An aluminum alloy process parameter regulation method based on machine learning

Abstract

The application discloses an aluminum alloy process parameter regulation and control method based on machine learning, relates to the technical field of process parameter regulation and control, and comprises the following steps: after a double-channel data processing path is constructed, a separate full connection layer is used to perform nonlinear physical mapping on a feature subset, a high-order physical state vector is generated, a Bayesian neural network model is trained by taking the high-order physical state vector as input and product performance data as a prediction target, a physical constraint mixed model is obtained, internal gradient information of the physical constraint mixed model is calculated, an uncertainty contribution degree vector is acquired, an optimization target is set, the physical constraint mixed model is taken as an evaluation function, the uncertainty contribution degree vector is taken as a constraint condition, a Pareto optimal solution set is searched for in a process parameter space, and after the Pareto optimal solution set is visualized, optimal process parameter combinations are screened out; and the aluminum alloy hot working process is intelligently, finely and reliably regulated and controlled through the parameter regulation and control method.

Description

Technical Field

[0001] This invention relates to the field of process parameter control technology, and in particular to a method for controlling aluminum alloy process parameters based on machine learning. Background Technology

[0002] Aluminum alloys, due to their excellent corrosion resistance and good formability, have become important structural materials in high-end manufacturing fields such as aerospace and new energy vehicles. With the continuous increase in demand for lightweight and high performance, the control of the mechanical properties of aluminum alloys increasingly depends on the precise control of the microstructure evolution during hot working and aging heat treatment, especially the synergistic optimization of the type, size and distribution of precipitates and grain morphology. In the past, the setting of process parameters was mainly based on metallurgical theory and empirical rules accumulated from a large number of experiments. In recent years, with the in-depth advancement of intelligent manufacturing, data-driven methods have shown great potential in material process optimization. Machine learning methods have been used to establish nonlinear mapping relationships between "composition-process-performance", improving process development efficiency. Especially in the exploration of high-dimensional parameter spaces and the modeling of complex nonlinear relationships, machine learning has shown the ability to surpass previous empirical models, providing a new path for intelligent process design of aluminum alloys.

[0003] Nevertheless, existing methods for controlling aluminum alloy process parameters still have room for improvement. First, most machine learning models do not explicitly consider the multi-scale coupling mechanism between precipitation strengthening and grain structure evolution in aluminum alloys, resulting in limited feature representation capabilities. Previous optimization strategies typically only use the average performance as the optimization objective, ignoring the uncertainty of model predictions and its impact on process robustness. They cannot effectively balance the multi-objective conflict between performance, stability, and energy consumption. Furthermore, there is a lack of quantification mechanisms for the sources of uncertainty in parameter sensitivity analysis and Pareto frontier exploration, making it difficult to identify perturbation factors and thus hindering intelligent control of aluminum alloy process parameters. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a machine learning-based method for controlling aluminum alloy process parameters to address the issues of impact on process robustness and the inability to intelligently control process parameters.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a machine learning-based method for controlling aluminum alloy process parameters, which includes collecting product performance data, raw material data, and process parameters from historical aluminum alloy production processes. Extract the physical features related to effective accumulated temperature from the process parameters and combine them with the raw material data to form a joint feature set; The joint feature set is identified and classified into feature subsets; After constructing a dual-channel data processing path, an independent fully connected layer is used to perform nonlinear physical mapping on the feature subset to generate a high-order physical state vector. Using high-order physical state vectors as input and product performance data as prediction targets, a Bayesian neural network model is trained to obtain a hybrid physical constraint model. Calculate the internal gradient information of the physical constraint hybrid model to obtain the uncertainty contribution vector; An optimization objective is set, and a physical constraint hybrid model is used as the evaluation function, with the uncertainty contribution vector as the constraint condition. The Pareto optimal solution set is then searched in the process parameter space. After visualizing the Pareto optimal solution set, the optimal combination of process parameters is selected.

[0007] As a preferred embodiment of the machine learning-based aluminum alloy process parameter control method of the present invention, the product performance data includes tensile strength, yield strength and elongation. The raw material data includes the weight percentages of magnesium, silicon, and copper. The process parameters include the extrusion speed, barrel temperature and die temperature during the extrusion stage, the cooling medium flow rate during the quenching stage, and the temperature-time curve of the entire aging heat treatment process.

[0008] As a preferred embodiment of the machine learning-based aluminum alloy process parameter control method of the present invention, the extraction of physical features related to effective accumulated temperature in the process parameters refers to using an aluminum alloy precipitation kinetic model to perform integral calculations on the temperature-time curves in the process parameters to obtain the effective accumulated temperature, and using the effective accumulated temperature as a newly added physical feature.

[0009] As a preferred embodiment of the machine learning-based aluminum alloy process parameter control method of the present invention, the combination with raw material data to form a joint feature set refers to horizontally combining the newly added physical features with the weight percentages of magnesium, silicon and copper elements in the raw material data to form a joint feature set.

[0010] As a preferred embodiment of the machine learning-based aluminum alloy process parameter control method of the present invention, wherein: the identification and classification of features in the joint feature set into feature subsets specifically includes: Features strongly correlated with precipitation dynamics in the joint feature set are classified as a subset of precipitation enhancement features; Features that are strongly correlated with grain morphology evolution in the joint feature set are classified as a subset of grain structure features.

