A multi-objective process parameter intelligent optimization method, device and equipment

Abstract

The application discloses a multi-target process parameter intelligent optimization method, device and equipment, and relates to the technical field of intelligent manufacturing and process optimization. The method comprises the following steps: acquiring process parameter configuration and performance index configuration, constructing training samples based on a synthetic data sampling channel or a real experiment data import channel, and training a proxy model between the process parameters and the performance indexes; generating a scalar target function according to performance index role information, and processing constraint conditions in combination with a augmented Lagrange method; calling a Bayesian optimization engine or a reinforcement learning engine through a unified optimization interface to perform parameter search, and outputting an optimized process parameter combination and performance index prediction results according to constraint update results. The application can realize intelligent parameter optimization of a complex multi-target, multi-constraint process.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and process optimization technology, and more specifically to a method, apparatus and equipment for intelligent optimization of multi-objective process parameters. Background Technology

[0002] With the development of materials preparation technology and intelligent manufacturing technology, the demand for parameter optimization of complex materials and processes is constantly increasing. In materials research and development and process design, product performance is usually influenced by multiple process parameters, such as raw material ratio, reaction temperature, reaction time, pressure, and additive concentration, all of which can affect the final performance indicators. Taking the preparation process of lithium battery fire extinguishing agents as an example, its fire extinguishing efficiency, production cost, pH, and cycle stability are all subject to complex nonlinear coupling relationships with the synthesis process parameters. Furthermore, different performance indicators often have mutual constraints, making the process parameter optimization process a complex optimization problem with multiple inputs, multiple outputs, multiple objectives, and constraints.

[0003] Currently, the optimization of the aforementioned process parameters typically employs a manual experimental iterative approach. This involves researchers setting initial process parameters based on experience, preparing samples and testing performance indicators, then continuously adjusting the parameters and repeating the experiment based on the test results. However, this method requires a large number of experiments, has a long development cycle, and high experimental costs. Furthermore, due to the complex nonlinear relationship between process parameters and performance indicators, it is difficult to determine the optimal parameter combination that satisfies multiple performance requirements based solely on manual experience.

[0004] To reduce experimental costs and improve optimization efficiency, existing technologies are gradually introducing machine learning models and optimization algorithms to assist in the design of process parameters. This typically involves collecting experimental data to train a surrogate model to establish a mapping relationship between process parameters and performance indicators, and then using optimization algorithms to search for combinations of process parameters that meet the target requirements based on the surrogate model.

[0005] However, existing surrogate model-based process parameter optimization methods still have the following shortcomings: On the one hand, existing methods usually rely on existing real experimental data to build training samples. In the early stages of research and development of new materials and processes, due to the lack of real experimental data, it is difficult to quickly establish an effective proxy model, thus limiting the application of optimization processes in the early stages of research and development.

[0006] On the other hand, existing methods typically use a single optimization algorithm for parameter search. When different process problems have different objective function characteristics, a single optimization algorithm is difficult to adapt to different search requirements and cannot flexibly switch optimization strategies according to the characteristics of the process optimization problem.

[0007] Furthermore, existing methods typically handle performance index constraints by deeming them infeasible and discarding them directly if they exceed the range. However, the surrogate model's prediction results have certain errors, which cause some candidate parameter combinations that are close to the feasible region and have potential optimization value to be eliminated in advance, affecting the optimization search process.

[0008] Meanwhile, existing methods usually require pre-setting target weights and optimization directions during the initialization phase of the optimization process. When developers need to adjust the importance of different performance indicators, they need to reconfigure or even re-execute the optimization process, making it difficult to quickly compare parameter combinations under different optimization requirements based on the same proxy model.

[0009] In addition, existing optimization methods can usually only perform a forward optimization process that maximizes or minimizes the objective function. When the actual need is to determine whether a set of performance indicators simultaneously meet a preset range and to further find the corresponding process parameters, there is a lack of a mechanism to use existing surrogate models for reverse parameter search.

[0010] Therefore, there is an urgent need for an intelligent process parameter optimization method that can initiate the optimization process in the absence of real experimental data, automatically construct optimization objectives based on the target relationships between multiple performance indicators, support flexible switching of multiple optimization algorithms, and utilize a unified model to achieve forward parameter optimization and reverse feasible region query. Summary of the Invention

[0011] The purpose of this invention is to provide a method, apparatus, and device for intelligent optimization of multi-objective process parameters. By constructing a dual-channel data-driven mechanism, it can complete surrogate model training based on synthetic data when real experimental data is lacking, and switch data channels once experimental data is available, thereby improving the applicability of the process parameter optimization process. By assigning roles to output performance indicators and automatically generating scalar objective functions based on normalized benchmarks, it achieves unified optimization calculations under multiple performance indicators and constraints. By introducing an augmented Lagrangian constraint handling mechanism, it avoids the direct rejection of candidate parameter combinations due to slight deviations in constraint indicators, enabling the optimization process to fully utilize information near the constraint boundaries. Simultaneously, it enables flexible invocation of different optimization algorithms through a unified dual-engine interface and completes iterative search of process parameters using a surrogate model, thereby achieving intelligent optimization of multi-objective parameter combinations under complex process conditions.

