Welding forming prediction and regulation method based on digital twinning technology
By combining genetic algorithms with backpropagation neural network models, a high-precision model for predicting and controlling welding forming dimensions is constructed, which solves the problems of high cost and low efficiency in existing technologies. It realizes accurate prediction of welding forming dimensions and optimized control of process parameters, thereby improving the intelligence and automation level of welding manufacturing technology.
Patent Information
- Application Number
- CN202410941724.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-15
- Publication Date
- 2026-01-20
AI Technical Summary
Existing technologies suffer from high costs and low efficiency in predicting and controlling welding forming dimensions, making it difficult to achieve accurate prediction and control.
By combining genetic algorithms and backpropagation neural network models, a high-precision model for predicting welding forming dimensions and controlling process parameters is constructed. Through data normalization, neural network training, and genetic algorithm optimization, a relationship model between welding forming dimensions and process parameters is established.
It enables accurate prediction of weld formation dimensions and directional control of process parameters, thereby improving the intelligence and automation level of welding manufacturing technology.
Smart Images
Figure CN121360907A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of welding, and particularly relates to a welding forming prediction and regulation method based on digital twinning technology BACKGROUND
[0002] Welding, as a key manufacturing and processing technology, plays an indispensable role in many industries. In the welding process, the accuracy of the forming size is crucial for the subsequent use and maintenance of the welded structure. The welding forming size and other features are closely related to the welding process parameters. Due to the uncertainty of the welding process, there are many difficulties in predicting and regulating the forming size. Using the traditional experimental trial-and-error method is costly and inefficient, so accurate prediction of the welding forming size is of great significance for optimizing the welding process parameters and improving work efficiency.
[0003] With the development of machine learning and artificial intelligence technologies, it is possible to use these technologies to predict and control the welding forming size. By collecting a large amount of experimental data and combining machine learning algorithms, a relationship model between the forming size and the process parameters can be established. This model can be used to predict the welding forming size and adjust the process parameters according to the prediction results, thereby achieving accurate control of the forming size. Currently, there are relatively few studies on the forming size prediction model for welded components. Developing a machine learning model suitable for the welding process that can accurately predict the welding forming size and inversely predict the process parameters to achieve directional regulation of the welding process is of great significance for improving the intelligence and automation level of welding manufacturing technology. SUMMARY
[0004] In view of the deficiencies of the prior art, the present application provides a welding forming prediction and regulation method based on digital twinning technology, which combines genetic algorithm (GA) with neural network model to further optimize model parameters, so as to obtain a high-precision forming size prediction and regulation model, optimize prediction accuracy and directional regulation efficiency of process parameters, and provide a new idea for efficient welding, directional regulation of welding forming size, and optimization of process parameters.
[0005] The technical solution for achieving the object of the present application is as follows:
[0006] A welding forming prediction and regulation method based on digital twinning technology, comprising the following steps:
[0007] Step 1: Perform welding experiments with different process parameters and collect corresponding process parameters and corresponding forming sizes.
[0008] The process parameters include wire feed speed and welding speed, and the forming sizes include weld width and weld excess height. The wire feed speed ranges from 2.5 to 4 m / min, and the welding speed ranges from 3 to 4.5 mm / s.
[0009] Step two, the collected welding process parameters, forming size data preprocessing.
[0010] In order to eliminate the influence of dimension and dimension unit, the data set data is normalized, and the specific operation expression is shown in formula (1). Through normalization processing, the characteristic value is scaled to the interval [0, 1].
[0011]
[0012] For normalized data, x (i) For original data, x max For the maximum value in this column feature, x min For the minimum value in this column feature
[0013] The data set processed according to the above steps is divided into training set and test set according to the set proportion, and the proportion is 8:2.
[0014] Step three, constructing a back propagation neural network (BPNN) machine learning prediction model to obtain a machine learning forward prediction welding forming size and reverse prediction welding process parameter model;
[0015] The machine learning model is a back propagation neural network model, and the neural network model is used for process parameter to welding forming size forward prediction and forming size to process parameter reverse prediction.
