spectral analysis of aggregates and products of time series construed of variable failure rates
TRANSCRIPT
This article was downloaded by: [University of California Santa Cruz]On: 17 November 2014, At: 21:11Publisher: Taylor & FrancisInforma Ltd Registered in England and Wales Registered Number: 1072954 Registered office: MortimerHouse, 37-41 Mortimer Street, London W1T 3JH, UK
Stochastic Analysis and ApplicationsPublication details, including instructions for authors and subscription information:http://www.tandfonline.com/loi/lsaa20
Spectral analysis of aggregates and products oftime series construed of variable failure ratesN. Singh a & V.S.S. Yadavalli ba Department of Mathematics , Monash University-Clayton , Vic. 3168, Australiab Department of Statistics , University of South Africa , P 0 Box 392, Pretoria, 0003,South AfricaPublished online: 03 Apr 2007.
To cite this article: N. Singh & V.S.S. Yadavalli (1997) Spectral analysis of aggregates and products oftime series construed of variable failure rates, Stochastic Analysis and Applications, 15:4, 629-641, DOI:10.1080/07362999708809498
To link to this article: http://dx.doi.org/10.1080/07362999708809498
PLEASE SCROLL DOWN FOR ARTICLE
Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”) containedin the publications on our platform. However, Taylor & Francis, our agents, and our licensors make norepresentations or warranties whatsoever as to the accuracy, completeness, or suitability for any purposeof the Content. Any opinions and views expressed in this publication are the opinions and views of theauthors, and are not the views of or endorsed by Taylor & Francis. The accuracy of the Content shouldnot be relied upon and should be independently verified with primary sources of information. Taylorand Francis shall not be liable for any losses, actions, claims, proceedings, demands, costs, expenses,damages, and other liabilities whatsoever or howsoever caused arising directly or indirectly in connectionwith, in relation to or arising out of the use of the Content.
This article may be used for research, teaching, and private study purposes. Any substantial or systematicreproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in anyform to anyone is expressly forbidden. Terms & Conditions of access and use can be found at http://www.tandfonline.com/page/terms-and-conditions
STOCHASTIC ANALYSIS AND APPLICATIONS, 15(4), 629-641 (1997)
SPECTRAL ANALYSIS OF AGGREGATES AND PRODUCTS OF TIME SERIES CONSTRUED OF VARIABLE FAILURE RATES
N. Singh
Department of Mathematics Monash University Clayton, Vic. 3168
Australia
and
V.S.S. Yadavalli
Department of Statistics University of South Africa P 0 Box 392, Pretoria 0003
South Africa
ABSTRACT
In this paper, the authors derive spectra of some linear and non-linear combinations of ARMA processes and discusses their application in reliability and surival analysis. Illustrative examples are given.
1. INTRODUCTION
Singh has recently [I] developed an unconventional
approach in time domain to the analysis of observed or
estimated failure rates of complex systems that operate
under changing operational and environmental conditions.
Since such failure rates often exhibit random changes in
level and slope, they can be construed as time series.
At times, the variable failure rates may also exhibit a
periodic (cyclic) pattern depending on the maintenance
and inspection procedures. Then obviously it would be
erroneous to analyse such failure rates using the
Copyright C 1997 by Marcel Dekker, Inc.
Dow
nloa
ded
by [
Uni
vers
ity o
f C
alif
orni
a Sa
nta
Cru
z] a
t 21:
11 1
7 N
ovem
ber
2014
630 SINGH AND YADAVALLI
standard failure distribution approach. Furthermore,
periodcities cannot be taken into account by the
latter approach; and, perhaps, for this reason the
periodicities have been ignored by the reliability
analysts.
Since engineers in general will be more interested
in the analysis of such variable failure rates in the
frequency domain, rather than the time domain, the
author obtains in this paper the spectra of aggregates
and products of time series made up of failure rates.
