ARTICLE IN PRESS

International Journal of Rock Mechanics & Mining Sciences 45 (2008) 1068–1081

www.elsevier.com/locate/ijrmms

Application of autocorrelation analysis for

interpreting acoustic emission in rock
Jan Vilhelma,b,, Vladimı́r Rudajevb, Tomáš Lokajı́čekb, Roman Živorb
a
Faculty of Science, Charles University in Prague, Albertov 6, 12843 Prague, Czech Republic
b
Institute of Geology of the ASCR, v.v.i., Puškinovo nám. 9, 16000 Prague 6, Czech Republic
Received 19 July 2007; received in revised form 8 November 2007; accepted 9 November 2007
Available online 21 February 2008

Abstract

The statistical properties of acoustic emission from rock samples were studied as a function of applied uniaxial load. It was found that
the parameters of the autocorrelation function of the acoustic emission event series change significantly near failure. An increase in the
values of the autocorrelation coefficients and a tendency to a linear decrease with time were observed. We propose that the increasing
autocorrelation of the acoustic emission series is an evidence of the increased affect that the individual acoustic emission sources have on
one another. This mutual effect of acoustic events arises as a result of the redistribution of stress in the sample during the fracturing
process at higher loads (more than 95% of ultimate strength). The results support the possibility of using autocorrelation analysis as a
failure warning sign or even to predict the sample’s total failure. Different rock materials and various loading patterns were used to
generalise the results obtained.
r 2007 Elsevier Ltd. All rights reserved.

Keywords: Compression; Deformation; Fractures; Laboratory measurement; Microseismicity; Rock fracture; Seismic-event rates; Statistical methods

1. Introduction in real-time scanning electron microscope observations [8],

The assessment of the stress–strain state of rocks, and
especially forecasting the occurrence of extensive brittle
fracturing, is currently one of the significant objectives of
geomechanics. The results of relevant experiments are of
importance in the evaluation of underground opening
stability and predicting induced seismic phenomena [1,2].
A number of methods and procedures, often based on
seismic or other geological and geophysical data, have been
used to study deformation processes in bodies of rock and
in forecasting sudden releases of seismic energy [3–6]. The
fracturing process can be studied on different scales.
Szwedzicki [7], for example, showed that based on
macroscopic analysis rock sample failure mechanism can
be assessed (tension, shear or coupling of tension and
shear). A microscopic approach was applied, for example,
Corresponding author at: Faculty of Science, Charles University in
Prague, Albertov 6, 12843 Prague, Czech Republic.
Tel.: +420 221 951 563; fax: +420 221 951 556.
E-mail address: vilhelm@natur.cuni.cz (J. Vilhelm).
Kawakata et al. [9] employed X-ray computer tomography,
Wulff et al. [10] and Couvreur et al. [11] studied the
attenuation of P- and S-waves, and Wei et al. [12] studied
strong stress fluctuations. The rock behaviour and fractur-
ing can be studied also by means of numerical modelling
[13–15]. Liu et al. [16] showed that numerical tests can
evaluate the process of rock samples fracturing by
determination of the crack nucleation and initiation, and
can differentiate between stable and unstable crack
propagation. Based on changes in the time series of
quantities studied, precursors of sudden failures may be
identified. However, the identification of these precursors
depends on the analysis of long-term measurements, that
is, on determining the basic trend of these quantities and
finding anomalous behaviour conditions, which can be
interpreted as precursors. In this paper, we deal with a
laboratory method of assessing the instability of rock
samples based on monitoring and interpreting acoustic
emissions (AE; discrete, burst-type signals [17], referred
also as ultrasonic emission, since the frequency content of
signals exceeds 100 kHz).

