Semi-Supervised Anomaly Detection for the Identification of Damages in an Aerospace Sandwich Structure Based on Synthetically Generated Strain Data

The structural health monitoring (SHM) of safety relevant composite components is becoming increasingly relevant as it enables in-service diagnosis and data acquisition capabilities, contributing to the optimization and efficient operation of the overall system and ultimately saving costs and resources. In this field, machine learning (ML) techniques are attracting growing attention due to their capability to recognize complex patterns, making them very suitable for the identification of damages in operating mechanical structures. However, the acquisition of sufficiently large amounts of labeled and representative data from both pristine and damaged structures is very costly. To address this, a ML-based SHM approach is proposed that identifies structural damage using only physics-based synthetic strain data generated from the structure’s numerical finite element model. It employs a semi-supervised anomaly detection approach, trained solely on synthetic pristine data, to identify deviations in experimental data indicating damage. The method is validated on an aircraft spoiler demonstrator made of a composite sandwich panel, instrumented with a strain gauge grid on its surface layer. The results show that the proposed SHM approach accurately classifies damaged and undamaged experimental data, independent of the prevailing load case, solely based on synthetic pristine strain data. It is also able to localize these damages in the form of a confidence area with respect to the sensor grid. This demonstrates the feasibility of using only synthetic pristine data for data-driven SHM of composite aerospace structures.