CN105574300A - Optimum design method for steel rail weld seam finish-milling machine tool beam body based on BP neural network and genetic algorithm - Google Patents

Abstract

The invention discloses an optimum design method for a steel rail weld seam finish-milling machine tool beam body based on a BP neural network and a genetic algorithm. The optimum design method includes the steps that firstly, the size of an auxiliary structure for supporting or reinforcing a whole in the beam body is selected as a design variable, and the optimization criterion of the beam body is to improve rigidity and reduce total weight on the premise of ensuring structural strength; secondly, the strength, rigidity and weight of the beam body are obtained by adopting an orthogonal test method to serve as sample data; thirdly, the neural network is designed and is trained with the sample data until the difference between a predicted value and a sample value is defined within an allowance error range; fourthly, a population is generated, and population fitness and constraint condition values are calculated by using the neural network so that genetic algorithm optimization solution can be conducted; fifthly, optimum obtained parameters are analyzed in a simulation mode to determine optimization result feasibility. On the premise of ensuring the structural strength, the structural rigidity is effectively improved, the structural weight is effectively reduced, and therefore the overall structural performance of a finish-milling machine tool is promoted.

Description

Based on the steel rail welding line finish-milling lathe cross girder Optimization Design of BP artificial neural network and genetic algorithms

Technical field

The present invention is specifically related to a kind of steel rail welding line finish-milling lathe cross girder Optimization Design based on BP artificial neural network and genetic algorithms.

Background technology

Steel rail welding line finish-milling numerically-controlled machine be a kind ofly integrate mechanical, electrical, liquid, observing and controlling supermatic new exclusive equipment, be mainly used in the shaping processing of gapless line postwelding working edge, the scope of operation.Cross girder as the important support parts of this lathe, its dimensioned precision and stable and reliable operation directly affecting lathe whether rational in infrastructure.Cross girder planform is complicated, mainly comprises main beam and secondary crossbeam, and each several part size arranges to have material impact for the intensity of entirety, rigidity and weight.How preferred arrangement each several part size is the major issue needing in cross girder design process to solve.At present, the optimization of cross girder is mainly based on Experience Design, and it often needs to be optimized by trial and error repeatedly, there is design cycle long, the unconspicuous shortcoming of effect of optimization.

Summary of the invention

The object of the present invention is to provide a kind of steel rail welding line finish-milling lathe cross girder Optimization Design based on BP artificial neural network and genetic algorithms, it is based on BP artificial neural network and genetic algorithms, complex optimum cross girder each several part dimensional structure, can under the prerequisite ensureing its structural strength, effective raising rigidity of structure and alleviate construction weight, thus promote Finish Milling Machine bed ensemble structural behaviour.

The technical solution adopted for the present invention to solve the technical problems is:

Based on a steel rail welding line finish-milling lathe cross girder Optimization Design for BP artificial neural network and genetic algorithms, comprise the following steps:

S1, determine design variable and optimization aim

The determination of S101, design variable: choose in cross girder for supporting or the supplementary structure size P of reinforcement unitarity ₁, P ₂..., P _nas design variable;

The determination of S102, optimization aim: the Optimality Criteria of cross girder is improve rigidity under the prerequisite of guarantee structural strength and alleviate general assembly (TW);

The foundation of S103, optimized mathematical model: maximum deformation quantity and the general assembly (TW) of choosing cross girder are objective function, and maximum stress is as constraint condition, and optimized mathematical model is

$\left\{\begin{array}{c}X={\[P_1,\mathrm{...},P_n\]}^T\\ \min F(e_{\max },m)\\ {\mathrm{\σ}}_{\max }\≤\[\mathrm{\σ}\]\\ X_{\min }\≤X\≤X_{\max }\end{array}\right.,$

Wherein, X is design variable, F (e _max, m) be objective function, F (e _max, m)=λ ₁f (e _max)+λ ₂g (m), e _maxfor maximum deformation quantity, [e], for allowing maximum deformation quantity, m is general assembly (TW), and [m] is maximum weight allowable, λ ₁, λ ₂for optimizing weight coefficient, λ ₁+ λ ₂=1, σ _maxfor maximum stress, [σ] is permissible stress, X _max, X _minfor the upper and lower limit of design variable;

S2, acquisition training sample

S201, the method for orthogonal test is taked to carry out sample acquisition, selected design variable is the factor of orthogonal test table, in the span of each design variable, choose several levels, design orthogonal test table, determine the design parameter of group number and each test group tested;

