Drill tool recognition and detection with SERep-CCNet: A lightweight model approach

Abstract

With the advancement of digital transformation in oilfields, the demand for automation in drilling tool management has become increasingly prominent, serving as a critical driver for improving operational efficiency and achieving refined management. In this context, intelligent inspection robots have been preliminarily deployed in some oilfield management applications. Despite significant progress in object detection algorithms for tool recognition and positioning, the implementation of these technologies on low-power edge devices remains constrained by limited computational resources. This poses challenges for achieving both efficient and accurate tool classification and positioning in resource-constrained environments, thereby limiting the overall level of operational automation. To address this issue, this study proposes an automatic identification method for lightweight drilling tools based on SERep-CCNet. The method is designed to optimize computing resource consumption while maintaining detection accuracy, facilitating efficient deployment on edge devices. Firstly, Squeeze and Excitation Network Version 2 (SENetV2) is integrated into the backbone model to enhance its ability to capture inter-channel dependencies by incorporating multi-layer perceptrons (MLPs), thereby improving feature representation while maintaining high parameter efficiency. Secondly, the high-speed prediction head (RepHead) module is utilized to reconfigure the You Only Look Once (YOLO) head, leveraging the predictive structural advantages of the YOLOX model to improve inference speed, making it particularly suitable for detecting fast-moving objects. Lastly, a cross-scale feature fusion (CCFF) module is introduced to further reduce the network's parameter count and computational complexity without compromising detection accuracy. The experimental results indicate that the lightweight automatic identification method for drilling tools based on SERep-CCNet significantly outperforms the baseline network YOLOv8n. The proposed method reduces the number of parameters (Params) to 1.77M and floating-point operations (FLOPs) to 5.5G, making it highly efficient for edge computing applications. On a custom drilling tool dataset specifically designed for this study, the model achieves a mean Average Precision (mAP) of 69.3 % at IoU thresholds ranging from 0.5 to 0.95 ([email protected]:.95), representing an 11.1 % improvement over YOLOv8n. These results highlight the effectiveness and robustness of SERep-CCNet in addressing the challenges of drilling tool detection in resource-constrained environments.