Περίληψη
Ο βασικός σκοπός της συγκεκριμένης διατριβής είναι η ανάλυση και αναγνώριση της σεισμικής συμπεριφοράς των τεκτονικών ρηγμάτων. Πιο συγκεκριμένα ο χωρικός εντοπισμός των asperities σε τεκτονικά ρήγματα επετεύχθη με την συνδυασμένη εφαρμογή θεωριών της Στατιστικής Σεισμολογίας και μεθοδολογιών της Τεχνητής νοημοσύνης. Μέσω της στατιστικής ανάλυσης σεισμικών ακολουθιών προσδιορίστηκαν τα κατάλληλα στατιστικά χαρακτηριστικά που μπορούν να υποδείξουν τις τοποθεσίες των asperities. Τέλος τεχνικές μηχανικής μάθησης τροφοδοτούμενες απο τα στατιστικά χαρακτηριστικά εφαρμόστηκαν για τον εντοπισμό των asperities
Περίληψη σε άλλη γλώσσα
Machine learning is a technological artifact well-known for its proven ability to classify different aspects of the same entity in different categories. The main goal of the present dissertation is to use the proper machine learning technics that will achieve the classification of rupture patches onto active fault surfaces in Asperities and Non-Asperity areas. For this purpose, the earthquakes had to be studied with the statistical principles that governs them. Therefore, a big part of this dissertation was devoted to the utilization of statistic tools for the analysis of earthquakes and also statistical features that are well-known for behaving differently on Asperity and Non-Asperity rupture patches, such as the b-value, earthquake density and recurrence interval of magnitude classes. This part of the dissertation in addition to the presentation of these statistical attributes, also leads to the creation of Self-cor, a method of calculating robust b-values for earthquake catalogs th ...
Machine learning is a technological artifact well-known for its proven ability to classify different aspects of the same entity in different categories. The main goal of the present dissertation is to use the proper machine learning technics that will achieve the classification of rupture patches onto active fault surfaces in Asperities and Non-Asperity areas. For this purpose, the earthquakes had to be studied with the statistical principles that governs them. Therefore, a big part of this dissertation was devoted to the utilization of statistic tools for the analysis of earthquakes and also statistical features that are well-known for behaving differently on Asperity and Non-Asperity rupture patches, such as the b-value, earthquake density and recurrence interval of magnitude classes. This part of the dissertation in addition to the presentation of these statistical attributes, also leads to the creation of Self-cor, a method of calculating robust b-values for earthquake catalogs that contain a small number of earthquakes. This method was tested in the seismicity of Corinth Gulf and the results were compared with well-known methods which proved to be efficient. Furthermore, the hypothesis that the b-value and earthquake density as the number of generated earthquakes are behaving differently in Asperity and Non-Asperity areas was tested in this area. Spatial distribution of b-value and earthquake density showed that the attributes appear to have significant lower values than the Asperity surrounding areas. Having as a baseline of attributes to be used for a feature vector the earthquake density, b-value, and recurrence interval of earthquakes, several experiments were conducted. These experiments meant to determine the best combination of preprocessing technics such as feature selection and class balancing, with machine learning algorithms to achieve robust classification of Asperity areas. In the conducted experiments Information Gain was used for feature selection. both oversampling and undersampling technics such as SMOTE and Tomek links were tested for balancing unbalanced data. Test subjects on these experiments were the regions of Hokkaido in Japan and the Central Ionian Islands in Greece. In a first attempt to check whether machine learning can face the task of classifying Asperity areas, various machine learning algorithms were used with data from the region of Hokkaido. The feature vector used in the experiments contained the b-value, earthquake density, and location of Hokkaido’s areas. Also, the SMOTE technique was used to balance the training data between the Asperity and Non-Asperity classes. To evaluate the results, an asperity map of the area was used as ground truth, and the experiment proved that machine learning algorithms can face the specific task with Random Forest algorithm achieving the best results. In an extension of the previous experiment both oversampling and under-sampling techniques were tested to balance the data from the region of Hokkaido. The feature vector this time contained the b-value and earthquake density. The experiment concluded that the use of under-sampling with the removal of Tomek links was the most efficient method. In a consecutive study, the addition of the recurrence interval of earthquakes in the feature vector was tested. Again, the data from the region of Hokkaido were used and the feature vector consisted from the b-value, the earthquake density and the recurrence interval. The removal of Tomek links was used to balance the data. Eventually, the best results were achieved with the use of Logistic and FT algorithms. The last experiment of this dissertation tried to identify the Asperity areas in the area of Central Ionian Islands. For this experiment, data from the Hokkaido region were used as training set. In order to compensate the different geophysical properties of the two regions, the data were normalized. The algorithm Logistic and FT were trained with data from Hokkaido that were balanced with the removal of Tomek Links. Then the trained models were used to identify Asperity locations in the Central Ionian Islands.
περισσότερα