| Earth dams and levees support industrial applications, flood defense, and irrigation infrastructure around the world. These structures naturally undergo erosion over time and massive erosion events can result in catastrophic failure. When these structures fail, disastrous floods can displace or even kill nearby residents. Thus, it is important to know when and where these structures have progressive internal erosion events so appropriate action can reduce the impact of the erosion. This work explores improvements on the performance of previous machine learning methods for the continuous health monitoring of earth dams and levees (EDLs). Specifically, we explore ensemble classification algorithms (Bagging, Boosting, and Random Forest) to combine the passive seismic data from multiple sensors; we note that previous work only considered the data from one sensor in the wired sensor network. By considering features extracted from the signals of multiple sensors, Boosting with support vector machines (SVMs) shows a 1.5 to 41.7% increase in F1 score over single support vector machine (SVM) models depending on the specific sensor chosen for the single SVM. We also explore the use of SVM models trained on data from distinct sensors for visualizing the locations of detected erosion events in the structure. |
