Twin-wire welding process parameter prediction method based on machine learning method

Abstract

The present invention relates to the technical field of intelligent welding and welding processes, and in particular to a twin-wire welding process parameter prediction method based on a machine learning method. The prediction method comprises the following steps: acquiring welding parameter data and weld bead profile parameter data by means of welding process tests, preprocessing the data, and establishing a dataset; constructing a back propagation neural network model on the basis of the dataset, constructing a support vector machine model, and using the dataset to train the back propagation neural network model and the support vector machine model; evaluating the trained back propagation neural network model and the trained support vector machine model, and selecting an optimal model as a model for actual industrial application; and inputting, into the selected model, parameters required for a weld bead to be welded, to obtain twin-wire welding process parameters. By predicting welding process parameters on the basis of the machine learning method, the present invention improves quality stability and efficiency of twin-wire welding.

Description

A Machine Learning-Based Method for Predicting Twin-Wire Welding Process Parameters Technical Field

[0001] This invention relates to the field of intelligent welding and welding process technology, and more specifically, to a method for predicting process parameters of dual-wire welding based on machine learning. Background Technology

[0002] The existing welding processes for large components in my country are still mainly based on traditional shielded metal arc welding (SMAW), submerged arc welding (SAW), and CO₂ gas shielded welding (MAG). These processes suffer from low mechanization, low welding efficiency, and large fluctuations in welding quality, making it difficult to meet the welding requirements of large components and hindering the development of China's heavy machinery equipment manufacturing industry. The reason for this is that the welding bevels of large components generally vary considerably, involving multi-layer, multi-pass welding with different filler amounts in each pass. Therefore, this paper proposes a method for predicting the process parameters of dual-wire welding based on machine learning. Summary of the Invention

[0003] The purpose of this invention is to provide a method for predicting the process parameters of twin-wire welding based on machine learning, in order to solve the problems mentioned in the background art of the complex process parameters in the existing large component welding technology, which makes it difficult to predict welding parameters in batches, thus leading to unstable welding quality and low welding efficiency.

[0004] To achieve the above objectives, the present invention aims to provide a method for predicting process parameters of twin-wire welding based on machine learning, comprising the following steps:

[0005] S1. Obtain welding parameters and weld morphology parameters through welding process experiments, preprocess the data, and establish a dataset;

[0006] S2. Construct a backpropagation neural network model based on the dataset, and construct a support vector machine model. Use the dataset to train the backpropagation neural network model and the support vector machine model.

[0007] S3. Evaluate the trained backpropagation neural network model and support vector machine model, select the optimal model as the model for actual industrial application, and introduce the input parameter sensitivity coefficient matrix and exponential weight allocation mechanism for optimization during the evaluation process;

[0008] S4. Input the required parameters for the weld bead to be welded into the selected model to obtain the dual-wire welding process parameters, and send the process parameters to the welding machine control system to complete the welding.

[0009] As a further improvement to this technical solution, in step S1, welding parameters and weld morphology parameters are obtained through welding process experiments, the data is preprocessed, and a dataset is established, including the following steps:

[0010] S1.1 Obtain welding condition parameters and welding process parameters, including the front wire feed speed, the rear wire feed speed and the welding speed.

[0011] S1.2 Cut the weld section, take pictures of the weld morphology using a microscope, and then obtain the weld reinforcement height and weld width data.

[0012] S1.3 Preprocess the obtained data on front wire feed speed, rear wire feed speed, welding speed, weld reinforcement height and weld width;

[0013] S1.4 Divide the dataset into training set, validation set and test set, and save the dataset.

[0014] As a further improvement to this technical solution, step S2, which involves constructing a backpropagation neural network model based on the dataset, includes the following steps:

[0015] S2.1. Set the number of input layer nodes according to the number of input features, determine the number of hidden layers and the number of nodes in each layer, and determine the number of output layer nodes. The input features include weld reinforcement height, weld width and welding speed.

[0016] S2.2, Randomly initialize the weights and biases between the nodes of the input layer, hidden layer, and output layer;

[0017] S2.3. Propagate the input features forward through the backpropagation neural network, calculate the output value of each input layer, hidden layer and output layer node, until the predicted values ​​of the front yarn feed speed and back yarn feed speed output by the backpropagation neural network are obtained.

[0018] S2.4 Starting from the output layer, calculate the gradient of each weight and bias with respect to the loss function layer by layer, and adjust the value of each weight and bias according to the gradient;

[0019] S2.5. Update the weights and biases using the gradient descent algorithm;

[0020] S2.6 Repeat steps S2.3 to S2.5 above until the maximum number of iterations t1 is reached.

[0021] As a further improvement to this technical solution, in S2.5, the gradient descent algorithm is as follows:

[0022] ;

[0023] ;

[0024] in, Indicates weight; , Indicates the node index; Indicates bias; Represents the learning rate; represents the current iteration number; Represents the loss function Weights The partial derivatives; Represents the loss function For bias The partial derivatives; This represents the loss function.

