CN113657040A - Laser scanning welding penetration prediction method based on neural evolution and application - Google Patents
Laser scanning welding penetration prediction method based on neural evolution and application Download PDFInfo
- Publication number
- CN113657040A CN113657040A CN202110974695.3A CN202110974695A CN113657040A CN 113657040 A CN113657040 A CN 113657040A CN 202110974695 A CN202110974695 A CN 202110974695A CN 113657040 A CN113657040 A CN 113657040A
- Authority
- CN
- China
- Prior art keywords
- welding
- laser scanning
- neural evolution
- neural
- network model
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/10—Geometric CAD
- G06F30/17—Mechanical parametric or variational design
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/02—Reliability analysis or reliability optimisation; Failure analysis, e.g. worst case scenario performance, failure mode and effects analysis [FMEA]
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Geometry (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Computer Hardware Design (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Analysis (AREA)
- Pure & Applied Mathematics (AREA)
- Mathematical Optimization (AREA)
- Computational Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Laser Beam Processing (AREA)
Abstract
The invention discloses a laser scanning welding penetration prediction method based on neural evolution and application thereof, which solve the problem that the existing prediction method in the prior art excessively depends on an initial network structure and random weight, have the beneficial effect of ensuring the penetration prediction precision, and have the following specific schemes: a laser scanning welding penetration prediction method based on neural evolution comprises the steps of preparing a welding test plate and welding equipment; carrying out a test according to the designed test parameters, obtaining related welding penetration data through the test, and establishing a laser scanning welding penetration data set; constructing a neural evolution network model, finishing the intersection and variation of a network topological structure and a connection weight in the training process, and continuously changing the network topological structure and the connection weight to obtain the network topological structure and the connection weight with the minimum total error; and forming a final neural evolution network model based on the network topology structure with the minimum total error and the connection weight, and realizing laser scanning welding penetration prediction.
Description
Technical Field
The invention relates to the technical field of intelligent welding, in particular to a laser scanning welding penetration prediction method based on neural evolution and application.
Background
The statements in this section merely provide background information related to the present disclosure and may not necessarily constitute prior art.
The laser scanning welding realizes the swing of a focused light beam by using a scanning galvanometer, and has the advantages of high welding precision, good weld quality and formability, low requirement on a plate splicing gap, high welding flexibility and the like compared with the traditional fixed light beam laser welding. Research shows that laser scanning welding can effectively increase the flow of a molten pool, inhibit splashing and formation of air holes and promote homogenization of welding seam alloy elements, and the width of a welding seam can be changed and the brittle phase of a welding seam fusion area is reduced by adjusting the swing frequency and amplitude of a vibrating mirror, so that the quality of the welding seam is improved.
The laser scanning welding of the double twisted line shape receives great attention in academia and industry in recent years, the stability of the welding pool of the laser scanning welding of the double twisted line shape is good, and compared with a linear oscillation mode and a circular oscillation mode, the laser scanning welding of the double twisted line shape presents fewer welding defects and better mechanical performance. However, the oscillation of the laser beam can change the criterion of welding mode conversion and change the welding seam forming mechanism, thereby causing difficult control of the penetration and influencing the welding quality.
Proper penetration helps ensure weld quality, however, the relationship between laser welding process parameters and penetration is unknown, non-linear, and complex, making it impractical, even for skilled operators, to determine visually the optimal process parameters. In addition, laser beam oscillation welding increases process parameters, making the welding process more complex.
Existing mathematical models such as polynomial response surfaces and radial basis functions can be used to establish the relationship between process parameters and weld geometry, but due to coupling effects between welding process parameters and non-linear causal relationships between welding process parameters and penetration, the predictive performance of the model cannot be guaranteed even with a large number of sample points. The excellent nonlinear processing capability of the neural network has obvious advantages in the aspects of predicting weld forming and mechanical properties, however, the mathematical principle of the neural network determines the inherent defect of easy falling into local optimization, and in addition, the training cost and the training effect of the neural network are excessively dependent on the network structure, initial random weight and threshold value.
