CN103886371A - Method for controlling component and thermal treatment technological process of pre-hardening plastic die steel - Google Patents

Abstract

The invention discloses a method for controlling the component and thermal treatment technological process of pre-hardening plastic die steel. The method includes the steps that firstly, an orthogonal test is designed according to structure characteristics and performance requirements of the pre-hardening plastic die steel, and thermal treatment technological parameters which have obvious influences on performance are found out through the test; then chemical components, the thermal treatment technological parameters and performance indexes are collected; an artificial neural network model is built, the model is trained, input of the model is the chemical components and the thermal treatment technological parameters, and output of the model is the performance indexes; the trained artificial neural network model is used for predicting steel performance and studying influence laws of input to output according to the chemical components and the thermal treatment technological parameters. According to the method, the complicated non-linear relationship between the chemical components, the thermal treatment technological parameters and mechanical properties is built, therefore, the component and thermal treatment technological process can be effectively controlled, the pre-hardening plastic die steel with excellent performance can be generated, and the service life of dies can be prolonged.

Description

A kind of pre-hardened plastic die steel composition and heat treatment process control method

Technical field

The present invention relates to ferrous materials and manufacture field, relate in particular to a kind of pre-hardened plastic die steel heat treatment process control method.

Background technology

At present, plastic products are widely used in the every field of national economy, and constantly to precise treatment, maximization and complicated future development.In plastics forming, use mould of plastics.But in many enterprises of China, the serviceable life of mould of plastics is also lower at present.The main failure mode of mould is surface abrasion, deformation and fracture.The overwhelming majority that lost efficacy mould causes because thermal treatment is improper, selection is improper.Therefore, the chemical composition of plastic die steel and the optimization of Technology for Heating Processing can make it have good mirror finish, suitable hardness, good wearing quality and corrosion stability etc.

Complicated, high-precision mould of plastics is heat-treated after pocket machining again, there will be the heat treating faults such as quenching distortion, cracking, decarburization, has a strong impact on surface of plastic products quality.The pre-hardened plastic die steel of widespread use is at present reprocessed into mould after the modifier treatment that completes steel or module, has effectively avoided heat treating fault.Therefore pre-hardened plastic die steel is particularly suitable for manufacturing complex-shaped, accurate big-and-middle-sized mould of plastics.

Mould steel is different from other surperficial purposes steel, and mould steel will be processed into complicated shape through die sinking, and therefore the inner arbitrary section of steel all likely becomes the workplace in use procedure.The microstructure of steel different parts and the difference of performance all likely affect the quality of product, and therefore the homogeneity of its Microstructure and properties just seems particularly important.Suitable heat treatment process control can effectively improve the Microstructure and properties homogeneity of mould.

Summary of the invention

The object of this invention is to provide a kind of pre-hardened plastic die steel composition and heat treatment process control method, can draw chemical composition, heat treatment process parameter and mould steel property relationship, then control composition and the Technology for Heating Processing of mould steel, to obtaining excellent performance, and the higher plastic die steel of performance uniformity.

The present invention adopts following technical proposals: a kind of pre-hardened plastic die steel composition and heat treatment process control method, comprise the following steps:

(1), according to the tissue signature of mould steel and performance requirement design orthogonal test, factor is heat treatment process parameter, therefrom find out the significant heat treatment process parameter of performance impact, the significant heat treatment process parameter of performance impact is comprised: one or more in quenching temperature, temperature retention time and the type of cooling, for the first time temperature and temperature retention time, for the second time temperature and temperature retention time; Mould steel performance index comprise: hardness, yield strength σ _0.2, tensile strength sigma _b,, length growth rate δ, reduction of area ψ and ballistic work A _kv;

, gather chemical composition, to the significant heat treatment process parameter of performance impact and mould steel performance index, wherein chemical composition is expressed by CE, $\mathrm{CE}=C+\frac{\mathrm{Si}}{11}+\frac{\mathrm{Mn}}{6}+\frac{\mathrm{Cr}+\mathrm{Mo}+V}{5}+\frac{\mathrm{Cu}+\mathrm{Ni}}{15}+5B;$

