Professional Documents
Culture Documents
Paper - Fatigue Lifetime of Welded Joints Under Random Loading - Colombini P., Dolinski K. - 2001
Paper - Fatigue Lifetime of Welded Joints Under Random Loading - Colombini P., Dolinski K. - 2001
www.elsevier.com/locate/probengmech
Fatigue lifetime of welded joints under random loading: rainflow cycle vs.
cycle sequence method
P. Colombi a,*, K. Doliñski b
a
Department of Structural Engineering, Milan Technical University, Piazza L. da Vinci 32, 20133 Milan, Italy
b
Institute of Fundamental Technological Research, Polish Academy of Sciences, ul. Swietokrzyska, 21 00-049 Warsaw, Poland
Received 1 March 1999; received in revised form 1 December 1999; accepted 1 February 2000
Abstract
The evaluation of fatigue lifetime is a complex problem due to the loading cycle sequence effect influencing the fatigue damage process. In
the literature two different approaches are usually adopted to solve the fatigue crack growth under time varying loading: the rainflow cycle
(RFC) and the cycle sequence (CSQ) model. In this paper the performance of the recently proposed stochastic version of these methods is
compared and applied to the evaluation of the fatigue lifetime of the welded cruciform tested in the literature. Non-Gaussianity and
bandwidth effect are correctly captured by both methods but only the CSQ model enables us to examine the influence of the load sequence
effect on the fatigue lifetime. Numerical results show that the retardation effect can be neglected for Gaussian and non-Gaussian loading as
the bandwidth increases. Finally, this work shows the greater performance and flexibility of the CSQ model compared to the RFC model.
䉷 2000 Elsevier Science Ltd. All rights reserved.
Keywords: Stochastic fatigue; Random loading; Cycle counting methods; Fatigue retardation; Fatigue reliability
account through the rainflow cycle algorithm itself. In the stress range is a crucial step in the fatigue lifetime evalua-
CSQ model the sequence effect can be considered explicitly tion. Among possible techniques to identify DSi [26], the
in fatigue crack propagation model. rainflow counting method is more often used in the literature
In a simplified analysis of the structure subjected to [4]. The damage in a single cycle can then be summed up by
stochastic loading only the mean value of the fatigue life- the Miner’s rule [5] over the total number of cycles N to
time is often evaluated while the external loading is evaluate the total damage D:
assumed to be Gaussian and narrow-banded. Due to the
X
N XN
DS m
i
non-linear behaviour of the structure and non-normal char- D Di
3
acter of the external loading the structural response may be i1 i1
K
highly non-normal and the frequency content of the
and the failure takes place when D 1: Under random
response power density function (PSD) not necessarily
loading the damage parameter D is a random variable and
narrow-banded. Recently Rychlik [24] and Frendahl et al.
is often approximated as a sum of the independent random
[18] have shown on a very wide numerical simulation basis
variables having the same probability distributions. The fail-
that a homogeneous Markov chain is a very good approx-
ure probability PF is then defined as:
imation of the random sequence of extremes of the station-
"N #
ary Gaussian and non-Gaussian processes with various XF DS m
i
bandwidth properties. The approximation of the random PF P ⬎1
4
i1
K
sequence of extremes of the loading by a homogeneous
Markov chain was the starting point for the development where N F is the random number of cycles to failure. It is
of two stochastic methods for the estimation of fatigue life- well known that the mean value E[NF] is given by:
time under random loading. Under the Markov chain 1 K
approximation of extremes of the external loading the prob- EN F
5
DS m
i EDS m
i
ability distribution of the rainflow cycles is evaluated in E
Refs. [2,9,24]. The Miner’s rule [5] is then used to compute K
the mean value of the fatigue lifetime. The CSQ method to and, for a great N F EN F value, the probability distribu-
compute the fatigue lifetime was employed in a mixed, tion of the random number of independent random variables
partially analytical, partially numerical approach originally in such a sum can be approximated by the inverse Gaussian
suggested in Ref. [7] and modified and improved in Refs. probability distribution [13–15].
[13–15]. The approach eventually leads to the analytical For narrow-band Gaussian processes the stress range
form of probability distribution of the fatigue lifetime and follows the Rayleigh probability distribution and then the
can be extended to compute the fatigue reliability under close form solution is available for E[NF] in Eq. (5):
random material properties and crack size and geometry [6,8].
