Application of Time-Series Analysis to Wireline Logs]

Sandra J. Kerford: Esso Resources Canada, Limited, Calgary, Canada

Daniel T. Georgi: Esso Resources Canada, Limited, Calgary, Canada

Abstract: Time-series analysis can be used to investigate rates. Several of our reservoirs are comprised of thin beds.
log repeatability and to quantitatively determine the vertical To better resolve these beds, we used higher sample rates
resolution of a log. This technique involvescomparingrepeat and reduced logging speeds. However, this increases the
runs of logs to differentiate between coherent signal and in- data volume and the length of time required for logging,
coherent noise. No detailed knowledge of the sonde char- as well as the costs of logging and data processing. Thus,
acteristics is required. The technique accounts for the deg-
radation of the logging data due to noise (e.g., counting to justify the additional expenditures and to enable us to
statistics) and sonde accelerations and decelerations, and, use these new logging techniques in a cost effective man-
hence, results are representative of the vertical resolution ner, we needed to quantify the differences in the log data
achieved under realistic logging conditions. quality.
In this paper, coherenceand phase measurements are dis- Examples are used to illustrate and explain the appli-
cussed, and two applications of these techniques to log anal- cation of time-series analysis to wireline logs. The first
ysis are described. Time-series analysis is used to quantita- examples show how time-series analysis can be used to
tively determine the optimum sample interval for a density determine the quality of logs. The second examples il-
and a neutron tool run at various logging speeds. The results lustrate how this type of analysis can be used to evaluate
indicate that at porosities of about 15% the density and neu- the benefits of different logging speeds and sample inter-
tron logs recorded at 550 m/hr. (1,800ft/hr.) should be sam- vals. Other possible applications of time-series analysis
pled every 0.1524 m (6 in). This sample interval enables all
the useful signals to be recorded. A smaller sample interval
to wireline logs are discussed in the summary.
does not improve the quality of the data because the signal
contained in the additional high frequency components can- TECHNIQUE OF APPLYING
not easily be distinguished from the noise. At a reduced TIME-SERIES ANALYSIS
logging speed of 160m/hr. (525 ft/hr.),the density logs should TO WIRELINE LOGS
be sampled using a smaller sample interval if high resolution
To quantify log repeatability and vertical resolution,
- data are required, because 0.1524 m (6 in) sampling does
not enable all the useful signal to be recorded. Regarding the we applied standard time-series analysis techniques (see
neutron log, 0.03048 m (1.2 in) sampling is not beneficial in Jenkins, 1963; Bendat and Piersol, 1971; Brigham, 1974).
higher neutron porosity formationseven ifthe log is recorded Although, these techniques are most commonly applied
at 160 m/hr.; however, 0.03048 m. (1.2 in) sampling of neu- to data that are sampled with respect to time (e.g., wind
tron logs may be beneficial in lower porosity formations. records, seismic traces, ocean currents, etc.), they can be
In addition to determining the degree of correlation be- applied equally well to data sampled with respect to depth.
tween logs and the vertical resolution of logs, time-series The major difference is that the results are expressed in
analysis can be used to determine the signal-to-noiseratio terms of spatial instead of temporal frequencies.
of log data and to quantify the effects of smoothing and We will not review the entire science of time-series
deconvolutionfilters.This technique can be applied routinely analysis but will only attempt to relate the basic concepts
because the only data required are two logging passes (main as they apply to our analysis of log data. The first step in
and repeat).
the analysis consists of transforming the log data (evenly
INTRODUCTION sampled in depth) to an evenly sampled series in spatial
Time-series analysis is a well-known process that has frequency. This transformed series is merely a Fourier
been used routinely for years in many areas of geophysics representation of the original data. The terms in the series
and engineering. In the case of wireline logs, the technique represent the amplitudes of sine waves of specific fre-
involves comparing, in the spatial-frequency domain, re- quencies and phases necessary to reconstruct the original
peat runs of logs recorded over the same interval to de- data. Figure 1 shows the first four sine waves necessary
termine the logs’ vertical resolution and to quantify their to approximate the simple square wave shown in the
repeatability. upper right-hand corner. By including additional high-
The impetus for this work was the desire to quantify frequency terms in the Fourier representation, we can
the advantages of slower logging speeds and higher sample improve the approximation of the original square wave.
The square of the Fourier coefficients, called the power
Originally presented at 1987 CWLS Formation Evaluation Sym- spectrum, is often presented on a log-log scale versus
posium, paper M. frequency. Figure 2 shows the power spectrum obtained