S202, enforcement orthogonal test scheme, corresponding cross girder model is set up according to the parameter of each test group, recycling finite element software carries out strength and distortion analyses to all cross girder models respectively, extract cross girder intensity, rigidity and weight result that simulation obtains, using maximum stress as the token state of intensity, using maximum deformation quantity as the token state of rigidity, using the training sample of these results as neural network;

S3, structure BP neural network

S301, by design variable P ₁, P ₂..., P _nas input layer, by m, e _max, σ _maxas output layer, middle layer adopts single hidden layer, neuron number wherein, n _ifor input layer number, n _ofor output layer nodes, q is constant, q ∈ [1,10];

S302, the training sample obtained in step S202 is utilized to train BP neural network, until the difference of predicted value and sample value is limited within the scope of permissible error;

S4, to solve with genetic algorithm optimization

S401, generation initial population;

S402, use BP neural computing fitness and constraint conditional value, constraint condition function is designed to

$Q\left(X\right)=F(e_{\max },m)+k\left(\max {\[0,\frac{{\mathrm{\σ}}_{\max }}{\[\mathrm{\σ}\]}-1\]}^2\right),(k\≥0),$

Fitness function is designed to

$Fit\left(X\right)=\frac{1}{1+c+Q\left(X\right)},(c\≥0,c+Q(X)\≥0);$

If S403 meets Optimality Criteria and constraint condition with regard to Output rusults, otherwise select the individuality that fitness is high, perform genetic manipulation and generate new population, then turn to step S402;

S5, determine Optimal Parameters: the parameter combinations that can obtain one group of effect optimum according to the optimum results of genetic algorithm, modeling and simulating is carried out to the cross girder under this group parameter and solves analysis, finally determine optimum results feasibility, if there is larger difference, the structure and the genetic algorithm optimization that re-start neural network solve.

By technique scheme, in step S101, choose the position dimension P of secondary crossbeam vertical edges thickness in cross girder ₁, secondary crossbeam transverse edge thickness position dimension P ₂, four rectangular dimension P in secondary crossbeam ₃× P ₄, secondary beam vertical is in the thickness P of drawing surface direction ₅and the position dimension P of main beam transverse edge thickness ₆as design variable.

By technique scheme, described P ₁span be 2010 ~ 2050mm, described P ₂span be 115 ~ 135mm, described P ₃span be 800 ~ 880mm, described P ₄span be 90 ~ 110mm, described P ₅span be 80 ~ 100mm, described P ₆span be 50 ~ 70mm.

By technique scheme, in step s 201, in the span of each design variable, choose 5 levels, design the orthogonal test table of six factor five levels, determine that the group number tested is 25 groups.

By technique scheme, described P ₁five levels be respectively 2010,2020,2030,2040,2050, described P ₂five levels be respectively 115,120,125,130,135, described P ₃five levels be respectively 880,860,840,820,800, described P ₄five levels be respectively 90,95,100,105,110, described P ₅five levels be respectively 80,85,90,95,100, described P ₆five levels be respectively 50,55,60,65,70.

By technique scheme, in step s 103, λ ₁=0.5, λ ₂=0.5.

By technique scheme, in step S301, ni=6, no=3, q=9, neuron number

By technique scheme, allow maximum deformation quantity [e]=0.5mm, maximum weight allowable [m]=1000kg, permissible stress [σ]=130MPa.

The beneficial effect that the present invention produces is: for Novel steel rail weld seam finish-milling numerically-controlled machine, for improving Machine Tool design level and strengthening product competitiveness in the market, being optimized design to device structure is inexorable trend, the present invention choose in cross girder for support or the supplementary structure size of reinforcement unitarity as design variable, set up optimized mathematical model, carry out multiple goal list constrained optimization, CAD parametric modeling and CAE finite element simulation is adopted to combine again, the method of orthogonal test obtains training sample, BP neural network containing single hidden layer is trained, The present invention gives the defining method of neural network neuron number, and with the addition of last Optimal Parameters check, the optimization of genetic algorithm is carried out again after training BP neural network, with BP neural network constraint IF condition and the fitness determined in genetic algorithm, genetic algorithm searches for optimum solution in gamut.BP artificial neural network and genetic algorithms combines by the present invention, fast and effeciently can be optimized design to cross girder, under the prerequisite that can require in proof strength, effectively improves the rigidity of structure and alleviates construction weight, is conducive to improving Finish Milling Machine bed ensemble performance.

