Controlled assembly of random threads based on large language models

Abstract

In precision assembly scenarios such as aerospace and automotive engineering, the random starting positions of internal and external threads pose a significant challenge. While achieving specified tightening torque ranges is critical for sealing integrity, precisely controlling the final orientation of threaded connections remains difficult for varying thread pairings. This study proposes a framework integrating visual feature extraction with pre-trained large language models (LLMs) to enable controlled assembly of randomly aligned threads. Using the directional tightening process of hydraulic cylinder barrels and pipe fittings as a case study, the method’s feasibility is validated: First, computer vision techniques extract thread assembly features; then, servo-driven tightening devices perform directional tightening experiments on different fittings, with results recorded. Through structured prompt engineering, assembly parameters, visual features, and experimental outcomes are input into the LLM, the gasket thickness and thread phase are regarded as the controlled input variables, while the collaborative condition judgment of tightening torque and end orientation serves as the output variables. Results demonstrate that pre-trained LLMs, unlike traditional deep learning methods, not only adapt to raw data but also accurately predict directional tightening outcomes for randomly selected shims without requiring additional training. This work provides a novel approach for applying LLMs in precision assembly.