, set up artificial nerve network model, and to model training, the CE that is input as mould steel of model and to the significant heat treatment process parameter of performance impact, model is output as mould steel performance index: hardness, yield strength σ _0.2, tensile strength sigma _b, length growth rate δ, reduction of area ψ and ballistic work A _kv;

(4), mould steel performance is predicted according to chemical composition and to the significant heat treatment process parameter of performance impact with the artificial nerve network model that trains, predicted value and actual value are compared;

, make arbitrary input of the artificial neural network training in reasonable value range, other inputs are fixing, draw the affect rule of this input on output, finally draw the affect rule of all inputs on output;

(6), according to the (4) contour map of the result drafting mould steel performance of middle mould steel performance prediction of step;

, the CE that affects the contour map control mould steel that (6) rule and step draw that (5) draws according to step, to the significant heat treatment process parameter of performance impact in suitable span, make produced mould steel performance meet standard-required.

It is to set up artificial BP neural network model that described step is set up artificial nerve network model in (3).

Thereby described step is trained for to model the process of adjusting BP neural network model weights and minimize BP neural network performance function in (3), and wherein performance function E is:

$E=\mathrm{\γ}\frac{1}{m}\underset{k=1}{\overset{m}{\mathrm{\Σ}}}\frac{1}{q}\underset{t=1}{\overset{q}{\mathrm{\Σ}}}{({T_t}^k-a_t^k)}^2+(1-\mathrm{\γ})\frac{1}{m}\underset{k=1}{\overset{m}{\mathrm{\Σ}}}\frac{1}{p\·q+n\·p+q+p}\underset{l=1}{\overset{p\·q+n\·p+q+p}{\mathrm{\Σ}}}{\left(w_l^k\right)}^2---\left(1\right)$

In formula (1) for the desired output of t output node of k pattern of network;

for the actual output of t output node of k pattern of network;

for l weight of k pattern of network; N is input number of nodes; P is hidden layer node number; Q is output node number; M is training mode number; γ is performance parameter;

The ground floor of BP neural network is input layer, and the second layer is hidden layer, and the 3rd layer is output layer, and the activation function that hidden layer and output layer adopt is respectively:

$f\left(n\right)=\frac{2}{1+\exp (-2n)}-1---\left(2\right)$

f(n)=n(3)

The adjustment of network weight and threshold value adopts Levenberg-Marquardt algorithm, and this algorithm has second order speed of convergence need not calculate Hessian matrix simultaneously, and Hessian matrix H and gradient g can be expressed as with following approximate matrix:

H=J ^TJ (4)

G=J ^tin e (5) formula, J is Jacobi matrix, J ^tfor transposed matrix, the vector that e is network error;

Weights or threshold value x calculate with following formula:

X _k+1=x _k-[J ^tj+ μ I] ^-1j ^tin e (6) formula, μ is scalar, and I is unit matrix;

BP neural metwork training process is as follows:

(a), initialization connection weight w _ji, v _tj, threshold value θ _j, γ _t, and give the random value of (1 ,+1), w _jifor input layer i is to the connection weights of hidden layer j unit, v _tjfor hidden layer j is to the connection weights of output layer t unit, θ _jfor the threshold value of hidden layer j unit, γ _tfor the threshold value of output layer t unit;

(b), choose at random a pattern to I _k, T _koffer network: network input

for CE, to the significant heat treatment process parameter of performance impact, k is 1,2 ..., m, m is training mode number; T _kfor desired output; Actual being output as for hardness, yield strength σ _0.2, tensile strength sigma _b, length growth rate δ, reduction of area ψ, ballistic work A _kv;

(c), calculate the input of hidden layer j unit: ${\mathrm{HidIn}}_j=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{w}_{\mathrm{ji}}\·I_i-{\mathrm{\θ}}_j---\left(7\right)$

The output of hidden layer j unit: ${\mathrm{HidOut}}_j=f\left({\mathrm{HidIn}}_j\right)=\frac{2}{1+\exp (-2{\mathrm{HidIn}}_j)}-1---\left(8\right)$

(d), calculate the input of output layer t unit: ${\mathrm{OutIn}}_t=\underset{j=1}{\overset{p}{\mathrm{\Σ}}}{v}_{\mathrm{tj}}\·{\mathrm{HidOut}}_j-{\mathrm{\γ}}_t---\left(9\right)$

Output layer t unit output: a _t=f (OutIn _t)=OutIn _t(10)

(e) get, at random next mode of learning pair, return to step (c), until whole m pattern is to having trained;

(f), adjust weights and threshold value according to Levenberg-Marquardt algorithm;

(g), again select at random one to return to (c) from m mode of learning centering, until minimize BP network performance function E;

(h), study finishes.

