Pipe bending springback angle multi-stage prediction method and prediction system based on machine learning

Abstract

The invention discloses a pipe bending springback angle multi-stage prediction method and system based on machine learning. The prediction method comprises the steps that a three-dimensional elastic-plastic finite element numerical simulation model of a circular pipe is constructed; through CAE forming simulation, sample data under five kinds of bending forming process parameters of the round pipe are obtained; performing data enhancement on the sample data; the method comprises the following steps: enhancing sample data of a circular pipe, constructing a multi-stage prediction model, training and testing the multi-stage prediction model by using the enhanced sample data to obtain a trained multi-stage prediction model, and then predicting the bending rebound angle of the circular pipe by using the trained multi-stage prediction model. The prediction method and the prediction system are high in prediction precision and high in prediction universality, and can be applied to an actual bent pipe production system.

Description

Technical Field

[0001] The present invention relates to the field of pipe bending springback angle prediction, and in particular to a pipe bending springback angle multi-stage prediction method and prediction system based on machine learning. Background Art

[0002] With the continuous development of aerospace technology, the performance and safety requirements for skid-type landing gear are increasing. Aluminum alloys are widely used in the design of such landing gears due to their light weight and excellent mechanical properties. However, since the landing gear is subjected to large bending stress during use, it may cause problems such as structural rupture, cross-sectional flattening, inner wrinkling and excessive rebound of the landing gear. Due to the small elastic modulus of aluminum alloy, its rebound is more significant, which seriously affects the performance of the pipe, and then affects its structural performance and safety. Therefore, it is of great significance to conduct a systematic study on the bending properties of the pipe.

[0003] At present, machine learning technology has shown significant potential in metal bending quality prediction, but its application also faces challenges. First, machine learning models require a large amount of high-quality data for effective training, but in the actual production process, the acquisition and quality of data are uncertain, which poses a challenge to the accuracy of the model. In addition, machine learning models lack generalization ability when facing new situations. In particular, the complexity of the springback problem increases the difficulty of the model to capture all influencing factors, which may lead to prediction errors.

[0004] Therefore, proposing an accurate, fast and more intelligent method for predicting the springback angle of pipe bending has become an urgent problem to be solved. Summary of the invention

[0005] In view of this, the present invention provides a multi-stage prediction method and prediction system for pipe bending springback angle based on machine learning to solve the problems raised in the above background technology.

[0006] To achieve the above objectives, the present invention provides, on one hand, a multi-stage prediction method for pipe bending springback angle based on machine learning, comprising:

[0007] S1: constructing a three-dimensional elastic-plastic finite element numerical simulation model of a round tube, wherein the three-dimensional elastic-plastic finite element numerical simulation model consists of four parts: an upper die, a lower die, a round tube and a mandrel;

[0008] S2: Design orthogonal experiments and obtain sample data under five bending forming process parameters: round tube yield strength, round tube wall thickness, round tube diameter, bending speed, and gap between mandrel and round tube through CAE forming simulation;

[0009] S3: Perform data enhancement on the sample data obtained in S2;

[0010] S4: constructing a multi-stage prediction model and using the enhanced sample data to train and test the multi-stage prediction model to obtain a trained multi-stage prediction model, and then using the trained multi-stage prediction model to predict the springback angle of the round tube bending;

[0011] Among them, the multi-stage prediction model includes a first-stage prediction model based on machine learning and a second-stage prediction model based on machine learning. The input of the first-stage prediction model is the gap between the mandrel and the round tube, the wall thickness of the round tube, the bending speed, the diameter of the round tube and the yield strength of the round tube, and the output is the round tube cross-sectional distortion rate, the thickening rate of the inner wall of the round tube and the thinning rate of the outer wall of the round tube. The input of the second-stage prediction model is the gap between the mandrel and the round tube, the wall thickness of the round tube, the bending speed, the diameter of the round tube, the yield strength of the round tube, the distortion rate of the round tube cross-sectional distortion rate, the thickening rate of the inner wall of the round tube and the thinning rate of the outer wall of the round tube, and the output is the bending rebound angle of the round tube.

[0012] Preferably, in S1, the three-dimensional elastic-plastic finite element simulation model of the aluminum alloy round tube is established using Abaqus / CAE, wherein the tube is simulated using a shell, the tube thickness is set to 3 mm, and the geometric models of the upper and lower molds are established using the modeling-sweep function provided in Abaqus, and are simplified into discrete rigid bodies, and the tube unit type is set to S4R, and 4 integration points are selected in the thickness direction.