150 The Log Analyst May- June, 1990

 Application of Time-Series Analysis to Wireline Logs

r--7
DEPTH DOMAIN I
L-A

FREQ. AMP PHASE

, Y
l T o

31 5 0
377

*
51 5&f 0

71 1
71T
0 -*
DEPTH - - DEPTH

FREQUENCY
q1 + , ,
. ’, ,,,,,., , ,,,,,,,, I
I
I , , , , . I, . I I

:++-*
DOMAIN ~ 0.01 0.1 1 10
SPATIAL FREOUENCY Il/m)
0
Figure 2 . Power spectrum of the counts recorded by the far
SPATIAL FREQUENCY a -7T SPATIAL FREQUENCY
detector of a neutron tool. The abrupt change in slope at 3
Figure 1: Fourier approximation of a square wave. The ap- cycledm may correspond to the vertical resolution of the tool.
proximation consists of transforming the square wave from the
depth domain to an evenly sampled series in spatial frequency.
The terms in the series represent the amplitudes and phases of
the sine waves of specific frequencies necessary to reconstruct function, known as coherence (see, for example, Bendat
the original data. This figure shows the first four sine waves and Piersol, 1971), it is possible to differentiate between
necessary to approximate the square shown in the upper right- coherent signal (i.e., the reservoir characteristics that do
hand corner. By including additional high-frequency terms in not change from logging run to logging run) and incoher-
the Fourier representation, the approximation of the square ent noise. For each spatial frequency, we have:
wave could be improved.

from the Fourier transform of count-rate data from the

far detector of a neutron tool. The data are typical of
geophysical data; most of the power (i.e., variance) occurs where Ccf)is the coherence squared, P, and Pw are the
at low frequencies, with the power spectrum falling off power spectra of the two input logs and I ( P , , ) I is the
rapidly as frequency increases. (Such spectra are often magnitude of the cross spectrum. The angled brackets,
referred to as “red.”) This particular power spectrum falls “( ),” represent band and/or ensemble averaging (for more
off rapidly as a function of frequency up to a spatial fre- details see Appendix).
quency of 3 cycles/m; at higher frequencies, the spectrum The coherence is the frequency domain analog to the
decreases only moderately with increasing frequency. One depth domain time-lagged correlation coefficients. As with
could surmise that this change in slope of the power correlation coefficients, the normalized coherence varies
spectrum is due to a transition from real signal to noise- between 0 and 1. If the coherence is computed for very
dominated signal and, thus, estimate the vertical reso- similar logs, then the coherence will be nearly equal to 1
lution of the log from the inverse of the frequency cor- for all frequencies.If, however, the coherence is computed
responding to the crossover of the two trends. However, for two very dissimilar logs, then the coherence will be
such an approach is, in general, unsatisfactory because nearly equal to 0 for all frequencies. In addition to the
not all power spectra are as easily subdivided into signal- coherence, we can also compute the phase, a.0, which is
dominated and noise-dominated frequencies. Further- a measure of the relative depth shift between the two logs.
more, the observed changes in slope could be related to If the main and repeat passes of the log are identical
formation characteristics and not to log quality. and on depth, then the computed coherence will be 1 at
By combining the Fourier representation of two logs all frequencies, and the phase will be 0 (see Figure 3). If,
across the same interval (e.g., main pass and repeat pass) however, the main pass and the repeat pass do not repeat
and computing the normalized cross-spectral density well, then the coherence will be low, and the phase will

May-June, 1990 The Log Analyst 151

 Kerford and Georgi

COHERENCE PHASE LITHOLOGY LOG

IDENTICAL DATA

0.0 I -180 TRANSITION

ZONE

VERY DISSIMILAR DATA I

0.0

THICKNESS OF TRANSITION ZONE * 112 WAVELENGTH OF THE

HIGH FREQUENCY CUTOFF lc.