In the existing method, a Chinese patent document (application number 201911358046.X) proposes an arc welding seam forming accurate prediction method based on deep learning, however, the inventor finds that in the prediction method designed by the above application document, a deep neural network structure and training parameters realize automatic optimization and self-learning of the parameters according to the requirements of data set size, prediction accuracy and calculation efficiency, but the improvement of the network structure and the optimization of the training parameters occur after the whole training process is finished, and whether improvement and optimization are needed is evaluated according to the training results of the whole training. In the training process, the deep neural network structure is kept unchanged, the network connection form is a full connection form, the full connection form is still kept after the improvement and optimization of the network structure, and the training effect is still excessively dependent on the network structure, the initial random weight and the threshold value in essence, so that the inherent defect that the local optimization is easy to fall into cannot be overcome, and the precision of the fusion depth prediction is influenced.
Disclosure of Invention
Aiming at the defects in the prior art, the invention aims to provide a laser scanning welding penetration prediction method based on neural evolution, which obviously reduces the training cost, overcomes the inherent defect of easy falling into local optimum of the traditional deep neural network, and realizes the accurate prediction of the double twisted line-shaped laser scanning welding penetration.
In order to achieve the purpose, the invention is realized by the following technical scheme:
a laser scanning welding penetration prediction method based on neural evolution comprises the following steps:
preparing a welding test plate and welding equipment;
carrying out a test according to the designed test parameters, obtaining related welding penetration data through the test, and establishing a laser scanning welding penetration data set;
constructing a neural evolution network model, training the neural evolution network model through a data set, finishing the intersection and variation of a network topology structure and connection weights in the training process, and continuously changing the network topology and the connection weights to obtain the network topology structure and the connection weights with the minimum total error;
and forming a final neural evolution network model based on the network topology structure with the minimum total error and the connection weight, and realizing laser scanning welding penetration prediction based on the final neural evolution network model.
In the prediction method, a neural evolution network model is constructed, and a method based on neural evolution in a network training process not only changes the topological structure of the network, but also changes the connection weight and the threshold value of the network, wherein the network connection form includes but is not limited to a full connection form. In this way, the network topological structure can be gradually evolved from the minimum structure, and a new topological structure is evolved only when necessary in the whole training process, so that the training cost is obviously reduced, the defect that the traditional deep neural network is easy to fall into local optimum is overcome, and the accurate prediction of the laser scanning welding penetration is realized.
According to the laser scanning welding penetration prediction method based on neural evolution, a plurality of groups of double-twisted line-shaped laser scanning welding process parameters are designed by adopting a Latin hypercube sampling method to form the test parameters, and the sampling method can accurately describe the distribution characteristics of original sample data by using a small sample amount and designing to obtain the test parameters by carrying out uniform, random and orthogonal sampling on a parameter design space.
According to the laser scanning welding penetration prediction method based on the neural evolution, a plurality of welding joints with the same or different penetration depths are obtained according to a test, a plurality of metallographic samples are obtained through the welding joints, the welding penetration of each metallographic sample is measured to obtain the welding penetration data, and the data set comprises the welding penetration of the test parameters corresponding to each sample.
The laser scanning welding penetration prediction method based on the neural evolution comprises the following steps of testing parameters, wherein the testing parameters comprise laser power P, welding speed V, scanning frequency SF, scanning amplitude SA, paraxial shielding gas flow and defocusing amount.
According to the laser scanning welding penetration prediction method based on neural evolution, in the neural evolution network model, laser power P, welding speed V, scanning frequency SF and scanning amplitude SA are selected as input parameters of the neural evolution network model, and welding penetration is selected as an output parameter of the neural evolution network model.
The laser scanning welding penetration prediction method based on the neural evolution comprises the data sets including a training data set used for training the neural evolution network model and a testing data set used for verifying the prediction accuracy of the final neural evolution network model.
According to the laser scanning welding penetration prediction method based on neural evolution, aiming at the neural evolution network model, an initial population scale and a fitness rule are set, each individual in the population represents one neural evolution network model, an output value corresponding to each group of input data in a training data set is calculated through the neural evolution network model, the square of an error between each output value and an actual value of the training data set is summed to obtain a total error, the fitness of each individual in the population is defined as subtracting the total error from the initial fitness, and the higher the fitness is, the smaller the total error is proved to be, the better the individual is.
