Unsupervised machine learning for bridge SHM using FPGA: proof of concept via full-scale experiments

Bridges are a critical component of civil infrastructure, enabling the movement of people, goods, and services. However, their structural integrity can degrade over time due to aging, overloading, and environmental factors. Structural health monitoring (SHM) is essential to ensure continued safety and performance. Traditional SHM systems often rely on costly setups, including sensors, data acquisition systems, and computing tools, which limits their practicality in remote or resource-constrained environments. To address these limitations, this study proposes a compact, hardware-based SHM solution using unsupervised machine learning (ML) implemented on a field-programmable gate array (FPGA). FPGA systems offer task-specific computation, low power consumption, and portability, making them suitable for real-time field deployment. A full-scale experiment was conducted on a decommissioned rural bridge (D041) in Nebraska, USA. Strain data was collected under controlled damage scenarios using vehicle loads. Principal component analysis (PCA) was implemented on the FPGA using singular value decomposition (SVD) of a covariance matrix derived from the strain data. The first principal component (PC), typically holding the most dominant damage features, was extracted from the first column of the left singular matrix. A normalized novelty index based on the first PC was defined to classify damage levels. Several parametric studies were conducted to assess the robustness and performance of the FPGA-based implementation. Results from the FPGA were compared against MATLAB's optimized SVD function to validate the approach. Overall, the study demonstrates that implementing unsupervised ML on an FPGA enables robust, real-time, and accurate damage classification for practical SHM applications.