Titanium alloy heat source dynamic regulation welding and joint performance hybrid prediction method

Abstract

The application discloses a titanium alloy heat source dynamic regulation and control welding and joint performance hybrid prediction method, relates to the technical field of welding, and comprises the following steps: aiming at the problem that the heat source of the titanium alloy medium plate welding is concentrated in the weld center and the side wall has a high unmelted rate, taking a heat source dispersion coefficient λ as a core regulation and control variable, constructing a physical-data hybrid driving model chain of the heat source form, the weld geometry, the microstructure and the tensile strength, defining the λ as the ratio of the side wall heat and the center heat, and realizing the active regulation of the λ in the safe range through a welding gun swing device; adopting a multivariate nonlinear regression and support vector machine regression hybrid mapping process parameter to the λ; adaptively fusing and predicting the weld penetration, the weld width and the remaining height by using MNR, BPNN and SVR; forcibly introducing physical intermediate variables such as the average width of α martensite and the proportion of low-angle grain boundaries to construct a tensile strength prediction model, and nonlinearly correcting the model by using the λ correction factor; solving a Pareto optimal process solution set; and establishing an online updating mechanism of the model chain causality consistency.

Description

Technical Field

[0001] This invention relates to the field of welding technology, and more specifically to a method for dynamically controlling the heat source of titanium alloys during welding and for predicting the performance of joints. Background Technology

[0002] Titanium alloys (such as TC4) are widely used in aerospace, armored vehicles, and marine engineering due to their high specific strength, excellent corrosion resistance, and biocompatibility. However, the welding of medium-thick titanium alloy plates has long faced the following technical challenges:

[0003] 1. Concentrated heat input: In traditional tungsten inert gas welding, the heat source is concentrated in the center of the weld, resulting in insufficient heat on the sidewall of the bevel, which easily leads to defects such as incomplete fusion and undercut.

[0004] 2. Poor pool fluidity: The titanium alloy melt has high viscosity and high surface tension. Insufficient stirring of the molten pool leads to prominent problems such as porosity, inclusions and grain coarsening. In particular, grain growth has a significant impact on toughness.

[0005] 3. Complex coupling of process parameters: Multiple variables such as welding current, welding speed, and wire feeding speed affect each other, exhibiting highly nonlinear and strongly coupled characteristics. On-site welders often need to repeatedly adjust dozens of sets of parameters to obtain a qualified weld.

[0006] 4. Lack of systematic performance prediction methods: Existing technologies mostly rely on destructive testing to obtain joint performance. The cost of a single joint can reach several thousand yuan, and the test cycle can last for several weeks, which seriously restricts the efficiency of process development.

[0007] In our previous explorations, we attempted to regulate the arc morphology using an external alternating magnetic field. However, in the welding of medium and thick plates in TC4, we found that the magnetic field parameters were severely coupled with the process parameters, resulting in highly unstable actual regulation effects. Furthermore, the magnetic field equipment occupied a large space, making it difficult to implement on-site. Therefore, this invention shifts to a more direct and easily engineerable approach: actively redistributing heat input through the heat source dispersion coefficient. In recent years, welding process parameter optimization methods based on surrogate models and multi-objective optimization algorithms have gradually attracted attention. For example, CN113642220A discloses a ship welding process optimization method based on RBF and MOPSO. This method uses a radial basis function neural network to construct a surrogate prediction model of welding process parameters and weld joint quality, and uses the weld formation coefficient and reinforcement coefficient as optimization objectives, employing a multi-objective particle swarm optimization algorithm to solve for the Pareto optimal solution set.

[0008] However, this type of existing technology has the following shortcomings:

[0009] First, the heat source morphology is fixed: existing methods treat the welding heat source as a fixed input and do not include the heat source distribution morphology as an independent variable that can be actively controlled in the optimization framework, making it difficult to fundamentally solve the sidewall fusion quality problem caused by uneven heat source distribution.

[0010] Second, the physical interpretability of the models is insufficient: existing technologies mostly use single data-driven surrogate models (such as RBF neural networks) to directly map process parameters and quality indicators. Such purely data-driven models lack clear physical mechanisms and cannot characterize microstructures (such as...). Intermediate physical processes such as the evolution of martensite, grain boundaries, etc., and the activation energy of phase transformation limit its generalization and extrapolation capabilities under complex working conditions.

[0011] Third, the optimization space is limited: Since the heat source form is not taken as an independent optimization variable, the optimization space of the process parameters is limited to traditional electrical and mechanical parameters, which restricts the further improvement of the overall performance of the joint.

[0012] Therefore, how to provide a method for precise and synergistic optimization of titanium alloy welding processes within a larger parameter space by using the heat source morphology as the core control variable and constructing a predictive model chain that integrates physical mechanisms and data-driven approaches is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0013] In view of the above problems, the present invention is proposed to provide a method for dynamically controlling the welding and joint performance of titanium alloys by heat source, thereby overcoming or at least partially solving the above problems.

[0014] To achieve the above objectives, the present invention adopts the following technical solution:

[0015] In a first aspect, embodiments of the present invention provide a method for dynamically controlling the heat source of titanium alloys during welding and for predicting the hybrid performance of joints, characterized in that it includes:

[0016] S1. Define the heat source dispersion coefficient λ as the ratio of the heat transferred to the sidewall of the groove to the heat transferred to the center of the weld. By adjusting the heat source dispersion coefficient λ, the active redistribution of welding heat input in the groove width direction can be achieved.

[0017] S2. Construct a hybrid mapping model for heat source dispersion coefficient. Based on process parameters, use a multivariate nonlinear regression model and a support vector machine regression model for prediction. Then, determine the fusion weights through a particle swarm optimization algorithm and output the heat source dispersion coefficient λ.

[0018] S3. Based on process parameters and heat source dispersion coefficient λ, establish a multi-source algorithm fusion prediction model for weld geometric features and output weld geometric features;

[0019] S4. Based on the heat source dispersion coefficient λ and the geometric characteristics of the weld, the average width of α martensite and the proportion of small-angle grain boundaries are calculated as intermediate variables to establish a hybrid prediction model for tensile strength. A dynamic adjustment correction factor for heat source distribution is introduced for correction, and the corrected tensile strength is output.