[0011] As a preferred embodiment of the machine learning-based aluminum alloy process parameter control method of the present invention, wherein: after constructing the dual-channel data processing path, an independent fully connected layer is used to perform nonlinear physical mapping on the feature subset to generate a high-order physical state vector, specifically: For the precipitation strengthening feature subset and the grain structure feature subset, a parallel dual-channel data processing path is established; Using a dual-channel data processing path, the enhanced feature subset and the grain structure feature subset are transmitted to independent fully connected layers and nonlinear transformations are performed to obtain the precipitate phase state vector and the grain structure state vector. The precipitated phase state vector and the grain structure state vector are concatenated to form a higher-order physical state vector.

[0012] As a preferred embodiment of the machine learning-based aluminum alloy process parameter control method of the present invention, wherein: the Bayesian neural network model is trained using a high-order physical state vector as input and product performance data as the prediction target to obtain a physical constraint hybrid model, specifically: The higher-order physical state vector is input into the Bayesian neural network model for forward propagation to obtain physical parameters and product performance prediction values. Calculate the physical constraint loss term using theoretical physics equations; The data fitting loss term is calculated based on the predicted product performance values ​​and the product performance data. The total loss function value is obtained by weighted summation of the data fitting loss term and the physical constraint loss term. The weight parameter distribution of the Bayesian neural network model is then optimized using the backpropagation algorithm to minimize the total loss function value, resulting in a hybrid physical constraint model.

[0013] As a preferred embodiment of the machine learning-based aluminum alloy process parameter control method of the present invention, wherein: the calculation of the gradient information within the physical constraint hybrid model to obtain the uncertainty contribution vector specifically involves: Input a set of process parameters and raw material data into the physical constraint hybrid model, and perform forward propagation calculations to obtain the uncertainty value; The automatic differentiation method is used to calculate the gradient of the uncertainty value with respect to process parameters and raw material data, and the absolute value of each gradient is taken and normalized. Arrange all the normalized absolute gradient values ​​to form a sequence of absolute gradient values ​​and use it as an uncertainty contribution vector.

[0014] As a preferred embodiment of the machine learning-based aluminum alloy process parameter control method of the present invention, the step of setting an optimization objective, using a physical constraint hybrid model as the evaluation function, and an uncertainty contribution vector as the constraint condition, and searching for a Pareto optimal solution set in the process parameter space, specifically involves: The optimization objectives are set as maximizing the mean performance predicted by the physical constraint hybrid model, minimizing the performance uncertainty predicted by the physical constraint hybrid model, and minimizing production energy consumption, with the physical constraint hybrid model as the evaluation function. A multi-objective evolutionary algorithm is used to generate a candidate population within the process parameter space through initialization, selection, crossover, and mutation operations. After combining each candidate population with the raw material data, the data is input into the physical constraint hybrid model for calculation, resulting in a set of average performance values, a set of performance uncertainty values, and a set of production energy consumption values. Based on the set of average performance values, the set of performance uncertainty values, the set of production energy consumption values, and the uncertainty contribution vector, the process parameters that contribute the most to uncertainty are identified. Through iterative selection, crossover, and mutation, the Pareto optimal solution set is obtained.

[0015] As a preferred embodiment of the machine learning-based aluminum alloy process parameter control method of the present invention, the step of visualizing the Pareto optimal solution set and then selecting the optimal combination of process parameters specifically involves: Plot the Pareto optimal solution set in a three-dimensional coordinate system and receive the user-defined performance mean, performance uncertainty value and production energy consumption value preference weights in the graphical interface. Calculate the comprehensive evaluation value of each solution in the Pareto optimal solution set based on preference weights, and select the solution with the highest comprehensive evaluation value as the optimal combination of process parameters. As a preferred embodiment of the machine learning-based aluminum alloy process parameter control method described in this invention, wherein: The beneficial effects of this invention are as follows: By extracting the physical feature of "effective accumulated temperature" related to aging precipitation kinetics and constructing a joint feature set with metallurgical significance in combination with raw material composition, the physical quantitative characterization of the process is realized. Furthermore, based on the strengthening mechanism, the features are decoupled into two subsets: precipitation strengthening and grain structure, effectively capturing the nonlinear laws of multi-scale microstructure evolution. On this basis, the JMAK precipitation kinetic equation (Johnson-Mel-Avramy-Kolmogorov precipitation kinetic equation) and the physical constraint loss term of recrystallization kinetics are introduced to train a Bayesian neural network to form a physical constraint hybrid model with both high-precision prediction capability and uncertainty quantification function. This supports the efficient search for Pareto optimal solutions that take into account high performance, high robustness and low energy consumption in the process parameter space, thereby realizing intelligent, refined and reliable control of aluminum alloy hot working process. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of a machine learning-based method for controlling aluminum alloy process parameters.

[0018] Figure 2 A flowchart for obtaining a hybrid model of physical constraints.

[0019] Figure 3 A flowchart for obtaining the optimal combination of process parameters.

[0020] Figure 4 A flowchart for obtaining the uncertainty contribution vector. Detailed Implementation

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0022] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0023] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0024] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for controlling aluminum alloy process parameters based on machine learning, including the following steps: S1. Collect product performance data, raw material data, and process parameters from the historical production process of aluminum alloys.

