CN104615821A - Automobile wire harness crosstalk pre-estimation method - Google Patents

Abstract

The invention relates to the technical field of automobile wire harness electromagnetic compatibility, and provides an automobile wire harness crosstalk pre-estimation method. According to the method, distributed capacitance and inductance of a wire harness are pre-estimated through a PSO-BP algorithm, corresponding wire harness design parameters are recorded, then the design parameters are optimized through a simplex algorithm to obtain the product of the minimum distributed capacitance and inductance per unit length, and the crosstalk pre-estimation value is obtained through a multi-conductor transmission theory and is compared with the standard value specified by the industry so as to determine whether the electromagnetic compatibility meets the standard or not. According to the method, real data collected by an experiment serve as sample data, statistics, analyzing and integration are conducted on the sample data, on the base of this, a pre-estimation model of the automobile wire harness crosstalk value based on the PSO-BP algorithm is built, the design parameters of the automobile wire harness are optimized through the simplex algorithm, the crosstalk value of the wire harness is reduced accordingly, and certain theory guidance is provided for design of the automobile wire harness so as to enable the electromagnetic compatibility of the automobile wire harness to meet the standard.

Description

The pre-estimation method of automotive wire bundle crosstalk

Technical field

The present invention relates to automotive wire bundle Electro Magnetic Compatibility technical field, be specifically related to a kind of pre-estimation method of automotive wire bundle crosstalk.

Background technology

In recent years, automotive engineering is constantly towards electronization and intelligent direction development, electronic equipment on automobile gets more and more, control to automatic radio equipment and vehicle interior temperature control system etc. from the engine of complexity and brake, the quantity of automobile inner electronic equipment and kind constantly increase, and make electromagnetic environment in car day by day complicated, there is various electromagnetic interference problem, the consequence produced by electromagnetic interference (EMI) is day by day serious, therefore, just seems more and more important to the electromagnetic compatibility Journal of Sex Research of vehicle electronics.

Electromagnetic compatibility refers to that equipment or system meet the requirements in its electromagnetic environment and runs and do not affect the normal work of miscellaneous equipment in environment.Because every bar wire also exists distributed capacitance, distributed inductance over the ground, so there is coupling between wire.Electromagnetic energy is by the capacitive coupling or inductive coupled between conductor, and by a conductive lines to other wires produces electromagnetic induction (curent change), electrostatic induction (change in voltage) is referred to as crosstalk.Wire for connecting various piece on automobile is a lot, and colligation is together, can produce crosstalk, and the crosstalk produced often through conductive lines to sensitive equipment, affect the misoperation of the driver module of performance as equipment, brake etc. of electrical equipment.In factors, the crosstalk of wire harness is the principal element affecting automotive EMC problem, and concrete manifestation is as follows:

(1) wire is the main path of Conduction coupling interference.The electromagnetic interference signal that the interference sources such as electric/electronic device produce directly can input in other electronic equipments by wire or cable, or by capacitive coupling or inductive coupledly enter control line and signal wire, affect the reliability of electronic device functions in car, and the coupling between the wire connecting distinct electronic apparatuses makes the electromagnetic compatibility problem of wire more complicated.

(2) radiation interference is also entered in electronic equipment by wire.Wire is the electromagnetic wave receiving antenna that efficiency is very high, and electromagnetic radiation can be directly coupled on electronic equipment, also can first be coupled on wire, and then conduction enters in electronic equipment.

(3) wire is the electromagenetic wave radiation antenna that efficiency is very high, the especially development of automotive electronic technology, and the frequency of operation of electronic equipment improves constantly, and the electromagnetic radiation that the wire of connection device produces is also day by day serious.

The maturation of automobile electronic system technology is while the comfortableness that improve automobile and convenience, also the also increasingly sophisticated problem of connection of the components and parts diversification more that automobile comprises in the finite space, various equipment is brought, such development trend will cause the deterioration of automobile electromagnetic environment, if electromagnetic interference (EMI) intolerable degradation, will cause Dynamic System malfunctioning or to personal safety produce threaten.Therefore, how to predict the electromagnetic interference (EMI) degree of automotive wire bundle during the design and find effective solution to seem more and more important.The unit length distribution parameter of actual measurement transmission line is the most accurate, but cannot measure before automotive wire bundle is produced.

