Interpretable machine learning approaches for damage identification in drilling risers

Detecting damages, especially micro-damages, early and accurately is crucial to the safety of drilling risers. This study presents a high-precision model for detecting micro-damages in drilling risers based on modal information using Machine Learning (ML) algorithms. The under-sampling technique is applied to ensure a balanced distribution of data. The modal information data set is used to analyze three machine learning models: Random Forest (RF), Extreme Learning Machine (ELM), and Back Propagation Neural Network (BPNN). The study investigates their applicability and interpretability in damage identification. SHapley Additive exPlanations (SHAP) analysis is utilized to unveil the decision-making process behind the models. The findings indicate that each model offers unique advantages, with the RF and BPNN models achieving a superior performance in terms of generalization ability. The study finally conducts feature optimization on the BPNN model based on SHAP values and develops a BPNN-4 model with enhanced stability and generalization capabilities. This research highlights the potential of applying interpretable ML techniques in structural health monitoring, offering a promising direction for future offshore engineering diagnostics.