CN112036062A - Metal material bending forming rebound angle prediction method - Google Patents
Metal material bending forming rebound angle prediction method Download PDFInfo
- Publication number
- CN112036062A CN112036062A CN202010789043.8A CN202010789043A CN112036062A CN 112036062 A CN112036062 A CN 112036062A CN 202010789043 A CN202010789043 A CN 202010789043A CN 112036062 A CN112036062 A CN 112036062A
- Authority
- CN
- China
- Prior art keywords
- metal material
- angle
- bending
- data
- neural network
- 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
- G06F30/23—Design optimisation, verification or simulation using finite element methods [FEM] or finite difference methods [FDM]
-
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/004—Artificial life, i.e. computing arrangements simulating life
- G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2111/00—Details relating to CAD techniques
- G06F2111/06—Multi-objective optimisation, e.g. Pareto optimisation using simulated annealing [SA], ant colony algorithms or genetic algorithms [GA]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2111/00—Details relating to CAD techniques
- G06F2111/10—Numerical modelling
-
- 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
- G06F2119/14—Force analysis or force optimisation, e.g. static or dynamic forces
Abstract
A metal material bending forming rebound angle prediction method comprises the following steps: establishing a bending resilience finite element model, and performing a virtual orthogonal test on the bending resilience of the metal by using finite element simulation software to obtain simulation data; performing range analysis on the numerical simulation data, and determining the main factor with the most obvious influence on the bending resilience angle; three levels are inspected by each main factor, an orthogonal test table is established, and each level of the main factor is randomly subjected to one experiment; calculating the ratio of experimental data and simulation data of the springback angle under the other same horizontal conditions of the main factor to obtain a correction coefficient of the springback angle under each horizontal condition of the main factor; multiplying the correction coefficient by the corresponding numerical simulation data to obtain a correction value; establishing a prediction model based on the neural network, and training and learning the neural network by using correction data under the condition of virtual orthogonal test working conditions; and predicting the bending resilience angle of the metal material under the working condition by using the trained neural network model according to the unknown working condition data.
Description
Technical Field
The invention relates to the field of metal material bending forming processes, in particular to a metal material bending forming rebound angle prediction method.
Background
At present, the metal bending part is easy to realize the light weight and the obdurability of products and meet the requirements of low consumption, high efficiency and accurate manufacturing, and is widely applied to the fields of aviation, aerospace, ships, vehicles, chemical engineering, medical treatment, energy sources and the like as a conveying path or a structure of gas and liquid.
Aiming at the problem of resilience of metal material in bending forming, scholars at home and abroad carry out a great deal of research, and at present, there are three main types of theory analysis, numerical simulation and experimental research.
The theoretical analytical method reveals and solves some practical problems in the bending production of metal materials to a certain degree, promotes the forward development of the bending analytical theory, but mostly bases on a large amount of assumptions and simplifications, for example, assuming the metal pipe as an ideal elastic-plastic material, neglecting the displacement of a stress-strain neutral layer, neglecting the change of material parameters such as shrinkage strain ratio, elastic modulus and the like in the bending process of the pipe, neglecting the shear stress on the cross section of a bent pipe and the like, lacks tightness, has large analytical error, and limits the popularization and application of the method in the practical production.
The introduction of a numerical analysis method enables the research of the bending process of the metal material to obtain a long-term progress, but a certain difference exists between the research and the production practice, mainly because the bending of the metal material has material nonlinearity, geometric nonlinearity and boundary nonlinearity, relates to the influence of many uncertain factors, is a very complex multiple nonlinear strong coupling deformation process, and the mechanism of the deformation process cannot be well mastered.
In the aspect of experimental research, the actual springback data can truly represent the influence of a plurality of nonlinear interactive coupling factors on springback, but the methods do not consider the influence of the incompleteness of a test sample on the prediction accuracy, or directly fill the finite element simulation data into sample data, cannot fill missing data by utilizing the internal rules of a bending forming system reflected by finite element analysis, and have a certain difference from the actual application.
In a word, in the existing research method, the analytic calculation method has too many assumptions, and the prediction accuracy is generally not high; the numerical simulation method has a certain gap from practical application because the bending deformation mechanism of the pipe cannot be well mastered, and the experimental research method cannot achieve the effect of practical application because the incompleteness of a test sample is not considered or finite element simulation data is directly used for filling sample data.
Disclosure of Invention
The invention aims to solve the problems in the prior art and provides a method for predicting the bending forming resilience angle of a metal material, missing data is filled by utilizing the internal rule of a bending forming system reflected by finite element analysis, the analysis error is small, the prediction accuracy is high, and the method can be popularized and applied in practical production.