Summary of the invention

Wire harness is that many transmission lines band together formation, so the forecast model set up should be the forecast model of multi-conductor transmission lines distributed inductance electric capacity.Distributed inductance electric capacity is made up of self-inductance electric capacity and mutual inductance electric capacity, but self-inductance electric capacity is much smaller than mutual inductance electric capacity, so distributed capacitance inductance depends primarily on mutual inductance electric capacity.The present invention proposes to study automotive wire bundle subsystem Electro Magnetic Compatibility, consider the significant impact of automotive EMC to vehicle safety and reliability, and the critical role of wire in automotive EMC problem, the present invention is on the basis that forefathers study, distributed inductance capacitance is estimated by particle cluster algorithm-BP neural network (PSO-BP) method, apply multi-conductor transmission lines theory (Multi-conductor TransmissionLine again, MTL) automotive wire bundle crossfire value is estimated, and the optimal design parameters meeting wire harness under EMC Requirements is tried to achieve by simplex algorithm, and then crosstalk is reduced.The spacing of transmission line unit length distribution parameter and transmission line, sectional area of wire, specific inductive capacity, , the parameters such as distance floor level determine conductor distribution parameter, therefore the spacing of transmission line, wire radius, specific inductive capacity, wire magnetic permeability, the parameters such as conductor spacing floor level are as input quantity, using the target sample of measured data as PSO-BP network training, the distributed inductance of conductor wire harness can be obtained by network training, the pre-estimation model of electric capacity, the relevant parameters of the automotive wire bundle of design phase is brought into this model and can learn whether design meets the standard of electromagnetic compatibility, and utilize simplex algorithm optimal design parameter on this basis, making the product of the distributed capacitance inductance of the unit length of wire harness minimum is that crossfire value is minimum, make the crossfire value between wire harness minimum.

The pre-estimation method of automotive wire bundle crosstalk of the present invention comprises the steps:

Step one: the sample data n group obtaining experiment: select a kind of automotive wire bundle by the radius r of its wire ₁, r ₂, separation d, conductor spacing floor level h ₁, h ₂, wire magnetic permeability μ, DIELECTRIC CONSTANT ε measure input data X as sample _i=(r _i1, r _i2, d _i, h _i1, h _i2, μ _i, ε _i) ^t, (i=1,2 ..., n), using the distributed inductance of wire harness and capacitance measurement out as the desired output T of sample _i=(t _i1, t _i2) ^t, (i=1,2 ... .., n; t _i1represent distributed inductance value, t _i2represent distributed capacitance), measure n group altogether;

Step 2: experimental data normalization: in sample data step one obtained, the maxima and minima of each row is found out and is designated as z respectively _max, z _min, by normalization formula each data in n group sample data are normalized;

Step 3: determine the nodes of the number of plies of BP neural network, each layer, excitation function, initially connect weights, threshold value;

3.1 numbers of plies determining BP neural network: have abundant hidden layer and node in hidden layer just can the characteristic of Nonlinear Function Approximation arbitrarily as long as have according to BP neural network, and can realize, to approaching of arbitrary function, the structure of BP neural network is defined as three layers of i.e. input layer, hidden layer, output layer through verification certificate hidden layer BP neural network according to existing;

The determination of 3.2BP neural network input layer nodes: the input quantity number according to BP neural network is 7, determines that the input layer number of BP neural network is 7;

The determination of 3.3BP neural network node in hidden layer: according to the same sample training in n group sample data, the node in hidden layer getting the minimum correspondence of error, as final node in hidden layer, determines that BP neural network node in hidden layer is 14;

The determination of 3.4BP neural network output layer nodes: the output quantity number according to BP neural network is 2, determines that the output layer nodes of BP neural network is 2;

The weight threshold initialization of 3.5BP neural network: initial weight threshold value adopts the method for random assignment to obtain, and input layer is w to the connection weights of hidden layer _ij(i=(1,2 ..., 7), j=(1,2 ..., 14)), hidden layer is w to the connection weights of output layer _jl(j=(1,2 ..., 14), l=(1,2)) and, the threshold value of input layer is zero, the threshold value θ of hidden layer node entirely _j=(α ₁, α ₂..., α ₁₄) be random number between 0 to 1;

