You are on page 1of 10

RAINFLOW CYCLES IN RANDOM LOADS

Igor Rychlik
Dept. of Math. Statistics, Univ. of Lund, Box 118, S-22100 Lund, Sweden

Abstract: A new approximation is developed for the expected fatigue


damage from the Miner-Palmgren rule and the rainflow cycle counting
method for random loads. The approach is based on an exact relation-
ship between the rainflow cycle count and the range-pair exceedances
count. This links the intensity of rainflow amplitudes to expectations
of marked crossings of the random load. The method is best suited for
the fatigue life prediction for Gaussian loads with slowly, or periodically,
varying means and covariance functions.

Introduction
The rainflow cycle counting method (RFC) is generally considered to be the
best approach for estimating the fatigue damage caused by randomly fluctuating
loading conditions, see Endo and Matsuishi [1] and Dowling [2]. A commonly
used damage measure D(t) in (0,£), developed by Palmgren and Miner, is
k(t) 1

m-E^ (i)
where the sum is extended over all cycles completed at time t, and Ns is the
cycle life obtained from S-N data from tests with the constant amplitude s. The
RFC-counting method can be used in Equation (1) to identify the amplitudes
and number of stress cycles.
In previous papers [3, 4, 5], we have presented a simple formulation of the
RFC count which lends itself to a mathematical analysis and a sequence of ap-
proximations to the RFC-distribution for ergodic loads, based on a Markov chain
approximation for the sequence of local maxima and minima. (This approach
was also used by Bishop and Sherratt [6].)
A different approach to the RFC-distribution, presented in Ford [7], was
based on the range pair exceedances to count cycles. It was postulated that this
approach is equivalent to the RFC-count. In Rychlik [8], the exact relationship
between these two counting methods was given and an exact formula for the
mean number of range pair exceedances with fixed amplitude and mid-value
was presented. Here, the formula is used to construct accurate approximations
for the expected damage in the case of Gaussian loads. A numerical example
for a simple non-stationary load illustrates the result.

21
SIGNAL *

y(x)

■k

t. i
t. =t t. TIMS
i i r l

Figure 1: Definition of a rainflow cycle.

Rainflow cycle counting m e t h o d


Many different algorithms for counting RFC-cycles have been proposed in the
literature. However, most of them have a complicated "sequential" structure
which makes them difficult to apply when their statistical properties are stud-
ied. The following definition of RFC-cycle, given in [3], is more convenient for
statistical analysis of RFC-cycles.
Definition 1 (Toplevel-up cycle count.) Let y(r), 0 < r < T, be a load func-
tion, and let {tk}, with 0 < t0 < t\ < . . . , be the times of the local maxima of
y(-). The toplevel-up count attaches to each maximum a cycle defined as follows.
For a local maximum at time ti, let tf be the time for the first upcrossing after ti
of the level y(ti) (or tf = T if no such upcrossing exists for t{ <r < T), and let
t~ be the time for the last downcrossing of y(t{) before ti (or t~ = 0 if no such
downcrossing exists for 0 < r < t). Suppose, the lowest minima in the intervals
(tj,ti) and (ti,tf) occur at t\, tr, respectively, and denote by t* the time when
the higher of the minima y(ti), y(tr) occurs, i.e.
f tr if y(U) < y(tT),
t* = { (2)
[ ti otherwise.
The toplevel-up cycle originating at t{ is defined as a pair of the maximum y(t{)
and the minimum y(t*) and the amplitude of the cycle is given by
Si = y(U) - y(t*), (3)
see Figure 1.