[0020] S5. Taking the weld geometry and the corrected tensile strength as objectives, the complete model chain constructed in S2-S4 is used as the fitness evaluation engine, and the process parameters are used as decision variables to construct a multi-objective collaborative optimization function and solve the Pareto optimal solution set.

[0021] S6. Collect newly added real welding data from the industrial site, remove abnormal data, and use effective incremental data to trigger adaptive updates of the complete model chain.

[0022] Furthermore, the heat source dispersion coefficient λ is specifically:

[0023]

[0024]

[0025] in To transfer heat to the bevel sidewall; To transfer heat to the center of the weld; Total heat input; These are parameters for heat source dispersion control.

[0026] Furthermore, in S2:

[0027] The process parameters are welding current, wire feed speed, and welding speed;

[0028] Multivariate nonlinear regression model for:

[0029]

[0030] Where the initial heat source dispersion coefficient =1.0 is a reference value; For welding current, wire feed speed, welding speed, and heat source dispersion control parameters, i = 1, 2, 3, 4; , , These are regression coefficients, dimensionless normalized regression coefficients; For special input indexes;

[0031] Support Vector Machine Regression Model for:

[0032]

[0033]

[0034] in It is a radial basis kernel function; For Lagrange multipliers; is the kernel width parameter; b is the bias term; N is the number of support vectors.

[0035] Furthermore, the specific process of determining the fusion weights using the particle swarm optimization algorithm in S2 is as follows:

[0036] An integrated multivariate nonlinear regression model using the weighted average method With support vector machine regression model :

[0037]

[0038] in Fusion weights and Fusion weights The solution is obtained by minimizing the prediction error using the particle swarm optimization algorithm.

[0039]

[0040] Where M is the number of validation samples; Let be the measured heat source dispersion coefficient of the k-th sample; These are the predicted values ​​from the integrated model.

[0041] Furthermore, the specific process of establishing a multi-source algorithm fusion prediction model for weld geometric features in S3 is as follows:

[0042] The geometric characteristics of the weld are the predicted penetration depth H, the predicted weld width B, and the predicted reinforcement height h.

[0043] S31. Establish three prediction sub-models respectively:

[0044] Multivariate nonlinear regression model MNR:

[0045]

[0046] in The weld geometry features predicted by MNR; These are the regression coefficients; For welding current, wire feed speed, welding speed, and heat source dispersion control parameters, i = 1, 2, 3, 4; For special input indexes;

[0047] Backpropagation Neural Network Model (BPNN):

[0048] A three-layer network structure is adopted, with the input layer consisting of 5 nodes. The hidden layer contains 8 nodes, and the output layer contains 3 nodes H, B, and h.

[0049] The activation function in the hidden layer is ReLU, and the optimization algorithm is Adam.

[0050]

[0051] in The network parameters are updated t times cumulatively in the activation function; The network parameters are updated t+1 times cumulatively in the activation function; t is the cumulative number of updates. The learning rate; The first moment deviation correction estimate for the gradient; The second moment deviation correction estimate for the gradient; For numerical stability terms; I is the welding current; V w V is the wire feeding speed; s Welding speed; These are parameters for heat source dispersion control.

[0052] Support Vector Machine Regression Model (SVR):

[0053] An SVR model is constructed using the radial basis function (RBF) to predict weld geometry features;

[0054] S32, Integrated Prediction Model F:

[0055] An adaptive weighted fusion algorithm is used to calculate the root mean square error (RMSE) of each prediction sub-model on the validation set, and weights are dynamically allocated accordingly.

[0056]

[0057] in The geometric features of the weld predicted by the k-th algorithm, MNR, BPNN, and SVR; The fusion weights of the k-th algorithm (MNR, BPNN, SVR) are proportional to the prediction accuracy; RMSE k The root mean square error (RMSE) of the k-th algorithm (MNR, BPNN, SVR) on the validation set. j Let represent the root mean square error of the j-th prediction algorithm on the validation set.

[0058] Furthermore, the specific process for establishing the hybrid prediction model for tensile strength in S4 is as follows:

[0059] S41. Based on the heat source dispersion coefficient λ and weld geometry, combined with welding thermal cycle theory, three key microstructure characterization variables were calculated:

[0060] α-martensite average width d α : with equivalent heat input There is a positive correlation, where nominal heat input I represents the welding current; V represents the welding current. s Welding speed; U is the arc voltage;

[0061] Small angle grain boundary ratio : with heat source dispersion coefficient Positively correlated;

[0062] S42. Establish a physical mechanism-driven model:

[0063]

[0064] in The tensile strength predicted by the physical model; These are the regression coefficients; This refers to the wire feeding speed; These are parameters for heat source dispersion control. Reference heat input; R is the activation energy for the β→α phase transition in titanium alloys; R is the gas constant.

[0065] The peak temperature is corrected by combining the heat conduction model with the heat source dispersion coefficient λ:

[0066]

[0067] in The initial temperature; denoted as thermal efficiency coefficient; x is thermal conductivity; r is distance from the center of the heat source; a is thermal diffusivity; m is the influence index of the heat source dispersion coefficient.

[0068] S43. Establish the digital-driven model BPNN:

[0069] Constructing a BP neural network:

[0070] 5 nodes in the input layer The hidden layer has 10 nodes, using ReLU as the activation function; the output layer has 1 node σ. b NN ; where σ b NN The tensile strength predicted by the data model;

[0071] S44. Employ a Stacking integration strategy to fuse the physical mechanism-driven model and the digital-driven model:

[0072]

[0073] in The predicted tensile strength; ∈[0,1] is the fusion coefficient.

[0074] Furthermore, the specific process of introducing a dynamic adjustment correction factor for heat source distribution is as follows:

[0075] Define the dynamic regulation correction factor for heat source distribution :

[0076]

[0077] Reference value of heat source dispersion coefficient =1; Minimum value of heat source dispersion coefficient =0.5; Maximum value of heat source dispersion coefficient =1.5; coefficients k1, k2, and k3 were determined by fitting experimental data using the least squares method;

[0078] The corrected tensile strength is:

[0079]

[0080] The tensile strength is adjusted by a correction factor. This represents the predicted tensile strength.

