Automatic downhole microseismic event location with an attention‑enhanced fully convolutional neural network

Abstract

Accurate and rapid localization of microseismic events in single‑well downhole data is highly significant for microseismic monitoring. Although traditional methods, such as diffraction stacking and grid search, achieve high localization accuracy, they are computationally expensive, limiting their applicability for real‑time microseismic monitoring. To address this limitation, we propose an attention‑enhanced fully convolutional neural network (FCN‑CBAM) for efficient single‑well microseismic event localization. The method ingests three‑component waveform data and outputs three one‑dimensional Gaussian distributions representing the probability of the source location along the X, Y, and Z axes. The model is trained using 11,500 theoretical samples generated using the geometry and velocity model of the field data. Compared to traditional grid search methods, which require picking arrival time, calculating back‑azimuth, and locating the 3D seismic source, the FCN‑CBAM model can predict the event locations for field data within seconds, with prediction accuracy comparable to traditional methods. Furthermore, velocity perturbation and signal‑to‑noise ratio tests are performed to demonstrate the robustness and efficiency of our method.