[0016] Specifically divided into four models, welding speed and wire feeding speed as input, weld width as output; welding speed and wire feeding speed as input, weld excess height as output; weld excess height and weld width as input, welding speed as output; weld excess height and weld width as input, wire feeding speed as output.
[0017] The number of neural network layers adopted is three, including input layer, hidden layer and output layer. The number of input layer and output layer neurons is determined by the number of features.
[0018] The number of hidden layer neurons is determined by formula (2), wherein: q is the number of hidden layer neurons; m and n are the number of nodes of input layer and output layer respectively; a is a constant, and the value is 1-10.
[0019]
[0020] The activation function selects ReLU function, which can make the network training faster, prevent gradient disappearance in the training process, and at the same time, it is a nonlinear function itself, which can increase the nonlinear mapping of the network when added to the neural network.
[0021] The Levenberg-Marquardt algorithm is selected as the optimization algorithm, which can make the training converge to the optimal solution faster.
[0022] The selection of the number of training rounds is also particularly important. If the number of training rounds is too small, underfitting may occur, and if the number of training rounds is too large, overfitting may occur, resulting in a decrease in the generalization ability of the model.
[0023] The forward prediction training round selection is 800, and the reverse prediction training round selection is 600.
[0024] By comparing the fitting degrees of models with different numbers of neurons, the model is evaluated, and the R 2 index is used as the model evaluation index, and its expression is formula (3).
[0025]
[0026] where y i is the actual value of the weld size, is the model prediction value, is the mean value of the weld size, and w is the sample size.
[0027] By comparing different neural network structures, the neural network structure of the weld width prediction model is finally determined as 2-9-1; the neural network structure of the weld reinforcement prediction model is 2-10-1; the neural network structure of the welding speed prediction model is 2-11-1; and the neural network structure of the wire feed speed prediction model is 2-8-1.
[0028] Step four, the forward forming size prediction model and the reverse process parameter prediction model based on BP neural network are optimized by genetic algorithm respectively to obtain the GA-BP machine learning prediction model, and the test set is used for welding forming size and process parameter prediction.
[0029] The weights and biases are encoded according to the neural network structure.
[0030] The root mean square error of the prediction value and the actual value is used as the fitness function to evaluate the training, and its expression is formula (4).
[0031]
[0032] where y i is the actual value, is the model prediction value, and w is the sample size.
[0033] The fitness of each individual is calculated according to the fitness function, and a number of individuals are selected as parent individuals according to the fitness; arithmetic crossover is used, so that each gene of each child individual is composed of the arithmetic average of the corresponding genes of the two parent individuals;
[0034] The variation operation adopts non-uniform variation, and the variation probability changes with the iteration, enhances the search precision, and meets the condition until the condition is met.
[0035] RMSE and R 2 The optimal model is obtained by comparing the precision of the BP neural network model and the GA-BP neural network model according to the evaluation index.
[0036] The final model is used for forward prediction of the welding size and reverse prediction of the process parameters of the test set.
[0037] Compared with the prior art, the present application has the following advantages:
[0038] The present application constructs a welding forming size forward prediction and process parameter reverse prediction model based on a back propagation neural network, and optimizes the weight and bias of the model by using a GA genetic algorithm, thereby improving the model precision. BRIEF DESCRIPTION OF DRAWINGS
[0039] Figure 1 It is the overall flowchart of the present application.
[0040] Figure 2 It is the GA-BP prediction model flowchart in the present application.
[0041] Figure 3 It is the neural network structure topological graph in the present application example (a) forward prediction of welding forming size (b) reverse prediction of welding parameters.
[0042] Figure 4 It is the GA-BP neural network fitness change line graph in the present application example (a) weld width fitness line (b) weld excess height fitness line (c) welding speed fitness line (d) wire feeding speed fitness line.
[0043] Figure 5 It is the GA-BP weld width prediction model real value and predicted value line graph in the present application example.
[0044] Figure 6 It is the GA-BP weld excess height prediction model real value and predicted value line graph in the present application example.
