07/09/2024
By Danielle Fretwell

The Francis College of Engineering, Department of Mechanical Engineering, invites you to attend a Doctoral Dissertation defense by Nitin Nagesh Kulkarni on: Full Field Expansion from Sparse Data Using Autoencoders and Physics-Informed Neural Networks

Candidate Name: Nitin Nagesh Kulkarni
Degree: Doctoral
Defense Date: Wednesday, 24 July, 2024
Time: 2 to 4 p.m.
Location: Perry-115

Committee:
Advisor: Dr. Alessandro Sabato, Assistant Professor, Department of Mechanical and Industrial Engineering, UML

Committee Members*
1. John Hunter Mack, Ph.D., Associate Professor, Department of Mechanical and Industrial Engineering, UML
2. Rozhin Hajian, Ph.D., Assistant Professor, Department of Mechanical and Industrial Engineering, UML
3. Yashar Eftekhar Azam, Ph.D., Assistant Professor, Department of Civil and Environmental Engineering, UNH


Brief Abstract:
Advancements in data processing have significantly enhanced the fields of structural health monitoring (SHM) and nondestructive evaluation (NDE). Contact-based methods, which rely on sparsely distributed sensors like accelerometers and strain gauges, face challenges in accurately detecting and localizing damage, especially when a limited number of sensors is available. Non-contact sensing techniques, such as computer vision, offer a promising alternative, particularly when integrated with deep learning (DL) models. While non-contact methods can automate damage detection tasks, they often suffer from lower signal-to-noise ratios (SNR) compared to contact-based methods. Therefore, this dissertation aims at: 1) enhancing the SNR of digital images for improved damage detection capabilities and 2) developing novel methodologies for full-field expansion of sparse or limited measurements.

To enhance the SNR of digital images, reduced-order models (ROMs) and DL-based methods are employed. ROM techniques like principal component analysis (PCA) are investigated for their potential to filter noise and extract critical structural features representative of damage. Moreover, a novel DL framework that combines long short-term memory (LSTM) autoencoders with a maximum correntropy loss function is created to preserve spatio-temporal characteristics of a system while denoising images, and address ROMs’ limitations for
non-linear systems.

For the full-field expansion of sparse measurements, autoencoders (AEs) and variational autoencoders (VAEs) are explored to overcome the constraints of existing interpolation methods. AEs are leveraged to capture the essential dynamic behavior of structures and expand sparse data to a full-field representation. However, as AEs struggle to reconstruct a system’s response when data from the system in a different configuration (e.g., damage) are used, a physics-informed variational autoencoder (PI-VAE) framework is developed. This hybrid approach enhances the network’s ability to adapt to changes in the system and improves the accuracy and robustness of the full-field expansion for damage detection.

This research demonstrates significant improvements in data quality and damage detection accuracy, paving the way for more effective SHM and NDE practices. The advanced techniques developed in this study can be applied to real-world structures to enable automated, efficient, and accurate assessments of their structural integrity even when a reduced number of contact-based sensors or lower resolution digital images are available.