[0025] As a further improvement to this technical solution, in step S2, the support vector machine model is constructed, including the following steps:

[0026] S2.7 Set the input variables as weld reinforcement height, weld width and welding speed, and set the output variables as front wire feed speed and rear wire feed speed;

[0027] S2.8 Select the radial basis function as the kernel function for the support vector machine model;

[0028] S2.9 Optimize the objective function and determine the optimization parameters;

[0029] S2.10. Use the radial basis function kernel to train the support vector regression model on the training set.

[0030] As a further improvement to this technical solution, in S2.9, the objective function is:

[0031] ;

[0032] Where represents the objective function; Represents the weight vector; Represents the regularization parameter; This represents the total number of samples in the dataset; Represents slack variables; This indicates the sample index of the dataset.

[0033] As a further improvement to this technical solution, in step S3, the trained backpropagation neural network model and support vector machine model are evaluated, and the optimal model is selected as the model for actual industrial applications. This includes the following steps:

[0034] S3.1 Load the pre-trained backpropagation neural network model and support vector machine model;

[0035] S3.2 Input the input data from the same test set into the backpropagation neural network model and the support vector machine model respectively, and obtain the prediction results of the backpropagation neural network model and the support vector machine model for the front yarn feed speed and the back yarn feed speed of the test set.

[0036] S3.3 Calculate the difference between the prediction results of the backpropagation neural network model and the support vector machine model and the true value by means of mean square error. In the mean square error calculation, the input parameter sensitivity coefficient matrix and the exponential weight allocation mechanism are introduced for optimization. The determination coefficients between the prediction results of the backpropagation neural network model and the support vector machine model and the true value are calculated.

[0037] S3.4 If the mean squared error between the prediction result and the true value of the backpropagation neural network model is less than the mean squared error between the prediction result and the true value of the support vector machine model, and the coefficient of determination of the backpropagation neural network model is greater than the coefficient of determination of the support vector machine model, then the backpropagation neural network model is selected; if the mean squared error between the prediction result and the true value of the backpropagation neural network model is greater than the mean squared error between the prediction result and the true value of the support vector machine model, and the coefficient of determination of the backpropagation neural network model is less than the coefficient of determination of the support vector machine model, then the support vector machine model is selected.

[0038] As a further improvement to this technical solution, in S3.3, the mean square error is:

[0039] ;

[0040] in, Indicates mean square error; Indicates the number of samples in the test set; Indicates the index of the test set samples; Represents the actual value; Indicates the prediction result;

[0041] To evaluate the strength of the influence of input parameters on output parameters, the local gradient of the input parameters with respect to the output is calculated, and the sensitivity coefficient matrix of the input parameters for each sample is also calculated. ;

[0042] The input parameters are weld reinforcement height, weld width, and welding speed, while the output parameters are the front wire feed speed and the rear wire feed speed.

[0043] ;

[0044] in, This represents the local gradient of the input parameters with respect to the output parameters; This represents the normalization factor for process parameters; Indicates the weight of the coupling effect; This represents the gradient of the effect of input parameters from samples not currently in the test set on the output. , Indicates the index of the test set samples;

[0045] An exponential weighting mechanism is introduced to account for the influence of the sensitivity of weld reinforcement, weld width, and welding speed parameters on the final error calculation.

[0046] ;

[0047] in, Indicates the dynamic adjustment coefficient; Indicates test set samples For output parameters Sensitivity; Indicates the index of the output parameter; Indicates the index of the test set samples; Indicates weight;

[0048] The optimized formula for calculating the mean squared error, after introducing an input parameter sensitivity coefficient matrix and an exponential weighting mechanism, is as follows:

[0049] ;

[0050] in, This represents the optimized mean squared error; Indicates the weld reinforcement; Indicates the weld width; Indicates welding speed; This indicates the priority coefficient.

[0051] As a further improvement to this technical solution, in S3.3, the determination coefficient is:

[0052] ;

[0053] in, This represents the mean of all true values; This represents the coefficient of determination.

[0054] As a further improvement to this technical solution, in step S4, the parameters required for the weld bead to be welded are input into the selected model to obtain the dual-wire welding process parameters, and the process parameters are sent to the welding machine control system to complete the welding, including the following steps:

[0055] S4.1 Obtain the size information of the actual welding target bevel to be filled;

[0056] S4.2 Obtain the number of weld passes to be filled and set the welding speed for each weld pass;

[0057] S4.3. Arrange the weld beads on the bevel and give the weld bead height and weld width for each weld bead;

[0058] S4.4 Input the morphology parameters of multiple weld beads and welding speeds of the corresponding target bevel to be filled into the selected model in sequence, and use the selected trained model to output the predicted front wire feed speed and rear wire feed speed.

[0059] S4.5 Input the welding speed and the predicted wire feed speeds output by the model into the TPS / i TWIN Push welding robot system.

