Modular Data-Driven Framework for Sensor Selection in Complex Systems
Restricted (Penn State Only)
- Author:
- Kulkarni, Amol
- Graduate Program:
- Industrial Engineering (PHD)
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- August 01, 2023
- Committee Members:
- Saurabh Basu, Major Field Member
Hui Yang, Major Field Member
Vittaldas Prabhu, Chair & Dissertation Advisor
Janis Terpenny, Special Member
Chris Rahn, Outside Unit & Field Member
Steven Landry, Program Head/Chair - Keywords:
- Sensor Selection
NLP
Fault Identification
Clustering
Unsupervised Machine Learning
Topic Modeling
Fuzzy Clustering
Ordered Clustering - Abstract:
- The current industrial revolution has enabled the digital transformation of manufacturing industries, making them look vastly different from the industries ten years ago. The advancement in sensor technology, data analytics, and machine learning have made it possible for manufacturers to choose and implement solutions that would enable increasing the utilization rate of equipment on the shop floor. A successful digital transformation yields higher returns, and industries across a wide range of sectors have seen anywhere from 30 to 50 percent reduction in machine downtime, according to a report titled “The Costs and Benefits of Advanced Maintenance in Manufacturing” by NIST. Despite the benefits of smart manufacturing/Industry 4.0, many industries have yet to fully embrace digitization, mainly due to the following challenges: • There are issues with scalability, primarily due to the high costs and how digital transformation typically does not provide short-term benefits to the organization. • Inadequate knowledge of digital capabilities prevents industries from enacting properly scaled transformative efforts. • Having a wide variety of equipment in the factory can require a variety of sensor types and configurations, which can cause confusion. Implementing the framework proposed in this dissertation will help address these challenges, especially the ones tied to scalability and dealing with different options in sensor technologies. After introducing the problem, motivation, and research questions in the first chapter, the second chapter demonstrates the effectiveness of having a human-in-the-loop approach to fault mode acquisition, thus standardizing maintenance vocabulary across industries and making it easier to identify required sensors. The third chapter focuses on sensor selection in a case when maintenance records are not available. The proposed OFFCaTS model takes advantage of Failure Mode, Mechanisms and Effects Analysis (FMMEA) to identify the types of sensors needed and their specification to monitor the health of a complex system. The effectiveness of this approach is demonstrated using a wind turbine gearbox as an example. The fourth chapter builds on the sensor selection method proposed in the third chapter, to enable the identification and selection of heterogeneous sensors, one of the limitations of the model proposed in chapter 3. A goal programming model is developed for sensor specification selection, while also establishing a pipeline to collect specification data on different type of sensors from multiple vendor website. Future directions of this work include developing a predictive maintenance model that works in tandem with the active learning model.