The invention adopts the technical scheme for solving the technical problems that: the method for predicting the bending forming rebound angle of the metal material comprises the following steps:
step 1: establishing a metal material bending resilience finite element model, and performing a virtual orthogonal test on the metal material bending resilience by using finite element simulation software to obtain numerical simulation data of the metal material bending resilience angle;
step 2: performing range analysis on numerical simulation data of the bending resilience angle of the metal material, and determining a main factor which has the most obvious influence on the bending resilience angle of the metal material;
and step 3: each main factor examines three levels, a field opening orthogonal test table is established, each level of the main factors is randomly selected from the virtual orthogonal test table to perform a real experiment, and experimental data of the bending resilience angle of the metal material are obtained;
and 4, step 4: dividing experimental data of the bending resilience angle of the metal material under a certain horizontal condition of the main factor by numerical simulation data to obtain a correction coefficient of the resilience angle under the horizontal condition of the main factor; calculating the ratio of experimental data and numerical simulation data of the springback angle under other same horizontal conditions of the main factor in sequence to obtain a correction coefficient of the springback angle under each horizontal condition of the main factor;
and 5: multiplying the correction coefficient of the main factor under each horizontal condition by the corresponding numerical simulation data to obtain the correction value of the virtual orthogonal test under each working condition;
step 6: establishing a prediction model of the bending resilience angle of the metal material based on the artificial neural network, training and learning the neural network by using correction data under the working condition of the virtual orthogonal test, inputting unknown working condition data of the bending resilience of the metal material, and obtaining a predicted value of the bending resilience angle of the metal material under the unknown working condition through calculation of the neural network.
For further improvement, the metal material includes a profile, a pipe, a plate, and the like.
For further improvement, the metal material is a Q235 pipe, the phase transformation general structure of the Q235 carbon steel is composed of ferrite and pearlite, the strength and the hardness are lower, and the plasticity and the toughness are better, so that the material is a good cold-bending processing material.
Further perfection, the stress that the metal material receives when bending is:
wherein σ is the actual stress, σ s is the yield limit, E is the Young's modulus, which is the actual strain, K is the hardening coefficient of the material, and n is the hardening index of the material.
Further perfection, the influence factors of the bending resilience angle of the metal material comprise a bending angle, a relative bending radius, a relative wall thickness and a friction coefficient.
Further perfecting, before step 6, the correction value in step 5 is normalized
Further perfecting, the normalization processing formula is as follows:
wherein Xij is the j value of the i input/output parameter after normalization, max (xi), min (xi) are the upper and lower limit values of the parameter respectively; xNmin, xNmax is the normalized threshold value, 0< xNmin < xNmax <1, where xNmin is 0.1 and xNmax is 0.9.
And further perfecting, wherein the artificial neural network in the step 6 is one or a combination of more of a BP neural network, a particle swarm optimization algorithm and a mixed particle swarm algorithm.
The invention has the beneficial effects that: the invention adopts an orthogonal test method, selects test influence factors as a bending angle, a relative bending radius, a relative wall thickness and a friction coefficient, takes 3 levels of each influence factor respectively, takes the size of a rebound angle as a test index, and has balanced dispersibility and uniformity comparability in the orthogonal test, thereby ensuring that the test times are less, the effect is good, the method is simple and convenient, and the efficiency is high; according to the method, an ABAQUS software is used for simulating a pipe fitting bending forming process and an unloading springback process on the basis of a finite element model of a bent pipe through an orthogonal test, simulated experiment data and real experiment data are obtained through the orthogonal test, so that a correction coefficient and a correction value of a springback angle under each level condition of a main factor are obtained, finally, unknown bending springback working condition data of a metal material are input after a neural network is trained and learned, and a predicted value of the bending springback angle of the metal material under the unknown working condition can be obtained through calculation of the neural network.
Drawings
FIG. 1 is a finite element model of a bent pipe according to the present invention;
FIG. 2 is a simplified finite element model of a bent pipe according to the present invention;
FIG. 3 is a reference point setting for a rigid body;
FIG. 4 is a neural network training error curve based on simulation experiment correction data;
FIG. 5 is a neural network sample fitting curve based on simulation experiment correction data;
FIG. 6 is a state diagram of neural network training based on simulation experiment correction data;
FIG. 7 is a neural network training sample regression analysis based on simulation experiment correction data;
FIG. 8 is a neural network training sample output error curve based on simulation experiment correction data;
FIG. 9 is a flow chart of a hybrid particle swarm algorithm;
FIG. 10 is a knowledge mining model of catheter bending resilience;
FIG. 11 is a neural network training error curve based on raw data from a simulation experiment;
FIG. 12 is a sample fitting curve for a neural network based on raw data from a simulation experiment;
FIG. 13 is a neural network fit error curve based on raw data from simulation experiments.
Description of reference numerals: 1. pipe material, 2, mould.
