Integrating Data-driven and Model-based Algorithms for Structural Health Monitoring

Abstract

Artificial Intelligence has rapidly supplanted classical statistical methods for data analysis and prediction-making in many scientific areas in the last few years due to substantial technological advances in hardware that enable the robust dedicated algorithms to produce reliable results in reasonable software execution times. Looking specifically at applications in the field of Structural Health Monitoring, automated algorithms aim to replace visual inspections for structural condition assessments, as they can reduce maintenance costs and identify damage not otherwise detectable. In this paper, we propose a well-structured approach covering the entire monitoring process, overcoming the critical issues that available algorithms suffer from, including the availability of suitable training data and compatibilities among the solutions used for different tasks throughout the entire anomaly detection process. This is possible through the adoption of a standard data storage format and the numerical construction of a digital twin of the structure under inspection. This way, ad hoc baseline patterns may be generated to feed artificial neural networks that are truly supervised, and any alarms created can, therefore, be checked by running dynamic simulations of the corresponding FEM model.