Beruflich Dokumente
Kultur Dokumente
Int. J, Fatigue Vol. 19, No. 5. pp. 415-419, 1997 g3 1997 Elsevier Science Limited All rights reserved. Printed in Great Britain 0142 1123/97/$17.00+.00
PIh S0142-1123(97)00037-0
INTRODUCTION Fatigue testing is time-consuming and costly. Therefore, research works have been carried out to estimate P - S - N curves from a limited number of fatigue tests by statistical methods, for example, see Nakazawa and Kodama ~ and Nishijima 2. Nakazawa and Kodama ~ introduced the JSME (Japan Society of Mechanical Engineers) Standard JSME S 002-1981 where 14 specimens are recommended for estimating S - N curves. In their method, the S - N curve is assumed to consist of two linear parts, a slope part and a horizontal part, in either a semi-log or a log-log scale. A maximum likelihood method is described in the ASTM Standard E739-913 for interpreting S - N data for the slope part. But, the maximum likelihood method is not suitable for the horizontal part. A more general method based on the maximum likelihood principle is proposed in this paper for estimating P - S - N curves based on a 3parameter non-linear expression for S - N curves 4. For the purpose of simplifying the analysis as in the ASTM Standard E739-913, the logarithm of fatigue life is assumed to follow a normal distribution although the log-normal distribution is fundamentally unsuited for the representation of the stress-life behavior because it predicts a decreasing failure rate for long lives. The proposed maximum likelihood method is used to estimate P - S - N curves from the test results of a carbon
steel. The estimated P - S - N curves are then compared with those obtained by the conventional method. FORMULATION The proposed maximum likelihood method for estimating P - S - N curves can be explained in the following. As shown in Figure 1, fatigue test results are presented in terms of the stress level Si ( i = 1,2 ..... n) and the logarithm of fatigue life N/. Initially, fatigue tests are designed to carry out at different stress levels. At each stress level Si, a fatigue life Ni can be obtained. Then, a stress level Sr is chosen as a reference stress level and a group of specimens are tested under this reference stress. Sr is chosen from S/in the range of fatigue
S'
0 *Permanent address is: College of Vehicle Engineering, Beijing Institute of Technology, Beijing, 100081, P.R. China.
log N
415
416
life of interest, say, 105 cycles. We denote the number of specimens in the reference group by r/. Then, the logarithms of fatigue lives of the r/ specimens, log Nj, j = 1,2 ..... r/, can be used to estimate the mean and standard deviation of the reference group as in Equations (1) and (2), respectively. ~r=~
j= 1
logN#
(1)
o'-=
~ - 1~
j= 1
(log ~ - ~")~
(2)
For simplicity, we take advantage of /x:,= 1.0 at p = 0.841. At p = 0.841, we replace the subscript p by 1. Equation (13) can then we rewritten as S " - So S - Sol o-(S) = O: +/3 log ~ _ S , 7 + / 3 , 1 O g S , . , S I , , (14)
In contrast to the popular assumption of bi-linear SN curves in a semi-log or a log-log scale 3, a more general 3-parameter expression for S - N curves 4 is assumed here
N(S - So)t3 = c~
(3)
where So, /3 and c~ are material constants. Equation (3) can be used to represent P - S - N curves at survival probability p as
With the assumption that the logarithm of fatigue life at an arbitrary stress level Si is normally distributed, the probability density function of the logarithm of fatigue life can be represented as
f(log Ni)
1 ~ e \,2~S,)
I logN~
~(,S' ) ]2
2,,2(s,~)
(15)
(4)
where Sop, /3p and %, are the material constants and N:, is the life corresponding to survival probability p. We can take the logarithms of the both sides of Equation (4). Then, Equation (4) becomes log N,, = log %,-/3:,log (S - Sop) (5) For p = 0.5, we arrive at the following equation by omitting the subscript p /x(S) = log No..so = log c, - / 3 l o g (S - So) (6) where /x(S) represents the mean of the logarithm of fatigue life at the stress level S. Since we assume that the logarithm of fatigue life is normally distributed, the logarithm of fatigue life at survival probability p, log N,, can be expressed as log N:, =/x(S) -/x:,o(S)
