A lightweight bearing defect detection model suitable for industrial scenarios

Bearings play a crucial role in mechanical systems, but surface defects can severely impact their lifespan and reliability, making defect detection vital for safe operations. However, existing bearing defect detection models suffer from issues in complex industrial settings, including high model complexity, significant computational resource consumption, and limited capability in identifying multi-directional, small-scale defects. To address these issues, this paper proposes a lightweight defect detection model for bearings, named LARD-YOLOv8, based on the YOLOv8n architecture. The model features a LiteShiftHead detection head with SPConv, REG, and CLS modules for efficient feature extraction and accurate classification regression while keeping the model lightweight. The ARConv module enhances adaptability to multi-directional defects through a convolutional kernel rotation mechanism and dynamic weight adjustment. The RepNCSPELAN4 module’s reparameterization technique further optimizes computational efficiency. Additionally, the Inner-DIoU loss function, with its dynamic adjustment of auxiliary bounding boxes, improves localization accuracy and convergence speed. Experimental results demonstrate that LARD-YOLOv8 achieves 96.3 % accuracy, 96.2 % recall, 98.4 % mAP0.5, and 71.1 % mAP0.5:0.95 on the bearing defect dataset (BR-DET). representing improvements of 2.4 %, 4.2 %, 2.1 %, and 6.0 % respectively over YOLOv8n. Concurrently, the model reduces parameter count by 19.5 % and computational load by 13.4 %, while maintaining a real-time detection speed of 89 FPS, meeting industrial inspection timeliness requirements. Moreover, compared to mainstream models such as YOLOv11n and YOLOv12n, LARD-YOLOv8 demonstrates significant advantages across all performance metrics. Cross-domain validation on public datasets including Northeastern University’s NEU-DET and DST-DET further confirms that this model possesses excellent generalisation capabilities and robustness while maintaining high accuracy, effectively meeting the demands of industrial real-time detection.