[0081] Furthermore, the multi-objective collaborative optimization function constructed in S5 is as follows:

[0082]

[0083] in The objective function vector; The tensile strength is adjusted by a correction factor. This is the predicted melting depth. The remaining height is the predicted value; process parameters I is the welding current; V is the wire feeding speed; s Welding speed; These are parameters for heat source dispersion control.

[0084] Furthermore, solving for the Pareto optimal solution set in S5 also includes:

[0085] A non-dominated sorting genetic algorithm assisted by particle swarm optimization was used to solve for the Pareto optimal process parameter solution set.

[0086] The non-dominated sorting genetic algorithm has a population size of 100, a crossover probability of 0.9, a mutation probability of 0.1, and a maximum number of iterations of 200. The initial population is pre-optimized using a particle swarm optimization algorithm.

[0087] Furthermore, the specific process of adaptive update in S6 is as follows:

[0088] For the backpropagation neural network in the complete model chain, the parameters are updated using online gradient descent:

[0089]

[0090] in This refers to the network parameters that are updated cumulatively t+1 times in the backpropagation neural network. The learning rate is the cumulative network parameter update t times in the backpropagation neural network. , =0.01, t is the cumulative number of updates; L is the loss function; x t y t Let be the input and output of the t-th new sample;

[0091] The incremental support vector machine regression algorithm is used. When a new sample falls into the support vector set or violates the KKT conditions, the parameters are updated incrementally by solving an augmented optimization problem.

[0092]

[0093] Where w is the weight vector; C is the regularization parameter; N is the original number of support vectors; and M is the number of new samples. and These are slack variables.

[0094] The beneficial effects of the above-described technical solutions provided in the embodiments of the present invention include at least the following:

[0095] 1. Heat source morphology as a core optimization variable: This invention is the first to incorporate the heat source dispersion coefficient λ as an independent optimization variable into a multi-objective optimization framework. By continuously adjusting λ within a safe range, the active redistribution of welding heat input in the groove width direction is achieved, fundamentally solving the process problems caused by heat source concentration, such as sidewall incomplete fusion and undercut, in the welding of medium-thick titanium alloy plates. Verification results show that the sidewall fusion quality is optimal when λ=1.05, increasing the sidewall penetration depth by approximately 15% compared to the conventional condition of λ=1.0.

[0096] 2. A physics-data hybrid driven model chain, combining interpretability and extrapolation capability: This invention constructs a complete causal chain of heat source morphology, weld geometry, microstructure, and mechanical properties, forcibly introducing physically interpretable intermediate variables such as the average width of α-martensite and the proportion of small-angle grain boundaries. This structure endows the model with stronger extrapolation capability: when inputting process parameters not covered by the training set, the model first infers the microstructure changes based on the physical mechanism of heat conduction, and then predicts the intensity based on the microstructure, avoiding the extrapolation distortion of a purely data-driven model. In the verification, when the welding current exceeds the maximum value of 10A in the training set, the prediction error of the model of this invention is 0.3%, while the error of the simple BPNN model is 2.1%.

[0097] Physics-data hybrid approach: The tensile strength prediction model integrates physical mechanisms and data-driven methods, which not only ensures the physical interpretability of the model, but also improves the prediction accuracy, with a prediction error of less than 0.7%.

[0098] 3. Multi-algorithm fusion, specifically designed for small sample sizes and strong nonlinearity. This invention employs adaptive weighted fusion of MNR, BPNN, and SVR in weld geometry prediction, and a physics-data hybrid driving approach in tensile strength prediction. This design is not a simple stacking of algorithms, but rather a tailor-made solution for the high cost of acquiring titanium alloy welding data, small sample size, and strong nonlinearity. The average prediction error of the integrated model on 30 validation samples is: weld penetration 0.2%, weld width 0.3%, and weld height 0.4%, which is more than 50% lower than that of a single algorithm.

[0099] 4. Causal Consistency Online Updates Ensure Long-Term Stability: Unlike conventional independent updates of each sub-model, this invention establishes a causal consistency update mechanism for the model chain. When the weld geometry model is updated, its output automatically triggers the calculation of microstructure parameters and the synchronous calibration of the tensile strength model; the heat source dispersion coefficient model is also fine-tuned based on feedback from the new data. This mechanism ensures that the model chain maintains physical causal consistency between input and output during long-term operation. After three months of continuous operation and accumulating 50 new data sets, the overall prediction accuracy of the model chain decreases by no more than 0.3%, while the accuracy decrease in the independent update mode is 1.2%.

[0100] 5. Heat source morphology as a core optimization variable: This invention uses the heat source dispersion coefficient λ as a core control and optimization variable. Compared with the existing method of parameter optimization under fixed heat source conditions, it broadens the process optimization space and can solve the core problems such as poor welding of sidewalls of titanium alloy medium and thick plates from the root of heat source distribution morphology.

[0101] 6. The fundamental difference between multi-algorithm fusion and a single surrogate model: This invention employs adaptive weighted fusion of MNR, BPNN, and SVR in weld geometry prediction, and a physics-data hybrid driving method in tensile strength prediction; neither of these is a simple application of a single algorithm. This multi-algorithm, multi-mechanism fusion strategy is specifically designed for titanium alloy welding, a highly nonlinear and strongly coupled system, and is fundamentally different from the general method using a single RBF surrogate model in existing technologies.

[0102] 7. Multi-objective optimization driven by heat source regulation: This invention incorporates the heat source dispersion coefficient λ as an independent optimization variable into the NSGA-II+PSO hybrid optimization framework. Compared with the existing technology that only optimizes traditional process parameters, it achieves synergistic optimization of heat source morphology and process parameters, and can obtain better comprehensive welding quality indicators. Attached Figure Description

[0103] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0104] Figure 1 This is an overall flowchart of the solution provided in the embodiments of the present invention;

[0105] Figure 2 This is a flowchart of the heat source dispersion coefficient hybrid mapping model provided in this embodiment of the invention;

[0106] Figure 3 This is a comparison chart of the predicted and measured weld width values ​​provided in the embodiments of the present invention;

[0107] Figure 4 This is a comparison chart of the predicted and measured weld penetration values ​​provided in the embodiments of the present invention.

