A LINEAR MODELING APPROACH ON LTGP SIGNALS AND SEISMIC ACTIVITY

International Journal of Chaos, Control, Modelling and Simulation (IJCCMS) Vol.7, No.1, March 2018
DOI:10.5121/ijccms.2018.7101 1
A LINEAR MODELING APPROACH ON LTGP
SIGNALS AND SEISMIC ACTIVITY
Dimitrios Dechouniotis1
and Apostolos Ifantis 2
1
School of Electrical and Computer Engineering, National Technical University of
Athens, Athens, Greece
2
Department of Electrical Engineering, Technical Education Institute of Western
Greece, Patras, Greece
ABSTRACT
This study illustrates presents a set of the Long – Term Geoelectric Potential (LTGP) measurements that
are collected for experimental investigation in Western Greece during a five–year period (1993 – 1997).
During this period, many major destructive earthquake events occurred that caused human casualties and
extended material damages. The collection and processing of geoelectric measurements was done by an
automated data acquisition system at the Seismological Laboratory of the University of Patras, Greece.
This novel study considers seismic activity of this area as a typical linear dynamic system and the dynamic
relationship between the magnitude of earthquakes and Long – Term Geoelectric Potential signals is
inferred by the Recursive Least Square algorithm. The results are encouraging and show that linear
dynamic systems, which are widely used in modern control theory, can describe efficiently the dynamic
behavior of seismic activity and become a useful interpretative tool of seismic phenomenon.
KEYWORDS
LTGP signals, Seismic Activity, LTI Modelling, System Identification
1. INTRODUCTION
Earthquakes’ disastrous effects on human lives has motivated scientists putting more effort on
earthquake prediction based on precursor events than on earthquake forecasting in a probabilistic
manner. A great variety of precursor signals have been proposed through years, among which
geoelectromagnetic ones, captured through geoelectrical measurements over a broad frequency
spectrum, have been promising on an eventual prediction of damaging earthquakes [1], [2], [3].
Earthquake event occurrence can be considered as the outcome of preseismic geotectonic
variations reflected on the geoelectric potential. In this way, the behavior of Long-Term
Geoelectrical Potential (LTGP) signals can be correlated to an upcoming earthquake event of
great importance. Research on this direction has given interesting results.
The chaotic behavior of the LTGP as a seismic event precursor signal has been explored and
confirmed using Lyapunov exponents and Takens estimator [4], [5]. The fractal analysis of the
Ultra-Low Frequency LTGP has revealed that strong earthquakes are shown to be preceded by a
decrease of the spectral power law exponent approaching unity [6], [7]. Also, in an experimental
analysis of seismic activity in Western Greece, a possible positive correlation of the Hurst
exponent of LTGP difference signal to the changes of the spectral power-law exponent of the
seismic activity has been found prior to major events [8].
LTGP power spectrum features relation to seismic activity has also been extensively investigated.
Initially, the Short Time Fourier Transform has been used [9]. The encouraging results motivated
scientists to deploy other advanced time-frequency analysis tools, such as [10] and Teager Huang
Transform [11].

2
In this paper, a linear dynamical system is used to model the dynamic relation between the
magnitude of an earthquake and LTGP signals. This type of modeling has been broadly adopted
in modern control theory for many different fields, e.g., electrical or mechanical systems,
computer and network systems [18], robotics etc. Raw LTGP data were collected at the
Earthquake Prediction Section of the University of Patras Seismological Laboratory (UPSL) [12]
during a five–year period, 1993–1997, in which significant events occurred close to the
observation station. To the best of our knowledge this is the first study that considers the seismic
activity as a dynamical system. The produced results are encouraging and show that a linear
dynamical system can be an accurate mathematical model of an earthquake.
The rest of this paper is structured as follows; in Section 2 the acquisition system along with the
data are described. Section 3 provides the algorithm for identifying the dynamic model of an
earthquake. Results are presented in Section 4. Finally, conclusions are drawn.
2. DATA ACQUISITION AND MEASUREMENTS
The geological procedures prior to an earthquake event include the application of stress on rocks,
which causes the redistribution of pore fluids, fluid flow, and the piezoelectric effect of quartz.
These explain the generation of electrical signals precursor to earthquake events. The
measurement of the LTGP signals is done using two pairs of Pb-PbCL2 electrodes, perpendicular
to each other, in NS (ch0) and EW (ch1) direction. In each dipole pair, electrodes are placed in a
100m distance. Dipole electrotelluric signals are directed to an electronic VAN device, converted
to digital with a sampling rate of 3 samples/min. Channel signals are in further anti-alias filtered
using a Butterworth low-pass filter sampled with a 32-bit resolution and transferred to a PC in the
control room via a dedicated line for further monitoring and processing. Also, a pen recording
system is used for reliability and illustration purposes. The structure of the data acquisition
system is presented in Figure 1.
Figure 1. Data Acquisition System

