APPLICATION OF SYNTHETIC WAVEFORMS FOR CREATING A DEEP LEARNING MODEL FOR IDENTIFYING FIRST ARRIVALS IN SEISMIC RECORDS

Abstract

The goal of this study was to develop a method for analyzing signal records from a group of stations acquired during microseismic monitoring using a neural network to identify the moments of first arrivals. The novelty of the proposed method lies in the use of Wavelet Scattering for feature extraction, which significantly simplifies the model training process compared to traditionally used convolutional neural networks. The Transformer architecture used in the model allows for information exchange between an arbitrary number of recording stations, thereby determining arrival times at individual stations based on information from the entire group. The model was trained on fully synthetic data. This approach, compared to training on real data, may lead to a model with better generalization capabilities, as synthetic data reflect the mechanics of the generation and propagation of elastic waves without being influenced by local conditions and signal registration features. Several data augmentation strategies were explored, including adding noise to increase the model’s robustness to noise. The best balance between noise robustness and model accuracy was achieved by adding Gaussian noise with random standard deviation. The results of this study can be used in the development of deep learning models trained on both synthetic and real seismic records.