An aluminum alloy laser self-fusion welding process based on bayesian and neural network collaborative optimization and a method thereof

Abstract

The application discloses an aluminum alloy laser self-melting welding process based on Bayesian and neural network cooperative optimization and a method thereof, and relates to the technical field of intelligent laser welding manufacturing.The initial data set of 6061 aluminum alloy laser welding is obtained through experiment design; a comprehensive evaluation function is constructed to convert the weld forming coefficient deviation and microhardness into a single optimization target; in the model construction stage, the Bayesian hyperparameter optimization strategy can be introduced to adaptively configure the MLP, or the MLP prediction model can be constructed by reusing the pre-determined hyperparameter combination; the nonlinear relationship between the process parameters and the welding quality is fitted by using the MLP model under the optimal hyperparameters; and then the global iterative optimization of the welding process parameters is carried out in the continuous parameter space by using the Gaussian process and the acquisition function.The double optimization strategy is adopted to reduce the trial and error cost, reduce the dependence on artificial experience, and improve the process stability of the preparation of high-quality welded joints.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology for laser welding, and in particular to a laser self-fusion welding process and method for aluminum alloys based on Bayesian and neural network co-optimization. Background Technology

[0002] 6061 aluminum alloy, a typical Al-Mg-Si heat-treatable aluminum alloy, is widely used in aerospace and automotive lightweighting fields due to its high specific strength and excellent corrosion resistance. However, laser welding of 6061 aluminum alloy faces technical challenges such as high reflectivity, low melting point, and susceptibility to porosity and cracking. Especially in laser self-fusion welding (without filler wire) mode, defects such as weld surface collapse, incomplete penetration, and joint softening are prone to occur, seriously affecting the service performance of components.

[0003] Existing methods for determining process parameters mainly rely on orthogonal experiments or full-factor experiments (trial and error). While reliable, these methods become exponentially more demanding when dealing with nonlinear systems involving multiple coupled variables such as laser power, welding speed, oscillation frequency, and oscillation amplitude. Furthermore, they fail to find a globally optimal solution within a continuous parameter space. For instance, obtaining a high-precision process window often requires dozens or even hundreds of physical welding experiments. This not only incurs high material costs and energy consumption but also results in extremely long development cycles due to the cumbersome post-weld metallographic sample preparation and mechanical testing processes.

[0004] In recent years, although genetic algorithms (GA) and particle swarm optimization (PSO) have been introduced into the welding field, these algorithms are inherently data-hungry, typically requiring a large amount of sample data to converge, making them unsuitable for small-sample laser welding experimental data. Therefore, there is an urgent need for an intelligent method with high sample efficiency that can achieve accurate modeling and global optimization with small sample data to solve the quality control problem of laser self-fusion welding of 6061 aluminum alloy. Summary of the Invention

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a laser self-fusion welding process and method for aluminum alloys based on Bayesian and neural network co-optimization, comprising the following steps: Step S1: Construct an initial process parameter dataset for laser autofusion welding of 6061 aluminum alloy; the process parameters include at least laser power, welding speed, oscillation frequency, and oscillation amplitude; the response values ​​of the dataset include weld formation coefficient and microhardness; the weld formation coefficient is the ratio of weld width to weld depth, wherein the weld width is the maximum width of the fusion zone in the weld cross-section, and the weld depth is the maximum depth of the fusion zone along the plate thickness direction; Step S2: Construct a comprehensive evaluation function for welding quality, and calculate the weld formation coefficient and the preset ideal weld formation coefficient. The absolute deviation is used to calculate the welding quality score for a single target, combined with microhardness. During model training, the measured weld formation coefficient W and measured microhardness H are used as supervision labels to train the MLP prediction model. During parameter optimization, the predicted values ​​output by the MLP are used as the basis for the calculation. and Calculate the welding quality score or the equivalent objective function to evaluate candidate process parameters; Step S3: Construct and train the MLP neural network prediction model; In the model construction stage, the hyperparameters of the MLP neural network can be adaptively optimized using a Bayesian optimization algorithm, or a pre-determined combination of hyperparameters can be used to construct the MLP neural network model. The configured MLP neural network is then used to fit the nonlinear mapping relationship between welding process parameters, weld formation coefficient, and microhardness to output the predicted value of the weld formation coefficient. Compared with the predicted value of microhardness ; Step S4: Globally optimize welding process parameters based on Bayesian optimization algorithm; randomly select several initial seed points from preset discrete grid points; the discrete grid points are used to initialize the continuous parameter space within the preset process parameter range to achieve cold start of Bayesian optimization; use a Gaussian process as a surrogate model to fit the posterior distribution of the objective function, and use the acquisition function to recommend the next set of candidate process parameters in the parameter space; wherein, the initial observation output of the Gaussian process surrogate model is the evaluation value corresponding to the initial seed point, and the evaluation value can be composed of virtual evaluation value calculated by MLP prediction value and / or measured evaluation value obtained by physical welding experiment; Step S5: Perform virtual iterative optimization based on the surrogate model; randomly generate a large number of candidate points in the continuous parameter space; calculate the acquisition function value of the candidate points using the acquisition function, which includes, but is not limited to, the expected improvement in EI or the upper confidence bound UCB; select the point with the highest acquisition function value as the next set of virtual evaluation points; use the MLP neural network prediction model to obtain the predicted weld formation coefficient value of the point. Compared with the predicted value of microhardness And calculate the predicted quality score Score accordingly; use this point and its virtual evaluation value calculated from the predicted value to update the posterior distribution of the Gaussian process surrogate model; wherein, the virtual evaluation value is the... Or, it can be an objective function equivalent to Score; the virtual evaluation value is a single scalar used as the observed output of Gaussian process regression; wherein, the virtual evaluation value is used as a low-confidence observation with preset observation noise as input to the Gaussian process surrogate model; the preset observation noise is used to characterize the prediction error between the virtual evaluation value and the actual measured value, and can be set as a constant; in one implementation, it can also be set with reference to cross-validation error or historical residuals; the update is used to improve the recommendation effect of the acquisition function for the next set of candidate process parameters, and the virtual evaluation value is the evaluation value calculated from the MLP prediction value, which is not equivalent to the actual weld measurement value; the iteration ends when the preset iteration round is reached; Step S6: Output the verified global optimal process parameter combination to the user interface, so that the operator can set and verify the process of laser self-fusion welding of 6061 aluminum alloy based on the parameters; optionally, compare the measured welding quality data obtained from the verification with the prediction results to evaluate the prediction error. This invention does not simply superimpose Bayesian optimization and neural network models, but rather introduces a virtual evaluation value mechanism with preset observation noise to achieve explicit modeling of prediction uncertainty by the surrogate model, thereby improving the stability and globality of process parameter optimization under small sample conditions.