[0012] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for intelligent optimization of multi-objective process parameters, the method comprising: Obtain the input parameter configuration and output performance index configuration corresponding to the process to be optimized. The input parameter configuration includes the parameter range of the process parameters to be optimized, and the output performance index configuration includes the role information corresponding to each output performance index. The role information includes the optimization target role, the constraint role, and the ignored role. Based on the input parameter configuration and output performance index configuration, either the synthetic data sampling channel or the real experimental data import channel is selected to obtain the training sample set. When the synthetic data sampling channel is selected, input parameter samples are generated within the input parameter range according to the preset synthesis function, and the corresponding output performance index is calculated. When the real experimental data import channel is selected, the imported experimental data is processed for abnormal data to form the training sample set. A proxy model is trained based on the mapping relationship between process parameters and performance indicators based on the training sample set, and the value range of each output performance indicator in the training sample set is obtained. The value range is used as a normalization benchmark and correspondingly saved with the proxy model. Based on the role information and normalized benchmark in the output performance index configuration, the role of each output performance index is analyzed to generate a scalar objective function for optimization calculation. For performance indices marked as optimization target roles, normalized weighting is performed according to the corresponding weights. For performance indices marked as constraint roles, the violation degree is calculated according to the corresponding constraint range, and the violation degree is introduced into the scalar objective function based on the augmented Lagrangian method. The preset optimization engine is invoked to iteratively search the surrogate model. The preset optimization engine includes a Bayesian optimization engine and a reinforcement learning engine. The Bayesian optimization engine and the reinforcement learning engine are called based on a unified optimization interface and use the same constraint update mechanism to search for the combination of process parameters that meet the constraints according to the scalar objective function. The Lagrange multipliers and penalty parameters are updated based on the constraint violations during the iteration process, and the optimized combination of process parameters and the corresponding performance index prediction results are output after the termination condition is met.

[0013] Secondly, the present invention also provides a multi-objective process parameter intelligent optimization device, the device comprising: The configuration acquisition module is used to acquire the input parameter configuration and output performance index configuration corresponding to the process to be optimized. The input parameter configuration includes the parameter range of the process parameters to be optimized, and the output performance index configuration includes the role information corresponding to each output performance index. The role information includes the optimization target role, the constraint role, and the ignored role. The training sample set acquisition module is used to select either a synthetic data sampling channel or a real experimental data import channel to acquire the training sample set based on the input parameter configuration and output performance index configuration. When the synthetic data sampling channel is selected, input parameter samples are generated within the input parameter range according to the preset synthesis function, and the corresponding output performance index is calculated. When the real experimental data import channel is selected, the imported experimental data is processed for abnormal data to form the training sample set. The model training module is used to train a proxy model that maps process parameters to performance indicators based on a set of training samples, and to obtain the range of values ​​for each output performance indicator in the set of training samples. The range of values ​​is then used as a normalization benchmark and saved in correspondence with the proxy model. The role parsing module is used to parse the roles of each output performance index according to the role information and normalization benchmark in the output performance index configuration, and generate a scalar objective function for optimization calculation. Specifically, for performance indices marked as optimization target roles, normalization weighting is performed according to the corresponding weights. For performance indices marked as constraint roles, the violation degree is calculated according to the corresponding constraint range, and the violation degree is introduced into the scalar objective function based on the augmented Lagrangian method. The iterative search module is used to call the preset optimization engine to perform iterative search on the surrogate model. The preset optimization engine includes a Bayesian optimization engine and a reinforcement learning engine. The Bayesian optimization engine and the reinforcement learning engine are called based on a unified optimization interface and use the same constraint update mechanism to search for the combination of process parameters that meet the constraints according to the scalar objective function. The parameter update module is used to update the Lagrange multipliers and penalty parameters according to the constraint violations during the iteration process, and outputs the optimized process parameter combination and the corresponding performance index prediction results after the termination condition is met.

[0014] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the intelligent optimization method for multi-objective process parameters provided in the first aspect.

[0015] Fourthly, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the intelligent optimization method for multi-objective process parameters provided in the first aspect.

[0016] Fifthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the intelligent optimization method for multi-objective process parameters provided in the first aspect.

[0017] The beneficial effects of this invention are as follows: The method of this invention, by constructing a dual-channel data-driven mechanism, can complete the training of the surrogate model based on synthetic data when real experimental data is lacking, and realize the data channel switching after experimental data is available, thereby improving the applicability of the process parameter optimization process; by configuring the output performance indicators in a role-based manner and automatically generating scalar objective functions in combination with normalized benchmarks, unified optimization calculation under multiple performance indicators and multiple constraints is realized; by introducing an augmented Lagrangian constraint processing mechanism, the direct rejection of candidate parameter combinations due to slight out-of-bounds constraints is avoided, enabling the optimization process to make full use of information near the constraint boundary; at the same time, the flexible invocation of different optimization algorithms is realized through a unified dual-engine interface, and the process parameter iterative search is completed in combination with the surrogate model, thereby realizing intelligent optimization of multi-objective parameter combinations under complex process conditions.

[0018] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating a multi-objective process parameter intelligent optimization method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a multi-objective process parameter intelligent optimization device according to an embodiment of the present invention; Figure 3 This is a schematic diagram of an electronic device structure provided in an embodiment of this application. Detailed Implementation

[0020] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] It should be noted that references to "an embodiment," "embodiment," "example embodiment," etc., in this specification refer to the described embodiment including specific features, structures, or characteristics; however, not every embodiment must include these specific features, structures, or characteristics. Furthermore, such expressions do not refer to the same embodiment. Moreover, when describing specific features, structures, or characteristics in conjunction with embodiments, whether or not explicitly described, it is indicated that incorporating such features, structures, or characteristics into other embodiments is within the knowledge of those skilled in the art.

[0022] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0023] Example 1 like Figure 1 As shown in this embodiment, a multi-objective intelligent optimization method for process parameters is provided. This method can be applied to scenarios such as material preparation, process parameter design, and intelligent manufacturing. It is used to automatically search for combinations of process parameters that meet multiple performance requirements based on the mapping relationship between process parameters and performance indicators. Taking the optimization process parameters for lithium battery fire extinguishing agent preparation as an example, this method addresses the problem of complex nonlinear correlations between multiple input process parameters such as raw material ratio, reaction temperature, reaction time, pressure, and additive concentration, and multiple output performance indicators such as fire extinguishing efficiency, production cost, pH, and cycle stability. By constructing a data-driven proxy model and combining it with a multi-objective optimization strategy, intelligent design of process parameters is achieved.

[0024] S101, obtain the input parameter configuration and output performance index configuration corresponding to the process to be optimized.