[0060] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0061] This machine learning-based method for predicting welding process parameters in dual-wire welding improves the quality stability and efficiency of the welding process. Compared to traditional manual arc welding, the predicted welding parameters are stable, do not rely on welder experience, and can ensure consistent quality in automated production. Furthermore, the design process for traditional process parameters is complex and cumbersome, with poor reusability; this invention significantly improves prediction efficiency and simplifies the workload of process engineers. Existing conventional machine learning parameter prediction in the welding field generally predicts welding quality based on welding parameters. This invention, however, breaks through this limitation by predicting welding process parameters through weld morphology and welding speed, broadening the application scope of machine learning in the welding field and enhancing its value. Attached Figure Description

[0062] Figure 1 is a flowchart of the overall method of the present invention;

[0063] Figure 2 is a structural diagram of the backpropagation neural network in the embodiment. Detailed Implementation

[0064] 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0065] Example: Please refer to Figures 1-2. This example provides a method for predicting process parameters of twin-wire welding based on machine learning, including the following steps:

[0066] S1. Obtain welding parameters and weld morphology parameters (welding speed, wire feeding speed of the first wire in double-wire welding, wire feeding speed of the second wire in double-wire welding, weld reinforcement and weld width) through welding process experiments, preprocess the data, and establish a dataset;

[0067] In this embodiment, the welding process experiment is a gas metal arc welding (GMAW) twin-wire welding. The process experiment is conducted by changing parameters such as welding speed, front wire feed speed, rear wire feed speed, weld angle, and plate tilt angle. After the experiment, the weld width and weld reinforcement dimensions are obtained. This embodiment uses the TPS / i TWIN Push twin-wire welding system. Furthermore, a single-wire feed speed welding method is adopted, so the obtained welding process parameters only include three categories: front wire feed speed, rear wire feed speed, and welding speed. These three categories of welding process parameters have significant influence and are highly operable; therefore, these three categories of process parameters are selected for machine learning.

[0068] The process involves obtaining welding parameters and weld morphology parameters (welding speed, wire feed speed of the first wire in double-wire welding, wire feed speed of the second wire in double-wire welding, weld reinforcement height, and weld width) through welding process experiments. The data is then preprocessed, and a dataset is established, including the following steps:

[0069] S1.1 Obtain welding condition parameters and welding process parameters. Welding condition parameters refer to the inherent welding properties of the welded joint and the base material, including the base material grade, welding wire grade, welding method, etc. The selection of the above welding parameters has a great impact on the welding quality. However, in the actual application scenario of multi-layer and multi-pass welding of large components, such parameters will not change and cannot be adjusted. Therefore, after obtaining them, they are used as usage conditions and are not involved in subsequent machine learning. Among them, welding process parameters include front wire feed speed, rear wire feed speed and welding speed.

[0070] S1.2 Cut the weld section, take pictures of the weld morphology using a microscope, and then use an image processing program (Image-Pro Plus) to obtain the weld reinforcement height and weld width data;

[0071] S1.3. Preprocess (clean) the obtained front wire feed speed, rear wire feed speed, welding speed, weld reinforcement height and weld width data, including removing outliers, removing duplicate data, and processing missing values ​​to obtain a dataset; outliers may be due to sensor failure or other reasons and should be removed; if duplicate data exists, only one copy should be kept.

[0072] S1.4 Divide the dataset into training, validation, and test sets, and save the dataset. The training set is used to train the machine learning model parameters, the validation set is used to adjust the hyperparameters of the machine learning model and evaluate the model performance, and the test set is used to finally evaluate the generalization ability of the machine learning model. Save the preprocessed dataset so that the machine learning model can be trained and applied in the future.

[0073] S2. Construct a backpropagation neural network model based on the dataset, and construct a support vector machine model. Use the dataset to train the backpropagation neural network model and the support vector machine model to establish the mapping relationship between welding speed, wire feeding speed and weld morphology.

[0074] Backpropagation (BP) neural networks are multilayer feedforward neural networks trained using an error backpropagation algorithm. They are among the most widely used neural network models, with gradient descent as their core principle. This algorithm consists of two processes: forward propagation of the signal and backward propagation of the error. During forward propagation, the input signal acts on the output node through the hidden layers, undergoing a nonlinear transformation to generate the output signal. If the actual output differs from the expected output, the error is propagated back along the original connection path, and the error signal is minimized by modifying the weights of each neuron. The structure of a BP neural network includes an input layer, several hidden layers, and an output layer. Neurons within each layer are fully connected, but there are no connections between neurons within the same layer, and no feedback connections between neurons across layers. Nodes and connections: Each node represents a neuron, and the connection strength between nodes is represented by weights. Additionally, each node has a threshold to control the activation state of the neuron.

[0075] Neural network models have a three-layer structure: input layer, hidden layers, and output layer. The input layer is the first layer of the neural network, used to receive and transmit input data from the dataset. Hidden layers, located between the input and output layers, can be multiple or a single layer. The number of neurons in the hidden layer is generally determined by empirical formulas. Hidden layers introduce nonlinear transformations, such as activation functions, to help the neural network model nonlinear relationships, thereby approximating arbitrary nonlinear functions. Furthermore, they can map input data to a higher-dimensional feature space, helping the network better understand the data by learning and extracting its features. The output layer is the last layer of the network model, used to output the results of the hidden layer operations. Each neuron adjusts its weights and thresholds through iterative training and transmits information between different layers. Neural networks can automatically learn and extract features from large amounts of sample data, thus enabling accurate predictions.