The excitation function of 3.6BP neural network input layer and hidden layer selects Log-sigmoid type function: (net _ifor the input of input layer, hidden layer node, being input as of hidden layer ); The excitation function of output layer selects purelin function (y _ilrepresent that the prediction of i-th sample of BP neural network exports; L represents the neuronic number of output layer);

Step 4: initialization m group weights and threshold is as initial population scale, and each group weights and threshold, as a particle, makes p _best=0 and G _best=0; Initial velocity value is produced, the error of permission with rands () (gbest-fit _ibe fitness minimum value in the particle colony of i-th iteration, k is iterations); Setting speed inertia weight w, Studying factors c ₁, c ₂size and setting v _max, x _maxvalue, structure fitness function, gets the performance index function of BP neural network (y _ilrepresent that the prediction of i-th sample of BP neural network exports; t _ilrepresent the desired output of i-th sample of BP neural network; L represents the neuronic number of output layer) as fitness function and J _p=E _p;

Step 5: sample inputs, and calculates the output valve of BP neural network, according to fitness function J according to the forward-propagating of BP neural network _p, calculate each ideal adaptation angle value pbest-fit, make position corresponding to the minimum fitness value of each particle be individual p _best, particle position corresponding when fitness value gbest-fit is minimum in population is designated as G _best;

Step 6: upgrade weights and threshold particle, according to the more translational speed of new particle and the position of formula below:

$v_{\mathrm{id}}^{k+1}=w*v_{\mathrm{id}}^k+c_1*\mathrm{rand}\left(\right)*(p_{\mathrm{id}}-x_{\mathrm{id}}^k)+c_2*\mathrm{rand}\left(\right)*(p_{\mathrm{gd}}-x_{\mathrm{id}}^k)$

$x_{\mathrm{id}}^{k+1}=x_{\mathrm{id}}^k+v_{\mathrm{id}}^{k+1}$

In formula, w is velocity inertia weight; c ₁for the weights coefficient of Particle tracking oneself history optimal value; c ₂for the weights coefficient of Particle tracking colony optimal value; Rand () is the random number between 0 to 1; p _idfor the history optimal value that particle oneself searches; p _gdfor the history optimal value that population searches; be i-th particle kth moment speed; be i-th particle kth moment position;

If v _id> v _max, then v is made _id=v _max; If v _id<-v _max, then v _id=-v _max, otherwise v _idget the value after renewal; If x _id> x _max, then x is made _id=x _max; If x _id<-x _max, then x _id=-x _max, otherwise x _idget the value after renewal;

Step 7: calculate the particle fitness value pbest-fit' after upgrading, and to compare with pbest-fit, if pbest-fit' is little, then by the particle position assignment of its correspondence to p _best, pbest-fit' assignment is to pbest-fit, otherwise p _bestconstant; Pbest-fit' and gbest-fit after this iteration is compared, if pbest-fit' is less than gbest-fit, then by its assignment to gbest-fit, and by position assignment corresponding for particle to G _best, otherwise remain unchanged;

Step 8: when training reaches permissible error requirement or maximum iteration time, training terminates, the now global optimum position G of particle _bestcorresponding weights and threshold is exactly best initial weights and the threshold value of BP neural network;

Step 9: select n from sample database ₁group data are as check data, check the forecast model determined by step 3 to step 8, if prediction exports with the difference of desired output in the scope that error allows, then forecast model is set up, otherwise return step 3 and rebuild forecast model, until prediction exports with the difference of desired output in the scope that error allows;

Step 10: obtain accurately after forecast model by step 9, uses simplex algorithm to the radius r of seven of BP neural network input parameters and wire ₁, r ₂, separation d, conductor spacing floor level h ₁, h ₂, wire magnetic permeability μ, DIELECTRIC CONSTANT ε carry out optimizing, brings the optimized parameter of trying to achieve into forecast model, to determine whether seven design parameters after optimizing make the product value of distributed capacitance and distributed inductance reduce; Detailed process is as follows:

10.1 in n group data Stochastic choice 8 data form one group of initial vertax X=(X as simplex algorithm ₁, X ₂..., X ₈), projection coefficient α, magnificationfactorβ, contraction coefficient γ are set;

