Machine-Learning-Based Vibration Classification of Inter-Module Connector Stiffness Loss in Modular Flood-Response Structures: A Simulation-Based Proof of Concept

Abstract

Ensuring the safety of modular facilities used for flood response and post-disaster sheltering requires reliable assessment of inter-module connections that may deteriorate under hydrodynamic loading, debris impact, repeated wetting, and corrosion. This study presents a simulation-based proof of concept for classifying connector residual stiffness from vibration response using machine learning. Flood-load analysis quantified hydrostatic, hydrodynamic, and debris-impact forces under five severity scenarios, which were hypothetically associated with seven prescribed connector-stiffness levels ranging from 100% to 5%. Only the stiffness-to-vibration relationship was physically simulated; the flood-exposure-to-stiffness relationship was assumed rather than experimentally validated. Finite element modal analysis in OpenSeesPy provided the first four natural frequencies and mode shapes, from which sixteen diagnostic features were derived. Gaussian perturbations generated 350 synthetic realizations from the seven deterministic stiffness states. To minimize data leakage, classifiers were evaluated using block-grouped cross-validation and compared with majority-class and second-mode-frequency baselines. Random Forest achieved the best multiclass performance, with a macro-F1 of 0.62 and balanced accuracy of 0.63, although early-stage deterioration classes remained difficult to distinguish. When reformulated as a binary severe/non-severe screening task, the model achieved 0.94 balanced accuracy with a 0.10 false-safe rate. These results demonstrate within-configuration separability under synthetic noise rather than field generalization. The proposed framework may support rapid condition screening of modular connections, but experimental validation of the degradation mapping, structural model, and sensing strategy is required before field implementation.