Increasing number of seismic stations located in close proximity to active faults allows analysis of seismic signals that are recorded in near fault regions. Unique seismic signals, called “impulsive” or “pulse shaped” signals, are captured in velocity waveforms in numerous large magnitude earthquakes. In such waveforms, the earthquake is recorded as a one or several long period high amplitude signals. Long period signals are important in engineering seismology due to their large loads on structures. Ground motion prediction equations and design codes fail to capture the amplitudes in long periods of the impulsive signals. In this thesis nature of impulsive signals and their spatial distribution in near fault regions are investigated. To do that two different algorithm are developed in order to distinguish impulsive signals from non-impulsive signals. Moreover, the probability of the pulse shaped signal occurrence is estimated. In order to investigate the impulsive signals, near fault records from major crustal earthquakes are merged into a dataset. It contains waveforms that are coming from well known seismogenic zones. Waveforms in the dataset are also analyzed by implementing several previous studies to make comparison. The first pulse shaped signal classification algorithm is developed using wavelet analysis. Wavelet analysis decomposes the signal into time-frequency domain which provides the energy variation with time and frequency. The wavelet power spectrum of velocity waveforms are analyzed by using Ricker and Morlet wavelets. A threshold of minimum amplitude is applied. A comparison is made between the total energy of a signal and the energy of the time incidence where peak ground velocity is measured. Furthermore time incidence where maximum spectral energy is located in time is also taken into consideration. Energy ratios are used for determination of impulsive signals. It is found that a Ricker wavelet explains the features of the impulsive part of the velocity waveforms more accurately than the Morlet wavelet. It can measure the period of the pulse and the phase shift of the impulsive parts of the waveform. Spectral features of the impulsive signals are also captured successfully using a Ricker wavelet. The second classification algorithm uses convolutional neural networks. In order to train the convolutional neural networks, synthetic impulsive signals are created. A model is developed using real non-impulsive velocity waveforms from the dataset and synthetic impulsive waveforms. Impulsive signals are manually labeled as impulsive or non-impulsive. The trained model is run on the real manually picked impulsive signals of the dataset and the performance of the convolutional neural network, the wavelet method and various previously published methods are benchmarked. The convolutional neural networks approach correctly identifies almost 97% of the impulsive signals. Accuracy rate of the model is superior than other models. In order to understand the probability of the impulsive signals on earthquakes, a multi-variate Bayes classifier method is implemented on the dataset. Various information about the fault, earthquake and station are analyzed and 3 parameters that are correlated with the impulsive signals are used for the probability calculations. Probability models are developed for normal, reverse and strike slip faults. The validity of this model is tested on the data set. Developed models can provide pulse probability distributions without requiring earthquake-specific parameters. A relation between the period of the pulses and the moment magnitude is also developed.