Accompanying drawing explanation

Below in conjunction with drawings and Examples, the invention will be further described, in accompanying drawing:

Fig. 1 is the structural representation of steel rail welding line finish-milling numerically-controlled machine in the embodiment of the present invention;

Fig. 2 is the secondary crossbeam of embodiment of the present invention middle cross beam body and the structural representation of main beam;

Fig. 3 is the process flow diagram of the embodiment of the present invention;

Fig. 4 is the structural representation of BP neural network in the embodiment of the present invention.

In figure: 1-lathe base; 2-slew gear; 3-cross girder; 4-cutter; The secondary crossbeam of 5-; 6-main beam.

Embodiment

In order to make object of the present invention, technical scheme and advantage clearly understand, below in conjunction with drawings and Examples, the present invention is further elaborated.Should be appreciated that specific embodiment described herein only in order to explain the present invention, be not intended to limit the present invention.

As shown in Figure 3, Figure 4, a kind of steel rail welding line finish-milling lathe cross girder Optimization Design based on BP artificial neural network and genetic algorithms, comprises the following steps:

S1, determine design variable and optimization aim

The determination of S101, design variable: choose in cross girder for supporting or the supplementary structure size P of reinforcement unitarity ₁, P ₂..., P _nas design variable;

The determination of S102, optimization aim: the Optimality Criteria of cross girder is improve rigidity under the prerequisite of guarantee structural strength and alleviate general assembly (TW);

The foundation of S103, optimized mathematical model: maximum deformation quantity and the general assembly (TW) of choosing cross girder are objective function, and maximum stress is as constraint condition, and optimized mathematical model is

$\left\{\begin{array}{c}X={\[P_1,\mathrm{...},P_n\]}^T\\ \min F(e_{\max },m)\\ {\mathrm{\σ}}_{\max }\≤\[\mathrm{\σ}\]\\ X_{\min }\≤X\≤X_{\max }\end{array}\right.,$

Wherein, X is design variable, F (e _max, m) be objective function, F (e _max, m)=λ ₁f (e _max)+λ ₂g (m), e _maxfor maximum deformation quantity, [e], for allowing maximum deformation quantity, m is general assembly (TW), and [m] is maximum weight allowable, λ ₁, λ ₂for optimizing weight coefficient, λ ₁+ λ ₂=1, σ _maxfor maximum stress, [σ] is permissible stress, X _max, X _minfor the upper and lower limit of design variable;

S2, acquisition training sample

S201, the method for orthogonal test is taked to carry out sample acquisition, selected design variable is the factor of orthogonal test table, in the span of each design variable, choose several levels, design orthogonal test table, determine the design parameter of group number and each test group tested;

S202, enforcement orthogonal test scheme, corresponding cross girder model is set up according to the parameter of each test group, recycling finite element software carries out strength and distortion analyses to all cross girder models respectively, extract cross girder intensity, rigidity and weight result that simulation obtains, using maximum stress as the token state of intensity, using maximum deformation quantity as the token state of rigidity, using the training sample of these results as neural network;

S3, structure BP neural network

S301, by design variable P ₁, P ₂..., P _nas input layer, by m, e _max, σ _maxas output layer, middle layer adopts single hidden layer, neuron number wherein, n _ifor input layer number, n _ofor output layer nodes, q is constant, q ∈ [1,10];

S302, the training sample obtained in step S202 is utilized to train BP neural network, until the difference of predicted value and sample value is limited within the scope of permissible error;

S4, to solve with genetic algorithm optimization

S401, generation initial population;

S402, use BP neural computing fitness and constraint conditional value, constraint condition function is designed to

$Q\left(X\right)=F(e_{\max },m)+k\left(\max {\[0,\frac{{\mathrm{\σ}}_{\max }}{\[\mathrm{\σ}\]}-1\]}^2\right),(k\≥0),$

Fitness function is designed to

$Fit\left(X\right)=\frac{1}{1+c+Q\left(X\right)},(c\≥0,c+Q(X)\≥0);$

If S403 meets Optimality Criteria and constraint condition with regard to Output rusults, otherwise select the individuality that fitness is high, perform genetic manipulation and generate new population, then turn to step S402;

S5, determine Optimal Parameters: the parameter combinations that can obtain one group of effect optimum according to the optimum results of genetic algorithm, modeling and simulating is carried out to the cross girder under this group parameter and solves analysis, finally determine optimum results feasibility, if there is larger difference, the structure and the genetic algorithm optimization that re-start neural network solve.

