"Pragmatic Approach Towards Predictive Maintenance Using Fault Detectio" by Joseph Mark H. Jimenez

Pragmatic Approach Towards Predictive Maintenance Using Fault Detection and Multi-Fault Classification Models Evaluated By Point-Based and Range-Based Metrics

Date of Award

12-1-2023

Document Type

Thesis

Degree Name

Master of Science in Electronics Engineering

First Advisor

Carlos M. Oppus

Abstract

Unplanned machine downtime due to machine breakdown tremendously lessens productivity of manufacturing industries. In response, deploying predictive maintenance system is seen as a solution. This research introduces major gaps found from the current literature, i.e., limitations of supervised learning methods for machine fault diagnosis and of current metrics for evaluation of machine learning models in the context of predictive maintenance. The nature of machine fault can be understood as a series (range) of instances in time-series data. However, point-based (instance-based) metrics is used as the conventional metric based on the literature found. In this research, range-based metrics is explored and compared against point-based (instance-based) metrics to determine which one is more suitable in the predictive maintenance domain for estimating model performance in terms of precision, recall, and Fβ-score and ranking these models accordingly.

A controlled testbed of an electric fan motor with induced machine faults is demonstrated here as a use case of predictive maintenance. Accelerometer signals and infrared thermal images are collected to monitor and describe condition of the electric fan. Fault diagnosis is performed through fault detection and classification. Various experiments are designed and simulated using data collected while machine is in normal and faulty condition. Scenario analysis is performed to compare and determine how the point-based and range-based metrics evaluate the machine learning models in each of the simulated experiments. In this thesis, fault detection is realized by semi-supervised learning methods. This offers a more pragmatic first-step solution that manufacturing industries can adopt due to its less time-consuming data collection in contrast to the commonly used supervised classification approach. On the other hand, supervised learning is used for fault classification. Semi-supervised models used include isolation forest, local outlier factor, and one-class support vector machine while supervised models include support vector machine, random forest, and k-nearest neighbor. Lastly, a qualitative cost-benefit analysis is conducted to provide insights into how predictive maintenance solutions can be adopted by manufacturing industries in a time and cost-efficient manner.

Share

COinS