You are on page 1of 4

2009 International Conference on E-Learning, E-Business, Enterprise Information Systems, and E-Government

Fitting and prediction for crack propagation rate based on machine


learning optimal algorithm

Yi WAN, Chengwen WU
College of Physics and Electronic Information Engineering
Wenzhou University
Wenzhou, China
E-mail: yiwan246@126.com, wuchengwen@yahoo.com

Abstract—Establishing fatigue crack propagation rate is the ultimate goal that designers obtain fatigue crack
key to forecasting structure fatigue lifetime, nine parameters propagation rate curve simply and reliably [1-4].
fatigue crack propagation rate model and McEvily model In the paper, we presented optimal common machine
are widely applied at present, but it is very complex to learning algorithm (least squares support vector
realize these models, partial derivative must be calculated machine-LSSVM) method for fatigue crack propagation
and there is large deviation between fitted static parameter rate forecast. Complicated and strong nonlinear fatigue
and actual value and physical conception isn’t clear. In crack propagation rate curve was simulated by network
accordance with the disadvantage above methods, Based on
design and conformation of LSSVM learning algorithm
optimum parameter selection with grid search and cross
and the optimized SVM parameters were selected by the
validation, we presented optimal common machine learning
algorithm (least squares support vector machine-LSSVM) method of network searching and cross validation.
method for fatigue crack propagation rate forecast. Compared the errors with output value of the optimized
Complicated and strong nonlinear fatigue crack propagation model and output value from nine parameters fatigue
rate curve was simulated by network design and crack propagation rate fitting model, LSSVM whose
conformation of LSSVM learning algorithm and the parameter was optimized with cross validation had
optimized SVM parameters were selected by the method of excellent ability of nonlinear modeling and generalization.
network searching and cross validation. Compared the It provided a simple and feasible intelligent approach for
errors with output value of the optimized model and output
material fatigue analysis.
value from nine parameters fatigue crack propagation rate
fitting model, LSSVM whose parameter was optimized with
cross validation had excellent ability of nonlinear modeling II. SVM ALGORITHM OF PARAMETER
and generalization. It provided a simple and feasible OPTIMUM SELECTION
intelligent approach for material fatigue analysis. A. Improved support vector algorithm-LSSVM
Keywords- Optimized common machine learning algorithm; Improved SVM algorithm is called least squares
grid search and cross validation; fatigue crack propagation support vector machine (LSSVM). It uses different
rate; material fatigue analysis, nine parameters fatigue crack optimized objective function and replaces equation
propagation rate model restriction by in-equation restriction [5-10].
Objective optimized function of LSSVM algorithm is:
I. INTRODUCTION 1 1 n 2
min J ( w, e) = W T W + γ ∑ ek (1)
More and more people understand fatigue crack w ,b , e 2 2 K =1
propagation rate curve is more basic curve than S-N curve
s.t. y k = W T φ ( x k ) + b + e k , k = 1,2, " , N
to describe material fatigue speciality in structure
reliability analysis. Establishing fatigue crack propagation where φ (⋅) : R n → R nh is kernel space mapping
rate is the key to forecasting structure fatigue lifetime,
function; w ∈ R nh is weight vector; ek ∈ R is error
and fatigue crack propagation rate is usually calculated by
some estimating model, such as McEvily model and W. C. variable; b is deflection variable; γ is adjustable
CUI presented a nine parameters fatigue crack parameter.
propagation rate model within a wide applicable range, Using equations (1), Lagrange function is obtained as
each parameter value of crack propagation rate model is
estimated by many experiment data for estimating model. N
In principle, all parameters are able to be established by {
L = J (w, e) − ∑αk W Tφ(xk ) + b + ek − yk } (2)
nonlinear multi-parameter regression method, but it is k =1
very complex to realize the model, complex partial where α k is Lagrange multiplier.
derivative need be calculated, in addition, there is large Partial derivative is solved to (2) and (3) is obtained by
deviation between fitted static parameter and actual value, transforming.
at the same time, physical conception isn’t clear, it is

