Seismic Phase Picking Using a Cross-Attention Network on NVIDIA Jetson Xavier NX

This paper introduces a neural network model for seismic phase picking tailored for edge intelligence. The model architecture is meticulously designed to accommodate the resource constraints of edge computing platforms, enabling real-time seismic phase picking at distributed instruments through the simultaneous extraction and analysis of seismic features in both time and frequency domains. The model uses a cross-attention mechanism to effectively integrate feature representations from time series and spectrograms. This integration enhances the model's ability to capture seismic signal context, improving the identification and focus on key seismic attributes, thereby increasing picking precision. Furthermore, cross-attention learning bolsters the model's generalization capabilities across varying geological conditions and signal-to-noise ratios (SNR). Testing on the Pacific Northwest (PNW) seismic dataset demonstrates that, under identical training conditions, our model achieves a 2.4% improvement in P phase picking precision and a 2.9% improvement in S phase picking precision compared to the earthquake transformer (EQTransformer). Inference performance tests on the NVIDIA Jetson Xavier NX platform show that our model's inference time is 20.6% shorter than that of EQTransformer, validating its efficiency and accuracy on edge computing platforms.