2026 Theses Master's

Using Minimal GRU Networks for Cross-Domain Damage Detection in Structural Health Monitoring

Starling, Riley

Identifying damage patterns in in-service structures for early maintenance and prevention remains a challenging task due to the limited availability of reliable, labeled operational data. Experimental datasets help bridge this gap by providing sufficient data to train reliable models; however, for practical deployment, these models must generalize effectively to out-of-domain structures and conditions. In this work, we address the generalization challenge by training neural network models on experimental structural data and evaluating their performance on the Z24 Bridge Benchmark. Structural vibrations are represented using cepstral coefficients, which capture underlying frequency characteristics and are used to extract damage-sensitive features. This study investigates two models: the first, a Minimal GRU (MinGRU), leverages a simplified recurrent architecture to efficiently learn temporal dependencies, while the second, a Bidirectional Minimal GRU (BiMinGRU), extends this approach to incorporate forward and backward processing to capture temporal dependencies in both directions. Model embeddings are analyzed using probabilistic linear discriminant analysis and evaluated through hypothesis-based damage detection. The models are initially evaluated on damage classification within the training domain, followed by cross-domain testing, in which deviations from the baseline condition are used to determine the damage state.

Keywords: Structural Health Monitoring, Deep Learning, Recurrent Neural Networks, Cepstral Coefficients, Damage Classification