Advanced POD-Based Evaluation of Load-Dependent Performance in Machine Learning Damage Diagnosis

The reliable diagnosis of damage in carbon fiber reinforced polymer (CFRP) structures under variable loading conditions remains a key challenge for ensuring structural integrity and extending service life. Acoustic Emission (AE)-based machine learning (ML) methods have shown great potential, their performance often depends on loading conditions, which are rarely quantified systematically. This study aims to quantitatively evaluate how loading conditions affect the performance of ML-based damage diagnosis models. An advanced probability of detection (POD) framework is developed mapping the diagnosis performance onto newly-introduced two-dimensional loaddependent POD surfaces, incorporating both load amplitude and cyclic frequency as process parameters as well as the likelihood of correct diagnosis statements as output. Acoustic Emission signals generated during controlled cyclic loading of CFRP plates were measured, preprocessed, and used to train and test supervised ML models for damage detection and classification. The resulting detection probabilities were computed for varying load configurations, enabling a detailed visualization and analysis of load sensitivity. The results reveal distinct regions of reduced detection reliability at specific combinations of load magnitude and frequency, highlighting the necessity of considering operational loading when validating ML-based diagnostic systems. The proposed two-dimensional POD approach represents a new and systematic quantification of load-dependent performance in AE-driven ML damage diagnosis. This approach enables performance assessment under varying loading conditions and stimulates future research on load-adaptive diagnostic models for complex composite structures.

Keywords: Acoustic Emission, Machine Learning, Probability of Detection, Damage Diagnosis, Reliability.