A fast unsupervised deep learning algorithm using seismic records of a single station for roadside rockfall recognition

The steep terrain in mountainous areas poses a significant threat to people's safety due to frequent geological hazards(e.g., rockfall and slope collapse), making effective management, monitoring, and timely issuance of alerts and warnings are crucial for highway authorities. Previous studies focus on studying the rainfall thresholds for possible rockfall occurrence. Recently, machine learning using seismic signals has been applied to detect rockfall events and monitor rockfall activity. However, supervised machine learning algorithms have relied on predefined labels, and the limited accumulation of data makes predicting model reliability challenging. The time-consuming model training can limit the practical application of the above models. In response to both aforementioned challenges, we first selected the roadside slope with relatively high activity of rockfalls and earthquakes as the study site and installed a seismic station on the crest of the slope. Then, we use an unsupervised machine learning framework to reveal patterns from unlabeled data and cluster seismic signals in continuous seismic records in the single three-component seismic station. Using continuous seismic data over one year, our approach combines a deep scattering network, features extraction, and features cluster to detect structures of signal segments. To illustrate the framework, a deep scattering network performs convolution and pooling on the three-component seismic signal data to extract multiscale information and construct scattering coefficients. For feature extraction, four different algorithms were employed: principal components analysis (PCA), independent components analysis (ICA), singular value decomposition (SVD), and Uniform Manifold Approximation and Projection (UMAP). Subsequently, we cluster the primary features using unsupervised learning algorithms such as K-means and Gaussian Mixture Model(GMM). We demonstrate the group categories belonging to rockfall events with in-situ data time-lapse images and videos. An approach proposed in this study could achieve rapid model training for building on-site rockfall warning systems using only single-station seismic records. Our high capability recognition model of rockfall events is ready to be implemented globally with high rockfall activity.

 

Keywords: unsupervised machine learning, deep scattering network, rockfall, seismic records, on-site early warning