Development of an AI-driven agent for continuous monitoring, root cause analysis, and adaptive decision support in smart manufacturing

This thesis investigates the development of an autonomous agent designed to manage the endtoend machine learning pipeline for lead time prediction in supply chain settings. The aim is to reduce the reliance on manual scripting and expert input by enabling the agent to make informed decisions across...

Descripción completa

Detalles Bibliográficos
Autor: Bosch Matas, Adrià
Tipo de recurso: tesis de maestría
Fecha de publicación:2025
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/447141
Acceso en línea:https://hdl.handle.net/2117/447141
Access Level:acceso abierto
Palabra clave:Artificial intelligence
Machine learning
Intelligent agents (Computer software)
Intel·ligència artificial
Aprenentatge automàtic
Agents intel·ligents (Programari)
Àrees temàtiques de la UPC::Informàtica
Descripción
Sumario:This thesis investigates the development of an autonomous agent designed to manage the endtoend machine learning pipeline for lead time prediction in supply chain settings. The aim is to reduce the reliance on manual scripting and expert input by enabling the agent to make informed decisions across key stages such as data preparation, feature engineering, model training, and evaluation. The solution combines structured logic with multimodal reasoning, integrating both language and visionbased tools to handle a variety of data types and conditions. To evaluate the agent’s behavior, we have developed a benchmarking framework based on 50 synthetic datasets that simulate realistic variability in industrial environments, including drift, irregularities, and structural change. The agent’s performance was evaluated across multiple dimensions, including tool usage accuracy, reasoning consistency, and overall execution stability. While occasional misclassifications—especially in earlystage vi sual decisions—caused deviations, the agent demonstrated strong generalization across a wide range of scenar ios. The results indicate that automated orchestration of data science workflows is both practical and effective in manufacturing contexts, providing a foundation for more adaptive and scalable forecasting systems.