Deep Learning Approach Based on Generative Adversarial Network for Automatic Diagnosis of Sucker-Rod Pump Operational Conditions

Abstract Early and accurate detection of operational anomalies in sucker-rod pumping wells is crucial. Maximizing production while minimizing downtime is essential. Traditionally, dynamometer (dyno) card classification has been labor-intensive and inefficient, especially with increasing high-frequency sensor data. To address this, a Semi-supervised Generative Adversarial Network (SSGAN) is proposed. It classifies sucker-rod pump conditions. The model uses both labeled and unlabeled dyno card data. It learns the underlying data distribution. Synthetic data is generated to improve the training set. This significantly enhances classification accuracy compared to traditional supervised Convolutional Neural Networks (CNNs), especially with limited data. SSGAN achieves an accuracy of 86% with limited labeled data and surpasses 92% accuracy with increased labeled data, outperforming the CNN model in both cases. By reducing the reliance on extensive manual labeling, this method offers a promising solution for improving the efficiency and effectiveness of sucker-rod pump monitoring and maintenance.