Case Study of Seismic Signals for Ghir Station before the Earthquake

In this paper, using two experimental results, chaotic signal modeling based on noise and detection of anomalies before an earthquake has been studied. First, a number of features such as frequency, statistical, entropy and FFT of the seismic signal were extracted, then the property matrix was investigated by multi-layered perceptron network (MLP). MLP can detect anomalies based on noise at a time interval of 5 minutes before an earthquake with an acceptable accuracy of 81.1404% for the magnitude 5-7 on the Richter scale from 21 earthquake recording stations in Iran (between 2004 and 2010). The recorded GHIR station signals were also investigated by this method, so MLP network determined the chaotic before an earthquake with an acceptable accuracy of 60.8696% for GHIR station. In this paper, it is tried to extract non-geologic properties from seismic signals of GHIR station and 21 stations. Then, the results of the MLP network are compared in terms of chaotic detection based on noise five minutes before an earthquake for each case. The results demonstrate that by using more recorded data of each station, MLP network is more capable for diagnosis of the chaotic based on noise before an earthquake.
The actual recorded seismic data is much better than simulated signals to detect seismic signal abnormalities before an earthquake and simulate the signal, in order to investigate the recorded signals of seismicity. In the recent study, recorded seismic signals were investigated.
The purpose of this study is to determine the most desirable characteristic matrix for detecting anomalies within five minutes before the earthquake.
Chaotic signal modeling based on noise and detection of anomalies before an earthquake using MLP network process contains four main stages: (1) Separate earthquake and slice signal, (2) The signals were processed using a high pass filter to reject the baseline drifts in signals, (3) Different types of features were extracted from the filtered signals and the feature vector was constructed, (4) The feature vector were fed to a neural network.
Our database contains 760 records for 21 stations and 148 records for GHIR station of IRAN earthquakes, from the International Institute of Earthquake Engineering and Seismology. The sampling frequency was 50 Hz.  . Fourth Order High Pass Butterworth Filter was used to annihilate low frequency (fc=0.04), and then the signal was normalized. The features were extracted from the filtered signals. Then the classifier was trained and tested. 70% of the data were randomly selected for training purposes and the rest 30% were used for testing for each experiment.
MLP networks were successfully trained using MATLAB (for both condition) and the results were presented for both conditions. The networks had one output in which every value at output uniquely representing each categories (1: earthquake, 0: no earthquake).
In this study, by using the results of two series of experiments, chaotic signal modeling based on noise and detection of anomalies before an earthquake has been investigated and the features matrix has been constructed for the following two conditions:

By using seismic signals in different geographic regions of the country (IRAN).
By using seismograph signals at a specific station (GHIR station).

The results show that MLP network produced better accuracy for 21 stations compared with GHIR station. Because the number of data was different in two ways or the features are non-geological. To achieve better results, the following suggestions are presented: noise reduction and use of more frequency features. We are now going to collect more and more earthquake data from all around the world, adding more chaos dependent parameters to the feature vector, using committee machines to increase the accuracy of classification and increasing prediction length from five minutes before to ten and more as well.