IDENTICAL DATA I ' O r 1801

I
DEPTH-SHIFTED

Figure 4: Relationship of the high-frequency cutoff to the

thickness of the transition zone on a log across an abrupt contact.
Figure 3: Interpretation of coherence and phase data. Coher-
ence and phase can be used to measure the degree of similarity
between repeat runs. Coherence results are not affected by sim-
ple depth shifts. tion and borehole) does not change between passes, then
the signal-to-noise ratio, S/N, for a large number of de-
grees of freedom is given by:
fluctuate wildly about 0". The coherence and phase func-
tions possess another interesting and useful property. If
the main and repeat passes are identical but are off-depth
relative to one another, then the coherence will still be
equal to 1 for all frequencies; however, the phase will no The signal-to-noise ratio is usually very high for resis-
longer be 0 but will increase linearly with increasing fre- tivity and acoustic tools and lower for tools that count
quency. radioactive decays (e.g., gamma ray, bulk density, neu-
Generally, the coherence is highest at low frequencies tron porosity, etc.). The signal-to-noise ratio can be used
and lowest at high frequencies. By quantifying the tran- to decompose the power spectra (P, and Pyy)into the
sition between coherent signal and incoherent noise, we noise and signal spectra (see Georgi, 1989). These spectra
are able to determine the vertical resolution of a log. As can then be integrated to obtain the coherent-signal vari-
with seismic data, vertical resolution is actually a function ance and the uncertainty due to the noise in the log data.
of bandwidth (Evenden et al., 1970); however, because The low frequency signal-to-noise ratios and the vertical
the response of most logging tools is good at low fre- resolutions of the logs make it possible to evaluate the
quencies, the bandwidth and the vertical resolution of effects of logging speed on log quality and to quantitatively
logs are proportional to the high frequency transition be- intercompare logs from different companies (see Georgi
tween coherent and incoherent signal. The relationship et al., 1989).
between the high frequency cutoff,f,, and the log response
observed at an abrupt lithology transition is shown in COMPUTATIONAL DETAILS
Figure 4. Because of the limited vertical resolution of the To increase the stability of the coherence and phase
log, the abrupt lithology change appears smoothed on the estimates, we employ both ensemble and band averaging.
log. The thickness of the apparent transition zone can be For example, if we have a 50 m repeat and main pass,
approximated by one-half the wavelength corresponding we subdivide each 50 m piece into two 25 m pieces and
to the high frequency cutoff. pool the results (Figure 5). This decreases the frequency
The magnitude of the observed coherence is related to resolution and results in the loss of one low frequency
the signal-to-noise ratio. When the repeat and main passes coherence estimate. To gain additional degrees of free-
agree perfectly, the coherence equals 1 and the signal-to- dom and to reduce the variance of the computed coher-
noise ratio is infinite. As the magnitude of incoherent ence and phase, we also pool results from neighboring
noise increases, the coherence decreases and the signal- frequencies using a Hanning filter (see, for example, Op-
to-noise ratio decreases. If it is assumed that the signal penheim and Shafer, 1975).
and the noise are statistically the same on the main and To minimize the effects of low frequencies with wave-
repeat passes and that the logging environment (forma- lengths longer than the lengths of the pieces being trans-

