CN101751487A - Bending displacement by utilizing artificial neural network - Google Patents
Bending displacement by utilizing artificial neural network Download PDFInfo
- Publication number
- CN101751487A CN101751487A CN200810185972A CN200810185972A CN101751487A CN 101751487 A CN101751487 A CN 101751487A CN 200810185972 A CN200810185972 A CN 200810185972A CN 200810185972 A CN200810185972 A CN 200810185972A CN 101751487 A CN101751487 A CN 101751487A
- Authority
- CN
- China
- Prior art keywords
- neural network
- artificial neural
- bending
- network
- bending displacement
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
Abstract
The invention discloses a bending displacement by utilizing artificial neural network and relates to a method for estimating the functional dependence of manufacturing parameters and the bending displacement in the bending process by an artificial neural network method; in the method, the weight value and the threshold value of the artificial neural network are calculated in advance by genetic algorithm, the weight value and the threshold value are given to the artificial neural network, and the artificial neural network is trained; in addition, the invention relates to a bending process by utilizing heat sources such as laser, electronics or arc beams and the like and an interface which carries out the method.
Description
Technical field
The present invention relates to a kind ofly estimate the Fabrication parameter in the BENDING PROCESS and the method for the functional dependence between the bending displacement by artificial neural network.
Background technology
Producing bending by the thermal source such as laser, electronics or arc bundle or UV-lamp is a kind of process technology very flexibly.It is a kind of noncontact forming technology, and thermal source is used as and is used for forming tool that the workpiece such as plate is formed in this technology, and without any need for mould.This technology is applicable to multiple field, particularly boats and ships, automobile and aerospace industry.
Because by the BENDING PROCESS that thermal source carried out such as laser, electronics or arc bundle is very complicated Nonlinear thermal physical process, so under the infusible condition of material surface, be difficult to prediction and control bending displacement, for example bending angle.
Summary of the invention
The purpose of this invention is to provide a kind of method that is used for crooked result being realized improved controllability in crooked field.
This purpose is by realizing according to the described method of independent claims.
The protection content that more useful progress is a dependent claims.
Result in order to control and to improve in BENDING PROCESS to be obtained can utilize artificial neural network (ANN), and described artificial neural network (ANN) is based on the mathematical model or the computation model of biological neural network.Utilize artificial neural network, the complex relationship between input (so-called input neuron) and the output (so-called output neuron) can be modeled.The advantage of artificial neural network is, need not to set up the mathematical model of total system, just can realize from a space R
n(n is the number of input neuron) is to another space R
mThe mapping of (m is the number of output neuron).In artificial neural network, a plurality of neurons interconnect so that form network based on specific function or rule.The single neuron that combination has a suitable simple functions can produce the network with complex characteristics generally, this network very accurately emulation by the nonlinear BENDING PROCESS of the complexity that local heat source carried out such as laser beam.
Counterpropagation network (BP network) is a multilayer feedforward neural network, and the precision of its mapping guarantees by using the sample that the BP network is trained.In the process of training, provide the feedback of the output quality that obtains about up to the present utilizing specific input to system.In general, neural network is configured to a plurality of layer, promptly be endowed input neuron input layer, be endowed the output layer of output neuron and at above-mentioned one or more hidden layers between two-layer.Should be noted that the BP network that has a hidden layer in theory can shine upon any function.Therefore, according to one embodiment of present invention, the BP network with three layers (input layer, a hidden layer and output layers) is just enough.
About the procedure of adaptation of BP network, utilize this network to determine to exist shortcoming to optimizing the result to BENDING PROCESS.Just, the BP network uses the local search approach with squared error function.This shortcoming that has is to have the excessive risk that runs into local minimum, slow speed of convergence and grow the training time.
Genetic algorithm (GA) is substantially similar to the process of biological evolution.Genetic algorithm is used by the technology that evolution biology produced, for example heredity, variation, selection and intersection.This helps to evolve so that find the optimization solution of particular problem by the population to the abstract representation (so-called chromosome) of candidate solution, thereby searches for the approximate or exact solution of some problem.The evolutionary process of genetic algorithm is usually from the population of individuality, and begins to evolve a plurality of generations from it.In the colony in each generation, carried out coding, duplicated, overlapping and change after calculate each individual adaptability.Then, each individuality is used to determine the individuality that a group is new according to specified conditions.
Genetic algorithm utilization searching method at random.When making genetic algorithm be adapted to BENDING PROCESS, the deficiency such as precocity convergence, low local search ability and slow speed of convergence appears.