Range pair exceedance counting m e t h o d


Briefly, see Ford [7], a range pair exceedance (rpe) of levels u and t>, u > v, is
completed when the load process y exceeds an upper threshold u, drops below

22
SIGNAL

TIME

Figure 2: Definition of a u-crossing amplitude.

to a lower level t;, and returns to the level u. The first cycle begins at the first
upcrossing of the level u by the load y. In our approach we shall slightly modify
the definition of the range pair exceedances of the levels u,t>, by attaching to
each downcrossing of the level u, by the load ?/, a u-crossing amplitude, defined
as follows.
Definition 2 (u-crossing amplitude count.) Let y(r), 0 < r < T, be a load
function, and let {T^}, with 0 < To < T\ < ..., be the times of the downcrossings
of the level u by ?/(•). For a u-downcrossing at time n, let T* be the time of the
first upcrossing after r,- of the level u (or T* = T if no such upcrossing exists for
T{ < T < T). The u-crossing amplitude originating at rt- is defined by
S? = u- min y(r) (4)
T.<T<T +

see Figure 2.
Observe that the number of range-pair exceedances of the levels u, v is equal
to the number of u-crossing amplitudes Su higher than u — v. Consequently, the
statistical properties of the rpe-method can be analyzed using the u-crossing
amplitudes.
Definition 3 Let y(r), 0 < r < t, be a load function. For fixed values u,v,
u > v, let Nrfc(u,v,t) be the number of rain flow cycles {(y(ti),y(t*))} counted
in y, such that the maximum of the cycle is higher than u and the attached
rainflow minimum is lower than v, i.e.
Nrfc(u, v; t) = #{ti] 0 < U < t such that y{U) > u and y(t*) < v}, (5)
where {t{} is the sequence of the local maxima times of y.
Further, let N(u,v;t) denote the number of u-crossing amplitudes S", such
that S? > u — v, i.e.
N(u, v; T) = # { r i ; 0 < r{ < T such that S? > u - v } , (6)
where {r t } is the sequence of u-downcrossing times ofy(r).

23
In [8] Lemma 4, we proved that, for any T and u,v fixed,

N(u, v; T) - 1 < iV r / c (u, v\ T) < N{u, v; T ) , (7)

which shows the equivalence of the rfc- and rpe-counting methods.

Expected damage
For fixed values u > v, consider the mean number of rfc-cycles counted in t/,
E[Nrfc(u,v;t)], such that the maximum of the cycle is higher than u and the
rfc-minimum is lower than v. Under some mild regularity assumptions on the
load y, the expected damage can be written as follows

E v
^ = - C L ir, ■ ikw^ > «M dvdu- w
Although, Equation (8) holds for smooth loads, the class of all non-stationary
loads is too wide to give one universal method to compute the expectation
E[D(t)]. In this paper we shall consider only smooth and locally stationary
loads, defined as follows.

Definition 4 The twice continuously differentiable load will be called locally sta-
tionary if the following conditions are satisfied;
(a) For any time t and all pairs (u,v) £ R2, u > v, E[NT*c(u,v\i)\ —
E[N(u,v;t)] and the following derivative exists

n(u,v;t) = -E[N(u,v;t)}. (9)

Further, the partial derivative -j^7^fi(u,v\t) is a continuous function of(u,v).


(b) The load y is a decomposable process, see Definition 1 in [9] for the
definition of decomposable processes.

Condition (a) is imposed to guarantee the equivalence of the rfc- and rpe-
counting methods, while Condition (b) specifies a class of processes for which an
explicit expression for the intensity /i(u, v\ i) can be given, e.g. the Rice's formula
for the expected number of level crossings is valid. The class of decomposable
processes is quite large and contains for example Gaussian processes, functions of
Gaussian vector processes, the sum of a Gaussian process and any independent
continuously differentiable process.
We turn now to the explicit formula for the damage intensity d(t) defined,
say, as the time derivative of the expected damage. Using (8), the damage
intensity can be written as

m = ±E[D(t)] = -CLiL- ^(^)<^ (10)


where Ns is obtained from S-N data. In applications, the function Ns is usually
given in explicit form, e.g. Ns = a • ,s - / ? , a > 0, /? > 1. If the function