[0025] It should be noted that during the mechanical property testing phase of the historical production process of aluminum alloy products, standard tensile specimens are prepared from finished or semi-finished aluminum alloy materials, and tensile tests are conducted on the standard tensile specimens using a universal testing machine. Simultaneously, tensile strength, yield strength, and elongation are recorded to obtain product performance data. During the raw material warehousing phase, a spectrometer is used to analyze the composition of each batch of aluminum ingots, recording the spectral analysis results of magnesium, silicon, and copper elements and converting them into weight percentages to obtain raw material data. During the production process, sensors installed on the extruder collect and record extrusion speed, barrel temperature, and die temperature in real time. A flow meter records the flow rate of the quenching cooling medium, and thermocouples and a data acquisition instrument continuously record the temperature changes inside the aging furnace, generating temperature-time curves to obtain process parameters.

[0026] S2. Extract the physical features related to effective accumulated temperature from the process parameters and combine them with the raw material data to form a joint feature set.

[0027] S2.1 It should be noted that the temperature-time curve of the entire aging heat treatment process is extracted from the process parameters; using the Arrhenius equation in the aluminum alloy precipitation kinetic model, the temperature-time curve is numerically integrated over the time interval of the entire aging process. The result of the numerical integration is the effective accumulated temperature, and the numerical integration formula is expressed as: ; in, Indicates effective accumulated temperature. Indicates the maximum number of points. This represents the natural exponential function. Indicates the precipitation activation energy. This represents the gas constant (e.g., 8.314). This indicates the time during the aging heat treatment process. The absolute temperature value of the aluminum alloy workpiece. Indicates time Perform differentiation; The precipitation activation energy is obtained by performing DSC (Differential Scanning Calorimetry) tests on aluminum alloy workpieces at different heating rates (e.g., 5, 10, and 20 K / min), forming characteristic peak temperatures, and then linearly fitting the Kissinger equation. The formula is as follows: ; in, Indicates the heating rate. Indicates the characteristic peak temperature. Represents the gas constant. Represents the constant term of the equation. Indicates the precipitation activation energy. This represents the natural logarithm function.

[0028] Differential scanning calorimetry (DSC) is a thermal analysis method that studies physical transformations and chemical reactions in samples by measuring changes in heat flow under a controlled temperature program. In the field of aluminum alloys, DSC is an important experimental method for obtaining precipitation kinetic parameters (such as precipitation activation energy).

[0029] S2.2 It should be noted that the effective accumulated temperature is aligned with the corresponding weight percentages of magnesium, silicon, and copper in the same batch to form a row vector, where the first column is the weight percentage of magnesium, the second column is the weight percentage of silicon, the third column is the weight percentage of copper, and the fourth column is the effective accumulated temperature; a corresponding row vector is generated for each production batch, and the row vectors of all batches are stacked vertically to form a two-dimensional numerical matrix, which is the joint feature set.

[0030] S3. Identify and classify the features in the joint feature set into feature subsets.

[0031] It should be noted that, based on the precipitation kinetics mechanism of aluminum alloys, the weight percentages of magnesium and silicon in the joint feature set are identified as the necessary solute element concentration characteristics for the formation of the main strengthening phase; the effective accumulated temperature is identified as the process parameter characteristic controlling the evolution kinetics of the precipitated phase; and the weight percentages of magnesium, silicon, and effective accumulated temperature are considered as a subset of precipitation strengthening characteristics.

[0032] Based on the theory of recrystallization and grain growth in aluminum alloys, the weight percentage of copper is identified as an alloying element characteristic that inhibits grain boundary migration through the solute drag effect; the extrusion speed and barrel temperature are identified as process parameter characteristics; and the weight percentage of copper, extrusion speed, and barrel temperature constitute a subset of grain structure characteristics.

[0033] S4. After constructing a dual-channel data processing path, use an independent fully connected layer to perform nonlinear physical mapping on the feature subset to generate a high-order physical state vector.

[0034] S4.1 Establish parallel dual-channel data processing paths for the precipitation strengthening feature subset and the grain structure feature subset.

[0035] It should be noted that a first independent data processing path (including an input layer, a fully connected layer, and an output layer) is created for the precipitation enhancement feature subset. This data processing path is designed specifically to receive and transmit the precipitation enhancement feature subset. At the same time, a second independent data processing path (including an input layer, a fully connected layer, and an output layer) is created for the grain structure feature subset. This data processing path is designed specifically to receive and transmit the grain structure feature subset. The first and second data processing paths maintain a parallel and independent relationship in structure. The two data processing paths do not share computational weights and intermediate results, and together they constitute a dual-channel data processing path.

[0036] S4.2 Using a dual-channel data processing path, the enhanced feature subset and the grain structure feature subset are transmitted to independent fully connected layers, and nonlinear transformations are performed to obtain the precipitate phase state vector and the grain structure state vector. It should be noted that after the precipitation strengthening feature subset and the grain structure feature subset are transmitted to the corresponding independent fully connected layer through the corresponding data processing path, a linear transformation is performed on the precipitation strengthening feature subset and the grain structure feature subset according to the weight matrix and bias vector in the independent fully connected layer. Specifically, the precipitation strengthening feature subset and the grain structure feature subset are multiplied by the weight matrix of the corresponding fully connected layer, and then the bias vector is added to obtain two linear transformation results. These two linear transformation results are nonlinearly activated through the ReLU activation function to output the precipitation phase state vector and the grain structure state vector. The precipitation phase state vector and the grain structure state vector are concatenated to form a higher-order physical state vector.