10.2 select target function f () the i.e. performance index function of BP neural network (y _ilrepresent that the prediction of i-th sample of BP neural network exports; t _ilrepresent the desired output of i-th sample of BP neural network; L represents the neuronic number of output layer), calculate the desired value size of each point, find out the some x that desired value is maximum _h, minimum some x _m, and arrange from small to large according to desired value;

10.3 calculate projected centre point, determine x according to projection coefficient α _r;

If 10.4 f (x _m) < f (x _r) < f (x _h), use x _rreplace x _hform a new simple point, return 10.2;

10.5 amplify simplex: make x _ε=γ x _r+ (1-γ) x, if f is (x _ε) < f (x _m), then amplify successfully, use x _εreplace x _hand form a new simplex; If f is (x _ε) > f (x _m), then amplify failure, use x _rreplace x _hreturn 10.2, continue projection process;

10.6 shrink simplex: if a little have f (x for the institute except i=h _r) > f (x _i) and f (x _r) < f (x _i), then use x _rreplace x _hand simplex is reduced: x _c=β x _h+ (1-β) x _o; If f is (x _r) > f (x _h), then reduce simplex, but do not change and previous be projected an x _h; If f is (x _c) < min [f (x _m), f (x _r)], then use x _creplace original being projected an x _h, then proceed projection process; If f is (x _c) > min [f (x _m), f (x _r)], then this contraction process failure, now uses (x _i+ x _m)/2 replace all x _i, then continue projection process;

If the relative error on 10.7 summits meets given accuracy requirement, then stop iteration, the centre of form of current simplex is optimum point;

Step 11: repeat step 10.1 to 10.7 and seek obtaining seven optimum points again;

Step 12: that step 10 and step 11 are sought amounts to eight optimum points as new one group of initial vertax, repeats the optimal design value that step 10.1 to 10.7 tries to achieve seven input parameters larger on the impact of automotive wire bundle crossfire value in step one.

The invention has the beneficial effects as follows: the method is based upon on existing production data basis, distributed capacitance and the inductance of wire harness is predicted by training pattern, thus determine the crossfire value of wire harness, be used to guide design and the production of automotive wire bundle, reference can be provided for automobile harness wiring, solve the bunch product quality problems because wire harness crosstalk brings to a certain extent.

Accompanying drawing explanation

Fig. 1 is the structural representation of BP neural network in the present invention.

Embodiment

Below in conjunction with accompanying drawing, the present invention is described in further details.

The present invention is for automotive wire bundle, corresponding strand design parameter is recorded by the distributed capacitance of PSO-BP algorithm pre-estimation wire harness and distributed inductance value, then optimize by simplex algorithm the product that this design parameter draws the minimum distributed capacitance inductance value of unit length, what draw crosstalk by many conductor propagation theory again estimates evaluation, then compares with industry specified standard value thus determines that whether Electro Magnetic Compatibility is up to standard.Method of the present invention tests the True Data of collection for sample data, sample data is added up, analyze, integrate, and set up the pre-estimation model of the automotive wire bundle crossfire value based on PSO-BP algorithm on this basis, and utilize simplex algorithm to optimize the design parameter of automotive wire bundle, and then reduce the crossfire value of wire harness, thus provide certain theoretical direction for the design of automotive wire bundle, make its Electro Magnetic Compatibility (crossfire value of automotive wire bundle controls within the specific limits) up to standard.

The pre-estimation method of automotive wire bundle crosstalk of the present invention comprises the steps:

Step one: the sample data n group obtaining experiment.Select a kind of automotive wire bundle by the radius r of its wire ₁, r ₂, separation d, distance floor level h ₁, h ₂, wire magnetic permeability μ, DIELECTRIC CONSTANT ε measure input data X as sample _i=(r _i1, r _i2, d _i, h _i1, h _i2, μ _i, ε _i) ^t, (i=1,2 ..., n), using the distributed inductance capacitance measurement of wire harness out as the desired output T of sample _i=(t _i1, t _i2) ^t, (i=1,2 ... .., n; t _i1represent distributed inductance value, t _i2represent distributed capacitance) measure n group altogether.

