Various Signal Transformation for Optimizing CNN Model Performance of Earthquake Detection

Given the frequency of earthquakes in Indonesia, it is crucial to accurately identify them in order to prevent unnecessary alarms over their mitigation. In the context of earthquake detection and classification, the development of artificial intelligence is ongoing to achieve a dependable system. The objective of this work is to compare the signal transformations as a preprocessing step in order to achieve optimal performance outcomes for the deep learning model in detecting seismic signals. The signal transformations used in this study are Fourier, wavelet, Hilbert, and probability of power spectral density (PPSD) methods to obtain a visual representation of the signal spectrum. The unprocessed data obtained from the Indonesia Tsunami Early Warning System (InaTEWS) server underwent a classification process including three distinct categories: earthquake, background noise, and instrument noise. Seismic signal identification is performed by utilizing the signal spectrum obtained from preprocessing the raw data on the optimized convolutional neural network (CNN) model. The assessment revealed that PPSD represents the optimal signal transformation approach, achieving a peak accuracy of 97.2% under RMSprop optimization, whereas Adam optimization exhibits slightly superior precision and F1 scores of 97.3 % and 97.2 %, respectively. It is believed that these favorable results will improve the dependability of the earthquake detection system based on artificial intelligence, therefore minimizing the occurrence of detection errors that may result in public widespread panic and chaos.