[0006] Furthermore, in step S1, the method for constructing the initial process parameter dataset includes: adopting a full factorial experimental design or an orthogonal experimental design scheme; collecting weld cross-sectional morphology data and microhardness data after the experiment; cleaning the collected data, removing abnormal data points with macroscopic porosity, incomplete penetration defects or severe surface collapse according to a preset defect judgment threshold, and standardizing the cleaned data.

[0007] Furthermore, in step S2, the comprehensive evaluation function for welding quality... The calculation formula is: ; In one implementation, it can be Maximize and rewrite as minimizing the objective function To adapt to a Bayesian optimization solver with minimization as the default setting, where: ; in, Microhardness value (unit: HV); is the weld formation coefficient, which is the ratio of weld width to weld depth, where weld width is the maximum width of the fusion zone in the weld cross section, and weld depth is the maximum depth of the fusion zone along the plate thickness direction; This is the ideal weld formation coefficient preset or calibrated based on the plate thickness and the desired weld penetration ratio; The hardness weighting coefficient is preferably set between 0.01 and 0.2, and can be calibrated based on the magnitude difference between the microhardness and the weld formation coefficient deviation term. This represents the penalty term for the absolute deviation between the weld formation coefficient and the preset ideal weld formation coefficient; the optimization objective is to find the parameter combination that maximizes the score value within the process parameter search space. Specifically, when constructing the training sample labels... and These are replaced by the measured microhardness H and the measured weld formation coefficient W, respectively; during the parameter optimization stage... and The predicted microhardness value output by the MLP model and predicted value of weld formation coefficient The physical meaning of this formula lies in maximizing microhardness while minimizing the weld formation coefficient compared to the preset ideal weld formation coefficient. The deviation between them.

[0008] Further, the specific implementation of step S3 is as follows: A search space for hyperparameters is defined; the mean square error of the welding quality score of the MLP model on the validation set, or the weighted combination of the weld formation coefficient and the microhardness prediction error, is used as the objective function for Bayesian optimization; the Gaussian process surrogate model and acquisition function are used to iteratively find the hyperparameter combination that minimizes the MSE within the search space; this optimal combination is then used to construct the final MLP prediction model. This step effectively avoids the subjectivity and blindness of manually setting the neural network structure, significantly improving the model's generalization ability. In practical applications, the hyperparameter optimization can be completed offline before system deployment to obtain a stable set of network structures and training parameters; in subsequent process parameter optimization, this set of hyperparameters can be directly reused to reduce computational overhead.

[0009] Furthermore, in step S4, the specific implementation of the Bayesian optimization algorithm includes: calculating the posterior mean and posterior variance of unsampled points using a Gaussian process regression model; constructing a sampling function and using an expectation enhancement (EI) strategy or a confidence upper bound (UCB) strategy to determine the next set of candidate process parameters.

[0010] Furthermore, the microhardness value ( The hardness of 6061 aluminum alloy is configured as a macroscopic proxy index reflecting the evolution of the microstructure of the weld joint. This invention utilizes the Hall-Petch relationship between the hardness and grain size of 6061 aluminum alloy, and under the globally optimal process parameters selected through the comprehensive evaluation function of welding quality, it can be used to obtain weld joints with a grain refinement trend and exhibiting an M-type microhardness distribution trend.

[0011] To achieve the above objectives, a second aspect of the present invention also proposes an adaptive optimization system for laser self-fusion welding process parameters of 6061 aluminum alloy based on Bayesian-neural network co-optimization, comprising: The experimental data entry and preprocessing module is configured to receive raw offline experimental data of laser self-fusion welding of 6061 aluminum alloy (e.g., 81 sets, including process parameters and corresponding weld formation coefficients and microhardness test values) through a user input interface, and to clean and standardize the data. The comprehensive welding quality assessment module is configured to integrate the multi-dimensional weld formation coefficient and microhardness index into a single-objective comprehensive welding quality score based on a weighted comprehensive evaluation function. The intelligent agent model construction and hyperparameter optimization module is configured to use a Bayesian hyperparameter optimization strategy to adaptively select the optimal topology and hyperparameters of the MLP neural network based on the cleaned effective dataset, and to build a high-precision process parameter-quality mapping model. The virtual optimization module for process parameters is configured to perform virtual iterative calculations in a continuous parameter space using Gaussian process regression and acquisition functions to discover the theoretically optimal global process parameters; and to input the virtual evaluation values ​​calculated from the MLP prediction values ​​as low-confidence observations with preset observation noise into the Gaussian process surrogate model to update its posterior distribution. The error assessment and feedback correction module is configured to compare the measured welding quality data with the predicted welding quality data, calculate the deviation, and determine whether it exceeds a preset threshold. In an optional implementation, when the threshold is exceeded, the model retraining and process parameter re-optimization process is triggered. The optimal parameter decision and visualization output module is configured to combine the verified global optimal process parameters and display them on a display device for visualization output, so that operators can read them and manually set the laser welding equipment based on the parameters.

[0012] To achieve the above objectives, a third aspect of the present invention also provides an electronic device, comprising: At least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the adaptive optimization method for laser self-fusion welding process parameters of 6061 aluminum alloy based on Bayesian-neural network co-optimization as described in any one of the first aspects above.

[0013] To achieve the above objectives, a fourth aspect of the present invention also provides a computer-readable storage medium. It stores a computer program, which, when executed by a processor, implements the adaptive optimization method for laser self-fusion welding process parameters of 6061 aluminum alloy based on Bayesian-neural network collaborative optimization as described in any of the first aspects above.

[0014] Compared with existing technologies, the beneficial effects of this invention are: Significantly reduces experimental costs: By leveraging the synergistic work of MLP neural networks and Bayesian optimization, it can quickly converge to the global optimum with only a small amount of initial experimental data, avoiding the trial-and-error costs of hundreds or thousands of trials in traditional full-factor experiments.

[0015] Solving the challenges of nonlinear modeling and model structure uncertainty: Through dual Bayesian optimization, the problem of determining hyperparameters in MLP neural networks is solved, ensuring the prediction accuracy of the model; and by using a data-driven black-box optimization strategy, welding quality can be accurately predicted without establishing complex physical equations.