K
The goal of this paper is to compare the fatigue lifetime EN F
6
p m
evaluation methods sketched above with reference to the
2 2·s s m · !
fatigue experiment results presented in Refs. [29,30]. The 2
capability and the performance of the two methods to where s s is the mean square root of the loading process, cf.
capture the non-Gaussianity, bandwidth parameter and Rychlik [25]. If the loading is not narrow-banded a coeffi-
retardation effects on the fatigue lifetime are finally illu- cient d is usually introduced:
strated and discussed. EN F wide-band loading
d
7
EN F narrow-band loading
2. The RFC method
The neglecting of the bandwidth effect produces conserva-
2.1. Fatigue lifetime evaluation tive results, i.e. d ⬎ 1 [20].
An additional coefficient g must be applied in the case
Under constant amplitude loading the fatigue results are when the process in non-Gaussian, namely:
usually represented by a S–N curve [5]:
EN F Gaussian loading
N K·DS⫺m
1 g
8
EN F non-Gaussian loading
where DS is the stress range while K and m are material The neglecting the non-Gaussianity produces conservative
constants. Under variable amplitude loading the expected results, i.e. g ⬍ 1; if the kurtosis of the loading process is
fatigue damage per cycle, Di, corresponding to a stress less than 3 and non-conservative results, i.e. g ⬎ 1; if the
range, DSi, can be determined as [5]: kurtosis of the loading process is greater than 3. The final
1 DS m expression for the estimation of the fatigue lifetime under a
i
Di
2 wide-band non-Gaussian process is then obtained by multi-
Ni K
plying Eq. (6) by d=g: In the literature several methods were
Under variable amplitude loading the identification of the proposed to estimate the coefficients d and g [4,6]. They
P. Colombi, K. Doliñski / Probabilistic Engineering Mechanics 16 (2001) 61–71 63
include some complicated expressions involving the irregu- … ⬍ un⫺1 ⬍ un ⬍ ⫹∞ one can define the discrete
larity factor and the central or Hermite moments of the sequence of the extremes M dk as:
loading process and the fatigue material parameters. The
proposed stochastic RFC method gives directly the fatigue Mkd ui if ui ⱖ Mkd ⬎ ui⫹1
11
lifetime without the need of any correction coefficient. Assuming that M dk
is a one step Markov chain defined by
the marginal distribution of a maximum M d0 ; P
M0d um ;
2.2. Distribution of the rainflow cycle stress range and the two transition probabilities matrices, the first one
In a general situation, cf. Eq. (5), the evaluation of from the maxima to the following minima P⫹
k; m and the
EDS im is crucial. In Ref. [23] a new definition of the second one from the minima to the next maxima P⫺
k; m;
classical rainflow counting algorithm was presented. From respectively:
Fig. 1 it is seen that the rainflow cycle range DS(t), with P⫹
k; m P
M1d um 兩M0d uk
12
reference to a maxima S
t M0 u; is given by:
DS
t minDH ⫹
t; DH ⫺
t
9 P⫺
k; m P
M2d um 兩M1d uk
In Fig. 1 t ⫺ and t ⫹ are the times of the last and first down- the conditional probability distribution of the discrete rain-
and up-crossing of the level u before and after t and flow cycle range is:
M⫺1 ; M1 … are the adjacent local minima as well as
M⫺2 ; M2 … are the adjacent local maxima. PDSd 兩 M d0
ui ⫺ uj 兩ui 1 ⫺ p2
ui ; uj 兩ui
13
The definition given in Eq. (9) and Fig. 1 is very conve- Details on the evaluation of p
ui ; uj 兩um are given in Refs.
nient [3,24] for the evaluation of the conditional probability [2,9,24]. As the conditional distribution PDSd 兩 M d0
h兩u is
distribution PDS 兩 M0
h兩u of the RFC range DS given M0 u: computed the unconditional distribution PDS
h of the
Such a conditional distribution is in fact the solution of the RFC range is calculated as follows:
following first passage problem of the sequence of local
maxima and minima Mk: X
n
PDS d
h PDS d 兩 M d
h兩ui P
M0d ui
14
0
PDS 兩 M0
h兩u 1 ⫺ P
M k crosses u ⫺ h before it crosses u as k i0
approximately statistically independent, crack length incre- the same probability distribution:
ments. Considering only one retardation block with the
X
MF
crack length increment, B(aol), and with a number of stress NF N Bi
22
cycles, NB(aol), the increment of the fatigue damage para- i1
meter, G B(aol), within a retardation block can be written, cf.