24
Ns is known explicitly, it is more convenient to rewrite Equation (10), using
integration by parts, and replace the derivative of fi by derivatives of ^ - . More
exactly, let g(s) = ~ be a twice continuously differentiate function. If #(0) = 0
and
+oo ru 0"
/ (ii)
/ \g'(u -v)\ • -^-z-n(u,v;t)dvdu < +oo,
-oo ./-00 OUOV
then the damage intensity d(t) is given by
+oo ru
/ g"(u-v)n(u,v;t)dvdu, (12)
/ -oo J—oo

where the intensity fi(u,v,t) is given by


+oo
E[y'{t)~y\s)-Iis{y)\y{t) = u,y{s) = v]fy{t)Ms){u,v)ds. (13)
/
Here x~ = max(0, —#), Its(y) is the following indicator function

1 if u > y(r) > v for all r, t < r < 5,


(14)
!
0 otherwise,
and fy(t),y(s)(uiv) i s a density of y(t),y(s).
The proof of (12-14) is given in [8]. (Note that, for loads with finite damage
intensity, \i Ns = a - s~^, then Condition (11) is satisfied.) Equation (12),
links the computation of the damage intensity d(t) to the intensity /z(u,t;;2),
which is given in an explicit way. However, since the indicator Its{y) depends
on the whole sample path of the load y in the interval (/,s), the calculation of
the intensity fi(u,v;t) is difficult. The approximation of fi(u,v]t) for Gaussian
loads is briefly described in the Appendix.

Numerical example
Let y(t) be a Gaussian load with mean E[y(t)] = sm(t) and a stationary covari-
ance function given by

sin(x/3|.s-*|)
Cov(y(t),y(s))= Xr—T - < 15 )
3
V I*-*I
Figure 3 (a) shows the level curves of the approximation fi(u,v; 1.57), while
Figure 3 (b), illustrates the damage intensity d(t) for the load y. Since the
covariance of y is stationary and the mean varies periodically, the expected
damage E[D(T)] can be easily obtained, for any T, by numerical integration of
the damage intensity. The approximation of expected damages during a period
of the mean, E[D(6.2S)] are: 11.7, 50.7, 247.4, 1325.1, for ^ = sp, p = 2 , 3 , 4 , 5 ,
respectively. Corresponding estimated damages (with 95 % confidence bounds)
from 100 realizations of y are: 11.2 ± 0 . 0 5 , 48.1 ± 0 . 3 , 232.0 ± 2 . 0 , 1228.6 ± 1 7 . 8 ,
respectively. The expected damages E[D(6.28)], for a zero mean Gaussian load

25
(a) (b)

750 n

500

250

Figure 3: Isolines of the approximate intensity /i(iz, v\ 1.57) (a); the damage
intensity d(t), for ~ = s5 (b); for the non-stationary Gaussian load with covari-
ance function (15) and mean s'm(t).

with the covariance function (15) are: 7.2, 25.5, 105.3 and 483.2, respectively.
The corresponding simulated damages (with 95 % confidence bounds) are: 7.2 ±
0.05, 25.5 ± 0.25, 103.8 ± 1.4, 469.4 ± 8.5, respectively. Results shows that the
expected damage strongly depends on the mean of the load. It can also be noted
that the theoretical results slightly overestimate the simulation calculation for
P = 5 in the stationary case. There are two reasons for the simulation results
to be slightly unconservative: the finite sample of the load and the discrete
representation of it. Note also that the theoretical analysis in the non-stationary
case is much more complex as demonstrated in the Appendix. For example, the
expected damages for stationary Gaussian load with covariance (15) computed
using the general algorithm for non-stationary loads are: 7.4, 26.4, 109.2, 501.8,
respectively, rather than 7.2, 25.5, 105.3, 483.2, obtained by the algorithm for
stationary loads. The difference between these values are 3-4%.