1365-1609/$ - see front matter r 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ijrmms.2007.11.004
 ARTICLE IN PRESS
J. Vilhelm et al. / International Journal of Rock Mechanics & Mining Sciences 45 (2008) 1068–1081
Deformation processes, taking place in situ, can be
modelled by laboratory methods under simplified condi-
tions. Specifically, the system of stress (uni-, bi- or tri-
axial), the rate of loading, etc., can be chosen, and the
responses of various types of rocks to loading can be
studied. Although a considerable simplification of the
actual deformation processes is involved, significant
similarity between laboratory experiments and phenomena
occurring in the natural environment can be observed. This
similarity applies not only to the manner of seismic energy
release (even though the dimensions of microfailures in
samples are of the order of millimetres, whereas in mining
tremor foci they are 104 times larger [18]) but also to the
form of the distribution functions, for example, energy–
frequency functions, after-shock sequence distribution
relations, etc. [19,20]. The parameters of these distribution
functions are not random, but depend on the physical
(material) properties of rock (for example, granularity and
primary degree of failure) and on the level of stress. The
problem of similarity between the sample and in situ
conditions lies in the substantially faster changes in the
state of stress under laboratory conditions, and in the finite
dimensions of the samples.
Time series of seismic events or AE events may be
analysed using methods of statistical physics, aiming at
predicting the time of failure. This approach is based on
assuming that an earthquake corresponds to a critical
point, but the stress state of the rock mass can approach, or
move away from this critical point.
Seismicity can be described using the framework of self-
organized criticality, which was originally derived from
cellular-automaton models [21]. This concept produces
power-law frequency–size statistics similar to the Guten-
berg–Richter relationship [22]. This relation seems to be
stable as long as the external driving force remains
constant. If some energy loss is introduced, the model
acquires a memory and this leads to seismic cycles [14,23].
The methods of forecasting total failure of rock samples
are based on analysing the distribution of the AE, and
mostly involve comparing these distributions under various
loads. The results of the experiments we conducted and the
study of time series, based on the application of autocorrela-
tion analysis, indicate, however, that changes in the
autocorrelation coefficients provide absolute criteria for
determining the approaching state of sudden failure which
does not require comparison with the overall preceding
fracturing process. The results of locating acoustic pulses, in
particular the migration and clustering of microfailure foci,
can also be used to forecast total failure, as well as energy
release analysis and especially the energy–frequency distribu-
tion and the pulse frequency as a function of load (or time).
2. Brief overview of methods and interpretation of acoustic
emission
The forecast of a sudden release of seismic energy
(whether severe earthquake, strong induced event—mining
1069
tremor, or failure of a loaded sample) consists of three
parts: (1) forecasting the place, (2) the time of origin and
(3) the seismic energy released. Forecasting the place is
based on locating the places of failure, their migration and
clustering in space. In the end, this also contributes to
forecasting the time. The estimate of the failure time is
based on seismic energy release analysis, the energy–
frequency distribution, analysis of shock sequences, and
on the autocorrelation properties of time sequences of
occurrence of seismic events. All these parameters are
function of the stress–strain condition of the rock mass,
or sample.
The location of foci of AE sources is based on multi-
channel recordings of acoustic events. To determine the
four unknown parameters of the focus (that is, the focus
coordinates x0, y0, z0 and time t0), at least four sensors are
required in the case of known P-wave’s velocity propaga-
tion [24]. This velocity may also be used to assess the
deformation of the sample and its failure state; such
information can also be derived from the attenuation of
energy [10]. For the purpose of locating the source of
emissions, however, velocity variations are important [25].
Location methods serve to determine the most fractured
parts of the rock sample. Correlation analysis can provide
a quantitative measure of the migration of foci during the
failure process [26,27]. The spatial distribution of foci can
also be classified in the terms of fractal dimension D
[28,29]. According to Lockner and Byerlee [30], the
decrease of D from a value close to 3 (applicable to a
uniform distribution of foci in space) indicates a change in
the spatial distribution of foci and possible creation of
fractures. Zang et al. [31], however, reported that the
decrease of parameter D neither characterize the nature of
any precursor to total sample failure in any of the types of
rock they studied nor for the various methods of loading.
Monitoring the AE from loaded rock samples also
makes it possible to determine the AE event energy. It is
frequently assumed [17,32,33] that the magnitude–fre-
quency distribution of emissions may be described by the
Gutenberg–Richter relationship, well known from earth-
quake research [22]. This relation can be expressed as
log NðMÞ ¼ a  bM, (1)
where N(M) is the number of earthquakes with magnitude
M or greater and variables a and b are the parameters of
the distribution. In the first approximation, seismic
magnitude M is proportional to the logarithm of energy
E of seismic events and Eq. (1) can be rearranged as
logðN i Þ ¼ A  g logðEi Þ, (2)
where Ni is the number of events within energy interval
Ei7DE and A and g are the parameters of the distri-
bution [34].
However, it has been shown that this distribution is valid
only over a restricted energy range. Holt et al. [35] showed
for tests on pressure vessel steels that the AE energy is
determined by the particle size together with the applied
 ARTICLE IN PRESS
1070 J. Vilhelm et al. / International Journal of Rock Mechanics & Mining Sciences 45 (2008) 1068–1081
stress, and that the energy distribution is determined
essentially by the particle size distribution. Rudajev and
Lokajı́ček [18] and Přikryl et al. [36] proved that the power-
law distribution on loaded rock samples is only valid for
energies E4Ec, where Ec is the characteristic energy of the
acoustic signals. Characteristic energy Ec is defined as the
maximum of energy–frequency distribution. The grain-size
distribution has a significant effect on this characteristic
energy Ec [37]. According to [17], Ec can be related to some
characteristic physical parameters of the deformation
mechanism.
Janssen et al. [38] have shown (on a granite rock) that
fractures originate not only between grains but also
through the failure of these grains. The number of
transgranular cracks increases and the number of grain
boundary fractures decreases on approaching failure.
According to [31], changes in parameter b are a precursor
of total failure of a loaded rock sample. Schorlemmer and
Wiemer [39] have shown that the b-value varies system-
atically for different styles of faulting.
During the laboratory experiments, an increased number
of AE events are usually observed when acting stress
exceeds 65% of the compressive strength. However, under
cyclic loading, the Kaiser’s effect [40,41] is frequently
observed [42,43]: when materials are subjected to cyclic
loading with increasing maximum stress, AE is not
detectable until the previously applied stress level is
exceeded [44]. Kaiser discovered this behaviour of AE in
a stressed metal sample. Later, it was also observed in rock
materials [45], but its existence in the case of Westerley
granite was seriously questioned by Sondergeld and Estey
[46]. According to Rudajev et al. [47] or Guarino et al. [42],
the Kaiser effect was not confirmed in extremely hetero-
geneous rock materials; however, this effect can be used to
investigate the stress history of rocks [48,49]. A quantita-
tive explanation of the Kaiser effect from the viewpoint of
damage mechanics of rocks by means of numerical
modelling is given by Tang et al. [15].
The onset of the AE depends on the properties of the
rock material being loaded, on the manner of loading
(varying rates of loading or deformation) as well as on the
parameters of AE recording (level of triggering). A
different AE occurrence rate was observed under stepwise
loading of rocks [50]. An increased rate of AE events was
observed as the load was increased to the next level. When
the load remained constant at the new higher level, the
number of AE began to decrease with time. The decrease of
AE under a constant load level is due to the time-
dependant subcritical deformation (long time strength
was not reached) that leads to both stress and strain
relaxation. This decay formally corresponds to Omori–
Utsu’s relation [13,51–53]. The magnitude of the coefficient
p in this relation depends on the material properties, on the
one hand, and on the state of stress of the rock sample, on
the other [54]. A smaller value of p (that is, the slower the
decrease of AE and the longer the after-shock sequence)
indicates a higher stress level in the rock sample. This
dependence is frequently exploited in researching induced
seismicity in mines for the purpose of assessing the state of
