Data-driven probabilistic quantification and assessment of the prediction error model in damage detection applications

Abstract

Advances in computer hardware and sensor technologies have led to a surge in the use of data-driven modeling and machine learning for structural engineering applications, with Structural Health Monitoring (SHM) being one of them. Despite considerable interest, it remains a research topic due to the difficulty in accurately quantifying aleatoric and epistemic uncertainty in SHM systems. Sources of uncertainty are related to operational and environmental variability, as well as measurement noise and the model prediction error associated with the data used to train damage identification algorithms. In this work, the authors aim to explicitly quantify the statistical structure of model prediction error and assess its influence on the detection performance of strain-based SHM architectures under the existence of aleatoric variability. A structural beam, subjected to probabilistic static loading is used as the reference structure and strain measurements as the damage-sensitive features. Model prediction error is quantified explicitly using robust statistical tools through available laboratory observations and synthetic (Finite Element) data. Monte Carlo simulations enabled the forward propagation of uncertainty to the feature space to generate training data for three binary detectors (Likelihood Ratio Test, Quadratic Discriminant Analysis and Mahalanobis Distance), based on statistical pattern recognition. Detection performance was compared between the explicitly quantified prediction model error and the commonly assumed white Gaussian noise model, showcasing the influence of systematic error (bias) and correlation on the robustness of an SHM system using real-world data.