Detailed Description
The invention will be further described with reference to the accompanying drawings in which:
with reference to the accompanying drawings: the method for predicting the rebound angle of the metal material during bending forming comprises the following steps:
step 1: establishing a finite element model of the bent pipe as shown in figure 1 based on an experimental pipe bender, firstly establishing a three-dimensional model of each mould of the bent pipe by using UG NX Software produced by Siemens PLM Software company, then leading the three-dimensional model into ABAQUS Software, establishing a pipe model, adding an Assembly module for Assembly, establishing a finite element model of the bent pipe for subsequent analysis, and performing a virtual orthogonal test on bending resilience of a metal material by using finite element simulation Software to obtain numerical simulation data of the bending resilience angle of the metal material;
as shown in fig. 2, for the thin-walled tube, the following assumptions were made: the wall thickness value of the thin-wall pipe is very small, the radial stress can be ignored, and the pipe is in a plane stress state; secondly, the mould is considered as a rigid body and can not be deformed; considering that the material is subjected to Mises yield criterion, isotropy and isotropic reinforcement, the research further simplifies and analyzes the finite element model according to the assumptions, for example, a thin-wall pipe is taken as a shell, a mould is taken as a discrete rigid body, a pipe forming mould is simplified into a cavity surface capable of realizing the same constraint function, and a bent pipe finite element simplified model is established;
as shown in fig. 3, taking a Q235 pipe with an outer diameter D of 42mm, a wall thickness t of 5mm, and a length l of 3380mm as an example, geometric modeling, defining analysis steps and outputs, defining interactions, defining boundary conditions, meshing, bending simulation, and unloading rebound are sequentially performed. The reference point is set as the rotation center of the bending die and the jaw during geometric modeling, and various performance parameters of the rigid body are unchanged during the movement around the rigid body;
when the analysis Step and the interaction are defined, besides the original analysis Step, two analysis steps of Step-1 bending and Step-2 rebounding are defined, wherein Step-1 is the rotation quantity of a jaw and a bending die around a reference point, the contact of the pipe with the jaw and the bending die is set to be rough, the contact of the pipe with a follow-up die is set to be initially 0.1, and the set time length is 1; step-2 is a displacement vector after unloading, the set time length is 3, the contact of the pipe with a jaw and a bending die is set to be invalid, namely, no constraint exists between the contact and the jaw and the bending die, and the simulation of a springback process can be realized;
in order to obtain simulation data of the bending resilience angle of the pipe fitting, ABAQUS software is used for simulating the bending forming process and the unloading resilience process of the pipe fitting. After the simulation of the two processes is finished, subtracting the front springback forming angle from the back springback forming angle to obtain a springback angle;
step 2: performing range analysis on numerical simulation data of the bending resilience angle of the metal material, and determining a main factor which has the most obvious influence on the bending resilience angle of the metal material;
and step 3: in combination with four main factors affecting the angle of springback of the pipe around bends: bending angle theta, relative bending radius R/D, relative wall thickness D/t and pipe die (follow-up die) friction coefficient t. The tested test index is the size of the rebound angle; examining 3 levels for each influence, as shown in the following tables 3-3, selecting an L9(34) Taokou orthogonal table, designing an experimental scheme, randomly selecting each level of a main factor from a virtual orthogonal test table to perform a real experiment, obtaining experimental data of the bending resilience angle of the metal material, forming 9 experiments, and obtaining a numerical simulation value of the resilience angle according to the steps as shown in the following tables 3-4:
TABLE 3-3 factor level table
TABLE 3-4 orthogonal test protocol and rebound Angle numerical simulation test data
And 4, step 4: the Q235 carbon steel pipe fitting adopting furnace cooling comprises the following chemical components: 0.14% w (C), 0.16% w (Si), 0.70% w (Mn), 0.015% w (P), and 0.008% w (S), and for a low carbon steel tube having an outer diameter of 42mm and a wall thickness of 4.6mm, according to national standard GB/T228.1-2010 part 1 of metallic Material tensile test: a tensile Q235 pipe sample is designed according to a room temperature test method, an AG-X100 kN microcomputer controlled electronic universal tester is used for carrying out a one-way tensile test at the speed of 1mm/min, and a power index hardening model is applied to fit the stress-strain relation of a plasticity stage.
And (3) taking the Q235 pipe as a research object, starting a man-machine interaction system of the experimental device according to actually measured bending resilience data acquired on a test table of the KM-A90-CNC-E320 pipe bender, and inputting or importing pipe shape data of the experimental bent pipe. And the processing technological parameters of the experimental bent pipe, such as the bending radius and the like, are set well on a 'parameter setting' interface of the upper computer human interaction system. And after the steps are completed, pressing a start button of the numerical control system, and finishing the bending processing work of the guide pipe by the pipe bender according to the process formula and the parameters to obtain the processed bending formed pipe fitting.
The forming angle of the pipe part which is formed by winding is obtained through measurement, and the value of the springback angle can be obtained by subtracting the target forming angle of the numerical control system. Corresponding physical measurements were performed according to the orthogonal test protocol of tables 3-4 to obtain the data of the springback test, as shown in table 4-1.
TABLE 4-1 catheter wrap-around springback test data
The pole difference analysis is performed according to the springback simulation data tables 3-4, and the influence of the bending angle on springback is significant in the four factors to be considered, so that the measured data of a certain bending angle is used for carrying out equal proportion correction on the simulation values of different parameters under the same bending angle. In the orthogonal experiment table of tables 3-4, a group of experiments are extracted for the same bending angle and measured on site, and the ratio of the measured rebound angle value to the simulated rebound angle value is calculated, so as to obtain the correction coefficient shown in table 5-1, and the result of correcting the rebound angle is obtained by the correction coefficient shown in table 5-2.