stress of the rock massif [55]. An attempt to establish a
physical basis for after-shock sequences by a continuum
damage rheology is presented in [56–59].
An analysis of time series of acoustic events serves to
determine how the source fractions of the individual AE
events affect one another. At the beginning of the loading,
the AE events are generated in the rock sample randomly
and do not affect one another, that is, their occurrence
depends on distribution of strength within the sample.
These sources are, above all, related to points of primary
weakening, i.e. to places where fractures occur. This is
where the state of stress increases and sudden irreversible
deformations, accompanied by emissions of acoustic
energy, occur. In rocks, containing a large number of
pre-existing microcracks, AE occur at relatively low values
of stress. These samples are unable to store a large amount
of energy, and gradually disintegrate without sudden brittle
failure occurring. On the other hand, in homogeneous
unbroken rocks, the occurrence of the first AE is observed
only at high values of stress (as much as 85% of the
compressive strength). These rocks are able to accumulate
strain energy and to release it suddenly in the form of total
failure.
Using correlation analysis of AE sequences, it is possible
to study how AE events affect one another. The increasing
mutual effect of acoustic events, characterized by an
increase in the values of autocorrelation coefficients, can
be used to estimate the time at which total failure of the
sample will occur.
3. Principle of correlation analysis
Consider two time functions x(t) and y(t). The cross-
correlation function of these functions is defined (assuming
that this integral converges) as
Z þ1
rxy ðtÞ ¼ xðtÞyðt þ tÞ dt. (3)
1
The cross-correlation function characterizes the similarity
of two functions x(t) and y(t) as a function of their offset t.
If the correlation of a single function (or x(t)=y(t)) is
being considered, one speaks of an autocorrelation
function:
Z þ1
rxx ðtÞ ¼ xðtÞxðt þ tÞ dt. (4)
1
Sometimes, it is more convenient to consider the
autocorrelation function of a centred function (i.e. after
subtracting the mean value m(x(t))), so that only fluctua-
tions remain. This is applied in this study. One then speaks
of an autocovariance function.
If instead of a continuous infinite function x(t), a
discrete, finite and centred sequence x(ti)=xi is considered,
the relation for the autocorrelation function (4) becomes
 ARTICLE IN PRESS
J. Vilhelm et al. / International Journal of Rock Mechanics & Mining Sciences 45 (2008) 1068–1081
the definition of the autocorrelation coefficients rxx(k):
1 NX
k
rxx ðkÞ ¼ xj xjþk ,
N j¼1
or
1 NX k
rxx ðkÞ ¼ xj xjþk ,
N  k j¼1
where k ¼ 0, 1, y, N1. Autocorrelation coefficients
rxx(k) calculated according to Eq. (5) are biased estimates,
because their expected value is not equal to their estimate
(for k6¼0). Relation (6) yields an unbiased estimate, but the
end points rxx(k) (near to k ¼ N1) suffer from large
variance because only a few data points are used in the
estimation.
Since rxx(k)prxx(0) for k ¼ 1, 2, y, N1 (when using
Eq. (5)), the normalised value Rxx(k) ¼ rxx(k)/rxx(0) is
frequently referred to as the autocorrelation coefficient. In
this case, it holds true that |Rxx(k)|p1 for k ¼ 1, 2, y,
N1. In case of the more often used unbiased coefficients
rxx(k) computed according to formula (6), the inequality
|Rxx(k)|p1 does not need to be true for k near to N1.
Taking into account this and considering statistical
reliability, we use only the first 10% of the autocorrelation
coefficients of their total number N.
Autocorrelation coefficients can be used to assess the
internal structure of a progression xi. If there is no relation
between the terms of the sequence (sequence xi consists of
random numbers—white noise) all autocorrelation coeffi-
cients Rxx(k) are zero and do not depend on offset k (k6¼0)
[60]. If there is a particular relation between the terms of
the sequence, i.e. for example, term xi+n+1 depends on
terms xi to xi+n, we find that the corresponding auto-
correlation coefficients are non-zero. The larger the value
of n, the more terms of the sequence affect the next term
and the larger number of autocorrelation coefficients are
non-zero (larger radius of correlation). If the following
term of the progression is affected by only one term (the
immediately preceding term), one speaks of the Markoff
process [61] and only the first autocorrelation coefficient
assumes a significant non-zero value (Rxx(0) ¼ 1 does
not count).
In assessing the AE from a loaded sample using
correlation analysis, it is possible to select the time series
of the number of AE events occurring per unit time
(acoustic rate), as sequence xi. One starts with the simple
notion of a loaded rock sample under uniformly increasing
load as the model of the failure process of a rock mass. If
we then wish to compare the processes of sample fracturing
state under various loads, we can choose time sub-intervals
DT of equal length, corresponding to these load levels, and
perform the autocorrelation analysis within these sub-
intervals independently. However, the problem with this
comparison may be that a significantly different total
number of AE events have occurred in each of the sub-
intervals. To be able to compare the results of the
1071
autocorrelation analyses directly, it is necessary to compute
the sequence of AE events rate, so that the average value of
the number of events of all series being compared is the
(5)
same. This can be done by choosing different lengths of
the elementary intervals dTi, depending on the rate of the
events recorded in the individual loading sub-intervals DTi.
The length of elementary intervals is determined according
(6) to the formula:
DT
dT i ¼ , (7)
ni
where ni is the number of elementary intervals dTi which
can be derived from the following expression:
DN i
ni ¼ . (8)
Nn
Here, N corresponds to the selected mean number
of AE events in all compared sub-intervals DTi and DNi is
the number of recorded AE events in individual sub-
interval DTi.
4. Experiment
Rock samples of migmatite and sandstone from the
Czech Republic are studied. The migmatite samples
originate from the region of Skalka in the Bohemian-
Moravian Uplands. Petrographic and mineralogical ana-
lyses indicate that the clusters of individual minerals in the
migmatite samples are arranged plane-parallel to linearly.
They contain rock-forming minerals quartz, potassium
feldspar, plagioclase, muscovite and biotite. The average
size of the quartz grains is 30  50 mm, grain aggregates
with undulating extinction are as large as 1.5  0.8 mm,
tabular aggregates of parallel intergrown grains of musco-
vite chips 2  2 mm and 0.65 mm thick. At the boundary
between mica and feldspar, there are infrequent occur-
rences of sericite aggregates and clay minerals, 60 mm in
size. The loading experiments are carried out in cylindrical
samples of diameter 50 mm and height 100 mm.
The sandstone samples were collected in the region of the
Kladno coal mines at a depth of 400 m below the surface,
that is, from a sandstone bank overlying the main coal
seam. The Kladno coal basin belongs to the Central
Bohemian region of the limnic carbon formation of the
Bohemian Massif. The sandstone samples display clastic
components, 0.25–0.50 mm in size, with sporadic occur-
rence of millimetre size grains. A minor feldspar com-
ponent occurs in trace quantities, as do fragments of
slate, lydite and spilite rocks. The rock structure is
psammitic of a basal nature and with a fine-grained
matrix: the individual grains are dispersed in the matrix
and are not in mutual contact. Prism-shaped samples,
100  100  200 mm in size, are tested.
The rock samples are uni-axially loaded with a noiseless
water press to maximum force of 500 kN. The loading was
effected by adding water in the tank, located on the acting
lever. Unloading is effected by draining the water from the
 ARTICLE IN PRESS
1072 J. Vilhelm et al. / International Journal of Rock Mechanics & Mining Sciences 45 (2008) 1068–1081
tank. The loading is carried out at a constant loading rate,
or by adding near-instantaneous incremental loads of
approximately 1–5% of the assumed compressive sample
strength [62]. Referring to the period of loading until final
failure, the experiments can be divided into short term (tens
of minutes) and long term (more than 10 h).
The basic deformation parameters of the sample studied
are measured in the course of the loading: the load applied,
and locally, at two measuring points (at half the height) and
the longitudinal and transverse deformations are measured
by two cross-strain gauges. The data from the gauges are
processed by a Hottinger Baldwin multi-channel tenso-
metric bridge and fed to a computer through an IEEE
interface. The deformation data are recorded at intervals of
100 s for the short-term measurements and 600 s for the
long-term measurements. Acoustic activity is monitored by
four broadband AE sensors, type WD, produced by the firm
PAC. The positions of the four sensors are shown in Fig. 1.
The signals are amplified by a broadband amplifier
(amplification of 40 dB). The AE is recorded by a special
PC interface card SF 41. The card enables independent and
continuous recording of AE in the frequency range from
5 kHz to 1.5 MHz. The card can record the arrival time of
the signal at the separate channels with an accuracy as high
as 125 ns, the integral value of the rectified signal (this value
is proportional to the energy) and other parameters [63].
