Transferring Random Samples in Actuator Systems for Binary Damage Detection

Data-driven models can accurately estimate the condition of systems, for example a hydraulic actuator. However, maintenance on the system can lower the predictive ability of condition models by changing the marginal and conditional distributions of the data. In this study, we propose to use transfer learning to address this issue in the context of a hydraulic actuator. Transfer learning aims to use knowledge from one system to improve modeling in another. This work uses random sampling to transfer samples between actuator rebuilds to predict a binary indicator of system damage in a rebuilt actuator. Features are selected based on distributional differences. We find that successful transfer using random sampling can occur when features are selected appropriately. Also, transferring only the damage data allows the model to improve as more baseline data from the rebuilt actuator becomes available.