Back to search

Democritus University of Thrace (DUTH)

Hybrid machine learning and deep learning methodologies for resilient, sustainable and interactive cyberphysical manufacturing systems: joint production, maintenance and quality control optimization

Abstract

dc:description

Advances in Artificial Intelligence led to the introduction of novel paradigms, e.g., Industry 4.0 and smart manufacturing, that endeavor to digitize processes and operations within production systems, integrated in manufacturing plants. This digital transformation is achieved through the integration of intelligent technologies. Such technologies are the cyberphysical systems, which attempt to create a “bridge” between the computing process and the physical manufacturing environments through their integrated sensor devices. In this regard, they interchange information with the production environment and collect feedback from it. To adapt themselves and carry out decision-making, the cyberphysical systems process that feedback using artificial intelligence methodologies, including machine learning and deep learning. As a result, useful knowledge regarding manufacturing components and aspects, such equipment condition and inventory level, could be extracted in an effort to aid decision-makers in devising effective schedules and plans for manufacturing activities. Toward this end, the present dissertation studies the joint control and planning problem within dynamic production systems that implement cyberphysical systems. The aim is the realization of sustainable, resilient, and interactive cyberphysical manufacturing systems. In this respect, the dissertation proposes a variety of hybrid machine learning and deep learning methodologies. These methodologies are incorporated into decision-making frameworks, which assume the behavior and functionality of cyberphysical systems. Specifically, the frameworks strive to derive optimal, or near optimal joint control policies for production, maintenance, and quality control activities in the context of stochastic and deteriorating production environments, involving single, or multiple processing machines and storage facilities. The behavior of these environments is specifically defined by a number of random events. Among these events, failures deteriorate the condition of the considered systems. The quality of the produced parts is affected as well by the deteriorated system condition. For the optimization of the systems mentioned above, the proposed frameworks are applied and summarized as follows. The first one follows reinforcement learning decision-making for the joint optimization of manufacturing, maintenance, and recycling operations in a single-stage production/inventory system, which produces and stockpiles a single-type of products. Afterwards, the behavior and functionality of this framework is expanded for the joint control of alternative production environments. In this respect, hybrid reinforcement learning methodologies are introduced to optimize circular manufacturing systems. In contrast to other production systems, the circular ones attempt to produce new products by reusing low-quality, or returned ones in an effort to reduce the generation of waste. Given that, the proposed methodologies involve either one, or two reinforcement learning agents to formulate strategies for the processing machines included in the studied systems. Their decision-making integrates well-known ad-hoc control policies. The incorporated policies include pull production policies, e.g., Kanban, and condition-based maintenance. The aim of this implementation is the minimization of redundant costs related with unsold items and unnecessary authorization of activities. Due to its promising performance, the two-agent hybrid reinforcement learning approach is also applied for addressing sustainable manufacturing. Within this context, green and sustainable practices are considered for reforming low-quality material. Lastly, the dissertation proposes a deep learning methodology, which involves a predictive model. The considered model strives to devise activity authorizations by identify fluctuating aspects of manufacturing systems, e.g., inventory level. The evaluation of the proposed hybrid machine learning and deep learning methodologies is performed through a series of experiments simulating the behavior of the studied manufacturing systems. Findings of this evaluation validate the performance and efficiency of introduced methodologies in terms of cost-efficiency, sustainability, material management, and system resilience.

Degree

thesis:*
Grantor dc:publisher
Democritus University of Thrace (DUTH)
Year dc:date
2023

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Paraschos, Panagiotis
  • Παράσχος, Παναγιώτης

Subjects

dc:subject × 12

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/55163

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

Paraschos, Panagiotis; Παράσχος, Παναγιώτης. Hybrid machine learning and deep learning methodologies for resilient, sustainable and interactive cyberphysical manufacturing systems: joint production, maintenance and quality control optimization. Democritus University of Thrace (DUTH), 2023. http://hdl.handle.net/10442/hedi/55163