To assess the dependency of the autocorrelation analysis
on the loading pattern, several different loading patterns
are selected.
Fig. 1. Diagram of location of AE sensors (S) and cross-resistance
tensometers (T) on the surface of cylindrical sample.
The loading patterns for the migmatite and sandstone
samples differ. The migmatite samples are loaded at a
constant loading rate. In the short-term tests, the rate is
about 25 kPa/s. In the long-term experiments, the rate is
about 2.75 kPa/s. For the slow increase of the load, several
small load steps are added near-instantaneously (these
additional loads amount to about 1–5% of the assumed
compressive strength). The purpose of this loading is to
further evaluate the independence of the correlation
analysis on the actual loading pattern. Sudden change of
stress may, moreover, be considered as a model of the
conditions following blasting in a mine, and we assume
that the study of the effect of this phenomenon on the
redistribution of stress and AE will receive attention in the
future. The time intervals between these additional loads
are not at regular intervals, but a roughly uniform coverage
is attempted of the part of the slow loading, during which
AE is already being observed. A similar combination of
linear slow loading increase and time variable additional
loading with a periodic time variation is used, e.g., by Yin
et al. [64] in modelling tidal triggering effects.
The sandstone samples are loaded in two loading cycles.
During the first one, the samples are loaded to about 80%
of their compressive strength at a rate of 8.3 kPa/s. During
the second loading cycle, a load close to that of the first
cycle is rapidly reached by applying a quadruple loading
rate, i.e. 33.3 kPa/s, and the sample is further loaded at a
rate of 8.3 kPa/s, similar to the first loading cycle. During
the second cycle, the activity of the AE confirms the Kaiser
effect. Until the previous value of the load applied, only
minimum activity of the AE is observed.
5. Autocorrelation analysis of AE series
The state of stress of the rock samples is assessed and the
moment of the total failure is estimated by monitoring
changes in the parameters of the autocorrelation function
of the sequence of emitted acoustic signals. For the purpose
of assessing the general validity of the conclusions drawn,
the effects of the structure on the AE, as well as the effect
of the different rates of loading are monitored. Sandstone
samples, which represent rocks of almost single-mineral
composition and thus display a more simple grain
distribution curve, are tested, as well as samples of
migmatite. The structure and mineral composition of the
latter are more complex (quartz, potassium feldspar,
plagioclase, muscovite and biotite) and its grain distribu-
tion curve is much more complex. The effect of the loading
rate on the parameters of the AE is also studied. For this
purpose, short-term tests, lasting minutes, and long-term
tests, lasting several hours, are run, and cyclic loading
is applied.
In all cases, the methods of processing are the same and
consist of dividing the entire loading period into partial
intervals to which autocorrelation analysis of the sequence
of emitted AE signals is applied. It is shown experimentally
that, under non-cyclic loading, the AE increases once
 ARTICLE IN PRESS
J. Vilhelm et al. / International Journal of Rock Mechanics & Mining Sciences 45 (2008) 1068–1081
roughly 65% of the samples’ compressive strength has been
reached. Under cyclic loading, the Kaiser effect is
observed, which means that the AE in the second cycle
increases only when the load applied on the sample in the
first cycle has been reached.
5.1. Short-term experiments
The short-term experiments, which take 120 min, are
carried out on migmatite samples. An example of the AE
under a uniform loading rate of 26 kPa/s is shown in
Fig. 2a. This figure P shows the cumulative curves of the
number of events ( N) recorded by all four sensors. Time
is plotted along the x-axis and the cumulative event
occurrence rate along the y-axis. The variation in the
number of events, recorded by all sensors, is practically
identical, and it is evident that an increase in AE activity
occurs at 4700 s, that is, when approximately 65% of the
samples’ compressive strength has been reached. The curve
of the cumulative number of events displays an exponential
increase, which agrees with previous results [65]. Specifi-
cally, the curve of the cumulative number of pulses,
recorded by sensor number 3 (the largest number of AE),
can be approximated by an exponential function, in the
interval from 6000 s to failure (see Fig. 2b):
X
NðtÞ ¼ 2504 þ 500  10ð0:00152ðt6000ÞÞ . (9)
Autocorrelation analysis is applied to consecutive time
sub-intervals of 120 s. The transformed sequence of the
number of events in the elementary intervals dT is
calculated, so that there is a mean value of four events to
an elementary interval. An example of the autocorrelation
functions, corresponding to three selected intervals A, B, C,
is shown in Fig. 2a (R(i) ¼ f(t)). The details P of the
cumulative number of events in subintervals ( N ¼ f(t))
display an increase: in sub-interval A 1668 pulses are
generated, while in interval B 3780 and in interval C 13,120
pulses. The non-cumulative number of events in elementary
sub-intervals (N/dT ¼ f(t)) displays chaotic variations. The
parameters of the autocorrelation functions are given in
Table 1. The first column of Table 1 shows the subinterval,
and the second the time of the beginning of the sub-
interval. The third column gives the load as a percentage of
the compressive strength. The fourth column shows the
correlation coefficient, corresponding to the linear approx-
imation of the autocorrelation function and the next
column gives the number of points fitted. The sixth column
shows the duration of the positive autocorrelation coeffi-
cients, that is, the time at which the approximation line
intersects level R(i) ¼ 0 (the correlation radius). The last
column gives the value of the first autocorrelation
coefficient. In the short-term tests, the value of the first
autocorrelation coefficient increases with the load, and the
tendency of the correlation function to decrease linearly,
characterized by the coefficient of correlation, is more
pronounced.
1073
A comparison of intervals A, B and C also clearly
indicates that the number of events increases in the
successive intervals. To assess the effect of this increase on
the results of the correlation analysis, a model calculation is
carried out based on fitting the curve of the cumulative
number of events to the exponential increase according to
Eq. (9). The autocorrelation analysis of the model of
cumulative number of events, Fig. 2b, is carried out in the
same intervals A, B and C as in the above experiment. The
same length of the sub-interval is used, DT ¼ 120 s, as well
as the average number of events (four events) per elementary
dT. The analysis indicates that the resultant autocorrelation
functions are in good mutual agreement for the three
intervals studied, and that the largest difference between
them is associated with the varying random scatter of values
R(i). This scatter is due to discretisation, the continuous
function of time increments between consecutive events
having been reduced to an integer number of events in the
separate elementary intervals. The larger the number of
events in the sub-interval analysed, the more elementary
intervals are obtained (i.e. the sequence of the number of
events in the elementary intervals is longer). This decreases
the probability of the occurrence of more significant
fluctuations in the autocorrelation coefficients.
In the case of the above model data, the increase in the
numbers of events in successive elementary intervals is small.
For the intervals in Fig. 2b, the mean difference between
successive intervals is less than one event. The transformed
sequence of the number of events consists of integers 3, 4
and 5. At the beginning, larger groups of integer 3 occur in
the sequence, rarely interrupted by the occurrence of one
integer 4. The alternation of integers 3 and 4 then
accelerates. As the integer 4 is approached in the sequence,
the groups of integer 3 diminish until its occurrence vanishes
from the sequence completely. The occurrence of one and
only one number different from the foregoing or subsequent
term of the sequence is relatively frequent. In calculating
autocorrelation coefficients, this is reflected in a significantly
lower correlation in shifting by exactly one point. The
decrease of the value of the first autocorrelation coefficient
R(1) can thus be considered an artefact, and in comparison
to real data it is more convenient to use coefficient R(2).
Table 2 summarises the results of the correlation analysis
of the model calculation. It is evident from Table 2 that
even there are significantly different numbers of events in
intervals A, B and C, the results of the correlation analysis
in the different intervals are approximately the same. The
autocorrelation function thus corresponds to the func-
tional relation of the input data, the number of events
itself, or the rate of increase of the number of events being
affected only by the scatter of the resultant coefficients R(i)
relative to linear approximation.
5.2. Long-term tests
The long-term test of the migmatite sample takes
1020 min, that is, about 17 h. Fig. 3 shows the strain curve
 ARTICLE IN PRESS
1074 J. Vilhelm et al. / International Journal of Rock Mechanics & Mining Sciences 45 (2008) 1068–1081

