Leveraging Large Language Models for Automated Insight Generation in Textile Manufacturing

Abstract

The textile manufacturing industry generates vast amounts of operational data, yet deriving actionable insights remains a manual and labor-intensive process due to the semantic gap between quantitative sensor logs and qualitative operator knowledge. While Artificial Intelligence (AI) has matured in the domain of defect detection via Computer Vision, current systems lack the reasoning capabilities to explain root causes or correlate visual defects with operational parameters. This dissertation proposes a Software Engineering (SE) architecture leveraging Large Language Models (LLMs) to automate insight generation in the knitting sector. A multi-agent system using Retrieval-Augmented Generation (RAG) and task-orchestration patterns is designed and implemented to abstract raw machine telemetry into human-readable narratives. Furthermore, specific software engineering challenges are addressed, including context-aware prompting for complex industrial schemas, deterministic metric aggregation, hallucination mitigation through execution-layer constraints, and secure architectural integration via the Model Context Protocol (MCP). The prototype is evaluated through automated tests for the Astra and MCP services and through a controlled comparison of seven OpenAI models generating the same daily factory report, measuring cost, runtime, token usage, completeness, and insight quality. The results show a functional and test-covered foundation for asynchronous daily report generation, while large-scale industrial validation remains future work.