Described step can also be that general regression neural network is trained to model training in (3): after the training sample of general regression neural network is determined, connection weights between corresponding network structure and each neuron are also thereupon definite, and the training of network is the process of determining the smooth factor;

The ground floor of network is input layer, and input vector is P, the dimension that neuron number R is input variable, P=[p ₁, p ₂..., p _r], p ₁, p ₂..., p _rfor the CE of mould steel with to the significant heat treatment process parameter of performance impact; The second layer of network is basic unit radially, and neuronic number is that training mode is counted Q, the weight matrix W of ground floor ¹be set as input sample I, deviation b ¹for the smooth factor, represent with σ, choice variable voluntarily, the weighting input ‖ dist ‖ of network represents input variable P and W ¹euclidean Norm, that is:

for W ¹the capable j column element of weight matrix i; The clean input n of network ¹for $n_i^1={\left|\right|\mathrm{dist}\left|\right|}_i\×b_i^1(i=\mathrm{1,2}...Q),$ This layer is output as $a_i^1=\exp (-\frac{{\left(n_i^1\right)}^2}{2\mathrm{\σ}})(i=\mathrm{1,2}...Q);$ The 3rd layer of network is linear output layer, and neuron number is that training mode is counted Q, W ²be set as output T; Basis function adopts regularization dot product function, the output of basis function

the linear activation function f of substitution (n)=n obtains the output of network $a_i^2=n_i^2(i=\mathrm{1,2}...Q).$

The present invention adopts artificial nerve network model can successfully set up the complex nonlinear relation between chemical composition, heat treatment process parameter and the mechanical property of pre-hardened plastic die steel, with the mechanical property of high-precision forecast mould steel, and can determine chemical composition and the affect rule of heat treatment process parameter on mechanical property, reduce Mechanical Fluctuation, thereby produce the mould steel of excellent performance, improved the serviceable life of mould.Specifically also have following advantage:

1, the present invention combines conventional orthogonal test design method and artificial neural network, has effectively set up the relation of composition, technique and performance.Adopt Orthogonal Experiment and Design from multiple factors, to pick out the significant heat treatment process parameter of performance impact, as the input of artificial neural network, can effectively reduce network input, simplified network structure, improve the precision of neural network forecast;

2, due to the environment for use of mould steel harshness, the contained chemical element of mould steel is more, and if every kind of element is all as an input of neural network, network is bulky complex too, and speed of convergence is slower, and precision of prediction reduces; The present invention on the basis of lot of experiments research and theoretical analysis, the degree of strength according to each element to mould steel performance impact, determine chemical composition by $\mathrm{CE}=(\mathrm{CE}=C+\frac{\mathrm{Si}}{11}+\frac{\mathrm{Mn}}{6}+\frac{\mathrm{Cr}+\mathrm{Mo}+V}{5}+\frac{\mathrm{Cu}+\mathrm{Ni}}{15}+5B)$ Express, and input using CE as network, further reduce the input of network, simplify network structure, the precision of raising neural network forecast;

3, can draw the contour map of mould steel performance according to the result of mould steel performance prediction, can be used for, according to concrete performance requirement, selecting suitable CE and heat treatment process parameter, make produced mould steel performance meet standard-required.

Accompanying drawing explanation

Fig. 1 is method flow diagram of the present invention;

Fig. 2 is BP network structure;

Fig. 3 is GRNN network structure;

Fig. 4 is the contour map of the hardness of 2 1050 ℃ of quenching mould steel of embodiment.

