A Comprehensive Framework for Automated Quality Control in the Automotive Industry

TL;DR

The work tackles the need for reliable, fast automated quality control in automotive manufacturing by developing a robotic inspection framework that uses two cobots, high-resolution vision, and optimized lighting to detect surface and thread defects on aluminum HPDC components. A dual defect-detection pipeline combines YOLO11n with SAHI and ensemble strategies to improve small-defect detection and reduce false positives, complemented by a defect-size measurement module calibrated against a data matrix. Key contributions include robust surface-and-thread inspection, advanced detection fusion, and precise defect sizing, demonstrated with real-time performance and scalable deployment prospects across production environments. The results show high detection accuracy, especially with SAHI-ensemble methods, and practical defect sizing to support automated sorting decisions, highlighting potential to enhance reliability and throughput in automotive QC.

Abstract

This paper presents a cutting-edge robotic inspection solution designed to automate quality control in automotive manufacturing. The system integrates a pair of collaborative robots, each equipped with a high-resolution camera-based vision system to accurately detect and localize surface and thread defects in aluminum high-pressure die casting (HPDC) automotive components. In addition, specialized lenses and optimized lighting configurations are employed to ensure consistent and high-quality image acquisition. The YOLO11n deep learning model is utilized, incorporating additional enhancements such as image slicing, ensemble learning, and bounding-box merging to significantly improve performance and minimize false detections. Furthermore, image processing techniques are applied to estimate the extent of the detected defects. Experimental results demonstrate real-time performance with high accuracy across a wide variety of defects, while minimizing false detections. The proposed solution is promising and highly scalable, providing the flexibility to adapt to various production environments and meet the evolving demands of the automotive industry.

Figures (7)

• Figure 1: The developed quality control solution for inspecting surfaces and threads of aluminum HPDC automotive components. Each system consists of (1) a HD camera-based vision system, (2) a specialized lighting setup, and (3) a collaborative robot (cobot). a) The experimental setup fo surface inspection, and b) the experimental setup for thread inspection. c) The rear, front, left, right, and bottom views (five of the six sides) of the HPDC component. The front view features areas with wide cavity that are included in the inspection process.
• Figure 2: Detections from two models. (a) Models trained on full-size images excel at detecting small, subtle defects, but struggle to detect large defects in slices. (b) Models trained solely on slices may misclassify holes, as defects in full-size images.
• Figure 3: The inspection workflow begins with the scanning of the aluminum HPDC component, followed by defect detection and measurement to assess the size and severity of any detected defect. The final step is the classification module, where the part is classified as defective if any considerable defect is found.
• Figure 4: (a) mAP50 and (b) mAP50-95 curves for YOLO11n in surface defect detection. After approximately epoch $125$, the metrics start to converge, indicating that further training does not yield significant improvements in model's performance.
• Figure 5: Comparison of surface defect detection across different training settings. The captured images are of high quality, with no reflections, proper lighting, and clear visibility of the defects. In SAHI-V1, the slicing technique is utilized, and only defective images are used during training. In SAHI-V2, non-defective images are incorporated into the training process. In SAHI-V3 and SAHI-V4, non-defective images of black stains and internal marks are utilized during training, respectively. Finally in Ensemble, only the common detections of SAHI-V3 and SAHI-V4, which are trained on different subsets of non-defective images, are retained. The Ensemble model successfully detects all defects correctly in these examples.