References
[1] Matsuishi M. and Endo T. Fatigue of metals subject to varying stress, paper
presented to Japan Soc. Mech. Engrs, Jukvoka, Japan 1968.
[2] Dowling N.E. Fatigue prediction for complicated stress-strain histories. J.
Materials, 1972, 7, 71-87.
[3] Rychlik I. A new definition of the rainflow cycle counting method. Internat.
J. Fatigue, 1987, 9, 119-121.
[4] Lindgren G. and Rychlik I. Rainflow cycle distribution for fatigue life pre-
diction under Gaussian load processes. Fatigue Fract. Engng. Mater and

26
Struct, 1987, 10, 251-260.
[5] Rychlik I. Rain flow cycle distribution for ergodic load processes. SIAM J.
Appl. Math., 1988, 48, 662-679.
[6] Bishop N. W. M. and Sherratt F. A theoretical Solution for the estimation
of "rainflow" ranges from power spectral density data. Fatigue Fract.
Engng. Mater and Struct, 1990, 13, 311-326.
[7] Ford D. Range-Mean-Pair exceedaces in stationary Gaussian processes. Re-
liability and Optimization of Structural Systems, Aalborg, Denmark,
Springer-Verlag, 1988, 119-139.
[8] Rychlik I. Rainflow cycles in Gaussian loads. Fatigue Fract. Engng. Mater
and Struct, 1991, to appear.
[9] Rychlik I. The two barrier problem for continuously differentiable processes.
Adv. Appl. Prob., 1992, to appear 1-34.

Appendix: approximation of the intensity fi(u,v]t) under


Gaussian loads
Consider continuously differentiable Gaussian load y, and let fi(u,v;t,s) denote
the integrand in (13), i.e.

fi(u,v,t,s) = E[y'(t)-y\s)-Its(y)\y(t) = u,y(s) = v]fy{t)Ms)(u,v), (16)

and
+oo
/ fi(u,v;t,s)ds. (17)
As we have mentioned before, the evaluation of (16-17) is intractable, since
the indicator Its(y) depends on the whole sample path of the load y in the interval
(t,s). (An additional difficulty is the infinite region of integration (£,+co) in
(17).) However, by approximating the integral Its{y) appearing in (16) one can
construct approximations to the function fi(u,v;t,s). For example, a simple
approximation to fi(u,v;t,s) is obtained by replacing the indicator 7*5(y), in
(16), by one, i.e.

H(u,v; t, s) « E[y'(t)-y'(s)-\y(t) = u,y(s) = v]fy{t)Ms)(u,v). (18)

Here we shall use the so called regression method, see [9], to approximate
the indicator. The main idea of the regression method is to find n random
variables Yi,...,Yn, containing the relevant information about the indicator
Its, and such that the joint density of y(t),y(s),y'(t),y,(s), Y\,..., Yn can be
computed. Then one approximates the load y(s\) by its regression curve yn{si),
i.e. yn(Sl) = ^ [ y ^ O l y ' W . n , . . . , ^ for n > 0, while y0(s1) = E[y(s1)\y'(t)].
Next, the intensity fi(u,v;t,s) is approximated by

fin(u, v; t, s) = E[y'(t)~yfn(s)~Its(yn)\y(t) = u, y(s) = v]/ y (0,y(*)K V)' (19)

27
The simplest approximation fio(u,v;t,s) can be computed analytically, while,
for n > 0, fin has to be computed numerically. (Observe that n corresponds to
the dimension of the integral involved in the computation of /i n .) The regression
algorithm presented in [9] can be used for any s and n; however, large values of
s require large n in order to achieve satisfactory accuracy of the approximation
and are difficult to compute.
In the case when s is close to £, or u to v, the intensity \i{u,v\t,s) is
accurately approximated by the explicit approximation tio(u, v\ t,s). Conse-
quently the algorithm finds a fixed time point, (dependent on the mean and
covariance of the load and the thresholds u , v ) , such that, for t < s < To,
^(u,i;;£,s) « // 0 (w,^; t,s). Then, a suitable constant n and the second time
point T\ are found, such that for T$ < s < T\ the intensity //(u,t>;£,s) can be
accurately approximated by /xn(u,t>; £,s), i.e.
[To tT\ r+oo
n{u,v\i) w / /i0(u,v;t,s)ds+ nn(u,v;t,s)ds + fi(uyv;tys)ds. (20)
Jt JTQ JTi