[0037] S5. Using the high-order physical state vector as input and product performance data as the prediction target, train the Bayesian neural network model to obtain a physical constraint hybrid model.

[0038] S5.1 Input the high-order physical state vector into the Bayesian neural network model for forward propagation to obtain physical parameters and product performance prediction values.

[0039] It should be noted that after the higher-order physical state vector is input into the Bayesian neural network model (the inference method is a probabilistic inference process), it is calculated layer by layer from the first hidden layer to the Nth hidden layer. In each hidden layer, the higher-order physical state vector is multiplied by the weight matrix, a bias vector is added, and then a nonlinear transformation is performed through the ReLU activation function. When the calculation reaches the Kth hidden layer of the Bayesian neural network model, the output vector of the Kth hidden layer is extracted as a physical parameter. At the same time, the output vector continues to propagate forward to the output layer of the Bayesian neural network model, and the output layer outputs the product performance prediction value. Here, K is an integer greater than 1 and less than N.

[0040] S5.2 Calculate the physical constraint loss term based on the theoretical physical equations.

[0041] It should be noted that, based on the theory of aluminum alloy precipitation kinetics, the theoretical precipitate fraction is calculated using the JMAK precipitation kinetic equation, and the formula is as follows: ; in, Indicates time The fraction of the precipitated phase, This represents the natural exponential function. Indicates the precipitation rate. Indicates the gestation period after separation. This represents the Avrami index.

[0042] The JMAK precipitation kinetic equation is a well-known kinetic model in the field of materials science used to describe phase transitions and precipitation processes.

[0043] The precipitation incubation period, also known as the "preparation phase" or "delay time" before the precipitation of the phase nucleus, is an important characteristic parameter in the precipitation kinetics process. It is introduced into the JMAK precipitation kinetic equations to more accurately describe this real-world physical process.

[0044] The Avrami index is a core kinetic parameter in the JMAK precipitation kinetic equation. It does not directly represent a physical quantity, but rather reveals a characteristic index of the precipitated nuclei and growth microscopic mechanisms. The process of obtaining it is as follows: measuring the phase transition fraction at different time points during the isothermal phase transition, then linearizing the JMAK precipitation kinetic equation by taking a double logarithm, and finally obtaining the Avrami index through linear regression.

[0045] Based on the grain state parameters, the theoretical grain size is calculated using the recrystallization kinetic equation, as follows: ; in, Indicates time Theoretical grain size, Indicates the initial grain size before recrystallization. This indicates the maximum increase in grain size after recrystallization. This represents the kinetic constant related to the recrystallization rate. This indicates the recrystallization kinetic index.

[0046] The kinetic constants and recrystallization kinetic index were obtained based on isothermal annealing experiments and data fitting. Specifically, aluminum alloy samples with the same deformation were subjected to isothermal annealing for different durations at multiple constant temperatures, and the recrystallized grain size at each time point was measured by metallographic image analysis. The measured grain size data were substituted into the recrystallization kinetic equation, and the kinetic constants and recrystallization kinetic index at each temperature were fitted using the nonlinear least squares method. The nonlinear least squares fitting algorithm was iteratively optimized to ensure that the change in the sum of squares of the residuals between the measured and theoretical grain sizes in two adjacent iterations was less than the convergence threshold, thus outputting the optimal kinetic constants and recrystallization kinetic index.

[0047] The mean square error between the precipitated phase state parameters and the theoretical precipitated phase fraction is calculated to obtain the precipitated phase constraint loss term; the mean square error between the grain state parameters and the theoretical grain size is calculated to obtain the grain constraint loss term; the weighted sum of the precipitated phase constraint loss term and the grain constraint loss term is obtained to obtain the physical constraint loss term.

[0048] The convergence threshold is set based on the measurement error between the measured and theoretical grain sizes. The value is the standard deviation of the measurement error, and its value is optimized during the iteration process based on the square of the residual between the measured and theoretical grain sizes.

[0049] S5.3 Calculate the data fitting loss term based on the predicted product performance value and the product performance data.

[0050] It should be noted that the predicted values ​​of product performance (including predicted tensile strength, predicted yield strength, and predicted elongation) are extracted; product performance data (including tensile strength, yield strength, and elongation) from the historical production process of aluminum alloys are extracted; the squared difference between the predicted tensile strength and the predicted yield strength, the squared difference between the predicted yield strength and the predicted elongation and the predicted elongation are calculated respectively; the average value of the sum of the squared differences in tensile strength, yield strength, and elongation is taken to obtain the data fitting loss term.

[0051] S5.4. The total loss function value is obtained by weighted summation of the data fitting loss term and the physical constraint loss term. The weight parameter distribution of the Bayesian neural network model is optimized by backpropagation algorithm to minimize the total loss function value and obtain the physical constraint hybrid model.

[0052] It should be noted that the data fitting loss term and the physical constraint loss term are weighted and summed to obtain the total loss function value. Based on the total loss function value, the partial derivatives with respect to the weight parameters and bias parameters of each layer of the Bayesian neural network model are calculated, i.e., the gradients of the weight parameters and the gradients of the bias parameters. The specific process is as follows: The formula for calculating the error vector of a Bayesian neural network model is as follows: ; in ， This represents the error vector of the output layer of a Bayesian neural network model. This represents the total loss function value. Represents the partial differential symbol. Indicates the activation output vector. This represents the first derivative of an activation function (such as ReLU). This indicates that the output vector is not activated. This represents the symbol for element-wise multiplication.