Embodiment

The invention discloses a kind of pre-hardened plastic die steel heat treatment process control method, as shown in Figure 1, specifically comprise the following steps:

(1), according to the tissue signature of mould steel and performance requirement design orthogonal test, factor is heat treatment process parameter, therefrom find out the significant heat treatment process parameter of performance impact, the significant heat treatment process parameter of performance impact is comprised: quenching temperature, temperature retention time and the type of cooling, temperature and temperature retention time for the first time, for the second time one or more in temperature and temperature retention time; Mould steel performance index comprise: hardness, yield strength σ _0.2, tensile strength sigma _b, length growth rate δ, reduction of area ψ, ballistic work A _kv;

, gather chemical composition and to the significant heat treatment process parameter of performance impact and mould steel performance index, wherein chemical composition is expressed by CE, $\mathrm{CE}=C+\frac{\mathrm{Si}}{11}+\frac{\mathrm{Mn}}{6}+\frac{\mathrm{Cr}+\mathrm{Mo}+V}{5}+\frac{\mathrm{Cu}+\mathrm{Ni}}{15}+5B;$

, set up artificial nerve network model, and to model training, the CE that is input as mould steel of model and to the significant heat treatment process parameter of performance impact, model is output as mould steel performance index: hardness, yield strength σ _0.2, tensile strength sigma _b, length growth rate δ, reduction of area ψ, ballistic work A _kv;

(4), mould steel performance is predicted according to chemical composition and to the significant heat treatment process parameter of performance impact with the artificial nerve network model that trains, predicted value and actual value are compared;

, make arbitrary input of the artificial neural network training in reasonable value range, other inputs are fixing, draw the affect rule of this input on output, finally draw the affect rule of all inputs on output;

(6), according to the (4) contour map of many groups result drafting mould steel performance of middle mould steel performance prediction of step; , the CE that affects the contour map control mould steel that (6) rule and step draw that (5) draws according to step, heat treatment process parameter in suitable span, make produced mould steel performance meet standard-required.

It is to set up artificial BP neural network model that described step is set up artificial nerve network model in (3); Thereby BP neural network model is trained for to the process of adjusting BP neural network model weights and minimize BP neural network performance function, wherein performance function E is:

$E=\mathrm{\γ}\frac{1}{m}\underset{k=1}{\overset{m}{\mathrm{\Σ}}}\frac{1}{q}\underset{t=1}{\overset{q}{\mathrm{\Σ}}}{({T_t}^k-a_t^k)}^2+(1-\mathrm{\γ})\frac{1}{m}\underset{k=1}{\overset{m}{\mathrm{\Σ}}}\frac{1}{p\·q+n\·p+q+p}\underset{l=1}{\overset{p\·q+n\·p+q+p}{\mathrm{\Σ}}}{\left(w_l^k\right)}^2---\left(1\right)$

In formula (1)

for the desired output of t output node of k pattern of network;

for the actual output of t output node of k pattern of network;

for l weight of k pattern of network; N is input number of nodes; P is hidden layer node number; Q is output node number; M is training mode number; γ is performance parameter;

As shown in Figure 2, the ground floor of BP neural network is input layer, and the second layer is hidden layer, and the 3rd layer is output layer, and the activation function that hidden layer and output layer adopt is respectively:

$f\left(n\right)=\frac{2}{1+\exp (-2n)}-1---\left(2\right)$

f(n)=n (3)

The adjustment of network weight and threshold value adopts Levenberg-Marquardt algorithm, and this algorithm has second order speed of convergence need not calculate Hessian matrix simultaneously, and Hessian matrix H and gradient g can be expressed as with following approximate matrix:

H=J ^TJ (4)

g=J ^Te (5)

In formula, J is Jacobi matrix, J ^tfor transposed matrix, the vector that e is network error;

Weights or threshold value x calculate with following formula:

x _k+1=x _k-[J ^TJ+μI] ^-1J ^Te (6)

In formula, μ is scalar, and I is unit matrix;

BP neural metwork training process is as follows:

(a), initialization connection weight w _ji, v _tj, threshold value θ _j, γ _t, and give the random value of (1 ,+1), w _jifor input layer i is to the connection weights of hidden layer j unit, v _tjfor hidden layer j is to the connection weights of output layer t unit, θ _jfor the threshold value of hidden layer j unit, γ _tfor the threshold value of output layer t unit;

(b), choose at random a pattern to I _k, T _koffer network: network input for CE, to the significant heat treatment process parameter of performance impact, k is 1,2 ..., m, m is training mode number; T _kfor desired output; Actual being output as

for hardness, yield strength σ _0.2, tensile strength sigma _b, length growth rate δ, reduction of area ψ, ballistic work A _kv;

