Vibro-Acoustic Testing and Machine Learning for Concrete Structures Damage Diagnosis

Abstract

A nondestructive evaluation (NDE) methodology is investigated for localizing internal (hidden) micro/macro cracking in large concrete structures due to the alkali-silica reaction (ASR). Four objectives are pursued: (1) damage diagnosis using Vibro-Acoustic Modulation (VAM) for a plain concrete specimen with known locations of damage inducement; (2) damage localization for a plain concrete specimen using approximate (2-dimensional) physics model-informed machine learning (ML) models; (3) development of a multi-fidelity physics informed ML approach for damage localization using 2-dimensional (2D) and 3-dimensional (3D) finite element models; and (4) investigation of physics model-based ML and transfer learning for damage localization in plain and reinforced concrete specimens with distributed damage inducement (unknown damage locations). The primary contribution of this research is the adaptation of VAM for probabilistic damage localization in thick concrete specimens. First, we conducted laboratory experiments on concrete block specimens containing aggregates at known locations, with different levels of alkali-silica reactivity. We investigated multiple VAM-based diagnostic information processing methodologies, such as simple averaging, Bayesian information fusion, and physics-informed ML techniques. We examined how ML models trained using physics simulation data can be effectively utilized for damage localization in a concrete slab specimen by learning the complex relationship between the damage and the damage index. These ML models were trained using data obtained from finite element (FE) simulations of the VAM test procedure. The diagnostic ML models provide a probabilistic estimate of the damage location, eliminate the need for a user-specified damage index threshold, and enable the localization of damage in three dimensions. Given a computational resource budget, we first investigated the performance of models trained using only 2D FE data and then studied the performance of models trained using data from models of varying fidelities (2D FE, 3D FE, and experimental data). Lastly, we examined ASR damage localization using VAM in concrete slab specimens (both plain and reinforced) where the aggregates are distributed throughout the specimen; thus, the damage locations are unknown. The VAM and ML-based damage localization results were validated using petrographic tests on cylindrical cores extracted from the specimens. Our results indicate that VAM testing and diagnostic ML models could be successfully used to localize internal (hidden) cracks in concrete structures.