AI Smart Jig Detects Micro Defects in 2.79 Seconds

Abstract

In the mass production of metal-based products such as automobiles, continuous welding and assembly processes are essential. The final product is created through multiple stages of welding, and the cumulative misalignment at each stage can lead to excessive residual stresses or dimensional defects in the product. To compensate for these issues, design modifications or significant post-processing costs have been required. Traditional dimensional inspection methods, whether manual or automated, are limited in their ability to keep pace with the speed required for mass production, as they focus on point-by-point measurements. While 3D vision-based methods offer a solution, they are often costly and primarily suited for macro-scale inspections. Here, we propose a machine learning-powered smart jig that enables precise, micro-level dimensional quality monitoring during production, without interrupting the continuous manufacturing process. This method, designed for direct integration into continuous assembly welding lines, reduces inspection time from 12 min to 2.79 s, enabling the detection of dimensional errors at the 500 μm level. Demonstrations conducted on the production line at a commercial automobile manufacturer confirm the feasibility of this approach for comprehensive subassembly inspections during mass production. This system is expected to be highly adaptable for various manufacturing domains utilizing assembly jigs, offering transformative potential in quality inspection processes.

A research team affiliated with UNIST has unveiled an innovative, high-precision AI-powered quality inspection system that reduces inspection time from 12 minutes to just under 3 seconds. This cutting-edge technology facilitates real-time, micron-scale defect detection during continuous manufacturing processes, paving the way for fully automated, inline quality control in high-speed production environments.

According to the research team, led by Professor Im Doo Jung in the Department of Mechanical Engineering at UNIST, the reported system employs a machine learning-powered smart jig that enables micro-level dimensional quality monitoring during production without disrupting the ongoing manufacturing process. By integrating advanced anomaly detection algorithms with 3D-printed sensor caps, the technology can identify minute dimensional errors within seconds, greatly enhancing inspection efficiency and accuracy.

Dimensional mismatches-such as micro-level gaps between assembled parts-can undermine product strength and quality, often stemming from manufacturing tolerances or deformation during handling. These defects are difficult to correct after assembly, making early detection vital to prevent costly rework, delays, or recalls.

During assembly, the smart jig precisely holds components in position. A specially designed clamp, fitted with a soft 3D-printed sensor cap on its contact surface, subtly deforms in response to the component's surface features. This deformation pattern is analyzed by AI algorithms to accurately detect anomalies, enabling instant assessment of whether a part has a dimensional defect.

Where traditional inspection methods took approximately 12 minutes per component, this new system completes inspection in just 2.79 seconds. It can detect defects as small as 500 micrometers, ensuring high-resolution quality control that operates seamlessly within automated production lines without halting operations.

Detected defects are visualized through easy-to-interpret heatmaps, allowing operators to quickly identify defect locations and severity levels, thereby facilitating immediate corrective measures and reducing rework time.

The AI model is trained exclusively on data from defect-free, normal products, eliminating the need for extensive defect datasets or manual labeling-common hurdles in manufacturing environments. Its low maintenance requirements and scalable design enable adaptation across diverse industries, including mobility, electronics, aerospace, and consumer appliances.

Professor Jung states, "This technology can be applied across high-precision assembly sectors, significantly reducing inspection personnel and time while boosting overall quality reliability. The system has the potential to save hundreds of millions of won annually by minimizing defects and streamlining inspection processes."

Led by Seobin Park and Taekyeong Kim, who served as first authors, this research will be published in the Journal of Manufacturing Systems (Impact Factor 14.2, JCR top 1%) online on July 10, 2025, prior to print.

This research was led by Seobin Park and Taekyeong Kim, who served as first authors. The findings are scheduled for publication in the Journal of Manufacturing Systems (Impact Factor 14.2, JCR top 1%) online on July 10, 2025, ahead of print.

Journal Reference

Seobin Park, Taekyeong Kim, Kyeong Min Kim, et al., "Quick dimensional inspection for continuous welding and assembly using machine learning-powered smart jig," J. Manuf. Syst., (2025).