(c), calculate the input of hidden layer j unit: ${\mathrm{HidIn}}_j=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{w}_{\mathrm{ji}}\·I_i-{\mathrm{\θ}}_j---\left(7\right)$

The output of hidden layer j unit: ${\mathrm{HidOut}}_j=f\left({\mathrm{HidIn}}_j\right)=\frac{2}{1+\exp (-2{\mathrm{HidIn}}_j)}-1---\left(8\right)$

(d), calculate the input of output layer t unit: ${\mathrm{OutIn}}_t=\underset{j=1}{\overset{p}{\mathrm{\Σ}}}{v}_{\mathrm{tj}}\·{\mathrm{HidOut}}_j-{\mathrm{\γ}}_t---\left(9\right)$

Output layer t unit output: a _t=f (OutIn _t)=OutIn _t(10)

(e) get, at random next mode of learning pair, return to step (c), until whole m pattern is to having trained;

(f), adjust weights and threshold value according to Levenberg-Marquardt algorithm;

(g), again select at random one to return to (c) from m mode of learning centering, until minimize BP network performance function E;

(h), study finishes.

In addition, described step can also be that general regression neural network (GRNN) is trained to model training in (3): after the training sample of general regression neural network is determined, connection weights between corresponding network structure and each neuron are also thereupon definite, and the training of network is the process of determining the smooth factor; As shown in Figure 3, the ground floor of network is input layer, and input vector is P, the dimension that neuron number R is input variable, P=[p ₁, p ₂..., p _r], p ₁, p ₂..., p _rfor the CE of mould steel with to the significant heat treatment process parameter of performance impact; The second layer of network is basic unit radially, and neuronic number is that training mode is counted Q, the weight matrix W of ground floor ¹be set as input sample I, deviation b ¹for the smooth factor, represent with σ, choice variable voluntarily, the weighting input ‖ dist ‖ of network represents input variable P and W ¹euclidean Norm, that is:

for W ¹the capable j column element of weight matrix i; The clean input n of network ¹for $n_i^1={\left|\right|\mathrm{dist}\left|\right|}_i\×b_i^1(i=\mathrm{1,2}...Q),$ This layer is output as $a_i^1=\exp (-\frac{{\left(n_i^1\right)}^2}{2\mathrm{\σ}})(i=\mathrm{1,2}...Q);$ The 3rd layer of network is linear output layer, and neuron number is that training mode is counted Q, W ²be set as output T; Basis function adopts regularization dot product function, the output of basis function the linear activation function f of substitution (n)=n obtains the output of network $a_i^2=n_i^2(i=\mathrm{1,2}...Q).$

The whole implementation procedure of method of the present invention is compiled into computer software with Visual C++, first gather and store the chemical composition of mould steel and to the significant heat treatment process technological parameter of performance impact and performance index, set up artificial nerve network model, artificial neural network does not need preset model, only, by the study of relation between data, just can reflect the relation between chemical composition and processing parameter and performance.The mould steel superior performance that adopts process control method of the present invention to produce, Microstructure and properties homogeneity is better.Specific embodiment is as described below.

Embodiment mono-: design orthogonal test, choose the significant heat treatment process parameter of performance impact by orthogonal test: quenching temperature, the type of cooling, temperature, as the input of artificial neural network (different mould steel be different to the significant heat treatment process parameter of performance impact, what choose in the present embodiment is quenching temperature, the type of cooling and temperature).Table 1 is for being used for the input data of network training; Table 2 is for being used for the output data of network training; The heat treatment process parameter that the mould steel chemical composition that table 3 is used for neural network forecast and neural network forecast are used; Table 4 is performance and the actual performance of neural network forecast.From table 4, estimated performance and actual performance are coincide better, BP network and GRNN network can be used for specializes in chemistry composition and on the significant heat treatment process parameter of performance impact on mould steel performance affect rule research, thereby the CE of control mould steel,

Heat treatment process parameter, in suitable span, makes produced mould steel performance meet standard-required.