[0025] Input process parameters refer to process variables that can be adjusted and participate in the search during the optimization process. The input process parameter configuration describes the name, upper and lower bounds, default value, and whether each input process parameter participates in the optimization search. The upper and lower bounds limit the range of values ​​that the corresponding input process parameter can be searched. The default value provides the initial value of the input process parameter when it is not specifically specified. Whether it participates in the optimization search distinguishes between input process parameters that can be adjusted by the optimization algorithm and input process parameters that remain fixed in the current task.

[0026] For example, in the preparation process of lithium battery fire extinguishing agents, the raw material ratio, reaction temperature, reaction time, pressing pressure, and additive concentration can be configured as input process parameters, and corresponding upper and lower bounds can be set for each. When an optimization task requires a fixed reaction temperature, the reaction temperature can be set as a fixed parameter, maintaining a specified value during this optimization process, while other input process parameters continue to search within their respective value ranges.

[0027] Output performance metric configurations describe the product or process performance determined by input process parameters, with each output performance metric assigned a corresponding role. Roles determine how the corresponding output performance metric is handled in the current optimization task, specifically including three roles: optimization objective, constraint, and ignore.

[0028] When an output performance metric is configured as an optimization objective, it participates in the construction of the scalar objective function and is associated with a weight that has a positive or negative sign. A positive weight indicates a desire to maximize the value of the corresponding performance metric, while a negative weight indicates a desire to minimize the value of the corresponding performance metric. For example, if fire extinguishing efficiency is set as the optimization objective and assigned a positive weight, the optimization process tends to search for combinations of process parameters with higher fire extinguishing efficiency; if production cost is set as the optimization objective and assigned a negative weight, the optimization process tends to search for combinations of process parameters with lower production costs.

[0029] When an output performance metric is configured as a constraint, it does not participate in the target weighting in the direction of increase or decrease. Instead, it is set with corresponding upper and lower bounds to ensure that its predicted value is within the allowable range formed by the upper and lower bounds. For example, pH can be set as a constraint and configured with allowable lower and upper bounds.

[0030] When an output performance metric is configured to be ignored, it is not included in the current optimization calculation, but it can still be used as the prediction output of the surrogate model and displayed to the user when performing single-point prediction. For example, when an optimization task mainly focuses on fire extinguishing effectiveness and production costs, cyclic stability can be temporarily set to be ignored.

[0031] Therefore, output performance indicators can be configured with specific roles based on different process development needs, and the construction method of the subsequent objective function can be directly determined by the configuration, without having to rewrite the objective function calculation program for different combinations of performance indicators. The above configurations can also be saved as independent process schemes for use in different optimization tasks.

[0032] S102, based on the input parameter configuration and output performance index configuration, select the synthetic data sampling channel or the real experimental data import channel to obtain the training sample set.

[0033] When the current process has not yet accumulated real experimental data, a synthesis function sampling channel is used to obtain training samples. The synthesis function is a pre-configured function used to generate output performance indicators. It can adopt different mathematical forms for different output performance indicators to simulate the response relationship formed by different performance indicators as input process parameters change. In this embodiment, synthesis functions such as Gaussian peak functions, quadratic parabolic functions, sine modulation functions, piecewise step functions, or Lorentz peak functions can be configured for different performance indicators to form output performance indicators with different variation characteristics.

[0034] After determining the synthesis function, N sets of input process parameter samples are generated using the Latin hypercube sampling method based on the value range in the input process parameter configuration.

[0035] Latin hypercube sampling is a hierarchical random sampling method that divides the value range of each input process parameter into multiple intervals, selects sampling values ​​in each interval, and recombines the sampling values ​​of different input dimensions so that the obtained N sets of input process parameter samples can cover the input parameter space more evenly.

[0036] For example, when the input process parameters include raw material ratio, reaction temperature, reaction time, pressing pressure, and additive concentration, N sets of sampling parameters are generated based on the upper and lower bounds of each input process parameter. Each set of sampling parameters includes a specific value of each of the above input process parameters. Subsequently, each set of sampling parameters is input into the synthesis function corresponding to the output performance index to obtain the output performance index data such as fire extinguishing efficiency, production cost, pH, and cycle stability corresponding to that set of input process parameters.

[0037] Furthermore, after obtaining the output performance index data from the synthesis function, a certain proportion of random noise is superimposed to simulate the measurement error in the real experimental process, thereby obtaining a training sample set containing the input process parameters and the corresponding output performance index.

[0038] By employing a synthetic function sampling channel, initial training samples for surrogate model training can be established before real experimental data has been accumulated, allowing the model training, prediction, and optimization processes to be validated before real experimental data is generated.

[0039] Once real experimental data has been accumulated, select the real experimental data import channel. Real experimental data is recorded in tabular form, with each row corresponding to a set of experimental records. Some data columns are used to record input process parameters, while others are used to record output performance indicators obtained through experimental testing.

[0040] For example, in multiple fire extinguishing agent preparation experiments, after each experiment was completed, the raw material ratio, reaction temperature, reaction time, pressing pressure and additive concentration used in that experiment were recorded, as well as the fire extinguishing efficiency, production cost, pH and cycle stability corresponding to the process conditions, thus forming multiple sets of real experimental data.

[0041] Before real experimental data enters the training process, outlier detection based on interquartile range (IQR) is performed on each output performance metric. IQR represents the data range between the first and third quartiles of the corresponding output performance metric. For each output performance metric, its first quartile, third quartile, and IQR are determined. An outlier judgment range is formed based on the IQR between the first quartile and a preset multiple, and the IQR between the third quartile and a preset multiple. When the data corresponding to an output performance metric exceeds this judgment range, it is identified as an outlier.

[0042] For identified outliers, either truncation or row removal is selected based on the actual data processing mode. When truncation is used, data exceeding the outlier threshold is restricted to the corresponding boundary; when row removal is used, the entire experimental record containing the outlier is deleted from the training sample set.

