EARTHQUAKE FAULT MECHANISM PREDICTION USING AI ALGORITHMS APPLIED TO STRONG MOTION RECORDS

Earthquake focal mechanisms are the results of a moment tensor solution for the earthquakes, and several methods such as the first motions polarity of P-waves, waveform inversion, amplitudes of P and/or S waves, full wave studies, etc., are applied to seismograms for understanding the earthquake rupture and the slip direction, including strike, dip, and rake angles. However, those procedures still need much more human interactions and avoiding full automation and efficiency. On the other hand, the number of available strong-motion earthquake records worldwide has increased over time and are used in many earthquake engineering studies. Today, several registers are available at different epicentral distances, depths, soil types, and tectonic regions. Could it be possible to analyse strong ground motion records in non-traditional feature spaces, such as shape, texture, and pixel intensity, for predicting the possible earthquake fault mechanism? The existence of numerous predictive applications based on artificial intelligence algorithms is now available, which allows analysing a set of information and extracting evidence of parameters and behaviours that are repeated given certain circumstances. It includes the possibility of establishing neural networks that analyse the energy content variation over time in different frequency ranges, which is developed over time during the seismic rupture, and which could be detected in the acceleration records. However, developed approaches are encountered in the traditional time-domain analysis, which incorporates considerable data pre-processing for filtering and cleaning noise. This study aims to explore a non-traditional workflow to classify the strong ground motion signals in different output categories (types of earthquakes based on focal mechanism) by applying machine learning-based approaches. The proposed workflow is a manifold procedure including: (1) transforming each signal from the time domain (1D space) to the frequency domain using the frequency spectrogram to generate 2D image representation of the event; (2) segmenting the event´s pattern in the spectrogram image to compute several shape, texture, and pixel intensity based descriptors; and (3) feeding the best feature subset to four machine learning classifiers of different taxonomies such as random forest, support vector machine, naive Bayes, and an artificial neural network to classify each record. If this method is possible, it could predict earthquake focal mechanisms in a short period of time, permitting automating and efficiency.