A Machine Learning and Ai Approach to Fault Diagnosis in the Context of Wire Manufacturing in the Philippines

Date of Award

12-2021

Document Type

Thesis

Degree Name

Master of Science in Computer Science, Straight

First Advisor

Jose Alfredo A. de Vera III

Abstract

AI and Machine Learning have become ubiquitous in many industries. Automated condition monitoring is a popular application for AI, and has been widely studied in recent years. This study provides a case for the usage of a Variational Autoencoder based approach for automated condition monitoring based on vibration data from accelerometers. We show that when trained in a semi-supervised manner, even with limited data, the VAE latent space is able to distinguish between normal and idle machine activity. We also show that the VAE is able to learn to group data from similar sensors together in latent space, even without a sensor label provided during training. This study likewise considers the particular context of applying AI for real world applications, and the associated challenges which may arise from this. In the case of this study, we encounter some data integrity issues which are commonly associated with real world data and we also encounter some challenges unique to this study given the operation of the data collection set up. We demonstrate the applicability of Joint Recurrence Plots as a possible solution to the issues encountered by the operation of the data collection set up. Likewise, this study also proposes the use of Gaussian Mixture Model for the creation of theoretical labels which may not be available given the limitations of the provided ground truth.

Link to Full Text

Share

COinS