In a preferred embodiment of the invention, as shown in Figure 1, steel rail welding line finish-milling lathe comprises lathe base 1, slew gear 2, cross girder 3 and cutter 4.As shown in Figure 2, cross girder comprises secondary crossbeam 5 and main beam 6, and the position dimension of secondary crossbeam 5 vertical edges thickness is decided to be P ₁, the position dimension of the transverse edge thickness of secondary crossbeam 5 and main beam 6 is decided to be P respectively ₂and P ₆, secondary, main beam body directly in the thickness of drawing surface direction for not to be defined as P ₅and P ₉, in secondary crossbeam 5, four rectangular dimension are defined as P ₃× P ₄, P ₈with P ₂association, ensures that the cross reinforcing thickness in the middle of secondary crossbeam is constant, P ₇with P ₆association, ensures that the cross reinforcing thickness in the middle of main beam is constant.Class elliptical aperture positions all in cross girder and size all can not change, main beam 6 middle left side rectangular configuration is for installing cross motor, hollow position, right side is used for the movement of the link of transverse ball lead screw and horizontal mobile device, these are fixed sturcture for structure that is fixing and that install each parts, size does not generally make an amendment in the design, other supports or the supplementary structure of reinforcement unitarity is varistructure, in process of optimization, mainly this part physical dimension is modified, therefore, these physical dimensions can be chosen targetedly as design variable.Therefore preferred, in step S101, choose the position dimension P of secondary crossbeam 5 vertical edges thickness in cross girder ₁, secondary crossbeam 5 transverse edge thickness position dimension P ₂, four rectangular dimension P in secondary crossbeam 5 ₃× P ₄, secondary crossbeam 5 is perpendicular to the thickness P of drawing surface direction ₅and the position dimension P of main beam 6 transverse edge thickness ₆as design variable.Wherein, P ₁span be 2010 ~ 2050mm, P ₂span be 115 ~ 135mm, P ₃span be 800 ~ 880mm, P ₄span be 90 ~ 110mm, P ₅span be 80 ~ 100mm, P ₆span be 50 ~ 70mm.

In a preferred embodiment of the invention, in step s 201, in the span of each design variable, choose 5 levels, design the orthogonal test table of six factor five levels, determine that the group number tested is 25 groups.Wherein, P ₁five levels be respectively 2010,2020,2030,2040,2050, P ₂five levels be respectively 115,120,125,130,135, P ₃five levels be respectively 880,860,840,820,800, P ₄five levels be respectively 90,95,100,105,110, P ₅five levels be respectively 80,85,90,95,100, P ₆five levels be respectively 50,55,60,65,70.

In a preferred embodiment of the invention, in step s 103, λ ₁=0.5, λ ₂=0.5.

In a preferred embodiment of the invention, in step S301, n _i=6, n _o=3, q=9, neuron number $n_1=\sqrt{n_i+n_o}+q=12.$

In a preferred embodiment of the invention, maximum deformation quantity [e]=0.5mm, maximum weight allowable [m]=1000kg, permissible stress [σ]=130MPa is allowed.

The present invention, when embody rule, comprises the following steps:

The determination of S1, design variable and optimization aim

S101, in optimal design, choose P ₁, P ₂, P ₃, P ₄, P ₅, P ₆for design variable;

S102, according to work request for utilization, under the prerequisite of structural strength is mainly ensured for the optimal design of cross girder, improve rigidity as far as possible, and reduce general assembly (TW), namely maximum deformation quantity and the general assembly (TW) of choosing cross girder are objective function, and using maximum stress as constraint condition, therefore can be expressed as the mathematical model optimized

Decision variable: X=[P ₁, P ₂, P ₃, P ₄, P ₅, P ₆] ^t,

Objective function: F (e _max, m)=λ ₁f (e _max)+λ ₂g (m), gets minimum value,

Constraint condition: σ _max≤ [σ], X _min≤ X≤X _max, λ ₁+ λ ₂=1,

Wherein, e _maxfor maximum deformation quantity, m is general assembly (TW), λ ₁, λ ₂for optimizing weight coefficient, getting 0.5,0.5 respectively, utilizing normalized function eliminate unit to target e _max, m impact, allow maximum deformation quantity [e]=0.5mm, maximum weight allowable [m]=1000kg, σ _maxfor maximum stress, [σ] is permissible stress (cross girder QT500, permissible stress [σ]=130MPa), X _min, X _maxfor the upper and lower limit of decision variable;

