<p>This study proposes a physical reservoir computing approach for structural damage detection, in which the target structure itself serves as the physical reservoir. Physical reservoir computing is a computational framework that leverages the dynamics of physical systems as computational resources for processing time series data. The proposed physical reservoir in this study is a nonlinear feedback system, which involves frequency virtual nodes realized by dividing the measured response of the target structure into multiple frequency bands. The divided signals are then respectively passed through nonlinear activation functions, followed by being summed and fed back to the target structure. The structure with the frequency virtual nodes and nonlinear feedback acts as a high-dimensional physical reservoir even with a single pair of sensor and actuator, which enables damage classification by detecting the changes in its dynamical behavior. To validate the proposed approach, both numerical and experimental verifications on a thin plate and a thin-walled tube with a single pair of sensor and actuator were conducted. As a result, the proposed method achieved comparable damage classification accuracy to conventional neural networks, while significantly reducing the training and inference costs, demonstrating the potential of a novel framework of structural health monitoring based on the physical reservoir computing.</p>