For small and moderated cycles, i.e., the barriers u,v close to each other, the
last integrand in (20) is usually close to zero and can be disregarded. However,
for high RFC-amplitudes , the integral has to be approximated.
For any T > t, the integral ft ft(u,v;t,s) ds is equal to the intensity of the
u-downcrossings of u, by y, at t followed by the downcrossing of the level v in
the interval [t; T], without crossing the level u in between. Consider an intensity
v defined by

v{u,v\t,&) = E[y'(t) y'(s)+Its(y)\y(t) = u,y(s) = u)fy{t)Ms)(u,u), (21)

then ft v(u, v; t, s) ds is the intensity of the u-downcrossings, by y, at t followed


by the upcrossing of the level u in the interval [t; T], without crossing the level
v in between, see [9] for more detailed discussion. By letting T go to -foo, we
obtain the following representation of the intensity of the u-downcrossings, by
V, at t
+oo r+oo
(22)
/
fi(u,v]t,s)ds +/ v(u,v;t1s)ds,
.e.
where fi(u]t) is the intensity of the u-downcrossings at t by the load t/, i.<
(23)
p(u;t) = E[y'(t)-\y(t) = u]fy{t)(u).

Consequently, for any T > t


H(u, v;t) = j fi(u, v; t, s) ds + P(T) ■ (fi(u; t) - J /j(w, v; t, s) + v(u, v\ t, s) ds),
(24)
where P ( T ) , 0 < P(T) < 1 is a proportion of sample paths oft/ which crossed the
level u at £, stayed between the barriers u and v in the interval [£,T], and finally
crossed u before v. The intensity i/(u, v ; / , s ) , t < s < T\ can be approximated
by the regression approximations J/0, i/n, say, obtained by replacing the y process
by its regression curves y0, yn, respectively, in the conditional expectation (21).

28
By choosing the factor P(T\) (T = Ti) equal to 1, 0, one obtains upper- and
lower-bounds for //(u,i>;£), respectively. As we have mentioned before, these
bounds are tight for the values of u,v that are relatively close to each other.
However, for u high and v low the majority of sample paths of the load y which
crossed u at t have not passed any of the barriers if, v in the interval (2, Ti], and
hence /x(u, v; t) « P(Ti)//(u; t).
The approximation of P(T\) is based on an observation that, for Gaussian
processes, the point processes of upcrossings of the level u and downcrossings of
v converge to independent Poisson processes, as u —» -f-oo and v —> — oo.
Consequently, we propose to approximate the factor P(T\) by the exact value
of -P(Ti), evaluated when the crossings of the levels v, u are independent Poisson
processes with appropriate crossing intensities, i.e.

P^{Tl) = [+°° XZ(t, s) ■ exp" A *•<">+**<'•*>* ds, (25)

where the intensities A", A+ are given by

Kfa3) = / i ( w ; 0 " 1 ^ b / ( 0 " 2 / / ( 5 ) " l 2 / ( 0 = u,y(s) = v]fy{t)Ms)(u,v),


l
\+(t,s) = tivt)- E\y'(t)-y'{*)+\y{t) = «,yW = "]/v(o.vwK")- (26)
The integrals in (25) have to be calculated numerically and hence the infinite
region of integration, [7\, -f oo] needs to be truncated to the interval [7\, T 2 ], say,
for suitably chosen time T 2 . By simple calculations Equation (26) can be written
as follows

FW(T,) = % X:(t,s) ■ e x p - A *•<«•*>+**<«■*> * ds

+P°?r(T2)-exp-%K{t'z)+XUt'z)d\ (27)

