Predicting Injection Moulded Plastic Fittings Quality Using Large Language Models

Abstract

Maintaining consistent quality in plastic injection moulding is crucial for product functionality and profitability. Traditional quality control methods often rely on manual inspections, are prone to errors and inefficiency, and result in defects and financial losses. This research explores the potential of large language models (LLMs) for real-time, data-driven quality assessments in plastic injection moulding. A Taguchi experiment design was conducted to optimise prompting techniques for an LLM, focusing on the accuracy, reliability, and interpretability of predictions. The experiments systematically evaluated the influence of data augmentation, reasoning methods, temperature settings, and the LLM’s personality and role in prediction quality. Results revealed that a specific combination of prompting techniques, including a knowledge base, chain-of-thought reasoning, and a defined LLM personality, significantly improved the LLM’s predictive accuracy, consistently yielding more than 75% prediction accuracy. This paper demonstrates the feasibility of using LLMs for real-time quality control in plastic injection moulding, offering a pathway to reduce scrap, enhance quality, and improve production efficiency.