[0016] Synergistic improvement of welding quality: By constructing a comprehensive evaluation function that includes weld formation coefficient and microhardness, without introducing additional mechanical testing costs, microhardness, a macroscopic proxy index that is highly correlated with various mechanical properties, is used to indirectly characterize the trend of comprehensive mechanical properties of welded joints. This allows for the consideration of both macroscopic forming quality and mechanical property evolution trends during the optimization of process parameters. Attached Figure Description

[0017] Figure 1 This is a flowchart of an aluminum alloy laser self-fusion welding process and method based on Bayesian and neural network co-optimization proposed in an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the principle of the Bayesian optimization algorithm in this invention; Figure 3 This is a flowchart of the hyperparameter optimization of the MLP neural network in this invention; Figure 4 This is a convergence curve of the objective function Score value during the Bayesian optimization iteration process in this embodiment of the invention; Figure 5 This is a macroscopic morphology diagram of the welded joint cross-section obtained using optimal process parameters in an embodiment of the present invention; Figure 6 These are metallographic images of the microstructure of each region of the welded joint obtained using optimal process parameters in an embodiment of the present invention. Figure 7 This is a microhardness distribution curve of the welded joint obtained using the optimal process parameters in an embodiment of the present invention; Figure 8 This is a schematic diagram of the module structure of the adaptive optimization system for process parameters provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0019] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0020] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. The invention will now be described in further detail with reference to the accompanying drawings.

[0021] Reference Figures 1 to 8 A laser self-fusion welding process and method for aluminum alloys based on Bayesian and neural network co-optimization includes the following steps: Step S1: Construct an initial process parameter dataset for laser autofusion welding of 6061 aluminum alloy; the process parameters include at least laser power, welding speed, oscillation frequency, and oscillation amplitude; the response values ​​of the dataset include weld formation coefficient and microhardness; the weld formation coefficient is the ratio of weld width to weld depth, wherein the weld width is the maximum width of the fusion zone in the weld cross-section, and the weld depth is the maximum depth of the fusion zone along the plate thickness direction; Step S2: Construct a comprehensive evaluation function for welding quality, and calculate the weld formation coefficient and the preset ideal weld formation coefficient. The absolute deviation is used to calculate the welding quality score of a single target based on microhardness; among which, during the model training phase, the measured weld formation coefficient is used. Compared with the measured microhardness The MLP prediction model is trained using the supervised labels; during the parameter optimization phase, the predicted values ​​output by the MLP are used as the supervisory labels. and Calculate welding quality score Alternatively, an equivalent objective function, Objective, can be used to evaluate candidate process parameters; Step S3: Construct and train the MLP neural network prediction model; in the model construction stage, the hyperparameters of the MLP neural network are adaptively optimized using a Bayesian optimization algorithm, or an MLP neural network model is constructed using a pre-determined combination of hyperparameters; and the configured MLP neural network is used to fit the nonlinear mapping relationship between welding process parameters, weld formation coefficient, and microhardness to output the predicted value of the weld formation coefficient. Compared with the predicted value of microhardness ; Step S4: Globally optimize welding process parameters based on Bayesian optimization algorithm; randomly select several initial seed points from preset discrete grid points; the discrete grid points are used to initialize the continuous parameter space within the preset process parameter range to achieve cold start of Bayesian optimization; use a Gaussian process as a surrogate model to fit the posterior distribution of the objective function, and use the acquisition function to recommend the next set of candidate process parameters in the parameter space; wherein, the initial observation output of the Gaussian process surrogate model is the evaluation value corresponding to the initial seed point, and the evaluation value can be composed of virtual evaluation value calculated by MLP prediction value and / or measured evaluation value obtained by physical welding experiment; Step S5: Perform virtual iterative optimization based on the surrogate model; randomly generate a large number of candidate points in the continuous parameter space; calculate the acquisition function value of the candidate points using the acquisition function, which includes, but is not limited to, the expected improvement in EI or the upper confidence bound UCB; select the point with the highest acquisition function value as the next set of virtual evaluation points; use the MLP neural network prediction model to obtain the predicted weld formation coefficient value of the point. Compared with the predicted value of microhardness And based on this, the predicted quality score is calculated. The point and its virtual evaluation value calculated from the predicted value are used to update the posterior distribution of the Gaussian process surrogate model; wherein, the virtual evaluation value is the Score, or an objective function equivalent to the Score; the virtual evaluation value is a single scalar used as the observed output of the Gaussian process regression; wherein, the virtual evaluation value is used as a low-confidence observation with preset observation noise as input to the Gaussian process surrogate model; the preset observation noise is used to characterize the prediction error between the virtual evaluation value and the actual measured value, and can be set as a constant; in one implementation, it can also be set with reference to cross-validation error or historical residuals; the update is used to improve the recommendation effect of the acquisition function for the next set of candidate process parameters, and the virtual evaluation value is the evaluation value calculated from the MLP predicted value, which is not equivalent to the actual welded value; the iteration ends when the preset iteration round is reached; Step S6: Output the verified global optimal process parameter combination to the user interface, so that the operator can set and verify the process of laser self-fusion welding of 6061 aluminum alloy based on the parameters; optionally, compare the measured welding quality data obtained from the verification with the prediction results to evaluate the prediction error.

[0022] In this invention, the method for constructing the initial process parameter dataset in step S1 includes: Multiple sets of solid welding experiments were conducted within a preset process window using full factorial experimental design or orthogonal experimental design schemes. Collect weld cross-sectional morphology data and microhardness data after the experiment; The collected data is cleaned, and abnormal data points with macroscopic porosity, incomplete penetration defects or severe surface collapse are removed according to the preset defect judgment threshold. The cleaned data is then standardized.

[0023] In this invention, in step S2, the comprehensive evaluation function for welding quality... The calculation logic is as follows: Construct a comprehensive score function oriented towards maximizing welding quality. : ; Among them, the Maximizing is equivalent to minimizing the following objective function: ; in, Microhardness value (unit: HV); is the weld formation coefficient, which is the ratio of weld width to weld depth, where weld width is the maximum width of the fusion zone in the weld cross section, and weld depth is the maximum depth of the fusion zone along the plate thickness direction; The preset ideal weld formation coefficient; The hardness weighting coefficient is preferably in the range of 0.01 to 0.2, and can be calibrated based on the magnitude difference between the microhardness and the weld formation coefficient deviation term. This represents the penalty term for the absolute deviation between the weld formation coefficient and the preset ideal weld formation coefficient; the optimization objective is to find the optimal weld formation coefficient within the process parameter search space. The parameter combination that maximizes the value, or equivalently, finding the combination that maximizes the value. The minimum combination of parameters; where, when constructing training sample labels, and The measured microhardness Compared with the measured weld formation coefficient Substitution; in the parameter optimization stage, and The predicted microhardness value output by the MLP model and predicted value of weld formation coefficient ;in, Used for dimensional matching of the deviation terms of microhardness and weld formation coefficient.

