Voltage-based unsupervised learning framework for bridge damage detection in simultaneous energy harvesting and sensing systems

In this study, piezoelectric energy harvesters are designed to offer a dual functionality in Structural Health Monitoring: harvesting electric power from bridge vibrations while serving as intrinsic damage sensors. This strategy utilises the voltage signal directly as the sensing input, eliminating the need for traditional sensing modules, thereby achieving a reduction in system complexity and energy consumption. This study also presents a bi-objective optimisation framework that maximises both power output and damage-detection accuracy of a harvester modelled as a composite cantilevered Kirchhoff–Love plate. Voltage responses under realistic bridge inputs are predicted via isogeometric analysis. We validate our approach in two scenarios: a numerical vehicle–bridge interaction model and a laboratory-scale beam test using a toy car, each evaluated under both healthy and damaged states. Unsupervised damage detection is achieved with a convolutional variational autoencoder trained solely on healthy voltage signatures. We then apply the Non-dominated Sorting Genetic Algorithm II to explore trade-offs between energy yield and sensing precision, conducting parametric studies on damage severity, location, and harvester geometry. Results reveal that optimised devices not only act as effective filters and sensing components, but also outperform traditional acceleration-based sensing by enhancing damage detection accuracy by 13% while reducing energy consumption by 98%. The case with a multi-parameter design space highlights the importance of bi-objective optimisation, due to observed deviations in the performance of harvesters even under resonant conditions. These findings demonstrate the feasibility of replacing traditional sensors with lightweight, self-powered piezoelectric energy harvesters and pave the way for sustainable, Simultaneous Energy Harvesting and Sensing systems.