[0108] Figure 5 This is a comparison chart of the predicted and measured values ​​of weld reinforcement height provided in this embodiment of the invention;

[0109] Figure 6 This is a comparison chart of the predicted and measured tensile strength values ​​provided in the embodiments of the present invention;

[0110] Figure 7 This is a welding hardness distribution curve provided in an embodiment of the present invention. Detailed Implementation

[0111] 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.

[0112] This invention discloses a method for dynamically controlling the heat source of titanium alloys during welding and for predicting the hybrid performance of joints, such as... Figure 1 As shown, it includes:

[0113] S1. Define the heat source dispersion coefficient λ as the ratio of the heat transferred to the sidewall of the groove to the heat transferred to the center of the weld. By adjusting the heat source dispersion coefficient λ, the active redistribution of welding heat input in the groove width direction can be achieved.

[0114] S2. Construct a hybrid mapping model for heat source dispersion coefficient. Based on process parameters, use a multivariate nonlinear regression model and a support vector machine regression model for prediction. Then, determine the fusion weights through a particle swarm optimization algorithm and output the heat source dispersion coefficient λ.

[0115] S3. Based on process parameters and heat source dispersion coefficient λ, establish a multi-source algorithm fusion prediction model for weld geometric features and output weld geometric features;

[0116] S4. Based on the heat source dispersion coefficient λ and the geometric characteristics of the weld, the average width of α martensite and the proportion of small-angle grain boundaries are calculated as intermediate variables to establish a hybrid prediction model for tensile strength. A dynamic adjustment correction factor for heat source distribution is introduced for correction, and the corrected tensile strength is output.

[0117] S5. Taking the weld geometry and the corrected tensile strength as objectives, the complete model chain constructed in S2-S4 is used as the fitness evaluation engine, and the process parameters are used as decision variables to construct a multi-objective collaborative optimization function and solve the Pareto optimal solution set.

[0118] S6. Collect newly added real welding data from the industrial site, remove abnormal data, and use effective incremental data to trigger adaptive updates of the complete model chain.

[0119] The specific implementation of this invention is as follows:

[0120] Step 1: Construct a dynamic control system for heat source distribution

[0121] This invention first constructs an explicit physical model of process parameters to heat source morphology. The heat source dispersion coefficient is defined as:

[0122]

[0123] λ: Heat source dispersion coefficient This indicates that the distribution is even on both sides; This indicates that the heat is concentrated in the center; This indicates that heat is drawn towards the sidewalls. Heat transferred to the bevel sidewall Heat transferred to the center of the weld.

[0124] Through experiments, we concluded that when λ=0.4, the center is almost burned through while the sidewalls remain cold; when λ=1.6, the sidewalls are overheated, leading to severe collapse of the molten pool. Therefore, [0.5, 1.5] is a feasible range for engineering applications.

[0125] By adjusting λ∈[0.5,1.5], the heat source distribution pattern can be continuously adjusted from a concentrated to a dispersed pattern. This heat source dispersion coefficient λ will serve as the core input variable for all subsequent prediction models, used to establish a quantitative mapping relationship between process parameters and heat source pattern. Here, ω is the heat source dispersion control parameter, used to adjust the energy distribution ratio of the heat source in the bevel width direction. This is achieved by changing the welding arc pattern or heat source oscillation parameters, with a value range of 0.5 to 1.2. In this embodiment, ω=0.8 is taken.

[0126] , Relationship with ω:

[0127]

[0128] in This represents the total heat input.

[0129] In this embodiment, the heat source dispersion coefficient λ is adjusted by the welding torch oscillation device: the oscillation frequency f is set to 2~5Hz, the oscillation amplitude A is set to 3~8mm, and a residence time is introduced. It can independently control the heat of the side walls. A mapping relationship between the system and the oscillation parameters (f, A, dwell time ratio) is established through calibration experiments. Operators only need to input the target λ value on the control interface, and the system automatically calculates the oscillation parameters. The dwell time ratio is defined as... ,in The mapping relationship between λ and ρ, A, and f is established through calibration experiments: in correspond The coefficients were obtained by fitting experimental data using the least squares method, and the typical values ​​are... After the operator inputs the target value λ, the system prioritizes adjustment. ,when If the values ​​exceed the safe range [0.6, 1.8], then further adjust A and f.

[0130] Step 2: Construct a hybrid mapping model for heat source dispersion coefficients

[0131] This step uses welding current I and wire feed speed V. w Welding speed V s Using the heat source dispersion control parameter ω as the input variable, multiple nonlinear regression (MNR) and support vector machine regression (SVR) models are constructed respectively. Then, particle swarm optimization (PSO) is used to determine the fusion weights, outputting the heat source dispersion coefficient λ. Specifically, a hybrid mapping model for the heat source dispersion coefficient is constructed as follows: Figure 2As shown, the heat source dispersion coefficient λ output in this step will serve as a key input to the weld geometry feature prediction model in step three, and also as a correction basis for the calculation of microstructure parameters and the prediction of tensile strength in step four. Multivariate nonlinear regression (MNR) and support vector machine regression (SVR) models are constructed respectively, and then particle swarm optimization (PSO) is used to determine the fusion weights, achieving complementary advantages of the two algorithms.

[0132] 2.1 Multivariate Nonlinear Regression Model (MNR)

[0133]

[0134] Where the initial heat source dispersion coefficient =1.0 is a reference value; (i=1,2,3,4) are the normalized values ​​of welding current, wire feed speed, welding speed, and heat source dispersion control parameters; , , These are regression coefficients, dimensionless normalized regression coefficients; This is a special input index. The normalized reference is: current I = 140A, wire feed speed V. w =8mm / s, welding speed V s =130mm / min, heat source dispersion control parameter =0.8.

