- Slides: 5
Application of Multi-Layer Perceptron (MLP) Neural Networks in Identification and Picking P-wave arrival Haijiang Zhang Department of Geology and Geophysics ECE 539 Project Presentation
Introduction Ø P-wave arrival: characterized by a rapid change in the amplitude and/or the arrival of high-frequency energy. Ø Quickly detecting and accurately picking the first-arrival of a P wave is of great importance in locating earthquakes and characterizing velocity structure. Ø The prior study of ANN on seismic phase picking -Input (1) The absolute seismic data (Dai et al. 1997) (2) Different attributes such as planarity, polarization, etc. (Wang et al. , 1997) -Output (1) Noise: 0 1 (2) P-wave arrival: 1 0 -Picking rule (1) A characteristic function is constructed from the ANN outputs. (2) P-wave arrival is chosen as some characteristic point.
MLP: Identification of the P-wave arrival Ø Configuration -30 inputs: 20 th sample corresponding to P-wave arrival -2 outputs: corresponding to the noise and P-wave arrival -1 hidden layer: 5 nodes -Learning rate: 0. 1, Momentum: 0. 8 Ø Results -Training set: including 18 P-wave arrival and noise segments -Classification rate: 94. 5% -Testing set: including 58 P-wave arrival and noise segments -Classification rate: 82%
MLP: Picking P-wave arrival Ø The characteristic function Ø The onset is chosen as a point whose value is greater than a threshold. Ø But it is difficult to choose such a point!!! The first, the maximum, the middle? ? Ø Long term, mid-term and short term to improve the picking accuracy (Zhao et al. , 1999) Ø My strategy Use Akaike Information Criteria (AIC) picker to pick the onset
Practical Application and conclusions ØApplication -The algorithm is tested on some seismograms from SAFOD. -90% P-wave arrivals are detected and picked. ØConclusions -It cannot discard spikes or glitches. -It is not very sensitive to S/N ratio -Comparing with former methods, this algorithm can pick the P-wave arrival more accurately (within 15 ms)