[0043] This reduces the impact of abnormal experimental measurements or other outliers on subsequent training of the surrogate model, while preserving normal experimental data.

[0044] After obtaining the training sample set through the synthesis function sampling channel or the real experimental data import channel, the training sample set is input into a unified training pipeline. First, the training sample set is divided into a training set, a validation set, and a test set according to a preset ratio. The training set is used to train the surrogate model, the validation set is used to evaluate the model state during the training process, and the test set is used to evaluate the surrogate model's predictive ability on data that was not used in the training process after training is completed.

[0045] Subsequently, the input and output data in the training samples are normalized. Normalization can be achieved using min-max normalization or standardization methods, ensuring that different input process parameters and different output performance indicators can participate in neural network training on a unified data scale.

[0046] S103, a proxy model is trained based on the training sample set to establish the mapping relationship between process parameters and performance indicators, and the range of values ​​for each output performance indicator in the training sample set is obtained. The range of values ​​is then used as a normalization benchmark and saved in correspondence with the proxy model.

[0047] When training the surrogate model, weighted mean squared error is used as the training loss function. The loss weight of each output performance index is set according to the reciprocal of the square of the ratio of the index's standard deviation to its numerical range. This ensures that the training influence of different output performance indices in the normalized space is balanced, avoiding an excessive weight for a particular performance index due to its large data range, while also preventing performance indices with small numerical ranges from being ignored during training.

[0048] Subsequently, an adaptive gradient optimization algorithm is used to iteratively update the neural network parameters, and the current network parameters are evaluated using a validation set after each training epoch. When the validation loss no longer shows significant improvement after several consecutive training epochs, training is stopped, and the network parameters with the best validation loss are saved, thereby reducing the model's overfitting to noise in the training samples.

[0049] After training, the true minimum and true maximum values ​​of each output performance metric in the training dataset are used as the normalization benchmarks for the corresponding output performance metrics, and are persistently saved together with the trained network weights and network structure parameters.

[0050] The normalized baseline, denoted as y_bounds, includes the true value range of each output performance metric in the current training dataset. This normalized baseline is not manually specified by the user during optimization, but is directly derived from the training data corresponding to the current proxy model.

[0051] When performing subsequent single-point performance predictions or constructing scalar objective functions, the value range of the corresponding performance index is read from y_bounds, which is jointly stored with the current surrogate model. This ensures that the data range on which the performance index is based remains consistent with the current surrogate model. If the configuration fields of the input process parameters or output performance indexes change, and the changed configuration is inconsistent with the saved model, the saved model is deemed invalid, and the surrogate model is required to be retrained to ensure that the correspondence between the current surrogate model, the normalized baseline, and the effective configuration is maintained.

[0052] The above approach enables the use of a unified training process to build neural network surrogate models from synthetic and real experimental data, and allows the normalized baseline to be saved along with the model, providing a unified data foundation for subsequent multi-objective function construction and dual-engine optimization.

[0053] S104: Based on the role information in the output performance index configuration and the normalized benchmark, perform role analysis on each output performance index to generate a scalar objective function for optimization calculation.

[0054] For the output performance index configured as the target role for optimization, first obtain its predicted value, the corresponding true value range of the training data, and the configured weights, and normalize it according to the corresponding normalization benchmark. Then, determine the direction of the index's role in the objective function based on the positive or negative sign of the weights.

[0055] For output performance metrics configured as constraints, the predicted value of the metric, along with the pre-configured lower and upper bounds of the constraints, is obtained, and the violation degree is calculated based on the position of the predicted value relative to the constraint interval.

[0056] The violation degree corresponding to the kth constraint is: ; Where x represents the current combination of input process parameters to be evaluated. This represents the k-th constraint performance index predicted by the surrogate model for the current combination of input process parameters. This indicates the lower bound of the performance index of the constraint. This represents the upper bound of the constraint performance index. When the predicted value is between the upper and lower bounds, both of the above maximum values ​​are 0, and therefore the corresponding violation degree is 0; when the predicted value is below the lower bound or above the upper bound, the corresponding violation degree is determined according to the degree to which the constraint range is exceeded.

[0057] After obtaining the normalized predicted values ​​of each optimization objective and the degree of violation of each constraint, a scalar objective function is automatically constructed based on the output performance index role, weight, and constraint parameters.

[0058] Let the predicted value of the i-th "optimization target" role indicator be... Its true value range on the training data is (i.e., normalized benchmark), with weight w i Let the predicted value of the k-th "constraint" role indicator be... The constraint interval is The scalar objective function J(x) is constructed as follows: ; Where J(x) is the scalar objective function maximized by the actual execution of the optimization algorithm, x is the current combination of input process parameters, and w i Let the signed weights be the values ​​corresponding to the i-th optimization objective. The performance index of the i-th optimization objective is predicted by the surrogate model. and The normalized baseline is formed by the true minimum and maximum values ​​of the performance metric in the training data; Let k be the degree of violation of the constraint. Let ρ be the Lagrange multiplier corresponding to the k-th constraint, and ρ be the penalty coefficient.

[0059] Therefore, for positive weight optimization objectives, the larger the normalized predicted value, the greater the contribution to the scalar objective function; for negative weight optimization objectives, the smaller the normalized predicted value, the more beneficial the contribution to the scalar objective function. For constraint role indicators, Lagrange multipliers and squared penalty terms are used to continuously deduct points from candidate parameters that violate constraints, rather than discarding them directly when they slightly exceed the limits.

[0060] For example, in the process of optimizing lithium battery fire extinguishing agent technology, fire extinguishing efficiency can be set as the optimization objective with positive weight, production cost as the optimization objective with negative weight, and pH as a constraint. If the pH corresponding to a candidate process parameter slightly exceeds the set upper and lower bounds, the candidate process parameter can still enter the optimization evaluation process, but its scalar objective function will be correspondingly reduced due to the non-zero violation.