[0135] 2.2 Support Vector Machine Regression Model (SVR)

[0136] The SVR model is constructed using the radial basis function (RBF):

[0137]

[0138] The radial basis kernel function , For Lagrange multipliers, : Kernel width parameter; b: Bias term; N: Number of support vectors.

[0139] 2.3 Integrated Mapping Model

[0140] The MNR and SVR models are integrated using a weighted average method, and the combination of models is achieved by optimizing the weights using the PSO algorithm.

[0141]

[0142] Among them, weight , The solution is obtained by minimizing the prediction error using the Particle Swarm Optimization (PSO) algorithm.

[0143]

[0144] for Integrating weights, for The fusion weights, where M is the number of validation samples; Let be the measured heat source dispersion coefficient of the k-th sample; These are the predicted values ​​from the integrated model.

[0145] Among them, weight , The Particle Swarm Optimization (PSO) algorithm is used to solve the problem, aiming to minimize the prediction error. The heat source dispersion coefficient λ output in this step will serve as a key input variable for the weld geometry prediction model in step three, and will also serve as a basis for correcting the calculation of microstructure parameters and tensile strength prediction in step four.

[0146] Step 3: Establish a multi-element algorithm fusion prediction model for weld geometric features

[0147] This step, as the second level of the model chain, expands its input variables into a five-dimensional vector, using welding current I and wire feed speed V as input variables. w Welding speed V s Heat source dispersion control parameters Using the heat source dispersion coefficient λ output from step two as the input variable, a hybrid mapping model for the heat source dispersion coefficient λ is established, integrating multivariate nonlinear regression and support vector machine regression, and employing particle swarm optimization to determine the fusion weights. A prediction model is used for weld penetration depth H (mm), weld width B (mm), and weld reinforcement height h (mm). The prediction results in this step not only serve as a direct evaluation index of weld formation quality but also as the basic input for calculating microstructure parameters in step four.

[0148] During model construction, significant differences in performance were observed among the models under varying data conditions: Under small sample conditions, MNR demonstrated good stability due to its explicit mathematical form and low variance; with increasing data volume, BPNN could capture the strong nonlinear relationship between process parameters and weld geometry, but it was prone to overfitting when the sample distribution was uneven; SVR showed superior generalization ability within a small to medium sample range. However, a single model struggles to maintain optimal performance across all operating conditions. Therefore, this invention employs an adaptive weighted fusion strategy, dynamically adjusting the weights based on the root mean square error of each model on the validation set.

[0149] 3.1 Multivariate Nonlinear Regression Model (MNR)

[0150]

[0151] in The weld geometry features predicted by MNR; These are the regression coefficients; (i=1,2,3,4) are the normalized values ​​of welding current, wire feed speed, welding speed, and heat source dispersion control parameters; For special input indexes;

[0152] This invention constructs a hybrid model that combines data-driven and physical mechanisms by introducing intermediate variables with clear physical meaning (such as the average width of α-martensite). This overcomes the shortcomings of purely data-driven models, such as poor physical interpretability and insufficient predictive ability. This is fundamentally different from existing methods that only use surrogate models such as single RBF.

[0153] 3.2 Backpropagation Neural Network Model (BPNN)

[0154] Construct a three-layer BP neural network structure: 5 nodes in the input layer The hidden layer has 8 nodes, and the output layer has 3 nodes (H, B, h).

[0155] The activation function uses ReLU, and the optimization algorithm uses Adam.

[0156]

[0157] in The network parameters are updated t times cumulatively in the activation function; The learning rate; The first moment deviation correction estimate for the gradient; The second moment deviation correction estimate for the gradient; It is a numerically stable term;

[0158] 3.3 Support Vector Machine Regression Model (SVR)

[0159] An SVR model is constructed using the RBF kernel function to predict weld geometry features, and the input is also expanded to a five-dimensional vector containing λ.

[0160] 3.4 Integrated Prediction Model

[0161] An adaptive weighted fusion algorithm is used:

[0162]

[0163] in The geometric features of the weld predicted by the k-th algorithm, MNR, BPNN, and SVR; The fusion weights of the k-th algorithm (MNR, BPNN, SVR) are proportional to the prediction accuracy; RMSE k The root mean square error (RMSE) of the k-th algorithm (MNR, BPNN, SVR) on the validation set.j Let represent the root mean square error of the j-th prediction algorithm on the validation set.

[0164] The weld geometry features (H, B, h) output in this step will be used as input for the calculation of microstructure parameters in step four, and also as one of the objective functions for multi-objective optimization in step six, to constrain the weld formation quality.

[0165] Step 4: Construct a microstructure-coupled prediction model for the tensile strength of welded joints

[0166] Based on the heat source dispersion coefficient λ output in step two and the weld geometry (H, B, h) output in step three, and combining the welding thermal cycle and microstructure evolution mechanism, this step introduces the α-martensite average width d, unlike the conventional "black box" proxy model that directly maps process parameters and tensile strength. α (μm) and proportion of small-angle grain boundaries (%) is used as an intermediate variable to establish tensile strength. A hybrid prediction model for (MPa) is used. These three variables are physically correlated with the heat source dispersion coefficient λ and weld geometry through a heat conduction model, and directly determine the strengthening mechanism of the titanium alloy. This hybrid structure, with the physical framework first and the neural network for residual correction, enables the model to maintain reasonable physical extrapolation ability even when the process parameters exceed the range of the training set.

[0167] 4.1 Physical Correlation Modeling of Microstructure Parameters

[0168] Based on the heat source dispersion coefficient λ output in step two and the weld geometry (H, B, h) output in step three, combined with the welding thermal cycle and microstructure evolution mechanism,

[0169] d α With equivalent heat input They are positively correlated, among which This is the nominal heat input.

[0170] It is positively correlated with the heat source dispersion coefficient λ;

[0171] Nominal heat input, in kJ / cm;

[0172] U: Arc voltage, unit V;

[0173] Equivalent heat input;

[0174] W HAZ Determined by the peak temperature distribution;

[0175] These parameters together form the physical bridge connecting the welding process and the joint performance.

