Research on Deep Learning-Based Diagnosis of Dynamometer Cards for Oil Wells

Dynamometer cards are a crucial basis for evaluating the working conditions of pumping units, and deep learning has become the mainstream method for dynamometer card diagnosis. This paper first performs a series of preprocessing steps on the dynamometer card dataset, including normalization, Gaussian filtering, and binarization, to enhance the contrast between the inner and outer regions of the load-displacement curve and make the features of the dynamometer cards more prominent. Subsequently, three models, Residual Network (ResNet), Vision Transformer (ViT), and Efficient Network (EfficientNet), are employed for dynamometer card diagnosis. Finally, a simple visualization interface is designed to make the recognition results more intuitive. Experimental results show that image preprocessing accelerates model convergence and improves prediction accuracy. The ResNet model achieves the highest diagnosis accuracy of 99.8%, while the ViT model has the smallest parameter count and memory usage. The study demonstrates that all three models can rapidly and accurately diagnose dynamometer cards. Additionally, filling the images effectively addresses issues such as model non-convergence caused by insufficient data for certain fault types and imbalanced sample distribution, proving the practical value of this approach.