[0061] Therefore, multiple output performance indicators with different dimensions and different optimization directions can be automatically configured to form a unified scalar objective function, while constraint indicators can participate in optimization evaluation in a soft constraint manner.

[0062] This part connects directly with the subsequent dual-engine search: the generated J(x) serves as a unified evaluation criterion for both the Bayesian optimization engine and the reinforcement learning engine, without the need to reconstruct the objective function for different optimization algorithms.

[0063] S105, invoke the preset optimization engine to perform iterative search on the proxy model.

[0064] In this embodiment, the optimization engine includes a Bayesian optimization engine and a reinforcement learning optimization engine. Both optimization engines call the proxy model, scalar objective function, and constraint processing module through a unified optimization interface, thereby enabling different optimization algorithms to reuse the same process parameter optimization task.

[0065] The unified optimization interface is used to shield the data structure differences between different optimization algorithms, so that the optimization engine only needs to input the current candidate input process parameter combination into the interface to obtain the performance index prediction results and the corresponding scalar objective function values ​​output by the surrogate model.

[0066] Specifically, after the optimization engine generates a set of candidate input process parameter combinations x, the unified optimization interface sends the input process parameter combination to the proxy model, which then outputs the corresponding predicted values ​​of multiple performance indicators. Subsequently, based on the role configuration of each performance indicator, the normalized benchmark, and the constraint information, the corresponding scalar objective function J(x) is calculated, and the calculation results are fed back to the optimization engine.

[0067] For example, in the process of optimizing the preparation process of fire extinguishing agents, the optimization engine generates a set of candidate parameter combinations including raw material ratio, reaction temperature, reaction time and additive concentration. The unified optimization interface inputs the parameter combination into the surrogate model to obtain the predicted values ​​of fire extinguishing performance, production cost and pH corresponding to the parameter combination. Then, the comprehensive evaluation result of the candidate parameter combination is calculated according to the scalar objective function.

[0068] By using a unified optimization interface, the surrogate model, objective function construction process, and constraint handling process are separated from the specific optimization algorithm. When it is necessary to replace the optimization algorithm, only the optimization engine needs to be replaced, without having to rebuild the data processing flow, surrogate model, and evaluation system.

[0069] When using a Bayesian optimization engine to perform parameter search, a probabilistic model between the input process parameter space and the objective function is established based on existing input process parameter samples and corresponding scalar objective function evaluation results. The acquisition function is then used to select the next round of input process parameter combinations to be evaluated.

[0070] Among them, the Bayesian optimization engine is suitable for process optimization scenarios where real-world experiments are costly and single performance evaluation takes a long time. Since the surrogate model can quickly predict the performance indicators corresponding to the input process parameters, the Bayesian optimization engine does not need to directly and frequently call real experiments, but instead completes large-scale parameter screening within the prediction space provided by the surrogate model.

[0071] Specifically, in the Bayesian optimization process, the existing set of evaluation samples is first obtained, where each evaluation sample includes a combination of input process parameters and the corresponding scalar objective function value.

[0072] Subsequently, a probabilistic prediction model was established based on existing evaluation samples to predict the possible trend of objective function changes in the unevaluated input process parameter region.

[0073] Based on this, the next round of input process parameters to be evaluated is selected according to the desired improvement of the acquisition function.

[0074] Among them, it is expected that the improved acquisition function will comprehensively consider the predicted values ​​of the objective function corresponding to the candidate input process parameters and the prediction uncertainty, so that the optimization process will not only focus on the region with good current prediction effect, but also explore the parameter region that has not been fully searched.

[0075] The desired improvement function is expressed as: ; Where μ(x) and σ(x) represent the predicted mean and predicted standard deviation for any candidate parameter point in the search space, respectively. This represents the currently observed optimal objective function value. , These are the cumulative distribution function and probability density function of the standard normal distribution, respectively. In each iteration, a candidate point that maximizes EI is selected, substituted into the surrogate model to predict the performance index, and the scalar objective function value is calculated. The observation point is then added to the observed set and a Gaussian process is refitted. This process is repeated until the set number of iterations is reached.

[0076] In this way, the Bayesian optimization engine can prioritize the selection of input process parameter combinations with high optimization potential for evaluation within a limited number of evaluation attempts.

[0077] For example, in the fire extinguishing agent preparation process, if experimental data shows that the combination of reaction temperature and additive concentration in a certain area has high fire extinguishing efficiency, the Bayesian optimization engine will continue to search in the vicinity of that area. At the same time, if a certain area has average current evaluation results but high prediction uncertainty, it may also be selected for further evaluation to avoid missing potentially excellent parameter combinations.

[0078] When using a reinforcement learning optimization engine to perform parameter search, the optimization process of input process parameters is constructed as a continuous decision-making process based on policy updates.

[0079] The reinforcement learning optimization engine includes state information, action information, and reward information.

[0080] Status information is used to describe the current state of the optimization process, including the current combination of input process parameters, historical search results, and information about the current optimization stage.

[0081] Action information is used to indicate operations that adjust the input process parameters, such as increasing the reaction temperature, decreasing the additive concentration, or changing the raw material ratio.

[0082] Reward information is used to evaluate the optimization effect of the current action, and it is determined based on the calculation results of the scalar objective function.

[0083] Specifically, in each round of optimization, the reinforcement learning optimization engine generates the next round of input process parameter adjustment actions through the policy network based on the current state information.

[0084] The policy network is used to describe the probabilistic mapping relationship between states and actions.

[0085] The next set of candidate input process parameters is selected based on the action probability output by the policy network, and the scalar objective function value corresponding to the candidate parameter is obtained through the unified optimization interface.

[0086] Subsequently, reward information is generated based on the evaluation results of the objective function, and the policy network is updated using the reward information, so that the policy network gradually tends to choose the input process parameter adjustment method that can obtain a higher objective function value.

[0087] The REINFORCE method is used for policy updates during the reinforcement learning optimization process.

