Back to search

Clemson University

Generating Training Data for Machine Learning Applications in Manufacturing Through Purposeful Failure

Abstract

dc:description.abstract

Manufacturing unexpected downtime contributes a significant cost drain to manufacturers. To reduce unexpected downtime, manufacturers invest in diagnostic and prognostic systems for determining equipment health. These systems help maintenance staff by providing predictions of equipment degradation over time. Based on these predictions, maintenance staff schedule repairs to prevent equipment failure from occurring during production. However, challenges still exist preventing the widespread implementation of these systems in the production environment. The amount and quality of data for training machine learning algorithms to predict these failures are two of these challenges. Existing methods of run-to-failure testing take too long to acquire the data necessary to represent manufacturing component failure. While methods exist to implement diagnostic and prognostic systems in manufacturing, there is no standard or methodology for generating data to train these systems. To address this challenge, a method, termed the Purposeful Failure Methodology (PFailM), is proposed to generate failure data in manufacturing equipment. A literature review was conducted around manufacturing equipment health to help formulate this methodology. Using previous papers, specific elements were identified from condition-based maintenance, prognostics and health management, failure methodologies, and design methodologies to create PFailM. These elements were formulated into a generalized methodology to allow for more widespread application to different manufacturing equipment types. Two different test cases are used to validate and test the PFailM approach. The first case was bearings. Two different damage protocols were created for bearings using PFailM. The first damage protocol involved the generation of damage to simulate spalling on the bearing races. The second damage protocol was to simulate contamination via lubrication injection on the bearing races and rolling elements. The second test case used a UR10 robot. The damage protocol simulated the overloading of the robot end-effector. Data were generated from the bearing spalling damage protocol and the overloading damage protocol for the robot. Data observations and methodology refinement were conducted to continue the standardization of steps for the methodology.

Degree

thesis:*
Name thesis:degree_name
Master of Science (MS)
Level thesis:degree_level
Thesis
Discipline thesis:degree_discipline
Mechanical Engineering

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Wescoat, Ethan
Contributors dc:contributor
  • Laine Mears
  • Cameron Turner
  • John Wagner

Identifiers

dc:identifier.*
Repository record dc:identifier
https://open.clemson.edu/all_theses/3547

Chain of custody

source
Harvested from
Clemson University
Base URL
open.clemson.edu/do/oai/
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
related terms
citation

Wescoat, Ethan. Generating Training Data for Machine Learning Applications in Manufacturing Through Purposeful Failure. Thesis thesis, https://open.clemson.edu/all_theses/3547