[0024] In this invention, the specific process of adaptively optimizing the hyperparameters of the MLP neural network in step S3 includes: setting a hyperparameter search space, wherein the hyperparameters include the number of hidden layers, the number of neurons in each hidden layer, the learning rate, and the number of iterations; using the prediction error of the MLP neural network on the validation set as the objective function; using a Bayesian optimization algorithm to iteratively search within the hyperparameter search space to find the hyperparameter combination that minimizes the prediction error, and using it as the final model configuration parameter; the prediction error is a weighted combination of the prediction errors of the weld formation coefficient W and the microhardness H; in another implementation, when the MLP uses Score as the single output regression label, the prediction error is the mean square error of Score.

[0025] In this invention, the specific implementation of the Bayesian optimization algorithm in step S4 includes: The posterior mean and posterior variance of the unsampled points were calculated using a Gaussian process regression model. Construct a data acquisition function, which employs an expectation boosting (EI) strategy or a confidence upper bound (UCB) strategy. The next set of candidate process parameters is determined by maximizing the value of the acquisition function, in order to balance global exploration and local development.

[0026] In this invention, the microhardness value ( As a macroscopic proxy indicator reflecting the evolution of the microstructure of the welded joint; the comprehensive evaluation function of welding quality maximizes the microhardness value ( Combined with weld formation coefficient constraints, the selected global optimal process parameter combination can be used to obtain welded joints with a grain refinement trend and an M-type microhardness distribution trend.

[0027] In this invention, an adaptive optimization system for laser self-fusion welding process parameters of 6061 aluminum alloy is provided, which runs on a general-purpose computer or workstation and includes: a memory, a processor, and a display and input device. The memory stores a computer program, which, when executed by the processor, implements the following functional modules: The experimental data entry and preprocessing module is configured to receive raw offline experimental data of laser self-fusion welding of 6061 aluminum alloy (e.g., 81 sets, including process parameters and corresponding weld formation coefficients and microhardness test values) through a user input interface, and to clean and standardize the data. The comprehensive welding quality assessment module is configured to integrate the multi-dimensional weld formation coefficient and microhardness index into a single-objective comprehensive welding quality score based on a weighted comprehensive evaluation function. The intelligent agent model construction and hyperparameter optimization module is configured to use a Bayesian hyperparameter optimization strategy to adaptively select the optimal topology and hyperparameters of the MLP neural network based on the cleaned effective dataset, and to build a high-precision process parameter-quality mapping model. The virtual optimization module for process parameters is configured to perform virtual iterative calculations in a continuous parameter space using Gaussian process regression and acquisition functions to discover the theoretically optimal global process parameters; and to input the virtual evaluation values ​​calculated from the MLP prediction values ​​as low-confidence observations with preset observation noise into the Gaussian process surrogate model to update its posterior distribution. The error assessment and feedback correction module is configured to compare the measured welding quality data with the predicted welding quality data, calculate the deviation, and determine whether it exceeds a preset threshold. In an optional implementation, when the threshold is exceeded, the model retraining and process parameter re-optimization process is triggered. The optimal parameter decision and visualization output module is configured to combine the verified global optimal process parameters and display them on a display device for visualization output, so that operators can read them and manually set the laser welding equipment based on the parameters.

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0029] Example: like Figure 1 As shown, this invention provides a laser self-fusion welding process, method, and system for aluminum alloys based on Bayesian and neural network co-optimization, including the following detailed steps: Step S1: Construct the initial process parameter dataset for laser autofusion welding of 6061 aluminum alloy.

[0030] In this embodiment, 6061-T6 aluminum alloy was selected as the base material, with a plate thickness of 2mm. The welding equipment used was a fiber laser, and the shielding gas was high-purity argon, with a flow rate set to 15L / min.

[0031] To construct a high-quality initial dataset, the inventors employed a full-factor experimental design. The selected input process parameters (variables) and their value ranges are as follows: Table 1: Design of process parameters for laser autofusion welding

[0032] Based on the above parameter combinations, a total of 81 sets of solid welding experiments were conducted. After welding, the 81 welded test plates were manually subjected to wire cutting, cold mounting, grinding, polishing, and etching treatments.

[0033] To ensure the accuracy and physical consistency of the training data, this embodiment establishes clear data cleaning and anomaly removal criteria. Specifically, 81 weld cross-sections are observed using an optical microscope, and a sample is identified as an "anomaly" and removed if it meets any of the following conditions: Based on the judgment principles for porosity, incomplete penetration, and surface formation defects in welded joint quality evaluation standards such as GB / T3323, and combined with the characteristics of laser autofusion welding process, the porosity diameter threshold and number threshold are used to reflect the acceptable range of macroscopic defect levels of welded joints at the metallographic cross-sectional scale, so as to avoid abnormal defect samples causing bias to the mapping model. 1. Severe porosity defects: More than 5 pores with a diameter >0.2 mm or micropores (diameter <0.2 mm) are observed on the weld cross-section; 2. Incomplete penetration defect: The weld penetration depth is less than the plate thickness (2mm), and the incomplete penetration depth exceeds 0.25mm (i.e., effective penetration depth <1.75mm); 3. Severe surface collapse: If the depth of the weld surface depression exceeds 0.4 mm, it is judged as an unqualified sample and is rejected.

[0034] The above 81 experimental samples were tested using an optical microscope and a microhardness tester, and the weld formation coefficient (W) and weld center microhardness (H) were obtained for each experimental group.

[0035] These 81 sets of original "process parameter-quality response" data were entered into the computer system through the dataset construction module. The system cleaned the data according to the aforementioned preset standards, removing abnormal samples containing severe porosity, incomplete penetration, or severe surface collapse.

[0036] After cleaning, 64 valid samples remained. It should be noted that although removing outliers reduces the size of the dataset, the Bayesian-neural network co-optimization strategy used in this invention has significant robustness to small samples. The 64 valid samples far exceed the minimum sample limit required for algorithm convergence, which is sufficient to support the model in building a high-precision mapping relationship.

[0037] The system then standardizes these 64 sets of valid data to construct the final training dataset.

[0038] Step S2: Construct a comprehensive evaluation function for welding quality.

[0039] To achieve synergistic optimization of macroscopic forming and mechanical properties, this embodiment constructs a single-objective comprehensive welding quality score function, Score.