3
Figure 2. Observed LTGP signals
In order to observe long term variation, the sampling (digitization) rate was in further decreased
to 1sample/hr. Thus, for the examination period of 1993-1997 in Western Greece a data set of
43824 points was obtained. The relevant signals of ch0 and ch1 are depicted in Figure 2.
Additionally, for the same period, the major earthquake events with magnitude Ms ≥ 4.8 close to
our recording station, as provided by the Institute of Geodynamics of National Observatory of
Athens [13], are recording for the exploration of their correlation to the precursor electrotelluric
signals. Figure 3 shows the exact location of these events, while Table 1 contains all the
necessary information about these major earthquakes.
Figure 3. Epicenter of major Earthquakes in Western Greece (1993-1997)

4
Table 1. Major Earthquakes occurred in Western Greece during the period 1993-1997
3. LINEAR TIME-INVARIANT STATE SPACE MODEL
In control theory, the dynamic evolution of a process is defined mathematically by a state–space
model. This kind of modeling is suitable for both continuous and discrete systems. In both cases,
a vector of inputs u consists of all control variables and the vector of states x contains all variables
that describe the dynamic evolution of the system. Generally, a state-space model is linear or non-
linear function of the state and input vectors f(x,u). In the case of discrete state-space model, a
difference equation denotes how the state variables of the next time interval evolve from the
current values of input and state vector, x(k + 1) = f(x(k),u(k)). A discrete model can be derived
directly from difference equations if they are available. On the contrary, the complexity, linear or
not, and the parameters of function f(x, u) can be selected so that an identification method
produce an explicit model from measurements of input and state variables. System identification
is an important area of systems theory [14] that provide black box methods for deriving state–
space model form system’s measurements. In this study, a Linear Time-Invariant (LTI) model is
chosen to describe the earthquake as a dynamic system, which mathematically defined by,
(1)
where is the state vector that contains ch0 and ch1 signals and is the input
vector that is the magnitude of the earthquake in every kth time interval respectively. The order of
the model is the number of the elements of vector xk. In order to increase the accuracy of the
model, past values of the input variables and states can be used. However, this increases the
complexity of the model. There is always a trade–off between the accuracy and complexity of the
model as the order of the model varies. The Recursive Least Square (RLS) algorithm [15] is the
most common identification algorithm for identifying linear model of processes in industry. In
this paper, this algorithm is chosen to compute the elements of matrices A and B of the LTI model

5
due to its simplicity and accuracy. Furthermore, it can be implemented online and does not
require significant computational resources.
4. VALIDATION AND RESULTS
In order to validate whether an LTI model can effectively capture the dynamics of an earthquake,
the acquired data set is separated geographically. In Figure 3, it is obvious that the major
earthquakes can be separated in three geographical areas. The first area, named Area1, is around
Lefkada island in North-Western Greece (Longitude: 20-21, Latitude: 38-39.5). The second area,
called Area2, lies on the North coast of Peloponnese (Longitude: 21.5-22.5 Latitude: 38-39),
which is closest to the data acquisition system. The marine area west of Peloponnese (Longitude:
20-22, Latitude: 37-38) is the third area of interest, called Area3.
The effectiveness of RLS algorithm is tested under two different scenarios. Initially, the RLS
algorithm is fed with the whole data set to produce a unique LTI model for the whole region of
Western Greece and an LTI model for each geographical area. Secondly, we focus on the major
earthquakes of Table 1. For each geographical area, a local data set is created by concatenating
measurements one week before and after a major earthquake. Then RLS algorithm used these
data sets to produce three LTI models for the week before the major earthquake, for the week
after the earthquake and for both weeks together for each geographical area. All derived models
are evaluated using the Best Fit Rate (BFR) [16], [17], that is a well-known metric of system
identification process. In order to produced unbiased results, different data were used for the
training and the evaluation process. For both scenarios, we derive second and four-order LTI
models. Higher order models do not improve the accuracy of the model. More specifically the
second and fourth order LTI models are given by Eq. (3)-(4) respectively,
Where 





=






k
k
k
k
ch
ch
x
x
1
0
2
1
,
Where












=












−
−
1
1
1
1
0
0
4
3
2
1
k
k
k
k
k
k
k
k
ch
ch
ch
ch
x
x
x
x
.
The RLS algorithm uses time-series of inputs and state variables in order to derive the
corresponding models. The recorded measurements of ch0 and ch1 with digitization rate one
sample per hour are the values of state variables. The input measurements are derived by the
earthquakes data set with digitization rate of 1 sample per hour. If more than one earthquake took
place during the one-hour interval, the magnitude of the strongest earthquake is used. If there is
no earthquake, the measurement of this time interval is set to zero.

