10/10/2023
By Danielle Fretwell

The Francis college of Engineering, Department of Mechanical and Industrial Engineering, invites you to attend a Doctoral dissertation proposal 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: Tuesday, Oct. 24, 2023
Time: 10 a.m. to noon
Location: Ball 326

Those interested in attending virtually should contact the student (nitin_kulkarni@student.uml.edu) and committee advisor (Alessandro_sabato@uml.edu) at least 24 hours prior to the defense to request access to the meeting.

Advisor: Alessandro Sabato, Committee chair, Asst. Professor, Department of Mechanical and Industrial Engineering, UMass Lowell

Committee Members

  • Hunter Mack, Assoc. Professor, Department of Mechanical and Industrial Engineering, UMass Lowell
  • Rozhin Hajian, Asst. Professor, Department of Mechanical and Industrial Engineering, UMass Lowell
  • Yashar Eftekhar Azam, Asst. Professor, Civil and Environmental Engineering, University of New Hampshire

Brief Abstract:

Recent advancements in machine learning algorithms, such as autoencoders (AE), have generated innovative paradigms for reconstructing full field data from sparse measurements and performing structural health monitoring (SHM). However, traditional AE methods suffer from inconsistent boundary condition constrains, which result in limited applicability for full field expansion. Therefore, knowledge gaps exist in the effective expansion of full field data that also incorporate denoising and damage detection capabilities. The proposed study aims to address these gaps and (1) create a framework capable of learning the spatio-temporal characteristics (STCs) of a targeted dynamic system and adapt to changes representative of damages, (2) develop a framework for expanding sparse sensor measurements to full field data using AE efficiently, and (3) enhance the accuracy of the AE-based framework using physics-based regularizers. The study will also address the use of reduced-order models for damage detection of large-scale structures.

This research develops a novel neural network framework referred to as Time Inferred Autoencoder (TIA). By using a novel concept of loss function (i.e., maximum correntropy loss function) to remove artifacts in the full field data created during the expansion, the TIA learns and reproduces the dynamics of a system dynamics (i.e., STCs) while simultaneously denoising the expanded data. Experiments have been conducted with different noise scenarios, demonstrating that TIA yields superior results compared to traditionally available denoising and expansion techniques. Furthermore, to address the limitation of currently used AEs and improve the accuracy of the expansion process even more, a physics-based regularizer is incorporated into the TIA framework to develop a new concept of physics-informed neural network. However, comprehensive testing is still necessary to conclusively validate its effectiveness and complete the creation of a tool that can be used for SHM and prediction of dynamically changing systems.

All interested students and faculty members are invited to attend the online defense via remote access.