Back to search

Aristotle University Of Thessaloniki (AUTH)

Non-supervised failure prediction in dynamic environments

Abstract

dc:description

In the era of digital manufacturing and the Internet of Things, predictive maintenance promises increased efficiency, reduced costs, and improved safety by predicting machine failures and guiding experts toward appropriate actions. However, in practice, industrial environments often consist of many complex machines monitored by sensors, each with unique characteristics and operating in a dynamic environment. A dynamic environment refers to changes in the internal or external conditions of a machine that lead to shifts in its behavior. Performing failure prediction and providing actionable guidance in such settings, using available sensory data and in the absence of prior knowledge, requires adaptive solutions that can continuously monitor machine states, issue early warnings, and ideally provide explanations to help experts understand the root causes of failures. Therefore, the three main goals of this thesis are: (i) to develop frameworks for early defect detection, (ii) to design online adaptive solutions that can automatically adjust to changing data conditions, and (iii) to enable context-awareness, with the aim of developing robust solutions for dynamic environments that also offer explainability. This thesis begins by presenting real-world case studies that motivated our research, followed by a discussion of the foundational concepts, definitions, and background related to failure prediction through early defect detection. We then present our contributions with respect to the three axes mentioned above. First, we propose adaptive, auto-configurable solutions for thresholding anomaly scores, distance-based anomaly detection, and online clustering of time series. In particular, for thresholding anomaly scores, we introduce a self-tuning approach and extend two existing dynamic thresholding techniques. For distance-based anomaly detection, we propose a method that is automatically configured and adaptable to streaming data. Regarding online clustering, we present BURST, a framework that renders partition-based clustering methods suitable for evolving time series. Second, we investigate the effects of dynamic environments in two real-world case studies, a cold-forming press and a vehicle fleet, and engineer appropriate solutions for early defect detection. This is preceded by a discussion on suitable evaluation metrics tailored to the early defect detection task. Finally, we introduce context-aware solutions based on two key concepts: wisdom of the crowd and fusion of heterogeneous data sources. In the first case, we propose an ensemble method that combines unsupervised distance-based and conformal anomaly detection techniques, along with the application of a deep learning approach over clustered data, leading to reduced maintenance management costs of a bus fleet. In the second case, we propose TECAD, a framework that enables contextualized decision-making and explainability by modeling context through a data matrix and a causal relationship graph.

Degree

thesis:*
Grantor dc:publisher
Aristotle University Of Thessaloniki (AUTH)
Year dc:date
2025

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Giannoulidis, Apostolos
  • Γιαννουλίδης, Απόστολος

Subjects

dc:subject × 14

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/59861

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Giannoulidis, Apostolos; Γιαννουλίδης, Απόστολος. Non-supervised failure prediction in dynamic environments. Aristotle University Of Thessaloniki (AUTH), 2025. http://hdl.handle.net/10442/hedi/59861