Enhanced Lightweight Bearing Defect Detection via Frequency Domain Analysis and Model Compression

Bearing defect detection is crucial for fault diagnosis and preventive maintenance of industrial equipment. Existing models often struggle to balance accuracy with computational efficiency. In this study, we propose an improved lightweight YOLO-FCMP model based on YOLOv7-tiny, which effectively addresses this trade-off. By proposing the FSC (FrFT-SpatialAttention-Conv) module, we enhance feature representation in the frequency domain, enabling the model to capture both local and global features of bearing surface defects with higher accuracy. Additionally, deformable convolution (DCNv2) is integrated to capture geometric deformations and complex shapes. We also present a novel CAMS attention mechanism, which improves upon the CBAM mechanism by incorporating multi-scale convolutional attention, mitigating the issue of shared weights in spatial attention. Further optimizations include the Diverse Branch Block (DBB) for re-parameterization and the lightweight VoVGSCSP module centered around GSConv convolution, which reduce computational complexity while maintaining high accuracy. We propose the Inner-MPDIoU loss function to improve bounding box regression accuracy and convergence speed. Model compression techniques, such as pruning and knowledge distillation, significantly reduce computational requirements, resulting in a model with an mAP of 99.4% and a computational cost of only 4.6 GFLOPs. This work presents an efficient and precise solution for bearing defect detection in industrial applications.The source code and dataset of our proposed method are available at:https://github.com/wudiw295/Bearing-Defect-Detection.git