Integración de un modelo de inteligencia artificial llm con enfoque rag para dar soporte al mantenimiento, operación y diagnóstico de la planta de procesos análogos (ppa)

Abstract

This undergraduate thesis presents the design, implementation, and validation of a conversational assistant that supports the maintenance, operation, and diagnosis of the Analog Process Plant (PPA) at Universidad Distrital Francisco José de Caldas, a didactic three-tank system interconnected with an Allen-Bradley CompactLogix PLC. As is the case in many plants with years of use, its technical documentation (manuals, practice logs, tag dictionaries, and schematics) has accumulated over different periods, making manual consultation time-consuming, especially for first-time users. The assistant combines a Large Language Model (LLM) with Retrieval-Augmented Generation (RAG): it answers technical questions by citing the source document, uses tool calling to query the actual sensor history captured from the PLC via EtherNet/IP, and generates maintenance alerts traceable to the documentation. The entire system runs locally on a computer without a GPU. It was validated using a custom dataset of 145 questions: overall accuracy increased from 0.29 to 0.624 using a three-level rubric, and from 0.618 to 0.723 using a six-level rubric after a second improvement cycle, while tool selection exceeded 84% from the first version.