According to the laser scanning welding penetration prediction method based on neural evolution, the R of the neural evolution network model is selected from the test data set verification of the prediction precision of the final neural evolution network model2(determination coefficient) and RMSE (root mean square error) are used as evaluation indexes to determine the prediction accuracy.
In the neural evolution network model, in the initial network topology structure, each input layer node is connected with each node in the intermediate layer and each output layer node, each node in the intermediate layer is connected with each output layer node, and the initial connections in the neural evolution network model are all enabled.
The laser scanning welding penetration prediction method based on the neural evolution comprises the steps of cutting, polishing and etching the welding joint to obtain a plurality of metallographic samples.
According to the laser scanning welding penetration prediction method based on neural evolution, the flow and the defocusing amount of the paraxial shield gas in the test parameters are set values.
In a second aspect, the invention further provides application of the laser scanning welding penetration prediction method based on neural evolution, and the method is applied to prediction of the laser scanning welding penetration of the twisted pair line shape.
The beneficial effects of the invention are as follows:
1) according to the method, the laser scanning welding penetration prediction method based on the neural evolution is provided, so that not only is the topological structure of the network changed, but also the connection weight and the threshold value of the network are changed. In this way, the network topological structure can be gradually increased from the minimum structure, and a new topological structure is updated only when necessary in the whole training process, so that the training cost is obviously reduced, the defect that the traditional deep neural network is easy to fall into local optimum is overcome, and the accuracy of laser scanning welding penetration prediction can be effectively ensured.
2) According to the method, a neural evolution network model is constructed, the network connection form includes but is not limited to a full connection form, each input layer node in an initial network structure is connected with each node in a middle layer and also connected with an output layer node, and in the training process, whether the connection in a network topology structure is enabled or not is continuously changed along with crossing and variation.
3) In the training process of the neural evolution network model, the network topology structure and the connection weight are crossed and varied through neural evolution, the network topology and the connection weight are continuously changed, and the network topology structure and the connection weight with the minimum total error are obtained, so that the final neural evolution network model is constructed.
Drawings
The accompanying drawings, which are incorporated in and constitute a part of this specification, are included to provide a further understanding of the invention, and are incorporated in and constitute a part of this specification, illustrate exemplary embodiments of the invention and together with the description serve to explain the invention and not to limit the invention.
FIG. 1 is a diagram illustrating steps of constructing a weld penetration prediction model according to an embodiment of the present invention;
FIG. 2 is a core logic diagram of a weld penetration prediction model in an embodiment of the present invention;
FIG. 3 is a schematic diagram of a twisted pair curve according to an embodiment of the present invention;
FIG. 4 is a diagram of a laser scanning welding static scanning trajectory in the form of twisted pairs in an embodiment of the present invention;
FIG. 5 is a diagram of a dynamic continuous scanning trajectory for twisted pair laser scanning welding in an embodiment of the present invention;
FIG. 6 is a schematic representation of the geometry of weld penetration in an embodiment of the present invention;
FIG. 7 is a weld penetration profile of a sample in an embodiment of the present invention;
FIG. 8 is a schematic diagram of the topology of the initial network model in an embodiment of the invention;
FIG. 9 is a schematic diagram of genotype expression and cross-matching of network topologies in an embodiment of the present invention;
FIG. 10 is a schematic diagram of two structural mutations of neural evolution in an embodiment of the present invention;
11(a), 11(b), 11(c), 11(d), 11(e), 11(f) and 11(g) are schematic diagrams of network topology and connection weight evolution after evolution through generations 25, 50, 75, 175, 275, 525 and 800 in the embodiment of the present invention;
FIG. 12 is a diagram illustrating population best fitness evolution in an embodiment of the present invention;
FIG. 13 is a diagram illustrating a final topology and connection weights of neural evolution in an embodiment of the present invention;
FIG. 14 is a diagram illustrating a comparison between a predicted value and an actual value of weld penetration in an embodiment of the present invention.
Wherein, 1-laser beam, 2-base material, 3-welding seam, 4-heat affected zone.
Detailed Description
It is to be understood that the following detailed description is exemplary and is intended to provide further explanation of the invention as claimed. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.