Step 2: experimental data normalization.The maxima and minima of dimension same in sample data (i.e. each row) is found out and is designated as z _max, z _min, by normalization formula each data are normalized.

Step 3: determine the nodes of the number of plies of BP neural network, each layer, excitation function, initially connect weights, threshold value, specific as follows:

3.1 numbers of plies determining BP neural network.As long as having according to BP neural network has abundant hidden layer and the number of hidden nodes just can the characteristic of Nonlinear Function Approximation arbitrarily, and according to existing through verification certificate hidden layer BP neural network can realize to arbitrary function approach and effect in actual use can be defined as three layers of i.e. input layer, hidden layer, output layer, shown in its structural drawing 1 the structure of BP neural network.

3.2 nodes determining BP neural network input layer.The input layer number of BP neural network is determined by the number of input quantity, because BP network has seven input quantities, therefore has seven nodes according to the feature input layer of BP network.

3.3 nodes determining BP neural network hidden layer.The effect of hidden layer node is that the inherent law in sample is extracted, and stored therein.Node in hidden layer is few, and be not enough to the inherent law of fully excavating sample, hidden layer node number is many, and the interference effect of improper data will be exaggerated, and the generalization ability of BP network is deteriorated, and the time of training also can increase.Usual hidden layer node number depends on the complexity of sample inherent law, to same sample training, gets the node in hidden layer of the minimum correspondence of error as final node in hidden layer.Getting hidden layer neuron number is herein 14.

3.4 nodes determining BP neural network output layer.Output layer nodes is number according to output quantity and fixed, and therefore the interstitial content of output layer is 2, and the structure of final BP neural network of establishing is 7-14-2.

The weight threshold initialization of 3.5BP neural network.Initial weight threshold value adopts the method for random assignment to obtain, and input layer is w to the connection weights of hidden layer _ij(i=(1,2 ..., 7), j=(1,2 ..., 14)), hidden layer is w to the connection weights of output layer _jl(j=(1,2 ..., 14), l=(1,2)).The threshold value of input layer is zero, the threshold value θ of hidden layer node entirely _j=(α ₁, α ₂..., α ₁₄) be random number between 0 to 1.

The input layer of 3.6BP neural network and the excitation function of hidden layer select Log-sigmoid type function: (net _ifor the input of input layer, hidden layer node, being input as of hidden layer ); The excitation function of output layer selects purelin function (y _ilrepresent that the prediction of i-th sample of BP neural network exports; L represents the neuronic number of output layer).

Step 4: initialization m group weights and threshold is as initial population scale, and each group weights and threshold, as a particle, makes p _best=0 and G _best=0; Initial velocity value is produced, the error of permission with rands () (gbest-fit _ibe fitness minimum value in the particle colony of i-th iteration, k is iterations); Setting speed inertia weight w, Studying factors c ₁, c ₂size and setting v _max, x _maxvalue, structure fitness function J _p, get the performance index function of neural network (y _ilrepresent that the prediction of i-th sample of BP neural network exports; t _ilrepresent the desired output of i-th sample of BP neural network; L represents the neuronic number of output layer) as fitness function and J _p=E _p.

Step 5: sample inputs, and calculates the output valve of neural network, according to fitness function J according to the forward-propagating of BP neural network _p, calculate each ideal adaptation angle value pbest-fit, make position corresponding to the minimum fitness value of each particle be individual p _best, particle position corresponding when fitness value gbest-fit is minimum in population is designated as G _best.

Step 6: upgrade weights and threshold particle.Come translational speed and the position of more new particle according to formula below, and consider that the particle rapidity after upgrading and position are whether in the scope limited:

$v_{\mathrm{id}}^{k+1}=w*v_{\mathrm{id}}^k+c_1*\mathrm{rand}\left(\right)*(p_{\mathrm{id}}-x_{\mathrm{id}}^k)+c_2*\mathrm{rand}\left(\right)*(p_{\mathrm{gd}}-x_{\mathrm{id}}^k)$

$x_{\mathrm{id}}^{k+1}=x_{\mathrm{id}}^k+v_{\mathrm{id}}^{k+1}$

In formula, w is velocity inertia weight; c ₁for the weights coefficient of Particle tracking oneself history optimal value; c ₂for the weights coefficient of Particle tracking colony optimal value; Rand () is the random number between 0 to 1; p _idfor the history optimal value that particle oneself searches; p _gdfor the history optimal value that population searches; be i-th particle kth moment speed; be i-th particle kth moment position;

If v _id> v _max, then v is made _id=v _max; If v _id<-v _max, then v _id=-v _max, otherwise v _idget updated value; If x _id> x _max, then x is made _id=x _max; If x _id<-x _max, then x _id=-x _max, otherwise x _idget updated value.