6
As mentioned earlier, at the first scenario a single model corresponds to the whole data set and an
LTI model for each area respectively. Eq. (2) is the simplest LTI model that it can be used. A
set of 30000 measurements is used for training and the rest data are used for the evaluation of the
model.
Figure 4 shows that the RLS algorithm converges quickly to the final values and Figure 5
illustrates that the produced models are very precise. The results for the produced fourth order
LTI model are similar and they are omitted for brevity.
Figure 4. Convergence of Second Order LTI Model
Figure 5. Real Measurements and Produced Data by LTI Model

7
Table 2. First Scenario: BFR Score of the RLS Algorithm
Data Set Second Order Model BFR (%) Fourth Order Model BFR (%)
Entire 85.02 85.11
Area1 85.66 85.79
Area2 87.96 88.13
Area3 86.48 86.57
Table 2 presents the BFR score of the second and fourth order LTI model for the entire data set
and each geographical area respectively. For each area the accuracy of the LTI model is high and
independent of the distance from the data acquisition system. As it was expected, the fourth order
LTI model is slightly improved than second order model.
At the second scenario, our work investigates the accuracy of LTI model for the major
earthquakes. The entire data set for Area1 and Area2 consists of 1685 measurements and the set of
Area3 includes 2022 samples. The produced models are derived from a portion of two third of the
entire data set for each area and the rest measurements are used for validation. Figures 6-7 present
the final value of the elements of matrices A and B of the four order LTI model of Area3. Figure 8
depicts the difference between real measurements and the produced values by the derived LTI
model. Although that the training sets are relatively small, the algorithm converges quickly to the
final values of the models (Figures 6-7) and its precision is still very high (Figure 8).
Table 3 presents the BFR score of the second and fourth order LTI model for the major
earthquakes of each geographical area. The columns of this table correspond to models for one-
week period before the earthquake, one week after the earthquake and for both weeks
respectively. The upper part of the Table 3 corresponds to second order LTI model and the lower
part corresponds to the fourth order model. It is worth-mentioned that models corresponding after
the major earthquake or for the two-week period are the most accurate. Also, the fourth order
model is slightly precise than the second order model.
Figure 6. Fourth Order LTI Model’s Parameters (Matrix A)

8
Figure 7. Fourth Order LTI Model’s Parameters (Matrix B)
Figure 8. Real Measurements and Produced Data by LTI Model
Table 3 – Second Scenario: BFR Score of the RLS Algorithm
This study is a “reverse-engineering” method compared with [4]; because the magnitude of the
earthquake is the cause and the LTGP signals are the results. From the results of both scenarios, it
is inferred that LTI models can satisfactorily describe the dynamic relationship between LTGP

9
signals and the earthquake’s magnitude. It is validated that the produced LTI models are accurate
independent of the presence of major earthquakes and the distance from the data acquisition
system. This study can be used to simulate the seismic activity of a specific area. Furthermore, it
can be combined with other prediction tools from modern control theory in order to predict
destructive earthquakes using historical data.
5. CONCLUSIONS AND FUTURE WORK
To the best of our knowledge this is the first study that investigates the dynamic relation between
the magnitude of major earthquakes and LTGP signals. The widely used RLS algorithm is used to
derive 2nd
and 4th
order discrete-time LTI models and tunes their parameters. The initial results
show that the derived models are precise independent of the presence of major earthquakes and
the distances from the data acquisition location. This kind of modeling combined with prediction
methods from modern control theory seems promising in order to extract useful information
about seismic activity of an area and the forecasting of destructive earthquakes.
REFERENCES
[1] G.A. Sobolev, (1975), Application of electric method to the tentative short-term forecast of
Kamchatka earthquakes. In Earthquake Prediction and Rock Mechanics, pages 229–235. Springer.
[2] T. Kawase, S. Uyeda, M. Uyeshima, and M. Kinoshita, (1993), Possible correlation between
geoelectric potential change in Izu-Oshima island and the earthquake swarm off the east Izu
peninsula, Japan. Tectonophysics, 224(1):83–93.
[3] G. Tselentis and A. Ifantis, , (1996), Geoelectric variations related to earthquakes observed during a
3-year independent investigation. Geophysical Research Letters, 23(11):1445–1448.
[4] A. Ifantis and K. Giannakopoulos, (2007), New experimental data reveal possible relation of chaotic
behavior of the long-term geoelectric potential difference to seismic activity in Western Greece.
Chaos, Solitons & Fractals, 34(3):717–726.
[5] K Giannakopoulos and A Ifantis, , (2008), Chaotic behavior of new experimental data of the LTGP
(2004-2006) confirm possible relation to seismic activity in western Greece. In Journal of Physics:
Conference Series, volume 96, page 012159. IOP Publishing.
[6] S. Uyeda, T. Nagao, Y. Orihara, T. Yamaguchi, and I. Takahashi, (2000), Geoelectric potential
changes: Possible precursors to earthquakes in Japan. Proceedings of the National Academy of
Sciences, 97(9):4561–4566.
[7] L. Telesca, V. Cuomo, V. Lapenna, and M. Macchiato, (2001), A new approach to investigate the
correlation between geoelectrical time fluctuations and earthquakes in a seismic area of southern
Italy. Geophysical Research Letters, 28(23):4375–4378.
[8] A Ifantis, C. Theoharatos, V. Tsagaris, and G. Tselentis, (2004), Experimental changes of the Hurst
exponent behavior of the geoelectrical potential and its possible relation to the changes of spectral
power-law exponent of the seismic activity in western Greece. In Signal Processing Conference, 2004
12th European, pages 777–780. IEEE.
[9] A Ifantis, G-A Tselentis, P Varotsos, and C Thanassoulas, (1993), Long term variations of the earth’s
electric field preceding two earthquakes in Greece. Journal Acta Geophysica Pol., 41(4):337–350.
[10] A. Ifantis, V. Nikolaidis, and G. Economou, (2007), An algorithmic investigation of power spectrum
features applied to geoelectric field signals and seismicity information. In Proceedings of the Fourth
conference on IASTED International Conference: Signal Processing, Pattern Recognition, and
Applications, pages 228–232. ACTA Press.
[11] K. Barkoula, A. Ifantis, and G. Economou, (2011), Long term geo-electric potential signal analysis
using the teager huang transform. In 8th IASTED International Conference on Signal Processing,
Pattern Recognition, and Applications, pages 245–252, 2011.