152 The Log Analyst May-June, 1990

 Application of Time-Series Analysis to Wireline Logs

LOG 1 LOG 2
RAW DATA

11x1

t MEAN AND TREND REMOVED

tI
--
Figure 5: Ensemble averaging technique employed to increase DATA TAPERED
stability of the computed coherence and phase. A
TAPERED v 2

Figure 6: Preprocessing technique employed to increase sta-

formed, we differentiate the logs, remove the mean and
bility of computed coherence and phase.
trend, and apply cosine tapers to the ends of the pieces.
These steps are summarized in Figure 6.
To compensate for the loss of data at the ends of the
To aid the interpretation of the coherence diagrams,
pieces due to the application of the cosine tapers, we use
we include the expected mean value of the coherence
overlapping pieces. Thus, the results from three pieces,
function (for 0 true coherence) based on the number of
as shown in Figure 5, are ensemble averaged to obtain
degrees of freedom (Goodman, 1957; Koopmans, 1974).
the coherence, phase, and average spectra. However, when
This is the 50% confidence level. Also included are the
calculating the number of degrees of freedom, only the
values of coherence that must be exceeded for the com-
number of independent pieces is counted (two in Figure
puted coherences to be different from 0 coherence at the
5). 95% and 90% confidence levels.
We stated earlier that the coherence for two unrelated
logs is equal to 0. This is strictly true only for an infinitely QUANTIFYING LOG QUALITY
large number of degrees of freedom. The coherence func- In one of Esso's fields, the dual-induction and the
tion is a biased estimator, and for a small number of gamma-ray tools are run in combination to determine
degrees of freedom, the mean value of coherence for two
intrareservoir characteristics. The spherically focused re-
unrelated logs can differ significantly from 0. The smaller
sistivity data are used to identify tight streaks, and the
the number of degrees of freedom, the larger the mean
gamma-ray data are used to quantify shaliness. To min-
value of the coherence function for two uncorrelated logs.
imize logging time, the logs are run at 1,100 m/hr. from
Bendat and Piersol (197 1,pp. 193-1 95) give the following
total depth to the surface casing. One repeat is routinely
formula for estimating the mean value of coherence for
collected across the formation of interest.
two uncorrelated logs:
Time-series analysis was applied to the repeat and main
E[CCf)]= tanh (l/[n - 21) pass log data to quantify vertical resolution and to estab-
lish log quality. Figure 7 shows the computed coherence
where E[Cy)] is the expected value of the coherence and data for the unaveraged spherically focused and gamma-
n is the number of degrees of freedom and can be ap- ray logs. The resistivity data exhibit excellent coherence
proximated by:
up to the highest recorded frequencies (3.3 cycledm). The
n N 2 x number of spectral components smoothed over observed coherence for the shallow resistivity data ex-
x number of pieces ensemble averaged over ceeds the coherence for the random noise at all frequen-

May-June, 1990 The Log Analyst 153

 Kerford and Georgi

\1
WAVELENGTH I m l WAVELENGTH ( m )

$
,y3L
230 100 50 20 10 0 5
- 1 0
SPHERICALLYLOG
FOCUSSED

ij
8
, , , ,,,,,
L?
02
c/o
Lq
00
*\
02 0s 01 0.2 05 10 20
SPATIAL FREQUENCY I l / m )

Figure 8: Coherence for openhole shallow resistivity and sonic

data and cased-hole-induced gamma-ray spectroscopy C/O and
. .
001, , , , , ,,,,,I , ', 1 H yield data.
0.1 1.0
SPATIAL F R E Q U E N C Y I l l m l