Step 7: calculate the particle fitness value pbest-fit' after upgrading, and to compare with pbest-fit, if pbest-fit' is little, then by the particle position assignment of its correspondence to p _best, pbest-fit' assignment is to pbest-fit otherwise p _bestconstant; Pbest-fit' and gbest-fit after this iteration is compared, if having less than gbest-fit, then by its assignment to gbest-fit, and by position assignment corresponding for particle to G _best, otherwise remain unchanged.

Step 8: when reaching permissible error requirement or maximum iteration time, training terminates.The global optimum position G of particle _bestbe exactly best initial weights and the threshold value of neural network.

Step 9: select n from sample database ₁organize data as check data, the reliability of the forecast model that inspection above-mentioned steps obtains.If prediction exports with the error of desired output in tolerance interval, then forecast model is set up, otherwise returns step 3 and rebuild forecast model, until error can accept.

Step 10: obtain accurately after forecast model, simplex algorithm is used to carry out optimizing to seven of BP neural network input parameters, bring the optimized parameter of trying to achieve into forecast model, to determine whether the design parameter after optimizing makes the product value of distributed capacitance inductance reduce, and detailed process is as follows:

10.1 in n group data Stochastic choice 8 data form one group of initial vertax X=(X as simplex algorithm ₁, X ₂..., X ₈), projection coefficient α, magnificationfactorβ, contraction coefficient γ are set.

(objective function is the performance index function of BP neural network to 10.2 select target function f () ) (y _ilrepresent that the prediction of i-th sample of BP neural network exports; t _ilrepresent the desired output of i-th sample of BP neural network; L represents the neuronic number of output layer), calculate the size of the desired value of each point, find out the some x that desired value is maximum _h, minimum some x _m, and be according to target worth and arrange from small to large.

10.3 calculate projected centre point, determine x according to projection coefficient α _r.

If 10.4 f (x _m) < f (x _r) < f (x _h), use x _rreplace x _hform a new simple point, return 10.2.

10.5 amplify simplex, make x _ε=γ x _r+ (1-γ) x, if f is (x _ε) < f (x _m), then amplify successfully, use x _εreplace x _hand form a new simplex.If f is (x _ε) > f (x _m), then amplify failure, still use x _rreplace x _hreturn 10.2, continue projection process.

10.6 contraction simplex.If a little have f (x for the institute except i=h _r) > f (x _i) and f (x _r) < f (x _i), then use x _rreplace x _hand simplex is reduced: x _c=β x _h+ (1-β) x _o.If f is (x _r) > f (x _h) still reduce simplex, but do not change and previous be projected an x _h; If f is (x _c) < min [f (x _m), f (x _r)], then use x _creplace original being projected an x _h, then proceed projection process: if f is (x _c) > min [f (x _m), f (x _r)], then this contraction process failure, now uses (x _i+ x _m)/2 replace all x _i, then continue projection process.

If the relative error on 10.7 summits meets given accuracy requirement, then stop iteration, the centre of form of current simplex is optimum point.

Step 11: repeat step 10.1 to 10.7 and seek obtaining seven optimum points again.

Step 12: eight optimum points of step 10 and step 11 being sought, as new one group of initial vertax, repeat step 10.1 to 10.7 and try to achieve optimal design value automotive wire bundle crossfire value being affected to seven larger input parameters.

Because transmission line cross-interference issue is occupied an leading position in automobile electromagnetic disturbing factor, therefore the present invention is based on transmission line theory, study and establish the transmission line crosstalk pre-estimation model of automobile electronic system, and the part optimized parameter of strand design is tried to achieve with method for optimally controlling, fruitful work has been done to automobile electronic system research and development test in early stage transmission line crosstalk aspect.