It is noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of exemplary embodiments according to the invention. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the invention expressly state otherwise, and it should be understood that when the terms "comprises" and/or "comprising" are used in this specification, it indicates the presence of the stated features, steps, operations, devices, components, and/or combinations thereof.
As introduced by the background art, the problem that the existing prediction method excessively depends on an initial network structure and random weight exists in the prior art, and in order to solve the technical problem, the invention provides a laser scanning welding penetration prediction method based on neural evolution and application thereof.
Example one
In an exemplary embodiment of the present invention, referring to fig. 1 and fig. 2, a laser scanning welding penetration prediction method based on neural evolution includes the following steps:
preparing a welding test plate and welding equipment;
carrying out a test according to the designed test parameters, obtaining related welding penetration data through the test, and establishing a laser scanning welding penetration data set;
constructing a neural evolution network model, training the neural evolution network model through a data set, finishing the intersection and variation of a network topology structure and connection weights in the training process, and continuously changing the network topology and the connection weights to obtain the network topology structure and the connection weights with the minimum total error;
and forming a final neural evolution network model based on the network topology structure with the minimum total error and the connection weight, and realizing laser scanning welding penetration prediction based on the final neural evolution network model.
Proper welding penetration is helpful for ensuring welding quality, laser beams are oscillated in laser scanning welding, however, the oscillation of the laser beams can change the criterion of welding mode conversion and change the welding seam forming mechanism, and further, the penetration is difficult to control, and the welding quality is influenced. The relationship between laser scanning welding process parameters and penetration is unknown, nonlinear and complex, even though it becomes impractical for a skilled operator to determine visually the optimum process parameters.
The method comprises the steps of obtaining a plurality of welding joints with the same or different penetration depths according to a test, cutting, polishing and etching the welding joints to obtain a plurality of metallographic samples, obtaining a plurality of metallographic samples through the welding joints, and measuring the welding penetration of each metallographic sample to obtain welding penetration data.
Specifically, in this embodiment, for example, the depth of penetration of the twisted pair laser scanning welding is predicted, and in order to obtain welded joints with different penetration depths, a surface of a 304 stainless steel plate with a thickness of 3mm is cleaned with acetone, and after drying, an oxide film on the surface is scraped off with a sand paper, and the surface is cleaned with absolute ethyl alcohol.
Further, a plurality of groups of double-twisted linear laser scanning welding process parameters are designed by adopting a Latin hypercube sampling method, 50-100 groups can be set in consideration of prediction precision, 58 groups are set in the embodiment, each group of process parameter data comprises laser power P (W), welding speed V (mm/s), scanning frequency SF (Hz) and scanning amplitude SA (mm), test parameters are formed, and the range of all selected test parameters is as follows:
800≤LP≤3000(W)
8≤WS≤30(mm/s)
10≤SF≤250(Hz)
0≤SA≤2.5(mm)
furthermore, in other test parameters, the flow of paraxial shield gas (argon) is set to be a fixed value of 15(L/min), the defocusing amount is set to be a fixed value of 0(mm), and the sampling method can accurately describe the distribution characteristics of original sample data by using a small amount of samples through uniform, random and orthogonal sampling on a parameter design space.
A double-twisted line-shaped laser scanning welding test is carried out by adopting 58 designed sets of process parameters, and welding joints with 58 different penetration depths are obtained. Wherein, in the laser scanning welding of the twisted pair shape, the laser beam is collimated and then irradiated onto the rotating mirror. The deflection of the beam deflector is driven by a motor to realize a twisted-pair shaped scanning trajectory, the twisted-pair shaped curve is shown in fig. 3, the length of segment AB is set to 2a, and if the moving point "P" satisfies formula (1), the trajectory of "P" is called lemniscate. When the welding speed V is 0, the trajectory of the laser beam is as shown in fig. 4, with the point "a" as the starting point of the welding cycle, and the laser beam continuously circulates along the path a (i) -b-c (g) -d-e-f-g (c) -h-i (a) at the scanning speed Ve. Ve is mainly determined by the Scanning Frequency (SF) and the Scanning Amplitude (SA) of the laser beam, and the Scanning Frequency (SF) is defined as shown in fig. 3. When the welding speed V > 0, the scanning trajectory of the laser is determined by the welding speeds V and Ve, and FIG. 5 shows the continuous scanning trajectory of the laser beam in a twisted-pair motion, the laser beam starting from the "m" point. After several passes, the welding process ends at "n" and during the dynamic continuous sweep, the actual velocity Va is a vector combination of Ve and V along the weld, which can be calculated using equation (2).
