BED-YOLO: An Enhanced YOLOv8 for High-Precision Real-Time Bearing Defect Detection

Abstract

In industrial production, precise detection of bearing defects is crucial for optimal machinery performance and maintenance, directly impacting the efficiency of industrial systems and the field of instrumentation and measurement. To tackle the diverse types and unique characteristics of bearing defects, we introduce an approach for identifying defects in bearings, named bearing enhanced detection you only look once (BED-YOLO), which is based on convolutional neural networks (CNNs). We propose the intelligent feature concentration (IFC) module, a lightweight adaptive downsampling technique that exploits the attention mechanism to accurately control the feature compression process, prioritizing the retention of key features through the generation and normalization of spatial attention maps. Additionally, we design the efficient feature fusion for scalable convolution (EFFSC) module to capture and fuse multiscale features through convolution kernels of different sizes and optimize the computational efficiency using grouped convolution, which significantly improves the model expressiveness and processing speed. To ensure the robustness and reliability of our model, we conducted k-fold cross-validation on our BRG-dataset, which allowed us to thoroughly evaluate the model’s performance and ensure its generalizability. The experimental results show that the BED-YOLO model demonstrates an excellent balance between performance and efficiency. The model achieves a mean average precision (mAP50) of 92.5%. Moreover, the model maintains high efficiency with a computational demand of only 7.7 GFLOPs and achieves processing speeds of 312.5 frames/s, while requiring only 2.5M parameters. These results highlight our model’s superiority in speed and accuracy, making it particularly suitable for real-time applications that require rapid and precise detection, and well-equipped to meet the rigorous demands of industrial defect detection. Further tests on the MS COCO dataset underscore the model’s remarkable adaptability and accuracy. Access to the methodology’s code is provided through GitHub at https://github.com/YOLO-dennis/BED-YOLO .