From a mathematical point of view, the spectrum and
autocovariance function contain equivalent information
concerning the underlying stationary random sequence
{ x } ; however, in practice the spectrum has a more
tangible interpretation in terms of the inherent varia-
tions about the mean. Furthermore, a great strength of
the spectral analysis lies in the independence of peri-
odogram ordinates at different Fourier frequencies.
This leads technically to a very straightforward method
of statistical analysis and testing. In addition, it is
known that the resulting methods are not critically
dependent on the assumption of the underlying random
sequences being normally distributed. However, the
restriction of spectral methods to stationary phenomena
is of ctitical importance. While using spectral methods,
it is presumed that any non-zero mean value, whether
constant or time-dependent, is absent; otherwise it
should be removed by detrending prior to the analysis.
2. BASIC CONCEPTS AND DEFINITIONS
2.1
Let { T , ; n = 1, 2 , . . . } be a sequence of pre-defined equidistant time points on the time
axis and let t, s t, s . . . be a sequence of failure times which constitute a serial time series. D
ownl
oade
d by
[U
nive
rsity
of
Cal
ifor
nia
Sant
a C
ruz]
at 2
1:11
17
Nov
embe
r 20
14
SPECTRAL ANALYSIS OF ARMA PROCESSES 63 1
I\
Let 2, denote the failure rate at time t and let 2,
denote the estimate of Z, defined by h 2, = X, if T , . , c t s T , , n = 1 , 2, . . . (2.1.1)
where # of failures in interval ( T , . ~ , T , ]
X, = . (2.1.2) # of survivals at 7 , - I
Then the estimated failure rate sequence {x,} may be
regarded as a time series at equidistant time points.
In general, the series {x,} may be nonstationary and periodic.
Example 2 .1 (Taken from Davis [2] )
A frequency distribution of failures of V 8 0 5 vacuum
tubes used in transmitters is given Figure2.1 and the
estimated failure rates using (2.1.2) are plotted in
Figure2.2 (see appendix I) . The plot of the number of failures in Figure 2.1
clearly indicates that the failure times follow an
exponential distribution and are obviously subjected
to decreasing trend and random fluctuations. This
phenomenon is also supported by the graph of the
failure rates in Figure2.2 which, instead of being
constant, seems to form a stationary time series.
2 . 2 LINEAR COMBINATIONS
Let {x, : t =0, +I, k2, . . . } denote a basic
univariate nonseasonal and stationary time series in the
original time scale t and let B denote the backshipt
operator (B1x, = x,.,) . We define the following:
1. An Overlapping Linear Combination
An overlapping linear combination of the x,'s in the
original time scale is defined by J-1
where the w ' s are constant weights.
Dow
nloa
ded
by [
Uni
vers
ity o
f C
alif
orni
a Sa
nta
Cru
z] a
t 21:
11 1
7 N
ovem
ber
2014
632 SINGH AND YADAVALLI
2. A Non-overlapping Linear Combination
A linear combination of the xt's in the new time scale
defined by J- 1
is called a non-overlapping linear combination, where
t = JT. The weights w,, w,, . . . , w,., in (2.2.1) and
(2.2.2) are assumed to be the known constants, with
w, e 0. Two interesting special cases of (2.2.2) are:
( i) Temporal Aggregation
When w, = w, = . . . =w,-, = w, V, in (2.2.2) is called
a temporal aggregation of the x,., namely,
V, = w(x, + x,-, + . . . +x~.~I;) (2.2.3)
(ii) Systematic Sample
If w, = 1 and w, = 0 for j = 1, 2, . . .J-1, then
V, , T = 0 , kl, . . . in (2.2.2) is called a systematic sample of the xtls in (2.2.1) namely
V, = x , , , T = O , + l , . . . (2.2.4)
3. Contemporaneous Aggregation
Let {x,, , i = 1, 2, . . . ,k} be k independent tirw series, then their linear combination
Zt = a, x,, + a, x,, + . . . + ak xk, (2.2.5)
is called the contemporaneous linear combination of the
xtls.