[0088] Among them, the REINFORCE method maximizes the expected reward corresponding to the sampling trajectory, so that the policy network parameters are gradually updated towards a better policy.

[0089] The REINFORCE loss function is expressed as: ; in, Let B be the moving average of historical rewards (baseline, used to reduce the variance of gradient estimation), and let B be the batch size for each sampling round. The mean μ and standard deviation σ of the policy are updated using the gradient of this loss function (with upper and lower bounds on the range of σ to prevent premature shrinkage or divergence of the distribution), causing the policy distribution to gradually shrink and focus towards the high-reward region. This process is repeated until the set number of iterations is reached. The final mean μ and standard deviation σ of the policy are also output as part of the results, which can be used to demonstrate the confidence level of the optimal parameter distribution. b For the b-th sampling candidate process parameter; r b Candidate process parameter x b The corresponding rewards.

[0090] For example, in the process of optimizing the preparation process of fire extinguishing agents, if the surrogate model predicts that the fire extinguishing efficiency is improved and the production cost is reduced after adjusting the reaction temperature and additive concentration, then this action will receive a higher reward; the reinforcement learning optimization engine will increase the selection probability of similar actions in the subsequent search process based on this reward.

[0091] Through multiple rounds of state updates, action execution, and reward feedback, the reinforcement learning optimization engine can gradually form a search strategy suitable for the current process optimization task.

[0092] S106 updates the Lagrange multipliers and penalty parameters based on constraint violations during the iteration process, and outputs the optimized combination of process parameters and corresponding performance index prediction results after the termination condition is met.

[0093] Both the Bayesian optimization engine and the reinforcement learning optimization engine employ a shared augmented Lagrange outer loop to handle constraints during the search process.

[0094] Among them, the shared augmented Lagrange outer loop is used to dynamically adjust the constraint processing parameters according to the current optimization results, so that the optimization process gradually approaches the feasible region that satisfies the constraint conditions.

[0095] Specifically, in each round of optimization, the optimization engine first obtains the corresponding constraint violation degree v based on the current parameter search results. k (x).

[0096] Subsequently, the Lagrange multipliers and penalty coefficients are updated based on the current constraint violation status.

[0097] Regardless of the internal engine used, the Lagrange multipliers corresponding to the constraints in the scalar objective function... With penalty coefficient The update logic is uniformly controlled by the outer loop, regardless of the specific engine used in the inner loop: the total number of iterations is divided into several "outer loops". Within each outer loop, the selected engine is called to execute a segment of inner loop iterative search. The current optimal candidate solution obtained from this segment of inner loop search is taken, and its violation degree on each constraint is calculated. Update multipliers, press Increase the penalty coefficient before entering the next outer loop; if the total violation of all constraints is less than the preset threshold at the end of an outer loop, terminate the outer loop early and use the remaining iteration count quota to supplement the convergence curve record of the output result, ensuring that the final returned iteration count is consistent with the iteration count requested by the user.

[0098] By dynamically updating the Lagrange multipliers and penalty coefficients, combinations of input process parameters that violate constraints are subject to stronger constraints in subsequent search processes. At the same time, some candidate parameters that are close to the feasible region are retained, so that the optimization process can avoid prematurely discarding potential high-quality solutions.

[0099] For example, when a set of process parameters can achieve high fire extinguishing efficiency, but its pH slightly exceeds the allowable range, the augmented Lagrange mechanism will not immediately delete the parameter combination, but will reduce its evaluation value according to the degree of violation, and gradually adjust it to a position that meets the constraints through subsequent searches.

[0100] After completing multiple rounds of optimization iterations, the optimization process is terminated based on preset termination conditions.

[0101] The termination conditions include reaching the maximum number of optimization iterations, or the scalar objective function changing to meet a preset stability condition during several consecutive rounds of optimization.

[0102] When the termination condition is not met, the optimization engine continues to generate new candidate combinations of input process parameters and repeats the surrogate model prediction, objective function calculation and constraint update process.

[0103] Once the termination condition is met, stop the optimization search and output the final optimization result.

[0104] The output results include: the optimized combination of input process parameters; the corresponding predicted output performance indicators; the evaluation value of the scalar objective function; and the iteration history information during the optimization process.

[0105] For example, the final output is a set of fire extinguishing agent preparation process parameters, including raw material ratio, reaction temperature, reaction time and additive concentration, and the corresponding fire extinguishing performance prediction, production cost prediction and constraint satisfaction status.

[0106] Through the above process, intelligent optimization of process parameters is achieved based on a proxy model, a multi-objective evaluation function, and the collaborative work of two optimization engines.

[0107] Furthermore, after completing the forward optimization, this embodiment also supports reverse feasible domain search based on the target performance index range.

[0108] Forward optimization refers to inputting process parameters, predicting output performance indicators through a surrogate model, and searching for the optimal combination of input process parameters that achieves the best performance indicators. Reverse feasible region search, on the other hand, involves finding the set of input process parameters that can produce the required performance indicators when the range of the target performance indicators is known.

[0109] For example, in practical applications, users may raise the following issues: The fire extinguishing effectiveness must reach the target range, while production costs must be limited to a preset value.

[0110] At this point, there is no need to retrain the proxy model; instead, the existing proxy model can be used directly for reverse search.

[0111] Specifically, the target performance index range input by the user is converted into a constraint, and the corresponding performance index is adjusted from the optimization target role to the constraint role.

[0112] For example, in the original task, fire extinguishing efficiency is the optimization objective, and it is desirable to have it as high as possible; in the reverse search task, fire extinguishing efficiency is converted into a constraint indicator, and a target range is set so that the search results satisfy the requirement that the fire extinguishing efficiency is within a specified range.

[0113] Subsequently, all target performance indicators and existing constraints are input into the reverse search module, and the differential evolution algorithm is used to find the combination of input process parameters that meet the constraints.

[0114] The differential evolution algorithm first generates multiple candidate parameter individuals in the input process parameter space, with each individual representing a set of possible process parameter combinations.