|PA|*|PB|＝a2 (1)
Further, the welded part is cut along the set cross section by a wire cutting machine to obtain a sample with a set size, for example, a sample with a size of 10mm × 3mm, the sample is polished and etched by a metallographic polishing machine and an iron trichloride solution to obtain a plurality of metallographic samples, macroscopic metallographic phase of a fusion region is observed and photographed under a 6XB three-mesh normal metallographic microscope to measure the weld penetration of each metallographic sample, the geometric characteristic dimension of the weld penetration of the sample is measured as shown in fig. 6, and in addition, as shown in fig. 7, the weld penetration distribution diagram of 58 metallographic samples presents relatively uniform random distribution.
Aiming at a data set, the data set comprises process parameters and welding penetration corresponding to each sample, the established data set is divided into a training data set and a testing data set according to a set proportion, for example, the ratio can be 50:8, the number of the training data sets is greater than that of the testing data sets, the training data sets are used for training a neural evolution network model, the testing data sets are used for verifying the prediction precision of the final neural evolution network model, and partial data of the data set are shown as a table:
further, in the neural evolution network model construction, the laser power p (w), the welding speed V (mm/s), the scanning frequency sf (hz) and the scanning amplitude sa (mm) are selected as input parameters of the neural evolution network model, and the welding penetration is selected as an output parameter of the neural evolution network model, as shown in fig. 8, in the network topology initial structure, the input layer has 4 nodes, the intermediate layer has 8 nodes, the output layer has 1 node, each input layer node is connected with each node in the intermediate layer and the output layer node, each node in the intermediate layer is connected with the output layer node, the initial connection in the neural evolution network model is enabled, and the number in the intermediate layer represents an innovation code.
As shown in fig. 9, each genome corresponds to a neural evolution network, and each genome includes a list of connected genes, each of which refers to two node genes connected in the neural evolution network. Each linking gene specifies an input node, an output node, a weight of the link, whether the linking gene is expressed (enabled), and an innovation code that facilitates identification of the corresponding gene in a later crossover process, the innovation code representing a schedule for each gene, the historical provenance of each gene can be monitored.
As shown in fig. 10, the mutation may change not only the connection weight but also the topology of the network. Structural mutations in the genome occur in two ways, linkage mutations and node mutations, by which genomes of different sizes are created, sometimes with completely different linkages at the same position.
As shown in fig. 9, when crossed, genes with the same innovative code in both genomes were aligned together. Unmatched genes are inherited from more suitable parents or, if they are equally suitable, randomly from both parents. Furthermore, unmatched genes are either disjoint or redundant depending on whether they appear within or outside the scope of the innovative coding of another gene.
Further, the population is divided into species according to historical labels and topological similarities. Compatible distances for different structures: (
) Can be calculated using equation (3), where E is the number of redundant genes, D is the number of disjoint genes, W is the difference in the average weights of the matching genes, and N is the number of genes in the larger genome; a1, a2, and a3 are coefficients that adjust the importance of these three factors.
Illustratively, the neural evolution approach allows individuals to compete primarily within their species, rather than the entire population, and thus, the individual has time to optimize their structure before having to compete with other individuals in the population, and it grows new structures only when necessary, in such a way that neural evolution searches within a minimum number of weight dimensions, thereby greatly reducing the number of genetic generations needed to find a solution.
And setting an initial population scale and a fitness rule aiming at the neural evolution network model.
In this embodiment, the initial population scale is set to 300, the maximum evolution algebra is 2000, each individual in the population represents a neural evolution network model, the output value corresponding to each set of input data in the training data set is calculated through the neural evolution network model, and the squares of the errors between each output value and the actual value of the training data set are summed to obtain a total error;
the fitness of each individual in the population is defined as the initial fitness minus the total error, the initial fitness is set to be 50, namely the higher the fitness is, the smaller the total error is proved to be, the better the individual is; in some examples, the fitness threshold is set to 49.8, i.e., the condition for the end of the entire training is that the maximum evolution algebra is reached or the fitness reaches the threshold.