[0040] In this invention, the weld formation coefficient It is used to characterize the macroscopic forming quality of the weld cross-section, and is defined as the ratio of weld width to weld depth, i.e. =Weld width / Weld depth Wherein, weld width is the maximum width of the fusion zone in the weld cross-section, and weld depth is the maximum depth of the fusion zone along the plate thickness direction; when the weld forming coefficient When the weld cross-section approaches 1.0, it exhibits a relatively balanced aspect ratio, typically corresponding to a more stable molten pool morphology and heat input distribution, which is beneficial for achieving better weld consistency and defect suppression. When the weld width is relatively large, it may indicate a relatively increased weld width but insufficient weld penetration, which may reduce the effective weld penetration and increase the risk of defects such as incomplete penetration and insufficient fusion. When the melt width is relatively small, it may indicate a relatively large melt depth but insufficient melt width or an unstable melt pool morphology, which may lead to defects such as porosity, cracks, or poor surface formation under specific process conditions; therefore, this invention will As the ideal target value for the weld formation coefficient.

[0041] It should be further noted that the ideal target value for the weld formation coefficient is not fixed at 1.0 under all working conditions. In this embodiment, for a 2mm thick 6061 aluminum alloy sheet, considering both the weld penetration requirement and the formation stability, the target value is... As the ideal target value.

[0042] When the plate thickness, welding mode, or heat input conditions change, those skilled in the art can adjust the target value of the ideal weld formation coefficient according to the ratio of plate thickness to desired penetration depth. For example, the target value can be set as a function or empirical calibration value related to plate thickness. This adjustment does not affect the basic technical idea of ​​the present invention to construct a comprehensive evaluation function by synergistically combining weld formation coefficient deviation and microhardness.

[0043] Targeting the characteristics of laser self-fusion welding of 6061 aluminum alloy, this invention aims to maximize the microhardness while simultaneously achieving a weld formation coefficient that is close to a preset ideal weld formation coefficient. The deviation is used as a penalty term for constraint. Therefore, the comprehensive score function is defined as follows: ; in, This is the microhardness value (unit: HV). Its physical dimensions are retained here instead of the normalized value to preserve its physical meaning. The weld formation coefficient is the ratio of weld width to weld depth. The hardness weighting coefficient can be calibrated within the range of 0.01 to 0.2 based on the order-of-magnitude difference between microhardness and forming factor; in this embodiment, it is set to 0.15. It should be noted that the Score function defined in this step serves as a unified evaluation criterion throughout the entire method, specifically in the model training phase of step S3. and The measured microhardness H and measured weld formation coefficient W are substituted into the model to construct the supervision signal for the training samples, while the score is not used as the direct regression output of the MLP model. In one implementation, the MLP uses the weld formation coefficient W and microhardness H as multi-output regression labels to predict... and Then calculate Score / Objective; in another implementation, Score can also be used directly as the single output regression label; in the optimization phase of steps S4 and S5, and The input is the predicted value output by the MLP model. and This is used to evaluate the merits of candidate process parameters. The final optimization objective is to maximize this Score value. In the solution implementation of this embodiment, to facilitate calculation using Gaussian process regression and Expectation Improvement (EI) within the minimization framework, Score maximization is equivalently rewritten as minimizing the Objective. The optimization is performed using the minimization of the Objective, and the final optimal solution is still interpreted and displayed in the Score equivalent form, where: ; It should be noted that the weighting coefficients in the above evaluation function... The setting logic is as follows: Regarding the weighting coefficients The physical meaning of (0.15): Since the microhardness of 6061 aluminum alloy welds is usually distributed between 50-100 HV, and the deviation term of the weld formation coefficient... Typically, these are dimensionless values ​​less than 1.0. A coefficient is introduced. =0.15 served the purposes of "aligning magnitudes" and "incentive dominance": 1. Alignment of scales: Convert the physical value of hardness into a numerical "bonus score" (e.g., 60HV × 0.15 = 9.0) to make it the baseline value for the score.

[0044] 2. Incentive-Driven Approach: Based on this baseline, forming deviation is fine-tuned as a "deduction item." This design ensures that the optimization algorithm prioritizes searching in "high-hardness regions," while using the deviation term to eliminate solutions with extremely poor forming (such as incomplete penetration or severe collapse). More importantly, according to the grain refinement strengthening mechanism (Hall-Petch relationship) in metallurgy, the hardness of a material is inversely proportional to the square root of its grain diameter. Therefore, by providing a positive incentive to the microhardness term, this invention essentially encourages the algorithm to search for process windows with faster cooling rates and moderate heat input; the weld microstructure under these windows is typically accompanied by grain refinement (such as the aforementioned fine equiaxed grains) and defect suppression.

[0045] It should be noted that this invention does not limit the weighting coefficient. The specific value, and its variation within the range, will not change the basic technical concept of this invention, which uses microhardness as a proxy index for microstructure and combines it with weld formation constraints to collaboratively optimize process parameters.

[0046] Experimental results show that the score value calculated under this coefficient setting not only accurately reflects the overall quality of the welded joint, but also achieves indirect screening of microstructure characteristics through macroscopic mechanical indicators.

[0047] Step S3: Build and train the MLP neural network prediction model.

[0048] It should be noted that in the preferred implementation of this embodiment, the MLP prediction model uses weld formation coefficient and microhardness as multi-output regression targets, and the comprehensive evaluation function Score / Objective is only used as the evaluation criterion in the process parameter optimization stage, and is not used as the direct regression label for MLP training. To improve the prediction accuracy and generalization ability of MLP neural network models, this invention employs a Bayesian optimization algorithm to adaptively tune network hyperparameters during the model construction phase, thereby avoiding the subjectivity caused by manual settings (e.g., ...). Figure 3(As shown), the specific process is as follows: Activation function selection: In MLP neural networks, the activation function is responsible for introducing nonlinearity into the neuron's output. Common activation functions include the sigmoid function, hyperbolic tangent function (Tanh), and rectified linear function (ReLU). The purpose of selecting an activation function is to enable the neural network to fit the complex nonlinear relationship between welding process parameters and welding quality.

[0049] The MLP hyperparameter search space is defined as follows: the number of nodes in the first hidden layer is set to [10, 100], the number of nodes in the second hidden layer is set to [10, 50], and the learning rate is set to [0.001, 0.1]. The hyperparameter search space is set to a loose coverage range based on the welding process parameter dimension and sample size. Its purpose is to provide sufficient exploration space for Bayesian optimization, rather than pre-assuming an optimal structure. Specific activation functions and optimizers are replaceable components and do not constitute a limitation of this invention.