S2, sample acquisition based on CAD/CAE modeling and simulating result

S201, be 6 factors according to design variable number determination orthogonal arrage factor number, P ₁value in (2010 ~ 2050) mm, P ₂value in (115 ~ 135) mm, P ₃value in (800 ~ 880) mm, P ₄value in (90 ~ 110) mm, P ₅value in (80 ~ 100) mm, P ₆value in (50 ~ 70) mm, each factor gets 5 levels, carries out 5 level 6 factorial experiments, selects L25 (5 ⁶) orthogonal arrage, amount to 25 groups of design variable experiment groups;

These 25 groups of test parameterss are carried out the parametric modeling of cross girder, then carry out strength and distortion analyses in AnsysWorkbench, obtain maximum stress σ by parameter that S202, basis determine respectively in Solidworks _max, maximum deformation quantity e _max, general assembly (TW) m is as the training sample of neural network;

S3, BP neural network builds

S301, by design variable P ₁, P ₂, P ₃, P ₄, P ₅, P ₆as input layer, using objective function as output layer, namely export m, e _max, σ _max, to the prediction of these values, middle layer adopts single hidden layer, neuron number n ₁select 12,

S302, the sample utilizing CAD/CAE to calculate acquisition are trained BP neural network, until the difference of predicted value and sample value is limited within the scope of permissible error;

S4, genetic algorithm optimization solve

Can carry out the optimization of genetic algorithm after training BP neural network, genetic algorithm optimization solution procedure is mainly as follows: a) produce initial population; B) with BP neural computing fitness and constraint conditional value; If c) meet Optimality Criteria and constraint condition with regard to Output rusults, otherwise select the individuality that fitness is high, perform genetic manipulation and generate new population, then turn to b;

Fitness function adopts lower bound structured approach structure:

$Fit\left(X\right)=\frac{1}{1+c+Q\left(X\right)},(c\≥0,c+Q(X)\≥0),$

Wherein, constraint condition function $Q\left(X\right)=F(e_{\max },m)+k\left(\max {\[0,\frac{{\mathrm{\σ}}_{\max }}{\[\mathrm{\σ}\]}-1\]}^2\right),(k\≥0),$

S5, Optimal Parameters are determined

Solve through above-mentioned genetic algorithm optimization, the parameter combinations of one group of effect optimum can be obtained, i.e. P ₁=2018, P ₂=136, P ₃=882, P ₄=112, P ₅=92, P ₆=68, under this condition, cross girder maximum stress σ _max=28MPa< [σ], the maximum deformation quantity e of Y-direction _max=0.365mm, general assembly (TW) m=665.2kg, compared with original cross girder structure, rigidity and weight obtain very large improvement, and solve through CAD/CAE modeling and simulating and analyze contrast, the two goodness of fit is higher, therefore can be used as the feasible prioritization scheme of this structure.

Should be understood that, for those of ordinary skills, can be improved according to the above description or convert, and all these improve and convert the protection domain that all should belong to claims of the present invention.

Claims (8)

1., based on a steel rail welding line finish-milling lathe cross girder Optimization Design for BP artificial neural network and genetic algorithms, it is characterized in that, comprise the following steps:

S1, determine design variable and optimization aim

The determination of S101, design variable: choose in cross girder for supporting or the supplementary structure size P of reinforcement unitarity ₁, P ₂..., P _nas design variable;

The determination of S102, optimization aim: the Optimality Criteria of cross girder is improve rigidity under the prerequisite of guarantee structural strength and alleviate general assembly (TW);

The foundation of S103, optimized mathematical model: maximum deformation quantity and the general assembly (TW) of choosing cross girder are objective function, and maximum stress is as constraint condition, and optimized mathematical model is

$\left\{\begin{array}{c}X={\&lsqb;P_1,...,P_n\&rsqb;}^T\\ \min F\left(e_{\max },m\right)\\ {\mathrm{\&sigma;}}_{\max }\&le;\&lsqb;\mathrm{\&sigma;}\&rsqb;\\ X_{min}\&le;X\&le;X_{max}\end{array}\right.,$