According to one embodiment of present invention, provide a kind of and estimate the Fabrication parameter in the BENDING PROCESS and the method for the functional dependence between the bending displacement by artificial neural network, described bending displacement is bending angle preferably, and this method may further comprise the steps: the weights and the threshold value that precompute artificial neural network by genetic algorithm; Give artificial neural network with weights and threshold value; And training of human artificial neural networks.The basis of this method is an artificial neural network, but passes through the combination of BP network and genetic algorithm, and this algorithm is modified.This method can be avoided the defective Training Capability of the slow speed of convergence and the genetic algorithm of BP network, and can guarantee in BENDING PROCESS stability and the precision imported and/or optimum prediction is carried out in output.Therefore, the advantage of BP network and genetic algorithm is combined, thereby produces a kind of improved BP network.As a result, can carry out BENDING PROCESS with the Fabrication parameter of optimizing, perhaps can predict bending displacement with specific Fabrication parameter for the bending displacement of expecting.For example, this method makes it possible to estimate or determines necessary Fabrication parameter (for example laser beam output power, sweep velocity, spot diameter and plate thickness) so that obtain the bending displacement of expectation, for example bending angle.Function below this method has also realized promptly when the bending displacement of expectation and plate thickness are known, is determined remaining Fabrication parameter, for example laser output power, sweep velocity, spot diameter based on this method.Also can estimate above-mentioned value conversely, promptly predict bending displacement based on known Fabrication parameter.
In addition, the invention provides a kind of bending method that makes the workpiece bending by the thermal source such as laser, electronics or arc bundle.Optimization to the prediction of Fabrication parameter and/or bending displacement can promote the application of bending techniques in actual production.
And, the invention provides a kind of interface that is used for determining Fabrication parameter and/or bending displacement in BENDING PROCESS.This provides a kind of close friend and man-machine dialog interface easy to use.In BENDING PROCESS, Fabrication parameter and/prediction of bending displacement is optimised and deformation displacement is predicted.Prediction " Fabrication parameter and/or bending displacement " means that several embodiments are possible.For example possible is that the bending displacement of expectation is known, and the necessary Fabrication parameter of this interface output, for example laser output power, sweep velocity, spot diameter and plate thickness.In another example, the bending displacement of expectation and plate thickness are known, and remaining Fabrication parameter is exported at this interface simultaneously, for example laser output power, sweep velocity, spot diameter.Possible in a further example is that the Fabrication parameter such as laser output power, sweep velocity, spot diameter and plate thickness is known, and the bending displacement that will utilize these Fabrication parameters to obtain is exported at this interface.
Description of drawings
Fig. 1 illustrates the overview flow chart of method according to an embodiment of the invention;
Fig. 2 is the process flow diagram of a part that illustrates in greater detail the process flow diagram of Fig. 1;
Fig. 3 illustrates method according to an embodiment of the invention and according to the synoptic diagram of the difference of the speed of convergence of the method for BP network;
Fig. 4 illustrates the synoptic diagram that the output data of artificial neural network and sample data are compared;
Fig. 5 illustrates sample data and by the comparison of output data of the neural network prediction of training; And
Fig. 6 illustrates the image of user interface according to an embodiment of the invention.
Embodiment
Further explain by artificial neural network to Fig. 6 below with reference to Fig. 1 of accompanying drawing and to estimate the manufacturing in the BENDING PROCESS or the method for the functional dependence between technical parameter and the bending angle.
Fig. 1 illustrates the overview flow chart of method according to an embodiment of the invention.
At first, in steps A, gather the experiment of sample data.Sample data can be divided into two parts: importation and output.The importation comprises the plate thickness of laser output power, sweep velocity, spot diameter and workpiece in this embodiment, and output comprises the bending angle of the plate that the corresponding input of utilization is obtained simultaneously.All these data all obtain from experiment.
It is neodymium-doped yttrium-aluminum garnet (Nd:YAG) the laser instrument emitted light beams of 3500W that laser beam is selected for use by wavelength X=1064nm, peak power.Plate is that length is that 150mm, width are that 100mm, thickness are the aluminium alloy AA6056 of 2.5mm.As cantilever slab, an end of this plate is fixed.Then, laser output power, sweep velocity and spot diameter correspondingly are provided with.Afterwards, laser beam begins to shine continuously along center line on width of the workpiece direction.Subsequently, measure bending angle.In this way, obtain 30 groups of sample datas, wherein one group of data by Fabrication parameter (plate thickness that comprises laser output power, laser scanning speed, laser spot diameter and workpiece) and corresponding bending angle form right.