[0050] Perform Bayesian optimization: The algorithm initializes a set of hyperparameters, trains the MLP model, and calculates the model's predictive performance using validation set data. In this embodiment, the weld formation coefficient W and microhardness H are preferably used as multi-output regression labels, and Bayesian optimization uses a weighted combination of the prediction errors of W and H on the validation set as the objective function. One implementation method is as follows: ; in, The number of samples in the validation set; , For the first Measured weld formation coefficient and microhardness of individual samples. , This corresponds to the MLP prediction value; , , where is the weighting coefficient; in another implementation, when Score is used as the single output regression label, the Bayesian optimization objective is the mean squared error of Score on the validation set.

[0051] Parameter Update and Locking: The Bayesian optimizer recommends the next set of hyperparameters that is more likely to reduce error based on historical MSE feedback. To balance search efficiency and model accuracy, this embodiment sets a dual termination condition: when the MSE value on the validation set drops to a preset threshold (e.g., ... When the number of iterations reaches a preset limit (e.g., 50), the optimization process automatically terminates. Subsequently, the algorithm locks the globally optimal hyperparameter combination corresponding to the current minimum MSE. In this embodiment, to reduce the computational overhead in each process parameter optimization process and improve system efficiency, the inventors, based on the previous independent hyperparameter optimization results, selected a set of stable MLP network hyperparameter combinations, and consistently used this set of hyperparameters for model training and prediction in subsequent process parameter optimization processes. 1. Network structure: 32 nodes in the first hidden layer and 16 nodes in the second hidden layer; 2. Key training parameters: The activation function used is the Rectified Linear Array (ReLU), the optimization solver is the Adaptive Moment Estimator (Adam), and the initial learning rate is set to 0.005. These parameters are used to construct the final MLP neural network prediction model.

[0052] Model training: The final MLP model is constructed using the optimal combination of hyperparameters. During model training, the input process parameters and output responses (weld forming coefficient, microhardness) can be standardized / de-standardized to improve training stability. However, in the evaluation function calculation and result output stage, microhardness still participates in the Score / Objective calculation and presentation in the physical dimension of HV.

[0053] Step S4: Global Optimization of Welding Process Parameters Based on Bayesian Algorithm. Using a trained MLP model as an "electronic black box," and the welding quality prediction score output by the MLP neural network as an approximate evaluation value of the objective function, the optimal solution is sought within a continuous process parameter space. In this step, the Bayesian optimization search space for the welding process parameters is defined as follows: Table 2: Optimization parameter setting range

[0054] To improve the efficiency of cold start optimization, this embodiment adopts a "grid seed initialization" strategy: the algorithm first divides the space into discrete grids and randomly selects 12 discrete points as initial seed points. The objective function evaluation values ​​(Score or equivalent objective function) of these 12 points are calculated using the MLP model, and they are used as the initial dataset for Bayesian optimization to construct a preliminary Gaussian process (GP) surrogate model.

[0055] Step S5: Perform virtual iterative optimization based on the surrogate model. Based on the initial dataset, the algorithm enters a Bayesian optimization iterative loop. It should be noted that this step is a virtual search process performed internally by the computer, aiming to discover the theoretical optimal solution, rather than a physical experimental loop. This embodiment sets a fixed number of iterations to 30 rounds, and each iteration includes the following core sub-steps: 1. Proxy Model Fitting: Utilizing all currently accumulated process parameters and their score data (including measured evaluation values ​​obtained from physical welding experiments and virtual evaluation values ​​predicted by MLP, wherein the virtual evaluation values ​​are based on predictions) and The calculated score, or equivalent objective, is used as the observed output of the Gaussian process. A Gaussian process (GP) regression model is fitted. To accurately capture the nonlinear relationship between welding process and quality and tolerate potential observation noise, this embodiment uses the Matern kernel function (…). The combination of a Gaussian process surrogate model and a white noise kernel function is used as a covariance function to predict the mean and uncertainty at any point in a continuous space. The white noise kernel function characterizes the prediction uncertainty / observation noise introduced by the virtual evaluation value, enabling the Gaussian process surrogate model to accommodate error fluctuations in the virtual evaluation value during posterior updates, thereby avoiding excessive bias in the optimization process caused by the virtual evaluation value. The noise level of the white noise kernel function can be set to a constant (e.g., ...). The noise level can be updated based on the newly added residuals when new measured samples are introduced in subsequent closed-loop validation, so that the surrogate model can maintain a "low confidence, correctable" fusion strategy for virtual evaluation values.

[0056] 2. Large-scale candidate point generation and acquisition: Within the continuously defined parameter search space, 4000 candidate parameter combinations are randomly generated. The potential value of each candidate point is calculated using an acquisition function. In this embodiment, the Expectation Improvement (EI) strategy is preferably used as the acquisition function; in other embodiments, the Confidence Upper Bound (UCB) strategy can also be used. Exploration parameters are then set. To balance utilization and exploration; 3. Decision-making and virtual evaluation: The point with the highest EI value among 4000 candidate points is selected as the recommendation parameter for this round. This parameter is input into the MLP neural network model for prediction, obtaining the predicted weld formation coefficient and predicted microhardness of this point, and calculating the predicted quality score accordingly, which is then incorporated into the dataset. 4. Model Update and Iteration: The recommended parameters and their predicted quality scores are used as virtual evaluation values ​​to update the posterior distribution of the Gaussian process surrogate model. These virtual evaluation values ​​are input into the Gaussian process surrogate model as low-confidence observations with preset observation noise. It should be noted that this update step aims only to improve the recommendation effect of the acquisition function in exploring the parameter space. The virtual evaluation value is calculated from the MLP prediction values ​​and is not equivalent to the actual measured values ​​of the solid welding. 5. Loop Termination: Repeat the above steps until the preset 30 iterations are completed, at which point the algorithm stops and outputs the historical best solution.

[0057] In the actual operation of this embodiment (e.g.) Figure 4 As shown in the figure, despite setting 30 iterations, monitoring data shows that the algorithm can usually find the global optimal score or a very close solution around the 25th round. This strategy of "small batch seed + large-scale candidate sampling" effectively balances "utilization" and "exploration", proving that the method can quickly lock the optimal process window with extremely limited computing resources (only calling the model prediction a few dozen times).

[0058] It should be noted that the number of initial seed points, the size of candidate points, and the number of iterations mentioned above are merely exemplary parameter settings in this embodiment. Those skilled in the art can adjust them according to computing resources, parameter space dimensions, and model complexity without affecting the implementation of the virtual optimization mechanism based on the proxy model of this invention.

[0059] Step S6: Optimal parameter output and closed-loop verification.

