Generalized neural network trained with a small amount of base samples: Application to event detection and phase picking in downhole microseismic monitoring

The deep learning method has been successfully applied to many geophysical problems to extract features from seismic big data. However, some applications may not have sufficient available data to directly train a generalized neural network. We apply data augmentation on a significantly small number of samples to train a generalized neural network for microseismic event detection and phase picking, which could be used in different project settings and areas. We use the U-Net architecture consisting of 2D convolutional layers to create the prediction function, and map the waveforms recorded by using multiple receivers to the P/S arrival time labels; thus, the neural network can learn the P/S moveout features from multiple receivers. The training set is generated by simulating various realizations of the data based on ten original samples from the beginning of a hydraulic fracturing stage. The trained neural network is then used to detect the events and pick the P/S phases from the continuous data for different stages and projects. A grid search from a precalculated traveltime table is performed to determine the event location after an event is detected. We build a real-time event detection and location workflow without human intervention by combining the neural network and grid search method, and apply the workflow to a different stage from the training events and a completely Geophysics independent project that the neural network has not encountered. The results show that microseismic events are successfully detected and located, and the picking performance of the neural network is superior to that of a traditional auto regression picker.