Figure 7: Coherence diagrams for gamma-ray and spherically cycledm and for the C/O ratio at 0.5 cycles/m. Because
focused logs recorded at 1,100 m/hr. using a 0.1524 (6 in) sample of the gradual transition to the expected level of coherence
interval. Note the high coherence for the SFL resistivity data
for random noise, the calculated vertical resolutions for
and the low coherence and poor repeatability of the gamma-ray
data. the H yield (0.36 m) and the C/O ratio (1 .O m) are some-
what ambiguous.
cies. In contrast, the gamma-ray coherences rarely exceed
the coherence expected for random noise. Only for spatial QUANTIFYING THE BENEFITS OF
frequencies <0.12 cycledm (8 m vertical wavelength) do DIFFERENT LOGGING SPEEDS
the coherences exceed the expected coherence of random AND SAMPLE INTERVALS
noise at the 90% confidence level. Clearly, the gamma- Wireline logging companies may record logs using var-
ray data collected at 1,100 m/hr. provide little infor- ious logging speeds and sample intervals. Reduced logging
mation other than the tops and bases of the formations speeds and smaller sample intervals are being used to
of interest. To improve the quality of the gamma-ray data, resolve thinner beds and to obtain more accurate log
it is necessary to decrease the logging speed or average measurements. Such techniques cannot be used arbitrar-
the data from the main pass and the repeat pass. ily because they do increase the data volume and the
Generally, the vertical resolution and repeatability of length of time required for logging, thereby increasing the
logs based on radioactivity is less than those for logs based costs of logging and data processing. Time-series analysis
on direct measurements of physical properties. This is can be used to quantify the benefits of logging a density
illustrated in Figure 8 with coherence results for log data and a neutron tool using different speeds and sample in-
from Cretaceous age uncemented Clearwater sands. The tervals.
coherences for open,hole shallow resistivity and sonic Figure 9a shows the coherence of density data that was
data are higher than those for the induced gamma-ray recorded using the routine logging speed of 550 m/hr. and
spectroscopy capture H yield and inelastic C/O data. The a 0.1524 m (6 in) sample interval. The coherence diagram
low-frequency (large vertical scales) coherence levels, indicates that the density data repeats well for frequencies
which reflect the log repeatability, are highest for the re- < 1.O cycle/m (wavelengthsgreater than 1 m). The vertical
sistivity data and lowest for the C/O data. The coherences resolution of the log may be determined by noting the
for the H yield are comparable to those for compensated frequency at which the mean value of the coherence equals
neutron porosity data. The coherence for the resistivity the 50% confidence level. This value occurs at 2.0 cycles/
data remains high to the highest recorded frequencies (3.3 m, which corresponds to a vertical resolution of about
cycledm). The decrease in coherence at high frequencies 0.25 m. For frequencies higher than 2.0 cycles/m, the
(>3.0 cycles/m) reflects the decrease in recorded variance signal content cannot be distinguished from the noise,
at high frequencies relative to low frequencies. However, indicating that more frequent sampling of this log is not
the S/N ratio remains well above the levels expected for beneficial at the routine logging speed. More frequent
random noise. The coherence for the sonic data decreases sampling will increase the number of high-frequency
to the level expected for random noise at 2.5 cycles/m. components that are recorded on the log; however, the
The corresponding vertical resolution, 0.2 m, is compa- amplitude of signal contained at these high-frequency
rable to the receiver spacing 0.3 m (1 ft). The coherence components does not exceed the amplitude of the inco-
for the H yield falls to the level of random noise at 1.4 herent noise and does not provide usable information.

154 The Log Analyst May-June, 1990

 Application of Time-Series Analysis to Wireline Logs

WAVELENGTH (m) WAVELENGTH lm) the coherence diagram is very similar to the coherence
10.0 1.0 10.0 1.0
1.0 diagram obtained for the density data recorded at 160
m/hr. with a 0.1524 m (6 in) sample interval; however,
because of the decreased sample interval, the coherence
diagram for the 0.03048 m (1.2 in) sampled data contains
YI