[0076] The neural network model in this embodiment has an input layer containing three neurons (as shown in Figure 2); a hidden layer containing eight neurons; and an output layer containing two neurons. The input values ​​are weld width, weld reinforcement, and welding speed, and the output values ​​are the front wire feed speed and the rear wire feed speed. Through training, the model can learn the relationship between different welding parameters and weld morphology, and predict the corresponding combination of front and rear wire feed speeds based on weld morphology parameters and welding speed.

[0077] During the training process in this embodiment, the model adjusts the network weights and biases through the backpropagation algorithm to minimize the loss function between the predicted result and the true label; enabling the neural network to learn the complex relationship between input features and output results.

[0078] In this embodiment, the backpropagation neural network model is constructed based on the dataset, including the following steps:

[0079] S2.1. Based on the number of input features, set the number of nodes in the input layer (three types of parameters: weld reinforcement height, weld width, and welding speed, then the input layer has 3 nodes), determine the number of hidden layers and the number of nodes in each layer, and determine the number of nodes in the output layer (corresponding to the number of predicted targets, two targets: front wire feed speed and rear wire feed speed, then the output layer has 2 nodes). The input features include weld reinforcement height, weld width, and welding speed.

[0080] S2.2, Randomly initialize the weights and biases between the nodes of the input layer, hidden layer, and output layer;

[0081] S2.3. The input features are propagated forward through the backpropagation neural network, and the output values ​​of each input layer, hidden layer and output layer node are calculated until the predicted values ​​of the front and back wire feed speeds output by the backpropagation neural network are obtained. This process involves multiplying the input signal with weights and then calculating the output of each neuron through the activation function.

[0082] S2.4 Starting from the output layer, calculate the gradient of each weight and bias with respect to the mean squared error (MSE) loss function layer by layer, and adjust the value of each weight and bias according to the gradient. This process is called backpropagation. It is based on the chain rule to calculate the contribution of each parameter to the final loss. Specifically, for each weight and bias, calculate its partial derivative with respect to the loss function.

[0083] S2.5 Update weights and biases using the gradient descent algorithm. Gradient descent is an iterative method for optimizing functions. Its main purpose is to minimize an objective function (usually called a loss function or cost function), which measures the difference between the model's predictions and the actual data. By adjusting the model parameters, the loss function is minimized, thereby achieving the best fit to the model.

[0084] Furthermore, the gradient descent algorithm is as follows:

[0085] ;

[0086] ;

[0087] in, Represents the weights, which are the parameters connecting the layers of a neural network; , Indicates the node index; The bias is an additional parameter for each neuron used to adjust the activation level of that neuron. Indicates the learning rate; Indicates the current iteration number; Represents the loss function Weights The partial derivative (gradient) represents the rate of change of the loss function with respect to that weight; Represents the loss function For bias The partial derivative (gradient) represents the rate of change of the loss function with respect to that bias; This represents the loss function, which is the mean squared error loss function.

[0088] S2.6 Repeat steps S2.3 to S2.5 above until the maximum number of iterations t1 is reached. Improve the performance of the neural network through multiple iterations of training. In each iteration, by inputting training data into the network, forward propagation, loss function calculation, backpropagation and parameter update are performed until the predetermined stopping condition is reached (such as reaching the maximum number of iterations or the loss function converges). Through the above steps, the backpropagation algorithm can iteratively adjust the weights and biases of the network, so that the neural network gradually approaches the optimal solution and improves the model's ability to learn the complex relationship between input features and output results.

[0089] Support Vector Machine (SVM) is a supervised learning model. For linearly separable datasets, the core idea of ​​SVM is to find a hyperplane (in a high-dimensional space) that can separate data points of different classes as much as possible, while maximizing the distance between this separating boundary and the nearest data point. SVM attempts to find an optimal hyperplane to completely separate the two classes of data. The selection of this hyperplane is based on the principle of maximizing the margin, that is, maximizing the distance to the nearest data point (support vector). When the data cannot be directly separated by a straight line or a hyperplane, SVM uses the kernel trick to map the original input space to a higher-dimensional space, where it searches for a hyperplane that can achieve linear separation of the data.

