Open AccessPreliminary Results of Automatic P-Wave Regional Earthquake Arrival Time Picking using Machine Learning with Kurtosis and Skewness as The Input Parameters
Author(s) -
Yosua Hotmaruli Lumban Gaol,
Renan Lobo,
S S Angkasa,
Ahmad Fikri Abdullah,
I Madrinovella,
S Widyanti,
Awali Priyono,
S K Suhardja,
Andri Dian Nugraha,
Zulfakriza Zulfakriza,
Sri Widiyantoro,
Moad Hakim,
Kadek Hendrawan Palgunadi,
Bambang Mujihardi
Publication year - 2021
Publication title -
iop conference series. earth and environmental science
Language(s) - English
Resource type - Journals
eISSN - 1755-1307
pISSN - 1755-1315
DOI - 10.1088/1755-1315/873/1/012061
Subject(s) - kurtosis , waveform , skewness , arrival time , computer science , amplitude , artificial neural network , moment (physics) , pattern recognition (psychology) , artificial intelligence , geodesy , statistics , mathematics , geology , engineering , telecommunications , physics , radar , classical mechanics , quantum mechanics , transport engineering
The traditional method in determining first arrival time of earthquake dataset is time consuming process due to waveform manual inspection. Additional waveform attributes can help determine events detection. One of the widely used attribute is The Short Term Averaging/Long Term Averaging (STA/LTA) which is simply division moving average of waveform amplitude between short time and longer time. Alternatively, waveform attribute can also be built using kurtosis and skewness. The kurtosis attribute is defined as sharpness of data distribution. By definition, the maximum signal should be at or close to the P wave arrival. The skewness is typically used to show normal distribution of the data. The uniqueness of this method is that it has an ability to determine whether the first P wave arrival has positive of negative number. The skewness calculation is similar to kurtosis but it uses the power of 3 instead of 4. With the objective of generating efficient scheme to pick first time arrival, we explore use artificial neural network and a combination of kurtosis and skewness attributes. We use a collection of magnitude events with moment magnitude larger than 3 located close to Moluccas island, Indonesia. We collected all events information from the Indonesian Agency of Meteorology, Climatology and Geophysics. The process is started with manually pick all P wave arrivals using manual inspection. Next, we trained the artificial neural network with attributes numbers as inputs and arrival time we manually picked as the output. In total we used 100 regional events that has clear P wave phases. Then, we validated the results until reaching 0.99 accuracy. In the last step, we tested the once trained procedures on new waveforms contained events. Current result shows an average of 0.4s different between manual pick and trained picked from machine learning. The accuracy can be improved by applying a band pass 0.1-2 Hz filtering with an average of 0.2s.