Claims (1)

1. the pre-estimation method of automotive wire bundle crosstalk, it is characterized in that, the method comprises the steps:

Step one: the sample data n group obtaining experiment: select a kind of automotive wire bundle by the radius r of its wire ₁, r ₂, separation d, conductor spacing floor level h ₁, h ₂, wire magnetic permeability μ, DIELECTRIC CONSTANT ε measure input data X as sample _i=(r _i1, r _i2, d _i, h _i1, h _i2, μ _i, ε _i) ^t, (i=1,2 ..., n), using the distributed inductance of wire harness and capacitance measurement out as the desired output T of sample _i=(t _i1, t _i2) ^t, (i=1,2 ... .., n; t _i1represent distributed inductance value, t _i2represent distributed capacitance), measure n group altogether;

Step 2: experimental data normalization: in sample data step one obtained, the maxima and minima of each row is found out and is designated as z respectively _max, z _min, by normalization formula each data in n group sample data are normalized;

Step 3: determine the nodes of the number of plies of BP neural network, each layer, excitation function, initially connect weights, threshold value;

3.1 numbers of plies determining BP neural network: have abundant hidden layer and node in hidden layer just can the characteristic of Nonlinear Function Approximation arbitrarily as long as have according to BP neural network, and can realize, to approaching of arbitrary function, the structure of BP neural network is defined as three layers of i.e. input layer, hidden layer, output layer through verification certificate hidden layer BP neural network according to existing;

The determination of 3.2BP neural network input layer nodes: the input quantity number according to BP neural network is 7, determines that the input layer number of BP neural network is 7;

The determination of 3.3BP neural network node in hidden layer: according to the same sample training in n group sample data, the node in hidden layer getting the minimum correspondence of error, as final node in hidden layer, determines that BP neural network node in hidden layer is 14;

The determination of 3.4BP neural network output layer nodes: the output quantity number according to BP neural network is 2, determines that the output layer nodes of BP neural network is 2;

The weight threshold initialization of 3.5BP neural network: initial weight threshold value adopts the method for random assignment to obtain, and input layer is w to the connection weights of hidden layer _ij(i=(1,2 ..., 7), j=(1,2 ..., 14)), hidden layer is w to the connection weights of output layer _jl(j=(1,2 ..., 14), l=(1,2)) and, the threshold value of input layer is zero, the threshold value θ of hidden layer node entirely _j=(α ₁, α ₂..., α ₁₄) be random number between 0 to 1;

The excitation function of 3.6BP neural network input layer and hidden layer selects Log-sigmoid type function: (net _ifor the input of input layer, hidden layer node, being input as of hidden layer the excitation function of output layer selects purelin function (y _ilrepresent that the prediction of i-th sample of BP neural network exports; L represents the neuronic number of output layer);

Step 4: initialization m group weights and threshold is as initial population scale, and each group weights and threshold, as a particle, makes p _best=0 and G _best=0; Initial velocity value is produced, the error of permission with rands () (gbest-fit _ibe fitness minimum value in the particle colony of i-th iteration, k is iterations); Setting speed inertia weight w, Studying factors c ₁, c ₂size and setting v _max, x _maxvalue, structure fitness function, gets the performance index function of BP neural network (y _ilrepresent that the prediction of i-th sample of BP neural network exports; t _ilrepresent the desired output of i-th sample of BP neural network; L represents the neuronic number of output layer) as fitness function and J _p=E _p;

Step 5: sample inputs, and calculates the output valve of BP neural network, according to fitness function J according to the forward-propagating of BP neural network _p, calculate each ideal adaptation angle value pbest-fit, make position corresponding to the minimum fitness value of each particle be individual p _best, particle position corresponding when fitness value gbest-fit is minimum in population is designated as G _best;

Step 6: upgrade weights and threshold particle, according to the more translational speed of new particle and the position of formula below:

$v_{\mathrm{id}}^{k+1}=w*v_{\mathrm{id}}^k+c_1*\mathrm{rand}\left(\right)*(p_{\mathrm{id}}-x_{\mathrm{id}}^k)+c_2*\mathrm{rand}\left(\right)*(p_{\mathrm{gd}}-x_{\mathrm{id}}^k)$