[0060] The parameter decision output module outputs the globally optimal combination of process parameters on the display interface: laser power 1780W, welding speed 48mm / s, oscillation frequency 100Hz, and oscillation amplitude 1.3mm.

[0061] Based on the output, the operator manually inputs the parameters into the laser welding equipment to perform a physical verification weld. Subsequently, the obtained form factor and microhardness data are entered into the system. This system has the following preset offline closed-loop correction logic: 1. Error Calculation and Judgment: The system automatically calculates the relative error between the entered measured score value and the predicted score value.

[0062] 2. Feedback Correction Mechanism (Adaptive): If the relative error exceeds a preset threshold (e.g., 5%), or if the entered quality inspection results show the presence of porosity / crack defects, the system will automatically execute the following "re-optimization" process. The error threshold is an example setting and can be adjusted according to the specific allowable range of welding quality: This set of "process parameters - measured response" was incorporated into the original dataset as a high-confidence sample; The updated dataset is used to re-trigger the model training in step S3 and the virtual optimization calculation in steps S4-S5; A new set of recommended process parameters is output for operators to verify again until the requirements are met.

[0063] 3. Termination process: If the relative error is less than the preset threshold and the weld defect index meets the preset qualification criteria (e.g., the size and number of pores, the depth of incomplete penetration, and the depth of surface collapse do not exceed the thresholds described in step S1), then the optimization is determined to be successful and the final process report is output. In one optional implementation, when multiple rounds of physical verification are introduced in actual industrial applications or multi-batch production scenarios, the newly obtained measured welding quality data can be used as high-confidence observation samples to update the Gaussian process surrogate model, so as to further correct the systematic bias introduced by the predicted evaluation value, thereby improving the stability and long-term consistency of subsequent process parameter optimization; it should be noted that the process parameter adaptive optimization method described in this embodiment does not require this update process as a necessary step, and does not impose mandatory limitations on this process; The specific verification results in this embodiment are as follows: 1. Macroscopic morphological analysis: like Figure 5 As shown, the optimized weld cross-section exhibits a typical "wine glass" shape. Measurements show that the weld formation factor is 1.003, which is close to the preset ideal weld formation factor. (In this embodiment) No pores or cracks were observed. This indicates that the optimized process parameters, through appropriate heat input and laser oscillation stirring, effectively prolonged the solidification time of the molten pool and promoted the escape of pores.

[0064] 2. Microstructure analysis: like Figure 6 As shown, high-magnification observation revealed that the central region of the weld is composed of equiaxed dendrites with an average grain size of 30-60 μm. In particular, the extremely narrow equiaxed fine-grained zone (FQZ) near the fusion line has an average grain size of approximately 5 μm and exhibits a non-dendritic morphology.

[0065] 3. Mechanical property analysis: like Figure 7 As shown, the microhardness distribution of the welded joint exhibits a significant M-shaped characteristic. The overall hardness of the weld is superior to that of the base metal, thus achieving joint strengthening.

[0066] It should be noted that the M-type microhardness distribution is a common microstructure and performance response characteristic of 6061 aluminum alloy welded joints under reasonable heat input conditions. This invention does not use this distribution morphology as a mandatory constraint condition for optimizing process parameters, but rather as one of the expected mechanical property distribution effects of the welded joint after optimization of the comprehensive evaluation function.

[0067] The embodiments described above validated the feasibility and effectiveness of the recommended process parameters through at least one physical verification. It should be noted that the adaptive optimization method and system for process parameters proposed in this invention do not depend on the specific number of single or multiple physical verifications. Its technical effect lies in the fact that the constructed optimization framework can be repeatedly executed under the same or similar process conditions to continuously output process parameter combinations that meet quality constraints. In practical industrial applications, the recommended parameters can be repeatedly verified according to production needs to further evaluate process stability; this invention does not impose any limitations on this.

[0068] Regarding the hardware implementation architecture of the system of this invention: like Figure 8 As shown, the adaptive optimization system for process parameters described in this embodiment is physically implemented based on a general-purpose computer or a high-performance computing workstation. The system adopts an "offline optimization-online verification" working mode. Logically, it is divided into a data layer, a computation layer, and an interaction layer, with each functional module deployed within one of these layers, as detailed below: 1. Data Layer: This layer primarily supports the experimental data entry and preprocessing module. Operators input the process parameters (including laser power, welding speed, etc.) of the 81 sets of physical welding experiments completed on the welding machine, along with their corresponding post-weld testing performance data, into the system via Excel spreadsheets or manual input. This module then cleans and standardizes the data, providing a high-quality data foundation for upper-level calculations.

[0069] 2. Computational Layer: This is the core algorithm engine of the system, integrating a comprehensive welding quality evaluation module, an intelligent proxy model construction and hyperparameter optimization module, a virtual process parameter optimization module, and an error evaluation and feedback correction module. The comprehensive welding quality assessment module is responsible for converting the preprocessed multidimensional data into a single score. The intelligent agent model building module utilizes CPU / GPU computing power to train a high-precision MLP neural network based on a Bayesian optimization strategy; The virtual optimization module for process parameters acts as a "virtual laboratory," simulating thousands of parameter iterations within a computer to uncover the theoretically optimal global solution.

[0070] The error assessment and feedback correction module is responsible for comparing the deviation between "actual measurement and prediction" and activating the closed-loop self-learning process of the system when the deviation is too large.

[0071] 3. Interaction Layer: Primarily responsible for optimal parameter decision-making and visualization output modules. This layer transforms the complex algorithm results output from the computation layer into intuitive combinations of process parameters (such as laser power and welding speed), and displays them on the screen. Operators then read these parameters and input them into the laser welding machine's control panel for final physical verification.

[0072] In summary, the system of this invention effectively reduces process development costs and improves the quality of laser self-fusion welding of 6061 aluminum alloy through the model of "data-driven modeling + Bayesian virtual optimization + manual entity verification".

[0073] It should be noted that, in the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0074] Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Unless otherwise expressly defined, "a plurality of" means at least two.

[0075] It should be understood that the method steps / processes described in the specification and the flowcharts in the accompanying drawings are merely illustrative examples, and the execution order of the steps shown is not a limitation of the present invention; without departing from the concept of the present invention, the execution order of each step can be adjusted as needed, or some steps can be executed in parallel.

[0076] The functional modules of the system / device / equipment described in this specification can be implemented by hardware, software, firmware, or a combination thereof. The hardware may include, but is not limited to: processors, controllers, application-specific integrated circuits (ASICs), programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), and combinations thereof; the software may include program instructions stored in memory and executed by the processor. The functional modules can be integrated into the same processing module, set up separately, or integrated in any combination.

