A Robust Deep Learning Framework for Real-Time Detection of Cavings in Water-Based Mud

Summary Real-time monitoring of cavings is critical for preventing borehole instability and ensuring drilling safety. Unlike previous studies that only focus on detecting cleaned cavings, this study aims to directly detect cavings at the discharge end of the shale shaker. However, this presents numerous challenges for object detection researchers, including harsh field conditions, heavy mud coating on cavings, low target-to-background contrast, high-speed motion blur, and variable illumination. To address these challenges, we propose YOLOv12-CavDet, a novel deep learning architecture for real-time cavings detection. The network builds upon the YOLOv12 framework and introduces a series of synergistic enhancements tailored for complex drilling environments. Specifically, an adaptive multiscale fusion (AMFusion) module is incorporated to alleviate feature confusion caused by low contrast and motion blur. In parallel, a contextual self-calibrated attention (CSCAttn) mechanism is designed to robustly extract salient features from multiscale, irregularly shaped cavings. Furthermore, a corner-aware localization loss (CALLoss) is developed to improve the bounding box accuracy of targets with indistinct boundaries. Experimental evaluations on a custom data set collected from an active drilling site demonstrate that YOLOv12-CavDet outperforms several state-of-the-art object detection models, achieving a precision exceeding 91.5%. Notably, YOLOv12-CavDet demonstrates remarkable robustness under extreme environmental conditions, such as dense fog and low-light scenarios, consistently achieving high-confidence detection results. This research provides a practical solution for the real-time, automated monitoring of wellbore stability, thereby enhancing the safety and efficiency of drilling operations.