[0053] The inactive output vector is obtained through a linear transformation. Specifically, for each layer of the Bayesian neural network model, the input vector is multiplied by the weight parameters of the corresponding layer, and then the bias parameters are added to obtain the inactive output vector of the corresponding layer.

[0054] The activated output vector is obtained by applying a nonlinear activation function to the inactive output vector. That is, each element in the inactive output vector is input into an activation function (such as ReLU) for nonlinear transformation, and the result of the transformation is the activated output vector of the corresponding layer.

[0055] The formula for calculating the partial derivatives of the weight parameters in a Bayesian neural network model is: ; in, This represents the partial derivative of the weighting parameters. This indicates the Bayesian neural network model's first... The error vector of the layer, This indicates the Bayesian neural network model's first... Layer activation output vector, This indicates the matrix transpose.

[0056] The formula for calculating the partial derivatives of the bias parameters of a Bayesian neural network model is as follows: ; in, This represents the partial derivative of the bias parameter. This indicates the Bayesian neural network model's first... Error vector of the layer.

[0057] Since the bias parameters act directly on the unactivated linear output in an additive manner, the partial derivatives of the bias parameters are equal to the error vector of the output layer of the Bayesian neural network model.

[0058] The partial derivatives of the weight parameters and the partial derivatives of the bias parameters are used as the gradients of the weight parameters and the bias parameters, respectively. Based on the calculated gradients (weight parameter gradients and bias parameter gradients), the distributions of the weight parameters and bias parameters of the Bayesian neural network model are updated using the stochastic gradient descent optimization algorithm. The process of calculating the total loss function value through forward propagation and updating the parameter distribution through backpropagation is performed iteratively. Iteration stops when the maximum number of iterations is reached (e.g., 20,000 times). The Bayesian neural network model trained at this point is the physically constrained hybrid model.

[0059] The maximum number of iterations is set based on the convergence characteristics of the loss function and the constraints of computational resources during the training of the Bayesian neural network model. It is used to terminate the training process when the loss function has basically converged but the benefits of continuing to iterate are limited.

[0060] S6. Calculate the gradient information inside the physical constraint hybrid model to obtain the uncertainty contribution vector.

[0061] It should be noted that, using the method described above for obtaining higher-order physical state vectors, new higher-order physical state vectors are re-acquired and input into the physical constraint hybrid model for forward propagation calculation (the forward propagation calculation process is the same as step S5.1) to obtain the probability distribution of the product performance prediction values. The probability distribution of the product performance prediction values ​​is jointly defined by the prediction mean and the prediction variance. The prediction variance is directly extracted from the probability distribution of the product performance prediction values, and the arithmetic square root of the prediction variance is calculated to obtain the standard deviation parameter. The standard deviation parameter is used as the uncertainty value output by the physical constraint hybrid model. For multiple output predictions (such as tensile strength prediction, yield strength prediction, and elongation prediction), their respective uncertainty values ​​are calculated separately.

[0062] The output layer of the physical constraint hybrid model provides the product performance prediction results in the form of a probability distribution. The probability distribution is determined by the prediction mean and the prediction variance, and the prediction variance is the direct output result of the output layer of the physical constraint hybrid model.

[0063] The extrusion speed, barrel temperature, die temperature, cooling medium flow rate, aging temperature-time curve process parameters, and the weight percentages of magnesium, silicon, and copper are used as input variables. The uncertainty value output by the physical constraint hybrid model is used as the output variable. The partial derivatives of the uncertainty value with respect to each input variable are calculated using the inverse mode differentiation method in the automatic differentiation method to obtain the original gradient vector, as shown in the formula: ; in, Represents partial derivatives, Represents the partial differential symbol. Indicates the value of uncertainty. Indicates the first One input variable, The output layer of the physical constraint hybrid model represents the first... The prediction variance of each neuron. This indicates the number of uncertainty values ​​output by the physical constraint hybrid model.

[0064] Take the absolute value of each gradient component in the original gradient vector to obtain the gradient absolute value vector; sum all components in the gradient absolute value vector to obtain the sum of absolute values; perform a quotient operation between the gradient absolute value vector and the sum of absolute values ​​to achieve normalization, obtaining the normalized gradient absolute value vector. Use the normalized gradient absolute value vector as the uncertainty contribution vector.

[0065] S7. Set the optimization objective, and use the physical constraint hybrid model as the evaluation function and the uncertainty contribution vector as the constraint condition to find the Pareto optimal solution set in the process parameter space.

[0066] It should be noted that the objectives include maximizing the mean performance predicted by the physical constraint mixture model, minimizing the performance uncertainty predicted by the physical constraint mixture model, and minimizing the production energy consumption; at the same time, the physical constraint mixture model is used as the evaluation function. The extrusion speed, barrel temperature, die temperature, quenching cooling medium flow rate, and aging temperature-time curve constitute the process parameter space. An initial candidate solution set is generated within the process parameter space by random sampling, forming an initial population. Each candidate solution in the initial population represents a specific combination of process parameters (including the specific values ​​of extrusion speed, barrel temperature, die temperature, quenching cooling medium flow rate, and aging temperature-time curve).