[0013] Further optimization, in S1, the tube bending-springback process selects a dual algorithm strategy and is divided into two analysis steps. During the bending process, a dynamic explicit algorithm is used to simulate the downward pressing action of the upper die by fixing the lower die and applying a speed to the upper die. After the bending is completed, the upper die is removed to simulate the steady-state behavior of the round tube when it rebounds.

[0014] Further optimization, in S2, L25 (5^6) is used to carry out a five-factor five-level orthogonal test, and five bending forming process parameters, namely, the yield strength of the round tube, the wall thickness of the round tube, the diameter of the round tube, the bending speed, and the gap between the mandrel and the round tube, are taken. The springback angle, the round tube cross-section distortion rate, the round tube inner wall thickening rate and the round tube outer wall thinning rate are used as evaluation indicators to carry out an orthogonal test.

[0015] Further preferably, in S2, the coordinate method is used to analyze the innermost and outermost node coordinates of the pipe, and a cross-sectional analysis is performed on the pipe. First, the pipe is cut along the part where the wall thickness changes most significantly to obtain the inner and outer ridges of the pipe; then, different positions of the inner and outer circles of the pipe are numbered, and the node coordinates of these positions are extracted. The innermost and outermost nodes on the cross section are selected, and the distance between the two points is calculated, wherein the wall thickness change rate I t =|t0-t1 / t0|×100%, t0 is the wall thickness at the outer ridge before bending, t1 is the wall thickness at the outer ridge after bending, and the cross-sectional distortion rate|D max -D min / D0|×100%, D0 is the nominal diameter of the round tube, D max and D min are the maximum and minimum outer diameters of the pipe cross section measured respectively, and the springback angle of the pipe after forming is Δα=α0-α1, where α0 and α1 represent the actual angle and ideal angle after bending respectively.

[0016] Further preferably, in S3, an unsupervised data enhancement method is used to enhance the data set obtained by the orthogonal experiment.

[0017] Further preferably, the data enhancement method is as follows: the number of samples to be sampled is randomly set between 2 and 5 for sampling without replacement, and after each random number of samples are drawn, the average value is taken as a new sample for data enhancement.

[0018] Further preferably, in S4, the first-stage prediction model and the second-stage prediction model both adopt linear regression models, as follows:

[0019] y=β0+β1x1+β2x2+...+β n x n +∈;

[0020] Among them, y is the output of the model, x1…x n is the input of the model, β0...β n are the parameters of the model, and ∈ is the error term.

[0021] Further preferably, in S4, both the first-stage prediction model and the second-stage prediction model adopt ridge regression model.

[0022] The present invention also provides a multi-stage prediction system for pipe bending springback angle based on machine learning, which is used to execute the above-mentioned multi-stage prediction method for pipe bending springback angle based on machine learning.

[0023] The multi-stage prediction method and prediction system for the tube bending springback angle based on machine learning provided by the present invention decompose the complex prediction task into two smaller and relatively independent prediction subtasks, wherein the first stage focuses on predicting the key factors that directly affect the forming quality (tube cross-section distortion rate, tube inner wall thickening rate and tube outer wall thinning rate), and the input of the second stage uses the output of the first stage, thereby improving the prediction accuracy of the final springback angle; by using data enhancement technology, without relying on time-consuming experiments, it avoids the problem that finite element simulation is highly dependent on high technical reserves and may face low computational efficiency in large-scale or high-complexity simulations, thereby effectively improving the model's processing capability for small sample data and its generalization.

[0024] The multi-stage prediction method and prediction system of pipe bending springback angle based on machine learning provided by the present invention have high prediction accuracy and strong prediction universality, can be applied to actual pipe bending production systems, and can be trained using real-time production data, which can provide strong technical support for the precise forming of pipes. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 A flow chart of a multi-stage prediction method for pipe bending springback angle based on machine learning provided by the present invention;

[0026] Figure 2 It is a schematic diagram of the assembly structure of the three-dimensional elastic-plastic finite element numerical simulation model;

[0027] Figure 3 Schematic diagram of the springback angle value distribution after pipe data enhancement. DETAILED DESCRIPTION

[0028] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0029] like Figure 1 As shown, the present invention provides a multi-stage prediction method for pipe bending springback angle based on machine learning, comprising:

[0030] S1: Construct a three-dimensional elastic-plastic finite element numerical simulation model of aluminum alloy round tube, such as Figure 2 As shown, the three-dimensional elastic-plastic finite element numerical simulation model consists of four parts: an upper die, a lower die, an aluminum alloy round tube and a mandrel;