3.SPECTRA OF LINEAR AND NON-LINEAR COMBINATIONS Assuming that the basic time series {x,) follow an
ARMA model, we first derive in this section the models
for various combinations of x, and then determine their
spectra.
3.1 SPECTRA OF OVERLAPPING LINEAR COMBINATIONS
Let {x,}be a nonseasonal and stationary AR(p)
process defined by Dow
nloa
ded
by [
Uni
vers
ity o
f C
alif
orni
a Sa
nta
Cru
z] a
t 21:
11 1
7 N
ovem
ber
2014
SPECTRAL ANALYSIS OF ARMA PROCESSES 633
@(B)x, = 6 , (3.1.1)
where
@(B) = 1 - p,B - p2B2 - . . . - p,BP is an AR operator and { E , } is a white-noise sequence with mean 0
variance aC2 . Theorem 3.1
Let {x,} be a nonseasonal and stationary A R ( p )
process defined in (3.1.1) and let V, = x, + x, + . . .+x,.~ be an overlapping linear combination of the
xtls. Then V, is an AR(p) process with the same
coefficients and the error variance equal to (J + l)a,'. The spectrum of V, is then given by
Theorem 3.1 suggests a smoothing technique which
corresponds to forming the overlapping.linear combina-
tion. This means that for any two different values of
t , V, may include some of the same x, . For t ranging
over half years, a simple case is when J = 1, that is,
a moving average of two six monthly values.
Theorem 3.2
Let {x,} be an M A ( q ) process defined by
where O(B) = 1 + 8,B + . . .+&JqBq is an MA operator and {e,} is defined in (3.1). Let V, = x, + x,., +
. . .+x,., be an overlapping linear combination of the
xtls, then V, is an MA(q) process with the same coeffi-
cients and the error variance equal to (J + 1)ue2 and
ae2 = (k + 1) (1 + 8,' + . . .+eq2)oe2. The spectrum of V, is given by
(J + 1) ae2 f,(w) = Is I
2 n
Dow
nloa
ded
by [
Uni
vers
ity o
f C
alif
orni
a Sa
nta
Cru
z] a
t 21:
11 1
7 N
ovem
ber
2014
SINGH AND YADAVALLI
Theorem 3.3
Let {x,} be an ARMA(p, q) process defined by
@(B)x, = O(B)e, (3.1.5)
and let V, = x, + x,., + . . .+x ,-,, then V, is an
ARMA(p,q), with the same coefficients and the error
variance (J + 1) u e 2 . The ue2 can be determined using the
well-known Yule and Walker equations. The spectrum of
V, is given by
3.2 SPECTRA OF NON-OVERLAPPING LINEAR COMBINATIONS
Suppose that the records about the failure times
and/or failure rates are maintained fortnightly and for
some reasons, one is able to analyse only the monthly
data by aggregating the fortnightly figures in the form:
V, = x,+x,-, , T = O , &I, k2, . . . (3.2.1)
where t = 2T. It may be noticed from (3.2.1) that the
linear conbination is non-overlapping.