TABLE 5-1 coefficient of resilience angle correction
And 5: multiplying the correction coefficient of the main factor under each horizontal condition by the corresponding numerical simulation data to obtain the correction value of the virtual orthogonal test under each working condition, as shown in table 5-2;
TABLE 5-2 Resilience Angle correction results
In order to avoid the network being unable to converge and to accelerate the convergence of the training network, the data is normalized, and the normalization method is as follows:
wherein Xij is the j value of the i input/output parameter after normalization, max (xi), min (xi) are the upper and lower limit values of the parameter respectively; xNmin, xNmax is the normalized threshold value, 0< xNmin < xNmax <1, where xNmin is 0.1 and xNmax is 0.9.
Normalizing the orthogonal simulation experiment result and the springback angle correction data according to the formula to obtain normalized orthogonal simulation experiment correction data, wherein the normalized orthogonal simulation experiment correction data are shown in a table 5-3;
TABLE 5-3 corrected data of normalized orthogonal simulation experiment
Step 6: a prediction model of the bending resilience angle of the metal material is established based on the artificial neural network, and neural network training samples are extracted according to the normalized orthogonal simulation experiment correction data in the table 5-3, as shown in the table 5-4.
The BP neural network is a multi-layer feedforward network trained according to an error back propagation algorithm. Theory has demonstrated that a three-layer BP network model enables arbitrary sequential mapping. The three-layer BP network model only comprises a hidden layer, and comprises an input layer, the hidden layer or an intermediate layer and an output layer; the input layer has i nodes, the hidden layer has j nodes, and the output layer has t nodes. The neurons of the upper layer and the lower layer are in full connection, namely, each unit of the lower layer is in weight connection with each unit of the upper layer, and the neurons of each layer are not in connection. The network learns according to the teaching of a teacher, when a pair of learning modes is provided for the network, the neuron activation value is transmitted from the input layer to the output layer through each intermediate layer, and each neuron of the output layer obtains the input response of the network. Thereafter, the connection weights are modified layer by layer from the output layer through the intermediate layers in a direction to reduce the error between the desired output and the actual output, and finally returned to the input layer.
The algorithm comprises the following steps:
(1) the initial weight coefficient w (0) is set to a small random non-zero value.
(2) Given an input/output sample pair, the output of the computing network:
let the input and output of the p-th group of samples be
up＝(u1p，u2p，…，unp)
dp＝(d1p，d2p，…，dnp)p＝1，2，…，L
When the p-th group of samples are input, the output of the node i is
In the formula IjP-the jth input of node i at the input of the pth set of samples; f is an excitation function, usually Sigmoid function, i.e.
The input of the network output layer node can be obtained from the input layer to the output layer through the hidden layer.
(3) Computing an objective function J of a network
Let Ep be the objective function of the network when the p-th group of samples is input, take the L2 norm
In the formula, ykp (t) -when the p group of samples is input, the output of the network is adjusted by the weight t times, and k is the kth node of the output layer.
The overall objective function of the network is:
as an evaluation of the network learning status.
And (3) discrimination: if it is
J≤ (5.6)
Wherein-predetermined, ≧ 0
The algorithm ends, otherwise, go to step (4).
(4) Back propagation computation
And (4) reversely calculating according to a gradient descent method by the output layer according to J, and adjusting the weight layer by layer.
Where η is the step size or learning rate.
The basic idea of the particle swarm optimization algorithm is that the potential solution of each optimization problem is equivalent to a bird in the search space, called a particle, all the particles have an adaptive Value (Fitness Value) determined by an optimization function, each particle also has a velocity vector to determine the flying direction and distance of the particle, and then the particles follow the current optimal particle to search in the solution space.
The PSO first initializes a population of random particles (random solution) and then finds the optimal solution through iteration. In each iteration, the particle updates itself by tracking two "extrema". One extreme value is the optimal solution found by the particle itself to the current moment and is called as an individual extreme value Pi; and the other is the optimal solution found from the whole population to the current moment, which is called as a global extreme value Pg. When the two optimal values are found, the particle updates the speed and the new position of the particle by using the information of the particle, the individual extreme value and the global extreme value. Therefore, the particle swarm optimization algorithm is also used for completing the search of the optimal solution in a complex search space based on individual cooperation and competition, and is an evolution calculation technology based on a swarm intelligence method.