[0031] S2: Design an orthogonal experiment and obtain sample data under five bending forming process parameters: round tube material properties (yield strength), round tube wall thickness, round tube diameter, bending speed, and gap between mandrel and round tube through CAE forming simulation;

[0032] S3: Perform data enhancement on the sample data obtained in S2;

[0033] S4: constructing a multi-stage prediction model and using the enhanced sample data to train and test the multi-stage prediction model to obtain a trained multi-stage prediction model, and then using the trained multi-stage prediction model to predict the springback angle of the round tube bending;

[0034] Among them, the multi-stage prediction model includes a first-stage prediction model based on machine learning and a second-stage prediction model based on machine learning. The input of the first-stage prediction model is the gap between the mandrel and the round tube, the wall thickness of the round tube, the bending speed, the diameter of the round tube and the yield strength of the round tube, and the output is the round tube cross-sectional distortion rate, the thickening rate of the inner wall of the round tube and the thinning rate of the outer wall of the round tube. The input of the second-stage prediction model is the gap between the mandrel and the round tube, the wall thickness of the round tube, the bending speed, the diameter of the round tube, the yield strength of the round tube, the distortion rate of the round tube cross-sectional distortion rate, the thickening rate of the inner wall of the round tube and the thinning rate of the outer wall of the round tube, and the output is the bending rebound angle of the round tube.

[0035] This multi-stage prediction method for the bending springback angle of tubes based on machine learning first uses the gap between the mandrel and the tube, the wall thickness of the tube, the bending speed, the tube diameter and the tube yield strength to predict the three factors that directly affect the forming quality: the tube cross-sectional distortion rate, the tube inner wall thickening rate and the tube outer wall thinning rate. Then, the second-stage prediction is carried out using the five factors of the gap between the mandrel and the tube, the tube wall thickness, the bending speed, the tube diameter and the tube yield strength and the three intermediate indicators of the first-stage prediction, which can more effectively predict the final bending springback angle of the tube.

[0036] In S1, the three-dimensional elastic-plastic finite element simulation model of the aluminum alloy round tube was established using Abaqus / CAE. The tube was simulated by a shell and the tube thickness was set to 3 mm. The geometric models of the upper and lower molds were established using the modeling-sweep function in Abaqus and simplified into discrete rigid bodies. The tube unit type was set to S4R, and 4 integration points were selected in the thickness direction. The dual algorithm strategy was selected for the tube bending-springback process and divided into two analysis steps. During the bending process, a dynamic explicit algorithm was used to simulate the downward pressing action of the upper mold by fixing the lower mold and applying a velocity to the upper mold. After the bending was completed, the upper mold was removed to simulate the steady-state behavior of the round tube during springback, so as to take into account both calculation accuracy and calculation efficiency. To simulate the actual working conditions, the end of the tube that was not bent was set as a fixed constraint with a length of 300 mm, and the mandrel constrained the same side end face to maintain the stability of the tube during the bending process.

[0037] In this embodiment, the experimental round tube is made of 7075 aluminum alloy, and the material properties are shown in Table 1.

[0038] Table 1:

[0039]

[0040] After CAE modeling, in S2, L25 (5^6) was used to conduct a five-factor five-level orthogonal test. Five bending forming process parameters, including round tube material properties, round tube wall thickness, round tube diameter, bending speed, and mandrel-to-round tube gap, were selected. The springback angle, section distortion rate, maximum inner wall thickening rate, and maximum outer thinning rate were used as evaluation indicators to conduct an orthogonal test.

[0041] Among them, the corresponding table of factor levels is shown in Table 2;

[0042] Table 2:

[0043]

[0044] The coordinate method is used to analyze the coordinates of the innermost and outermost nodes of the pipe, and the cross-section analysis of the pipe is performed. The pipe is cut along the part where the wall thickness changes most significantly to obtain the inner and outer ridges of the pipe; then the different positions of the inner and outer circles of the pipe are numbered, and the node coordinates of these positions are extracted. The innermost and outermost nodes on the cross section are selected, and the distance between the two points is calculated. The wall thickness change rate I t =|t0-t1 / t0|×100%, t0 is the wall thickness at the outer ridge before bending, t1 is the wall thickness at the outer ridge after bending, and the cross-sectional distortion rate|D max -D min / D0|×100%, D0 is the nominal diameter of the round tube, D max and D min They are the maximum and minimum outer diameters of the pipe cross section measured respectively. The springback angle Δα=α0-α1 is generally used for the springback of the pipe after forming. The difference between the actual angle after bending and the ideal angle is calculated to evaluate the forming accuracy.

