I n s t r u m e n t s

a n d

M e t h o d s

3. AN AUTOMATIC SYSTEM FOR SEISMIC SIGNAL
   
   IDENTIFICATION USINGFUZZY LOGIC
   
   The system we have set up forautomatic identification of
   the recordedthree-componentseismic signals proceedsin
   three steps (figure 1):signal analysis,information reduc-
   tion and decision. We will not discuss here the technical

Time domain: The histogram of the signal modulus is used
to discriminate very sharp signals,such as mining blasts
orfootsteps.A comparison between the smoothed signal
modulus and a typical earthquake model allows to recog-
nize short-rangeearthquakes.
Time-frequency domain:The ARCAP method,a com-
bination of Auto Regressive modelization and Capon esti-
mator (Dubesset et al., 1987), is used on a gliding window
to select the dominant frequencies and estimate the asso-
ciated power. We thus obtain a cloud of points represent-
ingthe powerdistributionofthesignalas afunction of
timeandfrequency.Thetime-frequencycontentofthe
signal allows to discriminate eventswith typical frequency
behavior (e.g. teleseismic or mid-range earthquakes, thun-
der).
Polarization domain:Capon's method isused to filter
each signalchannelat the dominant frequencies. A linear-
ity criterion is estimated on the filtered three-component
signals in order to locate lineargroundmotions in the time-
frequency plane.If a linear motion is detected in a given
signal window, its duration and azimuthare estimated. We
expect that this polarization study will allow us, in a near
future, to locateavalancheeventsin additionto recognizing
them.
The results of this signal analysis are the input of a sec-
ond programme,which estimates about 25features sum-
ming up the characteristics of the signal in each domain.
These features are derived from both the analysis results
and fuzzy sets (Zadeh, 1965). The fuzzy sets have been set
up accordingto the knowledge we have obtained from the
analysis ofpreviouslyidentified signals (cfpart 2).Each
feature is given a so-calledtruth value,that is,a number
between 0 (totallyfalse) and 1 (totally true).
Finally, the features are combined according to decision
rules by a third data fusion programme. These rules derive
from physical knowledge of both seismic waves generat-
ingphenomenaand propagation rules,aswellasmore
empirical knowledge resulting from the observation of the

Figure 1 Structure of the proposed system for automatic seismic sig-
nal identification.

details concerning each step. These are fully described in
(Leprettre, 1996) and (Leprettre et al., 1996) as well as in a
papersubmittedinJune1995toIEEETransactionson
Signal Processing by B. Leprettre, N. Martin, F. Glangeaud
and J.P. Navarre.

3.1.STRUCTURE OF THE ANALYSIS /DECISION

SYSTEM

The first programme performs a time-frequency-polariza-
tion analysis of the input three-component seismic signal:

FILE earthquake:8116 3C-samples
CRITERIA:truth value
Short signal:0.0
Steep ampl.hist.Z:0.0EW: 0.0NS:0.0
Earthquake shapeZ:0.0EW: 0.0NS:0.0
P-S frequency decrease:1.0
Global frequency decrease:1.0
Low mean amplitude:1.0
One component below 6Hz:1.0
One High Freq. component:0.0
Lowest comp.is Low Freq.:0.0
Helicopterfrequency:0.0
Spaced-out points:0.0
Broadband signal:0.0
Gentle ampl.hist.Z:1.0EW: 1.0NS:1.0
Farearthq.shapeZ:1.0EW: 1.0NS:1.0
Band-limited signal:1.0
Grouped points:0.6
CONCLUSION :Midrange earthquake:0.9
Avalanche:0.4
All others:0.0

130

Figure 2 Example of signal identification usingthe proposed system.