DEVELOPMENT OF MATHEMATICAL SOFTWARE FOR AUTOMATED DETECTION AND CLASSIFICATION OF SEISMIC EVENTS IN REAL TIME

Abstract

The processes of automated detection and classification of seismic events in real time are studied. Modern seismological networks generate large volumes of continuous data from hundreds of seismic stations. The decision-making process for identifying and classifying seismic events requires continuous processing of diverse signals characterized by high noise levels, variability of waveforms, and complex geological structures of the propagation medium. To solve this complex technical problem, it is necessary to address a scientific problem that involves the development of special mathematical software (SMS) based on machine learning methods, specifically deep neural networks, capable of accounting for multifactor characteristics of seismic signals and providing automated detection of P- and S-wave arrival phases and classification of seismic event types (tectonic earthquake, explosion, noise). The article proposes SMS based on a hybrid deep neural network architecture combining a Convolutional Neural Network (CNN) for local spatiotemporal feature extraction with an Attention-based Bidirectional Long Short-Term Memory network (BiLSTM) for temporal dependency modeling. The model was trained on the STEAD benchmark dataset containing over 1.2 million labeled records. The proposed SMS achieves F1-scores of 0.95 and 0.92 for P- and S-phase detection, respectively, with mean absolute errors of 0.11 s and 0.14 s. Event classification accuracy reaches a weighted average F1-score of 0.95. The inference time of 42 ms per 30-second window on GPU is twice faster than EQTransformer with comparable accuracy, enabling
real-time processing. Cross-domain validation on the independent INSTANCE dataset confirmed high generalization
capability with F1-score degradation not exceeding 9%.time of 42 ms per 30-second window on GPU is twice faster than EQTransformer with comparable accuracy, enabling real-time processing. Cross-domain validation on the independent INSTANCE dataset confirmed high generalization capability with F1-score degradation not exceeding 9%.