[0045] Figure 7Fig. 1 is a schematic diagram of the regression of the GA-BP inverse prediction process parameter model in the examples of the present application (a) wire feed speed regression schematic diagram (b) welding speed regression schematic diagram. DETAILED DESCRIPTION
[0046] The specific embodiments are described in further detail below in conjunction with the accompanying drawings and examples, which are used to illustrate the present application but not to limit the scope of the present application.
[0047] Step one, welding experiments with different process parameters are carried out, and the corresponding process parameters and corresponding forming dimensions are collected. The process parameters include wire feed speed and welding speed, and the forming dimensions include weld width and weld reinforcement.
[0048] Considering that the forming effect and forming dimension of the welded component are mainly related to heat input, different process parameter combinations are formed by changing the wire feed speed and welding speed, and finally the welding experiment process parameters are determined within the wire feed speed value range of 2.5-4 m / min; welding speed of 3-4.5 mm / s; under the premise of ensuring quality, within the wire feed speed value range of 2.5-4 m / min; welding speed of 3-4.5 mm / s, take small intervals to obtain more process parameter combinations. Finally, 126 groups of weld width and weld reinforcement are obtained respectively. Part of the data is shown in Table 1.
[0049] Table 1 shows part of the original data
[0050] Serial number Welding speed Wire feed speed Weld width Weld reinforcement 1 3 2.5 1.25 0.31 2 3 2.8 1.33 0.33 3 3 2.9 1.34 0.34 … … … … … 126 4.5 4 1.54 0.29
[0051] Step two, the collected different welding process parameters and forming dimension data are preprocessed, and the data set is divided into a training set and a test set;
[0052] In order to eliminate the influence of dimension and dimension unit, the data set data is normalized, and the specific operation expression is shown in formula (1), wherein is the normalized data, x (i) is the original data, x max is the maximum value in the column feature, x min is the minimum value in the column feature. Through normalization processing, the feature value is scaled to the interval [0, 1].
[0053]
[0054] The data set processed in the above step is divided into a training set and a test set according to the set proportion, and the proportion is 8:2.
[0055] Step three, a back propagation neural network (BPNN) machine learning prediction model is constructed, and a machine learning forward prediction welding forming dimension inverse prediction welding process parameter model is obtained;
[0056] Step three is specifically: the machine learning model is a back propagation neural network model, and the neural network model is used for forward prediction of process parameters to welding forming size and reverse prediction of forming size to forming process parameters. Specifically, it is divided into four models, welding speed and wire feeding speed as input, and weld width as output; welding speed and wire feeding speed as input, and weld reinforcement as output; weld width and weld reinforcement as input, and welding speed as output; weld width and weld reinforcement as input, and wire feeding speed as output.
[0057] The number of neural network layers adopted is three, including input layer, hidden layer and output layer. The number of neurons in the input layer and the output layer is determined by the number of features, and the number of neurons in the hidden layer is determined by formula (2), wherein: q is the number of neurons in the hidden layer; m and n are the numbers of nodes in the input layer and the output layer, respectively; a is a constant, and the value is 1-10, the model hidden layer parameters are as shown in Table 2, the activation function is selected as ReLU function, which can make the network training faster, prevent the gradient from disappearing in the training process, and at the same time, it is a nonlinear function, which can increase the nonlinear mapping of the network. The optimization algorithm is selected as Levenberg-Marquardt algorithm, which can make the training converge to the optimal solution faster. The selection of training rounds is also particularly important. If the training rounds are too few, underfitting may occur, and if the training is too much, overfitting may occur, resulting in a decrease in the generalization ability of the model. Finally, the forward prediction training rounds are selected as 800, and the reverse prediction training rounds are selected as 600.
[0058]
[0059] By comparing the fitting degrees of different numbers of neurons, the model is evaluated, and the model evaluation index adopts R 2 index; its expression is formula (3). Wherein y i is the actual value of the weld size, is the model prediction value, is the mean value of the weld size, and w is the sample number.