After this in step B, sample data is set up and by normalization, so that be used as the data of training of human artificial neural networks.The proper range and the quantity that this means sample data are set up.30 groups of sample datas that obtain in steps A are set up and are normalized in [1,1] scope, so that set up and the input value of artificial neural network and the corresponding relation of output valve, promptly make the scope of sample data and input value and output valve be in same grade.Use computing environment and programming language in this embodiment
Selected the premnmx function so as
Middle implementation step B.The premnmx function is existing software function, and it comes the pre-service sample data by the normalization input and output, so that they are in the interval [1,1].
In step C subsequently, artificial neural network is created as three layers of BP network with an input layer, a hidden layer and an output layer.The number of the node of input layer is set to 4, and hidden layer is set to comprise 15 nodes, and output layer is set to comprise a node.In addition, select the tansig function as the neuron transport function between middle layer and the output layer.The tansig function is a tanh S shape transport function, and it is used for being come by its network input the output of computation layer.This tansig function also is
In existing function.
After this in step D, by the genetic algorithm that illustrates in greater detail in the left hurdle that uses Fig. 2, artificial neural network is by training in advance or optimised.Therefore, the step S100 among Fig. 2 is performed in the step D of Fig. 1 to S107.In other words, utilize genetic algorithm to calculate in advance to be used to limit the weights and the threshold value of artificial neural network.In this genetic algorithm, by the population size be set to 50, the number in heredity generation is set to 100, crossover probability be set to 0.1 and the probability that makes a variation be set to 0.05 and in step S100, create initial population.Therefore, population comprises one group of weights and threshold value.Then, in step S101, determine the fitness of population, and whether the verification population meets the desired requirement (step S102).If meet the desired requirement, the process flow diagram of Fig. 2 then proceeds to step S107, this step S107 finishes genetic algorithm and the result's (weights and threshold value of promptly being used for artificial neural network) who is obtained is forwarded to step S108, and this step S108 is performed in the step e of the Fig. 1 that describes subsequently.If the result in the step 102 negates, the process flow diagram of Fig. 2 then carries out step S103, S104, S105 and S106 so that create new population, and described step S103, S104, S105 and S106 are the known functions of genetic algorithm.After having created new population, this process turns back to step S101.In this embodiment, come training in advance artificial neural network (promptly calculating weights and the threshold value that is used for ANN in advance) by using the GAOT function, thereby implement described step D.GAOT (GAOT is " genetic algorithm optimization tool box ") function is MATLAB
The tool box.At last, in step D, obtain to be used for the weights and the threshold value of artificial neural network, optimize described weights and threshold value by using genetic algorithm (being the GAOT function in this embodiment).
Then, in step e, weights that will obtain in step D and threshold value are given the BP that creates network in step C.For this purpose, the weights of acquisition and threshold matrix are decoded and be endowed the BP network.The step e of Fig. 1 is corresponding to the step S108 of Fig. 2.
BP network in step F subsequently.The number of times of training is set to 1000, and the target square error is 0.0001.When the number of times of training surpasses 1000 or when square error reaches 0~0.00001, training process will stop.The step F of Fig. 1 corresponding to the step S109 of Fig. 2 to S113, in the step S109 of Fig. 2 repetition training in the S113, up in step S113, meeting the desired square error.For the training of human artificial neural networks, select known Levenberg-Marquardt algorithm.
Fig. 3 is explanation according to the method for embodiment of the present invention with according to the synoptic diagram of the difference of the speed of convergence of the method for BP network.In the Fig. 3 that draws, horizontal ordinate is represented the epoch (epoch), and ordinate is represented square error.In Fig. 3, can see, for the present embodiment of weights that wherein use genetic algorithm to calculate in advance to be used for the BP network and threshold value, after 10 epoch, just reached expectation target (solid line), be that square error is 0.0001 straight line, yet need 28 epoch reach expectation target (dotted line) when not using genetic algorithm when only using the BP network.The convergence map of Fig. 3 is clearly shown that the method according to present embodiment has improved speed of convergence effectively.
The synoptic diagram that Fig. 4 illustrates the output data of artificial neural network (obtain network) that will training execution in step F after and the sample data that obtains in steps A compares.Can see that these two curves are almost consistent.
Refer back to Fig. 1, verification artificial neural network in step G.For this purpose, obtain new sample data by as steps A is described, experimentizing.The sample data that newly obtains comprises all possible sample situation, so that the performance of verification ANN fully.Generally, determine 14 groups of sample datas, and will utilize their output of the neural network prediction of being trained that after carrying out step F, exists.The step F of Fig. 1 is corresponding to the step S114 of Fig. 2.