[0090] Building a Support Vector Machine (SVM) model includes the following steps:

[0091] S2.7 Set the input variables as weld reinforcement height, weld width and welding speed, and set the output variables as front wire feed speed and rear wire feed speed;

[0092] S2.8. The Radial Basis Function (RBF) kernel is selected as the kernel function for the Support Vector Machine (SVM) model. The RBF kernel, also known as the Gaussian kernel, is a commonly used kernel function in SVM. It allows the original linear classification method to be applied to non-linearly separable datasets by mapping data in a low-dimensional space to a higher-dimensional space, where a hyperplane is found to separate the data. The RBF kernel function has the following form: ,in, Indicates sample and Kernel function values ​​between; This represents the parameter that controls the width of the kernel function, affecting how well the model fits the data;

[0093] S2.9 Optimize the objective function and determine the optimization parameters. The core purpose of optimizing the objective function is to find the optimal model parameter configuration so that the support vector regression (SVR) model can have good generalization ability while ensuring prediction accuracy. The optimization parameters include regularization parameters, kernel function parameters, etc.

[0094] The objective function is:

[0095] ;

[0096] in, Let represent the objective function, and let represent the function to be optimized. The goal is to find the optimal hyperplane that minimizes the regression error while ensuring that the model complexity is moderate. This represents the weight vector, used to define the regression hyperplane of Support Vector Regression (SVR); This represents the regularization parameter; a larger value indicates a higher regularization parameter. A larger penalty for error means the model tends to fit the training data more closely, but may overfit. It has a higher tolerance for error and the model is smoother, but it may underfit. This represents the total number of samples in the dataset; Represents slack variables. and Corresponding to samples respectively Exceeding above and below the actual value The scope part, For insensitive loss functions, It is a small positive number set by the user based on the specific needs of the problem and the characteristics of the data, used for definition. - Insensitive region, meaning that prediction error is ignored in this region; Indicates the sample index;

[0097] S2.10. Train the support vector regression model on the training set using the radial basis function (RBF).

[0098] S3. Evaluate the trained backpropagation neural network model and support vector machine model, select the optimal model as the model for actual industrial application, and introduce the input parameter sensitivity coefficient matrix and exponential weight allocation mechanism for optimization during the evaluation process;

[0099] In this embodiment, the trained backpropagation neural network model and support vector machine model are evaluated, and the optimal model is selected as the model for practical industrial applications. This includes the following steps:

[0100] S3.1 Load the pre-trained backpropagation neural network model and support vector machine model;

[0101] S3.2 Input the input data from the same test set into the backpropagation neural network model and the support vector machine model respectively, and obtain the prediction results of the backpropagation neural network model and the support vector machine model for the front yarn feed speed and the back yarn feed speed of the test set.

[0102] S3.3 Calculate the difference between the predicted results and the true values ​​of the backpropagation neural network model and the support vector machine model using mean square error (MSE). In the MSE calculation, an input parameter sensitivity coefficient matrix and an exponential weight allocation mechanism are introduced for optimization (the sensitivity coefficient matrix and weight allocation mechanism mainly affect the strength analysis of the influence of input parameters on the output results, and thus affect the optimization process of mean square error (MSE). In the model evaluation stage, by introducing these mechanisms, the quality of the model prediction results can be measured more accurately, especially considering the differences in the importance of different input parameters to the prediction of the final welding process parameters). Calculate the coefficients of determination between the predicted results and the true values ​​of the backpropagation neural network model and the support vector machine model.

[0103] Mean Squared Error (MSE) is a loss function that provides a direct and intuitive way to quantify the difference between model predictions and actual observations. Because it is based on the average of squared errors, it clearly reflects the overall performance of the model. Since MSE uses a squared form to calculate the error, it gives higher weight to larger prediction errors. This means that if the model's predictions for some samples deviate significantly from the true values, these errors will be significantly amplified in the final MSE calculation, making them easier to detect and analyze. This is particularly useful for identifying weaknesses in model performance.

[0104] The mean square error is:

[0105] ;

[0106] in, Indicates mean square error; Indicates the number of samples in the test set; Indicates the index of the test set samples; This represents the actual values, namely the actual front yarn feed speed and the rear yarn feed speed; This indicates the prediction results, namely the front and back yarn feed speeds predicted by the backpropagation neural network model or support vector machine model.

[0107] Welding process parameters (such as weld reinforcement, weld width, and welding speed) have varying degrees of influence on target parameters (front / back wire feed speed). A small change in weld width can lead to a significant adjustment in wire feed speed, while the impact of welding speed may be relatively minor. Treating the prediction errors of all parameters equally in model evaluation can mask errors in critical parameters, affecting the final weld quality. A sensitivity coefficient matrix is ​​generated by calculating local gradients (i.e., the intensity of the impact of changes in input parameters on the output). This matrix quantifies the sensitivity of each input parameter to the output parameter; an exponential weighting mechanism is introduced into the mean squared error (MSE) to make the model pay more attention to the parameters that have a greater impact on the output during training. The prediction error of highly sensitive parameters will be amplified, thereby forcing the model to prioritize optimizing the prediction accuracy of these parameters.