[0060]
[0061] Table 2 Number of neurons in hidden layer of different models
[0062]
[0063] Finally, the neural network structure of the weld width prediction model is determined as 2-9-1, the neural network structure of the weld reinforcement prediction model is 2-10-1, the neural network structure of the welding speed prediction model is 2-11-1, and the neural network structure of the wire feeding speed prediction model is 2-8-1.
[0064] Step 4: Optimize the forward forming size prediction model and the reverse process parameter prediction model based on BP neural network using a genetic algorithm to obtain the GA-BP machine learning prediction model. Use the test set to predict the welding forming size and process parameters.
[0065] The optimal initial weights and thresholds of the BP network are obtained through the global optimization capability of the genetic algorithm. The optimal initial weights and thresholds obtained by the optimization algorithm are then used as the initial weights and thresholds of the BP neural network, and then training is performed to avoid getting trapped in local minima.
[0066] First, the weights and biases are encoded according to the neural network structure.
[0067] Using the root mean square error (RMSE) as the fitness evaluation index, its expression is given by formula (4), where y i This represents the actual weld size. is the model's predicted value, and w is the number of samples. Figure Four This is a line graph showing the relationship between the number of iterations and the fitness value. As you can see, the fitness value remains constant as the number of iterations decreases.
[0068]
[0069] The selection operation of the genetic algorithm adopts the roulette wheel selection strategy. The roulette wheel selection algorithm can select based on the fitness value of individuals, avoiding the limitation of selecting solely according to fitness value sorting, and has a certain degree of randomness, which improves the global search capability of the algorithm.
[0070] Crossover uses arithmetic crossover, where each gene of each offspring is the arithmetic mean of the corresponding genes of the two parent individuals.
[0071] The mutation operation employs non-uniform mutation, and the mutation probability changes as the iteration progresses.
[0072] Individual models are selected through selection, crossover, and mutation operations until the conditions are met, resulting in the optimal model. The final model is then used to perform forward prediction of weld formation dimensions and inverse prediction of process parameters on a test set. The optimized GA-BP neural network model is compared with the original BP neural network model, using the evaluation metric R0. 2 Based on the number of hidden layers in each model, the RMSE value of the original BP neural network weld width prediction model is 0.0431, and the RMSE value of the GA-BP neural network is 0.0286. 2 The value increased from 0.94937 to 0.96384. The RMSE value of the original BP neural network weld reinforcement prediction model was 0.0931, and the RMSE value of the GA-BP neural network was 0.067. 2The value increases from 0.96453 to 0.983; compared with the model before and after optimization, the GA-BP prediction model optimized by the genetic algorithm has higher prediction accuracy. The R2 value of the GA-BP neural network wire feeding speed prediction model is 0.94621; the R2 value of the GA-BP neural network welding speed prediction model is 0.96352, the fitting effect is very good, and the prediction accuracy is high.
[0073] Therefore, the method for predicting the welding forming size based on machine learning and directionally regulating the welding process parameters is adopted, the genetic algorithm is used to optimize the weight and bias to construct the GA-BP high-precision prediction model, and the prediction accuracy is effectively increased. By acquiring different data in the welding process and adjusting the parameters, the nonlinear problem of the welding forming size prediction can be effectively solved, the accurate control of the welding size and the optimization selection of the process parameters can be realized, the production demand can be met, and a kind of idea is provided for efficient and intelligent welding.
[0074] The basic principles, main features and advantages of the present application are shown and described above. Those skilled in the art should understand that the scope of the application involved in the examples of the present disclosure is not limited to the technical solutions formed by the specific combination of the above technical features. Those skilled in the art can realize various equivalent modifications and replacements within the technical scope of the present application, and these modifications and replacements should be covered within the protection scope of the present application.
Claims
1. A welding forming prediction and regulation method based on digital twin technology, characterized in that, Comprise: Step one, carry out welding experiment of different process parameters, and collect corresponding process parameters and corresponding forming size; The process parameters include wire feeding speed, welding speed, and the forming size includes weld width and weld reinforcement; Step two, preprocess the collected different welding process parameters and forming size data, and divide the data set into training set and test set; Step three, construct a back propagation neural network machine learning prediction model to obtain a machine learning forward prediction welding forming size reverse prediction welding process parameter model; Step four, optimize the forward forming size prediction model and reverse process parameter prediction model based on BP neural network respectively by genetic algorithm to obtain GA-BP machine learning prediction model, and use the test set to predict the welding forming size and process parameter; Step five, model evaluation is carried out to obtain the optimal model, which is used for forming size and process parameter prediction.
