From laboratory experiments to in-service tests: Learning embeddings for damage identification in structural health monitoring

In-situ nondestructive structural tests provide invaluable insight into structural behavior under service conditions, but often are not directly applicable to damage detection in-service. We address the generalization gap between laboratory and in-service structures through (i) development of a neural network damage model learned from various laboratory test structures and (ii) identification of damage progression using this model for an in-service structure. We generate cepstral coefficients from accelerometer time histories across laboratory test structures and damage scenarios, which are classified by a time-delay neural network (TDNN) architecture to determine feature patterns associated with particular damage expressions. Classification of various damage classes over a large laboratory test structure corpus (the LANL SHM dataset, the IASC-ASCE benchmark, the U. Nevada & UCSD Seismic Bridge Columns test, the U. Houston NEESR Reinforced Columns test, the U. Illinois Urbana-Champaign 144-DoF Truss, and the Purdue Truss) provides the model a diversity of scenarios to cover the span of damage phenomena. Generalization of the damage model to in-service structures is performed through clustering of embeddings from this TDNN, which are more-general representations of damage in the laboratory structure domain. A model of the priors for labeled damage scenarios and corresponding embeddings is generated via probabilistic linear discriminant analysis (PLDA). We assess these techniques on the Z24 Bridge Benchmark and obtain strong results for both ambient and forced vibration tests, in (i) a-posteriori diagnosis of previous in-service damage to demonstrate the extent of separability through supervised clustering and (ii) hypothesis testing of the deviation of an unseen queried condition, from the baseline condition in the PLDA, spanned purely by laboratory test damage classes. Explainability is also investigated via the association of in-service damage with seen laboratory test structure damage conditions, demonstrating feasibility for interpretable output-only structural health assessment. • Lab test vibration measurements can serve as a basis for in-service damage detection. • A TDNN classifier is trained on lab damage cases as a damage-sensitive feature generator. • A PLDA model of lab TDNN DSFs allows hypothesis testing for DSFs of unseen structures. • The TDNN-PLDA system developed only on lab structures detects Z24 Bridge damage cases. • Diagnosis of in-service damage is possible using PLDA clusters of lab damage cases.