z
additional high-frequency components. The magnitude
EO5 of the coherence at these frequencies (2.0-5.5 cycledm)
I
s indicates that it is beneficial to record density data at
these reduced logging speeds (160 m/hr.) using a 0.03048
m (1.2 in) sample interval. This is in contrast to the results
obtained at normal logging speeds (550 m/hr.). However,
00 it must be realized that the coherence level for these ad-
0:1 1:o 0.1 l:o
SPATIAL FREQUENCY ( I l m l SPATIAL FREOUENC Y I I l m I ditional frequency components is not nearly as high as
the coherence level for the low-frequency components,
WAVELENGTH lm) indicating the presence of noise at the higher-frequency
10 0 I0 01
components. The upper limit of useful frequencies is 5.5
cycles/m, which corresponds to a vertical resolution of
0.09 m.
In summary, the coherence diagrams (Figures 9a, 9b,
95XCONFIDENCE LEVEL 9c) show that
1) slower logging increases the quality of density data,
8 2) 0.03048 m (1.2 in) sampling of the density log in this
formation is only beneficial if the tool is logged slowly,
-LOGGING 51
3) in this formation, the vertical resolution of the unav-
eraged density log when recorded at 550 m/hr. is 0.25
00
01 m,
SPATIAL FREOUENCY Illml 4) in this formation, the vertical resolution of the unav-
Figure 9: Coherencediagrams for density logs recorded at var- eraged density log when recorded at 160 m/hr. is 0.09
ious logging speeds and sample intervals. Note that a slower m.
logging speed increases the coherence of the density logs and
that 0.03048 (1.2 in) sampling as opposed to 0.1524 m (6 in) A similar analysis was also done for neutron log data
sampling is only beneficial if the density tool is logged slowly. (Figure 10). The average neutron porosity is 15%. In low-
er-porosity formations, the results may be different be-
cause the noise level on the log should be less.
Figure 9b compares coherence diagrams for density Figure 10a shows the coherence diagram of a neutron
data recorded using a 0.1524 m (6 in) sample interval log that was recorded using the routine logging speed of
and logging speeds of 160 m/hr. and 550 m/hr. The com- 550 m/hr. and a 0.1524 m (6 in) sample interval. The
parison indicates that the slower logging speed improves coherence diagram indicates that the neutron log repeats
the coherence at all frequencies. This result is expected well for frequencies <0.5 cycledm (a wavelength of 2.0
because the slower logging speed results in an increase in m). The uppermost useful frequency is 2.3 cycles/m, which
the number of radiation events measured by the detector, corresponds to a vertical resolution of 0.22 m. These
thereby decreasing the radiation noise. The coherence results indicate that more frequent sampling of this log,
diagram indicates that the density log recorded using the when it is recorded at 550 m/hr., is not beneficial, the
slower logging speed repeats well for frequencies up to signal contained in the additional high-frequency com-
2.0 cycles/m as opposed to 1.0 cycle/m for the normal ponents will not be distinguishable from the noise.
logging speed. The coherence diagram also proves, by the Figure 1Ob compares coherence diagrams of neutron
fact that the mean value of the coherence is greater than data recorded using a 0.1524 m (6 in) sample interval
the 50% confidence level for all frequencies, that all fre- and logging speeds of 550 m/hr. and 160 m/hr. As ex-
quencies on the log recorded at the slower logging speed pected, the slower logging speed improves the coherence
contain some signal that can be distinguished from ran- at all frequencies; however, it is also important to realize
dom noise. Hence, more frequent sampling of density that the amount of improvement in data quality is not
data recorded at 160 m/hr. is beneficial. the same for the neutron tool as for the density tool. The
Figure 9c shows the coherence for density data recorded slower logging speed greatly improves the log quality of
using a logging speed of 160 m/hr. and a 0.03048 m (1.2 the density data at high frequencies but only moderately
in) sample interval. At low frequencies (<2 cycledm), improves it at lower frequencies. In contrast, the log qual-

May-June, 1990 The Log Analyst 155

 Kerford and Georgi

WAVELENGTH I r n ) WAVELENGTH I rn 1 detailed knowledge of the sonde characteristics is re-

