Automated Dynamometer Chart Pattern Recognition of Sucker Rod Pumps Using Convolutional Neural Network Approach

Artificial lift systems play a vital role in maximizing hydrocarbon production efficiency, particularly in mature oil fields where natural reservoir pressure is insufficient. Among various lifting methods, sucker rod pumps (SRPs) are widely employed due to their reliability and operational simplicity. However, these systems are prone to various mechanical and operational faults that can significantly reduce production efficiency and increase maintenance costs. Manual diagnosis based on dynamometer card interpretation remains time-consuming and prone to subjectivity. To address these limitations, this study was conducted to develop a data-driven fault detection model using deep learning and image classification techniques, with the primary objective of enhancing the accuracy and speed of SRP fault diagnosis. In this work, dynamometer card data were collected from operational oil wells and were labeled into 25 distinct SRP fault categories. The raw image data were preprocessed through normalization, resizing, and augmentation steps, which included noise injection and cropping to standardize the input format for the neural network. Grayscale images were used to simplify computation without compromising critical visual features. A custom Convolutional Neural Network (CNN) architecture was designed and trained using a stratified split, with 65% of the dataset allocated for training and 35% for testing. The network incorporated two convolutional and pooling blocks, followed by fully connected layers and dropout regularization to prevent overfitting. The model was trained over 150 epochs, using categorical cross-entropy loss and the Adam optimizer. The trained model demonstrated strong generalization capabilities on the testing dataset. A test accuracy of 89% was achieved across the 25 fault classes. Precision, recall, and F1-score values exceeded 0.90 in the majority of the fault types, with a macro-averaged F1-score of 0.89 and a weighted precision of 0.92. Furthermore, AUC scores reached 1.00 for 18 out of 25 classes, while the lowest recorded AUC remained above 0.93, confirming the model's strong discriminatory power. Performance curves showed rapid convergence within the first 30 epochs, with training loss dropping below 0.01, and validation accuracy stabilizing near 87–90%. This study presents a novel and practical approach for automating SRP fault classification using dynamometer cards. Deep learning has been effectively integrated with real-world operational data, providing a scalable and accurate diagnostic solution for artificial lift monitoring in the oil and gas industry.