In general, a non-overlapping linear combination of
the x,'s in the new time scale T is defined by J-1
where t = JT. In (3.2.2) , the weight wj's are assumed to
be known. There are two special cases:
(a) Temporal Aggregation
When all weights in (3.2.2) are equal, we have a
temporal aggregation. Consider
Theorem 3.4
Let {x,} be an AR(1) process defined by
X, = cp x,-, + e, , (3.2.3)
where {e,} is a white-noise sequence with mean 0 and
varaince ae2 and let Dow
nloa
ded
by [
Uni
vers
ity o
f C
alif
orni
a Sa
nta
Cru
z] a
t 21:
11 1
7 N
ovem
ber
2014
SPECTRAL ANALYSIS OF ARMA PROCESSES 635
where t = JT, then V, is an ARMA (1,l) process defined by
approximately, where
u, = e, + e,-, + . . . + e, .,,.,, (3.2.6)
Hence a,' = J a,' , and
a$ = J 0,' (1 + p2 + 2pJt1) / (1 -* (p2J)
The spectrum of V, is then given by
J a,' (1 + 2 p cos w + p2) f,(w) = -
2 n (I - 2pJ cos w + p2J)
(b) Systematic Sampling
Suppose that failure rates of certain items are
recorded at the end of each semester in a year and
suppose the data are not complete in one of the
semesters; then one can form a yearly series consisting
of data of one semester only, that is,
x2,-, , for the first semester z, = LX2; T = 0, +I, . . . (3.2.8)
Theorem 3.5
Let {x,) be an -(I) process as defined in (3.2.3)
and let Z, be a process as defined in (3.2.8) ; then Z, is
an AR (1) process with error variance equal to ae2 (1 + ( p 2 )
and its spectrum is given by
ue2 (1 + (p2) f,(w) = {I + (p4 - 2p2 COS w}-I (3.2.9)
2 n
3.3 SPECTRA OF CONTEMPORANEOUS AGGREGATION
Theorem 3.6
Let { x ; i = 1, 2, . . . k} be k independent
stationary processes with spectra f,,(w),
i = 1, 2, . . .k; then the spectrum of the linear
combination
Dow
nloa
ded
by [
Uni
vers
ity o
f C
alif
orni
a Sa
nta
Cru
z] a
t 21:
11 1
7 N
ovem
ber
2014
636 SINGH AND YADAVALLI
is given by k
where the ails are constants.
The proof is simple and straightforward, hence
omtted.
Corollary 3 . 1
If a, = l / k ; i = 1 , 2, . . . , k and if f , , ( o ) =
f , , ( u ) = f , ( u ) ; i, j = 1 , 2 , . . . k ; i + j , then
f , ( u ) = f , (w) ( 3 . 3 . 3 )
Exam~le 3 . 1
Let { x i , i = 1 , 2 be two independent A R ( 1 )
processes defined by
Xit = (Pi X i . t - l + uLt ; i = 1 , 2 ( 3 . 3 . 4 )
where p < 1 , i = 1 , 2 ; the {ui,} are two independent
white-noise processes with variences auX2, i = 1 , 2. Then
the spectrum of
z, = Xl , + x2t ( 3 . 3 . 5 )
is given by f , ( w ) = f,, ( w ) + f,, ( w )
where f x i ( u ) = aui2{1 + qi2 - 2 p i cos w}-l/2n
when
(PI = (P2 = (P , U U I 2 = 0 ~ 2 ~ = au2
f,(~) = 2UU2 {l + (P2 - 2 (P COS a}-' / 271 ( 3 . 3 . 6 )
3 . 4 SPECTRA OF THE PRODUCT OF TWO TIME SERIES
Theorem 3 . 7
Let { x } , i = 1, 2 be two independent A R ( 1 )
processes defined at ( 3 . 2 . 3 ) . Then the spectrum of the
produce Z = XI, X,, is given by
Dow
nloa
ded
by [
Uni
vers
ity o
f C
alif
orni
a Sa
nta
Cru
z] a
t 21:
11 1
7 N
ovem
ber
2014
SPECTRAL ANALYSIS OF ARMA PROCESSES
Corollarv 3.2
When cp, = cp, = cp , aUl2 = = aU2, then
f,(w) = au2(1 + q12) / {(I - (p2)(l + p4 - 2p2 COS 0)}2n
(3.4.2)
Theorem 3.8
Let {x,} and {Y,} be two independent MA(1) processes
defined by X, = e, - 6, e,., and Y, = u, - 6, u,~,, (3.4.3)
where {e} and {u,} are two independent white-noise
processes with variances ae2 and aU2 respectively. Then
the spectrum of the product Z, = X,Y, is given by
f,(w) = ae2 aU2 {(I + 012) (1 + 0,') + 2 6, 0, cos w}/ 27~
Corollarv 3.3 (3.4.4)
If a,' = a, = au2 , 0, = 6, = 0, then
f,(~) = a4 {I + O4 + 2 02(1 + cos w)} (3.4.5)
4. EXAMPLES
4.1 Consider the following two m ( 1 ) processes:
X, = 0.8 X,-, + e, and Y, = 0.6 Y,-, + u, (4.4.1)
where e,, u, ,N(O, 1) . The spectra of their various
combinations are given in Appendix 11.
