AI-driven production scheduling based on multi-agent system

• Κύριος συγγραφέας: Σιάτρας, Βασίλειος
• Άλλοι συγγραφείς: Siatras, Vasileios
• Γλώσσα: English
• Έκδοση: 2022
• Θέματα: Production scheduling; Multi agent system; Artificial intelligence; Metaheuristics; Machine learning; Xρονοπρογραμματισμό παραγωγής; Σύστημα πολλαπλών αυτόνομων μονάδων; Τεχνητή νοημοσύνη; Μετα-ευρετικες μέθοδοι; Μηχανική μάθηση

Mass production industries and global firms depend on production management in order to compete in the modern volatile manufacturing marketplace. Achieving high manufacturing performance requires a compound of AI services and information technologies working in parallel across different business layers and ensure information transparency and precise decision-making. As research and innovation took off, an ecosystem of smart autonomous assets (so called agents) has been introduced in all manufacturing levels constructing the main idea behind Industry 4.0. With Multi-Agent Systems (MAS) appearing in the forefront of nowadays manufacturing research and innovation, the requirements of interoperability and granularity is the primary challenge for I4.0 stakeholders. The realization of this objective is found in the core components of I4.0, so called Asset Administration Shell (AAS), which provide abstraction in the production assets description and interactions. As a result, managing to design and develop AI services with increased abstraction in the applied environment enables the deployment of increased automation and flexibility in the business operations. This thesis focuses on exploiting AAS technology for the design and development of a scheduling meta-agent that uses AI in order to solve different production schedule optimization problems. For that reason there was proposed a toolbox of three independent AI agents, that act as a plugin to the meta-agent interface and allow decision-making in three different scheduling problems namely factory, logistics, and conveyor scheduling. (a) The development of the factory scheduler plugin a heuristic decision-making algorithm for scheduling was utilized; (b) for Logistics scheduler was based on a mathematical optimization problem that uses Fuzzy Logic and is solved by a Genetic Algorithm; (c) while for the Conveyor scheduler there were used both a Mixed Integer Programming model and two Machine Learning model (LSTM RNN, and a FFNN). The MAS was deployed in two different manufacturing cases i.e. automotive and bicycle manufacturing industries demonstrating improvement in the business KPIs. In overall, the MAS was proven efficient in the ability of managing the production environment with low human interference required.