Few-shot working condition recognition of pumping units based on physics-informed prototype network

Abstract In the context of intelligent oilfield operation and maintenance, the accurate recognition of pumping unit working conditions is crucial for ensuring production safety and efficiency. However, practical scenarios face challenges such as scarce labeled samples, imbalanced fault categories, and limited computational resources on edge devices, which constrain the application of deep learning methods. To address this, this paper proposes a lightweight working condition diagnosis method that integrates physical prior knowledge and few-shot learning—the physics-informed prototypical network (PIPK). This method builds a few-shot classification framework based on the prototypical network and embeds, as hard constraints into the feature learning process, the Gibbs one-dimensional wave equation describing the dynamic behavior of the sucker rod system and the mechanical model of polished rod load, thereby enhancing the model’s physical consistency and generalization capability. To further meet the requirements for edge deployment, an improved knowledge distillation architecture is designed to transfer knowledge from the complex teacher model containing physical information to a lightweight MobileNetV2 student network. Experiments on a real indicator diagram dataset show that PIPK achieves an average accuracy of 95.07% and an F 1 score of 94.46% using only 5 support samples per class, significantly outperforming existing few-shot learning methods.