Fig. 5 has illustrated the sample data that newly obtains and comparison by the output data (being bending angle) of the neural network prediction of training.Can see curve much at one.Relation or correlativity between this proof input value (Fabrication parameter such as laser output power, sweep velocity, spot diameter and plate thickness) and the output valve (bending angle) are correctly calculated.Method according to present embodiment can be similar to the output that calculates to a nicety, and the stability of this method and feasibility are proved to be.
At last, in step H, by inciting somebody to action
Combine with the programmed environment of visual C++ (VC++) and to write optimizer.At first, above-mentioned method conduct
In code implement.Then, comtool order be used to open the comtool dialogue (
In function).Code is added in the project, compiles then and run time version, and create their corresponding C ++ source file and C++ header file.Secondly, in VC++, create the MFC project, and C++ source and header file are added in the project that the first step is created.Design and develop as shown in Figure 6 friendly and man-machine dialog interface easily then.Utilize this interface, can emulation laser bending process and predict crooked result, as step S115 described of Fig. 2.
Network training part at the interface can be imported hidden layer and target error by corresponding button.Can also import sample data (comprising importation and output) by pressing corresponding button.Train and depositary's artificial neural networks by pressing corresponding other button then.Also there is the button that is used for the start-up parameter prediction.When input laser output power, sweep velocity, spot diameter and plate thickness, these level meters are calculated the bending angle that can directly be read.
Only describe an embodiment above in detail, and do not plan the present invention is limited to this embodiment.
Claims (7)
1. one kind is used for estimating the Fabrication parameter of BENDING PROCESS and the method for the functional dependence between the bending displacement by artificial neural network, and this method may further comprise the steps:
Calculate the weights and the threshold value of described artificial neural network in advance by genetic algorithm;
Give described artificial neural network with described weights and threshold value; And
Train described artificial neural network.
2. method according to claim 1, before the step of calculating described weights and threshold value in advance, this method is further comprising the steps of:
Gather the sample data of described functional dependence by experimentizing, wherein said sample data comprise a plurality of by Fabrication parameter and corresponding bending displacement form right; And
The described sample data of normalization.
3. method according to claim 2, further comprising the steps of:
Come the described artificial neural network of verification by using the more sample data.
4. according in the described method of one of preceding claim, wherein said artificial neural network is three layers of counterpropagation network that comprise an input layer, a hidden layer and an output layer.
5. according in the described method of one of preceding claim, wherein said Fabrication parameter comprises at least one in power of heat source, thermal source sweep velocity, thermal source size and the work plate thickness.
6. bending method that is used for making by thermal source the workpiece bending is wherein according to determining Fabrication parameter and/or bending displacement in one of preceding claim.
7. one kind is used for determining the Fabrication parameter of BENDING PROCESS and/or the interface of bending displacement, and wherein said interface is suitable for carrying out according to the described method of one of claim 1 to 5.
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN200810185972A CN101751487A (en) | 2008-12-18 | 2008-12-18 | Bending displacement by utilizing artificial neural network |
PCT/CN2009/001273 WO2010069121A1 (en) | 2008-12-18 | 2009-11-17 | Bending displacement with utilization of an artificial neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN200810185972A CN101751487A (en) | 2008-12-18 | 2008-12-18 | Bending displacement by utilizing artificial neural network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN101751487A true CN101751487A (en) | 2010-06-23 |
Family
ID=42268265
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN200810185972A Pending CN101751487A (en) | 2008-12-18 | 2008-12-18 | Bending displacement by utilizing artificial neural network |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN101751487A (en) |
WO (1) | WO2010069121A1 (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103381436A (en) * | 2012-05-04 | 2013-11-06 | 萨尔瓦尼尼意大利股份公司 | Apparatus and method for measuring the bending angle of a sheet |
CN108828885A (en) * | 2018-05-03 | 2018-11-16 | 合刃科技(深圳)有限公司 | Light source module group and optical projection system |
CN109359355A (en) * | 2018-09-05 | 2019-02-19 | 重庆创速工业有限公司 | A kind of design implementation method of normal structure module |
CN112632810A (en) * | 2020-11-30 | 2021-04-09 | 江苏科技大学 | Method for predicting pressing amount rule of upper die for bending rod piece |