[0115] For example, a candidate individual can be represented as: raw material ratio = a; reaction temperature = b; reaction time = c; additive concentration = d.

[0116] Subsequently, mutation, crossover, and selection operations were performed on different candidate individuals, and the performance indicators of each candidate individual were predicted using a surrogate model.

[0117] The degree of constraint satisfaction is calculated based on the prediction results, and candidate individuals that meet the target range or are closer to the target range are retained.

[0118] After multiple iterations, a set of feasible process parameters that meet the specified performance index range is obtained.

[0119] Unlike traditional methods that only output a single optimal parameter, this embodiment outputs feasible combinations of process parameters and details of violations through reverse feasible domain search, or outputs the reference candidate solution with the minimum total violation when it is not feasible, providing process engineers with more design options.

[0120] Based on the same inventive concept, this application also provides a multi-objective process parameter intelligent optimization device for implementing the multi-objective process parameter intelligent optimization method described above. The solution provided by this device is similar to the implementation scheme described in the above method. Therefore, the specific limitations in one or more embodiments of the multi-objective process parameter intelligent optimization device provided below can be found in the limitations of the multi-objective process parameter intelligent optimization method described above, and will not be repeated here.

[0121] In one embodiment, such as Figure 2 As shown, a multi-objective process parameter intelligent optimization device is provided, the device comprising: The configuration acquisition module 31 is used to acquire the input parameter configuration and output performance index configuration corresponding to the process to be optimized. The input parameter configuration includes the parameter range of the process parameters to be optimized, and the output performance index configuration includes the role information corresponding to each output performance index. The role information includes the optimization target role, the constraint role, and the ignored role. The training sample set acquisition module 32 is used to select either the synthetic data sampling channel or the real experimental data import channel to acquire the training sample set according to the input parameter configuration and the output performance index configuration. When the synthetic data sampling channel is selected, the input parameter samples are generated within the input parameter range according to the preset synthesis function, and the corresponding output performance index is calculated. When the real experimental data import channel is selected, the imported experimental data is processed for abnormal data to form the training sample set. The model training module 33 is used to train a proxy model that maps the process parameters to performance indicators based on the training sample set, and to obtain the range of values ​​for each output performance indicator in the training sample set. The range of values ​​is then saved as a normalization benchmark and corresponding to the proxy model. The role parsing module 34 is used to parse the role of each output performance index according to the role information in the output performance index configuration and the normalized benchmark, and generate a scalar objective function for optimization calculation. Among them, for the performance index marked as the optimization target role, normalized weighting is performed according to the corresponding weight. For the performance index marked as the constraint role, the violation degree is calculated according to the corresponding constraint range, and the violation degree is introduced into the scalar objective function based on the augmented Lagrangian method. The iterative search module 35 is used to call the preset optimization engine to perform iterative search on the surrogate model. The preset optimization engine includes a Bayesian optimization engine and a reinforcement learning engine. The Bayesian optimization engine and the reinforcement learning engine are called based on a unified optimization interface and use the same constraint update mechanism to search for the combination of process parameters that meet the constraints according to the scalar objective function. The parameter update module 36 is used to update the Lagrange multipliers and penalty parameters according to the constraint violation situation during the iteration process, and output the optimized process parameter combination and the corresponding performance index prediction results after the termination condition is met.

[0122] This application also provides an electronic device, in some embodiments, referring to... Figure 3 As shown, the electronic device 700 includes an input unit 710, a memory 720, a processor 730, and an output unit 740. The memory 720 stores program instructions that can be executed on the processor 730. The processor 730 can execute the intelligent optimization method and / or technical solution for multi-objective process parameters based on the aforementioned embodiments by calling the program instructions. This electronic device 700 can be a mobile terminal device such as a mobile phone or a computer.

[0123] Furthermore, embodiments of this application also provide a computer-readable storage medium for storing a computer program that executes a multi-objective process parameter intelligent optimization method. For example, computer program instructions, when executed by a computer, can invoke or provide the methods and / or technical solutions according to this application through the operation of the computer. The program instructions that invoke the methods of this application may be stored in a fixed or removable storage medium, and / or transmitted via data streams in broadcast or other signal carrying media, and / or stored in a storage medium that operates according to the program instructions.

[0124] Obviously, those skilled in the art should understand that the modules or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device, or fabricating them separately as individual integrated circuit modules, or fabricating multiple modules or steps as a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.

[0125] The technical features of the above embodiments can be arbitrarily integrated. For the sake of brevity, not all possible integrations of the technical features in the above embodiments are described. However, as long as the integration of these technical features does not contradict each other, they should be considered to be within the scope of this specification.

[0126] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for intelligent optimization of multi-objective process parameters, characterized in that, The method includes: Obtain the input parameter configuration and output performance index configuration corresponding to the process to be optimized. The input parameter configuration includes the parameter range of the process parameters to be optimized, and the output performance index configuration includes the role information corresponding to each output performance index. The role information includes the optimization target role, the constraint role, and the ignore role. Based on the input parameter configuration and the output performance index configuration, a training sample set is obtained by selecting either a synthetic data sampling channel or a real experimental data import channel. When the synthetic data sampling channel is selected, input parameter samples are generated within the input parameter range according to a preset synthesis function, and the corresponding output performance index is calculated. When the real experimental data import channel is selected, the imported experimental data is processed for abnormal data to form a training sample set. A proxy model is trained based on the training sample set to map the process parameters and performance indicators, and the range of values ​​for each output performance indicator in the training sample set is obtained. The range of values ​​is then used as a normalization benchmark and saved in correspondence with the proxy model. Based on the role information in the output performance index configuration and the normalization benchmark, role analysis is performed on each output performance index to generate a scalar objective function for optimization calculation. For performance indices marked as optimization target roles, normalization weighting is performed according to the corresponding weights. For performance indices marked as constraint roles, violation degree is calculated according to the corresponding constraint range, and the violation degree is introduced into the scalar objective function based on the augmented Lagrangian method. The surrogate model is iteratively searched by a preset optimization engine, which includes a Bayesian optimization engine and a reinforcement learning engine. The Bayesian optimization engine and the reinforcement learning engine are called based on a unified optimization interface and use the same constraint update mechanism to search for a combination of process parameters that meet the constraints according to the scalar objective function. The Lagrange multipliers and penalty parameters are updated based on the constraint violations during the iteration process, and the optimized combination of process parameters and the corresponding performance index prediction results are output after the termination condition is met.

