Author ORCID Identifier

https://orcid.org/0000-0002-5471-9040

Date Available

4-24-2025

Year of Publication

2023

Degree Name

Doctor of Philosophy (PhD)

Document Type

Doctoral Dissertation

College

Engineering

Department/School/Program

Electrical and Computer Engineering

First Advisor

Dr. Peng Wang

Second Advisor

Dr. Michael T. Johnson

Abstract

The future of smart manufacturing relies on predictive maintenance systems that intelligently minimize expensive downtime through timely assessment of machine condition. Deep Learning (DL) has achieved excellent performance in industrial condition monitoring experiments, but the constraints of the manufacturing environment prevent many algorithms from being practically deployed on the factory floor. Ubiquitous sensing from online machines generates high velocity data streams that require new techniques for efficient transmission and storage. Despite these ever-increasing data lakes, many applications still lack the data needed for training DL fault diagnosis and wear tracking models since most data is unlabeled and only from nominal operating conditions (i.e., it lacks the diversity needed to discriminate faults). In addition, even the best models are likely to encounter emerging faults that did not appear in the training data. Changes in machining process parameters further compound the deployment difficulty by limiting the reusability and generalization of the model. This research program develops a handbook for practical data-driven condition monitoring in view of the constraints of the factory floor. The proposed methods cover multiple aspects of training and deployment. Once sensor networks are installed, DL-based autoencoders can compress and reconstruct signals using novel Fault Division Autoencoder Multiplexing (FDAM) to reduce the volume of data sent to the cloud. Large amounts of streaming data may still fail to provide the necessary run-to-failure data sets for training degradation models. Thus, empirical models with Bayesian parameter estimation is introduced to estimate future degradation through a case study proposing a novel method for tracking degradation with nonmonotonic transient events using Compound Poisson Process. If manufacturers want to utilize DL to enhance condition monitoring, they should select application that highlight the strengths of DL. This study proposes deep normalizing flows as a general-purpose generative model for intelligent manufacturing, showing new applications of augmenting data sets with synthetic fault examples and performing fast amortized inference on empirical degradation models, while extending work on detecting anomalies when provided only normal operating data. To adapt to changing conditions and emerging faults, a new Self-Supervised Learning (SSL) approach for condition monitoring is adapted from Barlow Twins, and Mixed-Up Experience Replay is proposed as a novel method for for refining informative features learned from unlabeled data without catastrophically forgetting previously seen conditions. To facilitate reusable models, integrating SSL and Federated Learning for time-series condition monitoring can maximize the generalization of a static model trained on a fixed data set by learning a data-centric representation and efficiently sharing information. Overall, the proposed methods offer novel candidate solutions for a variety of data-related obstacles that may be encountered when building a condition monitoring pipeline. Future work should continue this focus on overcoming the practical barriers that prevent widespread adoption of data-driven methods in manufacturing environments.

Digital Object Identifier (DOI)

https://doi.org/10.13023/etd.2023.072

Funding Information

This work was supported by National Science Foundation Grant No. 2015889 from 2020-2023.

Available for download on Thursday, April 24, 2025

Share

COinS