If the integral JTi2 A~(£, z) -f A+(£, z) dz is large, the truncation error is negligible,
otherwise Papp(T2) has to be approximated.
For stationary loads, the intensities A~(£, s), AJ(£, s) converge to the station-
ary intensities of downcrossings of levels w,i;, //(u;2),/i(v; £), respectively, and
for T2 large enough
l
papV(T x= ffo*) = (28)

where /i(u; t), /i(t>; t) are downcrossing intensities given by (23). By substituting
in (27) the approximation (28) for Papp(T2) we obtain a new approximation for
P{T,).
For non-stationary loads the way of choosing the constant T2 and the estima-
tion of Papp(T2) depends heavily on the particular model for the non-stationary
properties of the load y. The algorithm chooses between a conservative estimate
of Papp(T2), in (27), i.e. Papp(T2) = 1, or assumes periodicity of the non-
stationary properties of y and modifies the simple approximation (28). More
exactly, if one assumes that for large values of s — t the intensities A~(2,s),

29
\+(t,s) are approximative^ equal to the intensities of downcrossings of the
level v and upcrossings of the level u at s, respectively. These intensities are
then given by Rice's formula, i.e.
fi~{v;s) = E[y'{s)-\y{s) = v]fy{s)(v),
f(u;s)= E[y'(s)+\y(s) = u}fy{s)(u). (29)
Next, assume that the non-stationary properties of the load y varies in a peri-
odical manner, i.e. there exists a positive period r, say, such that fi~(v;s) =
fi~(v;s + r ) and /z + (u;$) = // + (t/;3 + r ) . Consequently, for large T 2 , Papp(T2)
can be obtained from the following equation

+ F"»{T2) ■ e x p - ^ + % " ( ^ ) + " + ( u ; z H z . (30)


app
By solving (30) for P (T2) and substituting in (27), one obtains an approxi-
mation Papp(T\) for periodically non-stationary Gaussian loads. (One can easily
check that, for the stationary intensities fi~(v;s) = fi(v;t), // + (tr, s) = ^(u;£),
Papp(T2) defined by (28) satisfies Equation (30).)
Although the algorithm for the intensity ji(u,v,t) for non-stationary loads
is very similar to the algorithm for stationary loads, it is numerically more
complicated. For stationary zero mean loads, the following symmetry can be
used; //(u,v;0) = //(—v, — u\ 0). For example, consider a zero mean, variance
one, stationary Gaussian load; the intensity //(4, —1;0) ~ /*(4;0) is easy to
compute, while the opposite case /i(l,—4;0) w P(7\)/z(l;0) is computer time-
consuming. For non-stationary loads both intensities fi(u,v;t) and /i(—u, — u; t)
have to be computed separately.
Simpler approximations of the intensity //(u, v; t) can be obtained by choosing
the times To = t or T\ = To. For example, T\ = t leads to a Poisson type
approximation, which is related to the procedure proposed by Ford [7]. Tests
indicate that the simplest algorithm, based on a choice T\ = T 0 , perform better
than the Poisson type of approximation (T\ = t). (Observe that both simple
algorithms require similar amounts of computations.)
Finally, the damage intensity is computed using Equation (12). The expected
damage accumulated in the interval [0, T] is given by

[T d{t)dt.
E[D(T)]= (31)
Jo
The integral in (31) has to be computed numerically; hence, assume that one
can find a quadrature, i.e. nodes and weights (r,-, h{), i = 1 , . . . , k, such that

*W)]«£d(T,-)A,-. (32)
t= l

If the damage intensity d(t) changes slowly with t or varies in some regular
pattern, then it is possible to choose a relatively small number of nodes and
weights in (32). Otherwise, since for each r t , the approximation of J(r t ) requires
extensive numerical calculations, the sum (32) can be difficult to compute for
large values of T.

30

You might also like