[0108] To assess the influence of input parameters on output parameters, the local gradient of the input parameters on the output is calculated. (This is done using a finite difference model to simulate the welding process. The local gradient of the input parameters on the output is estimated based on the results of the welding thermodynamic model. This means that, firstly, the output results under a series of different input conditions are obtained using the welding thermodynamic model, and then the sensitivity of each input variable to the output variable, i.e., the local gradient, is estimated based on these data points. The welding thermodynamic model is as follows:) ,in, This indicates the temperature of the weld and its surrounding area at different points in time during the welding process. Indicates the density of the material; Indicates specific heat capacity; Indicates thermal conductivity; The term represents the heat source, indicating the heat generated per unit volume per unit time (including energy inputs such as laser beams and electric arcs), and calculates the sensitivity coefficient matrix of the input parameters for each sample. (Reflects the intensity of the impact of minute changes in parameters on the output);

[0109] The input parameters are weld reinforcement height, weld width, and welding speed, while the output parameters are the front wire feed speed and the rear wire feed speed.

[0110] ;

[0111] in, This represents the local gradient of the input parameters with respect to the output parameters, achieved through finite element simulation. This represents the normalization factor for process parameters; This represents the coupling effect weight, which adjusts the contribution weight of the synergistic effect of multiple parameters to sensitivity; This represents the gradient of the effect of input parameters from samples not currently in the test set on the output. , Indicates the index of the test set samples;

[0112] An exponential weighting mechanism is introduced to consider the impact of the sensitivity of weld reinforcement height, weld width, and welding speed parameters on the final error calculation (in actual welding, the requirements for parameter accuracy may differ at different process stages (the first weld bead needs strict control of the weld width, while the filler layer focuses more on the reinforcement height). The exponential weighting mechanism is introduced based on actual process requirements (priority coefficient). (Dynamically adjust the error weights of different output parameters (front yarn speed, back yarn speed) to make the model more flexible and adaptable to diverse scenarios):

[0113] ;

[0114] in, This represents the dynamic adjustment coefficient, which controls the intensity of the influence of sensitivity on the weight. Indicates test set samples For output parameters Sensitivity; Indicates the index of the output parameter; Indicates the index of the test set samples; Indicates weight;

[0115] The optimized formula for calculating the mean squared error, after introducing an input parameter sensitivity coefficient matrix and an exponential weighting mechanism, is as follows:

[0116] ;

[0117] in, This represents the optimized mean squared error; Indicates the weld reinforcement; Indicates the weld width; Indicates welding speed; This represents the priority coefficient, which sets the importance weight of different output parameters according to process requirements (e.g., the front yarn speed has a higher weight).

[0118] The coefficient of determination (COD) is a statistic that measures the goodness of fit of a regression model. It reflects the proportion of variability explained by the model relative to the total variability; specifically, it represents the percentage of data fluctuations the model can explain. Its value ranges between 0 and 1. The COD provides an intuitive way to measure a model's ability to explain data variation. It directly reflects the proportion of data fluctuations the model can explain; the closer the value is to 1, the better the model fits the data and the better it can explain changes in the target variable.

[0119] The coefficient of determination is:

[0120] ;

[0121] in, This represents the mean of all true values; The coefficient of determination measures the goodness of fit of a model, i.e., the model's ability to explain the variance of the data.

[0122] S3.4 If the mean squared error between the prediction result and the true value of the backpropagation neural network model is less than the mean squared error between the prediction result and the true value of the support vector machine model, and the coefficient of determination of the backpropagation neural network model is greater than the coefficient of determination of the support vector machine model, then the backpropagation neural network model is selected; if the mean squared error between the prediction result and the true value of the backpropagation neural network model is greater than the mean squared error between the prediction result and the true value of the support vector machine model, and the coefficient of determination of the backpropagation neural network model is less than the coefficient of determination of the support vector machine model, then the support vector machine model is selected.

[0123] In this embodiment, if a certain dual-wire welding test set contains the following 3 sets of samples, the actual process parameters and model prediction values ​​are as follows:

[0124] Sample 1: Input parameters: excess height 2.1mm, melt width 8mm, speed 0.8m / min; Actual output: front wire 6.2m / min, back wire 5.0m / min; BP prediction output: front wire 6.0, back wire 4.9; SVM prediction output: front wire 5.8, back wire 5.2;

[0125] Sample 2: Input parameters: excess height 1.8mm, melt width 7.5mm, speed 1.0m / min; Actual output: front wire 5.5m / min, back wire 4.6m / min; BP prediction output: front wire 5.6, back wire 4.5; SVM prediction output: front wire 5.3, back wire 4.8;

[0126] Sample 3: Input parameters: excess height 2.4mm, melt width 9mm, speed 0.7m / min; Actual output: front wire 6.8m / min, back wire 5.4m / min; BP prediction output: front wire 6.7, back wire 5.3; SVM prediction output: front wire 6.2, back wire 5.6;

[0127] The calculation steps are as follows: Calculate the traditional MSE:

[0128] BP model MSE (precursor silk): ;

[0129] SVM model MSE (front silk): ;

[0130] The above BP is clearly superior;

[0131] Introducing sensitivity optimization ( The following is:

[0132] Gradient calculations revealed that the excess height has the highest sensitivity coefficient to the front yarn speed. ), with melt width being the second largest ( ), lowest speed ( );

