Enhanced real-time defect detection for ceramic bearing balls using advanced feature extraction and optimization

Abstract Ceramic bearing balls, characterized by brittleness and low crack propagation resistance, require rigorous quality control during manufacture to ensure reliability and durability. To address the limitations of conventional image-processing methods in detecting intricate and variable surface defects, this study proposes an improved YOLOv11-based defect detection method tailored for ceramic bearing balls. Leveraging its advanced feature extraction capabilities and innovative network architecture, YOLOv11 achieves an optimal balance between real-time performance and high detection accuracy. The proposed model, named cross-stage switchable layer aggregation with You Only Look Once (CSLA-YOLO), incorporates a novel feature extraction module combining a general efficient layer aggregation network with a reparameterization-based switchable cross-stage partial module. Additionally, an optimized loss function was developed to align with the unique characteristics of the ceramic ball defect dataset. The experimental results demonstrate that CSLA-YOLO surpasses the baseline YOLOv11 model, achieving a 1.9% increase in average precision (AP) at an intersection over union (IoU) threshold of 0.5 (mAP 50 ), a 2.9% improvement in mean AP across IoU thresholds from 0.5 to 0.95 (mAP 50–95 ), and a 2.5% boost in total precision. Moreover, CSLA-YOLO outperforms mainstream models such as YOLOv5 and YOLOv8 in performance. When integrated into a custom detection platform, CSLA-YOLO achieves a detection speed of 40 balls per minute while accurately identifying defects as small as 0.1 mm. These results demonstrate its ability to meet the high-precision, real-time detection requirements of ceramic ball manufacturers, offering a robust and efficient solution for high-quality production control.