[0067] Based on the three objective function values ​​of the predicted tensile strength mean, predicted tensile strength uncertainty, and production energy consumption value corresponding to each candidate solution in the initial population, a fast non-dominated sorting method is used to perform hierarchical sorting of the initial population, and the crowding distance of each candidate solution is calculated. Specifically, for each candidate solution in the initial population, all other candidate solutions in the initial population are traversed, and the three objective function values ​​of the predicted tensile strength mean, predicted tensile strength uncertainty, and production energy consumption value are compared. If a candidate solution is not inferior to another candidate solution in all objective function values, and is strictly superior in at least one objective function value, then this candidate solution dominates the other candidate solution. The number of other candidate solutions dominated by each candidate solution is counted, called the domination count; at the same time, the set of other candidate solutions that dominate each candidate solution is recorded, called the dominated set; and the support is then divided into... Candidate solutions with a dominance count of zero are assigned to the first Pareto front and given the highest rank. For stratified candidate solutions, the dominance count of the candidate solutions in the dominated set is reduced. When the dominance count of the candidate solutions in the dominated set drops to zero, the candidate solution is assigned to the next Pareto front. This process is repeated until all candidate solutions are stratified. For all candidate solutions assigned to the same Pareto front, they are sorted in ascending order according to the three objective function values: the predicted mean of tensile strength, the predicted uncertainty of tensile strength, and the production energy consumption value. For each objective function value, the candidate solutions one position before and one position after the current candidate solution are identified as adjacent solutions. The absolute difference between the current candidate solution and its adjacent solutions at each objective function value is calculated and normalized. The three normalized absolute differences are summed, and the sum is the crowding distance of the current candidate solution.

[0068] The Pareto rank is a rank identifier obtained by classifying candidate solutions based on the multi-objective dominance relationship between them using the fast non-dominated ranking method. The smaller the Pareto rank value, the higher the overall superiority of the candidate solution in the multi-objective sense.

[0069] Based on the Pareto scores and crowding distances of candidate solutions, a binary tournament selection mechanism is used to select superior candidate solutions from the initial population as parent individuals. Specifically, two candidate solutions are randomly selected from the initial population as participants; the Pareto scores of the two participating candidate solutions are compared, and the candidate solution with the smaller Pareto score is selected as the winning individual; if the Pareto scores of the two participating candidate solutions are the same, the crowding distances of the two participating candidate solutions are compared, and the candidate solution with the larger crowding distance is selected as the winning individual; the winning candidate solution is added to the parent individual set; the steps of random selection, comparing Pareto scores, comparing crowding distances, and selecting the winning individual are repeated until the size of the parent individual set reaches a preset population size (e.g., 100); the preset population size is set according to multi-objective optimization theory.

[0070] A simulated binary crossover operator is applied to the selected parent individuals to exchange some process parameter genes with crossover probability, generating new offspring candidate solutions. A polynomial mutation operator is applied to the offspring candidate solutions to randomly perturb the process parameter gene values ​​with mutation probability, introducing population diversity. The new candidate solutions generated by selection, crossover, and mutation operations are merged with the initial population to form a temporary population. After re-sorting the temporary population for non-dominated ordering and calculating crowding, the iterative process of selection, crossover, and mutation is repeated until the maximum number of iterations (e.g., 500 times) is reached, obtaining a candidate population that approximates the true Pareto front.

[0071] The crossover and mutation probabilities are set according to the balance principle between maintaining population diversity and convergence efficiency in evolutionary algorithms. For example, the crossover probability is set to 0.9 and the mutation probability is set to 0.1. The values ​​are chosen to enhance the gene recombination ability between parent candidate solutions and accelerate Pareto front convergence.

[0072] The maximum number of iterations is set based on the iterative characteristics of the Pareto front gradually converging during multi-objective evolution, and the value is determined by ensuring the stability of the candidate solution set while also taking into account the consumption of computational resources.

[0073] A candidate solution is selected from the candidate population (including specific values ​​for extrusion speed, barrel temperature, die temperature, quenching cooling medium flow rate, and aging temperature-time curve parameters). The process parameters from the candidate solution are combined with raw material data, including the weight percentages of magnesium, silicon, and copper. The combined complete input data is then input into a physically constrained hybrid model for forward propagation calculation. The physically constrained hybrid model outputs three performance averages: the predicted mean of tensile strength, the predicted mean of yield strength, and the predicted mean of elongation, along with the corresponding predicted mean of tensile strength. The system calculates three performance uncertainties: certainty, yield strength prediction uncertainty, and elongation prediction uncertainty. Simultaneously, it calculates production energy consumption based on the aging temperature-time curve in the process parameters. Specifically, the entire aging process is discretized according to the aging temperature-time curve, dividing the aging time interval into several time steps (e.g., 1s or 10s). For each discrete moment, the corresponding measured temperature is read, and the instantaneous power (the product of the rated heating power of the aging furnace and the temperature correction coefficient) is calculated. The instantaneous power is numerically integrated to obtain the total electrical energy consumption. The total electrical energy consumption is then divided by the total mass of the aluminum alloy workpiece to obtain the production energy consumption value.

[0074] The temperature correction coefficient is obtained based on the measured temperature at each discrete moment. Specifically, the actual average power consumption corresponding to the measured temperature is recorded based on the measured temperature at each discrete moment (e.g., 100℃, 150℃, 200℃, 250℃, 300℃, etc.), and a set of actual power corresponding to different temperature points is obtained. The ratio of the actual power at each temperature point to the rated heating power of the furnace is calculated to obtain the temperature correction coefficient.