Let the D-dimensional search space have N particles, where the position of the ith particle (i-1, 2, …, N) is Xi (Xi1, Xi2, …, xiD) and the velocity is Vi (Vi1, Vi2, …, viD). And substituting the Xi into the objective function to calculate the adaptive value of the Xi, and measuring the quality of the Xi according to the adaptive value. The optimal position searched by the ith particle is represented by vector Pi ═ (Pi1, Pi2, …, piD), and the optimal position searched by the whole particle swarm is represented by vector Pg ═ (Pg1, Pg2, …, pgD). The particle state update operation is as follows:
the formulas (5.8) and (5.9) form a standard particle swarm optimization algorithm. Wherein D is 1, 2, …, D; k is the current iteration number; learning factor c1And c2Is a non-negative constant, and takes a value between 0 and 2; r is1And r2Is [0, 1 ]]Random numbers are used, and the random numbers are independent of each other and are used for keeping the diversity of the population; vid ∈ [ -Vmax, Vmax]Vmax is a constant set by the user, the velocity of the particle is limited to a maximum velocity Vmax range; the inertia weight omega is a non-negative number and describes the influence of the previous generation speed of the particle on the current generation speed, so that the particle can be protectedThe motion inertia is kept, so that the mobile terminal has the trend of expanding a search space and can search a new area; the inertial weight is selected to have a larger value, so that the global search capability of the algorithm is improved, and the local search capability of the algorithm is enhanced by selecting a smaller value; since different problems have different requirements on the global or local search capability of the algorithm, the balance relationship between the global search capability and the local search capability of the algorithm is preferably adjustable, that is, the inertia weight can be automatically adjusted according to different problems. In order to enable the algorithm to have stronger exploration capability in the initial stage of iteration, continuously search a new area, gradually enhance development capability and enable the algorithm to finely search around a possible optimal solution, a Linear Decreasing Weight (LDW) strategy for linearly decreasing omega value along with iteration is adopted,
namely:
in the formula, ω max and ω min are the maximum and minimum weights of ω, and typically take a value ω max of 0.9 to 1.4 and a value ω min of 0.4; k is the current evolutionary algebra and kmax is the maximum evolutionary algebra. And when the set maximum evolution algebra is reached or the optimal position searched by the particle swarm meets the set minimum adaptive threshold, the iteration is stopped.
The particle swarm optimization algorithm is a global search algorithm, convergence is rapid in the initial stage of global search, but near a global optimal value, the search process of the particle swarm optimization algorithm becomes very slow and sometimes stays before. On the contrary, the gradient descent algorithm is a local search algorithm, and a fast convergence rate can be obtained near the global optimal value, but the local optimal value is easy to be trapped. The combination of the two can better combine the stronger global search capability of the PSO algorithm with the stronger local search capability of the BP algorithm.
Among the derivative optimization-based BP algorithms, the L-M (Levenberg-Marquardt) optimization algorithm is one of the most successful typical algorithms. The L-M algorithm was developed from the classical Newton algorithm, which uses a nonlinear least squares derivation. The iterative formula of the L-M algorithm is as follows:
where I is the identity matrix and λ is a non-negative value. With the help of the amplitude variation of λ, the method smoothly varies between two extreme cases: namely the Newton method (λ → 0) and the standard gradient method (when λ → ∞). It follows that the L-M algorithm actually combines the advantages of both the Newton method and the standard gradient descent method.
Therefore, the particle swarm optimization algorithm and the L-M algorithm are introduced to carry out hybrid optimization on the neural network to form the PSOLM hybrid particle swarm optimization algorithm. The PSOLM hybrid particle swarm optimization method has the main idea that the PSO algorithm is used as a main frame, PSO optimization is firstly carried out, after a plurality of generations of evolution, individuals of particle swarm are randomly selected from the PSO algorithm to carry out L-M algorithm optimization searching for a plurality of steps, and local depth searching is carried out. Here, in order to avoid losing the diversity of the particles in the PSO search process, the diversity of the particles is improved by randomly selecting one particle and carrying out L-M search to update the position of the particle; in order to avoid that the PSO algorithm is slow in searching near the global optimum value, a heuristic algorithm from PSO searching to L-M searching is adopted; and meanwhile, introducing a minimum constant to improve the operability of the hair style algorithm.
The specific steps of the algorithm are as follows:
1) randomly initializing the speed and position of a group of particles;
2) evaluating the initial fitness of each particle by using the mean square error MSE as an adaptive function of the PSO, wherein the current position of each particle is used as the current optimal solution Pi of the particle, and Pg is set as the position of the best particle in the initialized particles;
3) if the algorithm runs to the maximum iteration number, or the training error E is smaller than the specified error standard E, the step 9) is carried out, otherwise, the step 4) is carried out;
4) storing the best current particles, updating the positions and the speeds of all the particles according to the formula (5.8) and the formula (5.9), and generating a new group of particles, namely adjusting the connection weight and the threshold of the neural network;
5) the fitness of each new particle is evaluated and the worst fitness particle is replaced by the best stored particle. If the new position of the ith particle is better than the value of Pi, the value of Pi is updated to be the position of the current particle. If the best position of all the new particles is better than that of the Pg, updating the Pg to be the best position of all the new particles;
6) if the change of the Pg value is very small (the absolute value of the difference value of the Pg values of two adjacent generations appears for 10 times continuously is smaller than a given minimum value constant), local depth search is carried out for a plurality of steps near the Pg by using an L-M algorithm, if the search result is better than the Pg, the search result is used for replacing the Pg, and the step 8 is executed. Otherwise, executing step 7);
7) randomly selecting one particle from the current particle swarm to carry out L-M search, and replacing the worst particle in the current particle swarm with an L-M search result;
8) linearly reducing the inertia weight omega according to the formula (5.10), and going to execute the step 3);
9) and outputting the optimal result Pg.