978-0-7695-3907-2/09 $26.00 © 2009 IEEE 93


DOI 10.1109/EEEE.2009.31
−1
T ⎡ 1 ⎤
1n ⎢Ω + I n ⎥ y
⎣ γ ⎦
b= −1
T ⎡ 1 ⎤
1 n ⎢ Ω + I n ⎥ 1n
⎣ γ ⎦
−1
⎡ 1 ⎤
α = ⎢Ω + I n ⎥ ( y − 1n b ) (3)
⎣ γ ⎦
So function of LSSVM decision-making classification is
N
y ( k ) = ∑ α k K ( x, x k ) + b (4) Fig.1 LSSVM intelligent model structure of fatigue crack propagation
k =1 rate prediction
where K ( x, x k ) is Gaussian Kernel Function,
B. Simulation and result
{
K ( x, x k ) = exp − x − x k
2
/σ 2
}. To evaluate the accuracy and practicality of fatigue
crack propagation rate intelligent model based on
B. Optimal selection of model parameter LSSVM, using material 1 and material 2 as simulation
Parameter pairs ( γ , σ ) are selected by grid search object [12], LSSVM model is trained by limit experimental
method at first, and then goal function is optimized using data, and then model prediction ability is validated by
cross validation until optimal parameter pairs are part remaining experiment data. LSSVM parameters are
searched [11]. optimized by grid search-cross validation, optimal result
A parameter pair ( γ , σ ) is selected by grid search, is γ = 12288.6006 and σ = 0.270370922 . Optimization
and the parameter pair is validated crosswise. The method
traversal process of parameters γ , σ based on grid
is that sample set D is divided to H
search-cross validation is shown in figure2.
groups {G1 , G 2 , " , G H } , Random H − 1 groups are
acted as sample set, remaining group is acted as
validation set, H times are able to be repeated by selecting
different validation set. And then cross validation is
progressed by selecting parameter pairs circularly and
mean square error MSE is calculated to every parameter
until grid search stop.

III. LSSVM INTELLIGENT MODEL OF FATIGUE


CRACK PROPAGATION RATE
A. Model structure design and data pretreatment
There are strong non-linear relations between material
crack propagation rate CPR (m/cycle) and the amplitude
value of stress intensity factor Δ K , and the characteristic Fig.2 Optimization traversal process of parameters γ,σ
relation is complex and uncertain, non-linear function based on grid search-cross validation
relation is described by (6). In nine parameters fitting model:
CPR = f ( Δ K ) (5) da = AM 2 (10)
Δ K = YΔ σ π a (6) Kc
( )
dN 1 − K max n

CPR = da / dN (7) ( )
M = Kmax(1 − R) − 1 − e−ka (Kop max − RKmax) − ΔKeffth (11)
Where Y is Geometry factor of crack factual length,
a is crack length (m), σ is material stress (Mpa). ⎛ σ ⎞⎛ ⎞
Kmax = π re⎜ Secπ max +1⎟⎜⎜1+Y(a) a ⎟⎟σmax (12)
Equation (5) can be represented as following equation. ⎝ 2 σV ⎠⎝ 2re ⎠
da / dN = f ( a ) (8) ΔK effth K op max
a is acted as network input layer, CPR is output, Where and are function of stress
output node number is one, intelligent model structure is ratio R, the two parameters can be described as following
shown in figure1. formula.
To improve network speciality, data are pretreated by ΔK effth = ΔK effth0 (1 − R) γ (13)
normalization, Gaussian kernel is used in the intelligent
model, namely: K op max = K op max 0 (1 − R) γ (14)
2 In the above model, there are nine parameters A, n,
K ( x, xi ) = exp(− x − xi /σ 2) (9)

94
σu, re, Kc, ΔK effth 0 , Kopmax0 , k and γ, they can be
solved by Mathematical theory of non-linear
multi-parameter regression.
Calculation and Prediction results of material 1
fatigue crack propagation rate values based on least
square support vector machine algorithm and nine
parameters fitting model are shown in figure3, Error of
LSSVM prediction values and calculating values of nine
parameters fitting model are shown in table1 and figure4,
mean relative error is 0.146% by calculating, it provided
an economical, practical and reliable approach for
material fatigue design.

Fig.4 Relative error of LSSVM prediction values and calculating


values of nine parameters fitting model

Fatigue crack propagation rate predictive curve of


material 1 and material 2 are obtained by trained and
optimized LSSVM intelligent model, their curve are
shown in figure5 and figure6.