[0075] Repeat the above combination, input, and calculation steps for each candidate solution in the candidate population until all candidate solutions have been calculated, and obtain a complete set of performance mean values, performance uncertainty values, and production energy consumption values.

[0076] The components of the uncertainty contribution vector are sorted by numerical value, and the process parameter corresponding to the component with the largest numerical value in the sorting result is determined as the process parameter with the largest contribution to uncertainty. In the mutation operation of the multi-objective evolutionary algorithm, a stronger mutation perturbation is applied to the identified process parameter with the largest contribution to uncertainty. Specifically, by increasing the mutation probability or mutation magnitude of the process parameter (with the largest contribution to uncertainty), the search is guided to explore regions with lower uncertainty. At the same time, based on the set of performance mean values, the set of performance uncertainty values, and the set of production energy consumption values, Pareto rank is divided for the candidate population through fast non-dominated sorting and crowding calculation. Based on the Pareto rank and crowding distance, a binary tournament selection mechanism is used to screen excellent candidate solutions. New solutions are generated by applying simulated binary crossover operator and polynomial mutation operation to the selected candidate solutions (consistent with the above selection, crossover, and mutation process). The selection, crossover, and mutation process is executed iteratively to obtain a Pareto optimal solution set that is balanced and optimized on the three objectives of tensile strength prediction mean, tensile strength prediction uncertainty, and production energy consumption value.

[0077] S8. After visualizing the Pareto optimal solution set, select the optimal combination of process parameters.

[0078] It should be noted that a three-dimensional Cartesian coordinate system is established with the tensile strength prediction mean as the X-axis, the tensile strength prediction uncertainty as the Y-axis, and the production energy consumption value as the Z-axis. The tensile strength prediction mean, tensile strength prediction uncertainty, and production energy consumption value of each candidate solution in the Pareto optimal solution set are located in the three-dimensional coordinate system, and a scatter plot is generated to form the Pareto front surface. Three slider controls are displayed next to the three-dimensional coordinate system, labeled as the tensile strength prediction mean preference weight slider, the tensile strength prediction uncertainty preference weight slider, and the production energy consumption value preference weight slider. Users can set the specific values ​​of the tensile strength prediction mean preference weight, the tensile strength prediction uncertainty preference weight, and the production energy consumption value preference weight by dragging the sliders. The graphical interface responds to user operations in real time and dynamically updates the display of preference weight values.

[0079] The user-defined weights for the tensile strength prediction mean, tensile strength prediction uncertainty, and production energy consumption are normalized to ensure that the sum of these three weights equals 1. For each candidate solution in the Pareto optimal solution set, the normalized weight for the tensile strength prediction mean is multiplied by the candidate solution's tensile strength prediction mean; the normalized weight for the tensile strength prediction uncertainty is multiplied by the negative value of the candidate solution's tensile strength prediction uncertainty (since uncertainty is the minimization objective); and the normalized weight for the production energy consumption is multiplied by the negative value of the candidate solution's production energy consumption (since production energy consumption is the minimization objective). The results of these three multiplications are added together to obtain the comprehensive evaluation value of the candidate solution. By calculating and comparing the comprehensive evaluation values ​​of all candidate solutions in the Pareto optimal solution set, the candidate solution with the highest comprehensive evaluation value is determined. The candidate solution with the highest comprehensive evaluation value is determined as the optimal combination of process parameters.

[0080] This embodiment also provides a computer device applicable to the machine learning-based aluminum alloy process parameter control method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the machine learning-based aluminum alloy process parameter control method proposed in the above embodiment.

[0081] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0082] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the method for controlling aluminum alloy process parameters based on machine learning as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0083] In summary, this invention achieves a physical quantitative characterization of the process by extracting the physical feature of "effective accumulated temperature" related to aging precipitation kinetics and constructing a joint feature set with metallurgical significance by combining it with the raw material composition. Furthermore, based on the strengthening mechanism, the features are decoupled into two subsets: precipitation strengthening and grain structure, effectively capturing the nonlinear laws of multi-scale microstructure evolution. On this basis, the JMAK precipitation kinetic equation (Johnson-Mel-Avramy-Kolmogorov precipitation kinetic equation) and the physical constraint loss term of recrystallization kinetics are introduced to train a Bayesian neural network to form a physical constraint hybrid model with both high-precision prediction capability and uncertainty quantification function. This supports the efficient search for Pareto optimal solutions that balance high performance, high robustness, and low energy consumption in the process parameter space, thereby realizing intelligent, refined, and reliable control of aluminum alloy hot working processes.

[0084] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for controlling aluminum alloy process parameters based on machine learning, characterized in that: include, Collect product performance data, raw material data, and process parameters from the historical production process of aluminum alloys; Extract the physical features related to effective accumulated temperature from the process parameters and combine them with the raw material data to form a joint feature set; The joint feature set is identified and classified into feature subsets; After constructing a dual-channel data processing path, an independent fully connected layer is used to perform nonlinear physical mapping on the feature subset to generate a high-order physical state vector. Using high-order physical state vectors as input and product performance data as prediction targets, a Bayesian neural network model is trained to obtain a hybrid physical constraint model. Calculate the internal gradient information of the physical constraint hybrid model to obtain the uncertainty contribution vector; An optimization objective is set, and a physical constraint hybrid model is used as the evaluation function, with the uncertainty contribution vector as the constraint condition. The Pareto optimal solution set is then searched in the process parameter space. After visualizing the Pareto optimal solution set, the optimal combination of process parameters is selected.