The number of blocks to failure, M F ; is usually suffi-
Eq. (17), as follows:
ciently great to apply the central limit theorem modified
Zaol B
aol da NBX
aol for a sum of a random number of random variables [13–
G B
aol DG
S n⫺ ; S⫹
n ; an 兩aol
18 15]. Thus, the probability distribution of the number of
aol g
a n1
cycles to failure N F ; can be approximated by the Gaussian
As mentioned before the strong inequalities B(aol) p aol probability distribution:
appears to be satisfied for any aol from a crack length inter- " #
n ⫺ M F ·N B
val [a0,aF] where a0 and aF denote the initial and admissible FN F
n PN F ⱕ n ⬇ F
23
ultimate crack lengths, respectively. Referring to the reset sN F
stress definition it is easily seen that the effect of the reset with mean value and variance:
stress depends on the ratio aol/a with aol ⱕ a ⱕ aol ⫹ B
aol ;
rather than on the location, aol, of the block within the inter- EN F M F ·N B
24
val [a0,aF] alone. Then, the number of cycles, NB, and the
gF
damage parameter increment, G B, in a retardation block can s N2 F M F ·s N2 B ⫹ s M2 F ·N B 2 ·
s N2 B ⫹ n G2B ·N B 2
be considered as independent of aol and P B;N B
bi ; n兩aol GB
PB;N B
bi ; n: Moreover, by making use of Eq. (18) it is possi- Additional calculation of the mean and standard deviation
ble to transform the joint probability distribution of (B, NB) of the time to failure are given in [13–15]. Moreover, the
into the joint probability distribution of (G B, NB). Statistical stochastic CSQ model was extended to the case when a
moments G Bk and N kB of the kth order of the damage para- sequence of loading trains of quite different statistical char-
meter, G B ; and of the number of cycles, N B ; within a single acteristics is applied to the structure [16]. It corresponds to
retardation block, B, are then easily calculated: the situations often met in application when a structure
X
∞ X
∞ subjected to stochastic loading has to sustain transient over-
G Bk g ki PG B ;N B
gi ; n loading periods of random duration and occurring occasion-
i1 n1 ally (storm for offshore structures, strong winds for high
19 buildings, slender columns or masts, take-off and landing
X
∞ X
∞
N Bk n k
PG B ;N B
gi ; n of aeroplanes, changing of operation modes of machinery
n1 i1 and so on).
The evaluation of the fatigue lifetime distribution given
The fatigue damage parameter, G
M after M blocks is
by Eq. (23) must be carried out for a given realisation of the
then the sum of the M independent random variables having
random vector of additional parameters x (material proper-
the same probability distribution. Until the fatigue failure,
ties, crack and structural detail geometry and so on). The
the fatigue damage parameter should be less than a critical
unconditional probability distribution of the lifetime, FNF ;
value, i.e. G F
M F ⬍ gF ; corresponding with the critical
can be calculated by a direct integration:
crack length, aF:
ZaF da Z
gF
20 FN F FN F
n兩x·fx
x dx
25
a0 g
a
where a0 is the initial crack length. Thus, the probability of where fx
x is the probability density function of the para-
failure is defined as: meters vector x or by using some approximation methods of
"M # reliability analysis [6,8].
XF
PF PG F
M F ⬎ gF P G Bi ⬎ gF
21
i1
4. Numerical example
where M F denotes the random number of blocks to failure.
For a great value of MF EM F ; the probability distribu- The cruciform joint, cf. Fig. 4, realised by a vertical leg
tion of the random number of independent random variables and two horizontal stems attached to the mid section by full
in such a sum can be approximated by the inverse Gaussian penetration fillet welds was extensively tested in [27–30]
probability distribution [13–15] with the mean M F gF =G B under both Gaussian and non-Gaussian stochastic loading.
and variance s M 2
F
s G2 B ·gF =G B3 nG2B ·M F : The total The specimen was excited axially in the vertical direction
number of cycles to failure, N F ; is then given as the sum and the experiments were ended when the current stiffness
of a random number M F of random variables N Bi (the of the specimen was one half of the original one. Each test
random number of cycles within a retardation block) having was repeated up to 5 times and the geometric mean was
66 P. Colombi, K. Doliñski / Probabilistic Engineering Mechanics 16 (2001) 61–71
Table 1 Table 3
Fatigue lifetime for gaussian wide-band process (unretarded solutions are in Fatigue lifetime for non-gaussian wide-band process
k 5 (unretarded
brackets) solutions are in brackets)
RFC method CSQ method Experiments RFC method CSQ method Experiments
where a is the scaling factor. These loading processes had k 2: All fatigue lifetime assessments (in duty cycles)
upward zero crossing rates of 0.20 and 0.39 occurrences/s for the welded cruciform joint, cf. Fig. 4, are shown in
and a peak rate of 0.24 and 1, respectively. Different levels Tables 1–6 below. The reference experimental results in
of root mean square (RMS) of the loading were used in form of the geometric means of the sample lifetimes are
order to investigate the effect of the load intensity on the quoted from Refs. [29,30] in the last column of each
resulting fatigue lifetime. table.