2. The intelligent optimization method for multi-objective process parameters as described in claim 1, characterized in that, The step of obtaining the training sample set based on the input parameter configuration and the output performance index configuration includes: When using a synthetic data sampling channel, multiple input parameter samples are generated in the input parameter space using Latin hypercube sampling. The synthesis function corresponding to each output performance index is called respectively, the corresponding output performance index is calculated according to the input parameter sample, and a preset noise amount is added to the calculation result to form a training sample set; When using real experimental data to import the channel, the interquartile range anomaly detection is performed on the output performance indicators in the experimental data, and the abnormal data is truncated or removed based on the detection results.

3. The intelligent optimization method for multi-objective process parameters as described in claim 1, characterized in that, The proxy model for the mapping relationship between the training process parameters and performance indicators includes: The training sample set is divided into a training set, a validation set, and a test set; The input parameters and output performance metrics are normalized respectively. Construct a neural network model that includes an input layer, multiple hidden layers, and an output layer, where the hidden layers use a non-linear activation function; The weighted mean square error is used as the model training loss function, and the corresponding loss weights are determined according to the relationship between the standard deviation and numerical range of each output performance index. The network parameters are updated using an adaptive gradient optimization algorithm, and training is stopped early based on the changes in the validation set loss.

4. The intelligent optimization method for multi-objective process parameters as described in claim 1, characterized in that, The step of calling the preset optimization engine for iterative search includes: Based on the algorithm selection information in the optimization request, select the target optimization engine from the Bayesian optimization engine and the reinforcement learning engine; When the Bayesian optimization engine is selected, the relationship between process parameters and scalar objective function is predicted using a probabilistic model, and the process parameters to be evaluated in the next round are determined based on the acquisition function. When selecting a reinforcement learning engine, a parameter policy distribution is constructed in the normalized input parameter space, and the policy parameters are updated based on the scalar objective function value output by the surrogate model as the reward signal. Both the Bayesian optimization engine and the reinforcement learning engine return process parameter results, performance index predictions, and iteration history information through a unified optimization interface.

5. The intelligent optimization method for multi-objective process parameters as described in claim 1, characterized in that, It also includes the reverse feasible region search process: Receive the target value range for the output performance index; Rebind the performance indicators corresponding to the optimization target roles in the original configuration to the constraint roles, and use the target value range as the corresponding constraint range; The total violation function is constructed by combining the performance metrics corresponding to the rebound constraint roles with those of the original constraint roles. Based on the surrogate model and the total violation function, a global search is performed in the input parameter space to obtain a combination of process parameters that satisfies the target value range.

6. The intelligent optimization method for multi-objective process parameters as described in claim 5, characterized in that, The global search employs a differential evolution algorithm, including: Initialize a population containing multiple candidate process parameter combinations; Differential mutation and crossover operations are performed on candidate process parameter combinations in the population to generate new candidate parameter combinations. The surrogate model is used to predict the performance index corresponding to the new candidate parameter combination, and the corresponding total violation is calculated; The candidate parameter combinations are selected and updated based on the total violation until a feasible parameter combination with a total violation less than a preset threshold is obtained, or after reaching a preset number of iterations, the candidate parameter combination with the minimum total violation is output as a reference result.

7. A multi-objective intelligent optimization device for process parameters, characterized in that, The device includes: The configuration acquisition module is used to acquire the input parameter configuration and output performance index configuration corresponding to the process to be optimized. The input parameter configuration includes the parameter range of the process parameters to be optimized, and the output performance index configuration includes the role information corresponding to each output performance index. The role information includes the optimization target role, the constraint role, and the ignore role. The training sample set acquisition module is used to select either a synthetic data sampling channel or a real experimental data import channel to acquire the training sample set according to the input parameter configuration and the output performance index configuration. When the synthetic data sampling channel is selected, input parameter samples are generated within the input parameter range according to a preset synthesis function, and the corresponding output performance index is calculated. When the real experimental data import channel is selected, the imported experimental data is processed for abnormal data to form the training sample set. The model training module is used to train a proxy model that maps process parameters to performance indicators based on the training sample set, and to obtain the value range of each output performance indicator in the training sample set, and to save the value range as a normalization benchmark corresponding to the proxy model. The role parsing module is used to parse the role information in the output performance index configuration and the normalization benchmark to generate a scalar objective function for optimization calculation. For performance indices marked as optimization target roles, normalization weighting is performed according to the corresponding weights. For performance indices marked as constraint roles, the violation degree is calculated according to the corresponding constraint range and introduced into the scalar objective function based on the augmented Lagrangian method. The iterative search module is used to call a preset optimization engine to perform iterative search on the surrogate model. The preset optimization engine includes a Bayesian optimization engine and a reinforcement learning engine. The Bayesian optimization engine and the reinforcement learning engine are called based on a unified optimization interface and use the same constraint update mechanism to search for a combination of process parameters that satisfy the constraint conditions according to the scalar objective function. The parameter update module is used to update the Lagrange multipliers and penalty parameters according to the constraint violations during the iteration process, and outputs the optimized process parameter combination and the corresponding performance index prediction results after the termination condition is met.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the intelligent optimization method for multi-objective process parameters as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the intelligent optimization method for multi-objective process parameters as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the intelligent optimization method for multi-objective process parameters as described in any one of claims 1 to 6.

Images