Intelligent Diagnosis System for Oil Well Underground Conditions Based on Convolutional Neural Network

The existing pumping unit downhole working condition diagnosis system has a high false alarm rate and a low accuracy rate of diagnosis for complex working conditions and abnormal working conditions. To address this problem, a diagnostic system of pumping unit workings is developed. First of all, the displacement-load data of the workover diagrams were converted into images, and then, through preliminary screening, manual review and data balancing from hundreds of millions of workover diagrams accumulated over the years, a sample database of 28 types of workover conditions, such as normal production, insufficient fluid supply, gas influence, rod breakage and tubing leakage, was established, with a total of about 760,000 samples, to compile a data set that is leading in quality and quantity in China. The project adopts “graphic + data” composite diagnosis, “graphic” corresponds to the power diagram, “data” refers to the electrical parameters, set pressure and other production parameters, and transforms the fault diagnosis problem of the power diagram into a deep learning-based image classification problem. Deep learning based image classification problem. A fault diagnosis method based on migration learning and category imbalance loss is designed. Better diagnostic results are obtained, with the single diagnostic accuracy of no less than 98% for common working conditions, 99.5% for normal production, 98.4% for insufficient fluid supply, and 97.2% for gas influence.