10/22/2021
By Bhavana Maradani

The Kennedy College of Sciences, Department of Computer Science, invites you to attend a doctoral dissertation defense by Bhavana Maradani on “Data-driven Analytics for Health and Safety." 

Ph.D. Candidate: Bhavana Maradani
Date: Wednesday, Nov. 3, 2021
Time: 7 a.m.
Location: This will be a virtual defense via Zoom. Those interested in attending please contact Committee Advisor haim_levkowitz@uml.edu and bhavana_maradani@student.uml.edu at least 24 hours prior to the defense to request access to the meeting.

Committee Members:

  • Haim Levkowitz (advisor), Associate Professor, Computer Science Department, University of Massachusetts Lowell
  • Jonatha Mwaura, Assistant Teaching Professor, Computer Science Department
  • Sashikala, Professor, Department of Operations and IT, ICFAI Business School, Hyderabad

Abstract:

Machine learning is a very powerful technique to analyze rapid data gleaned from various sources. Combined with visual analytics, good exploratory data analysis paves the way in the right direction. It both shortens the process of machine learning and provides accurate outcomes. However, managing and understanding this vast body of data to tease actionable insights in the shortest possible time and with less human intervention is critical. Our dissertation aims at addressing the challenges of providing solutions with less expertise intervention and early detection of life-threatening instances.

In this dissertation, we focus on the following two problems: 1) Analyzing the performance of a subject while performing physical activity during Tele-Rehabilitation; 2) Estimating the health status and remaining useful “life” (RUL) of an aircraft engine.

To address the problem for efficient visual analytics to provide insights into large streams of data, we considered the use case of Tele-rehabilitation. The data collected from the body sensor network needs to be presented in an understandable format to aid the caregiver to understand the movement of the subject. Most of the current research has developed models to estimate joint angles. However, research on visualizing the movement data has been limited and little has been done concerning the long-term performance of a subject. To fill this gap we propose an interactive visualization to evaluate subjects’ upper limb movements using data acquired while performing various rehabilitation exercises over a period of time. The main contribution of this task presents interactive visualizations to demonstrate the amount of deviation from the normal range of motion and the time taken to finish a task.

For the second task, we have considered the problem of detecting the health status and estimating the remaining useful life of an aircraft engine (RUL). Considering the high cost associated with engine failure, monitoring the degradation trend of an engine from operational and historical maintenance data can reduce operational expenses and, more important, the likelihood of — possibly catastrophic — failure during use, thus, potentially saving lives. To address the issue of detecting the health status of an engine, we propose an ensemble model to determine the state of an engine. The main contribution of this work is the automatic detection of the optimal number of health states and the development of a classifier to perform the health diagnosis of an engine. Further, we have predicted the RUL by constructing the health degradation curve (HI). Due to the complexities of physics-based models involved in the precise modeling of the degradation curve, the construction of HI solely from the historical data is gaining massive interest which doesn’t require any human expertise to model the physics of the machine. We proposed an attention-based deep-learning approach integrated with a modified similarity-based model to predict RUL. The main goal of this work is to extract the personalized HI curve without making any assumptions about the degradation. We evaluated the prognostic performance of the proposed ensemble approach and the end-to-end independent model on a publicly available NASA aircraft engine dataset. The experimental results demonstrate the competitiveness of the proposed RUL prediction model.