Abstract
As one of the most extensively employed artificial lifting equipment in oilfield development, the pumping unit requires effective working condition recognition to ensure production stability and economic viability in oilfield operations.However, the scarcity of specific fault condition dynamometer cards samples in realworld production environments severely limits the recognition performance of traditional deep learning methods. To overcome this limitation, this paper develops a Hierarchical Contrastive Learning and Prototypical Network Few-Shot Recognition (HC-PFR) method for pumping units. The core of this method lies in integrating instance-level and category-level contrastive learning to significantly enhance feature discriminability through dual-layer feature optimization. Simultaneously, by incorporating a prototypical network learning mechanism, it effectively leverages limited samples to construct category representations, achieving precise few-shot classification. To strengthen the model's generalization capacity under data-scarce conditions, a physics-constrained data augmentation strategy is introduced, ensuring that augmented dynamometer card data retains physical characteristics consistent with real-world working conditions. In multiple N-way K-shot experiments, the proposed method achieved an accuracy of 92% in 5-way 1-shot scenarios and 97.33% in 5-way 5-shot scenarios. Experimental findings indicate that the proposed method successfully mitigates the data scarcity issue encountered in pumping unit working condition recognition, provides technical support for intelligent oilfield production, and offers a novel research direction for few-shot fault diagnosis of pumping units.
Funding Information
  • Yanshan University Basic Innovation Research Cultivation Project (2024LGQL001)
  • Hubei Engineering Research Center for Intelligent Detection and Identification of Complex Parts (IDICP-KF-2024-10)
  • Science Research Project of Hebei Province Higher Education (CXZX2025024)

This publication has 40 references indexed in Scilit: