Comparative Analysis and Suggestions for Sucker Rod Pump Working Condition Diagnosis Using Machine Learning Techniques

Abstract This study explores the application of six state-of-the-art image classification models—ResNet, MobileNet, EfficientNet, ViT, EfficientViT, and MobileViT—in diagnosing sucker rod pump faults using dynamometer cards. The models, fine-tuned from pre-trained models developed on ImageNet-1k, are evaluated on a dataset of 11,941 dynamometer cards. Our findings demonstrate a significant inverse correlation between model accuracy and the number of trainable parameters, notably with EfficientViT achieving the highest accuracy (95.73%) despite having fewer parameters. This suggests that simpler, but more efficient models can outperform more complex architectures in specific tasks like SRP diagnosis. Through Grad-CAM visualization, we analyzed model predictions to identify critical features, highlighting ResNet's superior generalization ability and robustness. This work provides insights into model selection and data preparation, aiming to enhance the reliability and efficiency of SRP working condition diagnosis models.