10
[12] UPSL, University of Patras Seismological Laboratory.
http://seismo.geology.upatras.gr/heliplots/about.htm, 2017.
[13] GEIN, Institute of Geodynamics of National Observatory of Athens. http://www.gein.noa.gr/en/,
2017.
[14] L. Ljung, (1999), System Identification (2Nd Ed.): Theory for the User. Prentice Hall PTR, Upper
Saddle River, NJ, USA.
[15] P. Wellstead and M. Zarrop, (1991), Self-Tuning Systems: Control and Signal Processing. John Wiley
& Sons, Inc.
[16] N. Leontiou, D. Dechouniotis, N. Athanasopoulos, and S. Denazis, (2014), On load balancing and
resource allocation in cloud services. In Proceedings of 22nd IEEE Mediterranean Conference of
Control and Automation (MED), pp. 773-778.
[17] D. Dechouniotis, N. Leontiou, N. Athanasopoulos, A. Christakidis, and S. Denazis, (2015), A
control‐theoretic approach towards joint admission control and resource allocation of cloud
computing services. International Journal of Network Management, 25(3), pp.159-180.
[18] D. Dechouniotis, N. Leontiou, N. Athanasopoulos, G. Bitsoris, and S. Denazis, (2012), ACRA: A
unified admission control and resource allocation framework for virtualized environments. 8th
International Conference on Network and Service Management (CNSM), pp. 145-149.
AUTHOR
Dr. Dimitrios Dechouniotis is a Senior Researcher at NETMODE Lab, NTUA Greece.
He received his diploma from the Department of Electrical and Computer Engineering
(ECE) at the University of Patras, Greece, in 2004, his Master of Science degree from
the NTUA, Greece, in 2006, and his PhD degree from the Department of ECE at the
University of Patras, in 2014. He was Partner Lecturer at the Electrical Engineering
Department of Technical Educational Institute of Western Greece during 2007-2016.
His research interests are network monitoring, cloud computing, pervasive computing
and control theory.
Dr Apostolos Ifantis received the B.S. degree on Physic Sciences in 1976 and the
Ph.D. degree in 1994 from the University of Patras, Greece. Currently, he is
Professor in the Department of Electrical Engineering at the Technological
Educational Institute of West Greece (Patras) and collaborating teaching staff at the
University of Patras. Also, he is responsible of the Center of Pre-seismic Signals of
the Laboratory of Geophysics and Seismology of the University of Patras. His
research interests include automation control systems, data acquisition, signal analysis
and digital processing, image processing, data mining and sensors. He has published
more than 40 Journal and International Conference papers in the above fields and has over 50 references to
his research work.

A LINEAR MODELING APPROACH ON LTGP SIGNALS AND SEISMIC ACTIVITY

More Related Content

Similar to A LINEAR MODELING APPROACH ON LTGP SIGNALS AND SEISMIC ACTIVITY (20)

Recently uploaded (20)

A LINEAR MODELING APPROACH ON LTGP SIGNALS AND SEISMIC ACTIVITY