CN112912884A (en) * | 2018-10-30 | 2021-06-04 | 昭和电工株式会社 | Material designing device, material designing method, and material designing program |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10217237B1 (en) | 2018-06-21 | 2019-02-26 | 3D Med Ag | Systems and methods for forming a desired bend angle in an orthodontic appliance |
Family Cites Families (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP5085175B2 (en) * | 2007-03-30 | 2012-11-28 | 公益財団法人鉄道総合技術研究所 | Method for estimating dynamic characteristics of suspension system for railway vehicles |
CN101320400B (en) * | 2008-07-16 | 2010-04-21 | 桂林电子科技大学 | Optimization design method of micro-electron packaging device based on artificial neural network |
-
2008
- 2008-12-18 CN CN200810185972A patent/CN101751487A/en active Pending
-
2009
- 2009-11-17 WO PCT/CN2009/001273 patent/WO2010069121A1/en active Application Filing
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103381436A (en) * | 2012-05-04 | 2013-11-06 | 萨尔瓦尼尼意大利股份公司 | Apparatus and method for measuring the bending angle of a sheet |
CN103381436B (en) * | 2012-05-04 | 2017-03-01 | 萨尔瓦尼尼意大利股份公司 | Equipment for the angle of bend of sheet material measurement and method |
CN108828885A (en) * | 2018-05-03 | 2018-11-16 | 合刃科技(深圳)有限公司 | Light source module group and optical projection system |
CN109359355A (en) * | 2018-09-05 | 2019-02-19 | 重庆创速工业有限公司 | A kind of design implementation method of normal structure module |
CN112912884A (en) * | 2018-10-30 | 2021-06-04 | 昭和电工株式会社 | Material designing device, material designing method, and material designing program |
CN112912884B (en) * | 2018-10-30 | 2023-11-21 | 株式会社力森诺科 | Material designing apparatus, material designing method, and material designing program |
CN112632810A (en) * | 2020-11-30 | 2021-04-09 | 江苏科技大学 | Method for predicting pressing amount rule of upper die for bending rod piece |
Also Published As
Publication number | Publication date |
---|---|
WO2010069121A1 (en) | 2010-06-24 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN101751487A (en) | Bending displacement by utilizing artificial neural network | |
Chen et al. | Automatic design of hierarchical Takagi–Sugeno type fuzzy systems using evolutionary algorithms | |
Fu et al. | Springback prediction of high-strength sheet metal under air bending forming and tool design based on GA–BPNN | |
CN113569352A (en) | Additive manufacturing size prediction and process optimization method and system based on machine learning | |
CN103971162A (en) | Method for improving BP (back propagation) neutral network and based on genetic algorithm | |
CN109961186A (en) | Desulphurization system operating parameter prediction technique based on decision tree and BP neural network | |
Vagheesan et al. | Hybrid neural network–particle swarm optimization algorithm and neural network–genetic algorithm for the optimization of quality characteristics during CO2 laser cutting of aluminium alloy | |
Nukman et al. | Optimization of prediction error in CO2 laser cutting process by Taguchi artificial neural network hybrid with genetic algorithm | |
CN110232461A (en) | More interconnection vector machine water quality prediction methods based on quantum genetic algorithm optimization | |
CN102129259A (en) | Neural network proportion integration (PI)-based intelligent temperature control system and method for sand dust environment test wind tunnel | |
Cus et al. | Hybrid ANFIS-ants system based optimisation of turning parameters | |
CN113156900B (en) | Machining deformation control method based on meta reinforcement learning | |
CN113505519A (en) | Laser cutting machining modeling and parameter optimization selection method based on PSO-BP neural network and NSGA-II | |
Rohman et al. | Prediction and optimization of geometrical quality for pulsed laser cutting of non-oriented electrical steel sheet | |
Madic et al. | Application of cuckoo search algorithm for surface roughness optimization in co2 laser cutting | |
CN110287638A (en) | Winged saw life-span prediction method based on kalman-RNN neural network | |
CN104732067A (en) | Industrial process modeling forecasting method oriented at flow object | |
Goli et al. | An integration of neural network and shuffled frog-leaping algorithm for CNC machining monitoring | |
CN105389442A (en) | Reverse design method for coupling genetic algorithm, neural network and numerical simulation | |
CN109034478B (en) | High-precision prediction method for high-wind iterative competition along high-speed railway | |
Singh et al. | Modelling of direct metal laser sintering of EOS DM20 bronze using neural networks and genetic algorithms | |
Yan et al. | Time series forecasting with RBF neural network | |
Madić et al. | Optimization of surface roughness when turning polyamide using ANN-IHSA approach | |
Li et al. | Modeling and applying of RBF neural network based on fuzzy clustering and pseudo-inverse method | |
Zhang et al. | Milling force prediction of titanium alloy based on support vector machine and ant colony optimization |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
C02 | Deemed withdrawal of patent application after publication (patent law 2001) | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20100623 |