[0045] In S3, an unsupervised data augmentation method is used to enhance the dataset obtained by the orthogonal experiment.

[0046] The specific enhancement method is as follows: randomly set the number of samples to be sampled between 2 and 5 for sampling without replacement, and take the average value as the new sample after drawing a random number of samples each time to perform data enhancement, enrich the input space distribution of the training set, and enhance the generalization of the data set.

[0047] The enhanced data set is divided into a training set and a test set. In this embodiment, the training set generates 25 samples by orthogonal experiment (as shown in Table 3), 35 samples are fused by data enhancement, and 60 samples are generated by random sampling in the test set. Five key features of yield strength, wall thickness, tube diameter, bending speed and gap are recorded in each sample set, and three intermediate test results: distortion rate, thickening rate and thinning rate, as well as the final predicted bending springback angle of the aluminum alloy round tube.

[0048] Table 3:

[0049]

[0050]

[0051] Data augmentation technology can avoid the need to rely on traditional time-consuming experiments, and also avoid the problem that finite element simulation is highly dependent on high technical reserves and may face the problem of low computational efficiency in large-scale or high-complexity simulations; through data augmentation, a copy of an existing data set can be created, such as Figure 3 , adding data with a rebound angle in the range of 3 to 5 degrees. The improvement can not only enhance the model's ability to handle small sample data, but also significantly improve the generalization performance of the model.

[0052] The first-stage prediction model and the second-stage prediction model both use a linear regression model (LR), as follows:

[0053] y=β0+β1x1+β2x2+...+β n x n +∈;

[0054] Among them, y is the output of the model, x1…x n is the input of the model, β0...β n is the parameter of the model, ∈ is the error term; in the process of training the model, the value of the parameter is estimated by minimizing the residual sum of squares between the actual observation value and the model prediction value.

[0055] Further preferably, both the first-stage prediction model and the second-stage prediction model adopt ridge regression model.

[0056] Table 4 shows the performance comparison results of multi-stage prediction using linear regression model, ridge regression model and several other models. The model evaluation indicators are R2 score, mean absolute error (MAE) or mean absolute percentage error (MAPE). R2 (also known as the coefficient of determination) is an indicator to measure the goodness of fit of the model. Its value range is from negative infinity to 1. It can be interpreted as the degree to which the model explains the variability of the data. In R2 calculation, the total sum of squares (SST) is calculated first:

[0057]

[0058] Among them, y i is the ith observation, is the mean of the observed values, and n is the number of samples;

[0059] Secondly, calculate the regression sum of squares (SSR):

[0060]

[0061] in, is the i-th value predicted by the model;

[0062] Then, calculate the residual sum of squares (SSE):

[0063]

[0064] Finally calculate R2:

[0065]

[0066] MAE is an indicator used to measure the difference between the predicted value and the actual value. It is calculated by the average of the absolute values ​​of all prediction errors. The formula of MAE is as follows:

[0067]

[0068] Where n is the number of samples, y i is the ith observation, is the i-th value predicted by the model;

[0069] The calculation formula of MAPE is as follows:

[0070]

[0071] Table 4:

[0072]

[0073] It can be seen from Table 4 that after decomposing the rebound angle prediction task into two-stage subtasks, except for decision tree and XGBoost, the indicators of other machine learning models have been improved to a certain extent. Among them, the improvement of multi-layer perceptron and random forest is the most obvious, and the best performing linear regression also has a slight performance improvement, which proves that the above task decomposition method has a significant effect on improving performance.

[0074] It should be noted that the R2 correlation score of the springback angle on the test set of the multi-stage prediction model based on ridge regression reached 0.9669, which shows that the model has high prediction accuracy and strong universality in predicting pipe bending springback.

[0075] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is limited by the attached embodiments and their equivalents.

Claims

1. A multi-stage prediction method for pipe bending springback angle based on machine learning, characterized in that: include: S1: constructing a three-dimensional elastic-plastic finite element numerical simulation model of a round tube, wherein the three-dimensional elastic-plastic finite element numerical simulation model consists of four parts: an upper die, a lower die, a round tube and a mandrel; S2: Design orthogonal experiments and obtain sample data under five bending forming process parameters: round tube yield strength, round tube wall thickness, round tube diameter, bending speed, and gap between mandrel and round tube through CAE forming simulation; S3: Perform data enhancement on the sample data obtained in S2; S4: constructing a multi-stage prediction model and using the enhanced sample data to train and test the multi-stage prediction model to obtain a trained multi-stage prediction model, and then using the trained multi-stage prediction model to predict the springback angle of the round tube bending; Among them, the multi-stage prediction model includes a first-stage prediction model based on machine learning and a second-stage prediction model based on machine learning. The input of the first-stage prediction model is the gap between the mandrel and the round tube, the wall thickness of the round tube, the bending speed, the diameter of the round tube and the yield strength of the round tube, and the output is the round tube cross-sectional distortion rate, the thickening rate of the inner wall of the round tube and the thinning rate of the outer wall of the round tube. The input of the second-stage prediction model is the gap between the mandrel and the round tube, the wall thickness of the round tube, the bending speed, the diameter of the round tube, the yield strength of the round tube, the distortion rate of the round tube cross-sectional distortion rate, the thickening rate of the inner wall of the round tube and the thinning rate of the outer wall of the round tube, and the output is the bending rebound angle of the round tube.