Wherein, X is design variable, F (e _max, m) be objective function, F (e _max, m)=λ ₁f (e _max)+λ ₂g (m), e _maxfor maximum deformation quantity, [e], for allowing maximum deformation quantity, m is general assembly (TW), and [m] is maximum weight allowable, λ ₁, λ ₂for optimizing weight coefficient, λ ₁+ λ ₂=1, σ _maxfor maximum stress, [σ] is permissible stress, X _max, X _minfor the upper and lower limit of design variable;

S2, acquisition training sample

S201, the method for orthogonal test is taked to carry out sample acquisition, selected design variable is the factor of orthogonal test table, in the span of each design variable, choose several levels, design orthogonal test table, determine the design parameter of group number and each test group tested;

S202, enforcement orthogonal test scheme, corresponding cross girder model is set up according to the parameter of each test group, recycling finite element software carries out strength and distortion analyses to all cross girder models respectively, extract cross girder intensity, rigidity and weight result that simulation obtains, using maximum stress as the token state of intensity, using maximum deformation quantity as the token state of rigidity, using the training sample of these results as neural network;

S3, structure BP neural network

S301, by design variable P ₁, P ₂..., P _nas input layer, by m, e _max, σ _maxas output layer, middle layer adopts single hidden layer, neuron number wherein, n _ifor input layer number, n _ofor output layer nodes, q is constant, q ∈ [1,10];

S302, the training sample obtained in step S202 is utilized to train BP neural network, until the difference of predicted value and sample value is limited within the scope of permissible error;

S4, to solve with genetic algorithm optimization

S401, generation initial population;

S402, use BP neural computing fitness and constraint conditional value, constraint condition function is designed to

$Q\left(X\right)=F\left(e_{\max },m\right)+k\left(\max {\&lsqb;0,\frac{{\mathrm{\&sigma;}}_{\max }}{\&lsqb;\mathrm{\&sigma;}\&rsqb;}-1\&rsqb;}^2\right),\left(k\&GreaterEqual;0\right),$

Fitness function is designed to

$Fit\left(X\right)=\frac{1}{1+c+Q\left(X\right)},(c\&GreaterEqual;0,c+Q(X)\&GreaterEqual;0);$

If S403 meets Optimality Criteria and constraint condition with regard to Output rusults, otherwise select the individuality that fitness is high, perform genetic manipulation and generate new population, then turn to step S402;

S5, determine Optimal Parameters: the parameter combinations that can obtain one group of effect optimum according to the optimum results of genetic algorithm, modeling and simulating is carried out to the cross girder under this group parameter and solves analysis, finally determine optimum results feasibility, if there is larger difference, the structure and the genetic algorithm optimization that re-start neural network solve.

2. method according to claim 1, is characterized in that, in step S101, chooses the position dimension P of secondary crossbeam vertical edges thickness in cross girder ₁, secondary crossbeam transverse edge thickness position dimension P ₂, four rectangular dimension P in secondary crossbeam ₃× P ₄, secondary beam vertical is in the thickness P of drawing surface direction ₅and the position dimension P of main beam transverse edge thickness ₆as design variable.

3. method according to claim 2, is characterized in that, described P ₁span be 2010 ~ 2050mm, described P ₂span be 115 ~ 135mm, described P ₃span be 800 ~ 880mm, described P ₄span be 90 ~ 110mm, described P ₅span be 80 ~ 100mm, described P ₆span be 50 ~ 70mm.

4. method according to claim 3, is characterized in that, in step s 201, chooses 5 levels in the span of each design variable, designs the orthogonal test table of six factor five levels, determines that the group number tested is 25 groups.

5. method according to claim 4, is characterized in that, described P ₁five levels be respectively 2010,2020,2030,2040,2050, described P ₂five levels be respectively 115,120,125,130,135, described P ₃five levels be respectively 880,860,840,820,800, described P ₄five levels be respectively 90,95,100,105,110, described P ₅five levels be respectively 80,85,90,95,100, described P ₆five levels be respectively 50,55,60,65,70.

6. method according to claim 1, is characterized in that, in step s 103, and λ ₁=0.5, λ ₂=0.5.

7. method according to claim 2, is characterized in that, in step S301, and n _i=6, n _o=3, q=9, neuron number $n_1=\sqrt{n_i+n_o}+q=12.$

8. method according to claim 1, is characterized in that, allows maximum deformation quantity [e]=0.5mm, maximum weight allowable [m]=1000kg, permissible stress [σ]=130MPa.