The aim of this thesis is to investigate the possibility of using artificial neural networks to develop automatic processing techniques for picking and identifying seismic arrivals. These two key procedures of an earthquake analysis system are extremely labour intensive, and any automation would allow further processing of larger datasets. Two approaches based on the back-propagation neural network (BPNN) are developed to pick and identify seismic arrivals for the dataset which includes 762 three component (3-C) recordings from stations DP and AY of a local earthquake network in Turkey. A BPNN approach was developed to pick arrivals automatically from 3-C recordings and single component (1-C) recordings. In Chapter 4, this approach is applied to the vector modulus (amplitude) of 3-C recordings. A BPNN trained by <I>P</I>-arrivals with high signal-to-noise-ratio (SNR) and background noise from station DP can extend its ability to picking <I>S</I>-arrivals and picking arrivals from other stations and from seismograms with low SNRs. It successfully detects 94.3% of the <I>P</I>-arrivals and 86.4% of the <I>S</I>-arrivals, compared with manual picks. The onset times of 74.5% of the <I>P</I>-arrivals and 63.2% of the <I>S</I>-arrivals are successfully picked with an error of 10 ms (one sample increment). In Chapter 5, this method is adapted to pick seismic arrivals from the absolute value (amplitude) of 1-C recordings. A BPNN trained by <I>P</I>-arrivals and background noise from the vertical component of station DP can extend its ability to the other two horizontal components and other stations. The picking rates are 93.1%, 89.4%, and 83.1% for <I>P</I>-arrivals, and 75.0%, 90.9%, and 87.2% for <I>S</I>-arrivals from the Vertical, E-W, and N-S components respectively. With an error of 10 ms, 66.2%, 59.2% and 63.3% of the <I>P</I>-arrivals, and 52.7%, 61.2% and 57.7% of the S-arrivals are picked from Vertical, E-W and N-S components respectively. In Chapter 6, another BPNN approach is developed to identify <I>P</I>- and <I>S</I>-arrival types from local earthquake data, using the modified degree of polarization (DOP) of 3-C recordings. Only the arrival segments are input in this BPNN approach for processing. Compared with manual analysis, a BPNN trained with nine groups of training <I>P</I>-arrival, <I>S</I>-arrival, and noise burst segments of DOP from station DP can correctly identify 82.3% of the <I>P</I>-arrivals and 62.6% of the <I>S</I>-arrivals from station DP. Another BPNN trained with five groups of training datasets from station AY can correctly identify 76.6% of the <I>P</I>-arrivals and 60.5% of <I>S</I>-arrivals from station AY. In order to understand how the BPNN works, a weight map is designed in this thesis to show the weight patterns of a trained BPNN. This new finding would be applied to any BPNN application, enabling illumination of the "block-box" approach of BPNN analysis. Applying this map to three trained BPNNs shows that it is a useful tool to investigate the interior and performance of BPNNs. For example, the weight map of a BPNN applying to pick arrivals from 3-C recordings shows that its weight pattern is divided into two portions which have different functions in picking.
| Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:649072 |
| Date | January 1995 |
| Creators | Dai, Hengchang |
| Publisher | University of Edinburgh |
| Source Sets | Ethos UK |
| Detected Language | English |
| Type | Electronic Thesis or Dissertation |
| Source | http://hdl.handle.net/1842/12173 |
Page generated in 0.0026 seconds