Non-destructive evaluation (NDE) assumes a crucial role in examining infrastructure acrossdiverse industries that require ongoing monitoring to uphold product quality, sustain aging infrastructure, and ensure operational equipment safety. Recent advancements in data acquisition (sensing technologies) and data evaluation (data analytics) have prompted a significant shift in the NDE community towards fully automated, data-driven inspection routines. The primary goal of this dissertation is to develop intelligent, state-of-the-art NDE-based artificial intelligence (AI) methods. This research involves incorporating NDE theory and insights to adapt modern machine learning (ML) and AI algorithms, enabling efficient decision-making processes with high accuracy and reliability, while minimizing human intervention. Several cost-effective, miniaturized NDE sensing methodologies based on techniques like Magnetic Flux Leakage (MFL), Eddy Current (EC), Capacitive Imaging (CI), and laser profilometry are developed based on the materials being tested. The research focuses on the following key aims: (a) Developing spatially adaptive denoising algorithms to address uncertainties in the data collection process, enabling accurate defect detection and localization, (b) Dynamically updated defect tracking through mixture regression and optimally binned hypothesis testing, (c) Establishing an integrated diagnostic framework for inline inspection and automated defect characterization using a hybrid deep learning setup, (d) Efficient data augmentation and fusion techniques to combine information from multiple heterogeneous sensors, and improving defect diagnostics by twinning/registering experimental and simulated data through transfer learning principles, (e) Implementing sophisticated data compression techniques to achieve cost-effective and fast defect diagnostics without compromising efficiency.This dissertation develops new-age integrated NDE frameworks for data acquisition and evaluation,providing accurate and reliable defect diagnostics. NDE data used in practical applications often suffers from noise contamination and errors caused by various factors. In chapter 2, a Bayesian decision theory-based approach is established to understand the noise corruption in capacitive sensing data. A solely data driven spatially adaptive denoising algorithm is modelled that does not need any oracle information. In chapter 3, an integrated inline inspection (ILI) and automated defect detection framework for plastic pipelines using laser profilometry data and a hybrid deep learning approach are established. Chapter 4 proposes an under-sampling compressed scheme for fast and cost-effective scanning, providing accurate diagnosis. Spatially adaptive methods for identifying defect growth using dynamically updated transfer learning are introduced in subsequent chapters 5 and 6, leveraging information from less noisy baseline scan. Bivariate function estimation using mixture regression (TLMR) and transfer learning-based binned hypothesis testing (TLBH) are incorporated for tracking defect growth on later noisy scans. An inverse data-driven approach for estimating electrical parameters of the substrate is proposed in chapter 7, along with registrationaided machine learning models that adaptively utilize information from large synthetic datasets to improve predictions on scarce experimental data. Finally, an integrated multi-modal fusion setup is designed in chapter 8 for enhanced defect detection, incorporating optimal transport (OT) based registration of data from different sensing modalities.
Read
