A Generic Machine Learning Procedure for Detecting Seismic Events

Abstract:

Monitoring the subsurface is critical for evaluating hazards related to induced seismic events. Small or micro seismic events are difficult to detect using conventional methods as the amplitudes of signals are small. We developed two machine learning models to create an efficient and generic procedure to automatically detect and locate seismic events. Our algorithm is based on a convolution neural network and long short-term memory network. We use spectrograms of signals as inputs for both models to classify the seismic signal and locate their position in a confined region. The model is trained on millions of seismic records in Oklahoma at first. The performance of our algorithm is demonstrated on continuous dataset from Oklahoma, where seismicity has increased dramatically over the last decade. In addition, the parameters of the models are fine-tuned with local events in Decatur, Illinois. The tuned models show successful results when being applied to continuous data, demonstrating regional applicability of the models. Our procedure of seismic detection and location could be easily transferred to other study regions with light workload.