The non-Gaussian loading histories, Z
t; were obtained
by the following non-linear transformation of the corre-
sponding Gaussian loading histories X
t : 4.1. The RFC method results
b
sgn
X
兩X兩n The fatigue lifetimes obtained from the RFC method are
Z
t G
X
t X ⫹
35 given in the second column of each table. As the cycle
C
sequence effect cannot be taken into account explicitly in
where b and n are the appropriate transformation para- this method, no comparative information on the unretarded
meters and C is introduced to force the condition s x s z : fatigue lifetime is available.
The kurtosis: The RFC results seem to agree quite well with the experi-
mental ones. Under wide-band Gaussian loading the
E
Z ⫺ mz 4 maximum deviation is less than 20%. In the wide-band
kz
36
s z4 non-Gaussian cases the error is less than 35%. For the
narrow-band loading the deviation between the numerical
is a measure of the non-Gaussianity of the loading (for a and experimental results appears to be less than 15% for the
Gaussian process it is exactly equal to 3). Gaussian and non-Gaussian stochastic processes. Only the
The following sets of parameters (b , n, C) were adopted case with Gaussian loading and RMS 69 MPa produces a
in order to obtain two different values of the kurtosis k : larger error. The large disagreement between the experi-
(0.342, 2., 1.563) for k 5 and (1.735, 0.5, 2.257) for mental and numerical results for the non-Gaussian process
with RMS 103:5 MPa may be attributed to a fatigue rate
acceleration effect of large amplitude cycles occurring in
Table 2 that case of the loading when the stress maxima approach
Fatigue lifetime for non-gaussian wide-band process
k 2 (unretarded
solutions are in brackets)
the yield strength of the material.
Table 4
Fatigue lifetime for gaussian narrow-band process (unretarded solutions are in brackets)
RMS 34.5 (MPa) 5:4739 × 106 4:7680 × 106
3:0177 × 106 4:7097 × 106
RMS 51.75 (MPa) 1:2904 × 106 1:2546 × 106
7:7788 × 105 1:2441 × 106
RMS 69 (MPa) 4:6284 × 105 4:8026 × 105
3:0837 × 105 6:1160 × 105
RMS 103.5 (MPa) 1:0910 × 105 1:1820 × 105
7:1395 × 104 1:1750 × 105
RMS 155.25 (MPa) 2:5710 × 104 2:5908 × 104
1:7004 × 104 1:4400 × 104
case and the RMS level equal to 69 MPa. For the wide-band fatigue lifetimes produced by the Gaussian loading are
non-Gaussian loading the error was less than 15% except the longer than for the non-Gaussian loading with the same
bimodal spectrum case and the kurtosis k 2: For the RMS level and the kurtosis k 5: The reverse is true for
narrow-band loading a very good agreement with experi- non-Gaussian loading with the kurtosis k 2: It implies
mental results with an error less than 5% is observed for the that a special attention must be always paid when evaluating
Gaussian and non-Gaussian processes. Only the case with the fatigue lifetime of a structure under non-Gaussian
non-Gaussian loading and a RMS 69 MPa produces a loading. In fact, some traditional fatigue analysis techniques
larger error. Again the large disagreement between the experi- such as the Rayleigh approximation do not include this
mental and numerical results for the non-Gaussian process non-normality effect [4]. It produces a non-conservative
with RMS 103:5 MPa may be attributed to a fatigue rate fatigue lifetime estimation when the loading kurtosis is
acceleration effect of large amplitude cycles in this case. greater than 3. The results also reveal that with an increasing
It is very interesting to compare the unretarded fatigue bandwidth of the loading the fatigue lifetime increases as
lifetimes with the unretarded ones. As the bandwidth of well. Note that a simple dividing of the number of cycles to
loading becomes wider the retardation effect appears to failure by the peak rate drastically reduces the time to failure
become less significant. It is valid for both Gaussian and as the bandwidth increases for both the Gaussian and non-
non-Gaussian loading and results from a reduction of the Gaussian loading [30]. It means that the high frequency
mean length of the retardation block. As an example the components are important because they accelerate the rate
increment of the fatigue lifetime due to the retardation effect of fatigue damage accumulation. Moreover, this support the
is less than 9% for both Gaussian and non-Gaussian bimodal hypothesis [21], that an endurance fatigue strength limit
wide-band loading. For the narrow-band loading where the observed under constant amplitude loading does not exist
cluster effect produces large retardation blocks the differ- (or at least is suppressed to a lower level) under variable
ence between the retarded and unretarded lifetime values amplitude loading.