[0077] For the purposes of this specification, "computer-readable storage medium" means any medium capable of containing, storing, or carrying program instructions in any form, including but not limited to: random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, magnetic storage media, optical storage media, and any combination of the above media. When executed by a processor, the program instructions can implement all or part of the steps in the method embodiments of the present invention.

[0078] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A laser self-fusion welding process and method for aluminum alloys based on Bayesian and neural network co-optimization, characterized in that, Includes the following steps: Step S1: Construct an initial process parameter dataset for laser autofusion welding of 6061 aluminum alloy; the process parameters include at least laser power, welding speed, oscillation frequency, and oscillation amplitude; the response values ​​of the dataset include weld formation coefficient and microhardness; Step S2: Construct a comprehensive evaluation function for welding quality, calculate the absolute deviation between the weld formation coefficient and the ideal weld formation coefficient, and obtain a single-objective welding quality score by combining microhardness. During the training phase, the measured data is used as labels to train the MLP prediction model. During the optimization phase, the MLP prediction results are used to calculate the quality score and evaluate the candidate process parameters. Step S3: Construct and train the MLP neural network prediction model; when constructing the network, adaptively search for hyperparameters through Bayesian optimization, or build the network using a pre-determined combination of hyperparameters; use the trained MLP neural network to establish a nonlinear mapping between welding process parameters and weld formation coefficient and microhardness, and output the prediction results of the two indicators. Step S4: Globally optimize welding process parameters based on Bayesian optimization algorithm; select discrete grid points within the process parameter range, randomly select some grid points as initial seed points to complete the algorithm cold start, use Gaussian process to construct surrogate model to fit the posterior distribution of objective function, select the next set of candidate process parameters through acquisition function, and the initial observation value of Gaussian process surrogate model is the evaluation value corresponding to the seed point; Step S5: Perform virtual iterative optimization based on the agent model; Candidate points are generated in a continuous parameter space, and the optimal virtual evaluation point is selected by the acquisition function. The prediction results of weld formation coefficient and microhardness are obtained by using MLP neural network. The quality score is calculated as a virtual evaluation value of a single scalar. This virtual evaluation value is used as a low-confidence observation term with preset observation noise and input into the Gaussian process surrogate model to update the model posterior distribution. The iteration is terminated after a preset number of rounds. Step S6: Output the verified global optimal process parameter combination to the human-machine interface to complete the process configuration for laser self-fusion welding of 6061 aluminum alloy.

2. The method according to claim 1, characterized in that, In step S1, the method for constructing the initial process parameter dataset includes: Using full factorial experimental design or orthogonal experimental design schemes, physical welding experiments were conducted within a preset process window; Collect weld cross-sectional morphology data and microhardness data after the experiment; The collected data is cleaned, and abnormal data points are removed based on a preset defect judgment threshold. The cleaned data is then standardized.

3. The method according to claim 1, characterized in that, In step S2, the comprehensive evaluation function for welding quality The calculation logic is as follows: Construct a comprehensive score function oriented towards maximizing welding quality. : ； Among them, the Maximizing is equivalent to minimizing the following objective function: ； in, This refers to the microhardness value. This is the weld formation coefficient. The preset ideal weld formation coefficient, This is the hardness weighting coefficient. This represents the penalty term for the absolute deviation between the weld formation coefficient and the preset ideal weld formation coefficient.

4. The method according to claim 1, characterized in that, The specific process of adaptively optimizing the hyperparameters of the MLP neural network in step S3 includes: setting a hyperparameter search space, where the hyperparameters include the number of hidden layers, the number of hidden layer neurons, the learning rate, and the number of iterations; using the prediction error of the MLP neural network on the validation set as the objective function; and using Bayesian optimization to iteratively optimize within the search space to obtain the hyperparameter combination corresponding to the minimum prediction error, which is then used as the final configuration of the model.

5. The method according to claim 1, characterized in that, In step S4, the specific implementation of the Bayesian optimization algorithm includes: The posterior mean and posterior variance of the unsampled points were calculated using a Gaussian process regression model. Construct a data acquisition function, which employs an expectation boosting strategy or a confidence upper bound strategy; The next set of candidate process parameters is determined by maximizing the value of the acquisition function, in order to balance global exploration and local development.

6. The method according to claim 3, characterized in that, Microhardness serves as a proxy index for characterizing the evolution of the microstructure of welded joints. The comprehensive evaluation function for weld quality aims to maximize microhardness and incorporates weld formation coefficient constraints, enabling the optimized process parameters to produce welded joints with refined grains and M-shaped microhardness distribution.

7. An adaptive optimization system for laser self-fusion welding process parameters of 6061 aluminum alloy using the method described in any one of claims 1-6, characterized in that, Running on a general-purpose computer or workstation, including: memory, processor, and display and input devices; The memory stores a computer program, which, when executed by the processor, implements the following functional modules: The experimental data entry and preprocessing module is configured to receive raw offline experimental data of laser self-fusion welding of 6061 aluminum alloy through a user input interface, and to clean and standardize the data. The comprehensive welding quality assessment module is configured to integrate the multi-dimensional weld formation coefficient and microhardness index into a single-objective comprehensive welding quality score based on a weighted comprehensive evaluation function. The intelligent agent model construction and hyperparameter optimization module is used to adaptively select the network structure and hyperparameters of the MLP neural network based on the cleaned effective dataset and adopt the Bayesian hyperparameter optimization strategy to establish a high-precision mapping model between process parameters and welding quality. The virtual optimization module for process parameters is used to perform virtual iterative calculations in a continuous parameter space using Gaussian process regression and acquisition functions to search for globally optimal process parameters; the virtual evaluation values ​​calculated from MLP predictions are used as low-confidence observation data with preset observation noise and input into the Gaussian process surrogate model to update the posterior distribution; The error assessment and feedback correction module is configured to compare the measured welding quality data with the predicted welding quality data, calculate the deviation and determine whether it exceeds a preset threshold; when the threshold is exceeded, the model retraining and process parameter re-optimization process is triggered. The optimal parameter decision and visualization output module is configured to combine the verified global optimal process parameters and output them in a visual manner on a display device.

8. An electronic device, characterized in that, include: processor; And a memory for storing executable instructions of the processor; wherein the processor is configured to execute the adaptive optimization method for laser self-fusion welding process parameters of 6061 aluminum alloy based on Bayesian-neural network co-optimization as described in any one of claims 1-6 by executing the executable instructions.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the adaptive optimization method for laser self-fusion welding process parameters of 6061 aluminum alloy based on Bayesian-neural network collaborative optimization as described in any one of claims 1-6.

Images