or
~t(.'i)l-
,_,,--~s , i,
/16t
Taking the natural logarithms of the both sides of Equation (16) yields " { -[IogNi-/x(Si)] 2 } In L = - ~', In ,,2rr+ In cr(Si) + 2o_2(& )
i=1
(7)
" {
= ~ In o(Si) +
i=1
[logN,-tx(Si)] 2 }
2o.2(Si) . (18)
o(S) =
1
/xp
(8)
where o-(S) is the standard deviation of the logarithm of fatigue life under the stress level S, and &, is the standard normal deviate corresponding to survival probability p. /xp can easily be obtained for the normal distribution if p is given. Substituting Equations (5) and (6) into Equation (8) yields o-(S) =/xplog al '
(l)
-/x:,
/3 l o g ( S - S o ) + / 3 "it,, log(S-Sop)
Substituting Equations (12) and (14) into Equation (18), an explicit form of F can be obtained. It is obvious that the minimum point of F is exactly the maximum point of the likelihood function L. The maximum likelihood estimates of /3, /3~, So, and So, can be obtained by searching the minimum point of F. When /3, /3~, So and So, are available, /x(S) and o-(S), or P - S - N curves, can be obtained by substituting /3, /3j, So, and Sol into Equations (12), (14) and (7). APPLICATION Standard fatigue specimens with a theoretical stress concentration factor k, = 2.0 were made with #45 steel (Chinese steel). Two sets of fatigue tests were conducted under axial loads with the mean stress Sm=O (R = - 1 ) . One set was arranged according to the proposed maximum likelihood method, and the other was arranged by the conventional method 3. The results of the two sets of tests are listed in Tables 1 and 2. Following the maximum likelihood method discussed in the previous section, we have S r = 372.8 MPa. From Equations (1) and (2), we have/2"= 4.7065, &'= 0.093. Substituting Equations (12) and (14) with the values
(9)
Now, we consider the reference case such that S = S". Then, /x(S0 =/2 ~ and o-(Sr) = &" as defined in Equations (1) and (2). Equations (6) and (9) can then be written as /~"= log oe - / 3 log(S r - So) (10) and &" = /xp log /%, ( 1 1)
417
Stress amplitude
5, (MPa) 529.74 490.50 431.64 392.40 372.78
Sample size
I 1 1 I 6
arbitrary survival probability p can be calculated by I,~(S)- ~:,o(S) in Equation (7). For example, the logarithm of fatigue life with survival probability p = 0.841 is log
No.841
(S)
(21)
I 1 I
P - S - N curves can also be obtained from the test results according to the conventional method. Columns 5 and 6 in Table2 are the mean and the standard deviation of the logarithm of fatigue life, respectively. The logarithm of fatigue life with the survival probability of 0.841 can be obtained by / , ( S ) - o-(S). In Ref. 3, linear S-N curves are recommended to represent fatigue data in a log-log scale as
log N = A + B log S (22) where A and B are material constants. "['he relationship between the mean of the logarithm of fatigue life and the stress amplitude can be obtained by fitting Equation (22) to the data in column 1 and 5 of Table 2. The fitted equation is log No soo = 2 3 . 5 8 6 9 - 7 . 3 1 4 3 l o g S
(23)
of /2' and 6~ into Equation (18), the function F can be obtained. The minimum point of F can be found by a usual optimization method. The results are:
/3 = 1 . 6 7 5 2 /31 = 1.8075 So = 2 7 6 . 2 Soj = 2 6 4 . 9
These are the maximum likelihood estimates of/3, /3,, So, and Sin. Substituting these values into Equations (12) and (14) gives
/.L(S) = 8 . 0 3 1 7 - 1 . 6 7 5 2 l o g ( S - 2 7 6 . 2 ) (19)
The relationship between the logarithm of fatigue life with survival probability of 0.841 and the stress amplitude can be obtained by fitting Equation (22) to the data in column 1 and the difference of data in columns 5 and 6 of Table 2. The fitted equation is log No~41 = 2 1 . 7 4 3 2 - 6 . 6 3 2 0 log S
(24)
and
o-(S) = - 0 . 2 5 6 5 - 1.6752 log(S - 276.2) (20)
+ 1.8075 l o g ( S - 2 6 4 . 9 )
The mean and the standard deviation of the logarithm of fatigue life can be obtained from Equations (19) and (20). The logarithm of fatigue life with an
Table 2 Test results according to the conventional method
l o g Nu.841(S)
(25)
Therefore, the standard deviation of the logarithm of fatigue life as a function of the stress amplitude is
Stress amplitude S.