[0133] Index weighting (if) Sample 1 weights: (led by Yu Gao), the weighted BP's SWDMSE further decreased to 0.015, and SVM increased to 0.20;

[0134] Calculation of the coefficient of determination R²:

[0135] BP model R²: ;

[0136] SVM model R²: ;

[0137] The above, This demonstrates that BP can explain 98% of the data variation;

[0138] In summary, complex nonlinear coupling exists between the feed speed of the front and rear wires and the morphology parameters in dual-wire welding (dynamic interaction of the molten pool). BP neural networks can capture higher-order features through multiple hidden nodes. The welding height significantly affects the wire feed speed (for sample 1, a height of 2.1 mm requires precise control of the front wire). BP learns the sensitivity weights more accurately through gradient propagation, resulting in more stable performance in weighted error calculation. If the process test dataset is large (e.g., >1000 sets), the generalization ability of BP increases with the amount of data, while SVM is prone to overfitting in high-dimensional, small-sample scenarios. In this embodiment, the BP neural network is selected as the optimal model due to its lower optimization MSE (0.015 vs 0.20), higher R² (0.98 vs 0.85), and accurate modeling of parameter sensitivity.

[0139] S4. Input the required parameters of the weld bead to be welded into the selected model to obtain the dual-wire welding process parameters, and send the process parameters to the welding machine control system to complete the welding.

[0140] In this embodiment, the required parameters such as weld reinforcement height and weld width are input into the selected model to obtain the twin-wire welding process parameters, and the process parameters are sent to the welding machine control system to complete the welding, including the following steps:

[0141] S4.1 Obtain the dimensional information of the actual welding target bevel to be filled, including bevel type, bevel height, bevel gap and bevel angle;

[0142] S4.2 Obtain the number of weld passes to be filled and set the welding speed for each weld pass;

[0143] S4.3. Arrange the weld beads in the groove using a program (CAD) and assign the weld bead height and weld width to each weld bead.

[0144] S4.4 Input the morphology parameters of multiple weld beads and welding speeds of the corresponding target bevel to be filled into the selected model in sequence, and use the selected trained model to output the predicted front wire feed speed and rear wire feed speed.

[0145] S4.5 The welding speed and the predicted wire feeding speeds output by the model are input into the TPS / i TWIN Push welding robot system to achieve automated welding.

[0146] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A method for predicting process parameters of twin-wire welding based on machine learning, characterized in that, Includes the following steps: S1. Obtain welding parameters and weld morphology parameters through welding process experiments, preprocess the data, and establish a dataset; S2. Construct a backpropagation neural network model based on the dataset, and construct a support vector machine model. Use the dataset to train the backpropagation neural network model and the support vector machine model. S3. Evaluate the trained backpropagation neural network model and support vector machine model, select the optimal model as the model for actual industrial application, and introduce an input parameter sensitivity coefficient matrix and an exponential weight allocation mechanism for optimization during the evaluation process. S4. Input the required parameters for the weld bead to be welded into the selected model to obtain the dual-wire welding process parameters, and send the process parameters to the welding machine control system to complete the welding.

2. The method for predicting process parameters of twin-wire welding based on machine learning as described in claim 1, characterized in that: In step S1, welding parameters and weld morphology parameters are obtained through welding process experiments. The data is preprocessed and a dataset is established, including the following steps: S1.1 Obtain welding condition parameters and welding process parameters, including the front wire feed speed, the rear wire feed speed and the welding speed. S1.2 Cut the weld section, take pictures of the weld morphology using a microscope, and then obtain the weld reinforcement height and weld width data. S1.3 Preprocess the obtained data on front wire feed speed, rear wire feed speed, welding speed, weld reinforcement height and weld width; S1.4 Divide the dataset into training set, validation set and test set, and save the dataset.

3. The method for predicting process parameters of twin-wire welding based on machine learning as described in claim 2, characterized in that: In step S2, the backpropagation neural network model is constructed based on the dataset, including the following steps: S2.

1. Set the number of input layer nodes according to the number of input features, determine the number of hidden layers and the number of nodes in each layer, and determine the number of output layer nodes. The input features include weld reinforcement height, weld width and welding speed. S2.2, Randomly initialize the weights and biases between the nodes of the input layer, hidden layer, and output layer; S2.

3. Propagate the input features forward through the backpropagation neural network, calculate the output value of each input layer, hidden layer and output layer node, until the predicted values ​​of the front yarn feed speed and back yarn feed speed output by the backpropagation neural network are obtained. S2.4 Starting from the output layer, calculate the gradient of each weight and bias with respect to the loss function layer by layer, and adjust the value of each weight and bias according to the gradient; S2.

5. Update the weights and biases using the gradient descent algorithm; S2.6 Repeat steps S2.3 to S2.5 above until the maximum number of iterations t1 is reached.

4. The method for predicting process parameters of twin-wire welding based on machine learning as described in claim 3, characterized in that: In S2.5, the gradient descent algorithm is as follows: ； ； in, Indicates weight; 、 Indicates the node index; Indicates bias; Indicates the learning rate; Indicates the current iteration number; Represents the loss function Weights The partial derivatives; Represents the loss function For bias The partial derivatives; This represents the loss function.