P
Fig. 2. (a) Short-term loading test, migmatite: increase in cumulative rate of events with time (load). The detailed structure of the cumulative curve N,
the transformed progression N/dT and autocorrelation function R(i) are shown for selected sub-intervals A, B and C. (b) Short-term loading P test,
migmatite: the autocorrelation analysis of fitted exponential cumulative rate of events with time (load). The detailed structure of the cumulative curve N,
the transformed progression N/dT and autocorrelation function R(i) are shown for the same sub-intervals A, B and C as in Fig. 2a.
 ARTICLE IN PRESS
J. Vilhelm et al. / International Journal of Rock Mechanics & Mining Sciences 45 (2008) 1068–1081 1075

Table 1
Parameters of the autocorrelation analysis for samples of short-term loading test, migmatite (Fig. 2a)

Sub-interval Beginning of Load s (% of Coefficient of Number of fitted Correlation R(1)

interval (s) strength) correlation points radius (s)

A 6500 89.5 0.71 42 10 0.08

B 6800 93.7 0.77 94 13 0.36
C 7100 97.8 0.95 328 15 0.67

Table 2
Parameters of the autocorrelation analysis—model calculation (Fig. 2b)

Sub-interval Beginning of Load s (% of Coefficient of Number of Correlation R(1) R(2)

interval (s) strength) correlation fitted points radius (s)

A 6500 89.5 0.75 37 54 0.32 0.58

B 6800 93.7 0.92 107 54 0.32 0.56
C 7100 98.1 0.97 307 54 0.33 0.57

measured in the axial and radial directions. The graph
shows that the total axial deformation was about five times
larger than the total radial deformation. The linear loading
is accelerated for a short time in several cases. The middle
part of Fig. 3 shows the cumulative number of AE events
as a function of time (load). This cumulative event
occurrence rate displays short-term rapid increases in the
number of events in five cases. The overall trend of the
cumulative event occurrence rate is again exponential. The
autocorrelation analysis is carried out over a running sub-
interval of 120 s in the course of the loading. Several
autocorrelation functions in sub-intervals A–D are shown
in Fig. 3. Outside the loading steps, the intervals
correspond to uniform AE.
The curves in sub-intervals A and B indicate why
autocorrelation analysis could not be used due to the low
level of the AE. In long-term tests, autocorrelation analysis
can be applied from a load of roughly 90% of the
compressive strength. A short increase in the AE event
rate occurs as a result of a sudden increase in pressure,
which essentially models the dynamic additional loading of
the sample in situ such as may be caused by blasting in
studying rockbursts (mining tremors). In intervals of
uniform loading, the autocorrelation coefficients display a
distinct trend (Table 3). Specifically, as the first auto-
correlation coefficient increases, the coefficients tend to
decrease linearly with time (the tendency characterizes the
increase of the correlation coefficient for approximation by
a straight line) and the correlation radius increases
(increase in the number of non-negative autocorrelation
coefficients).
5.3. Comparison of the results from different sensors
The data in Table 3 is calculated for the events recorded
by sensor number 2. Similar trends of the autocorrelation
functions as a function of the load are, however, also
observed for the events recorded by the other sensors,
although these did not record the same total numbers of
events (Table 4). This result is illustrated in Fig. 4a and b,
where the autocorrelation functions for the events recorded
by the separate sensors, determined for sub-interval C in
Fig. 3 (the load corresponds to 91.9% of the compressive
strength) and for sub-interval D in Fig. 3 (immediately
prior to total failure) are presented.
5.4. Cyclic loading
Sandstone, which differs in composition as well as
structure from migmatite, is used as rock material in the
cyclic loading tests. The experiment is conducted under
rapid loading of the sample with a shape of a
100  100  200 mm right parallelepiped. Previous experi-
ence [66] has shown that the results of autocorrelation
analysis under rapid loading do not display the unique
trends observed in the slow tests. The reason is that the
short-term loading experiment involves the immediate
reaction of the sample material to the acting force without
the effects of the structure and especially the rheological
properties of the rock.
In the first loading cycle at a constant loading rate, a
loading of 85% of the compressive strength is reached
at approximately 2200 s. The sample is kept at this level
of stress for 800 s. Afterwards, the sample is unloaded
over a period of 600 s (Fig. 5). Correlation functions of
the recorded sequences of AE events are computed in
selected sub-intervals. These curves indicate that the
parameters of the autocorrelation function display no
pronounced trend. The parameters of the autocorrelation
function corresponding to the interval up to the maximum
state of stress reached in the first cycle, that is, 18 MPa
(87.5% of compressive strength) are given in Table 5.
The curve of this autocorrelation function is marked A in
Fig. 5.
 ARTICLE IN PRESS
1076 J. Vilhelm et al. / International Journal of Rock Mechanics & Mining Sciences 45 (2008) 1068–1081

Fig. 3. Long-term loading test, migmatite—upper graph: axial and radial strain; middle graph: increase in cumulative rate of eventsPwith time (load). The
empty triangles denote the short fast load increase. Lower part: the detailed structure of the cumulative number of events N, the transformed
progression N/dT and autocorrelation function R(i) for selected sub-intervals A, B, C and D.

Table 3
Parameters of the autocorrelation analysis for a migmatite sample under long-term load (Fig. 3)

Sub-interval Beginning of Load s (% of Coefficient of Number of fitted Correlation R(1)

interval (s) strength) correlation points radius (s)

A 42,720 69.0 4 10 0.25

B 51,000 82.4 0.67 6 25 0.25
C 56,880 91.9 0.57 47 52 0.38
D 61,680 99.6 0.93 237 28 0.75

The second loading cycle takes 1100 s and ends with the added until total failure at a rate of 8.3 kPa/s. The
total failure of the sample. At the beginning of this cycle, a maximum load in the first cycle (18 MPa) is reached after
loading rate of 33.3 kPa/s is applied up to a load of 16 MPa 720 s from the start of loading (Fig. 6). The autocorrelation
(480 s from the beginning of loading). Further loading was analysis is performed in three sub-intervals B, C and D.
 ARTICLE IN PRESS
J. Vilhelm et al. / International Journal of Rock Mechanics & Mining Sciences 45 (2008) 1068–1081 1077

The loading interval, from the start of the second cycle predicting the time of total failure of rock samples under
till time 750 s, is affected by the Kaiser effect, i.e. the uni-axial pressure by means of autocorrelation analysis of
number of AE events is relatively small and the auto- AE. We proposed and tested a hypothesis of increasing
correlation functions do not provide information about the causality in the AE series as the sample approaches a
relation between the load and the AE. Table 5 gives the critical state.
parameters of the autocorrelation functions for sub- Under low stress, inhomogeneities of the stress field at
intervals B, C and D. The results indicate that the values pre-existing fractures and at the edges of grains are
of the correlation coefficient of the approximation line and generated only gradually. In these places, isolated at the
the first values of autocorrelation coefficients R(1) in sub- beginning, the strength of the material is exceeded locally
interval A are significantly smaller than in the sub-intervals and through AE the expansion of existing fractures can be
close to total failure. observed, as well as the generation of new microfractures.
Values of autocorrelation coefficients near zero are usually
6. Discussion observed, indicating a more or less random occurrence of
AE events. As the stress increases, the number of locations
The experiments concentrated on looking for signs of where the strength has been locally exceeded and the
impending failure and on assessing the possibility of separate fractures or locations of failure may affect one
another. The values of the initial autocorrelation coeffi-
cients increase, the lengths of the intervals over which
Table 4 correlation is observed increase. The decay of positive
Analysis of sub-intervals C and D from Table 3 for all channels
values of the autocorrelation coefficients has a linear trend.
Channel Number Coefficient of Correlation R(1) Samples of sandstone (sediments) and migmatites
of events correlation radius (s) (crystalline rocks) are tested, with the aim of generalising
the results obtained for of rock material. The loading is
Sub-interval C (beginning 56,880 s ¼ 91.9% of strength)
1 1371 0.49 59 0.32 carried out over a wide range of loading rates, so that the
2 1863 0.57 52 0.38 experiments took from several tens of minutes to several
3 1183 0.39 59 0.34 hours. We also applied cyclic loading and staircase loading
4 1662 0.48 60 0.45 to prove that the results are not dependant on the loading
Sub-interval D (beginning 61,680 s ¼ 99.6% of strength) pattern applied. In all cases, the AE is measured and the
1 9107 0.92 22 0.69 sequences of AE events recorded are subjected to
2 9471 0.93 28 0.75 autocorrelation analysis.
3 11,457 0.94 20 0.65
It was observed that, under fast loading emission of
4 9107 0.89 22 0.75
acoustic signals increases once the load exceeds 65% of the