exceeds by 75% for the non-Gaussian loading with the Two different approaches to the problem of fatigue
kurtosis k 5: Note that neglecting the retardation effect crack growth under random loading have been presented
makes the fatigue crack propagation a memoryless and discussed in this paper. The stochastic RFC method
process. It makes also the whole analysis of the fatigue makes a use of the well-known rainflow cycle algorithm
process and the evaluation of the fatigue lifetime much to identify the stress cycles from the loading history. The
more simple [10]. Markov chain assumption for the sequence of extreme of
the random loading allows one to compute the probability
distribution of the rainflow stress range. Miner’s rule is
5. Concluding remarks then used to evaluate the damage accumulation and
finally the fatigue lifetime. In this approach the sequence
The examination of the numerical results indicate that the effect is only implicitly taken into account through the
Table 5
Fatigue lifetime for non-gaussian narrow-band process
k 2 (unretarded solutions are in bracket)
Table 6
Fatigue lifetime for non-gaussian narrow-band process
k 5 (unretarded solutions are in bracket)
RMS 34.5 (MPa) 3:4922 × 106 3:3197 × 106
1:8837 × 106 3:1699 × 106
RMS 51.75 (MPa) 8:7448 × 105 8:8040 × 105
5:0434 × 105 8:7450 × 105
RMS 69 (MPa) 3:1366 × 105 3:1764 × 105
1:6853 × 105 3:030 × 105
RMS 103.5 (MPa) 7:3939 × 104 7:1242 × 104
4:0547 × 104 2:930 × 104
random overloading pulse train. Comp Meth Appl Mech Engng [25] Rychlik I. On the narrow band approximation for the expected fatigue
1999;168:221–41. damage. Prob Engng Mech 1993;8:1–4.
[17] Dolinski K. Effect of random over- and underloads on fatigue life- [26] Rychlik I. Note on cycle counts for irregular loads. Fatigue Fract
time. ZAMM 1999 (in press). Engng Mater Struct 1993;16(4):377–90.
[18] Frendahl M, Rychlik I. Rainflow analysis—Markov method. [27] Sarkani S, Kihl DP, Beach JE. Stochastic fatigue damage in welded
Research Report, Department of Mathematical Statistics, University cruciform. In: Spanos, et al., editors. Computational stochastic
of Lund, 1992. mechanics, Amsterdam: Elsevier, 1991. p. 461–72.
[19] Hobbacher A. Stress intensity factor of welded joints. Engng Fract [28] Sarkani S, Kihl DP, Beach JE. Fatigue of welded joints under
Mech 1993;46(2):173–82. narrowband loadings. J Engng Mech (ASCE) 1992;116:2481–
[20] Lutes LD, Larsen CE. Improved spectral method for variable ampli- 95.
tude fatigue. J Struct Engng 1990;116(4):1149–64. [29] Sarkani S, Kihl DP, Beach JE. Fatigue of welded joints under
[21] Maddox SJ. Fatigue strength of welded joints. Cambridge: Abington narrowband non-Gaussian loadings. Prob Engng Mech 1994;9:179–
Publishing, 1991. 90.
[22] Peeker E. Extended numerical modelling of fatigue behaviour. PhD [30] Sarkani S, Kihl DP, Beach JE, et al. Stochastic fatigue of welded steel
thesis, Institute of Steel Structures, Polytechnic of Lausanne, No. joints. In: Guedes-Soares C, et al., editors. Offshore mechanics and
1617, 1997. arctic engineering’95, New York: ASME, 1995. p. 199–206.
[23] Rychlik I. A new definition of the rainflow cycle counting method. Int [31] Suresh S. Fatigue of materials. Cambridge: Cambridge University
J Fatigue 1987;9:119–21. Press, 1991.
[24] Rychlik I. Simple approximation of the rainflow cycle distribution for [32] Veers PJ. Fatigue crack growth due to random loading. Research
discretized random loads. Prob Engng Mech 1989;4:40–8. Report, Sandia National Laboratories, SAND87-S2037, 1987.