Transcript 地電流データの ニューラルネットによる解析

Reduction of Train Noise
from Telluric Current Data
by Neural Networks
Kazuki Joe (System Designer)
Toshiyasu Nagao (VAN Method Advisor)
Mika Koganeyama (Neural Network Implementor)
Moyo Sugita (Visualization Implementor)
Nara Women’s University
Tokai University
Earthquakes in Japan
（1990-1994）
What is short-term earthquake prediction?
Definition of Earthquake
Earthquake
Aftershock
Aftershock
Aftershock
Time
Electro-Magnetic Phenomena
Earthquake
Dec. 1994
Feb. 1995
3 months
90 days
VAN method
N
electrode
50m
Short dipoles（30～200m）
1km
Long dipoles（several km）
Designed by Greek physicists
 enable to observe SESs

Telluric Current Data(TCD)

Feeble current that flows in the earth surface
– potential difference between 2 points by burying electrodes
in the earth

observation points
– 42 points (Tokai and Hokuriku area)
– 8 channels or 16 channels for each observation point
– observe every 10 seconds (8640 data on one day)

seismic electric signals(SESs)
Seismic electric signals(SESs)

Current changes before earthquake
– earthquake is a kind of events of
destroying rocks
– current flows before rocks are
destroyed
20～30 minutes one-way
amplitude
 find the signals by specialists

about 160 frames（27 minutes）
300 frames（50 minutes）
Case Study

Big earthquake in Greece, Pirgos city（in March, 1993）
– Seismic electric signal was detected before the earthquake.
– By the prediction, some part of resident are evacuated.
• half of buildings (about 4000 ridge) were destroyed completely or partially
• no casualties
effectiveness of TCD
Investigate TCD in Japan
Problem of the use of
VAN method in Japan

TCD components in Japan
other noise
train noise
（about 90%）
SES
TCD
SES
Characteristics of
Train Noise
Similarity of the shape
Regularity of the appearance
can be learned & recognized
by Neural Networks
6:10 6:46 7:26 8:06
TCD
Up-train
Down-train
Timetable (Nagano railway Matsushiro station)
Train noise reduction filter
- Basic Idea Train noise + SES
train noise reduction filter
constructed by
neural network
SES
Problem of Constructing
the Filter by Neural Networks

NNs require training and supervising samples
– the TCD with train noise and SESs are very rare (only
several ten cases)
– no TCD with the same SESs without the train noise
Generate training and
supervising samples artificially
Artificial Generation of Training
& Supervising Samples

Pre-processed TCD (LF components are cut)
Train noise
120～250frames（about 20～40minutes）
300 frames
Natural noise
300 frames
Artificial Generation of Training
& Supervising Samples


Training data
＋
＋
＋
Train noise Natural noise
300 frames
Supervising data
SES
Natural noise
300 frames
SES
Artificial Generation of Training
& Supervising Samples

More shift-tolerant neural network to time series data
– train noise and SES are shifted right for several points as shown
below
Train noise
Supervising data
Training data
Experiment Result

After the learning, only train noise from
unknown TCD data could be removed.
– unknown TCD is generated artificially by train noise and an SES
Demonstration
 Artificial
generation of TCD with train
noise and an SES arbitrarily
 Train noise reduction of TCD with SESs
 Train noise reduction of unknown TCD
Conclusion and Future Work
It turned out that just train noise can be removed
from TCD by neural networks
 Can be a big progress toward automatic short-term
earthquake prediction
 More learning with other observation points
 Design and implementation of SES detector
 Other method...