2. The multi-stage prediction method for pipe bending springback angle based on machine learning according to claim 1 is characterized in that: In S1, the three-dimensional elastic-plastic finite element simulation model of the aluminum alloy round tube was established using Abaqus / CAE, in which the tube was simulated using a shell and the tube thickness was set to 3 mm. The geometric models of the upper and lower molds were established using the modeling-sweep function in Abaqus and simplified into discrete rigid bodies. The tube unit type was set to S4R, and 4 integration points were selected in the thickness direction.

3. The multi-stage prediction method for pipe bending springback angle based on machine learning according to claim 1 is characterized in that: In S1, the dual algorithm strategy is selected for the tube bending-springback process and is divided into two analysis steps. During the bending process, a dynamic explicit algorithm is used to simulate the downward pressing action by fixing the lower die and applying a velocity to the upper die. After the bending is completed, the upper die is removed to simulate the steady-state behavior of the round tube when it rebounds.

4. The multi-stage prediction method for pipe bending springback angle based on machine learning according to claim 1 is characterized in that: In S2, L25 (5^6) was used to carry out a five-factor five-level orthogonal test. Five bending forming process parameters, namely, yield strength of round tube, wall thickness of round tube, diameter of round tube, bending speed, and gap between mandrel and round tube, were selected. The springback angle, round tube cross-section distortion rate, round tube inner wall thickening rate, and round tube outer wall thinning rate were used as evaluation indicators for orthogonal test.

5. The multi-stage prediction method for pipe bending springback angle based on machine learning according to claim 1 is characterized in that: In S2, the coordinate method is used to analyze the coordinates of the innermost and outermost nodes of the pipe, and the cross-sectional analysis of the pipe is performed. First, the pipe is cut along the part where the wall thickness changes most significantly to obtain the inner and outer ridges of the pipe; then the different positions of the inner and outer circles of the pipe are numbered, and the node coordinates of these positions are extracted. The innermost and outermost nodes on the cross section are selected, and the distance between the two points is calculated. Among them, the wall thickness change rate I t =|t0-t1 / t0|×100%, t0 is the wall thickness at the outer ridge before bending, t1 is the wall thickness at the outer ridge after bending, and the cross-sectional distortion rate|D max -D min / D0|×100%, D0 is the nominal diameter of the round tube, D max and D min are the maximum and minimum outer diameters of the pipe cross section measured respectively, and the springback angle of the pipe after forming is Δα=α0-α1, where α0 and α1 represent the actual angle and ideal angle after bending respectively.

6. The multi-stage prediction method for pipe bending springback angle based on machine learning according to claim 1 is characterized in that: In S3, an unsupervised data augmentation method is used to enhance the dataset obtained by the orthogonal experiment.

7. The multi-stage prediction method for pipe bending springback angle based on machine learning according to claim 6 is characterized in that: The data enhancement method is as follows: the number of samples is randomly set between 2 and 5 for sampling without replacement. After each random number of samples is drawn, the average value is taken as the new sample for data enhancement.

8. The multi-stage prediction method for pipe bending springback angle based on machine learning according to claim 1 is characterized in that: In S4, both the first-stage prediction model and the second-stage prediction model adopt linear regression models, which are as follows: y=β0+β1x1+β2x2+...+β n x n +∈; Among them, y is the output of the model, x1…x n is the input of the model, β0...β n are the parameters of the model, and ∈ is the error term.

9. The multi-stage prediction method for pipe bending springback angle based on machine learning according to claim 1 is characterized in that: In S4, both the first-stage prediction model and the second-stage prediction model adopt the ridge regression model.

10. A multi-stage prediction system for pipe bending springback angle based on machine learning, characterized in that: Used to execute the multi-stage prediction method of tube bending springback angle based on machine learning as described in any one of claims 1 to 9.

Images