Seismic events are brittle failures mainly attributed to the reduction in effective stress. They are typically nonstationary signals with small moments (known as microseismic events) and are frequently exposed to various types of ambient noise, resulting in low Signal-to-Noise Ratio (SNR) sensor readings. However, an efficient and accurate (micro)seismic monitoring system is highly demanded, especially for unstable slopes such as landslides. Unlike the volume of literature that focuses primarily on earthquakes or volcanic seismic activities, the research on unstable slope monitoring requires addressing more challenging signal recordings from events such as quake/slide-quake considering microseismic signals, rockfall, earthquake, and anthropogenic noise. Thus, advanced signal processing approaches must be regarded as for signal denoising, event detection, feature engineering, and classification. This thesis first proposes an end-to-end platform containing denoising via Graph-Based Bilateral Filter (GraphBF), detection via Neyman-Pearson lemma, and classification via Graph Laplacian Regularisation (GLR) to identify the potential (micro)seismic events from raw observations. Secondly, an evaluation of feature engineering for (micro)seismic signal classification is proposed; the contribution concentrates on feature space optimisation with graph learning. Eventually, this thesis proposes a novel deep learning-based multitask learning to classify the (micro)seismic with little domain knowledge. For all proposed methods, the competitive performance is demonstrated in terms of accuracy and efficiency compared to state-of-the-art approaches, with the datasets collected at ongoing landslides.
|Date of Award||26 Apr 2023|
- University Of Strathclyde
|Sponsors||University of Strathclyde|
|Supervisor||Lina Stankovic (Supervisor) & Stella Pytharouli (Supervisor)|