Increasing number of seismic stations located in close proximity to active faults allows analysis of seismic signals that are recorded in near fault regions. Unique seismic signals, called “impulsive” or “pulse shaped” signals, are captured in velocity waveforms in numerous large magnitude earthquakes. In such waveforms, the earthquake is recorded as a one or several long period high amplitude signals. Long period signals are important in engineering seismology due to their large loads on structures. Ground motion prediction equations and design codes fail to capture the amplitudes in long periods of the impulsive signals. In this thesis nature of impulsive signals and their spatial distribution in near fault regions are investigated. To do that two different algorithm are developed in order to distinguish impulsive signals from non-impulsive signals. Moreover, the probability of the pulse shaped signal occurrence is estimated. In order to investigate the impulsive signals, near fault records from major crustal earthquakes are merged into a dataset. It contains waveforms that are coming from well known seismogenic zones. Waveforms in the dataset are also analyzed by implementing several previous studies to make comparison. The first pulse shaped signal classification algorithm is developed using wavelet analysis. Wavelet analysis decomposes the signal into time-frequency domain which provides the energy variation with time and frequency. The wavelet power spectrum of velocity waveforms are analyzed by using Ricker and Morlet wavelets. A threshold of minimum amplitude is applied. A comparison is made between the total energy of a signal and the energy of the time incidence where peak ground velocity is measured. Furthermore time incidence where maximum spectral energy is located in time is also taken into consideration. Energy ratios are used for determination of impulsive signals. It is found that a Ricker wavelet explains the features of the impulsive part of the velocity waveforms more accurately than the Morlet wavelet. It can measure the period of the pulse and the phase shift of the impulsive parts of the waveform. Spectral features of the impulsive signals are also captured successfully using a Ricker wavelet. The second classification algorithm uses convolutional neural networks. In order to train the convolutional neural networks, synthetic impulsive signals are created. A model is developed using real non-impulsive velocity waveforms from the dataset and synthetic impulsive waveforms. Impulsive signals are manually labeled as impulsive or non-impulsive. The trained model is run on the real manually picked impulsive signals of the dataset and the performance of the convolutional neural network, the wavelet method and various previously published methods are benchmarked. The convolutional neural networks approach correctly identifies almost 97% of the impulsive signals. Accuracy rate of the model is superior than other models. In order to understand the probability of the impulsive signals on earthquakes, a multi-variate Bayes classifier method is implemented on the dataset. Various information about the fault, earthquake and station are analyzed and 3 parameters that are correlated with the impulsive signals are used for the probability calculations. Probability models are developed for normal, reverse and strike slip faults. The validity of this model is tested on the data set. Developed models can provide pulse probability distributions without requiring earthquake-specific parameters. A relation between the period of the pulses and the moment magnitude is also developed.

Titolo: | Temporal and spatial analysis of near fault stations in terms of impulsive behavior |

Autori: | |

Relatore: | COSTA, GIOVANNI DE LORENZO, ANDREA |

Data di pubblicazione: | 20-mar-2020 |

Parole Chiave: | near fault; impulsive signal; wavelet analysis; machine learning; hazard model |

Settore Scientifico Disciplinare: | Settore GEO/10 - Geofisica della Terra Solida |

Anno Accademico: | 2018/2019 |

Ciclo: | 32 |

Editore: | Università degli Studi di Trieste |

Abstract: | Increasing number of seismic stations located in close proximity to active faults allows analysis... of seismic signals that are recorded in near fault regions. Unique seismic signals, called “impulsive” or “pulse shaped” signals, are captured in velocity waveforms in numerous large magnitude earthquakes. In such waveforms, the earthquake is recorded as a one or several long period high amplitude signals. Long period signals are important in engineering seismology due to their large loads on structures. Ground motion prediction equations and design codes fail to capture the amplitudes in long periods of the impulsive signals. In this thesis nature of impulsive signals and their spatial distribution in near fault regions are investigated. To do that two different algorithm are developed in order to distinguish impulsive signals from non-impulsive signals. Moreover, the probability of the pulse shaped signal occurrence is estimated. In order to investigate the impulsive signals, near fault records from major crustal earthquakes are merged into a dataset. It contains waveforms that are coming from well known seismogenic zones. Waveforms in the dataset are also analyzed by implementing several previous studies to make comparison. The first pulse shaped signal classification algorithm is developed using wavelet analysis. Wavelet analysis decomposes the signal into time-frequency domain which provides the energy variation with time and frequency. The wavelet power spectrum of velocity waveforms are analyzed by using Ricker and Morlet wavelets. A threshold of minimum amplitude is applied. A comparison is made between the total energy of a signal and the energy of the time incidence where peak ground velocity is measured. Furthermore time incidence where maximum spectral energy is located in time is also taken into consideration. Energy ratios are used for determination of impulsive signals. It is found that a Ricker wavelet explains the features of the impulsive part of the velocity waveforms more accurately than the Morlet wavelet. It can measure the period of the pulse and the phase shift of the impulsive parts of the waveform. Spectral features of the impulsive signals are also captured successfully using a Ricker wavelet. The second classification algorithm uses convolutional neural networks. In order to train the convolutional neural networks, synthetic impulsive signals are created. A model is developed using real non-impulsive velocity waveforms from the dataset and synthetic impulsive waveforms. Impulsive signals are manually labeled as impulsive or non-impulsive. The trained model is run on the real manually picked impulsive signals of the dataset and the performance of the convolutional neural network, the wavelet method and various previously published methods are benchmarked. The convolutional neural networks approach correctly identifies almost 97% of the impulsive signals. Accuracy rate of the model is superior than other models. In order to understand the probability of the impulsive signals on earthquakes, a multi-variate Bayes classifier method is implemented on the dataset. Various information about the fault, earthquake and station are analyzed and 3 parameters that are correlated with the impulsive signals are used for the probability calculations. Probability models are developed for normal, reverse and strike slip faults. The validity of this model is tested on the data set. Developed models can provide pulse probability distributions without requiring earthquake-specific parameters. A relation between the period of the pulses and the moment magnitude is also developed. |

