YOLO-RSMA: Enhanced wear particle detection in lubricating oil using adaptive attention mechanisms and second-order optimization

Abstract

Online detection of wear particles in marine machinery lubricants is critical for effective condition monitoring. However, current methods encounter significant difficulties in accurately identifying microscale wear debris owing to the high-frequency background noise and motion blur inherent in online ferrographic images. To address these challenges, this study proposes YOLO-RSMA, a novel detection framework based on YOLOv8 architecture. The model systematically integrates a Residual Efficient Multi-scale Attention (REMA) mechanism to retrieve global contextual features that are frequently attenuated in deep convolutional layers. Furthermore, the framework employs the Sophia optimizer to mitigate gradient instability within nonconvex loss landscapes, and incorporates the MPDIoU loss function to ensure precise geometric alignment for irregular particle morphologies. Validated on a custom dataset against industrial baselines and state-of-the-art methods, YOLO-RSMA demonstrated superior robustness under complex imaging conditions. The model achieves a mAP0.5 of 99.30%, representing a substantial improvement of 16.63% over the baseline, thereby offering an efficient and accurate solution for intelligent fault diagnosis.