07/27/2021
By Sokny Long
The Francis College of Engineering, Department of Mechanical Engineering, invites you to attend a Master’s thesis defense by Mark Todisco on “Structural Damage Classification of High-Rate Dynamic Systems via Hybrid Deep Learning.”
MSE Candidate: Mark Todisco
Defense Date: Tuesday, Aug. 3, 2021
Time: 10 to 11 a.m. EST
Location: This will be a virtual defense via Zoom. Those interested in attending should contact mark_todisco@student.uml.edu and committee advisor, zhu_mao@uml.edu, at least 24 hours prior to the defense to request access to the meeting.
Committee Chair (Advisor): Zhu Mao, Ph.D., Assistant Professor, Mechanical Engineering, University of Massachusetts Lowell
Committee Members:
- Peter Avitabile, Ph.D., Professor Emeritus, Mechanical Engineering, University of Massachusetts Lowell
- Jacob Dodson, Ph.D., Research Mechanical Engineer, Air Force Research Laboratory
- Adriane Moura, Ph.D., Staff Mechanical Engineer, Applied Research Associates
Brief Abstract:
The ability to classify the damage of a structure induced by high-rate high-acceleration (HRHA) impacts is critical for the mitigation of catastrophic system failure. These dynamic events are often unclear with nonstationary and broadband spectral energy. In addition to these challenges, high-acceleration events often destroy the structure, preventing repeatable experiments. As a side effect, HRHA data is also especially limited and sparse.
While the lack of high-fidelity physics in high-rate dynamic events has posed as a challenge for many years, Machine Learning (ML) is making significant breakthroughs in many structural health monitoring (SHM) applications. Some ML algorithms are able to extract features containing a component’s damage information, thereby reflecting the system’s overall health status. More specifically, data-driven techniques such as deep learning can overcome unclear system dynamics by learning a latent representation of the data.
This thesis work presents a hybrid damage classification method that fuses deep/shallow and unsupervised/supervised ML algorithms to classify six damage levels of a highly instrumented electronic assembly subjected to HRHA drop tower impacts. A deep unsupervised Convolutional Variational Autoencoder (CVAE) first extracts latent features containing the structure’s damage information. A shallow supervised support vector machine (SVM) classifier then uses this extracted damage information to predict the structure’s damage level. Wavelet transformations are also used to maintain adequate time-frequency information of the shock response and augment the presentation of its high-rate dynamics. Given only six training examples containing sparse data, the hybrid damage classification technique in this thesis classifies the severity of the structure’s damage. This research could be extended to remaining useful life (RUL) prediction in the future.
To address the severe lack of data intrinsic to HRHA impacts, an alternative approach is also investigated, where a finite element (FE) model simulates the dynamic response of the structure, thus generating a substantially larger training dataset. First, the CVAE is trained using the data generated by the simulated dynamic response of the FE model. Portions of the model’s parameters are then frozen, while the remaining parameters are fine-tuned with the experimental data. After initial research, it was determined that the quality of the FE model is essential in future research to obtain a better damage classification performance.
All those interested may attend remotely.