Microseismic event detection and classification based on convolutional neural network

Abstract

Microseismic event detection is a key step in the processing and interpretation of hydraulic fracturing monitoring. Low signal-to-noise ratio (S/N) properties of microseismic data, which makes the event detection of conventional methods not effective enough. In conventional methods, microseismic event detection is based on the difference of amplitude, frequency and correlation between effective signal and ambient noise. We have developed an event detection based on convolutional neural networks (CNN). To verify the feasibility of the method, we automatically create 4000 synthetic microseismic event and random noise records, and the corresponding labels for each record, which training result shown that is a good event detection method. Because the field data set is usually continuously collected by multiple geophones, we cut the data to 2000 sampling points as input data. In the training stage, we divide the synthetic microseismic data into three categories, namely, the (1) microseismic event, (2) single-phase event, and (3) ambient noise event, which are 2000, 1000, 1000, respectively. After training, the network architecture automatically learns to calculate the important features for event detection and saves the parameters of the highest training accuracy model. Field data examples of 11 stages indicate that CNN event detection is more accurate and effective than short-term average and long-term average ratio (STA/LTA) detection, and the CNN method can automatically classify events.