| Summary: | Microseismic monitoring is an important tool to characterize the reservoirs and delineate the growth of small-scale fractures. In addition, microseismic events are crucial for describing detailed fault geometries, stress changes, and spatial-temporal evolution of seismogenic activities. Deep learning has been extensively and successfully utilized for seismic event detection and phase picking. In this work, we propose an integrated workflow of waveform denoising, event detection, and seismic phase detection based on convolutional neural network (CNN) and unsupervised clustering, aiming at identifying and classifying microseismic P- and S-wave arrivals accurately. First, we preprocess (e.g., bandpass filter) the continuous waveforms and extract statistical features, and then feed the features into CNN for deep feature extraction and learning. The microseismic phase detection is then performed using clustering from waveform feature distance (similarity), which is retrieved from both statistical features of waveforms and the deep features extracted by CNN, yielding microseismic P- and S-wave arrival detections. We use synthetic data with different source mechanisms and signal-to-noise ratios, along with field microseismic data collected from the Hengill Geothermal area in Iceland, to verify the effectiveness of the proposed workflow in detecting weak microseismic phases. The refined phase detections also improved the resolution of stacking-based source imaging.
|
|---|