Training a neural evolution network model by using a training data set, and completing intersection and variation of a network topology structure and connection weights through neural evolution in a training process, wherein the network topology structure and the connection weights are evolved as shown in fig. 11(a) -11 (g), the network topology structure and the connection weights of the neural evolution network model are correspondingly evolved through 25, 50, 75, 175, 275, 525 and 800 generations, a solid line represents connection enabling, a dotted line represents connection disabling, and the network topology and the connection weights are continuously changed through multi-generation evolution, so that the network topology structure and the connection weights with the minimum total error are obtained. In addition, the list of activation functions of the network topology is an optional activation function in the evolved neural network structure, and the activation functions also have variation in the neural evolution process, and the variation range is in the list. In a multilayer network, a functional relationship exists between the output of an upper node and the input of a lower node, the function is called an activation function (also called an excitation function), the activation function can introduce a nonlinear factor, and the nonlinearity enables the network to learn a complex transformation relationship between the input and the output.
The list of activation functions for the network includes: the method comprises the following steps of a Softplus function, a Relu function and a Sigmoid function, wherein the default of an activation function is the Softplus function, the mutation probability of the activation function is 0.1, the sum function is sum, the probability of adding and deleting the connection in the network mutation is 0.2, the probability of adding and deleting the node is 0.2, the default connection is enabled, and the probability of whether the connection is enabled to be changed is 0.05.
It is understood that as population iteration, the best fitness in the population is continuously increased, as shown in fig. 12, and the continuous evolution of the population best fitness also means that the prediction error is continuously reduced, and after the population is subjected to 2000 generations of evolution, the final network structure, the connection weights and the threshold values of the neural evolution network model are shown in fig. 13, so that the final neural evolution network model is formed.
To verify the predictive performance of the model, 8 samples in the test dataset were utilizedThe data verifies the prediction accuracy of the final neural evolution network model. As shown in FIG. 14, the comparative analysis of the predicted value and the actual value of the laser scanning welding penetration of the twisted pair line shape shows good consistency between the predicted value and the experimental value, and in order to more intuitively evaluate the performance of the model, the R of the model is selected2And RMSE as an evaluation index, R of the model2The value is about 0.966, the model is proved to have good fitting performance, the RMSE of the welding penetration is about 0.079mm, the result shows that the model has higher prediction precision, and the established neural evolution network model is proved to be reliable, so that the accurate prediction of the double-twisted line-shaped laser scanning welding penetration can be realized by using the neural evolution method.
Aiming at the problems that the BP neural network and the deep neural network are inherently dependent on the network structure and the initial random weight too much and cannot overcome the inherent defect of easy falling into local optimization, the embodiment takes the twisted-pair laser scanning welding as an application object, and not only changes the topological structure of the network, but also changes the connection weight and the threshold value of the network based on a neural evolution method in the network training process, wherein the network connection form comprises but is not limited to a full connection form. In this way, the network topological structure can be gradually evolved from the minimum structure, and a new topological structure is evolved only when necessary in the whole training process, so that the training cost is obviously reduced, the defect that the traditional deep neural network is easy to fall into local optimum is overcome, and the accurate prediction of the laser scanning welding penetration is realized.
The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention, and various modifications and changes may be made by those skilled in the art. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims (10)
1. A laser scanning welding penetration prediction method based on neural evolution is characterized by comprising the following steps:
preparing a welding test plate and welding equipment;
designing test parameters, carrying out a test according to the designed test parameters, obtaining relevant welding penetration data through the test, and establishing a laser scanning welding penetration data set;
constructing a neural evolution network model, training the neural evolution network model through a data set, finishing the intersection and variation of a network topology structure and connection weights in the training process, and continuously changing the network topology structure and the connection weights to obtain the network topology structure and the connection weights with the minimum total error;
and forming a final neural evolution network model based on the network topology structure with the minimum total error and the connection weight, and realizing laser scanning welding penetration prediction based on the final neural evolution network model.