Fig. 4. Comparison of the autocorrelation functions of AE recorded by four different pick-ups. Long-term loading of migmatite sample (see Fig. 3). Part
(a) corresponds to interval C and part (b) to interval D.
 ARTICLE IN PRESS
1078 J. Vilhelm et al. / International Journal of Rock Mechanics & Mining Sciences 45 (2008) 1068–1081

Fig. 5. Cyclic loading test—first cycle, sandstone. Upper graph: loading; middle graph: axial and radial strain; lower graph: increase in cumulative rate of
events with time (load). The autocorrelation function is shown for selected sub-intervals.

Table 5
Parameters of the autocorrelation analysis for the sample of cyclically loaded sandstone (Figs. 5 and 6)

Sub-interval Beginning of interval (s) Load s (% of strength) Coefficient of correlation Correlation radius (s) R(1)

First cycle
A 2000 79.4 0.78 16 0.33
Second cycle
B 800 88.9 0.97 13 0.55
C 900 92.9 0.99 9 0.64
D 939 94.4 0.98 16 0.63

compressive strength. Under low loading rates, emission
increases only after the stress reaches 90% of the
compressive strength. Autocorrelation analysis of AE can
only be applied after these loads have been reached.
The principal parameters of the autocorrelation function
considered are the first values of the autocorrelation
coefficients R(1), the number of positive autocorrelation
coefficients (correlation radius) and the tendency of the
autocorrelation function to depend linearly on time. This
trend is expressed quantitatively in terms of the correlation
coefficient of the approximation line and it is affected also
by the number of events in the sub-interval tested. Thus,
this parameter also reflects the increase in the number of
AE events with increasing load during the interval within
which the sample is in a critical state.
Under short-term loading, the value of the correlation
coefficient of the approximation line increases significantly
after the load has reached 90% of the compressive strength.
The value of the first autocorrelation coefficient R(1) also
increases. However, the correlation radius remains practi-
cally the same under different loads (above 90% of the
compressive strength).
In the long-term tests, under constant loading rate, the
parameters studied display a nearly monotonic increase.
Under cyclic loading, it was found that the autocorrela-
tion method could be applied to a number of AE events in
 ARTICLE IN PRESS
J. Vilhelm et al. / International Journal of Rock Mechanics & Mining Sciences 45 (2008) 1068–1081 1079

Fig. 6. Cyclic loading test—second cycle, sandstone. Upper graph: loading; middle graph: axial and radial strain; lower graph: increase in cumulative rate
of events with time (load). The autocorrelation function is shown for selected sub-intervals.