5. The method for predicting process parameters of twin-wire welding based on machine learning according to claim 4, characterized in that: In step S2, constructing the support vector machine model includes the following steps: S2.7 Set the input variables as weld reinforcement height, weld width and welding speed, and set the output variables as front wire feed speed and rear wire feed speed; S2.8 Select the radial basis function as the kernel function for the support vector machine model; S2.9 Optimize the objective function and determine the optimization parameters; S2.

10. Use the radial basis function kernel to train the support vector regression model on the training set.

6. The method for predicting process parameters of twin-wire welding based on machine learning according to claim 5, characterized in that: In S2.9, the objective function is: ； in, Represent the objective function; Represents the weight vector; Represents the regularization parameter; This represents the total number of samples in the dataset; Represents slack variables; This indicates the sample index of the dataset.

7. The method for predicting process parameters of twin-wire welding based on machine learning as described in claim 6, characterized in that: In step S3, the trained backpropagation neural network model and support vector machine model are evaluated, and the optimal model is selected as the model for practical industrial applications. This includes the following steps: S3.1 Load the pre-trained backpropagation neural network model and support vector machine model; S3.2 Input the input data from the same test set into the backpropagation neural network model and the support vector machine model respectively, and obtain the prediction results of the backpropagation neural network model and the support vector machine model for the front yarn feed speed and the back yarn feed speed of the test set. S3.3 Calculate the difference between the prediction results of the backpropagation neural network model and the support vector machine model and the true value by means of mean square error. In the mean square error calculation, the input parameter sensitivity coefficient matrix and the exponential weight allocation mechanism are introduced for optimization. The determination coefficients between the prediction results of the backpropagation neural network model and the support vector machine model and the true value are calculated. S3.4 If the mean squared error between the prediction result and the true value of the backpropagation neural network model is less than the mean squared error between the prediction result and the true value of the support vector machine model, and the coefficient of determination of the backpropagation neural network model is greater than the coefficient of determination of the support vector machine model, then the backpropagation neural network model is selected; if the mean squared error between the prediction result and the true value of the backpropagation neural network model is greater than the mean squared error between the prediction result and the true value of the support vector machine model, and the coefficient of determination of the backpropagation neural network model is less than the coefficient of determination of the support vector machine model, then the support vector machine model is selected.

8. The method for predicting process parameters of twin-wire welding based on machine learning according to claim 7, characterized in that: In S3.3, the mean square error is: ； in, Indicates mean square error; Indicates the number of samples in the test set; Indicates the sample index of the test set; Represents the actual value; Indicates the prediction result; To evaluate the strength of the influence of input parameters on output parameters, the local gradient of the input parameters with respect to the output is calculated, and the sensitivity coefficient matrix of the input parameters for each sample is also calculated. ； The input parameters are weld reinforcement height, weld width, and welding speed, while the output parameters are the front wire feed speed and the rear wire feed speed. ； in, This represents the local gradient of the input parameters with respect to the output parameters; This represents the normalization factor for process parameters; Indicates the coupling effect weight; This represents the gradient of the effect of input parameters from samples not currently in the test set on the output. ， Indicates the sample index of the test set; An exponential weighting mechanism is introduced to account for the influence of the sensitivity of weld reinforcement, weld width, and welding speed parameters on the final error calculation. ； in, Indicates the dynamic adjustment coefficient; Indicates test set samples For output parameters Sensitivity; Indicates the index of the output parameter; Indicates the sample index of the test set; Indicates weight; The optimized formula for calculating the mean squared error, after introducing an input parameter sensitivity coefficient matrix and an exponential weighting mechanism, is as follows: ； in, This represents the optimized mean squared error; Indicates the weld reinforcement; Indicates the weld width; Indicates welding speed; This indicates the priority coefficient.

9. The method for predicting process parameters of twin-wire welding based on machine learning as described in claim 8, characterized in that: In S3.3, the coefficient of determination is: ； in, This represents the mean of all true values; This represents the coefficient of determination.

10. The method for predicting process parameters of twin-wire welding based on machine learning according to claim 9, characterized in that: In step S4, the parameters required for the weld bead to be welded are input into the selected model to obtain the dual-wire welding process parameters, and the process parameters are sent to the welding machine control system to complete the welding. This includes the following steps: S4.1 Obtain the size information of the actual welding target bevel to be filled; S4.2 Obtain the number of weld passes to be filled and set the welding speed for each weld pass; S4.

3. Arrange the weld beads on the bevel and give the weld bead height and weld width for each weld bead; S4.4 Input the morphology parameters of multiple weld beads and welding speeds of the corresponding target bevel to be filled into the selected model in sequence, and use the selected trained model to output the predicted front wire feed speed and rear wire feed speed. S4.5 Input the welding speed and the predicted wire feed speeds output by the model into the TPS / i TWIN Push welding robot system.

Images