2. The method for controlling aluminum alloy process parameters based on machine learning as described in claim 1, characterized in that: The product performance data includes tensile strength, yield strength, and elongation. The raw material data includes the weight percentages of magnesium, silicon, and copper. The process parameters include the extrusion speed, barrel temperature and die temperature during the extrusion stage, the cooling medium flow rate during the quenching stage, and the temperature-time curve of the entire aging heat treatment process.

3. The method for controlling aluminum alloy process parameters based on machine learning as described in claim 2, characterized in that: The physical feature related to effective accumulated temperature in the extraction process parameters refers to the use of an aluminum alloy precipitation kinetic model to perform integral calculations on the temperature-time curve in the process parameters to obtain the effective accumulated temperature, which is then used as a newly added physical feature.

4. The method for controlling aluminum alloy process parameters based on machine learning as described in claim 3, characterized in that: The combination of the new physical features with the raw material data to form a joint feature set refers to the horizontal combination of the newly added physical features with the weight percentages of magnesium, silicon and copper in the raw material data to form a joint feature set.

5. The method for controlling aluminum alloy process parameters based on machine learning as described in claim 4, characterized in that: The process of identifying and classifying features in the joint feature set into feature subsets specifically involves: Features strongly correlated with precipitation dynamics in the joint feature set are classified as a subset of precipitation enhancement features; Features that are strongly correlated with grain morphology evolution in the joint feature set are classified as a subset of grain structure features.

6. The method for controlling aluminum alloy process parameters based on machine learning as described in claim 5, characterized in that: After constructing the dual-channel data processing path, an independent fully connected layer is used to perform nonlinear physical mapping on the feature subset to generate a high-order physical state vector, specifically: For the precipitation strengthening feature subset and the grain structure feature subset, a parallel dual-channel data processing path is established; Using a dual-channel data processing path, the enhanced feature subset and the grain structure feature subset are transmitted to independent fully connected layers and nonlinear transformations are performed to obtain the precipitate phase state vector and the grain structure state vector. The precipitated phase state vector and the grain structure state vector are concatenated to form a higher-order physical state vector.

7. The method for controlling aluminum alloy process parameters based on machine learning as described in claim 6, characterized in that: The process involves training a Bayesian neural network model using a high-order physical state vector as input and product performance data as the prediction target to obtain a hybrid physical constraint model. Specifically: The higher-order physical state vector is input into the Bayesian neural network model for forward propagation to obtain physical parameters and product performance prediction values. Calculate the physical constraint loss term using theoretical physics equations; The data fitting loss term is calculated based on the predicted product performance values ​​and the product performance data. The total loss function value is obtained by weighted summation of the data fitting loss term and the physical constraint loss term. The weight parameter distribution of the Bayesian neural network model is then optimized using the backpropagation algorithm to minimize the total loss function value, resulting in a hybrid physical constraint model.

8. The method for controlling aluminum alloy process parameters based on machine learning as described in claim 7, characterized in that: The process of calculating the gradient information within the physical constraint hybrid model to obtain the uncertainty contribution vector is as follows: Input a set of process parameters and raw material data into the physical constraint hybrid model, and perform forward propagation calculations to obtain the uncertainty value; The automatic differentiation method is used to calculate the gradient of the uncertainty value with respect to process parameters and raw material data, and the absolute value of each gradient is taken and normalized. Arrange all the normalized absolute gradient values ​​to form a sequence of absolute gradient values ​​and use it as an uncertainty contribution vector.

9. The method for controlling aluminum alloy process parameters based on machine learning as described in claim 8, characterized in that: The optimization objective is set, and a hybrid physical constraint model is used as the evaluation function, with the uncertainty contribution vector as the constraint condition. Specifically, the Pareto optimal solution set is found in the process parameter space: The optimization objectives are set as maximizing the mean performance predicted by the physical constraint hybrid model, minimizing the performance uncertainty predicted by the physical constraint hybrid model, and minimizing production energy consumption, with the physical constraint hybrid model as the evaluation function. A multi-objective evolutionary algorithm is used to generate a candidate population within the process parameter space through initialization, selection, crossover, and mutation operations. After combining each candidate population with the raw material data, the data is input into the physical constraint hybrid model for calculation, resulting in a set of average performance values, a set of performance uncertainty values, and a set of production energy consumption values. Based on the set of average performance values, the set of performance uncertainty values, the set of production energy consumption values, and the uncertainty contribution vector, the process parameters that contribute the most to uncertainty are identified. Through iterative selection, crossover, and mutation, the Pareto optimal solution set is obtained.

10. The method for controlling aluminum alloy process parameters based on machine learning as described in claim 9, characterized in that: The process of visualizing the Pareto optimal solution set and then selecting the optimal combination of process parameters involves the following steps: Plot the Pareto optimal solution set in a three-dimensional coordinate system and receive the user-defined performance mean, performance uncertainty value and production energy consumption value preference weights in the graphical interface. The comprehensive evaluation value of each solution in the Pareto optimal solution set is calculated based on the preference weights, and the solution with the highest comprehensive evaluation value is selected as the optimal combination of process parameters.

Images