Machine learning metamodel and multi-fidelity neural network for guided wave-based structural health monitoring and prognostics

Chandran, Balaji

Machine learning metamodel and multi-fidelity neural network for guided wave-based structural health monitoring and prognostics

Publication Date

July 31, 2026

Creator

Chandran, Balaji

University of Nottingham

Abstract

This thesis presents a machine learning based framework for ultrasonic structural health monitoring (SHM) aimed at damage prediction, localization, and remaining useful life (RUL) estimation in engineering structures. Guided ultrasonic waves, particularly Lamb waves, are well suited for SHM due to their long range propagation and sensitivity to small defects. However, the interpretation of guided wave responses across varying structural configurations, materials, and damage scenarios remains challenging, motivating the use of data-driven approaches. In the initial phase of this study, a metamodel was developed comprising deterministic machine learning models including Polynomial Regression (PR), Artificial Neural Networks (ANN), and Support Vector Machines (SVM), alongside probabilistic models like Bayesian Regression (BR), and Stochastic Gradient Descent (SGD). The metamodel is designed to benchmark and select the most suitable learning algorithm based on dataset-specific performance metrics. Training data are derived from guided-wave experiments and corresponding finite element simulations of aluminium plates containing circular and slotted hole damage of varying sizes. Lamb waves are excited using tone-burst signals in the frequency range 300–500 kHz, and responses are recorded using a four-sensor network. Damage-sensitive features, namely time-of-flight and peak amplitude, are extracted and combined to form a damage index (DI), which serves as the primary learning target. Results show that ANN performs best when trained on limited experimental data, while SVM achieves superior performance when trained on fused experimental–simulation datasets. Although the metamodel provides valuable insight, directly combining low-fidelity simulation data with high-fidelity experimental data limits robustness. To address this, the second phase of this research introduces a Multi-Fidelity Neural Network (MFNN) framework that hierarchically integrates numerical simulations and experimental measurements. Finite element models are first validated against experimental waveforms using time alignment and amplitude scaling to ensure consistency. Within the MFNN architecture, simulation data provide broad coverage of the design space, while experimental data are used to learn a corrective residual that aligns predictions with physical reality. The MFNN demonstrates improved accuracy and robustness compared with the metamodel, particularly under limited experimental data conditions. Beyond damage severity estimation, the MFNN framework was extended to support damage localization through DI–based projection and remaining useful life estimation by coupling damage index trajectories with physics-informed fatigue models. The resulting prognostic trends remained physically consistent, demonstrating the suitability of the proposed framework for integrated detection, localization, and RUL estimation within ultrasonic SHM. Overall, this thesis contributes a novel metamodel and multi-fidelity neural network based framework for ultrasonic SHM, offering scalable, accurate and efficient solutions for structural diagnostics and prognostics.

Item Type

ethesis

Thesis Type

PhD

Supervisors

Pandey, Madhup

University of Nottingham, UK

Keywords

Machine Learning

Structural Health Mon...

SHM

Multi-fidelity Neural...