10 0 10 lo 0 10
10 quired. The technique accounts for the degradation of the
data due to noise (e.g., counting statistics) and sonde ac-
celerations and decelerations, and hence, the estimate of
the vertical resolution and repeatability are representative
of the results achieved under realistic logging conditions.
The described technique is very practical. Time-series
analysis can be applied to most logs because all that is
needed is a main pass and a repeat pass.
Time-series analysis can be used to quantify log quality.
The analysis can be used to evaluate the improvements
0.1 I:o 0.1 10
in vertical resolution and repeatability of logs run at re-
SPATIAL FREQUENCY 1 I l m l SPATIAL FREQUENCY I l l m ) duced speeds and recorded using a more frequent sam-
pling rate. For the density and neutron tools we used in
WAVELENGTH Im) our formation, 0.1524 m (6 in) sampling is sufficient if
10.0 1.0 0.I
these tools are logged at the routine speed of 550 m/hr.
At a reduced logging speed of 160 m/hr., 0.03048 m (1.2
in) sampling is beneficial for the density tool but not for
the neutron tool. It should be stressed that our results
were obtained from formations with average neutron po-
rosities of 15%; 0.03048 m (1.2 in) sampling of neutron
logs may be beneficial in lower-porosity formations.
Time-series analysis of wireline logs can be beneficial
in other aspects of formation evaluation as well. The
technique can be used to evaluate and compare logs pro-
o.o', """, 0.1
I 1 """,
I .o
1 t
I .
I I I I

10.0
vided by different service companies. Applied on a rou-
SPATIAL FREQUENCY ( I / m l
tine basis, the analysis can be used to monitor and main-
tain log quality. Time-series analysis can also aid in the
Figure 10: Coherence diagrams for neutron logs recorded at
various logging speeds and sample intervals. Note that a slower
evaluation of the effects of different smoothing filters and
logging speed increases the coherence of the neutron logs and the design of optimum smoothing filters. Logging com-
that 0.03048 (1.2 in) sampling as opposed to 0.1524 m (6 in) panies smooth logs to reduce the amount of noise on the
sampling is not beneficial in high porosity (- 15%) formations logs. This procedure is often beneficial; however, if the
even if the tool is logged at the reduced logging speed of 160 logs are oversmoothed, valuable information can be lost.
m/hr. Time-series analysis can be used to recognize and resolve
this problem.
ity of the neutron data is moderately improved at all Service companies are continually developing new tools
frequencies. and processing algorithms. Often, these improvements
Figure 1Oc shows the coherence for the neutron log are measured only subjectively by the log analyst or in-
recorded using a logging speed of 160 m/hr. and a 0.03048 terpreter, but time-series analysis can be used to quantify
m (1.2 in) sample interval. The coherence of the addi- these improvements. In summary, we wish to emphasize
tional high-frequency components recorded on the log, that time-series analysis can be applied to many aspects
on account of the smaller sample interval, equals the 50% of log analysis.
confidence level. This indicates that the magnitude of the
signal contained at these frequencies does not exceed the ACKNOWLEDGEMENTS
magnitude of the noise. It will be difficult to distinguish
the signal from the noise, and hence, 0.03048 m (1.2 in) We thank E. C. Trantham who suggested applying time-series
sampling is not beneficial. The uppermost useful fre- analysis to wireline logs and C. Keith who provided many valu-
able comments on the manuscript. We also thank Esso Re-
quency is 2.9 cycles/m, which corresponds to a vertical
sources Canada, Limited, for allowing us to pursue this research
resolution of 0.17 m. These results are dependent on li- and for providing permission to publish this paper.
thology and porosity, and more frequent sampling should
be beneficial in low neutron porosity formations.
REFERENCES
SUMMARY
Bendat, J. S., and Piersol, A. G . , 1971, Random data: analysis
Time-series analysis can be used to determine the ver- and measurement procedures: John Wiley & Sons, Inc., New
tical resolution and repeatability of a wireline log. No York, 406 p.