Table 1

Table 2

Table 3

Table 4

Embodiment bis-: design orthogonal test, choose the significant heat treatment process parameter of impact by orthogonal test: quenching temperature and the type of cooling, temperature, temperature for the second time for the first time, as the input of artificial neural network.Table 5 is for being used for the input data of network training; Table 6 is for being used for the output data of network training; The heat treatment process parameter that table 7 is used with chemical composition and the neural network forecast of steel for neural network forecast; Table 8 is performance and the actual performance of neural network forecast.Fig. 4 is the contour map of the hardness of 1050 ℃ of quenching mould steel, and other performance index of 1050 ℃ of quenching mould steel all can draw contour map.From table 8, estimated performance and actual performance are coincide better, BP network and GRNN network can be used for specializes in chemistry composition and the significant heat treatment process parameter of performance impact are affected to rule research to mould steel performance, thereby the CE of control mould steel, heat treatment process parameter, in suitable span, make produced mould steel performance meet standard-required.

Table 5

Table 6

Table 7

Table 8

Claims (4)

1. pre-hardened plastic die steel composition and a heat treatment process control method, is characterized in that: comprise the following steps:

(1), according to the tissue signature of mould steel and performance requirement design orthogonal test, factor is heat treatment process parameter, therefrom find out the significant heat treatment process parameter of performance impact, the significant heat treatment process parameter of performance impact is comprised: one or more in quenching temperature, temperature retention time and the type of cooling, for the first time temperature and temperature retention time, for the second time temperature and temperature retention time; Mould steel performance index comprise: hardness, yield strength σ _0.2, tensile strength sigma _b,, length growth rate δ, reduction of area ψ and ballistic work A _kv;

, gather chemical composition, to the significant heat treatment process parameter of performance impact and mould steel performance index, wherein chemical composition is expressed by CE, $\mathrm{CE}=C+\frac{\mathrm{Si}}{11}+\frac{\mathrm{Mn}}{6}+\frac{\mathrm{Cr}+\mathrm{Mo}+V}{5}+\frac{\mathrm{Cu}+\mathrm{Ni}}{15}+5B;$

, set up artificial nerve network model, and to model training, the CE that is input as mould steel of model and to the significant heat treatment process parameter of performance impact, model is output as mould steel performance index: hardness, yield strength σ _0.2, tensile strength sigma _b, length growth rate δ, reduction of area ψ and ballistic work A _kv;

(4), mould steel performance is predicted according to chemical composition and to the significant heat treatment process parameter of performance impact with the artificial nerve network model that trains, predicted value and actual value are compared;

, make arbitrary input of the artificial neural network training in reasonable value range, other inputs are fixing, draw the affect rule of this input on output, finally draw the affect rule of all inputs on output;

(6), according to the (4) contour map of the result drafting mould steel performance of middle mould steel performance prediction of step;

, the CE that affects the contour map control mould steel that (6) rule and step draw that (5) draws according to step, to the significant heat treatment process parameter of performance impact in suitable span, make produced mould steel performance meet standard-required.

2. mould steel composition according to claim 1 and heat treatment process control method, is characterized in that: it is to set up artificial BP neural network model that described step is set up artificial nerve network model in (3).

3. mould steel composition according to claim 2 and heat treatment process control method, it is characterized in that: thus described step is trained for to model the process of adjusting BP neural network model weights and minimize BP neural network performance function in (3), and wherein performance function E is:

$E=\mathrm{\γ}\frac{1}{m}\underset{k=1}{\overset{m}{\mathrm{\Σ}}}\frac{1}{q}\underset{t=1}{\overset{q}{\mathrm{\Σ}}}{({T_t}^k-a_t^k)}^2+(1-\mathrm{\γ})\frac{1}{m}\underset{k=1}{\overset{m}{\mathrm{\Σ}}}\frac{1}{p\·q+n\·p+q+p}\underset{l=1}{\overset{p\·q+n\·p+q+p}{\mathrm{\Σ}}}{\left(w_l^k\right)}^2---\left(1\right)$

In formula (1)

for the desired output of t output node of k pattern of network;

for the actual output of t output node of k pattern of network;

for l weight of k pattern of network; N is input number of nodes; P is hidden layer node number; Q is output node number; M is training mode number; γ is performance parameter;