TABLE 5-4 training sample of correction data for orthogonal simulation experiment
The knowledge of catheter wrap resilience (weights and thresholds of the neural network) is obtained by training of the neural network. The method comprises the steps of training a neural network model by using an orthogonal simulation experiment correction data training sample mentioned above and a gradient descent method with momentum, wherein a tan-sigmoid function is selected as a first-layer transfer function of a BP network, a linear function is used as a transfer function of an output layer, the maximum iteration time is set to be 5000 times, a target error is 10-4, a learning rate is 0.001, the learning rate is obtained through trial calculation, the number of hidden layer units is 5, and a 4-5-1 catheter bending springback knowledge mining model is established, as shown in FIG. 10.
By using the network structure and parameter setting, the BP neural network is trained, the BP neural network is rapidly converged through 3648 suboptimal calculation, the CPU takes 8 seconds, the objective function reaches 9.9969e-05, and a good effect is obtained, and a neural network training error curve, a training sample fitting curve, a training state, a training sample regression analysis and a sample neural network output fitting error curve are respectively shown in fig. 4, fig. 5, fig. 6, fig. 7 and fig. 8. Knowledge of catheter bending resilience (weights and thresholds) established by training learning is shown in tables 5-5;
TABLE 5-5 knowledge of catheter recoil from bending by training
Fitting the rebound angle of the known sample by using the excavated knowledge of the rebound of the catheter around the bend, wherein the average variance of the prediction of the rebound angle reaches 3.66 percent, as shown in tables 5-6; fitting the rebound angles of the unknown samples by using the excavated knowledge of the rebound of the catheter around the bend, wherein the average error of the prediction of the rebound angles is 10.89, and the prediction is shown in tables 5 to 7; from tables 5-6 and 5-7, it can be seen that learning of the knowledge of the rebound data by the BP neural network not only has higher convergence accuracy, but also has certain generalization capability.
Tables 5-6 fitting results for known samples
Tables 5-7 fitting results for unknown samples
In order to analyze the influence of the data correction method on the data mining result, firstly, the normalization formula is applied to normalize the original data of the orthogonal simulation experiment to obtain the normalized original data of the orthogonal simulation experiment, as shown in tables 5 to 8. The normalized orthogonal simulation experiment original data is used for learning of the neural network, the BP network adopts a 4-5-1 structure, a tan-sigmoid function is selected as a first-layer transfer function, a transfer function of an output layer is a linear function, the maximum iteration frequency is set to be 5000 times, the target error is 10-4, the learning rate is 0.001, and the BP neural network is trained by using a gradient descent method with momentum.
Tables 5-8 normalized raw data from orthogonal simulation experiments
By 3226, the suboptimal calculation is rapidly converged, the CPU takes 9 seconds, the objective function reaches 9.9947e-05, and a good effect is obtained, and the neural network training error curve, the training sample fitting curve and the neural network output fitting error curve are respectively shown in fig. 11, fig. 12 and fig. 13.
Tables 5-9 show the effect of different data samples on the fit results for the springback angle, and tables 5-10 show the effect of different data samples on the fit error for the springback angle.
TABLE 5-9 Effect of different data samples on the Return Angle fitting results
As can be seen from tables 5-9 and 5-10, when the simulation experiment correction data sample is used as the neural network learning sample, the maximum value of the resilience angle error reversely calculated by the network output is 4.38%, and the minimum value is 0.00%; the average error is 1.94%; when the original data sample of the simulation experiment is used as a neural network learning sample, the maximum value of rebound angle errors reversely calculated by the network output is 8.54 percent, the minimum value is 2.90 percent, the number of the 9 samples exceeding the maximum reverse calculation error of the correction data reaches 6, namely the reverse rebound angle errors of 66.7 percent of the samples exceed the maximum reverse calculation error of the correction data, the average error is 5.43 percent, and about 3 times of the average reverse calculation error of the correction data. Therefore, the finite element simulation experiment data is corrected by a method of correcting the simulation values of different parameters under the same bending angle in equal proportion according to the actual measurement data of a certain bending angle, and the precision of data mining of the neural network can be greatly improved.
TABLE 5-10 Effect of different data samples on springback Angle fitting error
In order to verify the effectiveness and superiority of the PSOLM hybrid particle swarm optimization, the 4-5-1 structure neural network established above is trained by using the BP algorithm and the PSOLM algorithm respectively under the same conditions. Here, layer 1 of the neural network uses a logarithmic sigmoid transfer function, layer 2 uses a linear transfer function, and the error criterion e is uniformly set to 10-4. For the BP algorithm, a momentum method improved fast learning algorithm is adopted, and the maximum iteration number is 1000. When a network is trained by using a PSOLM algorithm, the maximum number of times of L-M local depth search is set to be 20, the size of a PSO population is set to be 40, the position of each particle is clamped in the range of [ -2.5,2.5] in the searching process, and the speed of each particle is clamped in the range of [ -0.15,0.15 ]; the inertia weight is linearly reduced from 0.95 to 0.4; the constants c1 and c2 are both 1.49, and the sum of squared errors is taken as a fitness evaluation function of the particles; the maximum training algebra is 1000, and the minimum constant is 10-6.
