Deep learning for automated crack quantification with distributed fiber optic sensing: Addressing strain overlap and interface nonlinearity

Abstract

Distributed fiber optic sensors (DFOS) hold significant potential for automation in construction, particularly in identifying and quantifying cracks through strain distributions. However, interpreting these distributions is challenging, especially when strain peaks overlap and there is nonlinearity in the cable-structure interface. To address this problem, this paper develops a deep learning model, termed Physical-Constrained FiberNet (PC-FiberNet), to intelligently interpret strain distributions under multiple crack scenarios. PC-FiberNet accurately identifies the location and width of each crack while simultaneously estimating the material and interfacial parameters of the cable. These parameters facilitate the development of numerical models. To enhance the robustness and generalizability of the proposed model, a transfer learning approach is employed. The performance of PC-FiberNet is validated through extensive tests and simulations. This paper enhances the application of DFOS to structural health monitoring, offering an effective approach to crack quantification in a multi-cracking scenario.