2. The laser scanning welding penetration prediction method based on neural evolution as claimed in claim 1, wherein a plurality of sets of twisted pair laser scanning welding process parameters are designed by adopting a Latin hypercube sampling method to form the test parameters.
3. The laser scanning welding penetration depth prediction method based on neural evolution as claimed in claim 1, wherein a plurality of welding joints with the same or different penetration depths are obtained according to experiments, a plurality of metallographic samples are obtained through the welding joints, and the welding penetration depth of each metallographic sample is measured to obtain the welding penetration depth data.
4. The laser scanning welding penetration prediction method based on the neural evolution is characterized in that the experimental parameters comprise laser power P, welding speed V, scanning frequency SF, scanning amplitude SA, paraxial shield gas flow and defocus.
5. The laser scanning welding penetration prediction method based on neural evolution as claimed in claim 1, wherein in the neural evolution network model construction, laser power P, welding speed V, scanning frequency SF and scanning amplitude SA are selected as input parameters of the neural evolution network model, and welding penetration is selected as an output parameter of the neural evolution network model.
6. The laser scanning weld penetration prediction method based on neural evolution of claim 1, wherein the data sets comprise a training data set for training a neural evolution network model and a testing data set for verifying the prediction accuracy of a final neural evolution network model.
7. The laser scanning welding penetration prediction method based on neural evolution of claim 1, characterized in that, aiming at the neural evolution network model, an initial population scale and a fitness rule are set, each individual in the population represents one neural evolution network model, an output value corresponding to each group of input data in the training data set is calculated through the neural evolution network model, a total error is obtained by summing the squares of the error between each output value and the actual value of the training data set, and the fitness of each individual in the population is defined as the initial fitness minus the total error.
8. The laser scanning welding penetration prediction method based on neural evolution of claim 6, wherein in the verification of the prediction accuracy of the final neural evolution network model through the test data set, R of the neural evolution network model is selected2And RMSE as an evaluation index to judge the prediction accuracy.
9. The laser scanning welding penetration prediction method based on neural evolution as claimed in claim 3, characterized in that the welding joints are cut, polished and etched to obtain a plurality of metallographic samples.
10. The application of the laser scanning welding penetration prediction method based on the neural evolution is characterized by being applied to the prediction of the laser scanning welding penetration of the double twisted line shape.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110974695.3A CN113657040A (en) | 2021-08-24 | 2021-08-24 | Laser scanning welding penetration prediction method based on neural evolution and application |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110974695.3A CN113657040A (en) | 2021-08-24 | 2021-08-24 | Laser scanning welding penetration prediction method based on neural evolution and application |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN113657040A true CN113657040A (en) | 2021-11-16 |
Family
ID=78481780
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110974695.3A Pending CN113657040A (en) | 2021-08-24 | 2021-08-24 | Laser scanning welding penetration prediction method based on neural evolution and application |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113657040A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113828947A (en) * | 2021-11-23 | 2021-12-24 | 昆山宝锦激光拼焊有限公司 | BP neural network laser welding seam forming prediction method based on double optimization |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1529281A (en) * | 2003-10-21 | 2004-09-15 | 上海交通大学 | Neural network modelling method |
| CN104866904A (en) * | 2015-06-16 | 2015-08-26 | 中电科软件信息服务有限公司 | Parallelization method of BP neural network optimized by genetic algorithm based on spark |
| CN106650933A (en) * | 2017-01-12 | 2017-05-10 | 西安电子科技大学 | Deep neural network optimizing method based on coevolution and back propagation |
| CN110640316A (en) * | 2019-09-03 | 2020-01-03 | 温州大学 | Double-beam laser welding optimization method |
| CN110852435A (en) * | 2019-10-12 | 2020-02-28 | 沈阳航空航天大学 | Neural evolution calculation model |
