Optimisation and enhancement of deep learning model for seismo-acoustic event detection and framework development for future implementation of characteristics recognition

Abstract

The seismic network in the Netherlands is densely distributed over the Groningen gas field and close to populated areas due to the restricted landmass of the country. Combined with the history of natural gas extraction, the detection system is affected by man-made noise and struggles to pick up smaller seismo-acoustic events. This thesis describes the approach to optimise and enhance a convolutional neural network (CNN) event recognition model which can detect seismo-acoustic events from a single-station input time-series. Input types combinations are compared and the best performing CNN model is trained using low-processed time-series and spectrogram inputs. The model learns to recognise whether the input contains noise, a seismic event or an event of a different type (acoustic event, explosion, etc.) with a 98.8$\%$ accuracy, comparable to state of the art models in literature. Despite the high classification accuracy, the layer-specific training behaviours of the model are explored to find that deep hidden layers may be under-trained. The model's detection rates over randomly selected full days is compared to the current SeisComP3 detection system, showing a strong reduction in false positive hits but an inability to detect events with a magnitude of 0.8 or below. Recurrent neural networks architectures are also explored and preliminary results show a lower accuracy than the best CNN model. Finally, a proposal for the expansion of the project is discussed to create a workflow method composed of independently trained networks to work together and detect a seismo-acoustic event, identify its type and provide some predictions on major characteristics to aid in seismologists analysis.