156 The Log Analyst May-June, 1990

 Application of TimeSeries Analysis to Wireline Logs

Brigham, E. O., 1974, Thefast Fourier transform:Prentice-Hall, then the coherence squared is given by
Inc., Englewood Cliffs, New Jersey, 252 p.
Evenden, B. S., Stone, D. R., and Anstey, N. A., 1970, Seismic
prospecting instruments, v. 1: Gebruder Borntraeger, Berlin,
156 p.
Georgi, D. T., 1989, Application of time-series analysis to in- where XY and P are the complex conjugates of X and Y, re-
duced gamma ray spectroscopy logs from two cold lake heavy- spectively. The angled brackets "( )" represent band and/or
oil observation wells, SPE-19602, in SPE Annual Technical ensemble averaging. The two quantities in the denominator,
Conference and Exhibition Proceedings, vol. omega, For- XX* and Y P , are the power spectra P, and Pw, corresponding
mation Evaluation and Reservoir Geology: Society of Petro- to the logs x and y . The quantity I X P I is the magnitude of the
leum Engineers, p. 323-334. complex cross-correlation of x and y.
Georgi, D. T., Heavysage, R. G., Chen, S. T., and Erikson, E. It is convenient to express the complex cross-correlationXY*
A., 1989,Application of shear and compressionaltransit-time in terms of the real and imaginary parts:
data to cased hole carbonate reservoir evaluation, paper L,
in 12th Formation Evaluation Symposium Transactions: Ca-
XYY = CO + iQ
nadian Well Logging Society, p. Ll-L24. where the real part is often referred to as the coincident
Goodman, N. R., 1957, On joint estimation of the spectra, co- spectrum and the imaginary part as the quadrature spec-
spectrum and quadrature spectrum of a two-dimensional sta- trum. The ratio of the coincident and the quadrature
tionary Gaussian process: New York University, Engineering spectra define the phase spectrum:
Statistics Laboratory, scientific paper no. 10, 168 p.
Jenkins, G. M., 1963,Cross-spectral analysisand the estimation @(k) = arctan[Q(k)/CO(k)].
of linear open loop transfer functions, in M. Rosenblatt, ed.,
Symposium on time series analysis: John Wiley & Sons, Inc.,
London, p. 267-276. ABOUT THE AUTHORS
Koopmans, L. H., 1974, The spectral analysis of time series: Sandra Kerford received her B.Sc. and M.Sc. in Geophysics
Academic Press, Inc., New York, 366 p. from Queen's University in 1979 and 1981, respectively. Since
Oppenheim, A. V., and Shafer, R. W., 1975, Digital signalpro- then she has worked for Computalog Gearhart, Ltd., and for
cessing: Prentice-Hall, Inc., Englewood Cliffs, New Jersey, Esso Resources Canada, Limited. At Esso Resources, Sandra
585 p. has been primarily involved in the analysis of log data from a
signal processing point of view.
APPENDIX Dan Georgi received a B.A. in Physics from the University
of California in 1971 and an M.S. and Ph.D. from Columbia
The coherence, phase, and power spectra for two logs, x(z)
University in Earth Sciences in 1977. After working at Wood's
and y(z), are defined in terms of their discrete Fourier trans-
Hole Oceanographic Institute for 4 years, he joined Exxon Pro-
forms. The two logs are assumed to be evenly sampled, Az, in
duction Research Company, where he was responsiblefor petro-
the depth domain, z. Letting X ( k ) and Y(k) be the complex
physical evaluation of fractured reservoirs. In 1986, he trans-
Fourier representations for x(z)and y(z),respectively (see Brig-
ham, 1974, for details of the Fourier transform): ferred to Esso Resources Canada, Limited, in Calgary, where
he works in the Formation Evaluation group in the Research
and Technology Division.
Depth Fourier Frequency
domain
x(z)
Y(4
-
transform

-
domain
Xk)
Y(k)

May-June, 1990 The Log Analyst 157