| CN111177976A (en) * | 2019-12-25 | 2020-05-19 | 广东省焊接技术研究所（广东省中乌研究院） | Arc welding seam forming accurate prediction method based on deep learning |
| CN113094980A (en) * | 2021-03-26 | 2021-07-09 | 西安交通大学 | Solder paste printing quality prediction method and system based on IGA-DNN |
-
2021
- 2021-08-24 CN CN202110974695.3A patent/CN113657040A/en active Pending
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1529281A (en) * | 2003-10-21 | 2004-09-15 | 上海交通大学 | Neural network modelling method |
| CN104866904A (en) * | 2015-06-16 | 2015-08-26 | 中电科软件信息服务有限公司 | Parallelization method of BP neural network optimized by genetic algorithm based on spark |
| CN106650933A (en) * | 2017-01-12 | 2017-05-10 | 西安电子科技大学 | Deep neural network optimizing method based on coevolution and back propagation |
| CN110640316A (en) * | 2019-09-03 | 2020-01-03 | 温州大学 | Double-beam laser welding optimization method |
| CN110852435A (en) * | 2019-10-12 | 2020-02-28 | 沈阳航空航天大学 | Neural evolution calculation model |
| CN111177976A (en) * | 2019-12-25 | 2020-05-19 | 广东省焊接技术研究所（广东省中乌研究院） | Arc welding seam forming accurate prediction method based on deep learning |
| CN113094980A (en) * | 2021-03-26 | 2021-07-09 | 西安交通大学 | Solder paste printing quality prediction method and system based on IGA-DNN |
Non-Patent Citations (1)
| Title |
|---|
| 赵黎: ""面向服装流行色预测的层次协同演化模型与算法研究"", 《万方博士学位论文库》 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113828947A (en) * | 2021-11-23 | 2021-12-24 | 昆山宝锦激光拼焊有限公司 | BP neural network laser welding seam forming prediction method based on double optimization |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110116254B (en) | Method for predicting and controlling defects of annular welding joint of bimetal composite pipe for oil and gas gathering and transportation | |
| Sathiya et al. | Optimization of laser welding process parameters for super austenitic stainless steel using artificial neural networks and genetic algorithm | |
| Lostado et al. | Combining soft computing techniques and the finite element method to design and optimize complex welded products | |
| CN113569352A (en) | Additive manufacturing size prediction and process optimization method and system based on machine learning | |
| CN113657040A (en) | Laser scanning welding penetration prediction method based on neural evolution and application | |
| Chang et al. | Penetration quality prediction of asymmetrical fillet root welding based on optimized BP neural network | |
| Pal et al. | Optimization of quality characteristics parameters in a pulsed metal inert gas welding process using grey-based Taguchi method | |
| JP4795282B2 (en) | Machining condition search device | |
| Ai et al. | Process modeling and parameter optimization using radial basis function neural network and genetic algorithm for laser welding of dissimilar materials | |
| US20230101500A1 (en) | Additive manufacturing path generation apparatus, additive manufacturing path generation method, and machine learning apparatus | |
| CN116522806B (en) | Polishing process parameter optimization method, polishing system, electronic device and storage medium | |
| Hosseinpour et al. | Finite element modelling of castellated steel beams under lateral-distortional buckling mode | |
| CN111625997B (en) | Crystal cutting process optimization method and device | |
| Gowtham et al. | Intelligent modeling combining adaptive neuro fuzzy inference system and genetic algorithm for optimizing welding process parameters | |
| CN108763801A (en) | A kind of laser gain material remanufactures cladding layer geometric properties and dilution rate modeling method | |
| Azadi Moghaddam et al. | Simultaneous optimization of joint edge geometry and process parameters in gas metal arc welding using integrated ANN-PSO approach | |
| Vijayan et al. | Comparison of response surface methodology and genetic algorithm in parameter optimization of laser welding process | |
| Das et al. | Input–output modeling and multi-objective optimization of weld attributes in EBW | |
| CN101751487A (en) | Bending displacement by utilizing artificial neural network | |
| CN113774376A (en) | Laser cladding self-adaptive scanning path planning method based on transient temperature field feedback | |
| Nele et al. | Modeling of multiple characteristics of an arc weld joint | |
| CN110340354A (en) | Scraper for powder bed increasing material manufacturing judges automatically and selection method | |
| Saleem et al. | Optimization of process parameters on weld bead characteristics of mild steel in submerged arc welding process | |
| Lahoti et al. | Recurrent neural networks to model input-output relationships of metal inert gas (MIG) welding process | |
| CN113312825B (en) | Laser shot peening effect monitoring method and device |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |