multiple batches of this high-strength steel with thicknesses of 20mm, 40mm, and 50mm. This data is then used to construct a system with augmented inputs. A global context objective Gaussian process prior model is used, and the current context variable is fixed as... At this point, the current new size workpiece ( No samples have been processed yet (number of tests). Traditional methods face the challenge of cold start. This invention can learn the cross-size shared mapping relationship between process parameters and performance indicators under different size conditions, and directly output the historical prior prediction mean and prediction standard deviation.

[0128] Parameter configuration and dynamic fusion: Setting the historical attenuation factor .exist At that time, historical weight The first round of optimization was entirely based on historical mapping experience. The probability of the high-strength steel sample meeting the combined feasibility requirement of a yield strength greater than 1000 MPa and an elongation greater than 12% was calculated. This probability will serve as a hard penalty, filtering out invalid process parameters that have extremely high tensile strength but do not meet the yield strength and elongation requirements.

[0129] Three-state heterogeneous batch optimization and closed-loop experiment: setting robustness adjustment coefficients Improve margin Exploratory adjustment coefficient =2. In a single iteration, differential evolution optimization is performed simultaneously on three penalty-based acquisition functions: robust candidate points are derived as follows: (The expected tensile strength is moderate, but the probability of satisfying the constraints is extremely high); the compromise candidate points are calculated as follows: (Expected tensile strength is significantly improved); Exploratory candidate points show that: (Historical data blind spots, extremely large expected variance, possibly high potential particles).

[0130] The above three sets of parameters were sampled and tested in single batches and centrally. After obtaining accurate mechanical test reports, the data were added to the database and the number of tests was updated. At this point, the historical confidence weight decays to The adaptive model began to gradually take effect.

[0131] The above closed loop is continuously repeated. Real-world engineering verification shows that if traditional single-point migration-free Bayesian optimization is used (processing one sample at a time, exploring from scratch), due to a large number of early samples getting trapped in invalid regions that do not meet physical constraints, it takes an average of 20 rounds (more than 10 days) to find the optimal process parameters that meet the bottom-line constraints.

[0132] The scheme employing the combination of transfer learning and tri-state heterogeneous batch optimization of this invention, due to the initial historical prior guidance and the output of three complementary points that balance robustness, compromise, and exploration in a single iteration, only requires 5 iterations (processing 15 samples in total, taking approximately 3 days) to reach the convergent extreme value of tensile strength. This shortens the development cycle by 70% and, due to the combined probabilistic feasibility... The introduction of this technology has greatly reduced the scrap rate of samples due to unqualified yield and elongation.

[0133] In addition, embodiments of the present invention also provide an alloy process parameter optimization system based on transfer learning, such as... Figure 5 As shown, the alloy process parameter optimization system 500 based on transfer learning specifically includes:

[0134] The historical prior model pre-training module 510 is used to construct a historical dataset based on historical process parameters and alloy performance data, and to train the first Gaussian process regression model and the first black-box constraint model with augmented input parameters as input; based on the first Gaussian process regression model and the first black-box constraint model, the first round of optimization of the current parameter optimization task is completed;

[0135] The task adaptive model training module 520 is used to train the second Gaussian process regression model and the second black-box constraint model with process parameters as input based on the measured performance dataset obtained from the alloy processing samples after the first round of optimization.

[0136] The model fusion module 530 is used to combine the first Gaussian process regression model and the second Gaussian process regression model according to a custom historical confidence weight function to obtain a joint Gaussian process regression model; and to combine the first black-box constraint model and the second black-box constraint model to obtain a joint black-box constraint model.

[0137] The black-box constraint processing module 540 is used to calculate the joint feasibility probability of process parameters satisfying each constraint according to the joint black-box constraint model, and obtain the constraint penalty factor.

[0138] The process parameter Bayesian optimization module 550 is used to construct multiple acquisition functions with different search preferences within the process parameter search space based on the constraint penalty factor, the joint Gaussian process regression model, and the joint black-box constraint model, and to optimize the parameters of each acquisition function to obtain multiple sets of candidate process parameters; based on the multiple sets of candidate process parameters, alloy processing samples are processed and measured performance data is obtained, the measured performance data is added to the measured performance dataset, and the second Gaussian process regression model and the second black-box constraint model are retrained until the preset convergence condition is met, and the optimal process parameters are output.

[0139] The various embodiments in this invention are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0140] The foregoing has described specific embodiments of the present invention. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0141] The above description is merely an embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations can be made to the embodiments of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the embodiments of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for optimizing alloy process parameters based on transfer learning, characterized in that, The method includes: Based on historical process parameters and alloy property data, a historical dataset was constructed, and a first Gaussian process regression model and a first black-box constraint model were trained with augmented input parameters as input. Based on the first Gaussian process regression model and the first black-box constraint model, complete the first round of optimization for the current parameter optimization task; Based on the measured performance dataset obtained from alloy processing samples after the first round of optimization, a second Gaussian process regression model and a second black-box constraint model with process parameters as input are trained. Based on the custom historical confidence weight function, the first Gaussian process regression model and the second Gaussian process regression model are combined to obtain a joint Gaussian process regression model; the first black-box constraint model and the second black-box constraint model are combined to obtain a joint black-box constraint model. Based on the joint black-box constraint model, the joint feasibility probability of the process parameters satisfying each constraint is calculated, and the constraint penalty factor is obtained. Within the process parameter search space, multiple acquisition functions with different search preferences are constructed based on the constraint penalty factor, the joint Gaussian process regression model, and the joint black-box constraint model. Parameter optimization is performed on each acquisition function to obtain multiple sets of candidate process parameters. Based on the multiple sets of candidate process parameters, alloy processing samples are prepared and measured performance data is obtained. The measured performance data is then added to the measured performance dataset. The second Gaussian process regression model and the second black-box constraint model are retrained until the preset convergence condition is met, and the optimal process parameters are output.

2. The alloy process parameter optimization method based on transfer learning according to claim 1, characterized in that, Before constructing a historical dataset based on historical process parameters and alloy property data, the method further includes: Define the feature dimensions of the workpiece to be processed as context variables; define the set of process parameters according to the type of process parameters to be optimized; Based on the context variables and the set of process parameters, augmented input parameters are constructed; Based on maximizing the tensile strength of the workpiece to be processed, an objective function and black-box constraints are defined; wherein, the black-box constraints include yield strength constraints and elongation after fracture constraints. Based on the boundary values ​​of each process parameter in the set of process parameters, the process parameter search space is constructed, and the basic variables are defined.

3. The alloy process parameter optimization method based on transfer learning according to claim 2, characterized in that, Based on historical process parameters and alloy property data, a historical dataset was constructed, and a first Gaussian process regression model and a first black-box constraint model were trained using augmented input parameters. Specifically, this included: Collect historical process parameters and corresponding alloy property data corresponding to different feature dimensions of the workpiece to be processed, and construct the historical dataset. Using the augmented input parameters as model input data, and the objective function and black-box constraints as training constraints, a first Gaussian process regression model and a first black-box constraint model are constructed, and the models are trained using the historical dataset.

4. The alloy process parameter optimization method based on transfer learning according to claim 3, characterized in that, Based on the first Gaussian process regression model and the first black-box constraint model, the first round of optimization for the current parameter optimization task is completed, specifically including: Obtain the workpiece dimensions to be optimized corresponding to the current parameter optimization task; The context variable is set as the size of the workpiece to be optimized, and substituted into the first Gaussian process regression model to obtain the historical prior prediction mean and historical prior prediction standard deviation, which are used as the first round prediction mean and first round prediction standard deviation; at the same time, they are substituted into the first black box constraint model to calculate the joint feasibility probability and obtain the first round constraint penalty factor. Substitute the first-round prediction mean, the first-round prediction standard deviation, and the first-round constraint penalty factor into the acquisition function to perform parameter optimization and obtain the first-round optimized process parameters. Alloy processing samples were prepared based on the first round of optimized process parameters, and measured performance data were obtained to obtain the measured performance dataset after the first round of optimization.

5. The method for optimizing alloy process parameters based on transfer learning according to claim 1, characterized in that, Based on the measured performance dataset obtained from alloy processing samples after the first round of optimization, a second Gaussian process regression model and a second black-box constraint model were trained with process parameters as input. Specifically, this included: Based on the first round of process parameters obtained after the first round of optimization, alloy processing samples were prepared and measured performance datasets were obtained. A second Gaussian process regression model and a second black-box constraint model are constructed with only process parameters as input, and the models are trained using the first round of process parameters and the measured performance dataset.

6. The alloy process parameter optimization method based on transfer learning according to claim 1, characterized in that, Based on a custom historical confidence weight function, the first Gaussian process regression model and the second Gaussian process regression model are combined to obtain a joint Gaussian process regression model; the first black-box constraint model and the second black-box constraint model are combined to obtain a joint black-box constraint model, specifically including: Based on the number of trials n and the preset decay factor Define the historical confidence weight function. Among them, historical confidence weight As the number of trials n increases, it gradually decreases; Based on the historical confidence weight function, the first Gaussian process regression model and the second Gaussian process regression model are weighted and fused to obtain the joint Gaussian process regression model; wherein, the joint Gaussian process regression model is used to calculate the joint predicted mean and joint standard deviation of the objective function; Based on the historical confidence weight function, the first black-box constraint model and the second black-box constraint model are weighted and fused to obtain the joint black-box constraint model; wherein, the joint black-box constraint model is used to calculate the joint prediction mean and joint prediction standard deviation of the black-box constraint function.

