Anomaly Detection in Automotive Welding Systems using FedMKT

Abstract In modern automotive manufacturing, ensuring the reliability of high-speed robotic welding lines is a critical challenge. Frequent issues like tool wear, thermal drift, and mechanical imbalances often result in defective welds and costly production downtime. Traditional maintenance methods are often reactive and fail to detect early-stage faults. This paper proposes a novel Edge-LLM-based predictive maintenance framework that provides a distributed, intelligent, and self-adaptive solution for real-time failure prediction. The system integrates Smart Local Models (SLMs) at the edge, monitoring sensor modalities like grip pressure, temperature, and vibration, with a central Large Language Model (LLM) for decision-making. The SLMs perform localized anomaly detection and transmit only processed event data to the central LLM, which then aggregates signals, correlates patterns, and performs root cause analysis. This hybrid architecture enables the proactive scheduling of maintenance, enhancing tool lifespan, eliminating unplanned downtime, and improving overall productivity and quality in smart manufacturing environments. Keywords Predictive Maintenance, Edge Intelligence, Industrial AI, Large Language Models, Smart Manufacturing, Robotic Welding System