Using a Convolutional Neural Network To Differentiate Between Events and Production-related Seismic Signals: A case study from the Sultanate of Oman

Seismic instrumentation is used to monitor seismicity near oil and gas activities. The instrumentation records different signal types such as microseismic events, operations, and noise, which must be classified in real-time. We use a convolutional neural network approach to classify triggered waveforms into these three different signal types. The input data for the neural network are spectrograms created from the recorded waveforms. We manually parsed recorded data to identify the best quality training dataset and tested for generalization with time. We found that the neural network achieves an overall accuracy of 97.33%, which includes a 75.00% reduction in misclassified event signals compared to a simplified template matching approach. Importantly, the data characteristics changed with time which required adjustment of the training procedure to achieve the highest accuracy, highlighting how thorough testing and monitoring performance are both important parts of any neural network implementation, particularly for systems that change with time. Using the neural network improves the accuracy of classification approaches in real-time and would allow operators to make decisions faster based on the recorded signals.