One-Class Support Vector Machines for Structural Health Monitoring on Wave Energy Converters

Abstract

A wave energy converter (WEC) is a device that generates electricity from the kinetic energy of waves and the Department of Energy estimates that these devices could contribute significantly to energy production. However, WECs are not currently economically feasible due to high costs of repair and maintenance. Structural health monitoring (SHM) could be used to reduce these costs and make WECs efficient and deployable in large scale. This study investigates SHM techniques for the fiber reinforced plastics (FRPs) used to construct WECs. Data is collected using piezoelectric transducers on sample FRP plates. One of the primary hurdles with this problem is that the underlying structure of the FRPs is different for each plate, which results in significant variability in the collected data between plates. To overcome this issue, we propose using one-class support vector machines (OCSVMs) that can be trained solely on baseline examples from each plate. The OCSVM then detects anomalies and classifies them as damage. The numerical experiments demonstrate that the OCSVM can outperform traditional support vector machines (SVMs). Further, the OCSVM is only trained on baseline data from the plate being monitored while the tradition SVM requires data from the plate being monitored and the auxiliary data, both baseline and damage conditions, from other plates.