Use of transfer learning for detection of structural alterations

Abstract

Structural health monitoring is a discipline dedicated to the detection, identification, location and quantification of damage in structures based on performance indicators. Given the aim of the discipline, non-destructive method for data acquisition are always preferred. One such method is vibration-based testing, with which this work concerns itself. There has been widespread use of both machine learning applications when dealing with vibration data and use of computer vision-oriented machine learning models with pictures of the studied structure in order to address the concerns of structural health monitoring applications; this work propose a combination of the two. Since there are many pre-trained models for computer vision-oriented applications, this work successfully proposes a method for harnessing such models for processing of vibrational data through the use of transfer learning methodologies and finite element models. This can be achieved thanks to the visual nature of the Complex Frequency Domain Assurance Criterion (CFDAC) matrix, which can be obtained from vibrational data.