(MPa) 431.64 392.40
Sample size
2 4
Mean
4.3789 4.6174
Standard deviation
[).026~ 0.0626
372.78
4.7065
0.0932
353.16
4.8690
().175~
333.54
5.1395
[).1157
313.92
5.4019
().1027
418
Table 3
Difference of test results and CM results (%) -1.510 -0.039 1.530 1.670 -0.150 -1.435
450
450
'
'
'
'
'
'
'
'
'
'
'
'
\ ~"
\ ~ l ~
- -
--
-Maximum
Likelihood
Method
- -
Conventional
Method
a 400
Experiment
e~
400
<.~,~ 3 5 0
\ ~ ~
Q
<
300
,,,
4.2
4.4
5.4
i,,, 4.4
~,,,
4.6
~,,,
4.8
~,,,
5
, , ,
5.2 5.4
Logarithm of Fatigue Life F i g u r e 3 A comparison of the logarithm of fatigue life with p = 0.841 obtained by the two methods and test results
F i g u r e 2 A comparison of the means of the logarithm of fatigue life obtained by the two methods and test results
(26)
The logarithm of fatigue life with an arbitrary survival probability p can be calculated by /x(S) -/xpo-(S). Table 3 shows a comparison between the means of the logarithm of fatigue life obtained by the two methods and test results. It can be seen that the maximum difference between the means of the logarithm of fatigue life obtained by the maximum likelihood method and the test results is 1.033%, whereas the maximum difference between the means of the logarithm of fatigue life obtained by the conventional method and the test results is 1.670%. A comparison is also shown in Figure 2. Table 4 shows a comparison between the logarithms of fatigue life with survival probability p--0.841 obtained by the two methods and test results. It can be seen that the maximum difference between the logarithm of fatigue life with survival probability p - - 0 . 8 4 1 obtained by the maximum likelihood method and the test results is 1.836%, and the maximum
Table4
difference between the logarithm of fatigue life with survival probability p--0.841 obtained by the conventional method and the test results is 3.231%. A comparison is also shown in Figure 3. From Tables3 and 4 and Figures2 and 3, it is clear that the maximum likelihood method and the conventional method give nearly the same P-S-N curves. However, 27 specimens were used according to the conventional method, whereas only 13 specimens were used according to the maximum likelihood method. Thus, the maximum likelihood method requires a smaller number of specimens and consequently less testing time than the conventional method does but with a reasonable accuracy for engineering applications. CONCLUSIONS AND DISCUSSIONS We have presented a maximum likelihood method as a practical engineering method for estimating P-S-N
A comparison of the logarithm of fatigue life with p = 0 . 8 4 1 obtained by the two methods and lest results Difference of test results and MLM results (%) -1.836 -1.587 0.000 1.651 -I.091 -I.245
Difference of test results and CM results (%) -1.952 -0.290 1.648 3.231 -0.281 -2.168
419
Nakazawa, H. and Kodama, S., Statistical S-N Testing Method with 14 Specimens: JSME Standard Method for Determination of S-N Curves. Statistical Research on Fatigue and Fracture, Elsevier Applied Science, New York, 1987. pp. 59-69. Nishijima, S., Statistical Analysis of Small Sample Fatigue Data. Statistical Research on Fatigue and Fracture, Elsevier Applied Science. New York, 1987, pp. 1-19. ASTM, Standard Practice for Statistical Analysis of Linear or Linearized Stress-Life (S-N) and Strain-Life (e-N) Fatigue Data, ASTM Standard E739-91, American Society for Testing and Materials, Philadelphia, 1991. Weibull, W., Fatigue Testing and Analysis of Results, Macmillan Company, New York, 1961.