LLM-Powered Multi-Agent System for Automated Error Detection in Remanufacturing Inspection

The shift toward a circular economy and concepts such as the circular factory require inspection processes capable of handling high variability in product conditions and reliably assessing both surface quality and functional performance. While most automation approaches focus on visual surface defect detection, often powered by convolutional neural networks (CNNs), functional tests on test benches remain largely dependent on manual evaluation by experts. This paper introduces a novel approach that integrates Large Language Model (LLM) agents into the inspection process to automatically analyze data from functional tests. A multi-agent system (MAS) is employed to simulate operational states and manage data generation and coordination, while the LLM-based agent interprets functional test data to detect single and combined faults. Using an angle grinder as a representative case study, we evaluate the capability of this framework to classify complex defect patterns with a mean error rate below 30 %. Our approach complements traditional visual inspection by focusing on functional aspects, demonstrating the potential of combining MAS-based simulation with LLM reasoning for more intelligent and data-driven inspection strategies.