Data-Driven and Machine Learning Frameworks for Condition Assessment of Plate Structure using Elastic Waves - Short Abstract of Thesis

Sep 17, 2026

Uploaded by Levi

The thesis presents data-driven and machine learning frameworks for structural health monitoring to enhance condition assessment and extend the service life of engineering plate structures. It addresses defect-driven degradation of piezoelectric sensor performance by detecting and localizing anomalies in PZT substrates using a novel multiresolution dynamic mode decomposition approach with ultrasonic wave interactions. It further develops deep autoencoder-based anomaly detection, Bayesian filtering with an unscented Kalman filter for stiffness estimation, probabilistic multi-output Gaussian process regression for anisotropic composite elastic constants, and a multi-fidelity physics-informed neural network with transfer learning for acoustic emission source localization under measurement uncertainties.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages2

Reading Time5 min

AccessFree to download

Structural Health Monitoring Motivation

Piezoelectric Anomaly Detection (PZT)

Deep Autoencoder for Sensor Waveforms

Bayesian Stiffness Estimation (Unscented Kalman Filter)

Probabilistic ML for Composite Elastic Constants (moGPR)

Multi-Fidelity Physics-Informed Localization (mfPINN)

INDIAN INSTITUTE OF TECHNOLOGY GUWAHATI SHORT ABSTRACT OF THESIS Name of the Student Roll Number Programme of Study : Nur Mahammad Mussa Kalimullah : 186104113 : Ph. D. SHORT ABSTRACT The evolving domain of structural health monitoring (SHM) is crucial for ensuring the integrity and extending the service life of engineering structures. This thesis presents a suite of data-driven and machine learning frameworks developed to enhance the condition assessment of plate structures, particularly focusing on the complexities of piezoelectric materials and anisotropic composites. In a comprehensive exploration of data-driven and machine learning frameworks for assessing the condition of plate structures, this dissertation presents a series of interconnected studies, each contributing to the advanced insights and application in SHM, particularly focusing on piezoelectric materials and anisotropic composites. The effectiveness of piezoelectric sensors is deteriorated by the presence of defects, delamination, and corrosion, which must be identified and addressed to ensure the successful application of the SHM framework The research begins by addressing anomaly detection in piezoelectric Lead Zirconate Titanate (PZT) substrates through the development of a novel multiresolution dynamic mode decomposition (mrDMD) algorithm. This method, effective at diagnosing and localizing surface defects, utilizes the interaction of ultrasonic waves with microscale defects. It leverages sophisticated image registration techniques and Kullback Leibler divergence to enhance the precision and effectiveness of defect localization in complex material systems. Thesis Title: Data-Driven and Machine Learning Frameworks for Condition Assessment of Plate Structure using Elastic Waves Name of Thesis Supervisor(s) : Dr. Amit Shelke Thesis Submitted to the Department/ Center : Civil Engineering Date of completion of Thesis Viva-Voce Exam : 13/08/2024 Key words for description of Thesis Work : Structural Health Monitoring, Elastic Waves, Scientific Machine Learning Building on these foundational insights, the work progresses to the application of a deep autoencoder neural network, specifically tailored for anomaly detection in PZT sensors. This approach is beneficial at processing complex waveform data, identifies surface defects and delamination with high accuracy. The integration of image registration and peak signal-to-noise ratio metrics further refines the anomaly quantification, demonstrating a robust solution for handling the nonlinear interaction of waves with surface anomalies. The development of a probabilistic approach using the unscented Kalman filter, aimed at estimating the stiffness matrix of an anisotropic piezoelectric material, in particular, Lithium Niobate substrate. This segment of research presents a ground-breaking integration of experimental measurements with a physics-based mathematical model. Abstract-TH-3449_ 186104113 The implementation demonstrates the utility of Bayesian filtering in material characterization, especially under measurement uncertainties. In an extension of these methodologies, the exploration turns to an anisotropic carbon fibre composites plate structure. Here, a probabilistic ML based on multi-output Gaussian process regression (moGPR) is employed to estimate the elastic constants of woven fabric reinforced composites. This study marks a notable advancement in interconnecting numerical simulations with experimental observations, specifically through the analysis of Lamb wave dispersion curves. The ML model correlates between theoretical predictions and practical measurements in composite laminates. Being a probabilistic framework, the implementation considers the measurement uncertainties. The culmination of these methodologies is demonstrated in the final study, which introduces a multi-fidelity physicsinformed neural network (mfPINN) for acoustic emission source localization in a

FAQs

What is the central goal of the thesis?

To develop efficient, accurate, and robust data-driven and machine learning frameworks for condition assessment of plate structures in structural health monitoring.

How does the thesis detect defects in piezoelectric PZT substrates?

It introduces an anomaly detection method using a multiresolution dynamic mode decomposition (mrDMD) algorithm that uses ultrasonic waves interacting with microscale defects, together with image registration and Kullback-Leibler divergence for localization.

What probabilistic and physics-informed methods are used for composite materials and localization?

For anisotropic carbon-fiber composites, it employs multi-output Gaussian process regression to estimate elastic constants from Lamb wave dispersion curves. For final acoustic emission source localization, it uses a multi-fidelity physics-informed neural network with transfer learning to fuse low-fidelity physical models and sparse high-fidelity experimental data while accounting for sensor uncertainties.

Sep 17, 2026

Uploaded by Levi