7. The alloy process parameter optimization method based on transfer learning according to claim 6, characterized in that, Based on the joint black-box constraint model, the joint feasibility probability of the process parameters satisfying each constraint is calculated to obtain the constraint penalty factor, which specifically includes: Based on the joint prediction mean and joint prediction standard deviation of the black box constraint function obtained by the joint black box constraint model, the first feasibility probability of the process parameters of the current round satisfying the yield strength constraint condition and the second feasibility probability of satisfying the elongation after fracture constraint condition are calculated respectively. The first feasibility probability is multiplied by the second feasibility probability to obtain the joint feasibility probability, and the joint feasibility probability is used as the constraint penalty factor of the acquisition function in the subsequent optimization process.

8. The alloy process parameter optimization method based on transfer learning according to claim 1, characterized in that, Within the process parameter search space, multiple acquisition functions with different search preferences are constructed based on the constraint penalty factor, the joint Gaussian process regression model, and the joint black-box constraint model. Parameter optimization is then performed on each acquisition function to obtain multiple sets of candidate process parameters, specifically including: Within the process parameter search space, based on the constraint penalty factor, the joint Gaussian process regression model, and the joint black-box constraint model, at least three types of acquisition functions with search preferences—robust, compromise, and exploratory—are constructed. The differential evolution algorithm is used to perform a global search for each acquisition function, and the quasi-Newton method is used for local refinement to obtain the optimal process parameters for each acquisition function in the process parameter search space, resulting in at least three sets of candidate process parameters.

9. The method for optimizing alloy process parameters based on transfer learning according to claim 1, characterized in that, Based on the multiple sets of candidate process parameters, alloy processing samples are prepared and measured performance data is obtained. The measured performance data is then added to the measured performance dataset. The second Gaussian process regression model and the second black-box constraint model are retrained until the preset convergence condition is met, and the optimal process parameters are output. Specifically, this includes: Based on the aforementioned multiple candidate processes, single-batch processing is performed to obtain various alloy processing samples, and performance testing experiments are conducted to obtain the measured performance data; wherein, the measured performance data includes at least tensile strength, yield strength, and elongation after fracture; The measured performance data is added to the measured performance dataset, and the number of trials n is updated. The second Gaussian process regression model and the second black-box constraint model are retrained. The subsequent steps are executed iteratively until the preset convergence condition is met, and the multiple sets of process parameters output at this time are obtained as the optimal process parameters. The preset convergence condition is: the measured performance data reaches the preset target value, or the improvement value of the measured performance data after a preset number of consecutive iterations is lower than the preset threshold.

10. A system for optimizing alloy process parameters based on transfer learning, characterized in that, The system includes: The historical prior model pre-training module is used to construct a historical dataset based on historical process parameters and alloy performance data, and to train the first Gaussian process regression model and the first black-box constraint model with augmented input parameters as input; based on the first Gaussian process regression model and the first black-box constraint model, the first round of optimization for the current parameter optimization task is completed; The task adaptive model training module is used to train the second Gaussian process regression model and the second black-box constraint model with process parameters as input, based on the measured performance dataset obtained from the alloy processing samples after the first round of optimization. The model fusion module is used to combine the first Gaussian process regression model and the second Gaussian process regression model according to a custom historical confidence weight function to obtain a joint Gaussian process regression model; and to combine the first black-box constraint model and the second black-box constraint model to obtain a joint black-box constraint model. The black-box constraint processing module is used to calculate the joint feasibility probability of process parameters satisfying each constraint based on the joint black-box constraint model, and obtain the constraint penalty factor. The process parameter Bayesian optimization module is used to construct multiple acquisition functions with different search preferences within the process parameter search space, based on the constraint penalty factor, the joint Gaussian process regression model, and the joint black-box constraint model. It then optimizes the parameters of each acquisition function to obtain multiple sets of candidate process parameters. Based on these candidate process parameters, alloy processing samples are prepared and measured performance data is obtained. This measured performance data is then added to the measured performance dataset. The second Gaussian process regression model and the second black-box constraint model are retrained until a preset convergence condition is met, at which point the optimal process parameters are output.

Images