4.2 Spectra of the following two M ( 1 ) processes and
their various combinations are given in Appendix 111:
X, = -0.8 X,., + e, and Y, = 0.6 Y,., - u, (4.1.2)
4.3 Spectra of the following Ar(2) processes and their
various combinations are given in Appendix IV:
X, = 0. ax,., - 0 .4X,., + e, and
Y, = 0.5Y,., + 0 .7Y,., + U, (4.1.3)
REFERENCES
[I] N. Singh, "Stochastic Modelling of aggregates and products of variable failure rate", Statistical Research Report No.235, Department of Mathematics, Monash University, Clayton, Vic.3168, Australia.
[ 2 ] D.J. Davis, 'I An analysis of some failure data", J. Amer. Statist. Soc., Vo1.47, 1952, pp 113-150
Dow
nloa
ded
by [
Uni
vers
ity o
f C
alif
orni
a Sa
nta
Cru
z] a
t 21:
11 1
7 N
ovem
ber
2014
APPENDIX l
SINGH AND YADAVALLI
F I Q U ~ ~ 2 1 D~slribution of failure frequency w~ th operallng 11me - type v805 vacuum tube used n transm~tiei
0 0 0 o ~ I ~ ~ I ~ ~ I ~ I I ~ 1 I l I l 10 20 30 40 50 60 70 80 90 1 0 0 1 1 0 1 2 0 1 3 0 1 4 0 1 5 0 1 6 0 1 7 0 1 8 0 1 9 0 2 0 0
F~qure 2 2 The Fa~lure Rates
Dow
nloa
ded
by [
Uni
vers
ity o
f C
alif
orni
a Sa
nta
Cru
z] a
t 21:
11 1
7 N
ovem
ber
2014
SPECTRAL ANALYSIS OF ARMA PROCESSES
APPENDIX II
- F~gure II 1 F~gure 11.2 Spectrum of 4 = 0.84 . + e, Spectrum ol Y, = 0.W , + u,
0
F~gute ll 3 Spectrum of Ihe sum I = X + Y,
0
Figure 111 4 S~ec t rum of the product Z = X Y,
0
Fgure 1 5 Speclrum of the sum and product z = X I Y, + X Y,
Dow
nloa
ded
by [
Uni
vers
ity o
f C
alif
orni
a Sa
nta
Cru
z] a
t 21:
11 1
7 N
ovem
ber
2014
APPENDIX Ill
I
0 1 2 3 0
F~gure 111 1 Spectrum of X = -0 8 X , + e,
SINGH AND YADAVALLI
0
Fgure 111 2 Spectrum ol Y = 0 BY, + e
1 2
0 8 fh(o) 0 l0M 6 f2(0)iim
0 4
0 2 0 1 2 3 0
0 1 2 3
0 F~gure 111 3 Figure 111.4 Speclrum ol Ihe sum Z, = X, + Y, Spectrum of Ihe product Z, = X Y,
w Figure Ill 5 Spectrum of lhe sum and producl Z - X, + Y, I X, Y.
Dow
nloa
ded
by [
Uni
vers
ity o
f C
alif
orni
a Sa
nta
Cru
z] a
t 21:
11 1
7 N
ovem
ber
2014
SPECTRAL ANALYSIS OF ARMA PROCESSES
APPENDIX lV
F~gure IN1 Figure lV2 Spectrum ol X, = 0 8Y , - 0 4X, + e, Speclrum of Y = 0 5Y, , - 0 7Y, , + e,
0
Figure IV 1 Spectrum of the producl Z, = X, Y,
Dow
nloa
ded
by [
Uni
vers
ity o
f C
alif
orni
a Sa
nta
Cru
z] a
t 21:
11 1
7 N
ovem
ber
2014