Abstract: | Increasing number of seismic stations located in close proximity to active faults allows analysis... of seismic signals that are recorded in near fault regions. Unique seismic signals, called “impulsive” or “pulse shaped” signals, are captured in velocity waveforms in numerous large magnitude earthquakes. In such waveforms, the earthquake is recorded as a one or several long period high amplitude signals. Long period signals are important in engineering seismology due to their large loads on structures. Ground motion prediction equations and design codes fail to capture the amplitudes in long periods of the impulsive signals. In this thesis nature of impulsive signals and their spatial distribution in near fault regions are investigated. To do that two different algorithm are developed in order to distinguish impulsive signals from non-impulsive signals. Moreover, the probability of the pulse shaped signal occurrence is estimated. In order to investigate the impulsive signals, near fault records from major crustal earthquakes are merged into a dataset. It contains waveforms that are coming from well known seismogenic zones. Waveforms in the dataset are also analyzed by implementing several previous studies to make comparison. The first pulse shaped signal classification algorithm is developed using wavelet analysis. Wavelet analysis decomposes the signal into time-frequency domain which provides the energy variation with time and frequency. The wavelet power spectrum of velocity waveforms are analyzed by using Ricker and Morlet wavelets. A threshold of minimum amplitude is applied. A comparison is made between the total energy of a signal and the energy of the time incidence where peak ground velocity is measured. Furthermore time incidence where maximum spectral energy is located in time is also taken into consideration. Energy ratios are used for determination of impulsive signals. It is found that a Ricker wavelet explains the features of the impulsive part of the velocity waveforms more accurately than the Morlet wavelet. It can measure the period of the pulse and the phase shift of the impulsive parts of the waveform. Spectral features of the impulsive signals are also captured successfully using a Ricker wavelet. The second classification algorithm uses convolutional neural networks. In order to train the convolutional neural networks, synthetic impulsive signals are created. A model is developed using real non-impulsive velocity waveforms from the dataset and synthetic impulsive waveforms. Impulsive signals are manually labeled as impulsive or non-impulsive. The trained model is run on the real manually picked impulsive signals of the dataset and the performance of the convolutional neural network, the wavelet method and various previously published methods are benchmarked. The convolutional neural networks approach correctly identifies almost 97% of the impulsive signals. Accuracy rate of the model is superior than other models. In order to understand the probability of the impulsive signals on earthquakes, a multi-variate Bayes classifier method is implemented on the dataset. Various information about the fault, earthquake and station are analyzed and 3 parameters that are correlated with the impulsive signals are used for the probability calculations. Probability models are developed for normal, reverse and strike slip faults. The validity of this model is tested on the data set. Developed models can provide pulse probability distributions without requiring earthquake-specific parameters. A relation between the period of the pulses and the moment magnitude is also developed. |

Appare nelle tipologie: | 8.1 Tesi di dottorato |

###### File in questo prodotto:

File | Descrizione | Tipologia | Licenza | |
---|---|---|---|---|

PhD_Thesis.pdf | tesi di dottorato | N/A | Open Access Visualizza/Apri |