2. The welding formation prediction and regulation method based on digital twin technology according to claim 1, characterized in that, The wire feeding speed is in the range of 2.5-4 m / min, and the welding speed is 3-4.5 mm / s.
3. The welding formation prediction and regulation method based on digital twin technology according to claim 1, characterized in that, Step one is: welding method is one welding, that is, automatically changing wire feeding speed, welding current and welding voltage; different wire feeding speed and welding speed are used to obtain different process parameter combinations; the forming size of the welded component, including weld width and weld reinforcement, is measured.
4. The welding forming prediction and control method based on digital twin technology according to claim 1, characterized in that, The step two is specifically: different features in the data set have different dimensions and dimension units, in order to eliminate the influence of the dimensions and the dimension units, the data set data is normalized, and a specific operation expression is shown as formula (1) is the normalized data, x (i) is the original data, x max is the maximum value in the column feature, x min is the minimum value in the column feature; the feature value is scaled into the [0, 1] interval through the normalization processing; The data set processed in the above step is divided into training set and test set according to the set proportion.
5. The welding forming prediction and control method based on digital twin technology according to claim 1, characterized in that, Step three is: the machine learning model is a back propagation neural network model, and the neural network model is used for forward prediction of process parameters to welding forming size and reverse prediction of forming size to forming process parameters; Specifically, it is divided into four models, welding speed and wire feeding speed are input, and weld width is output; welding speed and wire feeding speed are input, and weld reinforcement is output; weld reinforcement and weld width are input, and welding speed is output; weld reinforcement and weld width are input, and wire feeding speed is output; the number of neural network layers adopted is three, including input layer, hidden layer and output layer; the number of neurons in the input layer and the output layer is determined by the number of features, and the number of neurons in the hidden layer is determined by formula (2) (3), wherein: q is the number of neurons in the hidden layer; m and n are the number of nodes in the input layer and the output layer respectively; a is a constant, and the value is 1-10; the activation function selects ReLU function, which can make the network training faster and prevent gradient disappearance in the training process; at the same time, it is a nonlinear function itself, which can increase the nonlinear mapping of the network when added to the neural network; the optimization algorithm selects Levenberg-Marquardt algorithm, which can make the training converge to the optimal solution faster; the number of training rounds for forward prediction is selected as 800, and the number of training rounds for reverse prediction is selected as 600; The model is evaluated by comparing the fitting degrees of different neuron numbers, and the evaluation index R is adopted 2 index; its expression is formula (3), wherein y i is the actual value of the weld size, is the predicted value of the model, is the mean value of the weld size, and w is the sample number; 6. The welding forming prediction and control method based on digital twin technology according to claim 1, characterized in that, The step four is specifically: obtaining the initial weight and threshold of the optimal BP network through the global optimization ability of the genetic algorithm, taking the optimal initial weight and threshold obtained by the optimization algorithm as the initial weight and threshold of the BP neural network, and then training to avoid falling into local minimum; the GA genetic algorithm includes selection, crossover, mutation and other operations; firstly, the weight and bias are coded according to the neural network structure; The root mean square of the error between the predicted value and the actual value is used as the fitness function to evaluate the training, and the expression is formula (4), wherein y i is the actual value, is the model predicted value, and w is the sample number; then the fitness of the individual is calculated according to the fitness function, and a plurality of individuals are selected as parent individuals according to the fitness; arithmetic crossover is used, so that each gene of each offspring individual is composed of the arithmetic average of the corresponding genes of the two parent individuals; the mutation operation adopts non-uniform mutation, and the mutation probability changes with the iteration, so as to enhance the search accuracy until the condition is met; the final GA-BP model is used to perform forward prediction of the weld size and reverse prediction of the process parameters on the test set;