Detection of Pumping Unit in Complex Scenes by YOLOv7 with Switched Atrous Convolution

The petroleum and natural gas industries exhibit a high dependency on lifting equipment for oil and gas. Any malfunction in these devices can lead to severe economic losses. Therefore, continuous and timely monitoring of the status of pumping equipment is of paramount importance to proactively prevent potential issues. In an effort to enhance this monitoring process, this study delves into multi-source data images at the well site and extends traditional information analysis methods. It introduces an improved YOLOv7 method based on switchable atrous convolution. While the YOLOv7 algorithm achieves a balance between speed and accuracy, its robustness in non-standard environments is suboptimal. To address this limitation, we propose the utilization of a switchable atrous convolution method for enhancement, thereby augmenting the adaptability of the model. Images of pumping units from diverse scenarios are actively collected and utilized to construct training, validation, and test sets. Different models, including YOLOv7SAC, YOLOv7, and YOLOv5-n, undergo testing, and their detection performances are systematically compared in complex environments. Experimental findings demonstrate that YOLOv7SAC consistently attains optimal detection results across various scenes. In conclusion, the study suggests that the combination of the YOLOv7 model with switchable atrous convolution proves effective for detecting pumping unit equipment in complex scenarios. This provides robust theoretical support for the detection and identification of pumping equipment issues under challenging conditions.