The Woodruff School

SUMMARY

Machine learning (ML) is an incredibly powerful tool for learning representations from data and, given the ever-growing amount of data made available by the Industrial Internet of Things, machine learning is of increasing interest in manufacturing. Despite its already proven success, deploying an ML model in a new application requires overcoming several barriers associated with resource allocation, data acquisition, and access to ML expertise. To lower these barriers to entry, there has been significant work in the machine learning field, including scalable computing platforms, open-source deep-learning software frameworks, and publicly available fundamental models. While these advances can also be used by ML practitioners in industrial or manufacturing areas, this field has several unique challenges that have limited widespread adoption of ML. These challenges include feature selection from unique data sources, model selection within a new domain, and resource constrained model deployment. This dissertation advances the state-of-the-art of democratized machine learning in industrial applications through three studies. First, to make full use of the advances from image classification, advanced convolution neural network architectures, developed for image classification, were adapted for sensor data to estimate remaining useful life of turbofan engines. In the next study, a methodology to use Bayesian optimization and image similarity was introduced for selecting the optimal time-frequency representation of sensor data without previous domain knowledge. This methodology was tested on two datasets whose objectives were to classify bearing health and classify the standoff distance in wire arc additive manufacturing. Finally, an architecture was proposed and tested to use anomaly detection as an input filter on edge devices to determine when data should be passed to cloud deployed supervised models. Together, the technical work in this dissertation develops an understanding of how to integrate ML with industrial applications, and this knowledge can be leveraged by manufacturers of all sizes to improve manufacturing processes and produced goods.

SUBJECT:	Ph.D. Dissertation Defense

BY:	Nathaniel DeVol

TIME:	Friday, April 19, 2024, 10:30 a.m.

PLACE:	GTMI (fomerly known as MaRC) Building, 114

TITLE:	Toward the Democratization of Industrial Machine Learning

COMMITTEE:	Dr. Christopher Saldana, Co-Chair (ME) Dr. Katherine Fu, Co-Chair (ME) Dr. Thomas Kurfess (ME) Dr. Andrew Henderson (Hendtech) Dr. Horacio Ahuett-Garza (ME - Monterrey)

SUMMARY Machine learning (ML) is an incredibly powerful tool for learning representations from data and, given the ever-growing amount of data made available by the Industrial Internet of Things, machine learning is of increasing interest in manufacturing. Despite its already proven success, deploying an ML model in a new application requires overcoming several barriers associated with resource allocation, data acquisition, and access to ML expertise. To lower these barriers to entry, there has been significant work in the machine learning field, including scalable computing platforms, open-source deep-learning software frameworks, and publicly available fundamental models. While these advances can also be used by ML practitioners in industrial or manufacturing areas, this field has several unique challenges that have limited widespread adoption of ML. These challenges include feature selection from unique data sources, model selection within a new domain, and resource constrained model deployment. This dissertation advances the state-of-the-art of democratized machine learning in industrial applications through three studies. First, to make full use of the advances from image classification, advanced convolution neural network architectures, developed for image classification, were adapted for sensor data to estimate remaining useful life of turbofan engines. In the next study, a methodology to use Bayesian optimization and image similarity was introduced for selecting the optimal time-frequency representation of sensor data without previous domain knowledge. This methodology was tested on two datasets whose objectives were to classify bearing health and classify the standoff distance in wire arc additive manufacturing. Finally, an architecture was proposed and tested to use anomaly detection as an input filter on edge devices to determine when data should be passed to cloud deployed supervised models. Together, the technical work in this dissertation develops an understanding of how to integrate ML with industrial applications, and this knowledge can be leveraged by manufacturers of all sizes to improve manufacturing processes and produced goods.