the second cycle only after the maximum load from the first
cycle had been achieved. The actual case mentioned
involved the short-term cyclic test. The time from the
beginning of loading in the second cycle until final failure
amounted to 20 min. This result, obtained on sandstones,
corresponds to the results of the short-term loading of
migmatite samples.
The nature of the AE, generated in the short-term tests,
is affected primarily by the immediate reaction of the
sample to the force applied. In the long-term tests, the AE
is then also affected by the rheological properties of
the rocks.
The increase in the values of the autocorrelation
coefficients R(i) with increasing load and the tendency to
linear decrease is an evidence of the increased effect that
the individual AE sources have on each other. This is
related to redistribution of stress within the sample, due to
the brittle microfracturing that generates the AE. This is in
agreement with the idea of mutual fracture interaction
during secondary and nucleation stage [62].
The problem of choosing a suitable length for the
subinterval and the mean value of the transformed
sequence of the number of events appears to be relatively
difficult. In the examples given, we chose the same length of
sub-intervals, DT ¼ 120 s and the same mean value of four
events per time dt (the length of the elementary interval) to
allow comparison of results of autocorrelation analyses at
different load levels and from different materials. Existing
experience indicates that the length of the sub-interval (and
subsequently the mean of the transformed sequence of the
number of events) must be chosen, so that the number of
elementary intervals is sufficient to carry out the auto-
correlation analysis, that is, the sub-interval should contain
at least 100 elementary intervals. The length of the sub-
interval, however, need not be chosen, so that the
exponential increase of the cumulative number of events
is evident. It is a significant advantage of our approach, in
comparison to, for example, rate-dependant material
failure by Voight [67,68] and Voight and Cornelius [69].
Their approach is based on the transformation of the
complete cumulative curve to the asymptotic power law
time-to-failure which is used for failure prediction. As
mentioned by Anifrani et al. [70], a disadvantage of this
approximation by a power law time-to-failure is narrow-
 ARTICLE IN PRESS
1080 J. Vilhelm et al. / International Journal of Rock Mechanics & Mining Sciences 45 (2008) 1068–1081
ness of the domain of its validity. According to Anifrani
et al. [70], Sornette and Sammis [23], Sammis and Smith [14],
this problem can be solved by applying a scaling theory of
rupture in terms of complex fractal exponents (correspond-
ing to log-periodic corrections). In this case, it is possible to
use a smaller number of input data for the prediction to carry
out the test over a broader interval of loads prior to ultimate
strength, and to consider time variable conditions of the load
applied (deformations). Autocorrelation analysis can also be
used in relatively broad interval of loads prior to ultimate
strength and allows relatively short intervals of AE
measurements to be used to estimate the state of failure.
This follows from the similarity of the two last named
methods. Anifrani et al. [70] claim that log-periodicity
implies a subtle correlation between successive AE hits
measured at constant pressure rate. This correlation, also
mentioned by Lei et al. [62], is the substance of the method
we tested, based on the autocorrelation function.
The result of the autocorrelation analysis, however, also
includes an assessment of the exponential increase in the
number of AE events. An important feature of the
precursory character of autocorrelation analysis is that
the resulting parameters are absolute and that they need
not be mutually compared in the course of the whole
loading process.
On the whole, it may be said that, under slow loading
lasting more than 10 h, the monitored parameters of the
autocorrelation function are of a precursory nature. If the
correlation coefficient of the approximation lines increases
above 0.9, the correlation radius approaches 20 s, and the
value of the first autocorrelation coefficient exceeds 0.7, it
can be said that the rock is being stressed close to its critical
state. The possibility of generalisation of these conclusions
was tested by comparing the analyses of the sequences of
AE events recorded by four independent sensors. In all
cases presented, although the numbers of events recorded
varied, the same results are obtained.
The conclusion drawn from testing the samples may be
useful in field measurements connected to research into
mining-induced seismic events, in which the seismo-
acoustic method is used.
Acknowledgements
The authors would like to thank the anonymous
reviewers for very valuable and helpful comments and
revising the English style of the manuscript, which
significantly improved its readability. This work was
partially supported by the Grant Agency of the Czech
Republic, project no. 205/06/0906, Research Project of
Ministry of Education no. MSM0021620855 and Institu-
tional Research Plan no. AVOZ 30130516.
References
[1] Lei X, Kusunose K, Rao MVMS, Nishizawa O, Satoh T. Quasi-static
fault growth and cracking in homogeneous brittle rock under triaxial
compression using acoustic emission monitoring. J Geophys Res
2000;105(B3):6127–39.
[2] Rudajev V, Čı́ž R, Lokajı́ček T, Vilhelm J. Geological structure,
seismic energy release and forecasting of rockbursts occurrence. In:
Proceedings of the 9th international congress on rock mechanics,
Paris, 1999. p. 1119–22.
[3] Bowman DD, Ouillon G, Sammis CG, Sornette A, Sornette D. An
observational test of the critical earthquake concept. J Geophys Res
1998;103(B10):24359–72.
[4] Main IG. Applicability of the time-to-failure analysis to accelerated
strain before earthquakes and volcanic eruptions. Geophys J Int
1999;139:F1–6.
[5] Davidson J, Goltz C. Are seismic waiting time distributions
universal? Geophys Res Lett 2004;31:L21612.
[6] Schorlemmer D, Wiemer S. Microseismicity data forecast rupture
area. Nature 2005;434:539–42.
[7] Szwedzicki T. A hypothesis on modes of failure of rock samples
tested in uniaxial compression. Rock Mech Rock Eng 2007;40(1):
97–104.
[8] Zhao Y. Crack pattern evolution and a fractal damage constitutive
model for rock. Int J Rock Mech Min Sci 1998;35(3):349–66.
[9] Kawakata H, Cho A, Kiyama T, Yanagidani T, Kusunose K,
Shimada M. Three-dimensional observations of faulting process in
Westerly granite under uniaxial and triaxial conditions by X-ray CT
scan. Tectonophysics 1999;313:293–305.
[10] Wulff AM, Hashida T, Watanabe K, Takahashi H. Attenuation
behaviour of tuffaceous sandstone and granite during microfractur-
ing. Geophys J Int 1999;139:395–409.
[11] Couvreur JF, Vervoort A, King MS, Lousberg E, Thimus JF.
Successive cracking steps of a limestone highlighted by ultrasonic
wave propagation. Geophys Prospect 2001;49(1):71–8.
[12] Wei YJ, Xia MF, Ke FJ, Yin XC, Bai YL. Evolution-induced
catastrophe and its predictability. Pure Appl Geophys 2000;157:
1945–57.
[13] Feng XT, Seto M. Neural network dynamic modeling of rock
microfracturing sequences under triaxial compressive stress condi-
tions. Tectonophysics 1998;292:293–309.
[14] Sammis CG, Smith S. Seismic cycles and the evolution of stress
correlation in cellular automaton models of finite fault networks.
Pure Appl Geophys 1999;155:307–34.
[15] Tang CA, Chen ZH, Xu XH, Li C. A theoretical model for Kaiser
effect in rock. Pure Appl Geophys 1997;150:203–15.
[16] Liu HY, Kou SQ, Lindqvist PA, Tang CA. Numerical modelling of
the heterogeneous rock fracture process using various test techniques.
Rock Mech Rock Eng 2007;40(2):107–44.
[17] Pollock AA. Acoustic emission amplitude distributions. Int Adv
Nondestr Test 1981;7:215–39.
[18] Rudajev V, Lokajı́ček T. Influence of rock grain size on energy
of acoustic energy emission. In: Proceedings of the 22nd
European conference on acoustic emission testing, Aberdeen, 1996.
p. 105–10.
[19] Shibazaki B, Matsu’ura M. Transition process from nucleation to
high-speed rupture propagation: scaling from stick-slip experiments
to natural earthquakes. Geophys J Int 1998;132:14–30.
[20] Diodati P, Bak P, Marchesoni F. Acoustic emission at the Stromboli
volcano: scaling laws and seismic activity. Earth Planet Sci Lett 2000;
182:253–8.