Fig.3 Prediction results of fatigue crack propagation rate based on


LSSVM optimal algorithm

Table1. Comparison between LSSVM predictive values and nine


parameters fitting values
M a/m Nine
LSSVM
at parameter
predictive error
er s fitting
values
ial values Figure5.Fatigue crack propagation rate predictive curve of material1
1 0.001 -4.50 -4.7959 -0.2959
0.004 -3.80 -4.0565 -0.2565
0.008 -3.50 -3.4359 0.0641
0.010 -3.30 -3.2515 0.0485
0.015 -3.10 -3.0343 0.0657
0.020 -3.00 -2.9754 0.0246
0.030 -2.60 -2.6272 -0.0272
0.031 -2.50 -2.5551 -0.0551
0.032 -2.40 -2.4762 -0.0762
0.034 -2.41 -2.2998 0.1102
2 0.020 -3.00 -2.9903 0.0097
0.024 -2.90 --2.8468 0.0532
0.026 -2.80 -2.8244 -0.0244
0.028 -2.70 -2.7927 -0.0927
0.030 -2.60 --2.6665 -0.0665
0.031 -2.50 -2.5303 -0.0303 Figure6.Fatigue crack propagation rate predictive curve of material2
0.033 -2.20 -2.0194 0.1806
0.034 -1.40 -1.6047 -0.2047

95
IV. CONCLUSION [2] C.Makabe, H.Kaneshiro. Consideration of evaluation of fatigue
crack propagation rate from effective stress intensity factor range,
In the paper, we presented optimal common machine Journal of Testing and Evaluation, May 2007, v 23, n 3, p 153-159.
learning algorithm (least squares support vector [3] X. M. Lu, Y.F. Ma. Correlation of the positive stress ratio effects on
machine-LSSVM) method for fatigue crack propagation fatigue crack propagation rate of aluminium alloys based on a new
parameter da/Ds, Journal of Harbin Institute of Technology (New
rate forecast. Complicated and strong nonlinear fatigue Series), June 2008, v 15, n 3, p 374-377.
crack propagation rate curve was simulated by network [4] F. S. Silva. Fatigue crack propagation after overloading and under
design and conformation of LSSVM learning algorithm loading at negative stress ratios, International Journal of Fatigue,
September/November 2007, v 29, n 9-11, p 1757-1771.
and the optimized SVM parameters were selected by the
[5] R Xiao , J. C Wang, Z X Sun, etc. An approach to incremental SVM
method of network searching and cross validation. learning algorithm. Journal of Nanjing University (Natural
Compared the errors with output value of the optimized Sciences) , 2002 ,38 (2) :152-157.
[6] C. D Guo, S. Z Li. Control - based Audio Classification and
model and output value from nine parameters fatigue Retrieval by Support Vector Machines. IEEE Trans. on Neural
crack propagation rate fitting model, LSSVM whose Network , 2003, 14 (1) , pp.209-115.
[7] Nello Cristianini, S T. John. An Introduction to Support Vector
parameter was optimized with cross validation had Machines and Other Kernel-based Learning Methods. Cambridge:
excellent ability of nonlinear modeling and generalization, University Press,2000.
[8] V N. Vapnik. Statistical Learning Theory . New York:
mean relative error of LSSVM prediction values and Springer-Verlag,2000.
calculating values of nine parameters fitting model is [9] C. W. Hsu, C. J. Lin. A comparison of methods for multi-class
support vector machines. IEEE Trans on Neural Networks, 2002,
0.146%. It provided a simple and feasible intelligent 13(2):415-425.
approach for material fatigue analysis. [10] H.F. Wu, Z.X. Li, J.L. Wu. Intelligent fault diagnosis for diesel
engine exhaust valve using support machine, Transactions of
CSICE, 2006, 24(5),pp.465-469.
ACKNOWLEDGEMENT [11] Zheng Shuibo. Application of LSSVMs in the automobile
dynamical system identification. Journal of Shanhai Jiaotong
This work is supported by foundation of Wenzhou
science and technology bureau (H20080051), supported University,2005,3(39):392-395.
[12] C. ZHOU. Non-linear fitting and sensitivity analysis of fatigue
by Foundation of Zhe’jiang Educational Committee of
crack propagation rate nine parameters model,
China (No.20070533).The authors are grateful for the http://www.baidu.com/s?
anonymous reviewers who made constructive comments. lm=0&si=&rn=10&ie=gb2312&ct=0&wd=%B7%C7%CF%DF%
D0%D4%C4%E2%BA%CF&pn=20&ver=0&cl=3&uim=0&usm=
0
REFERENCES
[1] F.S. Silva. The importance of compressive stresses on fatigue crack .
propagation rate, International Journal of Fatigue,
October/December 2005,v 27, n 10-12, p 1441-1452.

96

You might also like