[0176] 4.2 Physical Mechanism-Driven Model

[0177]

[0178] in The tensile strength predicted by the physical model; These are the regression coefficients; This refers to the wire feeding speed; These are parameters for heat source dispersion control. Reference heat input; R is the activation energy for the β→α phase transition in titanium alloys; R is the gas constant.

[0179] The peak temperature is corrected by combining the heat conduction model with the heat source dispersion coefficient λ:

[0180]

[0181] in The initial temperature; denoted as thermal efficiency coefficient; x as thermal conductivity; r as distance from the center of the heat source; a as thermal diffusivity; and m as the heat source dispersion coefficient. The influence index is obtained through heat source morphology calibration experiments, with a typical value range of 0.3 to 0.5.

[0182] 4.3 Data-Driven Neural Network (BPNN)

[0183] Construct a BP neural network with a 5-10-1 structure, with the input being... The output is the tensile strength σ predicted by the data model. b NN .

[0184] 4.4 Hybrid Model

[0185] Adopting a Stacking integration strategy:

[0186]

[0187] Among them ∈[0,1] represents the fusion coefficient, which is determined by grid search optimization on the training set using 5-fold cross-validation with the objective of minimizing the root mean square error of prediction. In this embodiment, the optimization result is... =0.6, : The final predicted tensile strength.

[0188] 4.5 Define the dynamic regulation correction factor for heat source distribution :

[0189]

[0190] Reference value of heat source dispersion coefficient =1.0 is the reference value, representing the minimum value of the heat source dispersion coefficient. =0.5, the maximum value of the heat source dispersion coefficient =1.5, and the coefficients k1, k2, and k3 were determined by fitting experimental data using the least squares method, with typical values ​​of k1=0.2, k2=0.05, and k3=0.03. The corrected tensile strength is:

[0191]

[0192] Tensile strength adjusted by a correction factor. Due to the periodic heat input generated by the welding torch oscillation on the bevel sidewall, a sine term is introduced into the correction factor to reflect this dynamic effect.

[0193] At this point, the complete predictive model chain constructed from steps two to four has formed a closed loop, providing a unified evaluation benchmark for the multi-objective optimization in step five.

[0194] Step 5: Establish a multi-objective collaborative optimization model

[0195] The weld geometry (H, h) output from step three and the corrected tensile strength output from step four are used. With the objective of using the complete model chain constructed in steps two through four as the fitness evaluation engine and the process parameters as decision variables, a multi-objective optimization function is constructed:

[0196]

[0197] in The objective function vector; The tensile strength is adjusted by a correction factor. This is the predicted melting depth. The remaining height is the predicted value; process parameters I is the welding current; V is the wire feeding speed; s Welding speed; These are the parameters for heat source dispersion and control; the constraints are the safety boundaries of the process parameters and the geometric quality requirements.

[0198] 5.1 Non-dominated sorting genetic algorithm (NSGA-II)

[0199] The NSGA-II algorithm is used to solve for the Pareto optimal solution set:

[0200] Population size N=100

[0201] Crossover probability p c =0.9

[0202] Mutation probability p m =0.1

[0203] Maximum number of iterations G max =200

[0204] 5.2 Particle Swarm Optimization (PSO)-Aided Optimization

[0205] To improve convergence speed and global search capability, the Pareto Optimization (PSO) algorithm is used to pre-optimize the initial population of NSGA-II. The PSO algorithm uses the model chain constructed in steps two through four as the fitness evaluation function and outputs the Pareto front solution set as the initial population input for NSGA-II, forming a hybrid optimization strategy. The particle velocity and position update formulas are as follows:

[0206]

[0207] in : The velocity of the i-th particle in the (t+1)-th iteration; : Position; w: Inertia weight, set to 0.7; Learning factor, set to 1.5; : A random number between [0,1]; : The optimal position of an individual; Global optimal position

[0208] Step Six: Online Model Update and Adaptive Algorithm

[0209] An incremental learning mechanism is introduced to update the aforementioned models online using newly acquired welding data. Furthermore, a causal consistency update mechanism for the model chain is constructed: after the weld geometry prediction model is updated, its predicted geometric features are automatically re-inputted into step four, prompting the tensile strength prediction model to be calibrated synchronously; simultaneously, the heat source dispersion coefficient mapping model is adjusted based on feedback from the new data, ensuring the causal consistency of the model chain and forming a complete closed loop of new data, updates to each sub-model, model chain calibration, and optimization strategy adjustment.

[0210] 6.1 BPNN Online Update

[0211] The model parameters are updated using online gradient descent.

[0212]

[0213] in This refers to the network parameters that are updated cumulatively t+1 times in the backpropagation neural network. The learning rate is the cumulative network parameter update t times in the backpropagation neural network. , =0.01, t is the cumulative number of updates; L is the loss function; x t y tLet be the input and output of the t-th new sample.

[0214] 6.2SVR Incremental Update

[0215] The incremental SVR algorithm is used to update the KKT conditions when a new sample falls into the support vector set:

[0216]

[0217] Where w is the weight vector; C is the regularization parameter; N is the original number of support vectors; and M is the number of new samples. and These are slack variables.

[0218] 6.3 Consistent Updates of the Model Chain

[0219] After the weld geometry prediction model in step three is updated, its predicted geometric features will be re-input into step four, triggering the synchronous calibration of the tensile strength prediction model. At the same time, the heat source dispersion coefficient mapping model in step two will also be fine-tuned based on the feedback of the new data to ensure the causal consistency of the model chain and form a complete closed-loop adaptive system of new data, updates of each sub-model, model chain calibration, and adjustment of optimization strategies.

[0220] Example 1:

[0221] 1. Experimental conditions

[0222] Base material: TC4 titanium alloy, 9mm thick, chemical composition conforms to GB / T3620.1 standard; welding wire: same material, 1.6mm diameter; shielding gas: argon, purity 99.99%, flow rate 25L / min.

[0223] 2. Parameter Settings

[0224] I=140A, =8mm / s, V s =130mm / min, =0.8.