The ground floor of BP neural network is input layer, and the second layer is hidden layer, and the 3rd layer is output layer, and the activation function that hidden layer and output layer adopt is respectively:

$f\left(n\right)=\frac{2}{1+\exp (-2n)}-1---\left(2\right)$

f(n)=n (3)

The adjustment of network weight and threshold value adopts Levenberg-Marquardt algorithm, and this algorithm has second order speed of convergence need not calculate Hessian matrix simultaneously, and Hessian matrix H and gradient g can be expressed as with following approximate matrix:

H=J ^TJ (4)

G=J ^tin e (5) formula, J is Jacobi matrix, J ^tfor transposed matrix, the vector that e is network error;

Weights or threshold value x calculate with following formula:

x _k+1=x _k-[J ^TJ+μI] ^-1J ^Te (6)

In formula, μ is scalar, and I is unit matrix;

BP neural metwork training process is as follows:

(a), initialization connection weight w _ji, v _tj, threshold value θ _j, γ _t, and give the random value of (1 ,+1), w _jifor input layer i is to the connection weights of hidden layer j unit, v _tjfor hidden layer j is to the connection weights of output layer t unit, θ _jfor the threshold value of hidden layer j unit, γ _tfor the threshold value of output layer t unit;

(b), choose at random a pattern to I _k, T _koffer network: network input

for CE, to the significant heat treatment process parameter of performance impact, k is 1,2 ..., m, m is training mode number; T _kfor desired output; Actual being output as

for hardness, yield strength σ _0.2, tensile strength sigma _b, length growth rate δ, reduction of area ψ, ballistic work A _kv;

(c), calculate the input of hidden layer j unit: ${\mathrm{HidIn}}_j=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{w}_{\mathrm{ji}}\·I_i-{\mathrm{\θ}}_j---\left(7\right)$

The output of hidden layer j unit: ${\mathrm{HidOut}}_j=f\left({\mathrm{HidIn}}_j\right)=\frac{2}{1+\exp (-2{\mathrm{HidIn}}_j)}-1---\left(8\right)$

(d), calculate the input of output layer t unit: ${\mathrm{OutIn}}_t=\underset{j=1}{\overset{p}{\mathrm{\Σ}}}{v}_{\mathrm{tj}}\·{\mathrm{HidOut}}_j-{\mathrm{\γ}}_t---\left(9\right)$

Output layer t unit output: a _t=f (OutIn _t)=OutIn _t(10)

(e) get, at random next mode of learning pair, return to step (c), until whole m pattern is to having trained;

(f), adjust weights and threshold value according to Levenberg-Marquardt algorithm;

(g), again select at random one to return to (c) from m mode of learning centering, until minimize BP network performance function E;

(h), study finishes.

4. pre-hardened plastic die steel composition according to claim 1 and heat treatment process control method, it is characterized in that: to model, training is that general regression neural network is trained to described step in (3): after the training sample of general regression neural network is determined, connection weights between corresponding network structure and each neuron are also thereupon definite, and the training of network is the process of determining the smooth factor;

The ground floor of network is input layer, and input vector is P, the dimension that neuron number R is input variable, P=[p ₁, p ₂..., p _r], p ₁, p ₂..., p _rfor the CE of mould steel with to the significant heat treatment process parameter of performance impact; The second layer of network is basic unit radially, and neuronic number is that training mode is counted Q, the weight matrix W of ground floor ¹be set as input sample I, deviation b ¹for the smooth factor, represent with σ, choice variable voluntarily, the weighting input ‖ dist ‖ of network represents input variable P and W ¹euclidean Norm, that is:

for W ¹the capable j column element of weight matrix i; The clean input n of network ¹for $n_i^1={\left|\right|\mathrm{dist}\left|\right|}_i\×b_i^1(i=\mathrm{1,2}...Q),$ This layer is output as $a_i^1=\exp (-\frac{{\left(n_i^1\right)}^2}{2\mathrm{\σ}})(i=\mathrm{1,2}...Q);$ Layer of network is linear output layer, and neuron number is that training mode is counted Q, W ²be set as output T; Basis function adopts regularization dot product function, the output of basis function

the linear activation function f of substitution (n)=n obtains the output of network $a_i^2=n_i^2(i=\mathrm{1,2}...Q).$

Images