$x_{\mathrm{id}}^{k+1}=x_{\mathrm{id}}^k+v_{\mathrm{id}}^{k+1}$

In formula, w is velocity inertia weight; c ₁for the weights coefficient of Particle tracking oneself history optimal value; c ₂for the weights coefficient of Particle tracking colony optimal value; Rand () is the random number between 0 to 1; p _idfor the history optimal value that particle oneself searches; p _gdfor the history optimal value that population searches; be i-th particle kth moment speed; be i-th particle kth moment position;

If v _id> v _max, then v is made _id=v _max; If v _id<-v _max, then v _id=-v _max, otherwise v _idget the value after renewal; If x _id> x _max, then x is made _id=x _max; If x _id<-x _max, then x _id=-x _max, otherwise x _idget the value after renewal;

Step 7: calculate the particle fitness value pbest-fit' after upgrading, and to compare with pbest-fit, if pbest-fit' is little, then by the particle position assignment of its correspondence to p _best, pbest-fit' assignment is to pbest-fit, otherwise p _bestconstant; Pbest-fit' and gbest-fit after this iteration is compared, if pbest-fit' is less than gbest-fit, then by its assignment to gbest-fit, and by position assignment corresponding for particle to G _best, otherwise remain unchanged;

Step 8: when training reaches permissible error requirement or maximum iteration time, training terminates, the now global optimum position G of particle _bestcorresponding weights and threshold is exactly best initial weights and the threshold value of BP neural network;

Step 9: select n from sample database ₁group data are as check data, check the forecast model determined by step 3 to step 8, if prediction exports with the difference of desired output in the scope that error allows, then forecast model is set up, otherwise return step 3 and rebuild forecast model, until prediction exports with the difference of desired output in the scope that error allows;

Step 10: obtain accurately after forecast model by step 9, uses simplex algorithm to the radius r of seven of BP neural network input parameters and wire ₁, r ₂, separation d, conductor spacing floor level h ₁, h ₂, wire magnetic permeability μ, DIELECTRIC CONSTANT ε carry out optimizing, brings the optimized parameter of trying to achieve into forecast model, to determine whether seven design parameters after optimizing make the product value of distributed capacitance and distributed inductance reduce; Detailed process is as follows:

10.1 in n group data Stochastic choice 8 data form one group of initial vertax X=(X as simplex algorithm ₁, X ₂..., X ₈), projection coefficient α, magnificationfactorβ, contraction coefficient γ are set;

10.2 select target function f () the i.e. performance index function of BP neural network (y _ilrepresent that the prediction of i-th sample of BP neural network exports; t _ilrepresent the desired output of i-th sample of BP neural network; L represents the neuronic number of output layer), calculate the desired value size of each point, find out the some x that desired value is maximum _h, minimum some x _m, and arrange from small to large according to desired value;

10.3 calculate projected centre point, determine x according to projection coefficient α _r;

If 10.4 f (x _m) < f (x _r) < f (x _h), use x _rreplace x _hform a new simple point, return 10.2;

10.5 amplify simplex: make x _ε=γ x _r+ (1-γ) x, if f is (x _ε) < f (x _m), then amplify successfully, use x _εreplace x _hand form a new simplex; If f is (x _ε) > f (x _m), then amplify failure, use x _rreplace x _hreturn 10.2, continue projection process;

10.6 shrink simplex: if a little have f (x for the institute except i=h _r) > f (x _i) and f (x _r) < f (x _i), then use x _rreplace x _hand simplex is reduced: x _c=β x _h+ (1-β) x _o; If f is (x _r) > f (x _h), then reduce simplex, but do not change and previous be projected an x _h; If f is (x _c) < min [f (x _m), f (x _r)], then use x _creplace original being projected an x _h, then proceed projection process; If f is (x _c) > min [f (x _m), f (x _r)], then this contraction process failure, now uses (x _i+ x _m)/2 replace all x _i, then continue projection process;

If the relative error on 10.7 summits meets given accuracy requirement, then stop iteration, the centre of form of current simplex is optimum point;

Step 11: repeat step 10.1 to 10.7 and seek obtaining seven optimum points again;

Step 12: that step 10 and step 11 are sought amounts to eight optimum points as new one group of initial vertax, repeats the optimal design value that step 10.1 to 10.7 tries to achieve seven input parameters in step one.