The performance of the BP and PSOLM data mining algorithms is compared by using the corrected data of the orthogonal simulation experiment shown in the table 5-4 as a training sample and the unknown sample data shown in the table 5-7 as a test sample, and the comparison of the training performance and the generalization performance of the BP and PSOLM algorithms is given in the table 5-11. As can be seen from tables 5-11, the average training error of the PSOLM algorithm is below 1/6 for the BP algorithm; the average test error of the PSOLM algorithm reaches about 1/8 of the BP algorithm with minimum. Therefore, the PSOLM algorithm takes less CPU time, but obtains higher convergence accuracy and prediction accuracy than the BP algorithm, that is, the PSOLM algorithm has better learning ability and generalization ability than the BP algorithm. In a word, the PSOLM algorithm has better average performance (including average training performance, average generalization performance and average CPU time) than the BP algorithm, and has obvious algorithm stability compared with the BP algorithm, that is, the PSOLM algorithm has better learning ability than the BP algorithm, and can obtain faster convergence speed and higher convergence accuracy.
TABLE 5-11 comparison of training and generalization Performance for two algorithms (statistical results of 20 experiments)
In summary, the PSOLM hybrid particle swarm algorithm provided herein has not only a fast convergence speed and a high convergence accuracy, but also a strong generalization capability, fully exerts the characteristics of PSO global search and L-M local fast optimization, makes up for their respective deficiencies, and realizes advantage complementation of the algorithm. The PSOLM algorithm is used for optimizing each parameter of the neural network, the modeling process is simple, the good convergence performance is achieved, the local minimum value is avoided, the global optimal solution of the problem is effectively found, the fast optimization performance is achieved, and the efficiency and the robustness which are better than those of the BP algorithm are obtained. The experimental result verifies the effectiveness and superiority of the algorithm.
While the invention has been shown and described with reference to a preferred embodiment, it will be understood by those skilled in the art that various changes in form and detail may be made therein without departing from the scope of the appended claims.
Claims (7)
1. A metal material bending forming rebound angle prediction method comprises the following steps:
step 1: establishing a metal material bending resilience finite element model, and performing a virtual orthogonal test on the metal material bending resilience by using finite element simulation software to obtain numerical simulation data of the metal material bending resilience angle;
step 2: performing range analysis on numerical simulation data of the bending resilience angle of the metal material, and determining a main factor which has the most obvious influence on the bending resilience angle of the metal material;
and step 3: each main factor examines three levels, a field opening orthogonal test table is established, each level of the main factors is randomly selected from the virtual orthogonal test table to perform a real experiment, and experimental data of the bending resilience angle of the metal material are obtained;
and 4, step 4: dividing experimental data of the bending resilience angle of the metal material under a certain horizontal condition of the main factor by numerical simulation data to obtain a correction coefficient of the resilience angle under the horizontal condition of the main factor; calculating the ratio of experimental data and numerical simulation data of the springback angle under other same horizontal conditions of the main factor in sequence to obtain a correction coefficient of the springback angle under each horizontal condition of the main factor;
and 5: multiplying the correction coefficient of the main factor under each horizontal condition by the corresponding numerical simulation data to obtain the correction value of the virtual orthogonal test under each working condition;
step 6: establishing a prediction model of the bending resilience angle of the metal material based on the artificial neural network, training and learning the neural network by using correction data under the working condition of the virtual orthogonal test, inputting unknown working condition data of the bending resilience of the metal material, and obtaining a predicted value of the bending resilience angle of the metal material under the unknown working condition through calculation of the neural network.
2. The method for predicting the bending return angle of a metal material according to claim 1, wherein: the metal material is Q235 pipe.
3. The method for predicting the bending return angle of a metal material according to claim 1 or 2, wherein: the stress to which the metal material is subjected when bent is as follows:
wherein σ is the actual stress, σ s is the yield limit, E is the Young's modulus, which is the actual strain, K is the hardening coefficient of the material, and n is the hardening index of the material.
4. The method for predicting the bending return angle of a metal material according to claim 1 or 2, wherein: the influence factors of the bending resilience angle of the metal material comprise a bending angle, a relative bending radius, a relative wall thickness and a friction coefficient.
5. The method for predicting the bending return angle of a metal material according to claim 1 or 2, wherein: the correction value in step 5 is normalized before step 6 is performed.
