Detecting Prevalent Faults in Sucker Rod Pumps Through Computer Vision, Deep Learning, and Dynamometer Analysis

Abstract

Listen

Translate article

Abstract This research explores the integration of dynamometer cards, computer vision, and deep learning techniques to revolutionize the monitoring and diagnosis of oil wells in artificial lift systems, particularly those utilizing sucker rod pumps (SRP). The dynamometer card, also known as the dynamometer card, provides essential insights into the performance of a downhole pump by visualizing the interaction between the load on the sucker rod and the position of the pump. Our approach goes beyond traditional dynamometer card analysis by leveraging advanced computer vision and deep learning algorithms to automate and enhance card interpretation. This combination holds great potential for improving the accuracy of identifying issues such as fluid pound, gas interference, and tubing leaks, while also enabling real-time monitoring and predictive maintenance. At the core of our methodology is the use of image-based analysis rather than relying solely on textual data, which is often not readily available. By harnessing state-of-the-art deep learning techniques in computer vision, computers can utilize the ability to "see" like humans, but with greater speed and accuracy, and to generate timely alerts for critical interventions. The development of the methodology involved a comparison with the traditional deep learning approach commonly used for fault detection in dynamometer card data. This traditional approach typically involves sampling and unifying steps. However, a major challenge with this approach is the variability and availability of suitable data, which can vary across different companies, particularly larger ones. Our proposed strategy addresses this challenge by utilizing the YOLO (You Only Look Once) library, which is renowned for its real-time object detection and classification capabilities in computer vision. This approach combines theoretical data with actual operational data by initially training the algorithm with more than 16 distinct fault shapes from dynamometer cards. To improve the model's performance when dealing with low-quality real-world data, various data augmentation techniques are incorporated, such as image rotation, scaling, flipping, cropping, brightness adjustments, noise injection, and elastic transformations. Following the augmentation and training process, the methodology culminates in the creation of diverse models that function as a ‘mixture of experts’, each corroborating the others’ findings. This approach shifts from a deterministic to a probabilistic one, relying on the combined results of these models to make more nuanced and accurate predictions. To make these advanced analytical capabilities accessible to a broader range of users, including those without programming skills, the models have been implemented through an API. This API is the backbone of a user-friendly web application, enabling power users to leverage the sophisticated analysis of the computer vision model. Obtaining promising results that highlight the effectiveness of our methodology, the testing process revealed an impressive accuracy rate range of 75-94% in fault detection from dynamometer card images. This exceptional level of precision remained consistent for diverse types of faults.