[21] Bak P, Tang C. Earthquakes as a self-organized critical phenomenon.
J Geophys Res 1989;94:15635–7.
[22] Gutenberg B, Richter CF. Magnitude and energy of earthquakes.
Ann Geofis 1956;9:1–15.
[23] Sornette D, Sammis CG. Complex critical exponents from renorma-
lization group theory of earthquakes: implications for earthquake
predictions. J Phys I France 1995;5:607–19.
[24] Veverka J, Vilhelm J, Rudajev V, Lokajı́ček T. Ultrasonic emission as
a tool for evaluation of rock instability—laboratory study. Rock-
bursts and seismicity in mines—RaSiM5. S Afr Inst Min Metall
2001:191–5.
 ARTICLE IN PRESS
J. Vilhelm et al. / International Journal of Rock Mechanics & Mining Sciences 45 (2008) 1068–1081
[25] Rudajev V, Vilhelm J, Kozák J. Simulation of aftershock series. Int
J Rock Mech Min Sci 1994;31(3):253–60.
[26] Zang A, Wagner FC, Dresen G. Acoustic emission, microstructure,
and damage model of dry and wet sandstone stressed to failure.
J Geophys Res 1996;101:17507–21.
[27] Eneva M, Ben-Zion Y. Techniques and parameters to analyze
seismicity patterns associated with large earthquakes. J Geophys Res
1997;102(B8):17785–95.
[28] Hirata T, Satoh T, Ito K. Fractal structure of spatial distribution of
microfracturing in rock. Geophys J Int 1987;90:369–74.
[29] Lei X, Nishizawa O, Kusunose K, Satoh T. Fractal structure of the
hypocenter distributions and focal mechanism solutions of acoustic
emission in two granites of different grain sizes. J Phys Earth 1992;
40:617–34.
[30] Lockner DA, Byerlee JD. Precursory AE patterns leading to rock
fracture. In: Hardy HR, editor. Proceedings of the 5th conference on
acoustic emission, 11–13 June 1995. Clausthal-Zellerfeld: Pennsylva-
nia State University, Trans-Tech Publications; 1995. p. 45–58.
[31] Zang A, Wagner FC, Stanchits S, Dresen G, Andresen R, Haidekker
M. Source analysis of acoustic emissions in Aue granite cores under
symmetric and asymmetric compressive loads. Geophys J Int 1998;
135:1113–30.
[32] Mogi K. Study of elastic shocks caused by the fracture of
heterogeneous materials and its relations to earthquake phenomena.
Bull Earthquake Res Inst 1962;40:125–73.
[33] Scholz CH. The frequency–magnitude relation of microfracturing in
rock and its relation to earthquakes. Bull Seis Soc Am 1968;58(1):
399–415.
[34] Riznitschenko JV. Methods of detailed investigation of seismicity. In:
Proceedings of the Institute of Physics of the Earth. Moscow: USSR
Academy of Sciences; 1960 [in Russian].
[35] Holt J, Palmer IG, Goddard DJ. The interpretation of acoustic
emission signals from the deformation of low alloy steels. In:
Schallemission, Symposium der Deutsche Gesellschaft für Metallk-
unde. München, 1974. p. 24–44.
[36] Přikryl R, Lokajı́ček T, Li C, Rudajev V. Acoustic emission
characteristics and failure of uniaxially stressed granitic rocks: the
effect of rock fabric. Rock Mech Rock Eng 2003;36(4):255–70.
[37] Rudajev V, Lokajı́ček T, Čı́ž R. Structure elements of rocks and
energy frequency distribution of seismic phenomena. Acta Montana
Ser A 1996;9(100):121–30.
[38] Janssen C, Wagner FC, Zang A, Dresen G. Fracture process zone in
granite: a microstructural analysis. Int J Earth Sci 2001;90(1):46–59.
[39] Schorlemmer D, Wiemer S. Microseismicity data forecast rupture
area. Nature 2005;434:539–42.
[40] Kaiser J. Erkenntnisse und Folgerungen aus der Messung von
Geräuschen bei Zugbeanspruchung von Metallischen Werkstoffen
[Information and conclusions from the measurement of noises in
tensile stressing of metallic materials]. Arch Eisenhuttenwesen 1953;
24:43–5.
[41] Lavrov A. The Kaiser effect in rocks: principles and stress estimation
techniques. Int J Rock Mech Min Sci 2003;40:151–71.
[42] Guarino A, Garcimartı́n A, Ciliberto S. An experimental test of the
critical behaviour of fracture precursors. Eur Phys J B 1998;6:13–24.
[43] Kawamoto S, William RS. Acoustic emission and acousto-ultrasonic
techniques for wood and wood-based composites—a review.
Gen Technical report FPL-GTR-134. Forest Products Laboratory,
US Department of Agriculture, Forest Service, Madison, WI, 2002.
p. 16.
[44] Panasiyan LL, Kolegov SA, Morgunov AN. Stress memory studies in
rocks by means of acoustic emission. In: Proceedings of the
international conference on mechanics of jointed and faulted rock.
Rotterdam: Balkema; 1990. p. 435–9.
[45] Goodman RE. Subaudible noise during compression of rocks. Bull
Geol Soc Am 1963;74:487–90.
[46] Sondergeld CH, Estey LH. Acoustic emission study of microfractur-
ing during the uniaxial cycling of Westerly granite. J Geophys Res
1981;86(B4):2915–24.
1081
[47] Rudajev V, Vilhelm J, Lokajı́ček T. Complex analysis of acoustic
emission from loaded rock samples. In: Kishi T, Mori Y, Enoki M,
editors. Progress in acoustic emission VII. Sapporo: Japan Society for
Non-Destructive Inspection; 1994, p. 243–8.
[48] Filimonov YL, Lavrov AV, Shkuratnik VL. Prospects of memory
effects for stress measurement applications in rock salt. In: Pro-
ceedings of EUROCK 2001. Rotterdam: Balkema; 2001. p. 59–63.
[49] Konečný P. Acoustic emission during cyclic loading of carboniferous
rocks and the manifestation of Kaiser effect. In: Proceedings of
EUROCK 2000. Essen: Verlag Gluckauf, 2000. p. 649–52.
[50] Rudajev V, Vilhelm J, Kozák J, Lokajı́ček T. Statistical precursors of
instability of loaded rock samples based on acoustic emission. Int
J Rock Mech Min Sci 1996;33(7):743–8.
[51] Utsu T. A statistical study on the occurrence of aftershocks. Geophys
Mag 1961;30:521–605.
[52] Lei X. How do asperities fracture? An experimental study of
unbroken asperities. Earth Planet Sci Lett 2003;213:347–59.
[53] Kawada Y, Nagahama H. Cumulative Benioff-strain release,
modified Omori’s law and transient behaviour of rocks. Tectono-
physics 2006;424:157–66.
[54] Wiemer S, Katsumata K. Spatial variability of seismicity parameters
in aftershock zones. J Geophys Res 1999;104(B6):13135–51.
[55] Kornowski JJ. Active seismoacoustic method of rockburst hazard
assessment: fundamentals and practice. In: Gibowitcz SJ, Lasocki S,
editors. Rockbursts and seismicity in mines. Rotterdam: Balkema;
1997. p. 325–9.
[56] Lyakhovsky V, Ben-Zion Y, Agnon A. Distributed damage, faulting
and friction. J Geophys Res 1997;102:27635–49.
[57] Lyakhovsky V, Ben-Zion Y, Agnon A. A visco-elastic damage
rheology and rate- and state-dependent friction. Geophys J Int 2005;
161:179–90.
[58] Hamiel Y, Liu Y, Lyakhovsky V, Ben-Zion Y, Lockner D. A visco-
elastic damage model with applications to stable and unstable
fracturing. Geophys J Int 2004;159(3):1155–65.
[59] Ben-Zion Y, Lyakhovsky V. Analysis of aftershocks in a lithospheric
model with seismogenic zone governed by damage rheology. Geophys
J Int 2006;165:197–210.
[60] Rektorys K. Survey of applicable mathematics, vol. II. Dordrecht:
Kluwer; 1994. p. 942.
[61] Papoulis A. Probability, random variables and stochastic processes.
2nd ed. New York: McGraw-Hill; 1984.
[62] Lei X, Masuda K, Nishizawa O, Jouniaux L, Liu L, Ma W, et al.
Detailed analysis of acoustic emission activity during catastrophic
fracture of faults in rock. J Struct Geol 2004;26:247–58.
[63] Lokajı́ček T, Vlk J. Complex multichannel system for continuous
monitoring of AE and related parameters. In: Rogers LM,
Tschelliesnig P, editors. Proceedings of the 22nd European con-
ference on acoustic emission, 1996. p. 271–5.
[64] Yin XC, Wang YC, Peng KY, Bai YL, Wang HT, Yin XF.
Development of new approach to earthquake prediction: load/unload
response ratio (LURR) theory. Pure Appl Geophys 2000;157:2365–83.
[65] Rudajev V, Vilhelm J, Lokajı́ček T. Analysis of acoustic emission
from rock samples loaded to long term strength limit. In: Proceedings
of the 8th international symposium on nondestructive character of
materials, 15–20 June 1997, Boulder, CO, 1997. p. 523–8.
[66] Rudajev V, Vilhelm J, Lokajı́ček T. Characteristics of acoustic
emission from rocks under different loading patterns. In: Proceedings
of the acoustic emission testing. Technical University of Vienna,
1998. p. 48–52.
[67] Voight B. A method for prediction of volcanic eruptions. Nature
1988;332:125–30.
[68] Voight B. A relation to describe rate-dependent material failure.
Science 1989;243(4888):200–3.
[69] Voight B, Cornelius RR. Prospects for eruption prediction in near
real-time. Nature 1991;350:695–8.
[70] Anifrani JC, Floc’h SL, Sornette D, Souillard B. Universal log-
periodic correction to renormalization group scaling for rupture stress
prediction from acoustic emissions. J Phys I France 1995;5:631–8.