Application of convolutional neural networks for automatic detection of microseismic events in local monitoring data of the North Caucasus

Abstract

Relevance. The North Caucasus is characterized by a complex geological structure, a heterogeneous seismic background, and natural and anthropogenic noise, which complicate automatic detection of weak microseismic events. Traditional threshold-based algorithms do not always reliably detect low-amplitude signals, especially under low signal-to-noise conditions. Aim of the study. The study aims to evaluate the use of convolutional neural networks for automatic detection of microseismic events in three-component local monitoring records from the North Caucasus and to compare their performance with the STA/LTA algorithm. Research methods. Digital three-component seismograms from local stations were resampled to a common sampling rate, bandpass filtered, normalized, and divided into time windows. A one-dimensional CNN was used to classify events and noise fragments by analyzing three-component time series. Performance was evaluated using accuracy, recall, F1-score, and the number of false alarms in comparison with STA/LTA. Results. The most stable results were obtained for three-component records, a 20 s window, a 100 Hz sampling rate, and 1–20 Hz filtering. The CNN achieved an accuracy of 0.93–0.95, recall of 0.91–0.92, and F1-score of 0.92–0.93. Compared with STA/LTA, the model provided a better balance between detection completeness and false alarms, especially at signal-to-noise ratios below 2.0. The results confirm the promise of CNN as a primary detection module in seismic monitoring systems.