[0225] According to the integrated model calculation in step two, the heat source dispersion coefficient λ = 1.05.

[0226] 3. Prediction of weld geometry features

[0227] The validation set consists of 30 samples; the typical results of sample group 15 are shown here. The comparisons between the predicted and measured values ​​of weld width, penetration, and reinforcement height in this embodiment are as follows: Figure 3 , Figure 4 , Figure 5 As shown, the prediction results of the ensemble model are as follows:

[0228] H=1.20mm, measured 1.20mm, error ≤0.1%

[0229] B = 9.02mm, actual measurement 9.03mm, error ≤ 0.1%

[0230] h=0.49mm, measured 0.49mm, error ≤0.1%

[0231] It should be noted that the above are test results for a typical sample. The average absolute percentage errors of the 30 sets of verification samples were: melt depth 0.35%, melt width 0.42%, and residual height 0.51%, with maximum errors of 1.1%, 1.3%, and 1.5%, respectively.

[0232] The prediction errors of each algorithm are compared in Table 1:

[0233] Table 1

[0234] MNR 1.2% 0.8% 1.5% BPNN 0.5% 0.3% 0.8% SVR 0.8% 0.5% 1.0% ensemble model ≤0.1% ≤0.1% ≤0.1%

[0235] 4. Microscopic tissue characterization

[0236] α-martensite average width d α =0.45μm, small-angle grain boundary ratio =15.8%.

[0237] 5. Tensile strength prediction and verification

[0238] Tensile specimens were prepared according to GB / T228.1 standard, and tensile tests were performed on a universal testing machine. The predicted and measured results are as follows:

[0239] Physical model prediction =945.2MPa;

[0240] BPNN predicts σ b NN =947.5MPa;

[0241] Hybrid model prediction =946.8MPa =0.6);

[0242] Actual measurement =946.1MPa, error ≤0.07%.

[0243] Overall effect verification: Figures 3 to 7 The predicted alignment and weld hardness distribution curves for the above representative samples are presented. For the overall validation set (30 samples), the average prediction error of this ensemble model is extremely low: 0.2% for weld penetration, 0.3% for weld width, 0.4% for weld height, and 0.5% for tensile strength, significantly outperforming single algorithms.

[0244] Example 2: Multi-objective optimization results

[0245] The Pareto optimal solution set was solved using a hybrid algorithm of NSGA-II and PSO. In the PSO warm-up phase, the population size was 30, and 50 iterations were performed. In the NSGA-II main optimization phase, the population size was 100, and 200 iterations were performed, with a crossover probability of 0.9 and a mutation probability of 0.1. The optimization results are shown in Table 2.

[0246] Table 2

[0247] 1 140 8 130 0.8 1.05 1.20 0.49 946.1 2 145 8 125 0.9 1.08 1.32 0.47 948.2 3 135 9 135 0.7 1.02 1.15 0.52 943.5 4 142 8.5 128 0.85 1.06 1.25 0.48 947.3

[0248] As shown in Table 2, the overall performance is better when λ is in the range of 1.02 to 1.08. Considering the melt depth, residual height, and tensile strength, the recommended process parameter combination is: I = 140A, V w =8mm / s, V s =130mm / min, =0.8, λ=1.05. This combination was experimentally reproduced and verified; the standard deviation of tensile strength after 5 tests was 2.1 MPa, indicating good process stability. The preferred operating conditions for the model chain determined in this embodiment are: welding speed Vs≤150 mm / min, and heat source dispersion coefficient λ∈[0.6,1.4]. When the above parameter boundaries are exceeded (e.g., higher welding speeds lead to drastic changes in the molten pool flow pattern), to ensure prediction accuracy (avoiding errors exceeding 2.5%~3.0%), the system can be extended and compensated online by introducing a molten pool fluid dynamics constraint module to adapt to more extreme industrial scenarios.

[0249] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0250] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for dynamically controlling the heat source of titanium alloys during welding and predicting the hybrid performance of joints, characterized in that, include: S1. Define the heat source dispersion coefficient λ as the ratio of the heat transferred to the sidewall of the groove to the heat transferred to the center of the weld. By adjusting the heat source dispersion coefficient λ, the active redistribution of welding heat input in the groove width direction can be achieved. S2. Construct a hybrid mapping model for heat source dispersion coefficient. Based on process parameters, use a multivariate nonlinear regression model and a support vector machine regression model for prediction. Then, determine the fusion weights through a particle swarm optimization algorithm and output the heat source dispersion coefficient λ. S3. Based on the process parameters and the heat source dispersion coefficient λ, establish a multi-source algorithm fusion prediction model for weld geometric features and output the weld geometric features; S4. Based on the heat source dispersion coefficient λ and the weld geometry, calculate the average width of α martensite and the proportion of small-angle grain boundaries as intermediate variables, establish a hybrid prediction model for tensile strength, introduce a dynamic adjustment correction factor for heat source distribution to correct it, and output the corrected tensile strength. S5. Taking the weld geometry and the corrected tensile strength as objectives, using the complete model chain constructed in S2-S4 as the fitness evaluation engine, and the process parameters as decision variables, construct a multi-objective collaborative optimization function and solve the Pareto optimal solution set. S6. Collect newly added real welding data from the industrial site, remove abnormal data, and use the effective incremental data to trigger the adaptive update of the complete model chain.

2. The method as described in claim 1, characterized in that, The heat source dispersion coefficient λ is specifically: ； ； in To transfer heat to the bevel sidewall; To transfer heat to the center of the weld; For total heat input; These are parameters for heat source dispersion control.

3. The method as described in claim 1, characterized in that, In S2: The process parameters are welding current, wire feed speed, and welding speed; The multivariate nonlinear regression model for: ； Where the initial heat source dispersion coefficient =1.0 is a reference value; For welding current, wire feed speed, welding speed, and heat source dispersion control parameters, i = 1, 2, 3, 4; , , These are regression coefficients, dimensionless normalized regression coefficients; For special input indexes; The support vector machine regression model for: ； ； in It is a radial basis kernel function; For Lagrange multipliers; is the kernel width parameter; b is the bias term; N is the number of support vectors.