6. The method for predicting the bending return angle of a metal material according to claim 5, wherein: the normalization processing formula is as follows:
wherein, XijIs the j (th) value of the i (th) input/output parameter after normalization, max (x)i)，min(xi) The upper and lower limit values of the parameter are respectively; x is the number ofNmin，xNmaxIs a normalized limit value of 0<xNmin<xNmax<1, taking xNmin＝0.1，xNmax＝0.9。
7. The method for predicting the bending return angle of a metal material according to claim 1 or 2, wherein: and 6, the artificial neural network is one or a combination of BP neural network, particle swarm optimization algorithm and mixed particle swarm optimization algorithm.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010789043.8A CN112036062A (en) | 2020-08-07 | 2020-08-07 | Metal material bending forming rebound angle prediction method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010789043.8A CN112036062A (en) | 2020-08-07 | 2020-08-07 | Metal material bending forming rebound angle prediction method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN112036062A true CN112036062A (en) | 2020-12-04 |
Family
ID=73582809
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010789043.8A Pending CN112036062A (en) | 2020-08-07 | 2020-08-07 | Metal material bending forming rebound angle prediction method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112036062A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113221398A (en) * | 2021-03-31 | 2021-08-06 | 成都飞机工业（集团）有限责任公司 | Method for predicting L-shaped composite material part curing deformation rebound angle |
| CN113221319A (en) * | 2021-03-31 | 2021-08-06 | 成都飞机工业（集团）有限责任公司 | Measurement and calculation method for C-shaped composite material part curing deformation resilience angle |
| CN117077312A (en) * | 2023-08-17 | 2023-11-17 | 捷安特(中国)有限公司 | Mold determining method, apparatus, electronic device and storage medium |
| CN117077312B (en) * | 2023-08-17 | 2024-04-26 | 捷安特(中国)有限公司 | Mold determining method, apparatus, electronic device and storage medium |
-
2020
- 2020-08-07 CN CN202010789043.8A patent/CN112036062A/en active Pending
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113221398A (en) * | 2021-03-31 | 2021-08-06 | 成都飞机工业（集团）有限责任公司 | Method for predicting L-shaped composite material part curing deformation rebound angle |
| CN113221319A (en) * | 2021-03-31 | 2021-08-06 | 成都飞机工业（集团）有限责任公司 | Measurement and calculation method for C-shaped composite material part curing deformation resilience angle |
| CN113221319B (en) * | 2021-03-31 | 2022-05-10 | 成都飞机工业（集团）有限责任公司 | Measurement and calculation method for C-shaped composite material part curing deformation resilience angle |
| CN113221398B (en) * | 2021-03-31 | 2022-08-12 | 成都飞机工业（集团）有限责任公司 | Method for predicting L-shaped composite material part curing deformation rebound angle |
| CN117077312A (en) * | 2023-08-17 | 2023-11-17 | 捷安特(中国)有限公司 | Mold determining method, apparatus, electronic device and storage medium |
| CN117077312B (en) * | 2023-08-17 | 2024-04-26 | 捷安特(中国)有限公司 | Mold determining method, apparatus, electronic device and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109508488B (en) | Shot peening forming process parameter prediction method based on genetic algorithm optimization BP neural network | |
| CN110640982A (en) | Multi-target optimization method for injection molding process parameters of thin-wall injection molding part | |
| CN105260532B (en) | Thin plate based on sequence near-optimal stretches pressure variable edge force uncertainty design method | |
| CN111401623A (en) | Shot peening strengthening surface integrity prediction method based on BP neural network | |
| CN112036062A (en) | Metal material bending forming rebound angle prediction method | |
| CN113449474B (en) | Improved gray wolf algorithm optimized BP neural network pipe forming quality prediction method | |
| CN111158237B (en) | Industrial furnace temperature multi-step prediction control method based on neural network | |
| CN110046408A (en) | Bounded-but-unknown uncertainty optimization method based on BP neural network differential and interval analysis | |
| CN109344477B (en) | 6-degree-of-freedom mechanical arm inverse kinematics solving method | |
| Gusel et al. | Application of genetic programming for modelling of material characteristics | |
| CN111310348A (en) | Material constitutive model prediction method based on PSO-LSSVM | |
| CN112231839A (en) | Intelligent exploration method for peak-tracking sampling of aircraft aiming at high time-consuming constraint | |
| Heng et al. | Towards intelligent design optimization: Progress and challenge of design optimization theories and technologies for plastic forming | |
| CN116502455A (en) | Process parameter determination method and system for laser selective melting technology | |
| CN113043278B (en) | Mechanical arm track planning method based on improved whale searching method | |
| Guo et al. | Bending angle prediction model based on BPNN-spline in air bending springback process | |
| CN111369039B (en) | Alloy heat treatment process optimization method based on multi-objective optimization assisted active learning | |
| CN109886405A (en) | It is a kind of inhibit noise based on artificial neural network structure's optimization method | |
| CN110210072B (en) | Method for solving high-dimensional optimization problem based on approximate model and differential evolution algorithm | |
| Zhang et al. | Rolling force prediction in heavy plate rolling based on uniform differential neural network | |
| CN113935556B (en) | Temperature sensor optimal arrangement method based on DNA genetic algorithm | |
| CN112241811B (en) | Layered hybrid performance prediction method for customized product in Internet +' environment | |
| CN114357878A (en) | Method for predicting water surface load of typical section of amphibious aircraft | |
| CN113434982A (en) | Inverse kinematics solution method of electric intelligent bionic climbing robot | |
| CN111275328A (en) | RNGRU (radio network unit) position marker part matching method based on comprehensive grey correlation sequence |
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 |