Thermoset state estimation using infrared spectroscopy and predictive analytics
Title Granting Institution
Technische Universität Hamburg
Place of Title Granting Institution
By design fiber-reinforced plastics (FRP) show extraordinary specific strength and stiffness performance when compared to classic structural materials. Unfortunately, the extraordinary properties come at the price of complexity in design, manufacturing and operation. Most notably, manifold failure modes and complicated detectability of inherent damages represent a challenge for composite engineers until now. Throughout the life-cycle of composites, it is crucial to understand and track the material state in order to prevent premature failure. Due to the high-cost sensitivity of the composites sector implementing material health monitoring systems is usually not practical in application. Instead, non-destructive testing (NDT) methods are usually applied to assure structural integrity and damages within allowed limitations at planned maintenance events. Nonetheless, currently available NDT methods are insufficient to fully cover all possible damage scenarios. In this work a new approach to determine material properties of the epoxy resin component using Fourier-transform infrared spectroscopy (FTIR)-spectroscopy is presented. It is shown that by post-processing the acquired molecular information the material state can be determined and based on this material state inherent material properties such as strength can be predicted. First, starting with the evaluation of manufacturing parameters it is shown that by using classic chemometrics techniques FTIR measurements can be used to quantify the mixing ratio of an epoxy resin fast and reliable. Thereafter, material changes due to mechanical loading in creep are observed and described by applying advanced feature extraction methods. Finally, specimens exposed to thermal loads are analyzed and it is confirmed, that the residual strength can be determined solely by applying FTIR measurements and machine learning algorithms. Furthermore, it is demonstrated, that the method is suitable to reveal the material exposure history. It is shown that an accurate prediction of mechanical properties, as well as the processing and degradation parameters, can be derived from the measurements and significantly improved by applying data processing and machine learning methods.