4. The method as described in claim 1, characterized in that, The specific process of determining the fusion weights using the particle swarm optimization algorithm in S2 is as follows: An integrated multivariate nonlinear regression model using the weighted average method With support vector machine regression model : ； in Fusion weights and Fusion weights The solution is obtained by minimizing the prediction error using the particle swarm optimization algorithm. ； Where M is the number of validation samples; Let be the measured heat source dispersion coefficient of the k-th sample; These are the predicted values ​​from the integrated model.

5. The method as described in claim 1, characterized in that, The specific process for establishing the multi-source algorithm fusion prediction model for weld geometric features in S3 is as follows: The weld geometry features include the predicted weld penetration H, the predicted weld width B, and the predicted weld height h. S31. Establish three prediction sub-models respectively: Multivariate nonlinear regression model MNR: ； in The weld geometry features predicted by MNR; These are the regression coefficients; For welding current, wire feed speed, welding speed, and heat source dispersion control parameters, i = 1, 2, 3, 4; For special input indexes; Backpropagation Neural Network Model (BPNN): A three-layer network structure is adopted, with the input layer consisting of 5 nodes. The hidden layer contains 8 nodes, and the output layer contains 3 nodes H, B, and h. The activation function in the hidden layer is ReLU, and the optimization algorithm is Adam. ； in The network parameters are updated t times cumulatively in the activation function; The network parameters are updated cumulatively t+1 times in the activation function; t represents the cumulative number of updates; The learning rate; The first moment deviation correction estimate for the gradient; The second moment deviation correction estimate for the gradient; For numerical stability terms; I is the welding current; V w V is the wire feeding speed; s Welding speed; These are parameters for heat source dispersion control. Support Vector Machine Regression Model (SVR): An SVR model is constructed using the radial basis function (RBF) to predict weld geometry features; S32, Integrated Prediction Model F: An adaptive weighted fusion algorithm is used to calculate the root mean square error (RMSE) of each prediction sub-model on the validation set, and weights are dynamically allocated. ； in The geometric features of the weld predicted by the k-th algorithm, MNR, BPNN, and SVR; The fusion weights of the k-th algorithm (MNR, BPNN, SVR) are proportional to the prediction accuracy; RMSE k RMSE represents the root mean square error (RMSE) of the k-th algorithm (MNR, BPNN, SVR) on the validation set. j Let represent the root mean square error of the j-th prediction algorithm on the validation set.

6. The method as described in claim 1, characterized in that, The specific process for establishing the hybrid prediction model for tensile strength in S4 is as follows: S41. Based on the heat source dispersion coefficient λ and the weld geometry, and combined with the welding thermal cycle theory, calculate three key microstructure characterization variables: α-martensite average width d α : with equivalent heat input There is a positive correlation, where nominal heat input I represents the welding current; V represents the welding current. s Welding speed; U is the arc voltage; Small-angle grain boundary ratio : with heat source dispersion coefficient Positively correlated; S42. Establish a physical mechanism-driven model: ； in The tensile strength predicted by the physical model; These are the regression coefficients; This refers to the wire feeding speed; These are parameters for heat source dispersion control. Reference heat input; The activation energy for the β→α phase transformation of titanium alloys; R is the gas constant; The peak temperature is corrected by combining the heat conduction model with the heat source dispersion coefficient λ: ； in The initial temperature; denoted as thermal efficiency coefficient; x is thermal conductivity; r is distance from the center of the heat source; a is thermal diffusivity; m is the influence index of the heat source dispersion coefficient. S43. Establish the digital-driven model BPNN: Constructing a BP neural network: 5 nodes in the input layer The hidden layer has 10 nodes, using ReLU as the activation function; the output layer has 1 node σ. b NN ;where σ b NN The tensile strength predicted by the data model; S44. Employ a Stacking integration strategy to fuse the physical mechanism-driven model and the digital-driven model: ； in The predicted tensile strength; ∈[0,1] represents the fusion coefficient.

7. The method as described in claim 1, characterized in that, The specific process of introducing a dynamic adjustment correction factor for heat source distribution is as follows: Define the dynamic regulation correction factor for heat source distribution : ； Reference value of heat source dispersion coefficient =1; Minimum value of heat source dispersion coefficient =0.5; Maximum value of heat source dispersion coefficient =1.5; coefficients k1, k2, and k3 were determined by fitting experimental data using the least squares method; The corrected tensile strength is: The tensile strength is adjusted by a correction factor. This represents the predicted tensile strength.

8. The method as described in claim 1, characterized in that, The multi-objective collaborative optimization function constructed in S5 is as follows: ； in The objective function vector; The tensile strength is adjusted by a correction factor. This is the predicted melting depth. The remaining height is the predicted value; process parameters I is the welding current; V is the wire feeding speed; s Welding speed; These are parameters for heat source dispersion control.

9. The method as described in claim 1, characterized in that, The solution to the Pareto optimal set in S5 also includes: A non-dominated sorting genetic algorithm assisted by particle swarm optimization was used to solve for the Pareto optimal process parameter solution set. The non-dominated sorting genetic algorithm has a population size of 100, a crossover probability of 0.9, a mutation probability of 0.1, and a maximum number of iterations of 200. The initial population is pre-optimized using a particle swarm optimization algorithm.

10. The method as described in claim 1, characterized in that, The specific process of adaptive update described in S6 is as follows: For the backpropagation neural network in the complete model chain, the parameters are updated using online gradient descent: ； in This refers to the network parameters that are updated cumulatively t+1 times in the backpropagation neural network. The learning rate is the cumulative network parameter update t times in the backpropagation neural network. , =0.01, t is the cumulative number of updates; L is the loss function; x t y t Let be the input and output of the t-th new sample; The incremental support vector machine regression algorithm is used. When a new sample falls into the support vector set or violates the KKT conditions, the parameters are updated incrementally by solving an augmented optimization problem. ； Where w is the weight vector; C is the regularization parameter; N is the original number of support vectors; and M is the number of new samples. and These are slack variables.

Images