Data-driven prediction of failure loads in damaged FRP composites under four-point flexure

Abstract

This study investigates the use of machine learning (ML) as a tool to make predictions of the criticality of delamination damage in fiber-reinforced polymer (FRP) composites subjected to four-point flexural loading. An extensive experimental campaign was conducted on polyester-glass FRP specimens. Most specimens were manufactured with a polytetrafluoroethylene film inserted at one interlaminar location to simulate delamination damage. Damage size, damage location through the laminate thickness, and the number of plies in the laminate, were each varied in the test matrix. The strength of damaged specimens was normalized against the strengths of corresponding pristine reference specimens to obtain a measure of damage criticality. Data augmentation techniques were subsequently utilized on the experimental data to synthetically generate a larger dataset for training, validating and testing the ML model. Output predictions of specific strength from the ML model proved very accurate for both the training dataset and the test dataset, meaning the ML model can accurately and near instantaneously predict the specific four-point flexure strength of new delamination damage cases. The method presented could be expanded to include new specimen characteristics and loading scenarios, or be combined with non-destructive testing techniques to enable data-backed, rapid decision making when delamination damage is detected in asset maintenance programs. The results highlight the effectiveness of data-driven methods for predicting the failure loads and apparent static strengths of damaged FRP composites and provide information on the most influential delamination features affecting the strength of FRP under flexure loads.